fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/sched/signal.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/sort.h>
  24 #include <linux/rcupdate.h>
  25 #include <linux/kthread.h>
  26 #include <linux/slab.h>
  27 #include <linux/ratelimit.h>
  28 #include <linux/percpu_counter.h>
  29 #include <linux/lockdep.h>
  30 #include <linux/crc32c.h>
  31 #include "tree-log.h"
  32 #include "disk-io.h"
  33 #include "print-tree.h"
  34 #include "volumes.h"
  35 #include "raid56.h"
  36 #include "locking.h"
  37 #include "free-space-cache.h"
  38 #include "free-space-tree.h"
  39 #include "math.h"
  40 #include "sysfs.h"
  41 #include "qgroup.h"
  42 #include "ref-verify.h"
  43
  44 #undef SCRAMBLE_DELAYED_REFS
  45
  46 /*
  47  * control flags for do_chunk_alloc's force field
  48  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  49  * if we really need one.
  50  *
  51  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  52  * if we have very few chunks already allocated.  This is
  53  * used as part of the clustering code to help make sure
  54  * we have a good pool of storage to cluster in, without
  55  * filling the FS with empty chunks
  56  *
  57  * CHUNK_ALLOC_FORCE means it must try to allocate one
  58  *
  59  */
  60 enum {
  61         CHUNK_ALLOC_NO_FORCE = 0,
  62         CHUNK_ALLOC_LIMITED = 1,
  63         CHUNK_ALLOC_FORCE = 2,
  64 };
  65
  66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  67                                struct btrfs_fs_info *fs_info,
  68                                 struct btrfs_delayed_ref_node *node, u64 parent,
  69                                 u64 root_objectid, u64 owner_objectid,
  70                                 u64 owner_offset, int refs_to_drop,
  71                                 struct btrfs_delayed_extent_op *extra_op);
  72 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  73                                     struct extent_buffer *leaf,
  74                                     struct btrfs_extent_item *ei);
  75 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  76                                       struct btrfs_fs_info *fs_info,
  77                                       u64 parent, u64 root_objectid,
  78                                       u64 flags, u64 owner, u64 offset,
  79                                       struct btrfs_key *ins, int ref_mod);
  80 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  81                                      struct btrfs_fs_info *fs_info,
  82                                      u64 parent, u64 root_objectid,
  83                                      u64 flags, struct btrfs_disk_key *key,
  84                                      int level, struct btrfs_key *ins);
  85 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  86                           struct btrfs_fs_info *fs_info, u64 flags,
  87                           int force);
  88 static int find_next_key(struct btrfs_path *path, int level,
  89                          struct btrfs_key *key);
  90 static void dump_space_info(struct btrfs_fs_info *fs_info,
  91                             struct btrfs_space_info *info, u64 bytes,
  92                             int dump_block_groups);
  93 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  94                                u64 num_bytes);
  95 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  96                                      struct btrfs_space_info *space_info,
  97                                      u64 num_bytes);
  98 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  99                                      struct btrfs_space_info *space_info,
 100                                      u64 num_bytes);
 101
 102 static noinline int
 103 block_group_cache_done(struct btrfs_block_group_cache *cache)
 104 {
 105         smp_mb();
 106         return cache->cached == BTRFS_CACHE_FINISHED ||
 107                 cache->cached == BTRFS_CACHE_ERROR;
 108 }
 109
 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 111 {
 112         return (cache->flags & bits) == bits;
 113 }
 114
 115 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 116 {
 117         atomic_inc(&cache->count);
 118 }
 119
 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 121 {
 122         if (atomic_dec_and_test(&cache->count)) {
 123                 WARN_ON(cache->pinned > 0);
 124                 WARN_ON(cache->reserved > 0);
 125
 126                 /*
 127                  * If not empty, someone is still holding mutex of
 128                  * full_stripe_lock, which can only be released by caller.
 129                  * And it will definitely cause use-after-free when caller
 130                  * tries to release full stripe lock.
 131                  *
 132                  * No better way to resolve, but only to warn.
 133                  */
 134                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE);
 232         set_extent_bits(&fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE);
 247         clear_extent_bits(&fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 271                                        bytenr, 0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(fs_info, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (!cache->caching_ctl) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         ctl = cache->caching_ctl;
 320         refcount_inc(&ctl->count);
 321         spin_unlock(&cache->lock);
 322         return ctl;
 323 }
 324
 325 static void put_caching_control(struct btrfs_caching_control *ctl)
 326 {
 327         if (refcount_dec_and_test(&ctl->count))
 328                 kfree(ctl);
 329 }
 330
 331 #ifdef CONFIG_BTRFS_DEBUG
 332 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 333 {
 334         struct btrfs_fs_info *fs_info = block_group->fs_info;
 335         u64 start = block_group->key.objectid;
 336         u64 len = block_group->key.offset;
 337         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 338                 fs_info->nodesize : fs_info->sectorsize;
 339         u64 step = chunk << 1;
 340
 341         while (len > chunk) {
 342                 btrfs_remove_free_space(block_group, start, chunk);
 343                 start += step;
 344                 if (len < step)
 345                         len = 0;
 346                 else
 347                         len -= step;
 348         }
 349 }
 350 #endif
 351
 352 /*
 353  * this is only called by cache_block_group, since we could have freed extents
 354  * we need to check the pinned_extents for any extents that can't be used yet
 355  * since their free space will be released as soon as the transaction commits.
 356  */
 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 358                        struct btrfs_fs_info *info, u64 start, u64 end)
 359 {
 360         u64 extent_start, extent_end, size, total_added = 0;
 361         int ret;
 362
 363         while (start < end) {
 364                 ret = find_first_extent_bit(info->pinned_extents, start,
 365                                             &extent_start, &extent_end,
 366                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 367                                             NULL);
 368                 if (ret)
 369                         break;
 370
 371                 if (extent_start <= start) {
 372                         start = extent_end + 1;
 373                 } else if (extent_start > start && extent_start < end) {
 374                         size = extent_start - start;
 375                         total_added += size;
 376                         ret = btrfs_add_free_space(block_group, start,
 377                                                    size);
 378                         BUG_ON(ret); /* -ENOMEM or logic error */
 379                         start = extent_end + 1;
 380                 } else {
 381                         break;
 382                 }
 383         }
 384
 385         if (start < end) {
 386                 size = end - start;
 387                 total_added += size;
 388                 ret = btrfs_add_free_space(block_group, start, size);
 389                 BUG_ON(ret); /* -ENOMEM or logic error */
 390         }
 391
 392         return total_added;
 393 }
 394
 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 396 {
 397         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 398         struct btrfs_fs_info *fs_info = block_group->fs_info;
 399         struct btrfs_root *extent_root = fs_info->extent_root;
 400         struct btrfs_path *path;
 401         struct extent_buffer *leaf;
 402         struct btrfs_key key;
 403         u64 total_found = 0;
 404         u64 last = 0;
 405         u32 nritems;
 406         int ret;
 407         bool wakeup = true;
 408
 409         path = btrfs_alloc_path();
 410         if (!path)
 411                 return -ENOMEM;
 412
 413         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 414
 415 #ifdef CONFIG_BTRFS_DEBUG
 416         /*
 417          * If we're fragmenting we don't want to make anybody think we can
 418          * allocate from this block group until we've had a chance to fragment
 419          * the free space.
 420          */
 421         if (btrfs_should_fragment_free_space(block_group))
 422                 wakeup = false;
 423 #endif
 424         /*
 425          * We don't want to deadlock with somebody trying to allocate a new
 426          * extent for the extent root while also trying to search the extent
 427          * root to add free space.  So we skip locking and search the commit
 428          * root, since its read-only
 429          */
 430         path->skip_locking = 1;
 431         path->search_commit_root = 1;
 432         path->reada = READA_FORWARD;
 433
 434         key.objectid = last;
 435         key.offset = 0;
 436         key.type = BTRFS_EXTENT_ITEM_KEY;
 437
 438 next:
 439         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 440         if (ret < 0)
 441                 goto out;
 442
 443         leaf = path->nodes[0];
 444         nritems = btrfs_header_nritems(leaf);
 445
 446         while (1) {
 447                 if (btrfs_fs_closing(fs_info) > 1) {
 448                         last = (u64)-1;
 449                         break;
 450                 }
 451
 452                 if (path->slots[0] < nritems) {
 453                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 454                 } else {
 455                         ret = find_next_key(path, 0, &key);
 456                         if (ret)
 457                                 break;
 458
 459                         if (need_resched() ||
 460                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 461                                 if (wakeup)
 462                                         caching_ctl->progress = last;
 463                                 btrfs_release_path(path);
 464                                 up_read(&fs_info->commit_root_sem);
 465                                 mutex_unlock(&caching_ctl->mutex);
 466                                 cond_resched();
 467                                 mutex_lock(&caching_ctl->mutex);
 468                                 down_read(&fs_info->commit_root_sem);
 469                                 goto next;
 470                         }
 471
 472                         ret = btrfs_next_leaf(extent_root, path);
 473                         if (ret < 0)
 474                                 goto out;
 475                         if (ret)
 476                                 break;
 477                         leaf = path->nodes[0];
 478                         nritems = btrfs_header_nritems(leaf);
 479                         continue;
 480                 }
 481
 482                 if (key.objectid < last) {
 483                         key.objectid = last;
 484                         key.offset = 0;
 485                         key.type = BTRFS_EXTENT_ITEM_KEY;
 486
 487                         if (wakeup)
 488                                 caching_ctl->progress = last;
 489                         btrfs_release_path(path);
 490                         goto next;
 491                 }
 492
 493                 if (key.objectid < block_group->key.objectid) {
 494                         path->slots[0]++;
 495                         continue;
 496                 }
 497
 498                 if (key.objectid >= block_group->key.objectid +
 499                     block_group->key.offset)
 500                         break;
 501
 502                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 503                     key.type == BTRFS_METADATA_ITEM_KEY) {
 504                         total_found += add_new_free_space(block_group,
 505                                                           fs_info, last,
 506                                                           key.objectid);
 507                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 508                                 last = key.objectid +
 509                                         fs_info->nodesize;
 510                         else
 511                                 last = key.objectid + key.offset;
 512
 513                         if (total_found > CACHING_CTL_WAKE_UP) {
 514                                 total_found = 0;
 515                                 if (wakeup)
 516                                         wake_up(&caching_ctl->wait);
 517                         }
 518                 }
 519                 path->slots[0]++;
 520         }
 521         ret = 0;
 522
 523         total_found += add_new_free_space(block_group, fs_info, last,
 524                                           block_group->key.objectid +
 525                                           block_group->key.offset);
 526         caching_ctl->progress = (u64)-1;
 527
 528 out:
 529         btrfs_free_path(path);
 530         return ret;
 531 }
 532
 533 static noinline void caching_thread(struct btrfs_work *work)
 534 {
 535         struct btrfs_block_group_cache *block_group;
 536         struct btrfs_fs_info *fs_info;
 537         struct btrfs_caching_control *caching_ctl;
 538         int ret;
 539
 540         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 541         block_group = caching_ctl->block_group;
 542         fs_info = block_group->fs_info;
 543
 544         mutex_lock(&caching_ctl->mutex);
 545         down_read(&fs_info->commit_root_sem);
 546
 547         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 548                 ret = load_free_space_tree(caching_ctl);
 549         else
 550                 ret = load_extent_tree_free(caching_ctl);
 551
 552         spin_lock(&block_group->lock);
 553         block_group->caching_ctl = NULL;
 554         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 555         spin_unlock(&block_group->lock);
 556
 557 #ifdef CONFIG_BTRFS_DEBUG
 558         if (btrfs_should_fragment_free_space(block_group)) {
 559                 u64 bytes_used;
 560
 561                 spin_lock(&block_group->space_info->lock);
 562                 spin_lock(&block_group->lock);
 563                 bytes_used = block_group->key.offset -
 564                         btrfs_block_group_used(&block_group->item);
 565                 block_group->space_info->bytes_used += bytes_used >> 1;
 566                 spin_unlock(&block_group->lock);
 567                 spin_unlock(&block_group->space_info->lock);
 568                 fragment_free_space(block_group);
 569         }
 570 #endif
 571
 572         caching_ctl->progress = (u64)-1;
 573
 574         up_read(&fs_info->commit_root_sem);
 575         free_excluded_extents(fs_info, block_group);
 576         mutex_unlock(&caching_ctl->mutex);
 577
 578         wake_up(&caching_ctl->wait);
 579
 580         put_caching_control(caching_ctl);
 581         btrfs_put_block_group(block_group);
 582 }
 583
 584 static int cache_block_group(struct btrfs_block_group_cache *cache,
 585                              int load_cache_only)
 586 {
 587         DEFINE_WAIT(wait);
 588         struct btrfs_fs_info *fs_info = cache->fs_info;
 589         struct btrfs_caching_control *caching_ctl;
 590         int ret = 0;
 591
 592         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 593         if (!caching_ctl)
 594                 return -ENOMEM;
 595
 596         INIT_LIST_HEAD(&caching_ctl->list);
 597         mutex_init(&caching_ctl->mutex);
 598         init_waitqueue_head(&caching_ctl->wait);
 599         caching_ctl->block_group = cache;
 600         caching_ctl->progress = cache->key.objectid;
 601         refcount_set(&caching_ctl->count, 1);
 602         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 603                         caching_thread, NULL, NULL);
 604
 605         spin_lock(&cache->lock);
 606         /*
 607          * This should be a rare occasion, but this could happen I think in the
 608          * case where one thread starts to load the space cache info, and then
 609          * some other thread starts a transaction commit which tries to do an
 610          * allocation while the other thread is still loading the space cache
 611          * info.  The previous loop should have kept us from choosing this block
 612          * group, but if we've moved to the state where we will wait on caching
 613          * block groups we need to first check if we're doing a fast load here,
 614          * so we can wait for it to finish, otherwise we could end up allocating
 615          * from a block group who's cache gets evicted for one reason or
 616          * another.
 617          */
 618         while (cache->cached == BTRFS_CACHE_FAST) {
 619                 struct btrfs_caching_control *ctl;
 620
 621                 ctl = cache->caching_ctl;
 622                 refcount_inc(&ctl->count);
 623                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 624                 spin_unlock(&cache->lock);
 625
 626                 schedule();
 627
 628                 finish_wait(&ctl->wait, &wait);
 629                 put_caching_control(ctl);
 630                 spin_lock(&cache->lock);
 631         }
 632
 633         if (cache->cached != BTRFS_CACHE_NO) {
 634                 spin_unlock(&cache->lock);
 635                 kfree(caching_ctl);
 636                 return 0;
 637         }
 638         WARN_ON(cache->caching_ctl);
 639         cache->caching_ctl = caching_ctl;
 640         cache->cached = BTRFS_CACHE_FAST;
 641         spin_unlock(&cache->lock);
 642
 643         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 644                 mutex_lock(&caching_ctl->mutex);
 645                 ret = load_free_space_cache(fs_info, cache);
 646
 647                 spin_lock(&cache->lock);
 648                 if (ret == 1) {
 649                         cache->caching_ctl = NULL;
 650                         cache->cached = BTRFS_CACHE_FINISHED;
 651                         cache->last_byte_to_unpin = (u64)-1;
 652                         caching_ctl->progress = (u64)-1;
 653                 } else {
 654                         if (load_cache_only) {
 655                                 cache->caching_ctl = NULL;
 656                                 cache->cached = BTRFS_CACHE_NO;
 657                         } else {
 658                                 cache->cached = BTRFS_CACHE_STARTED;
 659                                 cache->has_caching_ctl = 1;
 660                         }
 661                 }
 662                 spin_unlock(&cache->lock);
 663 #ifdef CONFIG_BTRFS_DEBUG
 664                 if (ret == 1 &&
 665                     btrfs_should_fragment_free_space(cache)) {
 666                         u64 bytes_used;
 667
 668                         spin_lock(&cache->space_info->lock);
 669                         spin_lock(&cache->lock);
 670                         bytes_used = cache->key.offset -
 671                                 btrfs_block_group_used(&cache->item);
 672                         cache->space_info->bytes_used += bytes_used >> 1;
 673                         spin_unlock(&cache->lock);
 674                         spin_unlock(&cache->space_info->lock);
 675                         fragment_free_space(cache);
 676                 }
 677 #endif
 678                 mutex_unlock(&caching_ctl->mutex);
 679
 680                 wake_up(&caching_ctl->wait);
 681                 if (ret == 1) {
 682                         put_caching_control(caching_ctl);
 683                         free_excluded_extents(fs_info, cache);
 684                         return 0;
 685                 }
 686         } else {
 687                 /*
 688                  * We're either using the free space tree or no caching at all.
 689                  * Set cached to the appropriate value and wakeup any waiters.
 690                  */
 691                 spin_lock(&cache->lock);
 692                 if (load_cache_only) {
 693                         cache->caching_ctl = NULL;
 694                         cache->cached = BTRFS_CACHE_NO;
 695                 } else {
 696                         cache->cached = BTRFS_CACHE_STARTED;
 697                         cache->has_caching_ctl = 1;
 698                 }
 699                 spin_unlock(&cache->lock);
 700                 wake_up(&caching_ctl->wait);
 701         }
 702
 703         if (load_cache_only) {
 704                 put_caching_control(caching_ctl);
 705                 return 0;
 706         }
 707
 708         down_write(&fs_info->commit_root_sem);
 709         refcount_inc(&caching_ctl->count);
 710         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 711         up_write(&fs_info->commit_root_sem);
 712
 713         btrfs_get_block_group(cache);
 714
 715         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 716
 717         return ret;
 718 }
 719
 720 /*
 721  * return the block group that starts at or after bytenr
 722  */
 723 static struct btrfs_block_group_cache *
 724 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 725 {
 726         return block_group_cache_tree_search(info, bytenr, 0);
 727 }
 728
 729 /*
 730  * return the block group that contains the given bytenr
 731  */
 732 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 733                                                  struct btrfs_fs_info *info,
 734                                                  u64 bytenr)
 735 {
 736         return block_group_cache_tree_search(info, bytenr, 1);
 737 }
 738
 739 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 740                                                   u64 flags)
 741 {
 742         struct list_head *head = &info->space_info;
 743         struct btrfs_space_info *found;
 744
 745         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 746
 747         rcu_read_lock();
 748         list_for_each_entry_rcu(found, head, list) {
 749                 if (found->flags & flags) {
 750                         rcu_read_unlock();
 751                         return found;
 752                 }
 753         }
 754         rcu_read_unlock();
 755         return NULL;
 756 }
 757
 758 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 759                              u64 owner, u64 root_objectid)
 760 {
 761         struct btrfs_space_info *space_info;
 762         u64 flags;
 763
 764         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 765                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 766                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 767                 else
 768                         flags = BTRFS_BLOCK_GROUP_METADATA;
 769         } else {
 770                 flags = BTRFS_BLOCK_GROUP_DATA;
 771         }
 772
 773         space_info = __find_space_info(fs_info, flags);
 774         ASSERT(space_info);
 775         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 776 }
 777
 778 /*
 779  * after adding space to the filesystem, we need to clear the full flags
 780  * on all the space infos.
 781  */
 782 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 783 {
 784         struct list_head *head = &info->space_info;
 785         struct btrfs_space_info *found;
 786
 787         rcu_read_lock();
 788         list_for_each_entry_rcu(found, head, list)
 789                 found->full = 0;
 790         rcu_read_unlock();
 791 }
 792
 793 /* simple helper to search for an existing data extent at a given offset */
 794 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 795 {
 796         int ret;
 797         struct btrfs_key key;
 798         struct btrfs_path *path;
 799
 800         path = btrfs_alloc_path();
 801         if (!path)
 802                 return -ENOMEM;
 803
 804         key.objectid = start;
 805         key.offset = len;
 806         key.type = BTRFS_EXTENT_ITEM_KEY;
 807         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 808         btrfs_free_path(path);
 809         return ret;
 810 }
 811
 812 /*
 813  * helper function to lookup reference count and flags of a tree block.
 814  *
 815  * the head node for delayed ref is used to store the sum of all the
 816  * reference count modifications queued up in the rbtree. the head
 817  * node may also store the extent flags to set. This way you can check
 818  * to see what the reference count and extent flags would be if all of
 819  * the delayed refs are not processed.
 820  */
 821 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 822                              struct btrfs_fs_info *fs_info, u64 bytenr,
 823                              u64 offset, int metadata, u64 *refs, u64 *flags)
 824 {
 825         struct btrfs_delayed_ref_head *head;
 826         struct btrfs_delayed_ref_root *delayed_refs;
 827         struct btrfs_path *path;
 828         struct btrfs_extent_item *ei;
 829         struct extent_buffer *leaf;
 830         struct btrfs_key key;
 831         u32 item_size;
 832         u64 num_refs;
 833         u64 extent_flags;
 834         int ret;
 835
 836         /*
 837          * If we don't have skinny metadata, don't bother doing anything
 838          * different
 839          */
 840         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 841                 offset = fs_info->nodesize;
 842                 metadata = 0;
 843         }
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return -ENOMEM;
 848
 849         if (!trans) {
 850                 path->skip_locking = 1;
 851                 path->search_commit_root = 1;
 852         }
 853
 854 search_again:
 855         key.objectid = bytenr;
 856         key.offset = offset;
 857         if (metadata)
 858                 key.type = BTRFS_METADATA_ITEM_KEY;
 859         else
 860                 key.type = BTRFS_EXTENT_ITEM_KEY;
 861
 862         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 863         if (ret < 0)
 864                 goto out_free;
 865
 866         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 867                 if (path->slots[0]) {
 868                         path->slots[0]--;
 869                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 870                                               path->slots[0]);
 871                         if (key.objectid == bytenr &&
 872                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 873                             key.offset == fs_info->nodesize)
 874                                 ret = 0;
 875                 }
 876         }
 877
 878         if (ret == 0) {
 879                 leaf = path->nodes[0];
 880                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 881                 if (item_size >= sizeof(*ei)) {
 882                         ei = btrfs_item_ptr(leaf, path->slots[0],
 883                                             struct btrfs_extent_item);
 884                         num_refs = btrfs_extent_refs(leaf, ei);
 885                         extent_flags = btrfs_extent_flags(leaf, ei);
 886                 } else {
 887 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 888                         struct btrfs_extent_item_v0 *ei0;
 889                         BUG_ON(item_size != sizeof(*ei0));
 890                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 891                                              struct btrfs_extent_item_v0);
 892                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 893                         /* FIXME: this isn't correct for data */
 894                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 895 #else
 896                         BUG();
 897 #endif
 898                 }
 899                 BUG_ON(num_refs == 0);
 900         } else {
 901                 num_refs = 0;
 902                 extent_flags = 0;
 903                 ret = 0;
 904         }
 905
 906         if (!trans)
 907                 goto out;
 908
 909         delayed_refs = &trans->transaction->delayed_refs;
 910         spin_lock(&delayed_refs->lock);
 911         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 912         if (head) {
 913                 if (!mutex_trylock(&head->mutex)) {
 914                         refcount_inc(&head->refs);
 915                         spin_unlock(&delayed_refs->lock);
 916
 917                         btrfs_release_path(path);
 918
 919                         /*
 920                          * Mutex was contended, block until it's released and try
 921                          * again
 922                          */
 923                         mutex_lock(&head->mutex);
 924                         mutex_unlock(&head->mutex);
 925                         btrfs_put_delayed_ref_head(head);
 926                         goto search_again;
 927                 }
 928                 spin_lock(&head->lock);
 929                 if (head->extent_op && head->extent_op->update_flags)
 930                         extent_flags |= head->extent_op->flags_to_set;
 931                 else
 932                         BUG_ON(num_refs == 0);
 933
 934                 num_refs += head->ref_mod;
 935                 spin_unlock(&head->lock);
 936                 mutex_unlock(&head->mutex);
 937         }
 938         spin_unlock(&delayed_refs->lock);
 939 out:
 940         WARN_ON(num_refs == 0);
 941         if (refs)
 942                 *refs = num_refs;
 943         if (flags)
 944                 *flags = extent_flags;
 945 out_free:
 946         btrfs_free_path(path);
 947         return ret;
 948 }
 949
 950 /*
 951  * Back reference rules.  Back refs have three main goals:
 952  *
 953  * 1) differentiate between all holders of references to an extent so that
 954  *    when a reference is dropped we can make sure it was a valid reference
 955  *    before freeing the extent.
 956  *
 957  * 2) Provide enough information to quickly find the holders of an extent
 958  *    if we notice a given block is corrupted or bad.
 959  *
 960  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 961  *    maintenance.  This is actually the same as #2, but with a slightly
 962  *    different use case.
 963  *
 964  * There are two kinds of back refs. The implicit back refs is optimized
 965  * for pointers in non-shared tree blocks. For a given pointer in a block,
 966  * back refs of this kind provide information about the block's owner tree
 967  * and the pointer's key. These information allow us to find the block by
 968  * b-tree searching. The full back refs is for pointers in tree blocks not
 969  * referenced by their owner trees. The location of tree block is recorded
 970  * in the back refs. Actually the full back refs is generic, and can be
 971  * used in all cases the implicit back refs is used. The major shortcoming
 972  * of the full back refs is its overhead. Every time a tree block gets
 973  * COWed, we have to update back refs entry for all pointers in it.
 974  *
 975  * For a newly allocated tree block, we use implicit back refs for
 976  * pointers in it. This means most tree related operations only involve
 977  * implicit back refs. For a tree block created in old transaction, the
 978  * only way to drop a reference to it is COW it. So we can detect the
 979  * event that tree block loses its owner tree's reference and do the
 980  * back refs conversion.
 981  *
 982  * When a tree block is COWed through a tree, there are four cases:
 983  *
 984  * The reference count of the block is one and the tree is the block's
 985  * owner tree. Nothing to do in this case.
 986  *
 987  * The reference count of the block is one and the tree is not the
 988  * block's owner tree. In this case, full back refs is used for pointers
 989  * in the block. Remove these full back refs, add implicit back refs for
 990  * every pointers in the new block.
 991  *
 992  * The reference count of the block is greater than one and the tree is
 993  * the block's owner tree. In this case, implicit back refs is used for
 994  * pointers in the block. Add full back refs for every pointers in the
 995  * block, increase lower level extents' reference counts. The original
 996  * implicit back refs are entailed to the new block.
 997  *
 998  * The reference count of the block is greater than one and the tree is
 999  * not the block's owner tree. Add implicit back refs for every pointer in
1000  * the new block, increase lower level extents' reference count.
1001  *
1002  * Back Reference Key composing:
1003  *
1004  * The key objectid corresponds to the first byte in the extent,
1005  * The key type is used to differentiate between types of back refs.
1006  * There are different meanings of the key offset for different types
1007  * of back refs.
1008  *
1009  * File extents can be referenced by:
1010  *
1011  * - multiple snapshots, subvolumes, or different generations in one subvol
1012  * - different files inside a single subvolume
1013  * - different offsets inside a file (bookend extents in file.c)
1014  *
1015  * The extent ref structure for the implicit back refs has fields for:
1016  *
1017  * - Objectid of the subvolume root
1018  * - objectid of the file holding the reference
1019  * - original offset in the file
1020  * - how many bookend extents
1021  *
1022  * The key offset for the implicit back refs is hash of the first
1023  * three fields.
1024  *
1025  * The extent ref structure for the full back refs has field for:
1026  *
1027  * - number of pointers in the tree leaf
1028  *
1029  * The key offset for the implicit back refs is the first byte of
1030  * the tree leaf
1031  *
1032  * When a file extent is allocated, The implicit back refs is used.
1033  * the fields are filled in:
1034  *
1035  *     (root_key.objectid, inode objectid, offset in file, 1)
1036  *
1037  * When a file extent is removed file truncation, we find the
1038  * corresponding implicit back refs and check the following fields:
1039  *
1040  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1041  *
1042  * Btree extents can be referenced by:
1043  *
1044  * - Different subvolumes
1045  *
1046  * Both the implicit back refs and the full back refs for tree blocks
1047  * only consist of key. The key offset for the implicit back refs is
1048  * objectid of block's owner tree. The key offset for the full back refs
1049  * is the first byte of parent block.
1050  *
1051  * When implicit back refs is used, information about the lowest key and
1052  * level of the tree block are required. These information are stored in
1053  * tree block info structure.
1054  */
1055
1056 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1057 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1058                                   struct btrfs_fs_info *fs_info,
1059                                   struct btrfs_path *path,
1060                                   u64 owner, u32 extra_size)
1061 {
1062         struct btrfs_root *root = fs_info->extent_root;
1063         struct btrfs_extent_item *item;
1064         struct btrfs_extent_item_v0 *ei0;
1065         struct btrfs_extent_ref_v0 *ref0;
1066         struct btrfs_tree_block_info *bi;
1067         struct extent_buffer *leaf;
1068         struct btrfs_key key;
1069         struct btrfs_key found_key;
1070         u32 new_size = sizeof(*item);
1071         u64 refs;
1072         int ret;
1073
1074         leaf = path->nodes[0];
1075         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1076
1077         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1078         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1079                              struct btrfs_extent_item_v0);
1080         refs = btrfs_extent_refs_v0(leaf, ei0);
1081
1082         if (owner == (u64)-1) {
1083                 while (1) {
1084                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1085                                 ret = btrfs_next_leaf(root, path);
1086                                 if (ret < 0)
1087                                         return ret;
1088                                 BUG_ON(ret > 0); /* Corruption */
1089                                 leaf = path->nodes[0];
1090                         }
1091                         btrfs_item_key_to_cpu(leaf, &found_key,
1092                                               path->slots[0]);
1093                         BUG_ON(key.objectid != found_key.objectid);
1094                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1095                                 path->slots[0]++;
1096                                 continue;
1097                         }
1098                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1099                                               struct btrfs_extent_ref_v0);
1100                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1101                         break;
1102                 }
1103         }
1104         btrfs_release_path(path);
1105
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1107                 new_size += sizeof(*bi);
1108
1109         new_size -= sizeof(*ei0);
1110         ret = btrfs_search_slot(trans, root, &key, path,
1111                                 new_size + extra_size, 1);
1112         if (ret < 0)
1113                 return ret;
1114         BUG_ON(ret); /* Corruption */
1115
1116         btrfs_extend_item(fs_info, path, new_size);
1117
1118         leaf = path->nodes[0];
1119         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1120         btrfs_set_extent_refs(leaf, item, refs);
1121         /* FIXME: get real generation */
1122         btrfs_set_extent_generation(leaf, item, 0);
1123         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1124                 btrfs_set_extent_flags(leaf, item,
1125                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1126                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1127                 bi = (struct btrfs_tree_block_info *)(item + 1);
1128                 /* FIXME: get first key of the block */
1129                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1130                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1131         } else {
1132                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1133         }
1134         btrfs_mark_buffer_dirty(leaf);
1135         return 0;
1136 }
1137 #endif
1138
1139 /*
1140  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1141  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1142  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1143  */
1144 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1145                                      struct btrfs_extent_inline_ref *iref,
1146                                      enum btrfs_inline_ref_type is_data)
1147 {
1148         int type = btrfs_extent_inline_ref_type(eb, iref);
1149         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1150
1151         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1152             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1153             type == BTRFS_SHARED_DATA_REF_KEY ||
1154             type == BTRFS_EXTENT_DATA_REF_KEY) {
1155                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1156                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1157                                 return type;
1158                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1159                                 ASSERT(eb->fs_info);
1160                                 /*
1161                                  * Every shared one has parent tree
1162                                  * block, which must be aligned to
1163                                  * nodesize.
1164                                  */
1165                                 if (offset &&
1166                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1167                                         return type;
1168                         }
1169                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1170                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1171                                 return type;
1172                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1173                                 ASSERT(eb->fs_info);
1174                                 /*
1175                                  * Every shared one has parent tree
1176                                  * block, which must be aligned to
1177                                  * nodesize.
1178                                  */
1179                                 if (offset &&
1180                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1181                                         return type;
1182                         }
1183                 } else {
1184                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1185                         return type;
1186                 }
1187         }
1188
1189         btrfs_print_leaf((struct extent_buffer *)eb);
1190         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1191                   eb->start, type);
1192         WARN_ON(1);
1193
1194         return BTRFS_REF_TYPE_INVALID;
1195 }
1196
1197 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1198 {
1199         u32 high_crc = ~(u32)0;
1200         u32 low_crc = ~(u32)0;
1201         __le64 lenum;
1202
1203         lenum = cpu_to_le64(root_objectid);
1204         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1205         lenum = cpu_to_le64(owner);
1206         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1207         lenum = cpu_to_le64(offset);
1208         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1209
1210         return ((u64)high_crc << 31) ^ (u64)low_crc;
1211 }
1212
1213 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1214                                      struct btrfs_extent_data_ref *ref)
1215 {
1216         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1217                                     btrfs_extent_data_ref_objectid(leaf, ref),
1218                                     btrfs_extent_data_ref_offset(leaf, ref));
1219 }
1220
1221 static int match_extent_data_ref(struct extent_buffer *leaf,
1222                                  struct btrfs_extent_data_ref *ref,
1223                                  u64 root_objectid, u64 owner, u64 offset)
1224 {
1225         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1226             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1227             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1228                 return 0;
1229         return 1;
1230 }
1231
1232 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1233                                            struct btrfs_fs_info *fs_info,
1234                                            struct btrfs_path *path,
1235                                            u64 bytenr, u64 parent,
1236                                            u64 root_objectid,
1237                                            u64 owner, u64 offset)
1238 {
1239         struct btrfs_root *root = fs_info->extent_root;
1240         struct btrfs_key key;
1241         struct btrfs_extent_data_ref *ref;
1242         struct extent_buffer *leaf;
1243         u32 nritems;
1244         int ret;
1245         int recow;
1246         int err = -ENOENT;
1247
1248         key.objectid = bytenr;
1249         if (parent) {
1250                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1251                 key.offset = parent;
1252         } else {
1253                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1254                 key.offset = hash_extent_data_ref(root_objectid,
1255                                                   owner, offset);
1256         }
1257 again:
1258         recow = 0;
1259         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1260         if (ret < 0) {
1261                 err = ret;
1262                 goto fail;
1263         }
1264
1265         if (parent) {
1266                 if (!ret)
1267                         return 0;
1268 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1269                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1270                 btrfs_release_path(path);
1271                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1272                 if (ret < 0) {
1273                         err = ret;
1274                         goto fail;
1275                 }
1276                 if (!ret)
1277                         return 0;
1278 #endif
1279                 goto fail;
1280         }
1281
1282         leaf = path->nodes[0];
1283         nritems = btrfs_header_nritems(leaf);
1284         while (1) {
1285                 if (path->slots[0] >= nritems) {
1286                         ret = btrfs_next_leaf(root, path);
1287                         if (ret < 0)
1288                                 err = ret;
1289                         if (ret)
1290                                 goto fail;
1291
1292                         leaf = path->nodes[0];
1293                         nritems = btrfs_header_nritems(leaf);
1294                         recow = 1;
1295                 }
1296
1297                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1298                 if (key.objectid != bytenr ||
1299                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1300                         goto fail;
1301
1302                 ref = btrfs_item_ptr(leaf, path->slots[0],
1303                                      struct btrfs_extent_data_ref);
1304
1305                 if (match_extent_data_ref(leaf, ref, root_objectid,
1306                                           owner, offset)) {
1307                         if (recow) {
1308                                 btrfs_release_path(path);
1309                                 goto again;
1310                         }
1311                         err = 0;
1312                         break;
1313                 }
1314                 path->slots[0]++;
1315         }
1316 fail:
1317         return err;
1318 }
1319
1320 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1321                                            struct btrfs_fs_info *fs_info,
1322                                            struct btrfs_path *path,
1323                                            u64 bytenr, u64 parent,
1324                                            u64 root_objectid, u64 owner,
1325                                            u64 offset, int refs_to_add)
1326 {
1327         struct btrfs_root *root = fs_info->extent_root;
1328         struct btrfs_key key;
1329         struct extent_buffer *leaf;
1330         u32 size;
1331         u32 num_refs;
1332         int ret;
1333
1334         key.objectid = bytenr;
1335         if (parent) {
1336                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1337                 key.offset = parent;
1338                 size = sizeof(struct btrfs_shared_data_ref);
1339         } else {
1340                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1341                 key.offset = hash_extent_data_ref(root_objectid,
1342                                                   owner, offset);
1343                 size = sizeof(struct btrfs_extent_data_ref);
1344         }
1345
1346         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1347         if (ret && ret != -EEXIST)
1348                 goto fail;
1349
1350         leaf = path->nodes[0];
1351         if (parent) {
1352                 struct btrfs_shared_data_ref *ref;
1353                 ref = btrfs_item_ptr(leaf, path->slots[0],
1354                                      struct btrfs_shared_data_ref);
1355                 if (ret == 0) {
1356                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1357                 } else {
1358                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1359                         num_refs += refs_to_add;
1360                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1361                 }
1362         } else {
1363                 struct btrfs_extent_data_ref *ref;
1364                 while (ret == -EEXIST) {
1365                         ref = btrfs_item_ptr(leaf, path->slots[0],
1366                                              struct btrfs_extent_data_ref);
1367                         if (match_extent_data_ref(leaf, ref, root_objectid,
1368                                                   owner, offset))
1369                                 break;
1370                         btrfs_release_path(path);
1371                         key.offset++;
1372                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1373                                                       size);
1374                         if (ret && ret != -EEXIST)
1375                                 goto fail;
1376
1377                         leaf = path->nodes[0];
1378                 }
1379                 ref = btrfs_item_ptr(leaf, path->slots[0],
1380                                      struct btrfs_extent_data_ref);
1381                 if (ret == 0) {
1382                         btrfs_set_extent_data_ref_root(leaf, ref,
1383                                                        root_objectid);
1384                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1385                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1386                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1387                 } else {
1388                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1389                         num_refs += refs_to_add;
1390                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1391                 }
1392         }
1393         btrfs_mark_buffer_dirty(leaf);
1394         ret = 0;
1395 fail:
1396         btrfs_release_path(path);
1397         return ret;
1398 }
1399
1400 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1401                                            struct btrfs_fs_info *fs_info,
1402                                            struct btrfs_path *path,
1403                                            int refs_to_drop, int *last_ref)
1404 {
1405         struct btrfs_key key;
1406         struct btrfs_extent_data_ref *ref1 = NULL;
1407         struct btrfs_shared_data_ref *ref2 = NULL;
1408         struct extent_buffer *leaf;
1409         u32 num_refs = 0;
1410         int ret = 0;
1411
1412         leaf = path->nodes[0];
1413         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1414
1415         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1416                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1417                                       struct btrfs_extent_data_ref);
1418                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1419         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1420                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1421                                       struct btrfs_shared_data_ref);
1422                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1423 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1424         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1425                 struct btrfs_extent_ref_v0 *ref0;
1426                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1427                                       struct btrfs_extent_ref_v0);
1428                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1429 #endif
1430         } else {
1431                 BUG();
1432         }
1433
1434         BUG_ON(num_refs < refs_to_drop);
1435         num_refs -= refs_to_drop;
1436
1437         if (num_refs == 0) {
1438                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1439                 *last_ref = 1;
1440         } else {
1441                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1442                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1443                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1444                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1445 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1446                 else {
1447                         struct btrfs_extent_ref_v0 *ref0;
1448                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1449                                         struct btrfs_extent_ref_v0);
1450                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1451                 }
1452 #endif
1453                 btrfs_mark_buffer_dirty(leaf);
1454         }
1455         return ret;
1456 }
1457
1458 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1459                                           struct btrfs_extent_inline_ref *iref)
1460 {
1461         struct btrfs_key key;
1462         struct extent_buffer *leaf;
1463         struct btrfs_extent_data_ref *ref1;
1464         struct btrfs_shared_data_ref *ref2;
1465         u32 num_refs = 0;
1466         int type;
1467
1468         leaf = path->nodes[0];
1469         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1470         if (iref) {
1471                 /*
1472                  * If type is invalid, we should have bailed out earlier than
1473                  * this call.
1474                  */
1475                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1476                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1477                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1478                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1479                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1480                 } else {
1481                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1482                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1483                 }
1484         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1485                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1486                                       struct btrfs_extent_data_ref);
1487                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1488         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1489                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1490                                       struct btrfs_shared_data_ref);
1491                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1492 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1493         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1494                 struct btrfs_extent_ref_v0 *ref0;
1495                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1496                                       struct btrfs_extent_ref_v0);
1497                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1498 #endif
1499         } else {
1500                 WARN_ON(1);
1501         }
1502         return num_refs;
1503 }
1504
1505 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1506                                           struct btrfs_fs_info *fs_info,
1507                                           struct btrfs_path *path,
1508                                           u64 bytenr, u64 parent,
1509                                           u64 root_objectid)
1510 {
1511         struct btrfs_root *root = fs_info->extent_root;
1512         struct btrfs_key key;
1513         int ret;
1514
1515         key.objectid = bytenr;
1516         if (parent) {
1517                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1518                 key.offset = parent;
1519         } else {
1520                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1521                 key.offset = root_objectid;
1522         }
1523
1524         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1525         if (ret > 0)
1526                 ret = -ENOENT;
1527 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1528         if (ret == -ENOENT && parent) {
1529                 btrfs_release_path(path);
1530                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1531                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1532                 if (ret > 0)
1533                         ret = -ENOENT;
1534         }
1535 #endif
1536         return ret;
1537 }
1538
1539 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1540                                           struct btrfs_fs_info *fs_info,
1541                                           struct btrfs_path *path,
1542                                           u64 bytenr, u64 parent,
1543                                           u64 root_objectid)
1544 {
1545         struct btrfs_key key;
1546         int ret;
1547
1548         key.objectid = bytenr;
1549         if (parent) {
1550                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1551                 key.offset = parent;
1552         } else {
1553                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1554                 key.offset = root_objectid;
1555         }
1556
1557         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1558                                       path, &key, 0);
1559         btrfs_release_path(path);
1560         return ret;
1561 }
1562
1563 static inline int extent_ref_type(u64 parent, u64 owner)
1564 {
1565         int type;
1566         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1567                 if (parent > 0)
1568                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1569                 else
1570                         type = BTRFS_TREE_BLOCK_REF_KEY;
1571         } else {
1572                 if (parent > 0)
1573                         type = BTRFS_SHARED_DATA_REF_KEY;
1574                 else
1575                         type = BTRFS_EXTENT_DATA_REF_KEY;
1576         }
1577         return type;
1578 }
1579
1580 static int find_next_key(struct btrfs_path *path, int level,
1581                          struct btrfs_key *key)
1582
1583 {
1584         for (; level < BTRFS_MAX_LEVEL; level++) {
1585                 if (!path->nodes[level])
1586                         break;
1587                 if (path->slots[level] + 1 >=
1588                     btrfs_header_nritems(path->nodes[level]))
1589                         continue;
1590                 if (level == 0)
1591                         btrfs_item_key_to_cpu(path->nodes[level], key,
1592                                               path->slots[level] + 1);
1593                 else
1594                         btrfs_node_key_to_cpu(path->nodes[level], key,
1595                                               path->slots[level] + 1);
1596                 return 0;
1597         }
1598         return 1;
1599 }
1600
1601 /*
1602  * look for inline back ref. if back ref is found, *ref_ret is set
1603  * to the address of inline back ref, and 0 is returned.
1604  *
1605  * if back ref isn't found, *ref_ret is set to the address where it
1606  * should be inserted, and -ENOENT is returned.
1607  *
1608  * if insert is true and there are too many inline back refs, the path
1609  * points to the extent item, and -EAGAIN is returned.
1610  *
1611  * NOTE: inline back refs are ordered in the same way that back ref
1612  *       items in the tree are ordered.
1613  */
1614 static noinline_for_stack
1615 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1616                                  struct btrfs_fs_info *fs_info,
1617                                  struct btrfs_path *path,
1618                                  struct btrfs_extent_inline_ref **ref_ret,
1619                                  u64 bytenr, u64 num_bytes,
1620                                  u64 parent, u64 root_objectid,
1621                                  u64 owner, u64 offset, int insert)
1622 {
1623         struct btrfs_root *root = fs_info->extent_root;
1624         struct btrfs_key key;
1625         struct extent_buffer *leaf;
1626         struct btrfs_extent_item *ei;
1627         struct btrfs_extent_inline_ref *iref;
1628         u64 flags;
1629         u64 item_size;
1630         unsigned long ptr;
1631         unsigned long end;
1632         int extra_size;
1633         int type;
1634         int want;
1635         int ret;
1636         int err = 0;
1637         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1638         int needed;
1639
1640         key.objectid = bytenr;
1641         key.type = BTRFS_EXTENT_ITEM_KEY;
1642         key.offset = num_bytes;
1643
1644         want = extent_ref_type(parent, owner);
1645         if (insert) {
1646                 extra_size = btrfs_extent_inline_ref_size(want);
1647                 path->keep_locks = 1;
1648         } else
1649                 extra_size = -1;
1650
1651         /*
1652          * Owner is our parent level, so we can just add one to get the level
1653          * for the block we are interested in.
1654          */
1655         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1656                 key.type = BTRFS_METADATA_ITEM_KEY;
1657                 key.offset = owner;
1658         }
1659
1660 again:
1661         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1662         if (ret < 0) {
1663                 err = ret;
1664                 goto out;
1665         }
1666
1667         /*
1668          * We may be a newly converted file system which still has the old fat
1669          * extent entries for metadata, so try and see if we have one of those.
1670          */
1671         if (ret > 0 && skinny_metadata) {
1672                 skinny_metadata = false;
1673                 if (path->slots[0]) {
1674                         path->slots[0]--;
1675                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1676                                               path->slots[0]);
1677                         if (key.objectid == bytenr &&
1678                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1679                             key.offset == num_bytes)
1680                                 ret = 0;
1681                 }
1682                 if (ret) {
1683                         key.objectid = bytenr;
1684                         key.type = BTRFS_EXTENT_ITEM_KEY;
1685                         key.offset = num_bytes;
1686                         btrfs_release_path(path);
1687                         goto again;
1688                 }
1689         }
1690
1691         if (ret && !insert) {
1692                 err = -ENOENT;
1693                 goto out;
1694         } else if (WARN_ON(ret)) {
1695                 err = -EIO;
1696                 goto out;
1697         }
1698
1699         leaf = path->nodes[0];
1700         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1702         if (item_size < sizeof(*ei)) {
1703                 if (!insert) {
1704                         err = -ENOENT;
1705                         goto out;
1706                 }
1707                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1708                                              extra_size);
1709                 if (ret < 0) {
1710                         err = ret;
1711                         goto out;
1712                 }
1713                 leaf = path->nodes[0];
1714                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1715         }
1716 #endif
1717         BUG_ON(item_size < sizeof(*ei));
1718
1719         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1720         flags = btrfs_extent_flags(leaf, ei);
1721
1722         ptr = (unsigned long)(ei + 1);
1723         end = (unsigned long)ei + item_size;
1724
1725         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1726                 ptr += sizeof(struct btrfs_tree_block_info);
1727                 BUG_ON(ptr > end);
1728         }
1729
1730         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1731                 needed = BTRFS_REF_TYPE_DATA;
1732         else
1733                 needed = BTRFS_REF_TYPE_BLOCK;
1734
1735         err = -ENOENT;
1736         while (1) {
1737                 if (ptr >= end) {
1738                         WARN_ON(ptr > end);
1739                         break;
1740                 }
1741                 iref = (struct btrfs_extent_inline_ref *)ptr;
1742                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1743                 if (type == BTRFS_REF_TYPE_INVALID) {
1744                         err = -EINVAL;
1745                         goto out;
1746                 }
1747
1748                 if (want < type)
1749                         break;
1750                 if (want > type) {
1751                         ptr += btrfs_extent_inline_ref_size(type);
1752                         continue;
1753                 }
1754
1755                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1756                         struct btrfs_extent_data_ref *dref;
1757                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1758                         if (match_extent_data_ref(leaf, dref, root_objectid,
1759                                                   owner, offset)) {
1760                                 err = 0;
1761                                 break;
1762                         }
1763                         if (hash_extent_data_ref_item(leaf, dref) <
1764                             hash_extent_data_ref(root_objectid, owner, offset))
1765                                 break;
1766                 } else {
1767                         u64 ref_offset;
1768                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1769                         if (parent > 0) {
1770                                 if (parent == ref_offset) {
1771                                         err = 0;
1772                                         break;
1773                                 }
1774                                 if (ref_offset < parent)
1775                                         break;
1776                         } else {
1777                                 if (root_objectid == ref_offset) {
1778                                         err = 0;
1779                                         break;
1780                                 }
1781                                 if (ref_offset < root_objectid)
1782                                         break;
1783                         }
1784                 }
1785                 ptr += btrfs_extent_inline_ref_size(type);
1786         }
1787         if (err == -ENOENT && insert) {
1788                 if (item_size + extra_size >=
1789                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1790                         err = -EAGAIN;
1791                         goto out;
1792                 }
1793                 /*
1794                  * To add new inline back ref, we have to make sure
1795                  * there is no corresponding back ref item.
1796                  * For simplicity, we just do not add new inline back
1797                  * ref if there is any kind of item for this block
1798                  */
1799                 if (find_next_key(path, 0, &key) == 0 &&
1800                     key.objectid == bytenr &&
1801                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1802                         err = -EAGAIN;
1803                         goto out;
1804                 }
1805         }
1806         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1807 out:
1808         if (insert) {
1809                 path->keep_locks = 0;
1810                 btrfs_unlock_up_safe(path, 1);
1811         }
1812         return err;
1813 }
1814
1815 /*
1816  * helper to add new inline back ref
1817  */
1818 static noinline_for_stack
1819 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1820                                  struct btrfs_path *path,
1821                                  struct btrfs_extent_inline_ref *iref,
1822                                  u64 parent, u64 root_objectid,
1823                                  u64 owner, u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct extent_buffer *leaf;
1827         struct btrfs_extent_item *ei;
1828         unsigned long ptr;
1829         unsigned long end;
1830         unsigned long item_offset;
1831         u64 refs;
1832         int size;
1833         int type;
1834
1835         leaf = path->nodes[0];
1836         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1837         item_offset = (unsigned long)iref - (unsigned long)ei;
1838
1839         type = extent_ref_type(parent, owner);
1840         size = btrfs_extent_inline_ref_size(type);
1841
1842         btrfs_extend_item(fs_info, path, size);
1843
1844         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1845         refs = btrfs_extent_refs(leaf, ei);
1846         refs += refs_to_add;
1847         btrfs_set_extent_refs(leaf, ei, refs);
1848         if (extent_op)
1849                 __run_delayed_extent_op(extent_op, leaf, ei);
1850
1851         ptr = (unsigned long)ei + item_offset;
1852         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1853         if (ptr < end - size)
1854                 memmove_extent_buffer(leaf, ptr + size, ptr,
1855                                       end - size - ptr);
1856
1857         iref = (struct btrfs_extent_inline_ref *)ptr;
1858         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1859         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1860                 struct btrfs_extent_data_ref *dref;
1861                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1862                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1863                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1864                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1865                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1866         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1867                 struct btrfs_shared_data_ref *sref;
1868                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1869                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1870                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1871         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1872                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1873         } else {
1874                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1875         }
1876         btrfs_mark_buffer_dirty(leaf);
1877 }
1878
1879 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1880                                  struct btrfs_fs_info *fs_info,
1881                                  struct btrfs_path *path,
1882                                  struct btrfs_extent_inline_ref **ref_ret,
1883                                  u64 bytenr, u64 num_bytes, u64 parent,
1884                                  u64 root_objectid, u64 owner, u64 offset)
1885 {
1886         int ret;
1887
1888         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1889                                            bytenr, num_bytes, parent,
1890                                            root_objectid, owner, offset, 0);
1891         if (ret != -ENOENT)
1892                 return ret;
1893
1894         btrfs_release_path(path);
1895         *ref_ret = NULL;
1896
1897         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1898                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1899                                             parent, root_objectid);
1900         } else {
1901                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1902                                              parent, root_objectid, owner,
1903                                              offset);
1904         }
1905         return ret;
1906 }
1907
1908 /*
1909  * helper to update/remove inline back ref
1910  */
1911 static noinline_for_stack
1912 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1913                                   struct btrfs_path *path,
1914                                   struct btrfs_extent_inline_ref *iref,
1915                                   int refs_to_mod,
1916                                   struct btrfs_delayed_extent_op *extent_op,
1917                                   int *last_ref)
1918 {
1919         struct extent_buffer *leaf;
1920         struct btrfs_extent_item *ei;
1921         struct btrfs_extent_data_ref *dref = NULL;
1922         struct btrfs_shared_data_ref *sref = NULL;
1923         unsigned long ptr;
1924         unsigned long end;
1925         u32 item_size;
1926         int size;
1927         int type;
1928         u64 refs;
1929
1930         leaf = path->nodes[0];
1931         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1932         refs = btrfs_extent_refs(leaf, ei);
1933         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1934         refs += refs_to_mod;
1935         btrfs_set_extent_refs(leaf, ei, refs);
1936         if (extent_op)
1937                 __run_delayed_extent_op(extent_op, leaf, ei);
1938
1939         /*
1940          * If type is invalid, we should have bailed out after
1941          * lookup_inline_extent_backref().
1942          */
1943         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1944         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1945
1946         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1947                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1948                 refs = btrfs_extent_data_ref_count(leaf, dref);
1949         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1950                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1951                 refs = btrfs_shared_data_ref_count(leaf, sref);
1952         } else {
1953                 refs = 1;
1954                 BUG_ON(refs_to_mod != -1);
1955         }
1956
1957         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1958         refs += refs_to_mod;
1959
1960         if (refs > 0) {
1961                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1962                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1963                 else
1964                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1965         } else {
1966                 *last_ref = 1;
1967                 size =  btrfs_extent_inline_ref_size(type);
1968                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1969                 ptr = (unsigned long)iref;
1970                 end = (unsigned long)ei + item_size;
1971                 if (ptr + size < end)
1972                         memmove_extent_buffer(leaf, ptr, ptr + size,
1973                                               end - ptr - size);
1974                 item_size -= size;
1975                 btrfs_truncate_item(fs_info, path, item_size, 1);
1976         }
1977         btrfs_mark_buffer_dirty(leaf);
1978 }
1979
1980 static noinline_for_stack
1981 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1982                                  struct btrfs_fs_info *fs_info,
1983                                  struct btrfs_path *path,
1984                                  u64 bytenr, u64 num_bytes, u64 parent,
1985                                  u64 root_objectid, u64 owner,
1986                                  u64 offset, int refs_to_add,
1987                                  struct btrfs_delayed_extent_op *extent_op)
1988 {
1989         struct btrfs_extent_inline_ref *iref;
1990         int ret;
1991
1992         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1993                                            bytenr, num_bytes, parent,
1994                                            root_objectid, owner, offset, 1);
1995         if (ret == 0) {
1996                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1997                 update_inline_extent_backref(fs_info, path, iref,
1998                                              refs_to_add, extent_op, NULL);
1999         } else if (ret == -ENOENT) {
2000                 setup_inline_extent_backref(fs_info, path, iref, parent,
2001                                             root_objectid, owner, offset,
2002                                             refs_to_add, extent_op);
2003                 ret = 0;
2004         }
2005         return ret;
2006 }
2007
2008 static int insert_extent_backref(struct btrfs_trans_handle *trans,
2009                                  struct btrfs_fs_info *fs_info,
2010                                  struct btrfs_path *path,
2011                                  u64 bytenr, u64 parent, u64 root_objectid,
2012                                  u64 owner, u64 offset, int refs_to_add)
2013 {
2014         int ret;
2015         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2016                 BUG_ON(refs_to_add != 1);
2017                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
2018                                             parent, root_objectid);
2019         } else {
2020                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2021                                              parent, root_objectid,
2022                                              owner, offset, refs_to_add);
2023         }
2024         return ret;
2025 }
2026
2027 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2028                                  struct btrfs_fs_info *fs_info,
2029                                  struct btrfs_path *path,
2030                                  struct btrfs_extent_inline_ref *iref,
2031                                  int refs_to_drop, int is_data, int *last_ref)
2032 {
2033         int ret = 0;
2034
2035         BUG_ON(!is_data && refs_to_drop != 1);
2036         if (iref) {
2037                 update_inline_extent_backref(fs_info, path, iref,
2038                                              -refs_to_drop, NULL, last_ref);
2039         } else if (is_data) {
2040                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2041                                              last_ref);
2042         } else {
2043                 *last_ref = 1;
2044                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2045         }
2046         return ret;
2047 }
2048
2049 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2050 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2051                                u64 *discarded_bytes)
2052 {
2053         int j, ret = 0;
2054         u64 bytes_left, end;
2055         u64 aligned_start = ALIGN(start, 1 << 9);
2056
2057         if (WARN_ON(start != aligned_start)) {
2058                 len -= aligned_start - start;
2059                 len = round_down(len, 1 << 9);
2060                 start = aligned_start;
2061         }
2062
2063         *discarded_bytes = 0;
2064
2065         if (!len)
2066                 return 0;
2067
2068         end = start + len;
2069         bytes_left = len;
2070
2071         /* Skip any superblocks on this device. */
2072         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2073                 u64 sb_start = btrfs_sb_offset(j);
2074                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2075                 u64 size = sb_start - start;
2076
2077                 if (!in_range(sb_start, start, bytes_left) &&
2078                     !in_range(sb_end, start, bytes_left) &&
2079                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2080                         continue;
2081
2082                 /*
2083                  * Superblock spans beginning of range.  Adjust start and
2084                  * try again.
2085                  */
2086                 if (sb_start <= start) {
2087                         start += sb_end - start;
2088                         if (start > end) {
2089                                 bytes_left = 0;
2090                                 break;
2091                         }
2092                         bytes_left = end - start;
2093                         continue;
2094                 }
2095
2096                 if (size) {
2097                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2098                                                    GFP_NOFS, 0);
2099                         if (!ret)
2100                                 *discarded_bytes += size;
2101                         else if (ret != -EOPNOTSUPP)
2102                                 return ret;
2103                 }
2104
2105                 start = sb_end;
2106                 if (start > end) {
2107                         bytes_left = 0;
2108                         break;
2109                 }
2110                 bytes_left = end - start;
2111         }
2112
2113         if (bytes_left) {
2114                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2115                                            GFP_NOFS, 0);
2116                 if (!ret)
2117                         *discarded_bytes += bytes_left;
2118         }
2119         return ret;
2120 }
2121
2122 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2123                          u64 num_bytes, u64 *actual_bytes)
2124 {
2125         int ret;
2126         u64 discarded_bytes = 0;
2127         struct btrfs_bio *bbio = NULL;
2128
2129
2130         /*
2131          * Avoid races with device replace and make sure our bbio has devices
2132          * associated to its stripes that don't go away while we are discarding.
2133          */
2134         btrfs_bio_counter_inc_blocked(fs_info);
2135         /* Tell the block device(s) that the sectors can be discarded */
2136         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2137                               &bbio, 0);
2138         /* Error condition is -ENOMEM */
2139         if (!ret) {
2140                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2141                 int i;
2142
2143
2144                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2145                         u64 bytes;
2146                         struct request_queue *req_q;
2147
2148                         if (!stripe->dev->bdev) {
2149                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2150                                 continue;
2151                         }
2152                         req_q = bdev_get_queue(stripe->dev->bdev);
2153                         if (!blk_queue_discard(req_q))
2154                                 continue;
2155
2156                         ret = btrfs_issue_discard(stripe->dev->bdev,
2157                                                   stripe->physical,
2158                                                   stripe->length,
2159                                                   &bytes);
2160                         if (!ret)
2161                                 discarded_bytes += bytes;
2162                         else if (ret != -EOPNOTSUPP)
2163                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2164
2165                         /*
2166                          * Just in case we get back EOPNOTSUPP for some reason,
2167                          * just ignore the return value so we don't screw up
2168                          * people calling discard_extent.
2169                          */
2170                         ret = 0;
2171                 }
2172                 btrfs_put_bbio(bbio);
2173         }
2174         btrfs_bio_counter_dec(fs_info);
2175
2176         if (actual_bytes)
2177                 *actual_bytes = discarded_bytes;
2178
2179
2180         if (ret == -EOPNOTSUPP)
2181                 ret = 0;
2182         return ret;
2183 }
2184
2185 /* Can return -ENOMEM */
2186 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2187                          struct btrfs_root *root,
2188                          u64 bytenr, u64 num_bytes, u64 parent,
2189                          u64 root_objectid, u64 owner, u64 offset)
2190 {
2191         struct btrfs_fs_info *fs_info = root->fs_info;
2192         int old_ref_mod, new_ref_mod;
2193         int ret;
2194
2195         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2196                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2197
2198         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2199                            owner, offset, BTRFS_ADD_DELAYED_REF);
2200
2201         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2202                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2203                                                  num_bytes, parent,
2204                                                  root_objectid, (int)owner,
2205                                                  BTRFS_ADD_DELAYED_REF, NULL,
2206                                                  &old_ref_mod, &new_ref_mod);
2207         } else {
2208                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2209                                                  num_bytes, parent,
2210                                                  root_objectid, owner, offset,
2211                                                  0, BTRFS_ADD_DELAYED_REF,
2212                                                  &old_ref_mod, &new_ref_mod);
2213         }
2214
2215         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2216                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2217
2218         return ret;
2219 }
2220
2221 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2222                                   struct btrfs_fs_info *fs_info,
2223                                   struct btrfs_delayed_ref_node *node,
2224                                   u64 parent, u64 root_objectid,
2225                                   u64 owner, u64 offset, int refs_to_add,
2226                                   struct btrfs_delayed_extent_op *extent_op)
2227 {
2228         struct btrfs_path *path;
2229         struct extent_buffer *leaf;
2230         struct btrfs_extent_item *item;
2231         struct btrfs_key key;
2232         u64 bytenr = node->bytenr;
2233         u64 num_bytes = node->num_bytes;
2234         u64 refs;
2235         int ret;
2236
2237         path = btrfs_alloc_path();
2238         if (!path)
2239                 return -ENOMEM;
2240
2241         path->reada = READA_FORWARD;
2242         path->leave_spinning = 1;
2243         /* this will setup the path even if it fails to insert the back ref */
2244         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2245                                            num_bytes, parent, root_objectid,
2246                                            owner, offset,
2247                                            refs_to_add, extent_op);
2248         if ((ret < 0 && ret != -EAGAIN) || !ret)
2249                 goto out;
2250
2251         /*
2252          * Ok we had -EAGAIN which means we didn't have space to insert and
2253          * inline extent ref, so just update the reference count and add a
2254          * normal backref.
2255          */
2256         leaf = path->nodes[0];
2257         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2258         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2259         refs = btrfs_extent_refs(leaf, item);
2260         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2261         if (extent_op)
2262                 __run_delayed_extent_op(extent_op, leaf, item);
2263
2264         btrfs_mark_buffer_dirty(leaf);
2265         btrfs_release_path(path);
2266
2267         path->reada = READA_FORWARD;
2268         path->leave_spinning = 1;
2269         /* now insert the actual backref */
2270         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2271                                     root_objectid, owner, offset, refs_to_add);
2272         if (ret)
2273                 btrfs_abort_transaction(trans, ret);
2274 out:
2275         btrfs_free_path(path);
2276         return ret;
2277 }
2278
2279 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2280                                 struct btrfs_fs_info *fs_info,
2281                                 struct btrfs_delayed_ref_node *node,
2282                                 struct btrfs_delayed_extent_op *extent_op,
2283                                 int insert_reserved)
2284 {
2285         int ret = 0;
2286         struct btrfs_delayed_data_ref *ref;
2287         struct btrfs_key ins;
2288         u64 parent = 0;
2289         u64 ref_root = 0;
2290         u64 flags = 0;
2291
2292         ins.objectid = node->bytenr;
2293         ins.offset = node->num_bytes;
2294         ins.type = BTRFS_EXTENT_ITEM_KEY;
2295
2296         ref = btrfs_delayed_node_to_data_ref(node);
2297         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2298
2299         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2300                 parent = ref->parent;
2301         ref_root = ref->root;
2302
2303         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2304                 if (extent_op)
2305                         flags |= extent_op->flags_to_set;
2306                 ret = alloc_reserved_file_extent(trans, fs_info,
2307                                                  parent, ref_root, flags,
2308                                                  ref->objectid, ref->offset,
2309                                                  &ins, node->ref_mod);
2310         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2311                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2312                                              ref_root, ref->objectid,
2313                                              ref->offset, node->ref_mod,
2314                                              extent_op);
2315         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2316                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2317                                           ref_root, ref->objectid,
2318                                           ref->offset, node->ref_mod,
2319                                           extent_op);
2320         } else {
2321                 BUG();
2322         }
2323         return ret;
2324 }
2325
2326 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2327                                     struct extent_buffer *leaf,
2328                                     struct btrfs_extent_item *ei)
2329 {
2330         u64 flags = btrfs_extent_flags(leaf, ei);
2331         if (extent_op->update_flags) {
2332                 flags |= extent_op->flags_to_set;
2333                 btrfs_set_extent_flags(leaf, ei, flags);
2334         }
2335
2336         if (extent_op->update_key) {
2337                 struct btrfs_tree_block_info *bi;
2338                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2339                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2340                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2341         }
2342 }
2343
2344 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2345                                  struct btrfs_fs_info *fs_info,
2346                                  struct btrfs_delayed_ref_head *head,
2347                                  struct btrfs_delayed_extent_op *extent_op)
2348 {
2349         struct btrfs_key key;
2350         struct btrfs_path *path;
2351         struct btrfs_extent_item *ei;
2352         struct extent_buffer *leaf;
2353         u32 item_size;
2354         int ret;
2355         int err = 0;
2356         int metadata = !extent_op->is_data;
2357
2358         if (trans->aborted)
2359                 return 0;
2360
2361         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2362                 metadata = 0;
2363
2364         path = btrfs_alloc_path();
2365         if (!path)
2366                 return -ENOMEM;
2367
2368         key.objectid = head->bytenr;
2369
2370         if (metadata) {
2371                 key.type = BTRFS_METADATA_ITEM_KEY;
2372                 key.offset = extent_op->level;
2373         } else {
2374                 key.type = BTRFS_EXTENT_ITEM_KEY;
2375                 key.offset = head->num_bytes;
2376         }
2377
2378 again:
2379         path->reada = READA_FORWARD;
2380         path->leave_spinning = 1;
2381         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2382         if (ret < 0) {
2383                 err = ret;
2384                 goto out;
2385         }
2386         if (ret > 0) {
2387                 if (metadata) {
2388                         if (path->slots[0] > 0) {
2389                                 path->slots[0]--;
2390                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2391                                                       path->slots[0]);
2392                                 if (key.objectid == head->bytenr &&
2393                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2394                                     key.offset == head->num_bytes)
2395                                         ret = 0;
2396                         }
2397                         if (ret > 0) {
2398                                 btrfs_release_path(path);
2399                                 metadata = 0;
2400
2401                                 key.objectid = head->bytenr;
2402                                 key.offset = head->num_bytes;
2403                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2404                                 goto again;
2405                         }
2406                 } else {
2407                         err = -EIO;
2408                         goto out;
2409                 }
2410         }
2411
2412         leaf = path->nodes[0];
2413         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2414 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2415         if (item_size < sizeof(*ei)) {
2416                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2417                 if (ret < 0) {
2418                         err = ret;
2419                         goto out;
2420                 }
2421                 leaf = path->nodes[0];
2422                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2423         }
2424 #endif
2425         BUG_ON(item_size < sizeof(*ei));
2426         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2427         __run_delayed_extent_op(extent_op, leaf, ei);
2428
2429         btrfs_mark_buffer_dirty(leaf);
2430 out:
2431         btrfs_free_path(path);
2432         return err;
2433 }
2434
2435 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2436                                 struct btrfs_fs_info *fs_info,
2437                                 struct btrfs_delayed_ref_node *node,
2438                                 struct btrfs_delayed_extent_op *extent_op,
2439                                 int insert_reserved)
2440 {
2441         int ret = 0;
2442         struct btrfs_delayed_tree_ref *ref;
2443         struct btrfs_key ins;
2444         u64 parent = 0;
2445         u64 ref_root = 0;
2446         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2447
2448         ref = btrfs_delayed_node_to_tree_ref(node);
2449         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2450
2451         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2452                 parent = ref->parent;
2453         ref_root = ref->root;
2454
2455         ins.objectid = node->bytenr;
2456         if (skinny_metadata) {
2457                 ins.offset = ref->level;
2458                 ins.type = BTRFS_METADATA_ITEM_KEY;
2459         } else {
2460                 ins.offset = node->num_bytes;
2461                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2462         }
2463
2464         if (node->ref_mod != 1) {
2465                 btrfs_err(fs_info,
2466         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2467                           node->bytenr, node->ref_mod, node->action, ref_root,
2468                           parent);
2469                 return -EIO;
2470         }
2471         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2472                 BUG_ON(!extent_op || !extent_op->update_flags);
2473                 ret = alloc_reserved_tree_block(trans, fs_info,
2474                                                 parent, ref_root,
2475                                                 extent_op->flags_to_set,
2476                                                 &extent_op->key,
2477                                                 ref->level, &ins);
2478         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2479                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2480                                              parent, ref_root,
2481                                              ref->level, 0, 1,
2482                                              extent_op);
2483         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2484                 ret = __btrfs_free_extent(trans, fs_info, node,
2485                                           parent, ref_root,
2486                                           ref->level, 0, 1, extent_op);
2487         } else {
2488                 BUG();
2489         }
2490         return ret;
2491 }
2492
2493 /* helper function to actually process a single delayed ref entry */
2494 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2495                                struct btrfs_fs_info *fs_info,
2496                                struct btrfs_delayed_ref_node *node,
2497                                struct btrfs_delayed_extent_op *extent_op,
2498                                int insert_reserved)
2499 {
2500         int ret = 0;
2501
2502         if (trans->aborted) {
2503                 if (insert_reserved)
2504                         btrfs_pin_extent(fs_info, node->bytenr,
2505                                          node->num_bytes, 1);
2506                 return 0;
2507         }
2508
2509         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2510             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2511                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2512                                            insert_reserved);
2513         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2514                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2515                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2516                                            insert_reserved);
2517         else
2518                 BUG();
2519         return ret;
2520 }
2521
2522 static inline struct btrfs_delayed_ref_node *
2523 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2524 {
2525         struct btrfs_delayed_ref_node *ref;
2526
2527         if (RB_EMPTY_ROOT(&head->ref_tree))
2528                 return NULL;
2529
2530         /*
2531          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2532          * This is to prevent a ref count from going down to zero, which deletes
2533          * the extent item from the extent tree, when there still are references
2534          * to add, which would fail because they would not find the extent item.
2535          */
2536         if (!list_empty(&head->ref_add_list))
2537                 return list_first_entry(&head->ref_add_list,
2538                                 struct btrfs_delayed_ref_node, add_list);
2539
2540         ref = rb_entry(rb_first(&head->ref_tree),
2541                        struct btrfs_delayed_ref_node, ref_node);
2542         ASSERT(list_empty(&ref->add_list));
2543         return ref;
2544 }
2545
2546 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2547                                       struct btrfs_delayed_ref_head *head)
2548 {
2549         spin_lock(&delayed_refs->lock);
2550         head->processing = 0;
2551         delayed_refs->num_heads_ready++;
2552         spin_unlock(&delayed_refs->lock);
2553         btrfs_delayed_ref_unlock(head);
2554 }
2555
2556 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2557                              struct btrfs_fs_info *fs_info,
2558                              struct btrfs_delayed_ref_head *head)
2559 {
2560         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2561         int ret;
2562
2563         if (!extent_op)
2564                 return 0;
2565         head->extent_op = NULL;
2566         if (head->must_insert_reserved) {
2567                 btrfs_free_delayed_extent_op(extent_op);
2568                 return 0;
2569         }
2570         spin_unlock(&head->lock);
2571         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2572         btrfs_free_delayed_extent_op(extent_op);
2573         return ret ? ret : 1;
2574 }
2575
2576 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2577                             struct btrfs_fs_info *fs_info,
2578                             struct btrfs_delayed_ref_head *head)
2579 {
2580         struct btrfs_delayed_ref_root *delayed_refs;
2581         int ret;
2582
2583         delayed_refs = &trans->transaction->delayed_refs;
2584
2585         ret = cleanup_extent_op(trans, fs_info, head);
2586         if (ret < 0) {
2587                 unselect_delayed_ref_head(delayed_refs, head);
2588                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2589                 return ret;
2590         } else if (ret) {
2591                 return ret;
2592         }
2593
2594         /*
2595          * Need to drop our head ref lock and re-acquire the delayed ref lock
2596          * and then re-check to make sure nobody got added.
2597          */
2598         spin_unlock(&head->lock);
2599         spin_lock(&delayed_refs->lock);
2600         spin_lock(&head->lock);
2601         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2602                 spin_unlock(&head->lock);
2603                 spin_unlock(&delayed_refs->lock);
2604                 return 1;
2605         }
2606         delayed_refs->num_heads--;
2607         rb_erase(&head->href_node, &delayed_refs->href_root);
2608         RB_CLEAR_NODE(&head->href_node);
2609         spin_unlock(&delayed_refs->lock);
2610         spin_unlock(&head->lock);
2611         atomic_dec(&delayed_refs->num_entries);
2612
2613         trace_run_delayed_ref_head(fs_info, head, 0);
2614
2615         if (head->total_ref_mod < 0) {
2616                 struct btrfs_block_group_cache *cache;
2617
2618                 cache = btrfs_lookup_block_group(fs_info, head->bytenr);
2619                 ASSERT(cache);
2620                 percpu_counter_add(&cache->space_info->total_bytes_pinned,
2621                                    -head->num_bytes);
2622                 btrfs_put_block_group(cache);
2623
2624                 if (head->is_data) {
2625                         spin_lock(&delayed_refs->lock);
2626                         delayed_refs->pending_csums -= head->num_bytes;
2627                         spin_unlock(&delayed_refs->lock);
2628                 }
2629         }
2630
2631         if (head->must_insert_reserved) {
2632                 btrfs_pin_extent(fs_info, head->bytenr,
2633                                  head->num_bytes, 1);
2634                 if (head->is_data) {
2635                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2636                                               head->num_bytes);
2637                 }
2638         }
2639
2640         /* Also free its reserved qgroup space */
2641         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2642                                       head->qgroup_reserved);
2643         btrfs_delayed_ref_unlock(head);
2644         btrfs_put_delayed_ref_head(head);
2645         return 0;
2646 }
2647
2648 /*
2649  * Returns 0 on success or if called with an already aborted transaction.
2650  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2651  */
2652 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2653                                              struct btrfs_fs_info *fs_info,
2654                                              unsigned long nr)
2655 {
2656         struct btrfs_delayed_ref_root *delayed_refs;
2657         struct btrfs_delayed_ref_node *ref;
2658         struct btrfs_delayed_ref_head *locked_ref = NULL;
2659         struct btrfs_delayed_extent_op *extent_op;
2660         ktime_t start = ktime_get();
2661         int ret;
2662         unsigned long count = 0;
2663         unsigned long actual_count = 0;
2664         int must_insert_reserved = 0;
2665
2666         delayed_refs = &trans->transaction->delayed_refs;
2667         while (1) {
2668                 if (!locked_ref) {
2669                         if (count >= nr)
2670                                 break;
2671
2672                         spin_lock(&delayed_refs->lock);
2673                         locked_ref = btrfs_select_ref_head(trans);
2674                         if (!locked_ref) {
2675                                 spin_unlock(&delayed_refs->lock);
2676                                 break;
2677                         }
2678
2679                         /* grab the lock that says we are going to process
2680                          * all the refs for this head */
2681                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2682                         spin_unlock(&delayed_refs->lock);
2683                         /*
2684                          * we may have dropped the spin lock to get the head
2685                          * mutex lock, and that might have given someone else
2686                          * time to free the head.  If that's true, it has been
2687                          * removed from our list and we can move on.
2688                          */
2689                         if (ret == -EAGAIN) {
2690                                 locked_ref = NULL;
2691                                 count++;
2692                                 continue;
2693                         }
2694                 }
2695
2696                 /*
2697                  * We need to try and merge add/drops of the same ref since we
2698                  * can run into issues with relocate dropping the implicit ref
2699                  * and then it being added back again before the drop can
2700                  * finish.  If we merged anything we need to re-loop so we can
2701                  * get a good ref.
2702                  * Or we can get node references of the same type that weren't
2703                  * merged when created due to bumps in the tree mod seq, and
2704                  * we need to merge them to prevent adding an inline extent
2705                  * backref before dropping it (triggering a BUG_ON at
2706                  * insert_inline_extent_backref()).
2707                  */
2708                 spin_lock(&locked_ref->lock);
2709                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2710                                          locked_ref);
2711
2712                 /*
2713                  * locked_ref is the head node, so we have to go one
2714                  * node back for any delayed ref updates
2715                  */
2716                 ref = select_delayed_ref(locked_ref);
2717
2718                 if (ref && ref->seq &&
2719                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2720                         spin_unlock(&locked_ref->lock);
2721                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2722                         locked_ref = NULL;
2723                         cond_resched();
2724                         count++;
2725                         continue;
2726                 }
2727
2728                 /*
2729                  * We're done processing refs in this ref_head, clean everything
2730                  * up and move on to the next ref_head.
2731                  */
2732                 if (!ref) {
2733                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2734                         if (ret > 0 ) {
2735                                 /* We dropped our lock, we need to loop. */
2736                                 ret = 0;
2737                                 continue;
2738                         } else if (ret) {
2739                                 return ret;
2740                         }
2741                         locked_ref = NULL;
2742                         count++;
2743                         continue;
2744                 }
2745
2746                 actual_count++;
2747                 ref->in_tree = 0;
2748                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2749                 RB_CLEAR_NODE(&ref->ref_node);
2750                 if (!list_empty(&ref->add_list))
2751                         list_del(&ref->add_list);
2752                 /*
2753                  * When we play the delayed ref, also correct the ref_mod on
2754                  * head
2755                  */
2756                 switch (ref->action) {
2757                 case BTRFS_ADD_DELAYED_REF:
2758                 case BTRFS_ADD_DELAYED_EXTENT:
2759                         locked_ref->ref_mod -= ref->ref_mod;
2760                         break;
2761                 case BTRFS_DROP_DELAYED_REF:
2762                         locked_ref->ref_mod += ref->ref_mod;
2763                         break;
2764                 default:
2765                         WARN_ON(1);
2766                 }
2767                 atomic_dec(&delayed_refs->num_entries);
2768
2769                 /*
2770                  * Record the must-insert_reserved flag before we drop the spin
2771                  * lock.
2772                  */
2773                 must_insert_reserved = locked_ref->must_insert_reserved;
2774                 locked_ref->must_insert_reserved = 0;
2775
2776                 extent_op = locked_ref->extent_op;
2777                 locked_ref->extent_op = NULL;
2778                 spin_unlock(&locked_ref->lock);
2779
2780                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2781                                           must_insert_reserved);
2782
2783                 btrfs_free_delayed_extent_op(extent_op);
2784                 if (ret) {
2785                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2786                         btrfs_put_delayed_ref(ref);
2787                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2788                                     ret);
2789                         return ret;
2790                 }
2791
2792                 btrfs_put_delayed_ref(ref);
2793                 count++;
2794                 cond_resched();
2795         }
2796
2797         /*
2798          * We don't want to include ref heads since we can have empty ref heads
2799          * and those will drastically skew our runtime down since we just do
2800          * accounting, no actual extent tree updates.
2801          */
2802         if (actual_count > 0) {
2803                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2804                 u64 avg;
2805
2806                 /*
2807                  * We weigh the current average higher than our current runtime
2808                  * to avoid large swings in the average.
2809                  */
2810                 spin_lock(&delayed_refs->lock);
2811                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2812                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2813                 spin_unlock(&delayed_refs->lock);
2814         }
2815         return 0;
2816 }
2817
2818 #ifdef SCRAMBLE_DELAYED_REFS
2819 /*
2820  * Normally delayed refs get processed in ascending bytenr order. This
2821  * correlates in most cases to the order added. To expose dependencies on this
2822  * order, we start to process the tree in the middle instead of the beginning
2823  */
2824 static u64 find_middle(struct rb_root *root)
2825 {
2826         struct rb_node *n = root->rb_node;
2827         struct btrfs_delayed_ref_node *entry;
2828         int alt = 1;
2829         u64 middle;
2830         u64 first = 0, last = 0;
2831
2832         n = rb_first(root);
2833         if (n) {
2834                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2835                 first = entry->bytenr;
2836         }
2837         n = rb_last(root);
2838         if (n) {
2839                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2840                 last = entry->bytenr;
2841         }
2842         n = root->rb_node;
2843
2844         while (n) {
2845                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2846                 WARN_ON(!entry->in_tree);
2847
2848                 middle = entry->bytenr;
2849
2850                 if (alt)
2851                         n = n->rb_left;
2852                 else
2853                         n = n->rb_right;
2854
2855                 alt = 1 - alt;
2856         }
2857         return middle;
2858 }
2859 #endif
2860
2861 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2862 {
2863         u64 num_bytes;
2864
2865         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2866                              sizeof(struct btrfs_extent_inline_ref));
2867         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2868                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2869
2870         /*
2871          * We don't ever fill up leaves all the way so multiply by 2 just to be
2872          * closer to what we're really going to want to use.
2873          */
2874         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2875 }
2876
2877 /*
2878  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2879  * would require to store the csums for that many bytes.
2880  */
2881 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2882 {
2883         u64 csum_size;
2884         u64 num_csums_per_leaf;
2885         u64 num_csums;
2886
2887         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2888         num_csums_per_leaf = div64_u64(csum_size,
2889                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2890         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2891         num_csums += num_csums_per_leaf - 1;
2892         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2893         return num_csums;
2894 }
2895
2896 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2897                                        struct btrfs_fs_info *fs_info)
2898 {
2899         struct btrfs_block_rsv *global_rsv;
2900         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2901         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2902         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2903         u64 num_bytes, num_dirty_bgs_bytes;
2904         int ret = 0;
2905
2906         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2907         num_heads = heads_to_leaves(fs_info, num_heads);
2908         if (num_heads > 1)
2909                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2910         num_bytes <<= 1;
2911         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2912                                                         fs_info->nodesize;
2913         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2914                                                              num_dirty_bgs);
2915         global_rsv = &fs_info->global_block_rsv;
2916
2917         /*
2918          * If we can't allocate any more chunks lets make sure we have _lots_ of
2919          * wiggle room since running delayed refs can create more delayed refs.
2920          */
2921         if (global_rsv->space_info->full) {
2922                 num_dirty_bgs_bytes <<= 1;
2923                 num_bytes <<= 1;
2924         }
2925
2926         spin_lock(&global_rsv->lock);
2927         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2928                 ret = 1;
2929         spin_unlock(&global_rsv->lock);
2930         return ret;
2931 }
2932
2933 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2934                                        struct btrfs_fs_info *fs_info)
2935 {
2936         u64 num_entries =
2937                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2938         u64 avg_runtime;
2939         u64 val;
2940
2941         smp_mb();
2942         avg_runtime = fs_info->avg_delayed_ref_runtime;
2943         val = num_entries * avg_runtime;
2944         if (val >= NSEC_PER_SEC)
2945                 return 1;
2946         if (val >= NSEC_PER_SEC / 2)
2947                 return 2;
2948
2949         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2950 }
2951
2952 struct async_delayed_refs {
2953         struct btrfs_root *root;
2954         u64 transid;
2955         int count;
2956         int error;
2957         int sync;
2958         struct completion wait;
2959         struct btrfs_work work;
2960 };
2961
2962 static inline struct async_delayed_refs *
2963 to_async_delayed_refs(struct btrfs_work *work)
2964 {
2965         return container_of(work, struct async_delayed_refs, work);
2966 }
2967
2968 static void delayed_ref_async_start(struct btrfs_work *work)
2969 {
2970         struct async_delayed_refs *async = to_async_delayed_refs(work);
2971         struct btrfs_trans_handle *trans;
2972         struct btrfs_fs_info *fs_info = async->root->fs_info;
2973         int ret;
2974
2975         /* if the commit is already started, we don't need to wait here */
2976         if (btrfs_transaction_blocked(fs_info))
2977                 goto done;
2978
2979         trans = btrfs_join_transaction(async->root);
2980         if (IS_ERR(trans)) {
2981                 async->error = PTR_ERR(trans);
2982                 goto done;
2983         }
2984
2985         /*
2986          * trans->sync means that when we call end_transaction, we won't
2987          * wait on delayed refs
2988          */
2989         trans->sync = true;
2990
2991         /* Don't bother flushing if we got into a different transaction */
2992         if (trans->transid > async->transid)
2993                 goto end;
2994
2995         ret = btrfs_run_delayed_refs(trans, async->count);
2996         if (ret)
2997                 async->error = ret;
2998 end:
2999         ret = btrfs_end_transaction(trans);
3000         if (ret && !async->error)
3001                 async->error = ret;
3002 done:
3003         if (async->sync)
3004                 complete(&async->wait);
3005         else
3006                 kfree(async);
3007 }
3008
3009 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3010                                  unsigned long count, u64 transid, int wait)
3011 {
3012         struct async_delayed_refs *async;
3013         int ret;
3014
3015         async = kmalloc(sizeof(*async), GFP_NOFS);
3016         if (!async)
3017                 return -ENOMEM;
3018
3019         async->root = fs_info->tree_root;
3020         async->count = count;
3021         async->error = 0;
3022         async->transid = transid;
3023         if (wait)
3024                 async->sync = 1;
3025         else
3026                 async->sync = 0;
3027         init_completion(&async->wait);
3028
3029         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3030                         delayed_ref_async_start, NULL, NULL);
3031
3032         btrfs_queue_work(fs_info->extent_workers, &async->work);
3033
3034         if (wait) {
3035                 wait_for_completion(&async->wait);
3036                 ret = async->error;
3037                 kfree(async);
3038                 return ret;
3039         }
3040         return 0;
3041 }
3042
3043 /*
3044  * this starts processing the delayed reference count updates and
3045  * extent insertions we have queued up so far.  count can be
3046  * 0, which means to process everything in the tree at the start
3047  * of the run (but not newly added entries), or it can be some target
3048  * number you'd like to process.
3049  *
3050  * Returns 0 on success or if called with an aborted transaction
3051  * Returns <0 on error and aborts the transaction
3052  */
3053 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3054                            unsigned long count)
3055 {
3056         struct btrfs_fs_info *fs_info = trans->fs_info;
3057         struct rb_node *node;
3058         struct btrfs_delayed_ref_root *delayed_refs;
3059         struct btrfs_delayed_ref_head *head;
3060         int ret;
3061         int run_all = count == (unsigned long)-1;
3062         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3063
3064         /* We'll clean this up in btrfs_cleanup_transaction */
3065         if (trans->aborted)
3066                 return 0;
3067
3068         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3069                 return 0;
3070
3071         delayed_refs = &trans->transaction->delayed_refs;
3072         if (count == 0)
3073                 count = atomic_read(&delayed_refs->num_entries) * 2;
3074
3075 again:
3076 #ifdef SCRAMBLE_DELAYED_REFS
3077         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3078 #endif
3079         trans->can_flush_pending_bgs = false;
3080         ret = __btrfs_run_delayed_refs(trans, fs_info, count);
3081         if (ret < 0) {
3082                 btrfs_abort_transaction(trans, ret);
3083                 return ret;
3084         }
3085
3086         if (run_all) {
3087                 if (!list_empty(&trans->new_bgs))
3088                         btrfs_create_pending_block_groups(trans);
3089
3090                 spin_lock(&delayed_refs->lock);
3091                 node = rb_first(&delayed_refs->href_root);
3092                 if (!node) {
3093                         spin_unlock(&delayed_refs->lock);
3094                         goto out;
3095                 }
3096                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3097                                 href_node);
3098                 refcount_inc(&head->refs);
3099                 spin_unlock(&delayed_refs->lock);
3100
3101                 /* Mutex was contended, block until it's released and retry. */
3102                 mutex_lock(&head->mutex);
3103                 mutex_unlock(&head->mutex);
3104
3105                 btrfs_put_delayed_ref_head(head);
3106                 cond_resched();
3107                 goto again;
3108         }
3109 out:
3110         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3111         return 0;
3112 }
3113
3114 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3115                                 struct btrfs_fs_info *fs_info,
3116                                 u64 bytenr, u64 num_bytes, u64 flags,
3117                                 int level, int is_data)
3118 {
3119         struct btrfs_delayed_extent_op *extent_op;
3120         int ret;
3121
3122         extent_op = btrfs_alloc_delayed_extent_op();
3123         if (!extent_op)
3124                 return -ENOMEM;
3125
3126         extent_op->flags_to_set = flags;
3127         extent_op->update_flags = true;
3128         extent_op->update_key = false;
3129         extent_op->is_data = is_data ? true : false;
3130         extent_op->level = level;
3131
3132         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3133                                           num_bytes, extent_op);
3134         if (ret)
3135                 btrfs_free_delayed_extent_op(extent_op);
3136         return ret;
3137 }
3138
3139 static noinline int check_delayed_ref(struct btrfs_root *root,
3140                                       struct btrfs_path *path,
3141                                       u64 objectid, u64 offset, u64 bytenr)
3142 {
3143         struct btrfs_delayed_ref_head *head;
3144         struct btrfs_delayed_ref_node *ref;
3145         struct btrfs_delayed_data_ref *data_ref;
3146         struct btrfs_delayed_ref_root *delayed_refs;
3147         struct btrfs_transaction *cur_trans;
3148         struct rb_node *node;
3149         int ret = 0;
3150
3151         cur_trans = root->fs_info->running_transaction;
3152         if (!cur_trans)
3153                 return 0;
3154
3155         delayed_refs = &cur_trans->delayed_refs;
3156         spin_lock(&delayed_refs->lock);
3157         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3158         if (!head) {
3159                 spin_unlock(&delayed_refs->lock);
3160                 return 0;
3161         }
3162
3163         if (!mutex_trylock(&head->mutex)) {
3164                 refcount_inc(&head->refs);
3165                 spin_unlock(&delayed_refs->lock);
3166
3167                 btrfs_release_path(path);
3168
3169                 /*
3170                  * Mutex was contended, block until it's released and let
3171                  * caller try again
3172                  */
3173                 mutex_lock(&head->mutex);
3174                 mutex_unlock(&head->mutex);
3175                 btrfs_put_delayed_ref_head(head);
3176                 return -EAGAIN;
3177         }
3178         spin_unlock(&delayed_refs->lock);
3179
3180         spin_lock(&head->lock);
3181         /*
3182          * XXX: We should replace this with a proper search function in the
3183          * future.
3184          */
3185         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3186                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3187                 /* If it's a shared ref we know a cross reference exists */
3188                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3189                         ret = 1;
3190                         break;
3191                 }
3192
3193                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3194
3195                 /*
3196                  * If our ref doesn't match the one we're currently looking at
3197                  * then we have a cross reference.
3198                  */
3199                 if (data_ref->root != root->root_key.objectid ||
3200                     data_ref->objectid != objectid ||
3201                     data_ref->offset != offset) {
3202                         ret = 1;
3203                         break;
3204                 }
3205         }
3206         spin_unlock(&head->lock);
3207         mutex_unlock(&head->mutex);
3208         return ret;
3209 }
3210
3211 static noinline int check_committed_ref(struct btrfs_root *root,
3212                                         struct btrfs_path *path,
3213                                         u64 objectid, u64 offset, u64 bytenr)
3214 {
3215         struct btrfs_fs_info *fs_info = root->fs_info;
3216         struct btrfs_root *extent_root = fs_info->extent_root;
3217         struct extent_buffer *leaf;
3218         struct btrfs_extent_data_ref *ref;
3219         struct btrfs_extent_inline_ref *iref;
3220         struct btrfs_extent_item *ei;
3221         struct btrfs_key key;
3222         u32 item_size;
3223         int type;
3224         int ret;
3225
3226         key.objectid = bytenr;
3227         key.offset = (u64)-1;
3228         key.type = BTRFS_EXTENT_ITEM_KEY;
3229
3230         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3231         if (ret < 0)
3232                 goto out;
3233         BUG_ON(ret == 0); /* Corruption */
3234
3235         ret = -ENOENT;
3236         if (path->slots[0] == 0)
3237                 goto out;
3238
3239         path->slots[0]--;
3240         leaf = path->nodes[0];
3241         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3242
3243         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3244                 goto out;
3245
3246         ret = 1;
3247         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3248 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3249         if (item_size < sizeof(*ei)) {
3250                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3251                 goto out;
3252         }
3253 #endif
3254         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3255
3256         if (item_size != sizeof(*ei) +
3257             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3258                 goto out;
3259
3260         if (btrfs_extent_generation(leaf, ei) <=
3261             btrfs_root_last_snapshot(&root->root_item))
3262                 goto out;
3263
3264         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3265
3266         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3267         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3268                 goto out;
3269
3270         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3271         if (btrfs_extent_refs(leaf, ei) !=
3272             btrfs_extent_data_ref_count(leaf, ref) ||
3273             btrfs_extent_data_ref_root(leaf, ref) !=
3274             root->root_key.objectid ||
3275             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3276             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3277                 goto out;
3278
3279         ret = 0;
3280 out:
3281         return ret;
3282 }
3283
3284 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3285                           u64 bytenr)
3286 {
3287         struct btrfs_path *path;
3288         int ret;
3289         int ret2;
3290
3291         path = btrfs_alloc_path();
3292         if (!path)
3293                 return -ENOENT;
3294
3295         do {
3296                 ret = check_committed_ref(root, path, objectid,
3297                                           offset, bytenr);
3298                 if (ret && ret != -ENOENT)
3299                         goto out;
3300
3301                 ret2 = check_delayed_ref(root, path, objectid,
3302                                          offset, bytenr);
3303         } while (ret2 == -EAGAIN);
3304
3305         if (ret2 && ret2 != -ENOENT) {
3306                 ret = ret2;
3307                 goto out;
3308         }
3309
3310         if (ret != -ENOENT || ret2 != -ENOENT)
3311                 ret = 0;
3312 out:
3313         btrfs_free_path(path);
3314         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3315                 WARN_ON(ret > 0);
3316         return ret;
3317 }
3318
3319 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3320                            struct btrfs_root *root,
3321                            struct extent_buffer *buf,
3322                            int full_backref, int inc)
3323 {
3324         struct btrfs_fs_info *fs_info = root->fs_info;
3325         u64 bytenr;
3326         u64 num_bytes;
3327         u64 parent;
3328         u64 ref_root;
3329         u32 nritems;
3330         struct btrfs_key key;
3331         struct btrfs_file_extent_item *fi;
3332         int i;
3333         int level;
3334         int ret = 0;
3335         int (*process_func)(struct btrfs_trans_handle *,
3336                             struct btrfs_root *,
3337                             u64, u64, u64, u64, u64, u64);
3338
3339
3340         if (btrfs_is_testing(fs_info))
3341                 return 0;
3342
3343         ref_root = btrfs_header_owner(buf);
3344         nritems = btrfs_header_nritems(buf);
3345         level = btrfs_header_level(buf);
3346
3347         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3348                 return 0;
3349
3350         if (inc)
3351                 process_func = btrfs_inc_extent_ref;
3352         else
3353                 process_func = btrfs_free_extent;
3354
3355         if (full_backref)
3356                 parent = buf->start;
3357         else
3358                 parent = 0;
3359
3360         for (i = 0; i < nritems; i++) {
3361                 if (level == 0) {
3362                         btrfs_item_key_to_cpu(buf, &key, i);
3363                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3364                                 continue;
3365                         fi = btrfs_item_ptr(buf, i,
3366                                             struct btrfs_file_extent_item);
3367                         if (btrfs_file_extent_type(buf, fi) ==
3368                             BTRFS_FILE_EXTENT_INLINE)
3369                                 continue;
3370                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3371                         if (bytenr == 0)
3372                                 continue;
3373
3374                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3375                         key.offset -= btrfs_file_extent_offset(buf, fi);
3376                         ret = process_func(trans, root, bytenr, num_bytes,
3377                                            parent, ref_root, key.objectid,
3378                                            key.offset);
3379                         if (ret)
3380                                 goto fail;
3381                 } else {
3382                         bytenr = btrfs_node_blockptr(buf, i);
3383                         num_bytes = fs_info->nodesize;
3384                         ret = process_func(trans, root, bytenr, num_bytes,
3385                                            parent, ref_root, level - 1, 0);
3386                         if (ret)
3387                                 goto fail;
3388                 }
3389         }
3390         return 0;
3391 fail:
3392         return ret;
3393 }
3394
3395 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3396                   struct extent_buffer *buf, int full_backref)
3397 {
3398         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3399 }
3400
3401 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3402                   struct extent_buffer *buf, int full_backref)
3403 {
3404         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3405 }
3406
3407 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3408                                  struct btrfs_fs_info *fs_info,
3409                                  struct btrfs_path *path,
3410                                  struct btrfs_block_group_cache *cache)
3411 {
3412         int ret;
3413         struct btrfs_root *extent_root = fs_info->extent_root;
3414         unsigned long bi;
3415         struct extent_buffer *leaf;
3416
3417         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3418         if (ret) {
3419                 if (ret > 0)
3420                         ret = -ENOENT;
3421                 goto fail;
3422         }
3423
3424         leaf = path->nodes[0];
3425         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3426         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3427         btrfs_mark_buffer_dirty(leaf);
3428 fail:
3429         btrfs_release_path(path);
3430         return ret;
3431
3432 }
3433
3434 static struct btrfs_block_group_cache *
3435 next_block_group(struct btrfs_fs_info *fs_info,
3436                  struct btrfs_block_group_cache *cache)
3437 {
3438         struct rb_node *node;
3439
3440         spin_lock(&fs_info->block_group_cache_lock);
3441
3442         /* If our block group was removed, we need a full search. */
3443         if (RB_EMPTY_NODE(&cache->cache_node)) {
3444                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3445
3446                 spin_unlock(&fs_info->block_group_cache_lock);
3447                 btrfs_put_block_group(cache);
3448                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3449         }
3450         node = rb_next(&cache->cache_node);
3451         btrfs_put_block_group(cache);
3452         if (node) {
3453                 cache = rb_entry(node, struct btrfs_block_group_cache,
3454                                  cache_node);
3455                 btrfs_get_block_group(cache);
3456         } else
3457                 cache = NULL;
3458         spin_unlock(&fs_info->block_group_cache_lock);
3459         return cache;
3460 }
3461
3462 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3463                             struct btrfs_trans_handle *trans,
3464                             struct btrfs_path *path)
3465 {
3466         struct btrfs_fs_info *fs_info = block_group->fs_info;
3467         struct btrfs_root *root = fs_info->tree_root;
3468         struct inode *inode = NULL;
3469         struct extent_changeset *data_reserved = NULL;
3470         u64 alloc_hint = 0;
3471         int dcs = BTRFS_DC_ERROR;
3472         u64 num_pages = 0;
3473         int retries = 0;
3474         int ret = 0;
3475
3476         /*
3477          * If this block group is smaller than 100 megs don't bother caching the
3478          * block group.
3479          */
3480         if (block_group->key.offset < (100 * SZ_1M)) {
3481                 spin_lock(&block_group->lock);
3482                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3483                 spin_unlock(&block_group->lock);
3484                 return 0;
3485         }
3486
3487         if (trans->aborted)
3488                 return 0;
3489 again:
3490         inode = lookup_free_space_inode(fs_info, block_group, path);
3491         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3492                 ret = PTR_ERR(inode);
3493                 btrfs_release_path(path);
3494                 goto out;
3495         }
3496
3497         if (IS_ERR(inode)) {
3498                 BUG_ON(retries);
3499                 retries++;
3500
3501                 if (block_group->ro)
3502                         goto out_free;
3503
3504                 ret = create_free_space_inode(fs_info, trans, block_group,
3505                                               path);
3506                 if (ret)
3507                         goto out_free;
3508                 goto again;
3509         }
3510
3511         /*
3512          * We want to set the generation to 0, that way if anything goes wrong
3513          * from here on out we know not to trust this cache when we load up next
3514          * time.
3515          */
3516         BTRFS_I(inode)->generation = 0;
3517         ret = btrfs_update_inode(trans, root, inode);
3518         if (ret) {
3519                 /*
3520                  * So theoretically we could recover from this, simply set the
3521                  * super cache generation to 0 so we know to invalidate the
3522                  * cache, but then we'd have to keep track of the block groups
3523                  * that fail this way so we know we _have_ to reset this cache
3524                  * before the next commit or risk reading stale cache.  So to
3525                  * limit our exposure to horrible edge cases lets just abort the
3526                  * transaction, this only happens in really bad situations
3527                  * anyway.
3528                  */
3529                 btrfs_abort_transaction(trans, ret);
3530                 goto out_put;
3531         }
3532         WARN_ON(ret);
3533
3534         /* We've already setup this transaction, go ahead and exit */
3535         if (block_group->cache_generation == trans->transid &&
3536             i_size_read(inode)) {
3537                 dcs = BTRFS_DC_SETUP;
3538                 goto out_put;
3539         }
3540
3541         if (i_size_read(inode) > 0) {
3542                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3543                                         &fs_info->global_block_rsv);
3544                 if (ret)
3545                         goto out_put;
3546
3547                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3548                 if (ret)
3549                         goto out_put;
3550         }
3551
3552         spin_lock(&block_group->lock);
3553         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3554             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3555                 /*
3556                  * don't bother trying to write stuff out _if_
3557                  * a) we're not cached,
3558                  * b) we're with nospace_cache mount option,
3559                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3560                  */
3561                 dcs = BTRFS_DC_WRITTEN;
3562                 spin_unlock(&block_group->lock);
3563                 goto out_put;
3564         }
3565         spin_unlock(&block_group->lock);
3566
3567         /*
3568          * We hit an ENOSPC when setting up the cache in this transaction, just
3569          * skip doing the setup, we've already cleared the cache so we're safe.
3570          */
3571         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3572                 ret = -ENOSPC;
3573                 goto out_put;
3574         }
3575
3576         /*
3577          * Try to preallocate enough space based on how big the block group is.
3578          * Keep in mind this has to include any pinned space which could end up
3579          * taking up quite a bit since it's not folded into the other space
3580          * cache.
3581          */
3582         num_pages = div_u64(block_group->key.offset, SZ_256M);
3583         if (!num_pages)
3584                 num_pages = 1;
3585
3586         num_pages *= 16;
3587         num_pages *= PAGE_SIZE;
3588
3589         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3590         if (ret)
3591                 goto out_put;
3592
3593         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3594                                               num_pages, num_pages,
3595                                               &alloc_hint);
3596         /*
3597          * Our cache requires contiguous chunks so that we don't modify a bunch
3598          * of metadata or split extents when writing the cache out, which means
3599          * we can enospc if we are heavily fragmented in addition to just normal
3600          * out of space conditions.  So if we hit this just skip setting up any
3601          * other block groups for this transaction, maybe we'll unpin enough
3602          * space the next time around.
3603          */
3604         if (!ret)
3605                 dcs = BTRFS_DC_SETUP;
3606         else if (ret == -ENOSPC)
3607                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3608
3609 out_put:
3610         iput(inode);
3611 out_free:
3612         btrfs_release_path(path);
3613 out:
3614         spin_lock(&block_group->lock);
3615         if (!ret && dcs == BTRFS_DC_SETUP)
3616                 block_group->cache_generation = trans->transid;
3617         block_group->disk_cache_state = dcs;
3618         spin_unlock(&block_group->lock);
3619
3620         extent_changeset_free(data_reserved);
3621         return ret;
3622 }
3623
3624 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3625                             struct btrfs_fs_info *fs_info)
3626 {
3627         struct btrfs_block_group_cache *cache, *tmp;
3628         struct btrfs_transaction *cur_trans = trans->transaction;
3629         struct btrfs_path *path;
3630
3631         if (list_empty(&cur_trans->dirty_bgs) ||
3632             !btrfs_test_opt(fs_info, SPACE_CACHE))
3633                 return 0;
3634
3635         path = btrfs_alloc_path();
3636         if (!path)
3637                 return -ENOMEM;
3638
3639         /* Could add new block groups, use _safe just in case */
3640         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3641                                  dirty_list) {
3642                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3643                         cache_save_setup(cache, trans, path);
3644         }
3645
3646         btrfs_free_path(path);
3647         return 0;
3648 }
3649
3650 /*
3651  * transaction commit does final block group cache writeback during a
3652  * critical section where nothing is allowed to change the FS.  This is
3653  * required in order for the cache to actually match the block group,
3654  * but can introduce a lot of latency into the commit.
3655  *
3656  * So, btrfs_start_dirty_block_groups is here to kick off block group
3657  * cache IO.  There's a chance we'll have to redo some of it if the
3658  * block group changes again during the commit, but it greatly reduces
3659  * the commit latency by getting rid of the easy block groups while
3660  * we're still allowing others to join the commit.
3661  */
3662 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3663 {
3664         struct btrfs_fs_info *fs_info = trans->fs_info;
3665         struct btrfs_block_group_cache *cache;
3666         struct btrfs_transaction *cur_trans = trans->transaction;
3667         int ret = 0;
3668         int should_put;
3669         struct btrfs_path *path = NULL;
3670         LIST_HEAD(dirty);
3671         struct list_head *io = &cur_trans->io_bgs;
3672         int num_started = 0;
3673         int loops = 0;
3674
3675         spin_lock(&cur_trans->dirty_bgs_lock);
3676         if (list_empty(&cur_trans->dirty_bgs)) {
3677                 spin_unlock(&cur_trans->dirty_bgs_lock);
3678                 return 0;
3679         }
3680         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3681         spin_unlock(&cur_trans->dirty_bgs_lock);
3682
3683 again:
3684         /*
3685          * make sure all the block groups on our dirty list actually
3686          * exist
3687          */
3688         btrfs_create_pending_block_groups(trans);
3689
3690         if (!path) {
3691                 path = btrfs_alloc_path();
3692                 if (!path)
3693                         return -ENOMEM;
3694         }
3695
3696         /*
3697          * cache_write_mutex is here only to save us from balance or automatic
3698          * removal of empty block groups deleting this block group while we are
3699          * writing out the cache
3700          */
3701         mutex_lock(&trans->transaction->cache_write_mutex);
3702         while (!list_empty(&dirty)) {
3703                 cache = list_first_entry(&dirty,
3704                                          struct btrfs_block_group_cache,
3705                                          dirty_list);
3706                 /*
3707                  * this can happen if something re-dirties a block
3708                  * group that is already under IO.  Just wait for it to
3709                  * finish and then do it all again
3710                  */
3711                 if (!list_empty(&cache->io_list)) {
3712                         list_del_init(&cache->io_list);
3713                         btrfs_wait_cache_io(trans, cache, path);
3714                         btrfs_put_block_group(cache);
3715                 }
3716
3717
3718                 /*
3719                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3720                  * if it should update the cache_state.  Don't delete
3721                  * until after we wait.
3722                  *
3723                  * Since we're not running in the commit critical section
3724                  * we need the dirty_bgs_lock to protect from update_block_group
3725                  */
3726                 spin_lock(&cur_trans->dirty_bgs_lock);
3727                 list_del_init(&cache->dirty_list);
3728                 spin_unlock(&cur_trans->dirty_bgs_lock);
3729
3730                 should_put = 1;
3731
3732                 cache_save_setup(cache, trans, path);
3733
3734                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3735                         cache->io_ctl.inode = NULL;
3736                         ret = btrfs_write_out_cache(fs_info, trans,
3737                                                     cache, path);
3738                         if (ret == 0 && cache->io_ctl.inode) {
3739                                 num_started++;
3740                                 should_put = 0;
3741
3742                                 /*
3743                                  * The cache_write_mutex is protecting the
3744                                  * io_list, also refer to the definition of
3745                                  * btrfs_transaction::io_bgs for more details
3746                                  */
3747                                 list_add_tail(&cache->io_list, io);
3748                         } else {
3749                                 /*
3750                                  * if we failed to write the cache, the
3751                                  * generation will be bad and life goes on
3752                                  */
3753                                 ret = 0;
3754                         }
3755                 }
3756                 if (!ret) {
3757                         ret = write_one_cache_group(trans, fs_info,
3758                                                     path, cache);
3759                         /*
3760                          * Our block group might still be attached to the list
3761                          * of new block groups in the transaction handle of some
3762                          * other task (struct btrfs_trans_handle->new_bgs). This
3763                          * means its block group item isn't yet in the extent
3764                          * tree. If this happens ignore the error, as we will
3765                          * try again later in the critical section of the
3766                          * transaction commit.
3767                          */
3768                         if (ret == -ENOENT) {
3769                                 ret = 0;
3770                                 spin_lock(&cur_trans->dirty_bgs_lock);
3771                                 if (list_empty(&cache->dirty_list)) {
3772                                         list_add_tail(&cache->dirty_list,
3773                                                       &cur_trans->dirty_bgs);
3774                                         btrfs_get_block_group(cache);
3775                                 }
3776                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3777                         } else if (ret) {
3778                                 btrfs_abort_transaction(trans, ret);
3779                         }
3780                 }
3781
3782                 /* if its not on the io list, we need to put the block group */
3783                 if (should_put)
3784                         btrfs_put_block_group(cache);
3785
3786                 if (ret)
3787                         break;
3788
3789                 /*
3790                  * Avoid blocking other tasks for too long. It might even save
3791                  * us from writing caches for block groups that are going to be
3792                  * removed.
3793                  */
3794                 mutex_unlock(&trans->transaction->cache_write_mutex);
3795                 mutex_lock(&trans->transaction->cache_write_mutex);
3796         }
3797         mutex_unlock(&trans->transaction->cache_write_mutex);
3798
3799         /*
3800          * go through delayed refs for all the stuff we've just kicked off
3801          * and then loop back (just once)
3802          */
3803         ret = btrfs_run_delayed_refs(trans, 0);
3804         if (!ret && loops == 0) {
3805                 loops++;
3806                 spin_lock(&cur_trans->dirty_bgs_lock);
3807                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3808                 /*
3809                  * dirty_bgs_lock protects us from concurrent block group
3810                  * deletes too (not just cache_write_mutex).
3811                  */
3812                 if (!list_empty(&dirty)) {
3813                         spin_unlock(&cur_trans->dirty_bgs_lock);
3814                         goto again;
3815                 }
3816                 spin_unlock(&cur_trans->dirty_bgs_lock);
3817         } else if (ret < 0) {
3818                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3819         }
3820
3821         btrfs_free_path(path);
3822         return ret;
3823 }
3824
3825 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3826                                    struct btrfs_fs_info *fs_info)
3827 {
3828         struct btrfs_block_group_cache *cache;
3829         struct btrfs_transaction *cur_trans = trans->transaction;
3830         int ret = 0;
3831         int should_put;
3832         struct btrfs_path *path;
3833         struct list_head *io = &cur_trans->io_bgs;
3834         int num_started = 0;
3835
3836         path = btrfs_alloc_path();
3837         if (!path)
3838                 return -ENOMEM;
3839
3840         /*
3841          * Even though we are in the critical section of the transaction commit,
3842          * we can still have concurrent tasks adding elements to this
3843          * transaction's list of dirty block groups. These tasks correspond to
3844          * endio free space workers started when writeback finishes for a
3845          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3846          * allocate new block groups as a result of COWing nodes of the root
3847          * tree when updating the free space inode. The writeback for the space
3848          * caches is triggered by an earlier call to
3849          * btrfs_start_dirty_block_groups() and iterations of the following
3850          * loop.
3851          * Also we want to do the cache_save_setup first and then run the
3852          * delayed refs to make sure we have the best chance at doing this all
3853          * in one shot.
3854          */
3855         spin_lock(&cur_trans->dirty_bgs_lock);
3856         while (!list_empty(&cur_trans->dirty_bgs)) {
3857                 cache = list_first_entry(&cur_trans->dirty_bgs,
3858                                          struct btrfs_block_group_cache,
3859                                          dirty_list);
3860
3861                 /*
3862                  * this can happen if cache_save_setup re-dirties a block
3863                  * group that is already under IO.  Just wait for it to
3864                  * finish and then do it all again
3865                  */
3866                 if (!list_empty(&cache->io_list)) {
3867                         spin_unlock(&cur_trans->dirty_bgs_lock);
3868                         list_del_init(&cache->io_list);
3869                         btrfs_wait_cache_io(trans, cache, path);
3870                         btrfs_put_block_group(cache);
3871                         spin_lock(&cur_trans->dirty_bgs_lock);
3872                 }
3873
3874                 /*
3875                  * don't remove from the dirty list until after we've waited
3876                  * on any pending IO
3877                  */
3878                 list_del_init(&cache->dirty_list);
3879                 spin_unlock(&cur_trans->dirty_bgs_lock);
3880                 should_put = 1;
3881
3882                 cache_save_setup(cache, trans, path);
3883
3884                 if (!ret)
3885                         ret = btrfs_run_delayed_refs(trans,
3886                                                      (unsigned long) -1);
3887
3888                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3889                         cache->io_ctl.inode = NULL;
3890                         ret = btrfs_write_out_cache(fs_info, trans,
3891                                                     cache, path);
3892                         if (ret == 0 && cache->io_ctl.inode) {
3893                                 num_started++;
3894                                 should_put = 0;
3895                                 list_add_tail(&cache->io_list, io);
3896                         } else {
3897                                 /*
3898                                  * if we failed to write the cache, the
3899                                  * generation will be bad and life goes on
3900                                  */
3901                                 ret = 0;
3902                         }
3903                 }
3904                 if (!ret) {
3905                         ret = write_one_cache_group(trans, fs_info,
3906                                                     path, cache);
3907                         /*
3908                          * One of the free space endio workers might have
3909                          * created a new block group while updating a free space
3910                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3911                          * and hasn't released its transaction handle yet, in
3912                          * which case the new block group is still attached to
3913                          * its transaction handle and its creation has not
3914                          * finished yet (no block group item in the extent tree
3915                          * yet, etc). If this is the case, wait for all free
3916                          * space endio workers to finish and retry. This is a
3917                          * a very rare case so no need for a more efficient and
3918                          * complex approach.
3919                          */
3920                         if (ret == -ENOENT) {
3921                                 wait_event(cur_trans->writer_wait,
3922                                    atomic_read(&cur_trans->num_writers) == 1);
3923                                 ret = write_one_cache_group(trans, fs_info,
3924                                                             path, cache);
3925                         }
3926                         if (ret)
3927                                 btrfs_abort_transaction(trans, ret);
3928                 }
3929
3930                 /* if its not on the io list, we need to put the block group */
3931                 if (should_put)
3932                         btrfs_put_block_group(cache);
3933                 spin_lock(&cur_trans->dirty_bgs_lock);
3934         }
3935         spin_unlock(&cur_trans->dirty_bgs_lock);
3936
3937         /*
3938          * Refer to the definition of io_bgs member for details why it's safe
3939          * to use it without any locking
3940          */
3941         while (!list_empty(io)) {
3942                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3943                                          io_list);
3944                 list_del_init(&cache->io_list);
3945                 btrfs_wait_cache_io(trans, cache, path);
3946                 btrfs_put_block_group(cache);
3947         }
3948
3949         btrfs_free_path(path);
3950         return ret;
3951 }
3952
3953 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3954 {
3955         struct btrfs_block_group_cache *block_group;
3956         int readonly = 0;
3957
3958         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3959         if (!block_group || block_group->ro)
3960                 readonly = 1;
3961         if (block_group)
3962                 btrfs_put_block_group(block_group);
3963         return readonly;
3964 }
3965
3966 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3967 {
3968         struct btrfs_block_group_cache *bg;
3969         bool ret = true;
3970
3971         bg = btrfs_lookup_block_group(fs_info, bytenr);
3972         if (!bg)
3973                 return false;
3974
3975         spin_lock(&bg->lock);
3976         if (bg->ro)
3977                 ret = false;
3978         else
3979                 atomic_inc(&bg->nocow_writers);
3980         spin_unlock(&bg->lock);
3981
3982         /* no put on block group, done by btrfs_dec_nocow_writers */
3983         if (!ret)
3984                 btrfs_put_block_group(bg);
3985
3986         return ret;
3987
3988 }
3989
3990 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3991 {
3992         struct btrfs_block_group_cache *bg;
3993
3994         bg = btrfs_lookup_block_group(fs_info, bytenr);
3995         ASSERT(bg);
3996         if (atomic_dec_and_test(&bg->nocow_writers))
3997                 wake_up_atomic_t(&bg->nocow_writers);
3998         /*
3999          * Once for our lookup and once for the lookup done by a previous call
4000          * to btrfs_inc_nocow_writers()
4001          */
4002         btrfs_put_block_group(bg);
4003         btrfs_put_block_group(bg);
4004 }
4005
4006 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
4007 {
4008         wait_on_atomic_t(&bg->nocow_writers, atomic_t_wait,
4009                          TASK_UNINTERRUPTIBLE);
4010 }
4011
4012 static const char *alloc_name(u64 flags)
4013 {
4014         switch (flags) {
4015         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4016                 return "mixed";
4017         case BTRFS_BLOCK_GROUP_METADATA:
4018                 return "metadata";
4019         case BTRFS_BLOCK_GROUP_DATA:
4020                 return "data";
4021         case BTRFS_BLOCK_GROUP_SYSTEM:
4022                 return "system";
4023         default:
4024                 WARN_ON(1);
4025                 return "invalid-combination";
4026         };
4027 }
4028
4029 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4030                              struct btrfs_space_info **new)
4031 {
4032
4033         struct btrfs_space_info *space_info;
4034         int i;
4035         int ret;
4036
4037         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4038         if (!space_info)
4039                 return -ENOMEM;
4040
4041         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4042                                  GFP_KERNEL);
4043         if (ret) {
4044                 kfree(space_info);
4045                 return ret;
4046         }
4047
4048         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4049                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4050         init_rwsem(&space_info->groups_sem);
4051         spin_lock_init(&space_info->lock);
4052         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4053         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4054         init_waitqueue_head(&space_info->wait);
4055         INIT_LIST_HEAD(&space_info->ro_bgs);
4056         INIT_LIST_HEAD(&space_info->tickets);
4057         INIT_LIST_HEAD(&space_info->priority_tickets);
4058
4059         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4060                                     info->space_info_kobj, "%s",
4061                                     alloc_name(space_info->flags));
4062         if (ret) {
4063                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4064                 kfree(space_info);
4065                 return ret;
4066         }
4067
4068         *new = space_info;
4069         list_add_rcu(&space_info->list, &info->space_info);
4070         if (flags & BTRFS_BLOCK_GROUP_DATA)
4071                 info->data_sinfo = space_info;
4072
4073         return ret;
4074 }
4075
4076 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4077                              u64 total_bytes, u64 bytes_used,
4078                              u64 bytes_readonly,
4079                              struct btrfs_space_info **space_info)
4080 {
4081         struct btrfs_space_info *found;
4082         int factor;
4083
4084         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4085                      BTRFS_BLOCK_GROUP_RAID10))
4086                 factor = 2;
4087         else
4088                 factor = 1;
4089
4090         found = __find_space_info(info, flags);
4091         ASSERT(found);
4092         spin_lock(&found->lock);
4093         found->total_bytes += total_bytes;
4094         found->disk_total += total_bytes * factor;
4095         found->bytes_used += bytes_used;
4096         found->disk_used += bytes_used * factor;
4097         found->bytes_readonly += bytes_readonly;
4098         if (total_bytes > 0)
4099                 found->full = 0;
4100         space_info_add_new_bytes(info, found, total_bytes -
4101                                  bytes_used - bytes_readonly);
4102         spin_unlock(&found->lock);
4103         *space_info = found;
4104 }
4105
4106 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4107 {
4108         u64 extra_flags = chunk_to_extended(flags) &
4109                                 BTRFS_EXTENDED_PROFILE_MASK;
4110
4111         write_seqlock(&fs_info->profiles_lock);
4112         if (flags & BTRFS_BLOCK_GROUP_DATA)
4113                 fs_info->avail_data_alloc_bits |= extra_flags;
4114         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4115                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4116         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4117                 fs_info->avail_system_alloc_bits |= extra_flags;
4118         write_sequnlock(&fs_info->profiles_lock);
4119 }
4120
4121 /*
4122  * returns target flags in extended format or 0 if restripe for this
4123  * chunk_type is not in progress
4124  *
4125  * should be called with either volume_mutex or balance_lock held
4126  */
4127 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4128 {
4129         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4130         u64 target = 0;
4131
4132         if (!bctl)
4133                 return 0;
4134
4135         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4136             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4137                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4138         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4139                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4140                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4141         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4142                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4143                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4144         }
4145
4146         return target;
4147 }
4148
4149 /*
4150  * @flags: available profiles in extended format (see ctree.h)
4151  *
4152  * Returns reduced profile in chunk format.  If profile changing is in
4153  * progress (either running or paused) picks the target profile (if it's
4154  * already available), otherwise falls back to plain reducing.
4155  */
4156 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4157 {
4158         u64 num_devices = fs_info->fs_devices->rw_devices;
4159         u64 target;
4160         u64 raid_type;
4161         u64 allowed = 0;
4162
4163         /*
4164          * see if restripe for this chunk_type is in progress, if so
4165          * try to reduce to the target profile
4166          */
4167         spin_lock(&fs_info->balance_lock);
4168         target = get_restripe_target(fs_info, flags);
4169         if (target) {
4170                 /* pick target profile only if it's already available */
4171                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4172                         spin_unlock(&fs_info->balance_lock);
4173                         return extended_to_chunk(target);
4174                 }
4175         }
4176         spin_unlock(&fs_info->balance_lock);
4177
4178         /* First, mask out the RAID levels which aren't possible */
4179         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4180                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4181                         allowed |= btrfs_raid_group[raid_type];
4182         }
4183         allowed &= flags;
4184
4185         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4186                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4187         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4188                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4189         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4190                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4191         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4192                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4193         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4194                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4195
4196         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4197
4198         return extended_to_chunk(flags | allowed);
4199 }
4200
4201 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4202 {
4203         unsigned seq;
4204         u64 flags;
4205
4206         do {
4207                 flags = orig_flags;
4208                 seq = read_seqbegin(&fs_info->profiles_lock);
4209
4210                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4211                         flags |= fs_info->avail_data_alloc_bits;
4212                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4213                         flags |= fs_info->avail_system_alloc_bits;
4214                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4215                         flags |= fs_info->avail_metadata_alloc_bits;
4216         } while (read_seqretry(&fs_info->profiles_lock, seq));
4217
4218         return btrfs_reduce_alloc_profile(fs_info, flags);
4219 }
4220
4221 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4222 {
4223         struct btrfs_fs_info *fs_info = root->fs_info;
4224         u64 flags;
4225         u64 ret;
4226
4227         if (data)
4228                 flags = BTRFS_BLOCK_GROUP_DATA;
4229         else if (root == fs_info->chunk_root)
4230                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4231         else
4232                 flags = BTRFS_BLOCK_GROUP_METADATA;
4233
4234         ret = get_alloc_profile(fs_info, flags);
4235         return ret;
4236 }
4237
4238 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4239 {
4240         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4241 }
4242
4243 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4244 {
4245         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4246 }
4247
4248 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4249 {
4250         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4251 }
4252
4253 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4254                                  bool may_use_included)
4255 {
4256         ASSERT(s_info);
4257         return s_info->bytes_used + s_info->bytes_reserved +
4258                 s_info->bytes_pinned + s_info->bytes_readonly +
4259                 (may_use_included ? s_info->bytes_may_use : 0);
4260 }
4261
4262 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4263 {
4264         struct btrfs_root *root = inode->root;
4265         struct btrfs_fs_info *fs_info = root->fs_info;
4266         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4267         u64 used;
4268         int ret = 0;
4269         int need_commit = 2;
4270         int have_pinned_space;
4271
4272         /* make sure bytes are sectorsize aligned */
4273         bytes = ALIGN(bytes, fs_info->sectorsize);
4274
4275         if (btrfs_is_free_space_inode(inode)) {
4276                 need_commit = 0;
4277                 ASSERT(current->journal_info);
4278         }
4279
4280 again:
4281         /* make sure we have enough space to handle the data first */
4282         spin_lock(&data_sinfo->lock);
4283         used = btrfs_space_info_used(data_sinfo, true);
4284
4285         if (used + bytes > data_sinfo->total_bytes) {
4286                 struct btrfs_trans_handle *trans;
4287
4288                 /*
4289                  * if we don't have enough free bytes in this space then we need
4290                  * to alloc a new chunk.
4291                  */
4292                 if (!data_sinfo->full) {
4293                         u64 alloc_target;
4294
4295                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4296                         spin_unlock(&data_sinfo->lock);
4297
4298                         alloc_target = btrfs_data_alloc_profile(fs_info);
4299                         /*
4300                          * It is ugly that we don't call nolock join
4301                          * transaction for the free space inode case here.
4302                          * But it is safe because we only do the data space
4303                          * reservation for the free space cache in the
4304                          * transaction context, the common join transaction
4305                          * just increase the counter of the current transaction
4306                          * handler, doesn't try to acquire the trans_lock of
4307                          * the fs.
4308                          */
4309                         trans = btrfs_join_transaction(root);
4310                         if (IS_ERR(trans))
4311                                 return PTR_ERR(trans);
4312
4313                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4314                                              CHUNK_ALLOC_NO_FORCE);
4315                         btrfs_end_transaction(trans);
4316                         if (ret < 0) {
4317                                 if (ret != -ENOSPC)
4318                                         return ret;
4319                                 else {
4320                                         have_pinned_space = 1;
4321                                         goto commit_trans;
4322                                 }
4323                         }
4324
4325                         goto again;
4326                 }
4327
4328                 /*
4329                  * If we don't have enough pinned space to deal with this
4330                  * allocation, and no removed chunk in current transaction,
4331                  * don't bother committing the transaction.
4332                  */
4333                 have_pinned_space = percpu_counter_compare(
4334                         &data_sinfo->total_bytes_pinned,
4335                         used + bytes - data_sinfo->total_bytes);
4336                 spin_unlock(&data_sinfo->lock);
4337
4338                 /* commit the current transaction and try again */
4339 commit_trans:
4340                 if (need_commit) {
4341                         need_commit--;
4342
4343                         if (need_commit > 0) {
4344                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4345                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4346                                                          (u64)-1);
4347                         }
4348
4349                         trans = btrfs_join_transaction(root);
4350                         if (IS_ERR(trans))
4351                                 return PTR_ERR(trans);
4352                         if (have_pinned_space >= 0 ||
4353                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4354                                      &trans->transaction->flags) ||
4355                             need_commit > 0) {
4356                                 ret = btrfs_commit_transaction(trans);
4357                                 if (ret)
4358                                         return ret;
4359                                 /*
4360                                  * The cleaner kthread might still be doing iput
4361                                  * operations. Wait for it to finish so that
4362                                  * more space is released.
4363                                  */
4364                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4365                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4366                                 goto again;
4367                         } else {
4368                                 btrfs_end_transaction(trans);
4369                         }
4370                 }
4371
4372                 trace_btrfs_space_reservation(fs_info,
4373                                               "space_info:enospc",
4374                                               data_sinfo->flags, bytes, 1);
4375                 return -ENOSPC;
4376         }
4377         data_sinfo->bytes_may_use += bytes;
4378         trace_btrfs_space_reservation(fs_info, "space_info",
4379                                       data_sinfo->flags, bytes, 1);
4380         spin_unlock(&data_sinfo->lock);
4381
4382         return ret;
4383 }
4384
4385 int btrfs_check_data_free_space(struct inode *inode,
4386                         struct extent_changeset **reserved, u64 start, u64 len)
4387 {
4388         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4389         int ret;
4390
4391         /* align the range */
4392         len = round_up(start + len, fs_info->sectorsize) -
4393               round_down(start, fs_info->sectorsize);
4394         start = round_down(start, fs_info->sectorsize);
4395
4396         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4397         if (ret < 0)
4398                 return ret;
4399
4400         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4401         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4402         if (ret < 0)
4403                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4404         else
4405                 ret = 0;
4406         return ret;
4407 }
4408
4409 /*
4410  * Called if we need to clear a data reservation for this inode
4411  * Normally in a error case.
4412  *
4413  * This one will *NOT* use accurate qgroup reserved space API, just for case
4414  * which we can't sleep and is sure it won't affect qgroup reserved space.
4415  * Like clear_bit_hook().
4416  */
4417 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4418                                             u64 len)
4419 {
4420         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4421         struct btrfs_space_info *data_sinfo;
4422
4423         /* Make sure the range is aligned to sectorsize */
4424         len = round_up(start + len, fs_info->sectorsize) -
4425               round_down(start, fs_info->sectorsize);
4426         start = round_down(start, fs_info->sectorsize);
4427
4428         data_sinfo = fs_info->data_sinfo;
4429         spin_lock(&data_sinfo->lock);
4430         if (WARN_ON(data_sinfo->bytes_may_use < len))
4431                 data_sinfo->bytes_may_use = 0;
4432         else
4433                 data_sinfo->bytes_may_use -= len;
4434         trace_btrfs_space_reservation(fs_info, "space_info",
4435                                       data_sinfo->flags, len, 0);
4436         spin_unlock(&data_sinfo->lock);
4437 }
4438
4439 /*
4440  * Called if we need to clear a data reservation for this inode
4441  * Normally in a error case.
4442  *
4443  * This one will handle the per-inode data rsv map for accurate reserved
4444  * space framework.
4445  */
4446 void btrfs_free_reserved_data_space(struct inode *inode,
4447                         struct extent_changeset *reserved, u64 start, u64 len)
4448 {
4449         struct btrfs_root *root = BTRFS_I(inode)->root;
4450
4451         /* Make sure the range is aligned to sectorsize */
4452         len = round_up(start + len, root->fs_info->sectorsize) -
4453               round_down(start, root->fs_info->sectorsize);
4454         start = round_down(start, root->fs_info->sectorsize);
4455
4456         btrfs_free_reserved_data_space_noquota(inode, start, len);
4457         btrfs_qgroup_free_data(inode, reserved, start, len);
4458 }
4459
4460 static void force_metadata_allocation(struct btrfs_fs_info *info)
4461 {
4462         struct list_head *head = &info->space_info;
4463         struct btrfs_space_info *found;
4464
4465         rcu_read_lock();
4466         list_for_each_entry_rcu(found, head, list) {
4467                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4468                         found->force_alloc = CHUNK_ALLOC_FORCE;
4469         }
4470         rcu_read_unlock();
4471 }
4472
4473 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4474 {
4475         return (global->size << 1);
4476 }
4477
4478 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4479                               struct btrfs_space_info *sinfo, int force)
4480 {
4481         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4482         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4483         u64 thresh;
4484
4485         if (force == CHUNK_ALLOC_FORCE)
4486                 return 1;
4487
4488         /*
4489          * We need to take into account the global rsv because for all intents
4490          * and purposes it's used space.  Don't worry about locking the
4491          * global_rsv, it doesn't change except when the transaction commits.
4492          */
4493         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4494                 bytes_used += calc_global_rsv_need_space(global_rsv);
4495
4496         /*
4497          * in limited mode, we want to have some free space up to
4498          * about 1% of the FS size.
4499          */
4500         if (force == CHUNK_ALLOC_LIMITED) {
4501                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4502                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4503
4504                 if (sinfo->total_bytes - bytes_used < thresh)
4505                         return 1;
4506         }
4507
4508         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4509                 return 0;
4510         return 1;
4511 }
4512
4513 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4514 {
4515         u64 num_dev;
4516
4517         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4518                     BTRFS_BLOCK_GROUP_RAID0 |
4519                     BTRFS_BLOCK_GROUP_RAID5 |
4520                     BTRFS_BLOCK_GROUP_RAID6))
4521                 num_dev = fs_info->fs_devices->rw_devices;
4522         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4523                 num_dev = 2;
4524         else
4525                 num_dev = 1;    /* DUP or single */
4526
4527         return num_dev;
4528 }
4529
4530 /*
4531  * If @is_allocation is true, reserve space in the system space info necessary
4532  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4533  * removing a chunk.
4534  */
4535 void check_system_chunk(struct btrfs_trans_handle *trans,
4536                         struct btrfs_fs_info *fs_info, u64 type)
4537 {
4538         struct btrfs_space_info *info;
4539         u64 left;
4540         u64 thresh;
4541         int ret = 0;
4542         u64 num_devs;
4543
4544         /*
4545          * Needed because we can end up allocating a system chunk and for an
4546          * atomic and race free space reservation in the chunk block reserve.
4547          */
4548         ASSERT(mutex_is_locked(&fs_info->chunk_mutex));
4549
4550         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4551         spin_lock(&info->lock);
4552         left = info->total_bytes - btrfs_space_info_used(info, true);
4553         spin_unlock(&info->lock);
4554
4555         num_devs = get_profile_num_devs(fs_info, type);
4556
4557         /* num_devs device items to update and 1 chunk item to add or remove */
4558         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4559                 btrfs_calc_trans_metadata_size(fs_info, 1);
4560
4561         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4562                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4563                            left, thresh, type);
4564                 dump_space_info(fs_info, info, 0, 0);
4565         }
4566
4567         if (left < thresh) {
4568                 u64 flags = btrfs_system_alloc_profile(fs_info);
4569
4570                 /*
4571                  * Ignore failure to create system chunk. We might end up not
4572                  * needing it, as we might not need to COW all nodes/leafs from
4573                  * the paths we visit in the chunk tree (they were already COWed
4574                  * or created in the current transaction for example).
4575                  */
4576                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4577         }
4578
4579         if (!ret) {
4580                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4581                                           &fs_info->chunk_block_rsv,
4582                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4583                 if (!ret)
4584                         trans->chunk_bytes_reserved += thresh;
4585         }
4586 }
4587
4588 /*
4589  * If force is CHUNK_ALLOC_FORCE:
4590  *    - return 1 if it successfully allocates a chunk,
4591  *    - return errors including -ENOSPC otherwise.
4592  * If force is NOT CHUNK_ALLOC_FORCE:
4593  *    - return 0 if it doesn't need to allocate a new chunk,
4594  *    - return 1 if it successfully allocates a chunk,
4595  *    - return errors including -ENOSPC otherwise.
4596  */
4597 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4598                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4599 {
4600         struct btrfs_space_info *space_info;
4601         int wait_for_alloc = 0;
4602         int ret = 0;
4603
4604         /* Don't re-enter if we're already allocating a chunk */
4605         if (trans->allocating_chunk)
4606                 return -ENOSPC;
4607
4608         space_info = __find_space_info(fs_info, flags);
4609         if (!space_info) {
4610                 ret = create_space_info(fs_info, flags, &space_info);
4611                 if (ret)
4612                         return ret;
4613         }
4614
4615 again:
4616         spin_lock(&space_info->lock);
4617         if (force < space_info->force_alloc)
4618                 force = space_info->force_alloc;
4619         if (space_info->full) {
4620                 if (should_alloc_chunk(fs_info, space_info, force))
4621                         ret = -ENOSPC;
4622                 else
4623                         ret = 0;
4624                 spin_unlock(&space_info->lock);
4625                 return ret;
4626         }
4627
4628         if (!should_alloc_chunk(fs_info, space_info, force)) {
4629                 spin_unlock(&space_info->lock);
4630                 return 0;
4631         } else if (space_info->chunk_alloc) {
4632                 wait_for_alloc = 1;
4633         } else {
4634                 space_info->chunk_alloc = 1;
4635         }
4636
4637         spin_unlock(&space_info->lock);
4638
4639         mutex_lock(&fs_info->chunk_mutex);
4640
4641         /*
4642          * The chunk_mutex is held throughout the entirety of a chunk
4643          * allocation, so once we've acquired the chunk_mutex we know that the
4644          * other guy is done and we need to recheck and see if we should
4645          * allocate.
4646          */
4647         if (wait_for_alloc) {
4648                 mutex_unlock(&fs_info->chunk_mutex);
4649                 wait_for_alloc = 0;
4650                 goto again;
4651         }
4652
4653         trans->allocating_chunk = true;
4654
4655         /*
4656          * If we have mixed data/metadata chunks we want to make sure we keep
4657          * allocating mixed chunks instead of individual chunks.
4658          */
4659         if (btrfs_mixed_space_info(space_info))
4660                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4661
4662         /*
4663          * if we're doing a data chunk, go ahead and make sure that
4664          * we keep a reasonable number of metadata chunks allocated in the
4665          * FS as well.
4666          */
4667         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4668                 fs_info->data_chunk_allocations++;
4669                 if (!(fs_info->data_chunk_allocations %
4670                       fs_info->metadata_ratio))
4671                         force_metadata_allocation(fs_info);
4672         }
4673
4674         /*
4675          * Check if we have enough space in SYSTEM chunk because we may need
4676          * to update devices.
4677          */
4678         check_system_chunk(trans, fs_info, flags);
4679
4680         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4681         trans->allocating_chunk = false;
4682
4683         spin_lock(&space_info->lock);
4684         if (ret < 0 && ret != -ENOSPC)
4685                 goto out;
4686         if (ret)
4687                 space_info->full = 1;
4688         else
4689                 ret = 1;
4690
4691         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4692 out:
4693         space_info->chunk_alloc = 0;
4694         spin_unlock(&space_info->lock);
4695         mutex_unlock(&fs_info->chunk_mutex);
4696         /*
4697          * When we allocate a new chunk we reserve space in the chunk block
4698          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4699          * add new nodes/leafs to it if we end up needing to do it when
4700          * inserting the chunk item and updating device items as part of the
4701          * second phase of chunk allocation, performed by
4702          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4703          * large number of new block groups to create in our transaction
4704          * handle's new_bgs list to avoid exhausting the chunk block reserve
4705          * in extreme cases - like having a single transaction create many new
4706          * block groups when starting to write out the free space caches of all
4707          * the block groups that were made dirty during the lifetime of the
4708          * transaction.
4709          */
4710         if (trans->can_flush_pending_bgs &&
4711             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4712                 btrfs_create_pending_block_groups(trans);
4713                 btrfs_trans_release_chunk_metadata(trans);
4714         }
4715         return ret;
4716 }
4717
4718 static int can_overcommit(struct btrfs_fs_info *fs_info,
4719                           struct btrfs_space_info *space_info, u64 bytes,
4720                           enum btrfs_reserve_flush_enum flush,
4721                           bool system_chunk)
4722 {
4723         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4724         u64 profile;
4725         u64 space_size;
4726         u64 avail;
4727         u64 used;
4728
4729         /* Don't overcommit when in mixed mode. */
4730         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4731                 return 0;
4732
4733         if (system_chunk)
4734                 profile = btrfs_system_alloc_profile(fs_info);
4735         else
4736                 profile = btrfs_metadata_alloc_profile(fs_info);
4737
4738         used = btrfs_space_info_used(space_info, false);
4739
4740         /*
4741          * We only want to allow over committing if we have lots of actual space
4742          * free, but if we don't have enough space to handle the global reserve
4743          * space then we could end up having a real enospc problem when trying
4744          * to allocate a chunk or some other such important allocation.
4745          */
4746         spin_lock(&global_rsv->lock);
4747         space_size = calc_global_rsv_need_space(global_rsv);
4748         spin_unlock(&global_rsv->lock);
4749         if (used + space_size >= space_info->total_bytes)
4750                 return 0;
4751
4752         used += space_info->bytes_may_use;
4753
4754         avail = atomic64_read(&fs_info->free_chunk_space);
4755
4756         /*
4757          * If we have dup, raid1 or raid10 then only half of the free
4758          * space is actually useable.  For raid56, the space info used
4759          * doesn't include the parity drive, so we don't have to
4760          * change the math
4761          */
4762         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4763                        BTRFS_BLOCK_GROUP_RAID1 |
4764                        BTRFS_BLOCK_GROUP_RAID10))
4765                 avail >>= 1;
4766
4767         /*
4768          * If we aren't flushing all things, let us overcommit up to
4769          * 1/2th of the space. If we can flush, don't let us overcommit
4770          * too much, let it overcommit up to 1/8 of the space.
4771          */
4772         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4773                 avail >>= 3;
4774         else
4775                 avail >>= 1;
4776
4777         if (used + bytes < space_info->total_bytes + avail)
4778                 return 1;
4779         return 0;
4780 }
4781
4782 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4783                                          unsigned long nr_pages, int nr_items)
4784 {
4785         struct super_block *sb = fs_info->sb;
4786
4787         if (down_read_trylock(&sb->s_umount)) {
4788                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4789                 up_read(&sb->s_umount);
4790         } else {
4791                 /*
4792                  * We needn't worry the filesystem going from r/w to r/o though
4793                  * we don't acquire ->s_umount mutex, because the filesystem
4794                  * should guarantee the delalloc inodes list be empty after
4795                  * the filesystem is readonly(all dirty pages are written to
4796                  * the disk).
4797                  */
4798                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4799                 if (!current->journal_info)
4800                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4801         }
4802 }
4803
4804 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4805                                         u64 to_reclaim)
4806 {
4807         u64 bytes;
4808         u64 nr;
4809
4810         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4811         nr = div64_u64(to_reclaim, bytes);
4812         if (!nr)
4813                 nr = 1;
4814         return nr;
4815 }
4816
4817 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4818
4819 /*
4820  * shrink metadata reservation for delalloc
4821  */
4822 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4823                             u64 orig, bool wait_ordered)
4824 {
4825         struct btrfs_space_info *space_info;
4826         struct btrfs_trans_handle *trans;
4827         u64 delalloc_bytes;
4828         u64 max_reclaim;
4829         u64 items;
4830         long time_left;
4831         unsigned long nr_pages;
4832         int loops;
4833
4834         /* Calc the number of the pages we need flush for space reservation */
4835         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4836         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4837
4838         trans = (struct btrfs_trans_handle *)current->journal_info;
4839         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4840
4841         delalloc_bytes = percpu_counter_sum_positive(
4842                                                 &fs_info->delalloc_bytes);
4843         if (delalloc_bytes == 0) {
4844                 if (trans)
4845                         return;
4846                 if (wait_ordered)
4847                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4848                 return;
4849         }
4850
4851         loops = 0;
4852         while (delalloc_bytes && loops < 3) {
4853                 max_reclaim = min(delalloc_bytes, to_reclaim);
4854                 nr_pages = max_reclaim >> PAGE_SHIFT;
4855                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4856                 /*
4857                  * We need to wait for the async pages to actually start before
4858                  * we do anything.
4859                  */
4860                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4861                 if (!max_reclaim)
4862                         goto skip_async;
4863
4864                 if (max_reclaim <= nr_pages)
4865                         max_reclaim = 0;
4866                 else
4867                         max_reclaim -= nr_pages;
4868
4869                 wait_event(fs_info->async_submit_wait,
4870                            atomic_read(&fs_info->async_delalloc_pages) <=
4871                            (int)max_reclaim);
4872 skip_async:
4873                 spin_lock(&space_info->lock);
4874                 if (list_empty(&space_info->tickets) &&
4875                     list_empty(&space_info->priority_tickets)) {
4876                         spin_unlock(&space_info->lock);
4877                         break;
4878                 }
4879                 spin_unlock(&space_info->lock);
4880
4881                 loops++;
4882                 if (wait_ordered && !trans) {
4883                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4884                 } else {
4885                         time_left = schedule_timeout_killable(1);
4886                         if (time_left)
4887                                 break;
4888                 }
4889                 delalloc_bytes = percpu_counter_sum_positive(
4890                                                 &fs_info->delalloc_bytes);
4891         }
4892 }
4893
4894 struct reserve_ticket {
4895         u64 bytes;
4896         int error;
4897         struct list_head list;
4898         wait_queue_head_t wait;
4899 };
4900
4901 /**
4902  * maybe_commit_transaction - possibly commit the transaction if its ok to
4903  * @root - the root we're allocating for
4904  * @bytes - the number of bytes we want to reserve
4905  * @force - force the commit
4906  *
4907  * This will check to make sure that committing the transaction will actually
4908  * get us somewhere and then commit the transaction if it does.  Otherwise it
4909  * will return -ENOSPC.
4910  */
4911 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4912                                   struct btrfs_space_info *space_info)
4913 {
4914         struct reserve_ticket *ticket = NULL;
4915         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4916         struct btrfs_trans_handle *trans;
4917         u64 bytes;
4918
4919         trans = (struct btrfs_trans_handle *)current->journal_info;
4920         if (trans)
4921                 return -EAGAIN;
4922
4923         spin_lock(&space_info->lock);
4924         if (!list_empty(&space_info->priority_tickets))
4925                 ticket = list_first_entry(&space_info->priority_tickets,
4926                                           struct reserve_ticket, list);
4927         else if (!list_empty(&space_info->tickets))
4928                 ticket = list_first_entry(&space_info->tickets,
4929                                           struct reserve_ticket, list);
4930         bytes = (ticket) ? ticket->bytes : 0;
4931         spin_unlock(&space_info->lock);
4932
4933         if (!bytes)
4934                 return 0;
4935
4936         /* See if there is enough pinned space to make this reservation */
4937         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4938                                    bytes) >= 0)
4939                 goto commit;
4940
4941         /*
4942          * See if there is some space in the delayed insertion reservation for
4943          * this reservation.
4944          */
4945         if (space_info != delayed_rsv->space_info)
4946                 return -ENOSPC;
4947
4948         spin_lock(&delayed_rsv->lock);
4949         if (delayed_rsv->size > bytes)
4950                 bytes = 0;
4951         else
4952                 bytes -= delayed_rsv->size;
4953         spin_unlock(&delayed_rsv->lock);
4954
4955         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4956                                    bytes) < 0) {
4957                 return -ENOSPC;
4958         }
4959
4960 commit:
4961         trans = btrfs_join_transaction(fs_info->extent_root);
4962         if (IS_ERR(trans))
4963                 return -ENOSPC;
4964
4965         return btrfs_commit_transaction(trans);
4966 }
4967
4968 /*
4969  * Try to flush some data based on policy set by @state. This is only advisory
4970  * and may fail for various reasons. The caller is supposed to examine the
4971  * state of @space_info to detect the outcome.
4972  */
4973 static void flush_space(struct btrfs_fs_info *fs_info,
4974                        struct btrfs_space_info *space_info, u64 num_bytes,
4975                        int state)
4976 {
4977         struct btrfs_root *root = fs_info->extent_root;
4978         struct btrfs_trans_handle *trans;
4979         int nr;
4980         int ret = 0;
4981
4982         switch (state) {
4983         case FLUSH_DELAYED_ITEMS_NR:
4984         case FLUSH_DELAYED_ITEMS:
4985                 if (state == FLUSH_DELAYED_ITEMS_NR)
4986                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4987                 else
4988                         nr = -1;
4989
4990                 trans = btrfs_join_transaction(root);
4991                 if (IS_ERR(trans)) {
4992                         ret = PTR_ERR(trans);
4993                         break;
4994                 }
4995                 ret = btrfs_run_delayed_items_nr(trans, nr);
4996                 btrfs_end_transaction(trans);
4997                 break;
4998         case FLUSH_DELALLOC:
4999         case FLUSH_DELALLOC_WAIT:
5000                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
5001                                 state == FLUSH_DELALLOC_WAIT);
5002                 break;
5003         case ALLOC_CHUNK:
5004                 trans = btrfs_join_transaction(root);
5005                 if (IS_ERR(trans)) {
5006                         ret = PTR_ERR(trans);
5007                         break;
5008                 }
5009                 ret = do_chunk_alloc(trans, fs_info,
5010                                      btrfs_metadata_alloc_profile(fs_info),
5011                                      CHUNK_ALLOC_NO_FORCE);
5012                 btrfs_end_transaction(trans);
5013                 if (ret > 0 || ret == -ENOSPC)
5014                         ret = 0;
5015                 break;
5016         case COMMIT_TRANS:
5017                 ret = may_commit_transaction(fs_info, space_info);
5018                 break;
5019         default:
5020                 ret = -ENOSPC;
5021                 break;
5022         }
5023
5024         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5025                                 ret);
5026         return;
5027 }
5028
5029 static inline u64
5030 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5031                                  struct btrfs_space_info *space_info,
5032                                  bool system_chunk)
5033 {
5034         struct reserve_ticket *ticket;
5035         u64 used;
5036         u64 expected;
5037         u64 to_reclaim = 0;
5038
5039         list_for_each_entry(ticket, &space_info->tickets, list)
5040                 to_reclaim += ticket->bytes;
5041         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5042                 to_reclaim += ticket->bytes;
5043         if (to_reclaim)
5044                 return to_reclaim;
5045
5046         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5047         if (can_overcommit(fs_info, space_info, to_reclaim,
5048                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5049                 return 0;
5050
5051         used = btrfs_space_info_used(space_info, true);
5052
5053         if (can_overcommit(fs_info, space_info, SZ_1M,
5054                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5055                 expected = div_factor_fine(space_info->total_bytes, 95);
5056         else
5057                 expected = div_factor_fine(space_info->total_bytes, 90);
5058
5059         if (used > expected)
5060                 to_reclaim = used - expected;
5061         else
5062                 to_reclaim = 0;
5063         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5064                                      space_info->bytes_reserved);
5065         return to_reclaim;
5066 }
5067
5068 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5069                                         struct btrfs_space_info *space_info,
5070                                         u64 used, bool system_chunk)
5071 {
5072         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5073
5074         /* If we're just plain full then async reclaim just slows us down. */
5075         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5076                 return 0;
5077
5078         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5079                                               system_chunk))
5080                 return 0;
5081
5082         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5083                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5084 }
5085
5086 static void wake_all_tickets(struct list_head *head)
5087 {
5088         struct reserve_ticket *ticket;
5089
5090         while (!list_empty(head)) {
5091                 ticket = list_first_entry(head, struct reserve_ticket, list);
5092                 list_del_init(&ticket->list);
5093                 ticket->error = -ENOSPC;
5094                 wake_up(&ticket->wait);
5095         }
5096 }
5097
5098 /*
5099  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5100  * will loop and continuously try to flush as long as we are making progress.
5101  * We count progress as clearing off tickets each time we have to loop.
5102  */
5103 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5104 {
5105         struct btrfs_fs_info *fs_info;
5106         struct btrfs_space_info *space_info;
5107         u64 to_reclaim;
5108         int flush_state;
5109         int commit_cycles = 0;
5110         u64 last_tickets_id;
5111
5112         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5113         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5114
5115         spin_lock(&space_info->lock);
5116         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5117                                                       false);
5118         if (!to_reclaim) {
5119                 space_info->flush = 0;
5120                 spin_unlock(&space_info->lock);
5121                 return;
5122         }
5123         last_tickets_id = space_info->tickets_id;
5124         spin_unlock(&space_info->lock);
5125
5126         flush_state = FLUSH_DELAYED_ITEMS_NR;
5127         do {
5128                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5129                 spin_lock(&space_info->lock);
5130                 if (list_empty(&space_info->tickets)) {
5131                         space_info->flush = 0;
5132                         spin_unlock(&space_info->lock);
5133                         return;
5134                 }
5135                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5136                                                               space_info,
5137                                                               false);
5138                 if (last_tickets_id == space_info->tickets_id) {
5139                         flush_state++;
5140                 } else {
5141                         last_tickets_id = space_info->tickets_id;
5142                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5143                         if (commit_cycles)
5144                                 commit_cycles--;
5145                 }
5146
5147                 if (flush_state > COMMIT_TRANS) {
5148                         commit_cycles++;
5149                         if (commit_cycles > 2) {
5150                                 wake_all_tickets(&space_info->tickets);
5151                                 space_info->flush = 0;
5152                         } else {
5153                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5154                         }
5155                 }
5156                 spin_unlock(&space_info->lock);
5157         } while (flush_state <= COMMIT_TRANS);
5158 }
5159
5160 void btrfs_init_async_reclaim_work(struct work_struct *work)
5161 {
5162         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5163 }
5164
5165 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5166                                             struct btrfs_space_info *space_info,
5167                                             struct reserve_ticket *ticket)
5168 {
5169         u64 to_reclaim;
5170         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5171
5172         spin_lock(&space_info->lock);
5173         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5174                                                       false);
5175         if (!to_reclaim) {
5176                 spin_unlock(&space_info->lock);
5177                 return;
5178         }
5179         spin_unlock(&space_info->lock);
5180
5181         do {
5182                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5183                 flush_state++;
5184                 spin_lock(&space_info->lock);
5185                 if (ticket->bytes == 0) {
5186                         spin_unlock(&space_info->lock);
5187                         return;
5188                 }
5189                 spin_unlock(&space_info->lock);
5190
5191                 /*
5192                  * Priority flushers can't wait on delalloc without
5193                  * deadlocking.
5194                  */
5195                 if (flush_state == FLUSH_DELALLOC ||
5196                     flush_state == FLUSH_DELALLOC_WAIT)
5197                         flush_state = ALLOC_CHUNK;
5198         } while (flush_state < COMMIT_TRANS);
5199 }
5200
5201 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5202                                struct btrfs_space_info *space_info,
5203                                struct reserve_ticket *ticket, u64 orig_bytes)
5204
5205 {
5206         DEFINE_WAIT(wait);
5207         int ret = 0;
5208
5209         spin_lock(&space_info->lock);
5210         while (ticket->bytes > 0 && ticket->error == 0) {
5211                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5212                 if (ret) {
5213                         ret = -EINTR;
5214                         break;
5215                 }
5216                 spin_unlock(&space_info->lock);
5217
5218                 schedule();
5219
5220                 finish_wait(&ticket->wait, &wait);
5221                 spin_lock(&space_info->lock);
5222         }
5223         if (!ret)
5224                 ret = ticket->error;
5225         if (!list_empty(&ticket->list))
5226                 list_del_init(&ticket->list);
5227         if (ticket->bytes && ticket->bytes < orig_bytes) {
5228                 u64 num_bytes = orig_bytes - ticket->bytes;
5229                 space_info->bytes_may_use -= num_bytes;
5230                 trace_btrfs_space_reservation(fs_info, "space_info",
5231                                               space_info->flags, num_bytes, 0);
5232         }
5233         spin_unlock(&space_info->lock);
5234
5235         return ret;
5236 }
5237
5238 /**
5239  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5240  * @root - the root we're allocating for
5241  * @space_info - the space info we want to allocate from
5242  * @orig_bytes - the number of bytes we want
5243  * @flush - whether or not we can flush to make our reservation
5244  *
5245  * This will reserve orig_bytes number of bytes from the space info associated
5246  * with the block_rsv.  If there is not enough space it will make an attempt to
5247  * flush out space to make room.  It will do this by flushing delalloc if
5248  * possible or committing the transaction.  If flush is 0 then no attempts to
5249  * regain reservations will be made and this will fail if there is not enough
5250  * space already.
5251  */
5252 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5253                                     struct btrfs_space_info *space_info,
5254                                     u64 orig_bytes,
5255                                     enum btrfs_reserve_flush_enum flush,
5256                                     bool system_chunk)
5257 {
5258         struct reserve_ticket ticket;
5259         u64 used;
5260         int ret = 0;
5261
5262         ASSERT(orig_bytes);
5263         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5264
5265         spin_lock(&space_info->lock);
5266         ret = -ENOSPC;
5267         used = btrfs_space_info_used(space_info, true);
5268
5269         /*
5270          * If we have enough space then hooray, make our reservation and carry
5271          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5272          * If not things get more complicated.
5273          */
5274         if (used + orig_bytes <= space_info->total_bytes) {
5275                 space_info->bytes_may_use += orig_bytes;
5276                 trace_btrfs_space_reservation(fs_info, "space_info",
5277                                               space_info->flags, orig_bytes, 1);
5278                 ret = 0;
5279         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5280                                   system_chunk)) {
5281                 space_info->bytes_may_use += orig_bytes;
5282                 trace_btrfs_space_reservation(fs_info, "space_info",
5283                                               space_info->flags, orig_bytes, 1);
5284                 ret = 0;
5285         }
5286
5287         /*
5288          * If we couldn't make a reservation then setup our reservation ticket
5289          * and kick the async worker if it's not already running.
5290          *
5291          * If we are a priority flusher then we just need to add our ticket to
5292          * the list and we will do our own flushing further down.
5293          */
5294         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5295                 ticket.bytes = orig_bytes;
5296                 ticket.error = 0;
5297                 init_waitqueue_head(&ticket.wait);
5298                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5299                         list_add_tail(&ticket.list, &space_info->tickets);
5300                         if (!space_info->flush) {
5301                                 space_info->flush = 1;
5302                                 trace_btrfs_trigger_flush(fs_info,
5303                                                           space_info->flags,
5304                                                           orig_bytes, flush,
5305                                                           "enospc");
5306                                 queue_work(system_unbound_wq,
5307                                            &fs_info->async_reclaim_work);
5308                         }
5309                 } else {
5310                         list_add_tail(&ticket.list,
5311                                       &space_info->priority_tickets);
5312                 }
5313         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5314                 used += orig_bytes;
5315                 /*
5316                  * We will do the space reservation dance during log replay,
5317                  * which means we won't have fs_info->fs_root set, so don't do
5318                  * the async reclaim as we will panic.
5319                  */
5320                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5321                     need_do_async_reclaim(fs_info, space_info,
5322                                           used, system_chunk) &&
5323                     !work_busy(&fs_info->async_reclaim_work)) {
5324                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5325                                                   orig_bytes, flush, "preempt");
5326                         queue_work(system_unbound_wq,
5327                                    &fs_info->async_reclaim_work);
5328                 }
5329         }
5330         spin_unlock(&space_info->lock);
5331         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5332                 return ret;
5333
5334         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5335                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5336                                            orig_bytes);
5337
5338         ret = 0;
5339         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5340         spin_lock(&space_info->lock);
5341         if (ticket.bytes) {
5342                 if (ticket.bytes < orig_bytes) {
5343                         u64 num_bytes = orig_bytes - ticket.bytes;
5344                         space_info->bytes_may_use -= num_bytes;
5345                         trace_btrfs_space_reservation(fs_info, "space_info",
5346                                                       space_info->flags,
5347                                                       num_bytes, 0);
5348
5349                 }
5350                 list_del_init(&ticket.list);
5351                 ret = -ENOSPC;
5352         }
5353         spin_unlock(&space_info->lock);
5354         ASSERT(list_empty(&ticket.list));
5355         return ret;
5356 }
5357
5358 /**
5359  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5360  * @root - the root we're allocating for
5361  * @block_rsv - the block_rsv we're allocating for
5362  * @orig_bytes - the number of bytes we want
5363  * @flush - whether or not we can flush to make our reservation
5364  *
5365  * This will reserve orgi_bytes number of bytes from the space info associated
5366  * with the block_rsv.  If there is not enough space it will make an attempt to
5367  * flush out space to make room.  It will do this by flushing delalloc if
5368  * possible or committing the transaction.  If flush is 0 then no attempts to
5369  * regain reservations will be made and this will fail if there is not enough
5370  * space already.
5371  */
5372 static int reserve_metadata_bytes(struct btrfs_root *root,
5373                                   struct btrfs_block_rsv *block_rsv,
5374                                   u64 orig_bytes,
5375                                   enum btrfs_reserve_flush_enum flush)
5376 {
5377         struct btrfs_fs_info *fs_info = root->fs_info;
5378         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5379         int ret;
5380         bool system_chunk = (root == fs_info->chunk_root);
5381
5382         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5383                                        orig_bytes, flush, system_chunk);
5384         if (ret == -ENOSPC &&
5385             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5386                 if (block_rsv != global_rsv &&
5387                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5388                         ret = 0;
5389         }
5390         if (ret == -ENOSPC) {
5391                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5392                                               block_rsv->space_info->flags,
5393                                               orig_bytes, 1);
5394
5395                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5396                         dump_space_info(fs_info, block_rsv->space_info,
5397                                         orig_bytes, 0);
5398         }
5399         return ret;
5400 }
5401
5402 static struct btrfs_block_rsv *get_block_rsv(
5403                                         const struct btrfs_trans_handle *trans,
5404                                         const struct btrfs_root *root)
5405 {
5406         struct btrfs_fs_info *fs_info = root->fs_info;
5407         struct btrfs_block_rsv *block_rsv = NULL;
5408
5409         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5410             (root == fs_info->csum_root && trans->adding_csums) ||
5411             (root == fs_info->uuid_root))
5412                 block_rsv = trans->block_rsv;
5413
5414         if (!block_rsv)
5415                 block_rsv = root->block_rsv;
5416
5417         if (!block_rsv)
5418                 block_rsv = &fs_info->empty_block_rsv;
5419
5420         return block_rsv;
5421 }
5422
5423 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5424                                u64 num_bytes)
5425 {
5426         int ret = -ENOSPC;
5427         spin_lock(&block_rsv->lock);
5428         if (block_rsv->reserved >= num_bytes) {
5429                 block_rsv->reserved -= num_bytes;
5430                 if (block_rsv->reserved < block_rsv->size)
5431                         block_rsv->full = 0;
5432                 ret = 0;
5433         }
5434         spin_unlock(&block_rsv->lock);
5435         return ret;
5436 }
5437
5438 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5439                                 u64 num_bytes, int update_size)
5440 {
5441         spin_lock(&block_rsv->lock);
5442         block_rsv->reserved += num_bytes;
5443         if (update_size)
5444                 block_rsv->size += num_bytes;
5445         else if (block_rsv->reserved >= block_rsv->size)
5446                 block_rsv->full = 1;
5447         spin_unlock(&block_rsv->lock);
5448 }
5449
5450 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5451                              struct btrfs_block_rsv *dest, u64 num_bytes,
5452                              int min_factor)
5453 {
5454         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5455         u64 min_bytes;
5456
5457         if (global_rsv->space_info != dest->space_info)
5458                 return -ENOSPC;
5459
5460         spin_lock(&global_rsv->lock);
5461         min_bytes = div_factor(global_rsv->size, min_factor);
5462         if (global_rsv->reserved < min_bytes + num_bytes) {
5463                 spin_unlock(&global_rsv->lock);
5464                 return -ENOSPC;
5465         }
5466         global_rsv->reserved -= num_bytes;
5467         if (global_rsv->reserved < global_rsv->size)
5468                 global_rsv->full = 0;
5469         spin_unlock(&global_rsv->lock);
5470
5471         block_rsv_add_bytes(dest, num_bytes, 1);
5472         return 0;
5473 }
5474
5475 /*
5476  * This is for space we already have accounted in space_info->bytes_may_use, so
5477  * basically when we're returning space from block_rsv's.
5478  */
5479 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5480                                      struct btrfs_space_info *space_info,
5481                                      u64 num_bytes)
5482 {
5483         struct reserve_ticket *ticket;
5484         struct list_head *head;
5485         u64 used;
5486         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5487         bool check_overcommit = false;
5488
5489         spin_lock(&space_info->lock);
5490         head = &space_info->priority_tickets;
5491
5492         /*
5493          * If we are over our limit then we need to check and see if we can
5494          * overcommit, and if we can't then we just need to free up our space
5495          * and not satisfy any requests.
5496          */
5497         used = btrfs_space_info_used(space_info, true);
5498         if (used - num_bytes >= space_info->total_bytes)
5499                 check_overcommit = true;
5500 again:
5501         while (!list_empty(head) && num_bytes) {
5502                 ticket = list_first_entry(head, struct reserve_ticket,
5503                                           list);
5504                 /*
5505                  * We use 0 bytes because this space is already reserved, so
5506                  * adding the ticket space would be a double count.
5507                  */
5508                 if (check_overcommit &&
5509                     !can_overcommit(fs_info, space_info, 0, flush, false))
5510                         break;
5511                 if (num_bytes >= ticket->bytes) {
5512                         list_del_init(&ticket->list);
5513                         num_bytes -= ticket->bytes;
5514                         ticket->bytes = 0;
5515                         space_info->tickets_id++;
5516                         wake_up(&ticket->wait);
5517                 } else {
5518                         ticket->bytes -= num_bytes;
5519                         num_bytes = 0;
5520                 }
5521         }
5522
5523         if (num_bytes && head == &space_info->priority_tickets) {
5524                 head = &space_info->tickets;
5525                 flush = BTRFS_RESERVE_FLUSH_ALL;
5526                 goto again;
5527         }
5528         space_info->bytes_may_use -= num_bytes;
5529         trace_btrfs_space_reservation(fs_info, "space_info",
5530                                       space_info->flags, num_bytes, 0);
5531         spin_unlock(&space_info->lock);
5532 }
5533
5534 /*
5535  * This is for newly allocated space that isn't accounted in
5536  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5537  * we use this helper.
5538  */
5539 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5540                                      struct btrfs_space_info *space_info,
5541                                      u64 num_bytes)
5542 {
5543         struct reserve_ticket *ticket;
5544         struct list_head *head = &space_info->priority_tickets;
5545
5546 again:
5547         while (!list_empty(head) && num_bytes) {
5548                 ticket = list_first_entry(head, struct reserve_ticket,
5549                                           list);
5550                 if (num_bytes >= ticket->bytes) {
5551                         trace_btrfs_space_reservation(fs_info, "space_info",
5552                                                       space_info->flags,
5553                                                       ticket->bytes, 1);
5554                         list_del_init(&ticket->list);
5555                         num_bytes -= ticket->bytes;
5556                         space_info->bytes_may_use += ticket->bytes;
5557                         ticket->bytes = 0;
5558                         space_info->tickets_id++;
5559                         wake_up(&ticket->wait);
5560                 } else {
5561                         trace_btrfs_space_reservation(fs_info, "space_info",
5562                                                       space_info->flags,
5563                                                       num_bytes, 1);
5564                         space_info->bytes_may_use += num_bytes;
5565                         ticket->bytes -= num_bytes;
5566                         num_bytes = 0;
5567                 }
5568         }
5569
5570         if (num_bytes && head == &space_info->priority_tickets) {
5571                 head = &space_info->tickets;
5572                 goto again;
5573         }
5574 }
5575
5576 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5577                                     struct btrfs_block_rsv *block_rsv,
5578                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5579 {
5580         struct btrfs_space_info *space_info = block_rsv->space_info;
5581         u64 ret;
5582
5583         spin_lock(&block_rsv->lock);
5584         if (num_bytes == (u64)-1)
5585                 num_bytes = block_rsv->size;
5586         block_rsv->size -= num_bytes;
5587         if (block_rsv->reserved >= block_rsv->size) {
5588                 num_bytes = block_rsv->reserved - block_rsv->size;
5589                 block_rsv->reserved = block_rsv->size;
5590                 block_rsv->full = 1;
5591         } else {
5592                 num_bytes = 0;
5593         }
5594         spin_unlock(&block_rsv->lock);
5595
5596         ret = num_bytes;
5597         if (num_bytes > 0) {
5598                 if (dest) {
5599                         spin_lock(&dest->lock);
5600                         if (!dest->full) {
5601                                 u64 bytes_to_add;
5602
5603                                 bytes_to_add = dest->size - dest->reserved;
5604                                 bytes_to_add = min(num_bytes, bytes_to_add);
5605                                 dest->reserved += bytes_to_add;
5606                                 if (dest->reserved >= dest->size)
5607                                         dest->full = 1;
5608                                 num_bytes -= bytes_to_add;
5609                         }
5610                         spin_unlock(&dest->lock);
5611                 }
5612                 if (num_bytes)
5613                         space_info_add_old_bytes(fs_info, space_info,
5614                                                  num_bytes);
5615         }
5616         return ret;
5617 }
5618
5619 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5620                             struct btrfs_block_rsv *dst, u64 num_bytes,
5621                             int update_size)
5622 {
5623         int ret;
5624
5625         ret = block_rsv_use_bytes(src, num_bytes);
5626         if (ret)
5627                 return ret;
5628
5629         block_rsv_add_bytes(dst, num_bytes, update_size);
5630         return 0;
5631 }
5632
5633 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5634 {
5635         memset(rsv, 0, sizeof(*rsv));
5636         spin_lock_init(&rsv->lock);
5637         rsv->type = type;
5638 }
5639
5640 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5641                                    struct btrfs_block_rsv *rsv,
5642                                    unsigned short type)
5643 {
5644         btrfs_init_block_rsv(rsv, type);
5645         rsv->space_info = __find_space_info(fs_info,
5646                                             BTRFS_BLOCK_GROUP_METADATA);
5647 }
5648
5649 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5650                                               unsigned short type)
5651 {
5652         struct btrfs_block_rsv *block_rsv;
5653
5654         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5655         if (!block_rsv)
5656                 return NULL;
5657
5658         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5659         return block_rsv;
5660 }
5661
5662 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5663                           struct btrfs_block_rsv *rsv)
5664 {
5665         if (!rsv)
5666                 return;
5667         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5668         kfree(rsv);
5669 }
5670
5671 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5672 {
5673         kfree(rsv);
5674 }
5675
5676 int btrfs_block_rsv_add(struct btrfs_root *root,
5677                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5678                         enum btrfs_reserve_flush_enum flush)
5679 {
5680         int ret;
5681
5682         if (num_bytes == 0)
5683                 return 0;
5684
5685         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5686         if (!ret) {
5687                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5688                 return 0;
5689         }
5690
5691         return ret;
5692 }
5693
5694 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5695 {
5696         u64 num_bytes = 0;
5697         int ret = -ENOSPC;
5698
5699         if (!block_rsv)
5700                 return 0;
5701
5702         spin_lock(&block_rsv->lock);
5703         num_bytes = div_factor(block_rsv->size, min_factor);
5704         if (block_rsv->reserved >= num_bytes)
5705                 ret = 0;
5706         spin_unlock(&block_rsv->lock);
5707
5708         return ret;
5709 }
5710
5711 int btrfs_block_rsv_refill(struct btrfs_root *root,
5712                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5713                            enum btrfs_reserve_flush_enum flush)
5714 {
5715         u64 num_bytes = 0;
5716         int ret = -ENOSPC;
5717
5718         if (!block_rsv)
5719                 return 0;
5720
5721         spin_lock(&block_rsv->lock);
5722         num_bytes = min_reserved;
5723         if (block_rsv->reserved >= num_bytes)
5724                 ret = 0;
5725         else
5726                 num_bytes -= block_rsv->reserved;
5727         spin_unlock(&block_rsv->lock);
5728
5729         if (!ret)
5730                 return 0;
5731
5732         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5733         if (!ret) {
5734                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5735                 return 0;
5736         }
5737
5738         return ret;
5739 }
5740
5741 /**
5742  * btrfs_inode_rsv_refill - refill the inode block rsv.
5743  * @inode - the inode we are refilling.
5744  * @flush - the flusing restriction.
5745  *
5746  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5747  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5748  * or return if we already have enough space.  This will also handle the resreve
5749  * tracepoint for the reserved amount.
5750  */
5751 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5752                                   enum btrfs_reserve_flush_enum flush)
5753 {
5754         struct btrfs_root *root = inode->root;
5755         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5756         u64 num_bytes = 0;
5757         int ret = -ENOSPC;
5758
5759         spin_lock(&block_rsv->lock);
5760         if (block_rsv->reserved < block_rsv->size)
5761                 num_bytes = block_rsv->size - block_rsv->reserved;
5762         spin_unlock(&block_rsv->lock);
5763
5764         if (num_bytes == 0)
5765                 return 0;
5766
5767         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5768         if (!ret) {
5769                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5770                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5771                                               btrfs_ino(inode), num_bytes, 1);
5772         }
5773         return ret;
5774 }
5775
5776 /**
5777  * btrfs_inode_rsv_release - release any excessive reservation.
5778  * @inode - the inode we need to release from.
5779  *
5780  * This is the same as btrfs_block_rsv_release, except that it handles the
5781  * tracepoint for the reservation.
5782  */
5783 static void btrfs_inode_rsv_release(struct btrfs_inode *inode)
5784 {
5785         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5786         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5787         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5788         u64 released = 0;
5789
5790         /*
5791          * Since we statically set the block_rsv->size we just want to say we
5792          * are releasing 0 bytes, and then we'll just get the reservation over
5793          * the size free'd.
5794          */
5795         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0);
5796         if (released > 0)
5797                 trace_btrfs_space_reservation(fs_info, "delalloc",
5798                                               btrfs_ino(inode), released, 0);
5799 }
5800
5801 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5802                              struct btrfs_block_rsv *block_rsv,
5803                              u64 num_bytes)
5804 {
5805         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5806
5807         if (global_rsv == block_rsv ||
5808             block_rsv->space_info != global_rsv->space_info)
5809                 global_rsv = NULL;
5810         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5811 }
5812
5813 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5814 {
5815         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5816         struct btrfs_space_info *sinfo = block_rsv->space_info;
5817         u64 num_bytes;
5818
5819         /*
5820          * The global block rsv is based on the size of the extent tree, the
5821          * checksum tree and the root tree.  If the fs is empty we want to set
5822          * it to a minimal amount for safety.
5823          */
5824         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5825                 btrfs_root_used(&fs_info->csum_root->root_item) +
5826                 btrfs_root_used(&fs_info->tree_root->root_item);
5827         num_bytes = max_t(u64, num_bytes, SZ_16M);
5828
5829         spin_lock(&sinfo->lock);
5830         spin_lock(&block_rsv->lock);
5831
5832         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5833
5834         if (block_rsv->reserved < block_rsv->size) {
5835                 num_bytes = btrfs_space_info_used(sinfo, true);
5836                 if (sinfo->total_bytes > num_bytes) {
5837                         num_bytes = sinfo->total_bytes - num_bytes;
5838                         num_bytes = min(num_bytes,
5839                                         block_rsv->size - block_rsv->reserved);
5840                         block_rsv->reserved += num_bytes;
5841                         sinfo->bytes_may_use += num_bytes;
5842                         trace_btrfs_space_reservation(fs_info, "space_info",
5843                                                       sinfo->flags, num_bytes,
5844                                                       1);
5845                 }
5846         } else if (block_rsv->reserved > block_rsv->size) {
5847                 num_bytes = block_rsv->reserved - block_rsv->size;
5848                 sinfo->bytes_may_use -= num_bytes;
5849                 trace_btrfs_space_reservation(fs_info, "space_info",
5850                                       sinfo->flags, num_bytes, 0);
5851                 block_rsv->reserved = block_rsv->size;
5852         }
5853
5854         if (block_rsv->reserved == block_rsv->size)
5855                 block_rsv->full = 1;
5856         else
5857                 block_rsv->full = 0;
5858
5859         spin_unlock(&block_rsv->lock);
5860         spin_unlock(&sinfo->lock);
5861 }
5862
5863 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5864 {
5865         struct btrfs_space_info *space_info;
5866
5867         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5868         fs_info->chunk_block_rsv.space_info = space_info;
5869
5870         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5871         fs_info->global_block_rsv.space_info = space_info;
5872         fs_info->trans_block_rsv.space_info = space_info;
5873         fs_info->empty_block_rsv.space_info = space_info;
5874         fs_info->delayed_block_rsv.space_info = space_info;
5875
5876         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5877         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5878         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5879         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5880         if (fs_info->quota_root)
5881                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5882         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5883
5884         update_global_block_rsv(fs_info);
5885 }
5886
5887 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5888 {
5889         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5890                                 (u64)-1);
5891         WARN_ON(fs_info->trans_block_rsv.size > 0);
5892         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5893         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5894         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5895         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5896         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5897 }
5898
5899
5900 /*
5901  * To be called after all the new block groups attached to the transaction
5902  * handle have been created (btrfs_create_pending_block_groups()).
5903  */
5904 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5905 {
5906         struct btrfs_fs_info *fs_info = trans->fs_info;
5907
5908         if (!trans->chunk_bytes_reserved)
5909                 return;
5910
5911         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5912
5913         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5914                                 trans->chunk_bytes_reserved);
5915         trans->chunk_bytes_reserved = 0;
5916 }
5917
5918 /* Can only return 0 or -ENOSPC */
5919 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5920                                   struct btrfs_inode *inode)
5921 {
5922         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5923         struct btrfs_root *root = inode->root;
5924         /*
5925          * We always use trans->block_rsv here as we will have reserved space
5926          * for our orphan when starting the transaction, using get_block_rsv()
5927          * here will sometimes make us choose the wrong block rsv as we could be
5928          * doing a reloc inode for a non refcounted root.
5929          */
5930         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5931         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5932
5933         /*
5934          * We need to hold space in order to delete our orphan item once we've
5935          * added it, so this takes the reservation so we can release it later
5936          * when we are truly done with the orphan item.
5937          */
5938         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5939
5940         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5941                         num_bytes, 1);
5942         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5943 }
5944
5945 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5946 {
5947         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5948         struct btrfs_root *root = inode->root;
5949         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5950
5951         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5952                         num_bytes, 0);
5953         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5954 }
5955
5956 /*
5957  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5958  * root: the root of the parent directory
5959  * rsv: block reservation
5960  * items: the number of items that we need do reservation
5961  * qgroup_reserved: used to return the reserved size in qgroup
5962  *
5963  * This function is used to reserve the space for snapshot/subvolume
5964  * creation and deletion. Those operations are different with the
5965  * common file/directory operations, they change two fs/file trees
5966  * and root tree, the number of items that the qgroup reserves is
5967  * different with the free space reservation. So we can not use
5968  * the space reservation mechanism in start_transaction().
5969  */
5970 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5971                                      struct btrfs_block_rsv *rsv,
5972                                      int items,
5973                                      u64 *qgroup_reserved,
5974                                      bool use_global_rsv)
5975 {
5976         u64 num_bytes;
5977         int ret;
5978         struct btrfs_fs_info *fs_info = root->fs_info;
5979         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5980
5981         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5982                 /* One for parent inode, two for dir entries */
5983                 num_bytes = 3 * fs_info->nodesize;
5984                 ret = btrfs_qgroup_reserve_meta(root, num_bytes, true);
5985                 if (ret)
5986                         return ret;
5987         } else {
5988                 num_bytes = 0;
5989         }
5990
5991         *qgroup_reserved = num_bytes;
5992
5993         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5994         rsv->space_info = __find_space_info(fs_info,
5995                                             BTRFS_BLOCK_GROUP_METADATA);
5996         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5997                                   BTRFS_RESERVE_FLUSH_ALL);
5998
5999         if (ret == -ENOSPC && use_global_rsv)
6000                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6001
6002         if (ret && *qgroup_reserved)
6003                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
6004
6005         return ret;
6006 }
6007
6008 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6009                                       struct btrfs_block_rsv *rsv)
6010 {
6011         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6012 }
6013
6014 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6015                                                  struct btrfs_inode *inode)
6016 {
6017         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6018         u64 reserve_size = 0;
6019         u64 csum_leaves;
6020         unsigned outstanding_extents;
6021
6022         lockdep_assert_held(&inode->lock);
6023         outstanding_extents = inode->outstanding_extents;
6024         if (outstanding_extents)
6025                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6026                                                 outstanding_extents + 1);
6027         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6028                                                  inode->csum_bytes);
6029         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6030                                                        csum_leaves);
6031
6032         spin_lock(&block_rsv->lock);
6033         block_rsv->size = reserve_size;
6034         spin_unlock(&block_rsv->lock);
6035 }
6036
6037 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6038 {
6039         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6040         struct btrfs_root *root = inode->root;
6041         unsigned nr_extents;
6042         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6043         int ret = 0;
6044         bool delalloc_lock = true;
6045
6046         /* If we are a free space inode we need to not flush since we will be in
6047          * the middle of a transaction commit.  We also don't need the delalloc
6048          * mutex since we won't race with anybody.  We need this mostly to make
6049          * lockdep shut its filthy mouth.
6050          *
6051          * If we have a transaction open (can happen if we call truncate_block
6052          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6053          */
6054         if (btrfs_is_free_space_inode(inode)) {
6055                 flush = BTRFS_RESERVE_NO_FLUSH;
6056                 delalloc_lock = false;
6057         } else {
6058                 if (current->journal_info)
6059                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6060
6061                 if (btrfs_transaction_in_commit(fs_info))
6062                         schedule_timeout(1);
6063         }
6064
6065         if (delalloc_lock)
6066                 mutex_lock(&inode->delalloc_mutex);
6067
6068         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6069
6070         /* Add our new extents and calculate the new rsv size. */
6071         spin_lock(&inode->lock);
6072         nr_extents = count_max_extents(num_bytes);
6073         btrfs_mod_outstanding_extents(inode, nr_extents);
6074         inode->csum_bytes += num_bytes;
6075         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6076         spin_unlock(&inode->lock);
6077
6078         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6079                 ret = btrfs_qgroup_reserve_meta(root,
6080                                 nr_extents * fs_info->nodesize, true);
6081                 if (ret)
6082                         goto out_fail;
6083         }
6084
6085         ret = btrfs_inode_rsv_refill(inode, flush);
6086         if (unlikely(ret)) {
6087                 btrfs_qgroup_free_meta(root,
6088                                        nr_extents * fs_info->nodesize);
6089                 goto out_fail;
6090         }
6091
6092         if (delalloc_lock)
6093                 mutex_unlock(&inode->delalloc_mutex);
6094         return 0;
6095
6096 out_fail:
6097         spin_lock(&inode->lock);
6098         nr_extents = count_max_extents(num_bytes);
6099         btrfs_mod_outstanding_extents(inode, -nr_extents);
6100         inode->csum_bytes -= num_bytes;
6101         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6102         spin_unlock(&inode->lock);
6103
6104         btrfs_inode_rsv_release(inode);
6105         if (delalloc_lock)
6106                 mutex_unlock(&inode->delalloc_mutex);
6107         return ret;
6108 }
6109
6110 /**
6111  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6112  * @inode: the inode to release the reservation for.
6113  * @num_bytes: the number of bytes we are releasing.
6114  *
6115  * This will release the metadata reservation for an inode.  This can be called
6116  * once we complete IO for a given set of bytes to release their metadata
6117  * reservations, or on error for the same reason.
6118  */
6119 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes)
6120 {
6121         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6122
6123         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6124         spin_lock(&inode->lock);
6125         inode->csum_bytes -= num_bytes;
6126         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6127         spin_unlock(&inode->lock);
6128
6129         if (btrfs_is_testing(fs_info))
6130                 return;
6131
6132         btrfs_inode_rsv_release(inode);
6133 }
6134
6135 /**
6136  * btrfs_delalloc_release_extents - release our outstanding_extents
6137  * @inode: the inode to balance the reservation for.
6138  * @num_bytes: the number of bytes we originally reserved with
6139  *
6140  * When we reserve space we increase outstanding_extents for the extents we may
6141  * add.  Once we've set the range as delalloc or created our ordered extents we
6142  * have outstanding_extents to track the real usage, so we use this to free our
6143  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6144  * with btrfs_delalloc_reserve_metadata.
6145  */
6146 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes)
6147 {
6148         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6149         unsigned num_extents;
6150
6151         spin_lock(&inode->lock);
6152         num_extents = count_max_extents(num_bytes);
6153         btrfs_mod_outstanding_extents(inode, -num_extents);
6154         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6155         spin_unlock(&inode->lock);
6156
6157         if (btrfs_is_testing(fs_info))
6158                 return;
6159
6160         btrfs_inode_rsv_release(inode);
6161 }
6162
6163 /**
6164  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6165  * delalloc
6166  * @inode: inode we're writing to
6167  * @start: start range we are writing to
6168  * @len: how long the range we are writing to
6169  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6170  *            current reservation.
6171  *
6172  * This will do the following things
6173  *
6174  * o reserve space in data space info for num bytes
6175  *   and reserve precious corresponding qgroup space
6176  *   (Done in check_data_free_space)
6177  *
6178  * o reserve space for metadata space, based on the number of outstanding
6179  *   extents and how much csums will be needed
6180  *   also reserve metadata space in a per root over-reserve method.
6181  * o add to the inodes->delalloc_bytes
6182  * o add it to the fs_info's delalloc inodes list.
6183  *   (Above 3 all done in delalloc_reserve_metadata)
6184  *
6185  * Return 0 for success
6186  * Return <0 for error(-ENOSPC or -EQUOT)
6187  */
6188 int btrfs_delalloc_reserve_space(struct inode *inode,
6189                         struct extent_changeset **reserved, u64 start, u64 len)
6190 {
6191         int ret;
6192
6193         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6194         if (ret < 0)
6195                 return ret;
6196         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6197         if (ret < 0)
6198                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6199         return ret;
6200 }
6201
6202 /**
6203  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6204  * @inode: inode we're releasing space for
6205  * @start: start position of the space already reserved
6206  * @len: the len of the space already reserved
6207  * @release_bytes: the len of the space we consumed or didn't use
6208  *
6209  * This function will release the metadata space that was not used and will
6210  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6211  * list if there are no delalloc bytes left.
6212  * Also it will handle the qgroup reserved space.
6213  */
6214 void btrfs_delalloc_release_space(struct inode *inode,
6215                                   struct extent_changeset *reserved,
6216                                   u64 start, u64 len)
6217 {
6218         btrfs_delalloc_release_metadata(BTRFS_I(inode), len);
6219         btrfs_free_reserved_data_space(inode, reserved, start, len);
6220 }
6221
6222 static int update_block_group(struct btrfs_trans_handle *trans,
6223                               struct btrfs_fs_info *info, u64 bytenr,
6224                               u64 num_bytes, int alloc)
6225 {
6226         struct btrfs_block_group_cache *cache = NULL;
6227         u64 total = num_bytes;
6228         u64 old_val;
6229         u64 byte_in_group;
6230         int factor;
6231
6232         /* block accounting for super block */
6233         spin_lock(&info->delalloc_root_lock);
6234         old_val = btrfs_super_bytes_used(info->super_copy);
6235         if (alloc)
6236                 old_val += num_bytes;
6237         else
6238                 old_val -= num_bytes;
6239         btrfs_set_super_bytes_used(info->super_copy, old_val);
6240         spin_unlock(&info->delalloc_root_lock);
6241
6242         while (total) {
6243                 cache = btrfs_lookup_block_group(info, bytenr);
6244                 if (!cache)
6245                         return -ENOENT;
6246                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6247                                     BTRFS_BLOCK_GROUP_RAID1 |
6248                                     BTRFS_BLOCK_GROUP_RAID10))
6249                         factor = 2;
6250                 else
6251                         factor = 1;
6252                 /*
6253                  * If this block group has free space cache written out, we
6254                  * need to make sure to load it if we are removing space.  This
6255                  * is because we need the unpinning stage to actually add the
6256                  * space back to the block group, otherwise we will leak space.
6257                  */
6258                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6259                         cache_block_group(cache, 1);
6260
6261                 byte_in_group = bytenr - cache->key.objectid;
6262                 WARN_ON(byte_in_group > cache->key.offset);
6263
6264                 spin_lock(&cache->space_info->lock);
6265                 spin_lock(&cache->lock);
6266
6267                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6268                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6269                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6270
6271                 old_val = btrfs_block_group_used(&cache->item);
6272                 num_bytes = min(total, cache->key.offset - byte_in_group);
6273                 if (alloc) {
6274                         old_val += num_bytes;
6275                         btrfs_set_block_group_used(&cache->item, old_val);
6276                         cache->reserved -= num_bytes;
6277                         cache->space_info->bytes_reserved -= num_bytes;
6278                         cache->space_info->bytes_used += num_bytes;
6279                         cache->space_info->disk_used += num_bytes * factor;
6280                         spin_unlock(&cache->lock);
6281                         spin_unlock(&cache->space_info->lock);
6282                 } else {
6283                         old_val -= num_bytes;
6284                         btrfs_set_block_group_used(&cache->item, old_val);
6285                         cache->pinned += num_bytes;
6286                         cache->space_info->bytes_pinned += num_bytes;
6287                         cache->space_info->bytes_used -= num_bytes;
6288                         cache->space_info->disk_used -= num_bytes * factor;
6289                         spin_unlock(&cache->lock);
6290                         spin_unlock(&cache->space_info->lock);
6291
6292                         trace_btrfs_space_reservation(info, "pinned",
6293                                                       cache->space_info->flags,
6294                                                       num_bytes, 1);
6295                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6296                                            num_bytes);
6297                         set_extent_dirty(info->pinned_extents,
6298                                          bytenr, bytenr + num_bytes - 1,
6299                                          GFP_NOFS | __GFP_NOFAIL);
6300                 }
6301
6302                 spin_lock(&trans->transaction->dirty_bgs_lock);
6303                 if (list_empty(&cache->dirty_list)) {
6304                         list_add_tail(&cache->dirty_list,
6305                                       &trans->transaction->dirty_bgs);
6306                                 trans->transaction->num_dirty_bgs++;
6307                         btrfs_get_block_group(cache);
6308                 }
6309                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6310
6311                 /*
6312                  * No longer have used bytes in this block group, queue it for
6313                  * deletion. We do this after adding the block group to the
6314                  * dirty list to avoid races between cleaner kthread and space
6315                  * cache writeout.
6316                  */
6317                 if (!alloc && old_val == 0) {
6318                         spin_lock(&info->unused_bgs_lock);
6319                         if (list_empty(&cache->bg_list)) {
6320                                 btrfs_get_block_group(cache);
6321                                 list_add_tail(&cache->bg_list,
6322                                               &info->unused_bgs);
6323                         }
6324                         spin_unlock(&info->unused_bgs_lock);
6325                 }
6326
6327                 btrfs_put_block_group(cache);
6328                 total -= num_bytes;
6329                 bytenr += num_bytes;
6330         }
6331         return 0;
6332 }
6333
6334 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6335 {
6336         struct btrfs_block_group_cache *cache;
6337         u64 bytenr;
6338
6339         spin_lock(&fs_info->block_group_cache_lock);
6340         bytenr = fs_info->first_logical_byte;
6341         spin_unlock(&fs_info->block_group_cache_lock);
6342
6343         if (bytenr < (u64)-1)
6344                 return bytenr;
6345
6346         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6347         if (!cache)
6348                 return 0;
6349
6350         bytenr = cache->key.objectid;
6351         btrfs_put_block_group(cache);
6352
6353         return bytenr;
6354 }
6355
6356 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6357                            struct btrfs_block_group_cache *cache,
6358                            u64 bytenr, u64 num_bytes, int reserved)
6359 {
6360         spin_lock(&cache->space_info->lock);
6361         spin_lock(&cache->lock);
6362         cache->pinned += num_bytes;
6363         cache->space_info->bytes_pinned += num_bytes;
6364         if (reserved) {
6365                 cache->reserved -= num_bytes;
6366                 cache->space_info->bytes_reserved -= num_bytes;
6367         }
6368         spin_unlock(&cache->lock);
6369         spin_unlock(&cache->space_info->lock);
6370
6371         trace_btrfs_space_reservation(fs_info, "pinned",
6372                                       cache->space_info->flags, num_bytes, 1);
6373         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6374         set_extent_dirty(fs_info->pinned_extents, bytenr,
6375                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6376         return 0;
6377 }
6378
6379 /*
6380  * this function must be called within transaction
6381  */
6382 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6383                      u64 bytenr, u64 num_bytes, int reserved)
6384 {
6385         struct btrfs_block_group_cache *cache;
6386
6387         cache = btrfs_lookup_block_group(fs_info, bytenr);
6388         BUG_ON(!cache); /* Logic error */
6389
6390         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6391
6392         btrfs_put_block_group(cache);
6393         return 0;
6394 }
6395
6396 /*
6397  * this function must be called within transaction
6398  */
6399 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6400                                     u64 bytenr, u64 num_bytes)
6401 {
6402         struct btrfs_block_group_cache *cache;
6403         int ret;
6404
6405         cache = btrfs_lookup_block_group(fs_info, bytenr);
6406         if (!cache)
6407                 return -EINVAL;
6408
6409         /*
6410          * pull in the free space cache (if any) so that our pin
6411          * removes the free space from the cache.  We have load_only set
6412          * to one because the slow code to read in the free extents does check
6413          * the pinned extents.
6414          */
6415         cache_block_group(cache, 1);
6416
6417         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6418
6419         /* remove us from the free space cache (if we're there at all) */
6420         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6421         btrfs_put_block_group(cache);
6422         return ret;
6423 }
6424
6425 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6426                                    u64 start, u64 num_bytes)
6427 {
6428         int ret;
6429         struct btrfs_block_group_cache *block_group;
6430         struct btrfs_caching_control *caching_ctl;
6431
6432         block_group = btrfs_lookup_block_group(fs_info, start);
6433         if (!block_group)
6434                 return -EINVAL;
6435
6436         cache_block_group(block_group, 0);
6437         caching_ctl = get_caching_control(block_group);
6438
6439         if (!caching_ctl) {
6440                 /* Logic error */
6441                 BUG_ON(!block_group_cache_done(block_group));
6442                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6443         } else {
6444                 mutex_lock(&caching_ctl->mutex);
6445
6446                 if (start >= caching_ctl->progress) {
6447                         ret = add_excluded_extent(fs_info, start, num_bytes);
6448                 } else if (start + num_bytes <= caching_ctl->progress) {
6449                         ret = btrfs_remove_free_space(block_group,
6450                                                       start, num_bytes);
6451                 } else {
6452                         num_bytes = caching_ctl->progress - start;
6453                         ret = btrfs_remove_free_space(block_group,
6454                                                       start, num_bytes);
6455                         if (ret)
6456                                 goto out_lock;
6457
6458                         num_bytes = (start + num_bytes) -
6459                                 caching_ctl->progress;
6460                         start = caching_ctl->progress;
6461                         ret = add_excluded_extent(fs_info, start, num_bytes);
6462                 }
6463 out_lock:
6464                 mutex_unlock(&caching_ctl->mutex);
6465                 put_caching_control(caching_ctl);
6466         }
6467         btrfs_put_block_group(block_group);
6468         return ret;
6469 }
6470
6471 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6472                                  struct extent_buffer *eb)
6473 {
6474         struct btrfs_file_extent_item *item;
6475         struct btrfs_key key;
6476         int found_type;
6477         int i;
6478
6479         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6480                 return 0;
6481
6482         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6483                 btrfs_item_key_to_cpu(eb, &key, i);
6484                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6485                         continue;
6486                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6487                 found_type = btrfs_file_extent_type(eb, item);
6488                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6489                         continue;
6490                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6491                         continue;
6492                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6493                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6494                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6495         }
6496
6497         return 0;
6498 }
6499
6500 static void
6501 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6502 {
6503         atomic_inc(&bg->reservations);
6504 }
6505
6506 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6507                                         const u64 start)
6508 {
6509         struct btrfs_block_group_cache *bg;
6510
6511         bg = btrfs_lookup_block_group(fs_info, start);
6512         ASSERT(bg);
6513         if (atomic_dec_and_test(&bg->reservations))
6514                 wake_up_atomic_t(&bg->reservations);
6515         btrfs_put_block_group(bg);
6516 }
6517
6518 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6519 {
6520         struct btrfs_space_info *space_info = bg->space_info;
6521
6522         ASSERT(bg->ro);
6523
6524         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6525                 return;
6526
6527         /*
6528          * Our block group is read only but before we set it to read only,
6529          * some task might have had allocated an extent from it already, but it
6530          * has not yet created a respective ordered extent (and added it to a
6531          * root's list of ordered extents).
6532          * Therefore wait for any task currently allocating extents, since the
6533          * block group's reservations counter is incremented while a read lock
6534          * on the groups' semaphore is held and decremented after releasing
6535          * the read access on that semaphore and creating the ordered extent.
6536          */
6537         down_write(&space_info->groups_sem);
6538         up_write(&space_info->groups_sem);
6539
6540         wait_on_atomic_t(&bg->reservations, atomic_t_wait,
6541                          TASK_UNINTERRUPTIBLE);
6542 }
6543
6544 /**
6545  * btrfs_add_reserved_bytes - update the block_group and space info counters
6546  * @cache:      The cache we are manipulating
6547  * @ram_bytes:  The number of bytes of file content, and will be same to
6548  *              @num_bytes except for the compress path.
6549  * @num_bytes:  The number of bytes in question
6550  * @delalloc:   The blocks are allocated for the delalloc write
6551  *
6552  * This is called by the allocator when it reserves space. If this is a
6553  * reservation and the block group has become read only we cannot make the
6554  * reservation and return -EAGAIN, otherwise this function always succeeds.
6555  */
6556 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6557                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6558 {
6559         struct btrfs_space_info *space_info = cache->space_info;
6560         int ret = 0;
6561
6562         spin_lock(&space_info->lock);
6563         spin_lock(&cache->lock);
6564         if (cache->ro) {
6565                 ret = -EAGAIN;
6566         } else {
6567                 cache->reserved += num_bytes;
6568                 space_info->bytes_reserved += num_bytes;
6569
6570                 trace_btrfs_space_reservation(cache->fs_info,
6571                                 "space_info", space_info->flags,
6572                                 ram_bytes, 0);
6573                 space_info->bytes_may_use -= ram_bytes;
6574                 if (delalloc)
6575                         cache->delalloc_bytes += num_bytes;
6576         }
6577         spin_unlock(&cache->lock);
6578         spin_unlock(&space_info->lock);
6579         return ret;
6580 }
6581
6582 /**
6583  * btrfs_free_reserved_bytes - update the block_group and space info counters
6584  * @cache:      The cache we are manipulating
6585  * @num_bytes:  The number of bytes in question
6586  * @delalloc:   The blocks are allocated for the delalloc write
6587  *
6588  * This is called by somebody who is freeing space that was never actually used
6589  * on disk.  For example if you reserve some space for a new leaf in transaction
6590  * A and before transaction A commits you free that leaf, you call this with
6591  * reserve set to 0 in order to clear the reservation.
6592  */
6593
6594 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6595                                      u64 num_bytes, int delalloc)
6596 {
6597         struct btrfs_space_info *space_info = cache->space_info;
6598         int ret = 0;
6599
6600         spin_lock(&space_info->lock);
6601         spin_lock(&cache->lock);
6602         if (cache->ro)
6603                 space_info->bytes_readonly += num_bytes;
6604         cache->reserved -= num_bytes;
6605         space_info->bytes_reserved -= num_bytes;
6606
6607         if (delalloc)
6608                 cache->delalloc_bytes -= num_bytes;
6609         spin_unlock(&cache->lock);
6610         spin_unlock(&space_info->lock);
6611         return ret;
6612 }
6613 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6614 {
6615         struct btrfs_caching_control *next;
6616         struct btrfs_caching_control *caching_ctl;
6617         struct btrfs_block_group_cache *cache;
6618
6619         down_write(&fs_info->commit_root_sem);
6620
6621         list_for_each_entry_safe(caching_ctl, next,
6622                                  &fs_info->caching_block_groups, list) {
6623                 cache = caching_ctl->block_group;
6624                 if (block_group_cache_done(cache)) {
6625                         cache->last_byte_to_unpin = (u64)-1;
6626                         list_del_init(&caching_ctl->list);
6627                         put_caching_control(caching_ctl);
6628                 } else {
6629                         cache->last_byte_to_unpin = caching_ctl->progress;
6630                 }
6631         }
6632
6633         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6634                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6635         else
6636                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6637
6638         up_write(&fs_info->commit_root_sem);
6639
6640         update_global_block_rsv(fs_info);
6641 }
6642
6643 /*
6644  * Returns the free cluster for the given space info and sets empty_cluster to
6645  * what it should be based on the mount options.
6646  */
6647 static struct btrfs_free_cluster *
6648 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6649                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6650 {
6651         struct btrfs_free_cluster *ret = NULL;
6652
6653         *empty_cluster = 0;
6654         if (btrfs_mixed_space_info(space_info))
6655                 return ret;
6656
6657         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6658                 ret = &fs_info->meta_alloc_cluster;
6659                 if (btrfs_test_opt(fs_info, SSD))
6660                         *empty_cluster = SZ_2M;
6661                 else
6662                         *empty_cluster = SZ_64K;
6663         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6664                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6665                 *empty_cluster = SZ_2M;
6666                 ret = &fs_info->data_alloc_cluster;
6667         }
6668
6669         return ret;
6670 }
6671
6672 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6673                               u64 start, u64 end,
6674                               const bool return_free_space)
6675 {
6676         struct btrfs_block_group_cache *cache = NULL;
6677         struct btrfs_space_info *space_info;
6678         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6679         struct btrfs_free_cluster *cluster = NULL;
6680         u64 len;
6681         u64 total_unpinned = 0;
6682         u64 empty_cluster = 0;
6683         bool readonly;
6684
6685         while (start <= end) {
6686                 readonly = false;
6687                 if (!cache ||
6688                     start >= cache->key.objectid + cache->key.offset) {
6689                         if (cache)
6690                                 btrfs_put_block_group(cache);
6691                         total_unpinned = 0;
6692                         cache = btrfs_lookup_block_group(fs_info, start);
6693                         BUG_ON(!cache); /* Logic error */
6694
6695                         cluster = fetch_cluster_info(fs_info,
6696                                                      cache->space_info,
6697                                                      &empty_cluster);
6698                         empty_cluster <<= 1;
6699                 }
6700
6701                 len = cache->key.objectid + cache->key.offset - start;
6702                 len = min(len, end + 1 - start);
6703
6704                 if (start < cache->last_byte_to_unpin) {
6705                         len = min(len, cache->last_byte_to_unpin - start);
6706                         if (return_free_space)
6707                                 btrfs_add_free_space(cache, start, len);
6708                 }
6709
6710                 start += len;
6711                 total_unpinned += len;
6712                 space_info = cache->space_info;
6713
6714                 /*
6715                  * If this space cluster has been marked as fragmented and we've
6716                  * unpinned enough in this block group to potentially allow a
6717                  * cluster to be created inside of it go ahead and clear the
6718                  * fragmented check.
6719                  */
6720                 if (cluster && cluster->fragmented &&
6721                     total_unpinned > empty_cluster) {
6722                         spin_lock(&cluster->lock);
6723                         cluster->fragmented = 0;
6724                         spin_unlock(&cluster->lock);
6725                 }
6726
6727                 spin_lock(&space_info->lock);
6728                 spin_lock(&cache->lock);
6729                 cache->pinned -= len;
6730                 space_info->bytes_pinned -= len;
6731
6732                 trace_btrfs_space_reservation(fs_info, "pinned",
6733                                               space_info->flags, len, 0);
6734                 space_info->max_extent_size = 0;
6735                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6736                 if (cache->ro) {
6737                         space_info->bytes_readonly += len;
6738                         readonly = true;
6739                 }
6740                 spin_unlock(&cache->lock);
6741                 if (!readonly && return_free_space &&
6742                     global_rsv->space_info == space_info) {
6743                         u64 to_add = len;
6744
6745                         spin_lock(&global_rsv->lock);
6746                         if (!global_rsv->full) {
6747                                 to_add = min(len, global_rsv->size -
6748                                              global_rsv->reserved);
6749                                 global_rsv->reserved += to_add;
6750                                 space_info->bytes_may_use += to_add;
6751                                 if (global_rsv->reserved >= global_rsv->size)
6752                                         global_rsv->full = 1;
6753                                 trace_btrfs_space_reservation(fs_info,
6754                                                               "space_info",
6755                                                               space_info->flags,
6756                                                               to_add, 1);
6757                                 len -= to_add;
6758                         }
6759                         spin_unlock(&global_rsv->lock);
6760                         /* Add to any tickets we may have */
6761                         if (len)
6762                                 space_info_add_new_bytes(fs_info, space_info,
6763                                                          len);
6764                 }
6765                 spin_unlock(&space_info->lock);
6766         }
6767
6768         if (cache)
6769                 btrfs_put_block_group(cache);
6770         return 0;
6771 }
6772
6773 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6774                                struct btrfs_fs_info *fs_info)
6775 {
6776         struct btrfs_block_group_cache *block_group, *tmp;
6777         struct list_head *deleted_bgs;
6778         struct extent_io_tree *unpin;
6779         u64 start;
6780         u64 end;
6781         int ret;
6782
6783         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6784                 unpin = &fs_info->freed_extents[1];
6785         else
6786                 unpin = &fs_info->freed_extents[0];
6787
6788         while (!trans->aborted) {
6789                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6790                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6791                                             EXTENT_DIRTY, NULL);
6792                 if (ret) {
6793                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6794                         break;
6795                 }
6796
6797                 if (btrfs_test_opt(fs_info, DISCARD))
6798                         ret = btrfs_discard_extent(fs_info, start,
6799                                                    end + 1 - start, NULL);
6800
6801                 clear_extent_dirty(unpin, start, end);
6802                 unpin_extent_range(fs_info, start, end, true);
6803                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6804                 cond_resched();
6805         }
6806
6807         /*
6808          * Transaction is finished.  We don't need the lock anymore.  We
6809          * do need to clean up the block groups in case of a transaction
6810          * abort.
6811          */
6812         deleted_bgs = &trans->transaction->deleted_bgs;
6813         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6814                 u64 trimmed = 0;
6815
6816                 ret = -EROFS;
6817                 if (!trans->aborted)
6818                         ret = btrfs_discard_extent(fs_info,
6819                                                    block_group->key.objectid,
6820                                                    block_group->key.offset,
6821                                                    &trimmed);
6822
6823                 list_del_init(&block_group->bg_list);
6824                 btrfs_put_block_group_trimming(block_group);
6825                 btrfs_put_block_group(block_group);
6826
6827                 if (ret) {
6828                         const char *errstr = btrfs_decode_error(ret);
6829                         btrfs_warn(fs_info,
6830                            "discard failed while removing blockgroup: errno=%d %s",
6831                                    ret, errstr);
6832                 }
6833         }
6834
6835         return 0;
6836 }
6837
6838 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6839                                 struct btrfs_fs_info *info,
6840                                 struct btrfs_delayed_ref_node *node, u64 parent,
6841                                 u64 root_objectid, u64 owner_objectid,
6842                                 u64 owner_offset, int refs_to_drop,
6843                                 struct btrfs_delayed_extent_op *extent_op)
6844 {
6845         struct btrfs_key key;
6846         struct btrfs_path *path;
6847         struct btrfs_root *extent_root = info->extent_root;
6848         struct extent_buffer *leaf;
6849         struct btrfs_extent_item *ei;
6850         struct btrfs_extent_inline_ref *iref;
6851         int ret;
6852         int is_data;
6853         int extent_slot = 0;
6854         int found_extent = 0;
6855         int num_to_del = 1;
6856         u32 item_size;
6857         u64 refs;
6858         u64 bytenr = node->bytenr;
6859         u64 num_bytes = node->num_bytes;
6860         int last_ref = 0;
6861         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6862
6863         path = btrfs_alloc_path();
6864         if (!path)
6865                 return -ENOMEM;
6866
6867         path->reada = READA_FORWARD;
6868         path->leave_spinning = 1;
6869
6870         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6871         BUG_ON(!is_data && refs_to_drop != 1);
6872
6873         if (is_data)
6874                 skinny_metadata = false;
6875
6876         ret = lookup_extent_backref(trans, info, path, &iref,
6877                                     bytenr, num_bytes, parent,
6878                                     root_objectid, owner_objectid,
6879                                     owner_offset);
6880         if (ret == 0) {
6881                 extent_slot = path->slots[0];
6882                 while (extent_slot >= 0) {
6883                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6884                                               extent_slot);
6885                         if (key.objectid != bytenr)
6886                                 break;
6887                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6888                             key.offset == num_bytes) {
6889                                 found_extent = 1;
6890                                 break;
6891                         }
6892                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6893                             key.offset == owner_objectid) {
6894                                 found_extent = 1;
6895                                 break;
6896                         }
6897                         if (path->slots[0] - extent_slot > 5)
6898                                 break;
6899                         extent_slot--;
6900                 }
6901 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6902                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6903                 if (found_extent && item_size < sizeof(*ei))
6904                         found_extent = 0;
6905 #endif
6906                 if (!found_extent) {
6907                         BUG_ON(iref);
6908                         ret = remove_extent_backref(trans, info, path, NULL,
6909                                                     refs_to_drop,
6910                                                     is_data, &last_ref);
6911                         if (ret) {
6912                                 btrfs_abort_transaction(trans, ret);
6913                                 goto out;
6914                         }
6915                         btrfs_release_path(path);
6916                         path->leave_spinning = 1;
6917
6918                         key.objectid = bytenr;
6919                         key.type = BTRFS_EXTENT_ITEM_KEY;
6920                         key.offset = num_bytes;
6921
6922                         if (!is_data && skinny_metadata) {
6923                                 key.type = BTRFS_METADATA_ITEM_KEY;
6924                                 key.offset = owner_objectid;
6925                         }
6926
6927                         ret = btrfs_search_slot(trans, extent_root,
6928                                                 &key, path, -1, 1);
6929                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6930                                 /*
6931                                  * Couldn't find our skinny metadata item,
6932                                  * see if we have ye olde extent item.
6933                                  */
6934                                 path->slots[0]--;
6935                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6936                                                       path->slots[0]);
6937                                 if (key.objectid == bytenr &&
6938                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6939                                     key.offset == num_bytes)
6940                                         ret = 0;
6941                         }
6942
6943                         if (ret > 0 && skinny_metadata) {
6944                                 skinny_metadata = false;
6945                                 key.objectid = bytenr;
6946                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6947                                 key.offset = num_bytes;
6948                                 btrfs_release_path(path);
6949                                 ret = btrfs_search_slot(trans, extent_root,
6950                                                         &key, path, -1, 1);
6951                         }
6952
6953                         if (ret) {
6954                                 btrfs_err(info,
6955                                           "umm, got %d back from search, was looking for %llu",
6956                                           ret, bytenr);
6957                                 if (ret > 0)
6958                                         btrfs_print_leaf(path->nodes[0]);
6959                         }
6960                         if (ret < 0) {
6961                                 btrfs_abort_transaction(trans, ret);
6962                                 goto out;
6963                         }
6964                         extent_slot = path->slots[0];
6965                 }
6966         } else if (WARN_ON(ret == -ENOENT)) {
6967                 btrfs_print_leaf(path->nodes[0]);
6968                 btrfs_err(info,
6969                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6970                         bytenr, parent, root_objectid, owner_objectid,
6971                         owner_offset);
6972                 btrfs_abort_transaction(trans, ret);
6973                 goto out;
6974         } else {
6975                 btrfs_abort_transaction(trans, ret);
6976                 goto out;
6977         }
6978
6979         leaf = path->nodes[0];
6980         item_size = btrfs_item_size_nr(leaf, extent_slot);
6981 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6982         if (item_size < sizeof(*ei)) {
6983                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6984                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6985                                              0);
6986                 if (ret < 0) {
6987                         btrfs_abort_transaction(trans, ret);
6988                         goto out;
6989                 }
6990
6991                 btrfs_release_path(path);
6992                 path->leave_spinning = 1;
6993
6994                 key.objectid = bytenr;
6995                 key.type = BTRFS_EXTENT_ITEM_KEY;
6996                 key.offset = num_bytes;
6997
6998                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6999                                         -1, 1);
7000                 if (ret) {
7001                         btrfs_err(info,
7002                                   "umm, got %d back from search, was looking for %llu",
7003                                 ret, bytenr);
7004                         btrfs_print_leaf(path->nodes[0]);
7005                 }
7006                 if (ret < 0) {
7007                         btrfs_abort_transaction(trans, ret);
7008                         goto out;
7009                 }
7010
7011                 extent_slot = path->slots[0];
7012                 leaf = path->nodes[0];
7013                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7014         }
7015 #endif
7016         BUG_ON(item_size < sizeof(*ei));
7017         ei = btrfs_item_ptr(leaf, extent_slot,
7018                             struct btrfs_extent_item);
7019         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7020             key.type == BTRFS_EXTENT_ITEM_KEY) {
7021                 struct btrfs_tree_block_info *bi;
7022                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7023                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7024                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7025         }
7026
7027         refs = btrfs_extent_refs(leaf, ei);
7028         if (refs < refs_to_drop) {
7029                 btrfs_err(info,
7030                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7031                           refs_to_drop, refs, bytenr);
7032                 ret = -EINVAL;
7033                 btrfs_abort_transaction(trans, ret);
7034                 goto out;
7035         }
7036         refs -= refs_to_drop;
7037
7038         if (refs > 0) {
7039                 if (extent_op)
7040                         __run_delayed_extent_op(extent_op, leaf, ei);
7041                 /*
7042                  * In the case of inline back ref, reference count will
7043                  * be updated by remove_extent_backref
7044                  */
7045                 if (iref) {
7046                         BUG_ON(!found_extent);
7047                 } else {
7048                         btrfs_set_extent_refs(leaf, ei, refs);
7049                         btrfs_mark_buffer_dirty(leaf);
7050                 }
7051                 if (found_extent) {
7052                         ret = remove_extent_backref(trans, info, path,
7053                                                     iref, refs_to_drop,
7054                                                     is_data, &last_ref);
7055                         if (ret) {
7056                                 btrfs_abort_transaction(trans, ret);
7057                                 goto out;
7058                         }
7059                 }
7060         } else {
7061                 if (found_extent) {
7062                         BUG_ON(is_data && refs_to_drop !=
7063                                extent_data_ref_count(path, iref));
7064                         if (iref) {
7065                                 BUG_ON(path->slots[0] != extent_slot);
7066                         } else {
7067                                 BUG_ON(path->slots[0] != extent_slot + 1);
7068                                 path->slots[0] = extent_slot;
7069                                 num_to_del = 2;
7070                         }
7071                 }
7072
7073                 last_ref = 1;
7074                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7075                                       num_to_del);
7076                 if (ret) {
7077                         btrfs_abort_transaction(trans, ret);
7078                         goto out;
7079                 }
7080                 btrfs_release_path(path);
7081
7082                 if (is_data) {
7083                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7084                         if (ret) {
7085                                 btrfs_abort_transaction(trans, ret);
7086                                 goto out;
7087                         }
7088                 }
7089
7090                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7091                 if (ret) {
7092                         btrfs_abort_transaction(trans, ret);
7093                         goto out;
7094                 }
7095
7096                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7097                 if (ret) {
7098                         btrfs_abort_transaction(trans, ret);
7099                         goto out;
7100                 }
7101         }
7102         btrfs_release_path(path);
7103
7104 out:
7105         btrfs_free_path(path);
7106         return ret;
7107 }
7108
7109 /*
7110  * when we free an block, it is possible (and likely) that we free the last
7111  * delayed ref for that extent as well.  This searches the delayed ref tree for
7112  * a given extent, and if there are no other delayed refs to be processed, it
7113  * removes it from the tree.
7114  */
7115 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7116                                       u64 bytenr)
7117 {
7118         struct btrfs_delayed_ref_head *head;
7119         struct btrfs_delayed_ref_root *delayed_refs;
7120         int ret = 0;
7121
7122         delayed_refs = &trans->transaction->delayed_refs;
7123         spin_lock(&delayed_refs->lock);
7124         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7125         if (!head)
7126                 goto out_delayed_unlock;
7127
7128         spin_lock(&head->lock);
7129         if (!RB_EMPTY_ROOT(&head->ref_tree))
7130                 goto out;
7131
7132         if (head->extent_op) {
7133                 if (!head->must_insert_reserved)
7134                         goto out;
7135                 btrfs_free_delayed_extent_op(head->extent_op);
7136                 head->extent_op = NULL;
7137         }
7138
7139         /*
7140          * waiting for the lock here would deadlock.  If someone else has it
7141          * locked they are already in the process of dropping it anyway
7142          */
7143         if (!mutex_trylock(&head->mutex))
7144                 goto out;
7145
7146         /*
7147          * at this point we have a head with no other entries.  Go
7148          * ahead and process it.
7149          */
7150         rb_erase(&head->href_node, &delayed_refs->href_root);
7151         RB_CLEAR_NODE(&head->href_node);
7152         atomic_dec(&delayed_refs->num_entries);
7153
7154         /*
7155          * we don't take a ref on the node because we're removing it from the
7156          * tree, so we just steal the ref the tree was holding.
7157          */
7158         delayed_refs->num_heads--;
7159         if (head->processing == 0)
7160                 delayed_refs->num_heads_ready--;
7161         head->processing = 0;
7162         spin_unlock(&head->lock);
7163         spin_unlock(&delayed_refs->lock);
7164
7165         BUG_ON(head->extent_op);
7166         if (head->must_insert_reserved)
7167                 ret = 1;
7168
7169         mutex_unlock(&head->mutex);
7170         btrfs_put_delayed_ref_head(head);
7171         return ret;
7172 out:
7173         spin_unlock(&head->lock);
7174
7175 out_delayed_unlock:
7176         spin_unlock(&delayed_refs->lock);
7177         return 0;
7178 }
7179
7180 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7181                            struct btrfs_root *root,
7182                            struct extent_buffer *buf,
7183                            u64 parent, int last_ref)
7184 {
7185         struct btrfs_fs_info *fs_info = root->fs_info;
7186         int pin = 1;
7187         int ret;
7188
7189         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7190                 int old_ref_mod, new_ref_mod;
7191
7192                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7193                                    root->root_key.objectid,
7194                                    btrfs_header_level(buf), 0,
7195                                    BTRFS_DROP_DELAYED_REF);
7196                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7197                                                  buf->len, parent,
7198                                                  root->root_key.objectid,
7199                                                  btrfs_header_level(buf),
7200                                                  BTRFS_DROP_DELAYED_REF, NULL,
7201                                                  &old_ref_mod, &new_ref_mod);
7202                 BUG_ON(ret); /* -ENOMEM */
7203                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7204         }
7205
7206         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7207                 struct btrfs_block_group_cache *cache;
7208
7209                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7210                         ret = check_ref_cleanup(trans, buf->start);
7211                         if (!ret)
7212                                 goto out;
7213                 }
7214
7215                 pin = 0;
7216                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7217
7218                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7219                         pin_down_extent(fs_info, cache, buf->start,
7220                                         buf->len, 1);
7221                         btrfs_put_block_group(cache);
7222                         goto out;
7223                 }
7224
7225                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7226
7227                 btrfs_add_free_space(cache, buf->start, buf->len);
7228                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7229                 btrfs_put_block_group(cache);
7230                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7231         }
7232 out:
7233         if (pin)
7234                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7235                                  root->root_key.objectid);
7236
7237         if (last_ref) {
7238                 /*
7239                  * Deleting the buffer, clear the corrupt flag since it doesn't
7240                  * matter anymore.
7241                  */
7242                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7243         }
7244 }
7245
7246 /* Can return -ENOMEM */
7247 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7248                       struct btrfs_root *root,
7249                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7250                       u64 owner, u64 offset)
7251 {
7252         struct btrfs_fs_info *fs_info = root->fs_info;
7253         int old_ref_mod, new_ref_mod;
7254         int ret;
7255
7256         if (btrfs_is_testing(fs_info))
7257                 return 0;
7258
7259         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7260                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7261                                    root_objectid, owner, offset,
7262                                    BTRFS_DROP_DELAYED_REF);
7263
7264         /*
7265          * tree log blocks never actually go into the extent allocation
7266          * tree, just update pinning info and exit early.
7267          */
7268         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7269                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7270                 /* unlocks the pinned mutex */
7271                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7272                 old_ref_mod = new_ref_mod = 0;
7273                 ret = 0;
7274         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7275                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7276                                                  num_bytes, parent,
7277                                                  root_objectid, (int)owner,
7278                                                  BTRFS_DROP_DELAYED_REF, NULL,
7279                                                  &old_ref_mod, &new_ref_mod);
7280         } else {
7281                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7282                                                  num_bytes, parent,
7283                                                  root_objectid, owner, offset,
7284                                                  0, BTRFS_DROP_DELAYED_REF,
7285                                                  &old_ref_mod, &new_ref_mod);
7286         }
7287
7288         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7289                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7290
7291         return ret;
7292 }
7293
7294 /*
7295  * when we wait for progress in the block group caching, its because
7296  * our allocation attempt failed at least once.  So, we must sleep
7297  * and let some progress happen before we try again.
7298  *
7299  * This function will sleep at least once waiting for new free space to
7300  * show up, and then it will check the block group free space numbers
7301  * for our min num_bytes.  Another option is to have it go ahead
7302  * and look in the rbtree for a free extent of a given size, but this
7303  * is a good start.
7304  *
7305  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7306  * any of the information in this block group.
7307  */
7308 static noinline void
7309 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7310                                 u64 num_bytes)
7311 {
7312         struct btrfs_caching_control *caching_ctl;
7313
7314         caching_ctl = get_caching_control(cache);
7315         if (!caching_ctl)
7316                 return;
7317
7318         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7319                    (cache->free_space_ctl->free_space >= num_bytes));
7320
7321         put_caching_control(caching_ctl);
7322 }
7323
7324 static noinline int
7325 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7326 {
7327         struct btrfs_caching_control *caching_ctl;
7328         int ret = 0;
7329
7330         caching_ctl = get_caching_control(cache);
7331         if (!caching_ctl)
7332                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7333
7334         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7335         if (cache->cached == BTRFS_CACHE_ERROR)
7336                 ret = -EIO;
7337         put_caching_control(caching_ctl);
7338         return ret;
7339 }
7340
7341 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7342         [BTRFS_RAID_RAID10]     = "raid10",
7343         [BTRFS_RAID_RAID1]      = "raid1",
7344         [BTRFS_RAID_DUP]        = "dup",
7345         [BTRFS_RAID_RAID0]      = "raid0",
7346         [BTRFS_RAID_SINGLE]     = "single",
7347         [BTRFS_RAID_RAID5]      = "raid5",
7348         [BTRFS_RAID_RAID6]      = "raid6",
7349 };
7350
7351 static const char *get_raid_name(enum btrfs_raid_types type)
7352 {
7353         if (type >= BTRFS_NR_RAID_TYPES)
7354                 return NULL;
7355
7356         return btrfs_raid_type_names[type];
7357 }
7358
7359 enum btrfs_loop_type {
7360         LOOP_CACHING_NOWAIT = 0,
7361         LOOP_CACHING_WAIT = 1,
7362         LOOP_ALLOC_CHUNK = 2,
7363         LOOP_NO_EMPTY_SIZE = 3,
7364 };
7365
7366 static inline void
7367 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7368                        int delalloc)
7369 {
7370         if (delalloc)
7371                 down_read(&cache->data_rwsem);
7372 }
7373
7374 static inline void
7375 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7376                        int delalloc)
7377 {
7378         btrfs_get_block_group(cache);
7379         if (delalloc)
7380                 down_read(&cache->data_rwsem);
7381 }
7382
7383 static struct btrfs_block_group_cache *
7384 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7385                    struct btrfs_free_cluster *cluster,
7386                    int delalloc)
7387 {
7388         struct btrfs_block_group_cache *used_bg = NULL;
7389
7390         spin_lock(&cluster->refill_lock);
7391         while (1) {
7392                 used_bg = cluster->block_group;
7393                 if (!used_bg)
7394                         return NULL;
7395
7396                 if (used_bg == block_group)
7397                         return used_bg;
7398
7399                 btrfs_get_block_group(used_bg);
7400
7401                 if (!delalloc)
7402                         return used_bg;
7403
7404                 if (down_read_trylock(&used_bg->data_rwsem))
7405                         return used_bg;
7406
7407                 spin_unlock(&cluster->refill_lock);
7408
7409                 /* We should only have one-level nested. */
7410                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7411
7412                 spin_lock(&cluster->refill_lock);
7413                 if (used_bg == cluster->block_group)
7414                         return used_bg;
7415
7416                 up_read(&used_bg->data_rwsem);
7417                 btrfs_put_block_group(used_bg);
7418         }
7419 }
7420
7421 static inline void
7422 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7423                          int delalloc)
7424 {
7425         if (delalloc)
7426                 up_read(&cache->data_rwsem);
7427         btrfs_put_block_group(cache);
7428 }
7429
7430 /*
7431  * walks the btree of allocated extents and find a hole of a given size.
7432  * The key ins is changed to record the hole:
7433  * ins->objectid == start position
7434  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7435  * ins->offset == the size of the hole.
7436  * Any available blocks before search_start are skipped.
7437  *
7438  * If there is no suitable free space, we will record the max size of
7439  * the free space extent currently.
7440  */
7441 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7442                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7443                                 u64 hint_byte, struct btrfs_key *ins,
7444                                 u64 flags, int delalloc)
7445 {
7446         int ret = 0;
7447         struct btrfs_root *root = fs_info->extent_root;
7448         struct btrfs_free_cluster *last_ptr = NULL;
7449         struct btrfs_block_group_cache *block_group = NULL;
7450         u64 search_start = 0;
7451         u64 max_extent_size = 0;
7452         u64 empty_cluster = 0;
7453         struct btrfs_space_info *space_info;
7454         int loop = 0;
7455         int index = btrfs_bg_flags_to_raid_index(flags);
7456         bool failed_cluster_refill = false;
7457         bool failed_alloc = false;
7458         bool use_cluster = true;
7459         bool have_caching_bg = false;
7460         bool orig_have_caching_bg = false;
7461         bool full_search = false;
7462
7463         WARN_ON(num_bytes < fs_info->sectorsize);
7464         ins->type = BTRFS_EXTENT_ITEM_KEY;
7465         ins->objectid = 0;
7466         ins->offset = 0;
7467
7468         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7469
7470         space_info = __find_space_info(fs_info, flags);
7471         if (!space_info) {
7472                 btrfs_err(fs_info, "No space info for %llu", flags);
7473                 return -ENOSPC;
7474         }
7475
7476         /*
7477          * If our free space is heavily fragmented we may not be able to make
7478          * big contiguous allocations, so instead of doing the expensive search
7479          * for free space, simply return ENOSPC with our max_extent_size so we
7480          * can go ahead and search for a more manageable chunk.
7481          *
7482          * If our max_extent_size is large enough for our allocation simply
7483          * disable clustering since we will likely not be able to find enough
7484          * space to create a cluster and induce latency trying.
7485          */
7486         if (unlikely(space_info->max_extent_size)) {
7487                 spin_lock(&space_info->lock);
7488                 if (space_info->max_extent_size &&
7489                     num_bytes > space_info->max_extent_size) {
7490                         ins->offset = space_info->max_extent_size;
7491                         spin_unlock(&space_info->lock);
7492                         return -ENOSPC;
7493                 } else if (space_info->max_extent_size) {
7494                         use_cluster = false;
7495                 }
7496                 spin_unlock(&space_info->lock);
7497         }
7498
7499         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7500         if (last_ptr) {
7501                 spin_lock(&last_ptr->lock);
7502                 if (last_ptr->block_group)
7503                         hint_byte = last_ptr->window_start;
7504                 if (last_ptr->fragmented) {
7505                         /*
7506                          * We still set window_start so we can keep track of the
7507                          * last place we found an allocation to try and save
7508                          * some time.
7509                          */
7510                         hint_byte = last_ptr->window_start;
7511                         use_cluster = false;
7512                 }
7513                 spin_unlock(&last_ptr->lock);
7514         }
7515
7516         search_start = max(search_start, first_logical_byte(fs_info, 0));
7517         search_start = max(search_start, hint_byte);
7518         if (search_start == hint_byte) {
7519                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7520                 /*
7521                  * we don't want to use the block group if it doesn't match our
7522                  * allocation bits, or if its not cached.
7523                  *
7524                  * However if we are re-searching with an ideal block group
7525                  * picked out then we don't care that the block group is cached.
7526                  */
7527                 if (block_group && block_group_bits(block_group, flags) &&
7528                     block_group->cached != BTRFS_CACHE_NO) {
7529                         down_read(&space_info->groups_sem);
7530                         if (list_empty(&block_group->list) ||
7531                             block_group->ro) {
7532                                 /*
7533                                  * someone is removing this block group,
7534                                  * we can't jump into the have_block_group
7535                                  * target because our list pointers are not
7536                                  * valid
7537                                  */
7538                                 btrfs_put_block_group(block_group);
7539                                 up_read(&space_info->groups_sem);
7540                         } else {
7541                                 index = btrfs_bg_flags_to_raid_index(
7542                                                 block_group->flags);
7543                                 btrfs_lock_block_group(block_group, delalloc);
7544                                 goto have_block_group;
7545                         }
7546                 } else if (block_group) {
7547                         btrfs_put_block_group(block_group);
7548                 }
7549         }
7550 search:
7551         have_caching_bg = false;
7552         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7553                 full_search = true;
7554         down_read(&space_info->groups_sem);
7555         list_for_each_entry(block_group, &space_info->block_groups[index],
7556                             list) {
7557                 u64 offset;
7558                 int cached;
7559
7560                 /* If the block group is read-only, we can skip it entirely. */
7561                 if (unlikely(block_group->ro))
7562                         continue;
7563
7564                 btrfs_grab_block_group(block_group, delalloc);
7565                 search_start = block_group->key.objectid;
7566
7567                 /*
7568                  * this can happen if we end up cycling through all the
7569                  * raid types, but we want to make sure we only allocate
7570                  * for the proper type.
7571                  */
7572                 if (!block_group_bits(block_group, flags)) {
7573                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7574                                 BTRFS_BLOCK_GROUP_RAID1 |
7575                                 BTRFS_BLOCK_GROUP_RAID5 |
7576                                 BTRFS_BLOCK_GROUP_RAID6 |
7577                                 BTRFS_BLOCK_GROUP_RAID10;
7578
7579                         /*
7580                          * if they asked for extra copies and this block group
7581                          * doesn't provide them, bail.  This does allow us to
7582                          * fill raid0 from raid1.
7583                          */
7584                         if ((flags & extra) && !(block_group->flags & extra))
7585                                 goto loop;
7586                 }
7587
7588 have_block_group:
7589                 cached = block_group_cache_done(block_group);
7590                 if (unlikely(!cached)) {
7591                         have_caching_bg = true;
7592                         ret = cache_block_group(block_group, 0);
7593                         BUG_ON(ret < 0);
7594                         ret = 0;
7595                 }
7596
7597                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7598                         goto loop;
7599
7600                 /*
7601                  * Ok we want to try and use the cluster allocator, so
7602                  * lets look there
7603                  */
7604                 if (last_ptr && use_cluster) {
7605                         struct btrfs_block_group_cache *used_block_group;
7606                         unsigned long aligned_cluster;
7607                         /*
7608                          * the refill lock keeps out other
7609                          * people trying to start a new cluster
7610                          */
7611                         used_block_group = btrfs_lock_cluster(block_group,
7612                                                               last_ptr,
7613                                                               delalloc);
7614                         if (!used_block_group)
7615                                 goto refill_cluster;
7616
7617                         if (used_block_group != block_group &&
7618                             (used_block_group->ro ||
7619                              !block_group_bits(used_block_group, flags)))
7620                                 goto release_cluster;
7621
7622                         offset = btrfs_alloc_from_cluster(used_block_group,
7623                                                 last_ptr,
7624                                                 num_bytes,
7625                                                 used_block_group->key.objectid,
7626                                                 &max_extent_size);
7627                         if (offset) {
7628                                 /* we have a block, we're done */
7629                                 spin_unlock(&last_ptr->refill_lock);
7630                                 trace_btrfs_reserve_extent_cluster(fs_info,
7631                                                 used_block_group,
7632                                                 search_start, num_bytes);
7633                                 if (used_block_group != block_group) {
7634                                         btrfs_release_block_group(block_group,
7635                                                                   delalloc);
7636                                         block_group = used_block_group;
7637                                 }
7638                                 goto checks;
7639                         }
7640
7641                         WARN_ON(last_ptr->block_group != used_block_group);
7642 release_cluster:
7643                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7644                          * set up a new clusters, so lets just skip it
7645                          * and let the allocator find whatever block
7646                          * it can find.  If we reach this point, we
7647                          * will have tried the cluster allocator
7648                          * plenty of times and not have found
7649                          * anything, so we are likely way too
7650                          * fragmented for the clustering stuff to find
7651                          * anything.
7652                          *
7653                          * However, if the cluster is taken from the
7654                          * current block group, release the cluster
7655                          * first, so that we stand a better chance of
7656                          * succeeding in the unclustered
7657                          * allocation.  */
7658                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7659                             used_block_group != block_group) {
7660                                 spin_unlock(&last_ptr->refill_lock);
7661                                 btrfs_release_block_group(used_block_group,
7662                                                           delalloc);
7663                                 goto unclustered_alloc;
7664                         }
7665
7666                         /*
7667                          * this cluster didn't work out, free it and
7668                          * start over
7669                          */
7670                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7671
7672                         if (used_block_group != block_group)
7673                                 btrfs_release_block_group(used_block_group,
7674                                                           delalloc);
7675 refill_cluster:
7676                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7677                                 spin_unlock(&last_ptr->refill_lock);
7678                                 goto unclustered_alloc;
7679                         }
7680
7681                         aligned_cluster = max_t(unsigned long,
7682                                                 empty_cluster + empty_size,
7683                                               block_group->full_stripe_len);
7684
7685                         /* allocate a cluster in this block group */
7686                         ret = btrfs_find_space_cluster(fs_info, block_group,
7687                                                        last_ptr, search_start,
7688                                                        num_bytes,
7689                                                        aligned_cluster);
7690                         if (ret == 0) {
7691                                 /*
7692                                  * now pull our allocation out of this
7693                                  * cluster
7694                                  */
7695                                 offset = btrfs_alloc_from_cluster(block_group,
7696                                                         last_ptr,
7697                                                         num_bytes,
7698                                                         search_start,
7699                                                         &max_extent_size);
7700                                 if (offset) {
7701                                         /* we found one, proceed */
7702                                         spin_unlock(&last_ptr->refill_lock);
7703                                         trace_btrfs_reserve_extent_cluster(fs_info,
7704                                                 block_group, search_start,
7705                                                 num_bytes);
7706                                         goto checks;
7707                                 }
7708                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7709                                    && !failed_cluster_refill) {
7710                                 spin_unlock(&last_ptr->refill_lock);
7711
7712                                 failed_cluster_refill = true;
7713                                 wait_block_group_cache_progress(block_group,
7714                                        num_bytes + empty_cluster + empty_size);
7715                                 goto have_block_group;
7716                         }
7717
7718                         /*
7719                          * at this point we either didn't find a cluster
7720                          * or we weren't able to allocate a block from our
7721                          * cluster.  Free the cluster we've been trying
7722                          * to use, and go to the next block group
7723                          */
7724                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7725                         spin_unlock(&last_ptr->refill_lock);
7726                         goto loop;
7727                 }
7728
7729 unclustered_alloc:
7730                 /*
7731                  * We are doing an unclustered alloc, set the fragmented flag so
7732                  * we don't bother trying to setup a cluster again until we get
7733                  * more space.
7734                  */
7735                 if (unlikely(last_ptr)) {
7736                         spin_lock(&last_ptr->lock);
7737                         last_ptr->fragmented = 1;
7738                         spin_unlock(&last_ptr->lock);
7739                 }
7740                 if (cached) {
7741                         struct btrfs_free_space_ctl *ctl =
7742                                 block_group->free_space_ctl;
7743
7744                         spin_lock(&ctl->tree_lock);
7745                         if (ctl->free_space <
7746                             num_bytes + empty_cluster + empty_size) {
7747                                 if (ctl->free_space > max_extent_size)
7748                                         max_extent_size = ctl->free_space;
7749                                 spin_unlock(&ctl->tree_lock);
7750                                 goto loop;
7751                         }
7752                         spin_unlock(&ctl->tree_lock);
7753                 }
7754
7755                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7756                                                     num_bytes, empty_size,
7757                                                     &max_extent_size);
7758                 /*
7759                  * If we didn't find a chunk, and we haven't failed on this
7760                  * block group before, and this block group is in the middle of
7761                  * caching and we are ok with waiting, then go ahead and wait
7762                  * for progress to be made, and set failed_alloc to true.
7763                  *
7764                  * If failed_alloc is true then we've already waited on this
7765                  * block group once and should move on to the next block group.
7766                  */
7767                 if (!offset && !failed_alloc && !cached &&
7768                     loop > LOOP_CACHING_NOWAIT) {
7769                         wait_block_group_cache_progress(block_group,
7770                                                 num_bytes + empty_size);
7771                         failed_alloc = true;
7772                         goto have_block_group;
7773                 } else if (!offset) {
7774                         goto loop;
7775                 }
7776 checks:
7777                 search_start = ALIGN(offset, fs_info->stripesize);
7778
7779                 /* move on to the next group */
7780                 if (search_start + num_bytes >
7781                     block_group->key.objectid + block_group->key.offset) {
7782                         btrfs_add_free_space(block_group, offset, num_bytes);
7783                         goto loop;
7784                 }
7785
7786                 if (offset < search_start)
7787                         btrfs_add_free_space(block_group, offset,
7788                                              search_start - offset);
7789                 BUG_ON(offset > search_start);
7790
7791                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7792                                 num_bytes, delalloc);
7793                 if (ret == -EAGAIN) {
7794                         btrfs_add_free_space(block_group, offset, num_bytes);
7795                         goto loop;
7796                 }
7797                 btrfs_inc_block_group_reservations(block_group);
7798
7799                 /* we are all good, lets return */
7800                 ins->objectid = search_start;
7801                 ins->offset = num_bytes;
7802
7803                 trace_btrfs_reserve_extent(fs_info, block_group,
7804                                            search_start, num_bytes);
7805                 btrfs_release_block_group(block_group, delalloc);
7806                 break;
7807 loop:
7808                 failed_cluster_refill = false;
7809                 failed_alloc = false;
7810                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7811                        index);
7812                 btrfs_release_block_group(block_group, delalloc);
7813                 cond_resched();
7814         }
7815         up_read(&space_info->groups_sem);
7816
7817         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7818                 && !orig_have_caching_bg)
7819                 orig_have_caching_bg = true;
7820
7821         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7822                 goto search;
7823
7824         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7825                 goto search;
7826
7827         /*
7828          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7829          *                      caching kthreads as we move along
7830          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7831          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7832          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7833          *                      again
7834          */
7835         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7836                 index = 0;
7837                 if (loop == LOOP_CACHING_NOWAIT) {
7838                         /*
7839                          * We want to skip the LOOP_CACHING_WAIT step if we
7840                          * don't have any uncached bgs and we've already done a
7841                          * full search through.
7842                          */
7843                         if (orig_have_caching_bg || !full_search)
7844                                 loop = LOOP_CACHING_WAIT;
7845                         else
7846                                 loop = LOOP_ALLOC_CHUNK;
7847                 } else {
7848                         loop++;
7849                 }
7850
7851                 if (loop == LOOP_ALLOC_CHUNK) {
7852                         struct btrfs_trans_handle *trans;
7853                         int exist = 0;
7854
7855                         trans = current->journal_info;
7856                         if (trans)
7857                                 exist = 1;
7858                         else
7859                                 trans = btrfs_join_transaction(root);
7860
7861                         if (IS_ERR(trans)) {
7862                                 ret = PTR_ERR(trans);
7863                                 goto out;
7864                         }
7865
7866                         ret = do_chunk_alloc(trans, fs_info, flags,
7867                                              CHUNK_ALLOC_FORCE);
7868
7869                         /*
7870                          * If we can't allocate a new chunk we've already looped
7871                          * through at least once, move on to the NO_EMPTY_SIZE
7872                          * case.
7873                          */
7874                         if (ret == -ENOSPC)
7875                                 loop = LOOP_NO_EMPTY_SIZE;
7876
7877                         /*
7878                          * Do not bail out on ENOSPC since we
7879                          * can do more things.
7880                          */
7881                         if (ret < 0 && ret != -ENOSPC)
7882                                 btrfs_abort_transaction(trans, ret);
7883                         else
7884                                 ret = 0;
7885                         if (!exist)
7886                                 btrfs_end_transaction(trans);
7887                         if (ret)
7888                                 goto out;
7889                 }
7890
7891                 if (loop == LOOP_NO_EMPTY_SIZE) {
7892                         /*
7893                          * Don't loop again if we already have no empty_size and
7894                          * no empty_cluster.
7895                          */
7896                         if (empty_size == 0 &&
7897                             empty_cluster == 0) {
7898                                 ret = -ENOSPC;
7899                                 goto out;
7900                         }
7901                         empty_size = 0;
7902                         empty_cluster = 0;
7903                 }
7904
7905                 goto search;
7906         } else if (!ins->objectid) {
7907                 ret = -ENOSPC;
7908         } else if (ins->objectid) {
7909                 if (!use_cluster && last_ptr) {
7910                         spin_lock(&last_ptr->lock);
7911                         last_ptr->window_start = ins->objectid;
7912                         spin_unlock(&last_ptr->lock);
7913                 }
7914                 ret = 0;
7915         }
7916 out:
7917         if (ret == -ENOSPC) {
7918                 spin_lock(&space_info->lock);
7919                 space_info->max_extent_size = max_extent_size;
7920                 spin_unlock(&space_info->lock);
7921                 ins->offset = max_extent_size;
7922         }
7923         return ret;
7924 }
7925
7926 static void dump_space_info(struct btrfs_fs_info *fs_info,
7927                             struct btrfs_space_info *info, u64 bytes,
7928                             int dump_block_groups)
7929 {
7930         struct btrfs_block_group_cache *cache;
7931         int index = 0;
7932
7933         spin_lock(&info->lock);
7934         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7935                    info->flags,
7936                    info->total_bytes - btrfs_space_info_used(info, true),
7937                    info->full ? "" : "not ");
7938         btrfs_info(fs_info,
7939                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7940                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7941                 info->bytes_reserved, info->bytes_may_use,
7942                 info->bytes_readonly);
7943         spin_unlock(&info->lock);
7944
7945         if (!dump_block_groups)
7946                 return;
7947
7948         down_read(&info->groups_sem);
7949 again:
7950         list_for_each_entry(cache, &info->block_groups[index], list) {
7951                 spin_lock(&cache->lock);
7952                 btrfs_info(fs_info,
7953                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7954                         cache->key.objectid, cache->key.offset,
7955                         btrfs_block_group_used(&cache->item), cache->pinned,
7956                         cache->reserved, cache->ro ? "[readonly]" : "");
7957                 btrfs_dump_free_space(cache, bytes);
7958                 spin_unlock(&cache->lock);
7959         }
7960         if (++index < BTRFS_NR_RAID_TYPES)
7961                 goto again;
7962         up_read(&info->groups_sem);
7963 }
7964
7965 /*
7966  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7967  *                        hole that is at least as big as @num_bytes.
7968  *
7969  * @root           -    The root that will contain this extent
7970  *
7971  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7972  *                      is used for accounting purposes. This value differs
7973  *                      from @num_bytes only in the case of compressed extents.
7974  *
7975  * @num_bytes      -    Number of bytes to allocate on-disk.
7976  *
7977  * @min_alloc_size -    Indicates the minimum amount of space that the
7978  *                      allocator should try to satisfy. In some cases
7979  *                      @num_bytes may be larger than what is required and if
7980  *                      the filesystem is fragmented then allocation fails.
7981  *                      However, the presence of @min_alloc_size gives a
7982  *                      chance to try and satisfy the smaller allocation.
7983  *
7984  * @empty_size     -    A hint that you plan on doing more COW. This is the
7985  *                      size in bytes the allocator should try to find free
7986  *                      next to the block it returns.  This is just a hint and
7987  *                      may be ignored by the allocator.
7988  *
7989  * @hint_byte      -    Hint to the allocator to start searching above the byte
7990  *                      address passed. It might be ignored.
7991  *
7992  * @ins            -    This key is modified to record the found hole. It will
7993  *                      have the following values:
7994  *                      ins->objectid == start position
7995  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7996  *                      ins->offset == the size of the hole.
7997  *
7998  * @is_data        -    Boolean flag indicating whether an extent is
7999  *                      allocated for data (true) or metadata (false)
8000  *
8001  * @delalloc       -    Boolean flag indicating whether this allocation is for
8002  *                      delalloc or not. If 'true' data_rwsem of block groups
8003  *                      is going to be acquired.
8004  *
8005  *
8006  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8007  * case -ENOSPC is returned then @ins->offset will contain the size of the
8008  * largest available hole the allocator managed to find.
8009  */
8010 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8011                          u64 num_bytes, u64 min_alloc_size,
8012                          u64 empty_size, u64 hint_byte,
8013                          struct btrfs_key *ins, int is_data, int delalloc)
8014 {
8015         struct btrfs_fs_info *fs_info = root->fs_info;
8016         bool final_tried = num_bytes == min_alloc_size;
8017         u64 flags;
8018         int ret;
8019
8020         flags = get_alloc_profile_by_root(root, is_data);
8021 again:
8022         WARN_ON(num_bytes < fs_info->sectorsize);
8023         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8024                                hint_byte, ins, flags, delalloc);
8025         if (!ret && !is_data) {
8026                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8027         } else if (ret == -ENOSPC) {
8028                 if (!final_tried && ins->offset) {
8029                         num_bytes = min(num_bytes >> 1, ins->offset);
8030                         num_bytes = round_down(num_bytes,
8031                                                fs_info->sectorsize);
8032                         num_bytes = max(num_bytes, min_alloc_size);
8033                         ram_bytes = num_bytes;
8034                         if (num_bytes == min_alloc_size)
8035                                 final_tried = true;
8036                         goto again;
8037                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8038                         struct btrfs_space_info *sinfo;
8039
8040                         sinfo = __find_space_info(fs_info, flags);
8041                         btrfs_err(fs_info,
8042                                   "allocation failed flags %llu, wanted %llu",
8043                                   flags, num_bytes);
8044                         if (sinfo)
8045                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8046                 }
8047         }
8048
8049         return ret;
8050 }
8051
8052 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8053                                         u64 start, u64 len,
8054                                         int pin, int delalloc)
8055 {
8056         struct btrfs_block_group_cache *cache;
8057         int ret = 0;
8058
8059         cache = btrfs_lookup_block_group(fs_info, start);
8060         if (!cache) {
8061                 btrfs_err(fs_info, "Unable to find block group for %llu",
8062                           start);
8063                 return -ENOSPC;
8064         }
8065
8066         if (pin)
8067                 pin_down_extent(fs_info, cache, start, len, 1);
8068         else {
8069                 if (btrfs_test_opt(fs_info, DISCARD))
8070                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8071                 btrfs_add_free_space(cache, start, len);
8072                 btrfs_free_reserved_bytes(cache, len, delalloc);
8073                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8074         }
8075
8076         btrfs_put_block_group(cache);
8077         return ret;
8078 }
8079
8080 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8081                                u64 start, u64 len, int delalloc)
8082 {
8083         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8084 }
8085
8086 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8087                                        u64 start, u64 len)
8088 {
8089         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8090 }
8091
8092 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8093                                       struct btrfs_fs_info *fs_info,
8094                                       u64 parent, u64 root_objectid,
8095                                       u64 flags, u64 owner, u64 offset,
8096                                       struct btrfs_key *ins, int ref_mod)
8097 {
8098         int ret;
8099         struct btrfs_extent_item *extent_item;
8100         struct btrfs_extent_inline_ref *iref;
8101         struct btrfs_path *path;
8102         struct extent_buffer *leaf;
8103         int type;
8104         u32 size;
8105
8106         if (parent > 0)
8107                 type = BTRFS_SHARED_DATA_REF_KEY;
8108         else
8109                 type = BTRFS_EXTENT_DATA_REF_KEY;
8110
8111         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8112
8113         path = btrfs_alloc_path();
8114         if (!path)
8115                 return -ENOMEM;
8116
8117         path->leave_spinning = 1;
8118         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8119                                       ins, size);
8120         if (ret) {
8121                 btrfs_free_path(path);
8122                 return ret;
8123         }
8124
8125         leaf = path->nodes[0];
8126         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8127                                      struct btrfs_extent_item);
8128         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8129         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8130         btrfs_set_extent_flags(leaf, extent_item,
8131                                flags | BTRFS_EXTENT_FLAG_DATA);
8132
8133         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8134         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8135         if (parent > 0) {
8136                 struct btrfs_shared_data_ref *ref;
8137                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8138                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8139                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8140         } else {
8141                 struct btrfs_extent_data_ref *ref;
8142                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8143                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8144                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8145                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8146                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8147         }
8148
8149         btrfs_mark_buffer_dirty(path->nodes[0]);
8150         btrfs_free_path(path);
8151
8152         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8153                                           ins->offset);
8154         if (ret)
8155                 return ret;
8156
8157         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8158         if (ret) { /* -ENOENT, logic error */
8159                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8160                         ins->objectid, ins->offset);
8161                 BUG();
8162         }
8163         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8164         return ret;
8165 }
8166
8167 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8168                                      struct btrfs_fs_info *fs_info,
8169                                      u64 parent, u64 root_objectid,
8170                                      u64 flags, struct btrfs_disk_key *key,
8171                                      int level, struct btrfs_key *ins)
8172 {
8173         int ret;
8174         struct btrfs_extent_item *extent_item;
8175         struct btrfs_tree_block_info *block_info;
8176         struct btrfs_extent_inline_ref *iref;
8177         struct btrfs_path *path;
8178         struct extent_buffer *leaf;
8179         u32 size = sizeof(*extent_item) + sizeof(*iref);
8180         u64 num_bytes = ins->offset;
8181         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8182
8183         if (!skinny_metadata)
8184                 size += sizeof(*block_info);
8185
8186         path = btrfs_alloc_path();
8187         if (!path) {
8188                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8189                                                    fs_info->nodesize);
8190                 return -ENOMEM;
8191         }
8192
8193         path->leave_spinning = 1;
8194         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8195                                       ins, size);
8196         if (ret) {
8197                 btrfs_free_path(path);
8198                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8199                                                    fs_info->nodesize);
8200                 return ret;
8201         }
8202
8203         leaf = path->nodes[0];
8204         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8205                                      struct btrfs_extent_item);
8206         btrfs_set_extent_refs(leaf, extent_item, 1);
8207         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8208         btrfs_set_extent_flags(leaf, extent_item,
8209                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8210
8211         if (skinny_metadata) {
8212                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8213                 num_bytes = fs_info->nodesize;
8214         } else {
8215                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8216                 btrfs_set_tree_block_key(leaf, block_info, key);
8217                 btrfs_set_tree_block_level(leaf, block_info, level);
8218                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8219         }
8220
8221         if (parent > 0) {
8222                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8223                 btrfs_set_extent_inline_ref_type(leaf, iref,
8224                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8225                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8226         } else {
8227                 btrfs_set_extent_inline_ref_type(leaf, iref,
8228                                                  BTRFS_TREE_BLOCK_REF_KEY);
8229                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8230         }
8231
8232         btrfs_mark_buffer_dirty(leaf);
8233         btrfs_free_path(path);
8234
8235         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8236                                           num_bytes);
8237         if (ret)
8238                 return ret;
8239
8240         ret = update_block_group(trans, fs_info, ins->objectid,
8241                                  fs_info->nodesize, 1);
8242         if (ret) { /* -ENOENT, logic error */
8243                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8244                         ins->objectid, ins->offset);
8245                 BUG();
8246         }
8247
8248         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8249                                           fs_info->nodesize);
8250         return ret;
8251 }
8252
8253 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8254                                      struct btrfs_root *root, u64 owner,
8255                                      u64 offset, u64 ram_bytes,
8256                                      struct btrfs_key *ins)
8257 {
8258         struct btrfs_fs_info *fs_info = root->fs_info;
8259         int ret;
8260
8261         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8262
8263         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8264                            root->root_key.objectid, owner, offset,
8265                            BTRFS_ADD_DELAYED_EXTENT);
8266
8267         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8268                                          ins->offset, 0,
8269                                          root->root_key.objectid, owner,
8270                                          offset, ram_bytes,
8271                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8272         return ret;
8273 }
8274
8275 /*
8276  * this is used by the tree logging recovery code.  It records that
8277  * an extent has been allocated and makes sure to clear the free
8278  * space cache bits as well
8279  */
8280 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8281                                    struct btrfs_fs_info *fs_info,
8282                                    u64 root_objectid, u64 owner, u64 offset,
8283                                    struct btrfs_key *ins)
8284 {
8285         int ret;
8286         struct btrfs_block_group_cache *block_group;
8287         struct btrfs_space_info *space_info;
8288
8289         /*
8290          * Mixed block groups will exclude before processing the log so we only
8291          * need to do the exclude dance if this fs isn't mixed.
8292          */
8293         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8294                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8295                                               ins->offset);
8296                 if (ret)
8297                         return ret;
8298         }
8299
8300         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8301         if (!block_group)
8302                 return -EINVAL;
8303
8304         space_info = block_group->space_info;
8305         spin_lock(&space_info->lock);
8306         spin_lock(&block_group->lock);
8307         space_info->bytes_reserved += ins->offset;
8308         block_group->reserved += ins->offset;
8309         spin_unlock(&block_group->lock);
8310         spin_unlock(&space_info->lock);
8311
8312         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8313                                          0, owner, offset, ins, 1);
8314         btrfs_put_block_group(block_group);
8315         return ret;
8316 }
8317
8318 static struct extent_buffer *
8319 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8320                       u64 bytenr, int level)
8321 {
8322         struct btrfs_fs_info *fs_info = root->fs_info;
8323         struct extent_buffer *buf;
8324
8325         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8326         if (IS_ERR(buf))
8327                 return buf;
8328
8329         btrfs_set_header_generation(buf, trans->transid);
8330         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8331         btrfs_tree_lock(buf);
8332         clean_tree_block(fs_info, buf);
8333         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8334
8335         btrfs_set_lock_blocking(buf);
8336         set_extent_buffer_uptodate(buf);
8337
8338         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8339                 buf->log_index = root->log_transid % 2;
8340                 /*
8341                  * we allow two log transactions at a time, use different
8342                  * EXENT bit to differentiate dirty pages.
8343                  */
8344                 if (buf->log_index == 0)
8345                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8346                                         buf->start + buf->len - 1, GFP_NOFS);
8347                 else
8348                         set_extent_new(&root->dirty_log_pages, buf->start,
8349                                         buf->start + buf->len - 1);
8350         } else {
8351                 buf->log_index = -1;
8352                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8353                          buf->start + buf->len - 1, GFP_NOFS);
8354         }
8355         trans->dirty = true;
8356         /* this returns a buffer locked for blocking */
8357         return buf;
8358 }
8359
8360 static struct btrfs_block_rsv *
8361 use_block_rsv(struct btrfs_trans_handle *trans,
8362               struct btrfs_root *root, u32 blocksize)
8363 {
8364         struct btrfs_fs_info *fs_info = root->fs_info;
8365         struct btrfs_block_rsv *block_rsv;
8366         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8367         int ret;
8368         bool global_updated = false;
8369
8370         block_rsv = get_block_rsv(trans, root);
8371
8372         if (unlikely(block_rsv->size == 0))
8373                 goto try_reserve;
8374 again:
8375         ret = block_rsv_use_bytes(block_rsv, blocksize);
8376         if (!ret)
8377                 return block_rsv;
8378
8379         if (block_rsv->failfast)
8380                 return ERR_PTR(ret);
8381
8382         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8383                 global_updated = true;
8384                 update_global_block_rsv(fs_info);
8385                 goto again;
8386         }
8387
8388         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8389                 static DEFINE_RATELIMIT_STATE(_rs,
8390                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8391                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8392                 if (__ratelimit(&_rs))
8393                         WARN(1, KERN_DEBUG
8394                                 "BTRFS: block rsv returned %d\n", ret);
8395         }
8396 try_reserve:
8397         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8398                                      BTRFS_RESERVE_NO_FLUSH);
8399         if (!ret)
8400                 return block_rsv;
8401         /*
8402          * If we couldn't reserve metadata bytes try and use some from
8403          * the global reserve if its space type is the same as the global
8404          * reservation.
8405          */
8406         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8407             block_rsv->space_info == global_rsv->space_info) {
8408                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8409                 if (!ret)
8410                         return global_rsv;
8411         }
8412         return ERR_PTR(ret);
8413 }
8414
8415 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8416                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8417 {
8418         block_rsv_add_bytes(block_rsv, blocksize, 0);
8419         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8420 }
8421
8422 /*
8423  * finds a free extent and does all the dirty work required for allocation
8424  * returns the tree buffer or an ERR_PTR on error.
8425  */
8426 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8427                                              struct btrfs_root *root,
8428                                              u64 parent, u64 root_objectid,
8429                                              const struct btrfs_disk_key *key,
8430                                              int level, u64 hint,
8431                                              u64 empty_size)
8432 {
8433         struct btrfs_fs_info *fs_info = root->fs_info;
8434         struct btrfs_key ins;
8435         struct btrfs_block_rsv *block_rsv;
8436         struct extent_buffer *buf;
8437         struct btrfs_delayed_extent_op *extent_op;
8438         u64 flags = 0;
8439         int ret;
8440         u32 blocksize = fs_info->nodesize;
8441         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8442
8443 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8444         if (btrfs_is_testing(fs_info)) {
8445                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8446                                             level);
8447                 if (!IS_ERR(buf))
8448                         root->alloc_bytenr += blocksize;
8449                 return buf;
8450         }
8451 #endif
8452
8453         block_rsv = use_block_rsv(trans, root, blocksize);
8454         if (IS_ERR(block_rsv))
8455                 return ERR_CAST(block_rsv);
8456
8457         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8458                                    empty_size, hint, &ins, 0, 0);
8459         if (ret)
8460                 goto out_unuse;
8461
8462         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8463         if (IS_ERR(buf)) {
8464                 ret = PTR_ERR(buf);
8465                 goto out_free_reserved;
8466         }
8467
8468         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8469                 if (parent == 0)
8470                         parent = ins.objectid;
8471                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8472         } else
8473                 BUG_ON(parent > 0);
8474
8475         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8476                 extent_op = btrfs_alloc_delayed_extent_op();
8477                 if (!extent_op) {
8478                         ret = -ENOMEM;
8479                         goto out_free_buf;
8480                 }
8481                 if (key)
8482                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8483                 else
8484                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8485                 extent_op->flags_to_set = flags;
8486                 extent_op->update_key = skinny_metadata ? false : true;
8487                 extent_op->update_flags = true;
8488                 extent_op->is_data = false;
8489                 extent_op->level = level;
8490
8491                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8492                                    root_objectid, level, 0,
8493                                    BTRFS_ADD_DELAYED_EXTENT);
8494                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8495                                                  ins.offset, parent,
8496                                                  root_objectid, level,
8497                                                  BTRFS_ADD_DELAYED_EXTENT,
8498                                                  extent_op, NULL, NULL);
8499                 if (ret)
8500                         goto out_free_delayed;
8501         }
8502         return buf;
8503
8504 out_free_delayed:
8505         btrfs_free_delayed_extent_op(extent_op);
8506 out_free_buf:
8507         free_extent_buffer(buf);
8508 out_free_reserved:
8509         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8510 out_unuse:
8511         unuse_block_rsv(fs_info, block_rsv, blocksize);
8512         return ERR_PTR(ret);
8513 }
8514
8515 struct walk_control {
8516         u64 refs[BTRFS_MAX_LEVEL];
8517         u64 flags[BTRFS_MAX_LEVEL];
8518         struct btrfs_key update_progress;
8519         int stage;
8520         int level;
8521         int shared_level;
8522         int update_ref;
8523         int keep_locks;
8524         int reada_slot;
8525         int reada_count;
8526         int for_reloc;
8527 };
8528
8529 #define DROP_REFERENCE  1
8530 #define UPDATE_BACKREF  2
8531
8532 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8533                                      struct btrfs_root *root,
8534                                      struct walk_control *wc,
8535                                      struct btrfs_path *path)
8536 {
8537         struct btrfs_fs_info *fs_info = root->fs_info;
8538         u64 bytenr;
8539         u64 generation;
8540         u64 refs;
8541         u64 flags;
8542         u32 nritems;
8543         struct btrfs_key key;
8544         struct extent_buffer *eb;
8545         int ret;
8546         int slot;
8547         int nread = 0;
8548
8549         if (path->slots[wc->level] < wc->reada_slot) {
8550                 wc->reada_count = wc->reada_count * 2 / 3;
8551                 wc->reada_count = max(wc->reada_count, 2);
8552         } else {
8553                 wc->reada_count = wc->reada_count * 3 / 2;
8554                 wc->reada_count = min_t(int, wc->reada_count,
8555                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8556         }
8557
8558         eb = path->nodes[wc->level];
8559         nritems = btrfs_header_nritems(eb);
8560
8561         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8562                 if (nread >= wc->reada_count)
8563                         break;
8564
8565                 cond_resched();
8566                 bytenr = btrfs_node_blockptr(eb, slot);
8567                 generation = btrfs_node_ptr_generation(eb, slot);
8568
8569                 if (slot == path->slots[wc->level])
8570                         goto reada;
8571
8572                 if (wc->stage == UPDATE_BACKREF &&
8573                     generation <= root->root_key.offset)
8574                         continue;
8575
8576                 /* We don't lock the tree block, it's OK to be racy here */
8577                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8578                                                wc->level - 1, 1, &refs,
8579                                                &flags);
8580                 /* We don't care about errors in readahead. */
8581                 if (ret < 0)
8582                         continue;
8583                 BUG_ON(refs == 0);
8584
8585                 if (wc->stage == DROP_REFERENCE) {
8586                         if (refs == 1)
8587                                 goto reada;
8588
8589                         if (wc->level == 1 &&
8590                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8591                                 continue;
8592                         if (!wc->update_ref ||
8593                             generation <= root->root_key.offset)
8594                                 continue;
8595                         btrfs_node_key_to_cpu(eb, &key, slot);
8596                         ret = btrfs_comp_cpu_keys(&key,
8597                                                   &wc->update_progress);
8598                         if (ret < 0)
8599                                 continue;
8600                 } else {
8601                         if (wc->level == 1 &&
8602                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8603                                 continue;
8604                 }
8605 reada:
8606                 readahead_tree_block(fs_info, bytenr);
8607                 nread++;
8608         }
8609         wc->reada_slot = slot;
8610 }
8611
8612 /*
8613  * helper to process tree block while walking down the tree.
8614  *
8615  * when wc->stage == UPDATE_BACKREF, this function updates
8616  * back refs for pointers in the block.
8617  *
8618  * NOTE: return value 1 means we should stop walking down.
8619  */
8620 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8621                                    struct btrfs_root *root,
8622                                    struct btrfs_path *path,
8623                                    struct walk_control *wc, int lookup_info)
8624 {
8625         struct btrfs_fs_info *fs_info = root->fs_info;
8626         int level = wc->level;
8627         struct extent_buffer *eb = path->nodes[level];
8628         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8629         int ret;
8630
8631         if (wc->stage == UPDATE_BACKREF &&
8632             btrfs_header_owner(eb) != root->root_key.objectid)
8633                 return 1;
8634
8635         /*
8636          * when reference count of tree block is 1, it won't increase
8637          * again. once full backref flag is set, we never clear it.
8638          */
8639         if (lookup_info &&
8640             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8641              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8642                 BUG_ON(!path->locks[level]);
8643                 ret = btrfs_lookup_extent_info(trans, fs_info,
8644                                                eb->start, level, 1,
8645                                                &wc->refs[level],
8646                                                &wc->flags[level]);
8647                 BUG_ON(ret == -ENOMEM);
8648                 if (ret)
8649                         return ret;
8650                 BUG_ON(wc->refs[level] == 0);
8651         }
8652
8653         if (wc->stage == DROP_REFERENCE) {
8654                 if (wc->refs[level] > 1)
8655                         return 1;
8656
8657                 if (path->locks[level] && !wc->keep_locks) {
8658                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8659                         path->locks[level] = 0;
8660                 }
8661                 return 0;
8662         }
8663
8664         /* wc->stage == UPDATE_BACKREF */
8665         if (!(wc->flags[level] & flag)) {
8666                 BUG_ON(!path->locks[level]);
8667                 ret = btrfs_inc_ref(trans, root, eb, 1);
8668                 BUG_ON(ret); /* -ENOMEM */
8669                 ret = btrfs_dec_ref(trans, root, eb, 0);
8670                 BUG_ON(ret); /* -ENOMEM */
8671                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8672                                                   eb->len, flag,
8673                                                   btrfs_header_level(eb), 0);
8674                 BUG_ON(ret); /* -ENOMEM */
8675                 wc->flags[level] |= flag;
8676         }
8677
8678         /*
8679          * the block is shared by multiple trees, so it's not good to
8680          * keep the tree lock
8681          */
8682         if (path->locks[level] && level > 0) {
8683                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8684                 path->locks[level] = 0;
8685         }
8686         return 0;
8687 }
8688
8689 /*
8690  * helper to process tree block pointer.
8691  *
8692  * when wc->stage == DROP_REFERENCE, this function checks
8693  * reference count of the block pointed to. if the block
8694  * is shared and we need update back refs for the subtree
8695  * rooted at the block, this function changes wc->stage to
8696  * UPDATE_BACKREF. if the block is shared and there is no
8697  * need to update back, this function drops the reference
8698  * to the block.
8699  *
8700  * NOTE: return value 1 means we should stop walking down.
8701  */
8702 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8703                                  struct btrfs_root *root,
8704                                  struct btrfs_path *path,
8705                                  struct walk_control *wc, int *lookup_info)
8706 {
8707         struct btrfs_fs_info *fs_info = root->fs_info;
8708         u64 bytenr;
8709         u64 generation;
8710         u64 parent;
8711         u32 blocksize;
8712         struct btrfs_key key;
8713         struct extent_buffer *next;
8714         int level = wc->level;
8715         int reada = 0;
8716         int ret = 0;
8717         bool need_account = false;
8718
8719         generation = btrfs_node_ptr_generation(path->nodes[level],
8720                                                path->slots[level]);
8721         /*
8722          * if the lower level block was created before the snapshot
8723          * was created, we know there is no need to update back refs
8724          * for the subtree
8725          */
8726         if (wc->stage == UPDATE_BACKREF &&
8727             generation <= root->root_key.offset) {
8728                 *lookup_info = 1;
8729                 return 1;
8730         }
8731
8732         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8733         blocksize = fs_info->nodesize;
8734
8735         next = find_extent_buffer(fs_info, bytenr);
8736         if (!next) {
8737                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8738                 if (IS_ERR(next))
8739                         return PTR_ERR(next);
8740
8741                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8742                                                level - 1);
8743                 reada = 1;
8744         }
8745         btrfs_tree_lock(next);
8746         btrfs_set_lock_blocking(next);
8747
8748         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8749                                        &wc->refs[level - 1],
8750                                        &wc->flags[level - 1]);
8751         if (ret < 0)
8752                 goto out_unlock;
8753
8754         if (unlikely(wc->refs[level - 1] == 0)) {
8755                 btrfs_err(fs_info, "Missing references.");
8756                 ret = -EIO;
8757                 goto out_unlock;
8758         }
8759         *lookup_info = 0;
8760
8761         if (wc->stage == DROP_REFERENCE) {
8762                 if (wc->refs[level - 1] > 1) {
8763                         need_account = true;
8764                         if (level == 1 &&
8765                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8766                                 goto skip;
8767
8768                         if (!wc->update_ref ||
8769                             generation <= root->root_key.offset)
8770                                 goto skip;
8771
8772                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8773                                               path->slots[level]);
8774                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8775                         if (ret < 0)
8776                                 goto skip;
8777
8778                         wc->stage = UPDATE_BACKREF;
8779                         wc->shared_level = level - 1;
8780                 }
8781         } else {
8782                 if (level == 1 &&
8783                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8784                         goto skip;
8785         }
8786
8787         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8788                 btrfs_tree_unlock(next);
8789                 free_extent_buffer(next);
8790                 next = NULL;
8791                 *lookup_info = 1;
8792         }
8793
8794         if (!next) {
8795                 if (reada && level == 1)
8796                         reada_walk_down(trans, root, wc, path);
8797                 next = read_tree_block(fs_info, bytenr, generation);
8798                 if (IS_ERR(next)) {
8799                         return PTR_ERR(next);
8800                 } else if (!extent_buffer_uptodate(next)) {
8801                         free_extent_buffer(next);
8802                         return -EIO;
8803                 }
8804                 btrfs_tree_lock(next);
8805                 btrfs_set_lock_blocking(next);
8806         }
8807
8808         level--;
8809         ASSERT(level == btrfs_header_level(next));
8810         if (level != btrfs_header_level(next)) {
8811                 btrfs_err(root->fs_info, "mismatched level");
8812                 ret = -EIO;
8813                 goto out_unlock;
8814         }
8815         path->nodes[level] = next;
8816         path->slots[level] = 0;
8817         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8818         wc->level = level;
8819         if (wc->level == 1)
8820                 wc->reada_slot = 0;
8821         return 0;
8822 skip:
8823         wc->refs[level - 1] = 0;
8824         wc->flags[level - 1] = 0;
8825         if (wc->stage == DROP_REFERENCE) {
8826                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8827                         parent = path->nodes[level]->start;
8828                 } else {
8829                         ASSERT(root->root_key.objectid ==
8830                                btrfs_header_owner(path->nodes[level]));
8831                         if (root->root_key.objectid !=
8832                             btrfs_header_owner(path->nodes[level])) {
8833                                 btrfs_err(root->fs_info,
8834                                                 "mismatched block owner");
8835                                 ret = -EIO;
8836                                 goto out_unlock;
8837                         }
8838                         parent = 0;
8839                 }
8840
8841                 if (need_account) {
8842                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8843                                                          generation, level - 1);
8844                         if (ret) {
8845                                 btrfs_err_rl(fs_info,
8846                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8847                                              ret);
8848                         }
8849                 }
8850                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8851                                         parent, root->root_key.objectid,
8852                                         level - 1, 0);
8853                 if (ret)
8854                         goto out_unlock;
8855         }
8856
8857         *lookup_info = 1;
8858         ret = 1;
8859
8860 out_unlock:
8861         btrfs_tree_unlock(next);
8862         free_extent_buffer(next);
8863
8864         return ret;
8865 }
8866
8867 /*
8868  * helper to process tree block while walking up the tree.
8869  *
8870  * when wc->stage == DROP_REFERENCE, this function drops
8871  * reference count on the block.
8872  *
8873  * when wc->stage == UPDATE_BACKREF, this function changes
8874  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8875  * to UPDATE_BACKREF previously while processing the block.
8876  *
8877  * NOTE: return value 1 means we should stop walking up.
8878  */
8879 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8880                                  struct btrfs_root *root,
8881                                  struct btrfs_path *path,
8882                                  struct walk_control *wc)
8883 {
8884         struct btrfs_fs_info *fs_info = root->fs_info;
8885         int ret;
8886         int level = wc->level;
8887         struct extent_buffer *eb = path->nodes[level];
8888         u64 parent = 0;
8889
8890         if (wc->stage == UPDATE_BACKREF) {
8891                 BUG_ON(wc->shared_level < level);
8892                 if (level < wc->shared_level)
8893                         goto out;
8894
8895                 ret = find_next_key(path, level + 1, &wc->update_progress);
8896                 if (ret > 0)
8897                         wc->update_ref = 0;
8898
8899                 wc->stage = DROP_REFERENCE;
8900                 wc->shared_level = -1;
8901                 path->slots[level] = 0;
8902
8903                 /*
8904                  * check reference count again if the block isn't locked.
8905                  * we should start walking down the tree again if reference
8906                  * count is one.
8907                  */
8908                 if (!path->locks[level]) {
8909                         BUG_ON(level == 0);
8910                         btrfs_tree_lock(eb);
8911                         btrfs_set_lock_blocking(eb);
8912                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8913
8914                         ret = btrfs_lookup_extent_info(trans, fs_info,
8915                                                        eb->start, level, 1,
8916                                                        &wc->refs[level],
8917                                                        &wc->flags[level]);
8918                         if (ret < 0) {
8919                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8920                                 path->locks[level] = 0;
8921                                 return ret;
8922                         }
8923                         BUG_ON(wc->refs[level] == 0);
8924                         if (wc->refs[level] == 1) {
8925                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8926                                 path->locks[level] = 0;
8927                                 return 1;
8928                         }
8929                 }
8930         }
8931
8932         /* wc->stage == DROP_REFERENCE */
8933         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8934
8935         if (wc->refs[level] == 1) {
8936                 if (level == 0) {
8937                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8938                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8939                         else
8940                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8941                         BUG_ON(ret); /* -ENOMEM */
8942                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8943                         if (ret) {
8944                                 btrfs_err_rl(fs_info,
8945                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8946                                              ret);
8947                         }
8948                 }
8949                 /* make block locked assertion in clean_tree_block happy */
8950                 if (!path->locks[level] &&
8951                     btrfs_header_generation(eb) == trans->transid) {
8952                         btrfs_tree_lock(eb);
8953                         btrfs_set_lock_blocking(eb);
8954                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8955                 }
8956                 clean_tree_block(fs_info, eb);
8957         }
8958
8959         if (eb == root->node) {
8960                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8961                         parent = eb->start;
8962                 else
8963                         BUG_ON(root->root_key.objectid !=
8964                                btrfs_header_owner(eb));
8965         } else {
8966                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8967                         parent = path->nodes[level + 1]->start;
8968                 else
8969                         BUG_ON(root->root_key.objectid !=
8970                                btrfs_header_owner(path->nodes[level + 1]));
8971         }
8972
8973         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8974 out:
8975         wc->refs[level] = 0;
8976         wc->flags[level] = 0;
8977         return 0;
8978 }
8979
8980 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8981                                    struct btrfs_root *root,
8982                                    struct btrfs_path *path,
8983                                    struct walk_control *wc)
8984 {
8985         int level = wc->level;
8986         int lookup_info = 1;
8987         int ret;
8988
8989         while (level >= 0) {
8990                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8991                 if (ret > 0)
8992                         break;
8993
8994                 if (level == 0)
8995                         break;
8996
8997                 if (path->slots[level] >=
8998                     btrfs_header_nritems(path->nodes[level]))
8999                         break;
9000
9001                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9002                 if (ret > 0) {
9003                         path->slots[level]++;
9004                         continue;
9005                 } else if (ret < 0)
9006                         return ret;
9007                 level = wc->level;
9008         }
9009         return 0;
9010 }
9011
9012 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9013                                  struct btrfs_root *root,
9014                                  struct btrfs_path *path,
9015                                  struct walk_control *wc, int max_level)
9016 {
9017         int level = wc->level;
9018         int ret;
9019
9020         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9021         while (level < max_level && path->nodes[level]) {
9022                 wc->level = level;
9023                 if (path->slots[level] + 1 <
9024                     btrfs_header_nritems(path->nodes[level])) {
9025                         path->slots[level]++;
9026                         return 0;
9027                 } else {
9028                         ret = walk_up_proc(trans, root, path, wc);
9029                         if (ret > 0)
9030                                 return 0;
9031
9032                         if (path->locks[level]) {
9033                                 btrfs_tree_unlock_rw(path->nodes[level],
9034                                                      path->locks[level]);
9035                                 path->locks[level] = 0;
9036                         }
9037                         free_extent_buffer(path->nodes[level]);
9038                         path->nodes[level] = NULL;
9039                         level++;
9040                 }
9041         }
9042         return 1;
9043 }
9044
9045 /*
9046  * drop a subvolume tree.
9047  *
9048  * this function traverses the tree freeing any blocks that only
9049  * referenced by the tree.
9050  *
9051  * when a shared tree block is found. this function decreases its
9052  * reference count by one. if update_ref is true, this function
9053  * also make sure backrefs for the shared block and all lower level
9054  * blocks are properly updated.
9055  *
9056  * If called with for_reloc == 0, may exit early with -EAGAIN
9057  */
9058 int btrfs_drop_snapshot(struct btrfs_root *root,
9059                          struct btrfs_block_rsv *block_rsv, int update_ref,
9060                          int for_reloc)
9061 {
9062         struct btrfs_fs_info *fs_info = root->fs_info;
9063         struct btrfs_path *path;
9064         struct btrfs_trans_handle *trans;
9065         struct btrfs_root *tree_root = fs_info->tree_root;
9066         struct btrfs_root_item *root_item = &root->root_item;
9067         struct walk_control *wc;
9068         struct btrfs_key key;
9069         int err = 0;
9070         int ret;
9071         int level;
9072         bool root_dropped = false;
9073
9074         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9075
9076         path = btrfs_alloc_path();
9077         if (!path) {
9078                 err = -ENOMEM;
9079                 goto out;
9080         }
9081
9082         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9083         if (!wc) {
9084                 btrfs_free_path(path);
9085                 err = -ENOMEM;
9086                 goto out;
9087         }
9088
9089         trans = btrfs_start_transaction(tree_root, 0);
9090         if (IS_ERR(trans)) {
9091                 err = PTR_ERR(trans);
9092                 goto out_free;
9093         }
9094
9095         if (block_rsv)
9096                 trans->block_rsv = block_rsv;
9097
9098         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9099                 level = btrfs_header_level(root->node);
9100                 path->nodes[level] = btrfs_lock_root_node(root);
9101                 btrfs_set_lock_blocking(path->nodes[level]);
9102                 path->slots[level] = 0;
9103                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9104                 memset(&wc->update_progress, 0,
9105                        sizeof(wc->update_progress));
9106         } else {
9107                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9108                 memcpy(&wc->update_progress, &key,
9109                        sizeof(wc->update_progress));
9110
9111                 level = root_item->drop_level;
9112                 BUG_ON(level == 0);
9113                 path->lowest_level = level;
9114                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9115                 path->lowest_level = 0;
9116                 if (ret < 0) {
9117                         err = ret;
9118                         goto out_end_trans;
9119                 }
9120                 WARN_ON(ret > 0);
9121
9122                 /*
9123                  * unlock our path, this is safe because only this
9124                  * function is allowed to delete this snapshot
9125                  */
9126                 btrfs_unlock_up_safe(path, 0);
9127
9128                 level = btrfs_header_level(root->node);
9129                 while (1) {
9130                         btrfs_tree_lock(path->nodes[level]);
9131                         btrfs_set_lock_blocking(path->nodes[level]);
9132                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9133
9134                         ret = btrfs_lookup_extent_info(trans, fs_info,
9135                                                 path->nodes[level]->start,
9136                                                 level, 1, &wc->refs[level],
9137                                                 &wc->flags[level]);
9138                         if (ret < 0) {
9139                                 err = ret;
9140                                 goto out_end_trans;
9141                         }
9142                         BUG_ON(wc->refs[level] == 0);
9143
9144                         if (level == root_item->drop_level)
9145                                 break;
9146
9147                         btrfs_tree_unlock(path->nodes[level]);
9148                         path->locks[level] = 0;
9149                         WARN_ON(wc->refs[level] != 1);
9150                         level--;
9151                 }
9152         }
9153
9154         wc->level = level;
9155         wc->shared_level = -1;
9156         wc->stage = DROP_REFERENCE;
9157         wc->update_ref = update_ref;
9158         wc->keep_locks = 0;
9159         wc->for_reloc = for_reloc;
9160         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9161
9162         while (1) {
9163
9164                 ret = walk_down_tree(trans, root, path, wc);
9165                 if (ret < 0) {
9166                         err = ret;
9167                         break;
9168                 }
9169
9170                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9171                 if (ret < 0) {
9172                         err = ret;
9173                         break;
9174                 }
9175
9176                 if (ret > 0) {
9177                         BUG_ON(wc->stage != DROP_REFERENCE);
9178                         break;
9179                 }
9180
9181                 if (wc->stage == DROP_REFERENCE) {
9182                         level = wc->level;
9183                         btrfs_node_key(path->nodes[level],
9184                                        &root_item->drop_progress,
9185                                        path->slots[level]);
9186                         root_item->drop_level = level;
9187                 }
9188
9189                 BUG_ON(wc->level == 0);
9190                 if (btrfs_should_end_transaction(trans) ||
9191                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9192                         ret = btrfs_update_root(trans, tree_root,
9193                                                 &root->root_key,
9194                                                 root_item);
9195                         if (ret) {
9196                                 btrfs_abort_transaction(trans, ret);
9197                                 err = ret;
9198                                 goto out_end_trans;
9199                         }
9200
9201                         btrfs_end_transaction_throttle(trans);
9202                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9203                                 btrfs_debug(fs_info,
9204                                             "drop snapshot early exit");
9205                                 err = -EAGAIN;
9206                                 goto out_free;
9207                         }
9208
9209                         trans = btrfs_start_transaction(tree_root, 0);
9210                         if (IS_ERR(trans)) {
9211                                 err = PTR_ERR(trans);
9212                                 goto out_free;
9213                         }
9214                         if (block_rsv)
9215                                 trans->block_rsv = block_rsv;
9216                 }
9217         }
9218         btrfs_release_path(path);
9219         if (err)
9220                 goto out_end_trans;
9221
9222         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9223         if (ret) {
9224                 btrfs_abort_transaction(trans, ret);
9225                 err = ret;
9226                 goto out_end_trans;
9227         }
9228
9229         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9230                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9231                                       NULL, NULL);
9232                 if (ret < 0) {
9233                         btrfs_abort_transaction(trans, ret);
9234                         err = ret;
9235                         goto out_end_trans;
9236                 } else if (ret > 0) {
9237                         /* if we fail to delete the orphan item this time
9238                          * around, it'll get picked up the next time.
9239                          *
9240                          * The most common failure here is just -ENOENT.
9241                          */
9242                         btrfs_del_orphan_item(trans, tree_root,
9243                                               root->root_key.objectid);
9244                 }
9245         }
9246
9247         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9248                 btrfs_add_dropped_root(trans, root);
9249         } else {
9250                 free_extent_buffer(root->node);
9251                 free_extent_buffer(root->commit_root);
9252                 btrfs_put_fs_root(root);
9253         }
9254         root_dropped = true;
9255 out_end_trans:
9256         btrfs_end_transaction_throttle(trans);
9257 out_free:
9258         kfree(wc);
9259         btrfs_free_path(path);
9260 out:
9261         /*
9262          * So if we need to stop dropping the snapshot for whatever reason we
9263          * need to make sure to add it back to the dead root list so that we
9264          * keep trying to do the work later.  This also cleans up roots if we
9265          * don't have it in the radix (like when we recover after a power fail
9266          * or unmount) so we don't leak memory.
9267          */
9268         if (!for_reloc && !root_dropped)
9269                 btrfs_add_dead_root(root);
9270         if (err && err != -EAGAIN)
9271                 btrfs_handle_fs_error(fs_info, err, NULL);
9272         return err;
9273 }
9274
9275 /*
9276  * drop subtree rooted at tree block 'node'.
9277  *
9278  * NOTE: this function will unlock and release tree block 'node'
9279  * only used by relocation code
9280  */
9281 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9282                         struct btrfs_root *root,
9283                         struct extent_buffer *node,
9284                         struct extent_buffer *parent)
9285 {
9286         struct btrfs_fs_info *fs_info = root->fs_info;
9287         struct btrfs_path *path;
9288         struct walk_control *wc;
9289         int level;
9290         int parent_level;
9291         int ret = 0;
9292         int wret;
9293
9294         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9295
9296         path = btrfs_alloc_path();
9297         if (!path)
9298                 return -ENOMEM;
9299
9300         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9301         if (!wc) {
9302                 btrfs_free_path(path);
9303                 return -ENOMEM;
9304         }
9305
9306         btrfs_assert_tree_locked(parent);
9307         parent_level = btrfs_header_level(parent);
9308         extent_buffer_get(parent);
9309         path->nodes[parent_level] = parent;
9310         path->slots[parent_level] = btrfs_header_nritems(parent);
9311
9312         btrfs_assert_tree_locked(node);
9313         level = btrfs_header_level(node);
9314         path->nodes[level] = node;
9315         path->slots[level] = 0;
9316         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9317
9318         wc->refs[parent_level] = 1;
9319         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9320         wc->level = level;
9321         wc->shared_level = -1;
9322         wc->stage = DROP_REFERENCE;
9323         wc->update_ref = 0;
9324         wc->keep_locks = 1;
9325         wc->for_reloc = 1;
9326         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9327
9328         while (1) {
9329                 wret = walk_down_tree(trans, root, path, wc);
9330                 if (wret < 0) {
9331                         ret = wret;
9332                         break;
9333                 }
9334
9335                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9336                 if (wret < 0)
9337                         ret = wret;
9338                 if (wret != 0)
9339                         break;
9340         }
9341
9342         kfree(wc);
9343         btrfs_free_path(path);
9344         return ret;
9345 }
9346
9347 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9348 {
9349         u64 num_devices;
9350         u64 stripped;
9351
9352         /*
9353          * if restripe for this chunk_type is on pick target profile and
9354          * return, otherwise do the usual balance
9355          */
9356         stripped = get_restripe_target(fs_info, flags);
9357         if (stripped)
9358                 return extended_to_chunk(stripped);
9359
9360         num_devices = fs_info->fs_devices->rw_devices;
9361
9362         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9363                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9364                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9365
9366         if (num_devices == 1) {
9367                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9368                 stripped = flags & ~stripped;
9369
9370                 /* turn raid0 into single device chunks */
9371                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9372                         return stripped;
9373
9374                 /* turn mirroring into duplication */
9375                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9376                              BTRFS_BLOCK_GROUP_RAID10))
9377                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9378         } else {
9379                 /* they already had raid on here, just return */
9380                 if (flags & stripped)
9381                         return flags;
9382
9383                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9384                 stripped = flags & ~stripped;
9385
9386                 /* switch duplicated blocks with raid1 */
9387                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9388                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9389
9390                 /* this is drive concat, leave it alone */
9391         }
9392
9393         return flags;
9394 }
9395
9396 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9397 {
9398         struct btrfs_space_info *sinfo = cache->space_info;
9399         u64 num_bytes;
9400         u64 min_allocable_bytes;
9401         int ret = -ENOSPC;
9402
9403         /*
9404          * We need some metadata space and system metadata space for
9405          * allocating chunks in some corner cases until we force to set
9406          * it to be readonly.
9407          */
9408         if ((sinfo->flags &
9409              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9410             !force)
9411                 min_allocable_bytes = SZ_1M;
9412         else
9413                 min_allocable_bytes = 0;
9414
9415         spin_lock(&sinfo->lock);
9416         spin_lock(&cache->lock);
9417
9418         if (cache->ro) {
9419                 cache->ro++;
9420                 ret = 0;
9421                 goto out;
9422         }
9423
9424         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9425                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9426
9427         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9428             min_allocable_bytes <= sinfo->total_bytes) {
9429                 sinfo->bytes_readonly += num_bytes;
9430                 cache->ro++;
9431                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9432                 ret = 0;
9433         }
9434 out:
9435         spin_unlock(&cache->lock);
9436         spin_unlock(&sinfo->lock);
9437         return ret;
9438 }
9439
9440 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9441                              struct btrfs_block_group_cache *cache)
9442
9443 {
9444         struct btrfs_trans_handle *trans;
9445         u64 alloc_flags;
9446         int ret;
9447
9448 again:
9449         trans = btrfs_join_transaction(fs_info->extent_root);
9450         if (IS_ERR(trans))
9451                 return PTR_ERR(trans);
9452
9453         /*
9454          * we're not allowed to set block groups readonly after the dirty
9455          * block groups cache has started writing.  If it already started,
9456          * back off and let this transaction commit
9457          */
9458         mutex_lock(&fs_info->ro_block_group_mutex);
9459         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9460                 u64 transid = trans->transid;
9461
9462                 mutex_unlock(&fs_info->ro_block_group_mutex);
9463                 btrfs_end_transaction(trans);
9464
9465                 ret = btrfs_wait_for_commit(fs_info, transid);
9466                 if (ret)
9467                         return ret;
9468                 goto again;
9469         }
9470
9471         /*
9472          * if we are changing raid levels, try to allocate a corresponding
9473          * block group with the new raid level.
9474          */
9475         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9476         if (alloc_flags != cache->flags) {
9477                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9478                                      CHUNK_ALLOC_FORCE);
9479                 /*
9480                  * ENOSPC is allowed here, we may have enough space
9481                  * already allocated at the new raid level to
9482                  * carry on
9483                  */
9484                 if (ret == -ENOSPC)
9485                         ret = 0;
9486                 if (ret < 0)
9487                         goto out;
9488         }
9489
9490         ret = inc_block_group_ro(cache, 0);
9491         if (!ret)
9492                 goto out;
9493         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9494         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9495                              CHUNK_ALLOC_FORCE);
9496         if (ret < 0)
9497                 goto out;
9498         ret = inc_block_group_ro(cache, 0);
9499 out:
9500         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9501                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9502                 mutex_lock(&fs_info->chunk_mutex);
9503                 check_system_chunk(trans, fs_info, alloc_flags);
9504                 mutex_unlock(&fs_info->chunk_mutex);
9505         }
9506         mutex_unlock(&fs_info->ro_block_group_mutex);
9507
9508         btrfs_end_transaction(trans);
9509         return ret;
9510 }
9511
9512 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9513                             struct btrfs_fs_info *fs_info, u64 type)
9514 {
9515         u64 alloc_flags = get_alloc_profile(fs_info, type);
9516
9517         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9518 }
9519
9520 /*
9521  * helper to account the unused space of all the readonly block group in the
9522  * space_info. takes mirrors into account.
9523  */
9524 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9525 {
9526         struct btrfs_block_group_cache *block_group;
9527         u64 free_bytes = 0;
9528         int factor;
9529
9530         /* It's df, we don't care if it's racy */
9531         if (list_empty(&sinfo->ro_bgs))
9532                 return 0;
9533
9534         spin_lock(&sinfo->lock);
9535         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9536                 spin_lock(&block_group->lock);
9537
9538                 if (!block_group->ro) {
9539                         spin_unlock(&block_group->lock);
9540                         continue;
9541                 }
9542
9543                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9544                                           BTRFS_BLOCK_GROUP_RAID10 |
9545                                           BTRFS_BLOCK_GROUP_DUP))
9546                         factor = 2;
9547                 else
9548                         factor = 1;
9549
9550                 free_bytes += (block_group->key.offset -
9551                                btrfs_block_group_used(&block_group->item)) *
9552                                factor;
9553
9554                 spin_unlock(&block_group->lock);
9555         }
9556         spin_unlock(&sinfo->lock);
9557
9558         return free_bytes;
9559 }
9560
9561 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9562 {
9563         struct btrfs_space_info *sinfo = cache->space_info;
9564         u64 num_bytes;
9565
9566         BUG_ON(!cache->ro);
9567
9568         spin_lock(&sinfo->lock);
9569         spin_lock(&cache->lock);
9570         if (!--cache->ro) {
9571                 num_bytes = cache->key.offset - cache->reserved -
9572                             cache->pinned - cache->bytes_super -
9573                             btrfs_block_group_used(&cache->item);
9574                 sinfo->bytes_readonly -= num_bytes;
9575                 list_del_init(&cache->ro_list);
9576         }
9577         spin_unlock(&cache->lock);
9578         spin_unlock(&sinfo->lock);
9579 }
9580
9581 /*
9582  * checks to see if its even possible to relocate this block group.
9583  *
9584  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9585  * ok to go ahead and try.
9586  */
9587 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9588 {
9589         struct btrfs_root *root = fs_info->extent_root;
9590         struct btrfs_block_group_cache *block_group;
9591         struct btrfs_space_info *space_info;
9592         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9593         struct btrfs_device *device;
9594         struct btrfs_trans_handle *trans;
9595         u64 min_free;
9596         u64 dev_min = 1;
9597         u64 dev_nr = 0;
9598         u64 target;
9599         int debug;
9600         int index;
9601         int full = 0;
9602         int ret = 0;
9603
9604         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9605
9606         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9607
9608         /* odd, couldn't find the block group, leave it alone */
9609         if (!block_group) {
9610                 if (debug)
9611                         btrfs_warn(fs_info,
9612                                    "can't find block group for bytenr %llu",
9613                                    bytenr);
9614                 return -1;
9615         }
9616
9617         min_free = btrfs_block_group_used(&block_group->item);
9618
9619         /* no bytes used, we're good */
9620         if (!min_free)
9621                 goto out;
9622
9623         space_info = block_group->space_info;
9624         spin_lock(&space_info->lock);
9625
9626         full = space_info->full;
9627
9628         /*
9629          * if this is the last block group we have in this space, we can't
9630          * relocate it unless we're able to allocate a new chunk below.
9631          *
9632          * Otherwise, we need to make sure we have room in the space to handle
9633          * all of the extents from this block group.  If we can, we're good
9634          */
9635         if ((space_info->total_bytes != block_group->key.offset) &&
9636             (btrfs_space_info_used(space_info, false) + min_free <
9637              space_info->total_bytes)) {
9638                 spin_unlock(&space_info->lock);
9639                 goto out;
9640         }
9641         spin_unlock(&space_info->lock);
9642
9643         /*
9644          * ok we don't have enough space, but maybe we have free space on our
9645          * devices to allocate new chunks for relocation, so loop through our
9646          * alloc devices and guess if we have enough space.  if this block
9647          * group is going to be restriped, run checks against the target
9648          * profile instead of the current one.
9649          */
9650         ret = -1;
9651
9652         /*
9653          * index:
9654          *      0: raid10
9655          *      1: raid1
9656          *      2: dup
9657          *      3: raid0
9658          *      4: single
9659          */
9660         target = get_restripe_target(fs_info, block_group->flags);
9661         if (target) {
9662                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9663         } else {
9664                 /*
9665                  * this is just a balance, so if we were marked as full
9666                  * we know there is no space for a new chunk
9667                  */
9668                 if (full) {
9669                         if (debug)
9670                                 btrfs_warn(fs_info,
9671                                            "no space to alloc new chunk for block group %llu",
9672                                            block_group->key.objectid);
9673                         goto out;
9674                 }
9675
9676                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9677         }
9678
9679         if (index == BTRFS_RAID_RAID10) {
9680                 dev_min = 4;
9681                 /* Divide by 2 */
9682                 min_free >>= 1;
9683         } else if (index == BTRFS_RAID_RAID1) {
9684                 dev_min = 2;
9685         } else if (index == BTRFS_RAID_DUP) {
9686                 /* Multiply by 2 */
9687                 min_free <<= 1;
9688         } else if (index == BTRFS_RAID_RAID0) {
9689                 dev_min = fs_devices->rw_devices;
9690                 min_free = div64_u64(min_free, dev_min);
9691         }
9692
9693         /* We need to do this so that we can look at pending chunks */
9694         trans = btrfs_join_transaction(root);
9695         if (IS_ERR(trans)) {
9696                 ret = PTR_ERR(trans);
9697                 goto out;
9698         }
9699
9700         mutex_lock(&fs_info->chunk_mutex);
9701         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9702                 u64 dev_offset;
9703
9704                 /*
9705                  * check to make sure we can actually find a chunk with enough
9706                  * space to fit our block group in.
9707                  */
9708                 if (device->total_bytes > device->bytes_used + min_free &&
9709                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9710                         ret = find_free_dev_extent(trans, device, min_free,
9711                                                    &dev_offset, NULL);
9712                         if (!ret)
9713                                 dev_nr++;
9714
9715                         if (dev_nr >= dev_min)
9716                                 break;
9717
9718                         ret = -1;
9719                 }
9720         }
9721         if (debug && ret == -1)
9722                 btrfs_warn(fs_info,
9723                            "no space to allocate a new chunk for block group %llu",
9724                            block_group->key.objectid);
9725         mutex_unlock(&fs_info->chunk_mutex);
9726         btrfs_end_transaction(trans);
9727 out:
9728         btrfs_put_block_group(block_group);
9729         return ret;
9730 }
9731
9732 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9733                                   struct btrfs_path *path,
9734                                   struct btrfs_key *key)
9735 {
9736         struct btrfs_root *root = fs_info->extent_root;
9737         int ret = 0;
9738         struct btrfs_key found_key;
9739         struct extent_buffer *leaf;
9740         int slot;
9741
9742         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9743         if (ret < 0)
9744                 goto out;
9745
9746         while (1) {
9747                 slot = path->slots[0];
9748                 leaf = path->nodes[0];
9749                 if (slot >= btrfs_header_nritems(leaf)) {
9750                         ret = btrfs_next_leaf(root, path);
9751                         if (ret == 0)
9752                                 continue;
9753                         if (ret < 0)
9754                                 goto out;
9755                         break;
9756                 }
9757                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9758
9759                 if (found_key.objectid >= key->objectid &&
9760                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9761                         struct extent_map_tree *em_tree;
9762                         struct extent_map *em;
9763
9764                         em_tree = &root->fs_info->mapping_tree.map_tree;
9765                         read_lock(&em_tree->lock);
9766                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9767                                                    found_key.offset);
9768                         read_unlock(&em_tree->lock);
9769                         if (!em) {
9770                                 btrfs_err(fs_info,
9771                         "logical %llu len %llu found bg but no related chunk",
9772                                           found_key.objectid, found_key.offset);
9773                                 ret = -ENOENT;
9774                         } else {
9775                                 ret = 0;
9776                         }
9777                         free_extent_map(em);
9778                         goto out;
9779                 }
9780                 path->slots[0]++;
9781         }
9782 out:
9783         return ret;
9784 }
9785
9786 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9787 {
9788         struct btrfs_block_group_cache *block_group;
9789         u64 last = 0;
9790
9791         while (1) {
9792                 struct inode *inode;
9793
9794                 block_group = btrfs_lookup_first_block_group(info, last);
9795                 while (block_group) {
9796                         spin_lock(&block_group->lock);
9797                         if (block_group->iref)
9798                                 break;
9799                         spin_unlock(&block_group->lock);
9800                         block_group = next_block_group(info, block_group);
9801                 }
9802                 if (!block_group) {
9803                         if (last == 0)
9804                                 break;
9805                         last = 0;
9806                         continue;
9807                 }
9808
9809                 inode = block_group->inode;
9810                 block_group->iref = 0;
9811                 block_group->inode = NULL;
9812                 spin_unlock(&block_group->lock);
9813                 ASSERT(block_group->io_ctl.inode == NULL);
9814                 iput(inode);
9815                 last = block_group->key.objectid + block_group->key.offset;
9816                 btrfs_put_block_group(block_group);
9817         }
9818 }
9819
9820 /*
9821  * Must be called only after stopping all workers, since we could have block
9822  * group caching kthreads running, and therefore they could race with us if we
9823  * freed the block groups before stopping them.
9824  */
9825 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9826 {
9827         struct btrfs_block_group_cache *block_group;
9828         struct btrfs_space_info *space_info;
9829         struct btrfs_caching_control *caching_ctl;
9830         struct rb_node *n;
9831
9832         down_write(&info->commit_root_sem);
9833         while (!list_empty(&info->caching_block_groups)) {
9834                 caching_ctl = list_entry(info->caching_block_groups.next,
9835                                          struct btrfs_caching_control, list);
9836                 list_del(&caching_ctl->list);
9837                 put_caching_control(caching_ctl);
9838         }
9839         up_write(&info->commit_root_sem);
9840
9841         spin_lock(&info->unused_bgs_lock);
9842         while (!list_empty(&info->unused_bgs)) {
9843                 block_group = list_first_entry(&info->unused_bgs,
9844                                                struct btrfs_block_group_cache,
9845                                                bg_list);
9846                 list_del_init(&block_group->bg_list);
9847                 btrfs_put_block_group(block_group);
9848         }
9849         spin_unlock(&info->unused_bgs_lock);
9850
9851         spin_lock(&info->block_group_cache_lock);
9852         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9853                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9854                                        cache_node);
9855                 rb_erase(&block_group->cache_node,
9856                          &info->block_group_cache_tree);
9857                 RB_CLEAR_NODE(&block_group->cache_node);
9858                 spin_unlock(&info->block_group_cache_lock);
9859
9860                 down_write(&block_group->space_info->groups_sem);
9861                 list_del(&block_group->list);
9862                 up_write(&block_group->space_info->groups_sem);
9863
9864                 /*
9865                  * We haven't cached this block group, which means we could
9866                  * possibly have excluded extents on this block group.
9867                  */
9868                 if (block_group->cached == BTRFS_CACHE_NO ||
9869                     block_group->cached == BTRFS_CACHE_ERROR)
9870                         free_excluded_extents(info, block_group);
9871
9872                 btrfs_remove_free_space_cache(block_group);
9873                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9874                 ASSERT(list_empty(&block_group->dirty_list));
9875                 ASSERT(list_empty(&block_group->io_list));
9876                 ASSERT(list_empty(&block_group->bg_list));
9877                 ASSERT(atomic_read(&block_group->count) == 1);
9878                 btrfs_put_block_group(block_group);
9879
9880                 spin_lock(&info->block_group_cache_lock);
9881         }
9882         spin_unlock(&info->block_group_cache_lock);
9883
9884         /* now that all the block groups are freed, go through and
9885          * free all the space_info structs.  This is only called during
9886          * the final stages of unmount, and so we know nobody is
9887          * using them.  We call synchronize_rcu() once before we start,
9888          * just to be on the safe side.
9889          */
9890         synchronize_rcu();
9891
9892         release_global_block_rsv(info);
9893
9894         while (!list_empty(&info->space_info)) {
9895                 int i;
9896
9897                 space_info = list_entry(info->space_info.next,
9898                                         struct btrfs_space_info,
9899                                         list);
9900
9901                 /*
9902                  * Do not hide this behind enospc_debug, this is actually
9903                  * important and indicates a real bug if this happens.
9904                  */
9905                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9906                             space_info->bytes_reserved > 0 ||
9907                             space_info->bytes_may_use > 0))
9908                         dump_space_info(info, space_info, 0, 0);
9909                 list_del(&space_info->list);
9910                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9911                         struct kobject *kobj;
9912                         kobj = space_info->block_group_kobjs[i];
9913                         space_info->block_group_kobjs[i] = NULL;
9914                         if (kobj) {
9915                                 kobject_del(kobj);
9916                                 kobject_put(kobj);
9917                         }
9918                 }
9919                 kobject_del(&space_info->kobj);
9920                 kobject_put(&space_info->kobj);
9921         }
9922         return 0;
9923 }
9924
9925 static void link_block_group(struct btrfs_block_group_cache *cache)
9926 {
9927         struct btrfs_space_info *space_info = cache->space_info;
9928         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9929         bool first = false;
9930
9931         down_write(&space_info->groups_sem);
9932         if (list_empty(&space_info->block_groups[index]))
9933                 first = true;
9934         list_add_tail(&cache->list, &space_info->block_groups[index]);
9935         up_write(&space_info->groups_sem);
9936
9937         if (first) {
9938                 struct raid_kobject *rkobj;
9939                 int ret;
9940
9941                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9942                 if (!rkobj)
9943                         goto out_err;
9944                 rkobj->raid_type = index;
9945                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9946                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9947                                   "%s", get_raid_name(index));
9948                 if (ret) {
9949                         kobject_put(&rkobj->kobj);
9950                         goto out_err;
9951                 }
9952                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9953         }
9954
9955         return;
9956 out_err:
9957         btrfs_warn(cache->fs_info,
9958                    "failed to add kobject for block cache, ignoring");
9959 }
9960
9961 static struct btrfs_block_group_cache *
9962 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9963                                u64 start, u64 size)
9964 {
9965         struct btrfs_block_group_cache *cache;
9966
9967         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9968         if (!cache)
9969                 return NULL;
9970
9971         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9972                                         GFP_NOFS);
9973         if (!cache->free_space_ctl) {
9974                 kfree(cache);
9975                 return NULL;
9976         }
9977
9978         cache->key.objectid = start;
9979         cache->key.offset = size;
9980         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9981
9982         cache->fs_info = fs_info;
9983         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9984         set_free_space_tree_thresholds(cache);
9985
9986         atomic_set(&cache->count, 1);
9987         spin_lock_init(&cache->lock);
9988         init_rwsem(&cache->data_rwsem);
9989         INIT_LIST_HEAD(&cache->list);
9990         INIT_LIST_HEAD(&cache->cluster_list);
9991         INIT_LIST_HEAD(&cache->bg_list);
9992         INIT_LIST_HEAD(&cache->ro_list);
9993         INIT_LIST_HEAD(&cache->dirty_list);
9994         INIT_LIST_HEAD(&cache->io_list);
9995         btrfs_init_free_space_ctl(cache);
9996         atomic_set(&cache->trimming, 0);
9997         mutex_init(&cache->free_space_lock);
9998         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9999
10000         return cache;
10001 }
10002
10003 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10004 {
10005         struct btrfs_path *path;
10006         int ret;
10007         struct btrfs_block_group_cache *cache;
10008         struct btrfs_space_info *space_info;
10009         struct btrfs_key key;
10010         struct btrfs_key found_key;
10011         struct extent_buffer *leaf;
10012         int need_clear = 0;
10013         u64 cache_gen;
10014         u64 feature;
10015         int mixed;
10016
10017         feature = btrfs_super_incompat_flags(info->super_copy);
10018         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10019
10020         key.objectid = 0;
10021         key.offset = 0;
10022         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10023         path = btrfs_alloc_path();
10024         if (!path)
10025                 return -ENOMEM;
10026         path->reada = READA_FORWARD;
10027
10028         cache_gen = btrfs_super_cache_generation(info->super_copy);
10029         if (btrfs_test_opt(info, SPACE_CACHE) &&
10030             btrfs_super_generation(info->super_copy) != cache_gen)
10031                 need_clear = 1;
10032         if (btrfs_test_opt(info, CLEAR_CACHE))
10033                 need_clear = 1;
10034
10035         while (1) {
10036                 ret = find_first_block_group(info, path, &key);
10037                 if (ret > 0)
10038                         break;
10039                 if (ret != 0)
10040                         goto error;
10041
10042                 leaf = path->nodes[0];
10043                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10044
10045                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10046                                                        found_key.offset);
10047                 if (!cache) {
10048                         ret = -ENOMEM;
10049                         goto error;
10050                 }
10051
10052                 if (need_clear) {
10053                         /*
10054                          * When we mount with old space cache, we need to
10055                          * set BTRFS_DC_CLEAR and set dirty flag.
10056                          *
10057                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10058                          *    truncate the old free space cache inode and
10059                          *    setup a new one.
10060                          * b) Setting 'dirty flag' makes sure that we flush
10061                          *    the new space cache info onto disk.
10062                          */
10063                         if (btrfs_test_opt(info, SPACE_CACHE))
10064                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10065                 }
10066
10067                 read_extent_buffer(leaf, &cache->item,
10068                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10069                                    sizeof(cache->item));
10070                 cache->flags = btrfs_block_group_flags(&cache->item);
10071                 if (!mixed &&
10072                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10073                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10074                         btrfs_err(info,
10075 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10076                                   cache->key.objectid);
10077                         ret = -EINVAL;
10078                         goto error;
10079                 }
10080
10081                 key.objectid = found_key.objectid + found_key.offset;
10082                 btrfs_release_path(path);
10083
10084                 /*
10085                  * We need to exclude the super stripes now so that the space
10086                  * info has super bytes accounted for, otherwise we'll think
10087                  * we have more space than we actually do.
10088                  */
10089                 ret = exclude_super_stripes(info, cache);
10090                 if (ret) {
10091                         /*
10092                          * We may have excluded something, so call this just in
10093                          * case.
10094                          */
10095                         free_excluded_extents(info, cache);
10096                         btrfs_put_block_group(cache);
10097                         goto error;
10098                 }
10099
10100                 /*
10101                  * check for two cases, either we are full, and therefore
10102                  * don't need to bother with the caching work since we won't
10103                  * find any space, or we are empty, and we can just add all
10104                  * the space in and be done with it.  This saves us _alot_ of
10105                  * time, particularly in the full case.
10106                  */
10107                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10108                         cache->last_byte_to_unpin = (u64)-1;
10109                         cache->cached = BTRFS_CACHE_FINISHED;
10110                         free_excluded_extents(info, cache);
10111                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10112                         cache->last_byte_to_unpin = (u64)-1;
10113                         cache->cached = BTRFS_CACHE_FINISHED;
10114                         add_new_free_space(cache, info,
10115                                            found_key.objectid,
10116                                            found_key.objectid +
10117                                            found_key.offset);
10118                         free_excluded_extents(info, cache);
10119                 }
10120
10121                 ret = btrfs_add_block_group_cache(info, cache);
10122                 if (ret) {
10123                         btrfs_remove_free_space_cache(cache);
10124                         btrfs_put_block_group(cache);
10125                         goto error;
10126                 }
10127
10128                 trace_btrfs_add_block_group(info, cache, 0);
10129                 update_space_info(info, cache->flags, found_key.offset,
10130                                   btrfs_block_group_used(&cache->item),
10131                                   cache->bytes_super, &space_info);
10132
10133                 cache->space_info = space_info;
10134
10135                 link_block_group(cache);
10136
10137                 set_avail_alloc_bits(info, cache->flags);
10138                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10139                         inc_block_group_ro(cache, 1);
10140                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10141                         spin_lock(&info->unused_bgs_lock);
10142                         /* Should always be true but just in case. */
10143                         if (list_empty(&cache->bg_list)) {
10144                                 btrfs_get_block_group(cache);
10145                                 list_add_tail(&cache->bg_list,
10146                                               &info->unused_bgs);
10147                         }
10148                         spin_unlock(&info->unused_bgs_lock);
10149                 }
10150         }
10151
10152         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10153                 if (!(get_alloc_profile(info, space_info->flags) &
10154                       (BTRFS_BLOCK_GROUP_RAID10 |
10155                        BTRFS_BLOCK_GROUP_RAID1 |
10156                        BTRFS_BLOCK_GROUP_RAID5 |
10157                        BTRFS_BLOCK_GROUP_RAID6 |
10158                        BTRFS_BLOCK_GROUP_DUP)))
10159                         continue;
10160                 /*
10161                  * avoid allocating from un-mirrored block group if there are
10162                  * mirrored block groups.
10163                  */
10164                 list_for_each_entry(cache,
10165                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10166                                 list)
10167                         inc_block_group_ro(cache, 1);
10168                 list_for_each_entry(cache,
10169                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10170                                 list)
10171                         inc_block_group_ro(cache, 1);
10172         }
10173
10174         init_global_block_rsv(info);
10175         ret = 0;
10176 error:
10177         btrfs_free_path(path);
10178         return ret;
10179 }
10180
10181 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10182 {
10183         struct btrfs_fs_info *fs_info = trans->fs_info;
10184         struct btrfs_block_group_cache *block_group, *tmp;
10185         struct btrfs_root *extent_root = fs_info->extent_root;
10186         struct btrfs_block_group_item item;
10187         struct btrfs_key key;
10188         int ret = 0;
10189         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10190
10191         trans->can_flush_pending_bgs = false;
10192         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10193                 if (ret)
10194                         goto next;
10195
10196                 spin_lock(&block_group->lock);
10197                 memcpy(&item, &block_group->item, sizeof(item));
10198                 memcpy(&key, &block_group->key, sizeof(key));
10199                 spin_unlock(&block_group->lock);
10200
10201                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10202                                         sizeof(item));
10203                 if (ret)
10204                         btrfs_abort_transaction(trans, ret);
10205                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10206                                                key.offset);
10207                 if (ret)
10208                         btrfs_abort_transaction(trans, ret);
10209                 add_block_group_free_space(trans, fs_info, block_group);
10210                 /* already aborted the transaction if it failed. */
10211 next:
10212                 list_del_init(&block_group->bg_list);
10213         }
10214         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10215 }
10216
10217 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10218                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10219                            u64 type, u64 chunk_offset, u64 size)
10220 {
10221         struct btrfs_block_group_cache *cache;
10222         int ret;
10223
10224         btrfs_set_log_full_commit(fs_info, trans);
10225
10226         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10227         if (!cache)
10228                 return -ENOMEM;
10229
10230         btrfs_set_block_group_used(&cache->item, bytes_used);
10231         btrfs_set_block_group_chunk_objectid(&cache->item,
10232                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10233         btrfs_set_block_group_flags(&cache->item, type);
10234
10235         cache->flags = type;
10236         cache->last_byte_to_unpin = (u64)-1;
10237         cache->cached = BTRFS_CACHE_FINISHED;
10238         cache->needs_free_space = 1;
10239         ret = exclude_super_stripes(fs_info, cache);
10240         if (ret) {
10241                 /*
10242                  * We may have excluded something, so call this just in
10243                  * case.
10244                  */
10245                 free_excluded_extents(fs_info, cache);
10246                 btrfs_put_block_group(cache);
10247                 return ret;
10248         }
10249
10250         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10251
10252         free_excluded_extents(fs_info, cache);
10253
10254 #ifdef CONFIG_BTRFS_DEBUG
10255         if (btrfs_should_fragment_free_space(cache)) {
10256                 u64 new_bytes_used = size - bytes_used;
10257
10258                 bytes_used += new_bytes_used >> 1;
10259                 fragment_free_space(cache);
10260         }
10261 #endif
10262         /*
10263          * Ensure the corresponding space_info object is created and
10264          * assigned to our block group. We want our bg to be added to the rbtree
10265          * with its ->space_info set.
10266          */
10267         cache->space_info = __find_space_info(fs_info, cache->flags);
10268         if (!cache->space_info) {
10269                 ret = create_space_info(fs_info, cache->flags,
10270                                        &cache->space_info);
10271                 if (ret) {
10272                         btrfs_remove_free_space_cache(cache);
10273                         btrfs_put_block_group(cache);
10274                         return ret;
10275                 }
10276         }
10277
10278         ret = btrfs_add_block_group_cache(fs_info, cache);
10279         if (ret) {
10280                 btrfs_remove_free_space_cache(cache);
10281                 btrfs_put_block_group(cache);
10282                 return ret;
10283         }
10284
10285         /*
10286          * Now that our block group has its ->space_info set and is inserted in
10287          * the rbtree, update the space info's counters.
10288          */
10289         trace_btrfs_add_block_group(fs_info, cache, 1);
10290         update_space_info(fs_info, cache->flags, size, bytes_used,
10291                                 cache->bytes_super, &cache->space_info);
10292         update_global_block_rsv(fs_info);
10293
10294         link_block_group(cache);
10295
10296         list_add_tail(&cache->bg_list, &trans->new_bgs);
10297
10298         set_avail_alloc_bits(fs_info, type);
10299         return 0;
10300 }
10301
10302 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10303 {
10304         u64 extra_flags = chunk_to_extended(flags) &
10305                                 BTRFS_EXTENDED_PROFILE_MASK;
10306
10307         write_seqlock(&fs_info->profiles_lock);
10308         if (flags & BTRFS_BLOCK_GROUP_DATA)
10309                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10310         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10311                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10312         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10313                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10314         write_sequnlock(&fs_info->profiles_lock);
10315 }
10316
10317 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10318                              struct btrfs_fs_info *fs_info, u64 group_start,
10319                              struct extent_map *em)
10320 {
10321         struct btrfs_root *root = fs_info->extent_root;
10322         struct btrfs_path *path;
10323         struct btrfs_block_group_cache *block_group;
10324         struct btrfs_free_cluster *cluster;
10325         struct btrfs_root *tree_root = fs_info->tree_root;
10326         struct btrfs_key key;
10327         struct inode *inode;
10328         struct kobject *kobj = NULL;
10329         int ret;
10330         int index;
10331         int factor;
10332         struct btrfs_caching_control *caching_ctl = NULL;
10333         bool remove_em;
10334
10335         block_group = btrfs_lookup_block_group(fs_info, group_start);
10336         BUG_ON(!block_group);
10337         BUG_ON(!block_group->ro);
10338
10339         /*
10340          * Free the reserved super bytes from this block group before
10341          * remove it.
10342          */
10343         free_excluded_extents(fs_info, block_group);
10344         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10345                                   block_group->key.offset);
10346
10347         memcpy(&key, &block_group->key, sizeof(key));
10348         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10349         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10350                                   BTRFS_BLOCK_GROUP_RAID1 |
10351                                   BTRFS_BLOCK_GROUP_RAID10))
10352                 factor = 2;
10353         else
10354                 factor = 1;
10355
10356         /* make sure this block group isn't part of an allocation cluster */
10357         cluster = &fs_info->data_alloc_cluster;
10358         spin_lock(&cluster->refill_lock);
10359         btrfs_return_cluster_to_free_space(block_group, cluster);
10360         spin_unlock(&cluster->refill_lock);
10361
10362         /*
10363          * make sure this block group isn't part of a metadata
10364          * allocation cluster
10365          */
10366         cluster = &fs_info->meta_alloc_cluster;
10367         spin_lock(&cluster->refill_lock);
10368         btrfs_return_cluster_to_free_space(block_group, cluster);
10369         spin_unlock(&cluster->refill_lock);
10370
10371         path = btrfs_alloc_path();
10372         if (!path) {
10373                 ret = -ENOMEM;
10374                 goto out;
10375         }
10376
10377         /*
10378          * get the inode first so any iput calls done for the io_list
10379          * aren't the final iput (no unlinks allowed now)
10380          */
10381         inode = lookup_free_space_inode(fs_info, block_group, path);
10382
10383         mutex_lock(&trans->transaction->cache_write_mutex);
10384         /*
10385          * make sure our free spache cache IO is done before remove the
10386          * free space inode
10387          */
10388         spin_lock(&trans->transaction->dirty_bgs_lock);
10389         if (!list_empty(&block_group->io_list)) {
10390                 list_del_init(&block_group->io_list);
10391
10392                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10393
10394                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10395                 btrfs_wait_cache_io(trans, block_group, path);
10396                 btrfs_put_block_group(block_group);
10397                 spin_lock(&trans->transaction->dirty_bgs_lock);
10398         }
10399
10400         if (!list_empty(&block_group->dirty_list)) {
10401                 list_del_init(&block_group->dirty_list);
10402                 btrfs_put_block_group(block_group);
10403         }
10404         spin_unlock(&trans->transaction->dirty_bgs_lock);
10405         mutex_unlock(&trans->transaction->cache_write_mutex);
10406
10407         if (!IS_ERR(inode)) {
10408                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10409                 if (ret) {
10410                         btrfs_add_delayed_iput(inode);
10411                         goto out;
10412                 }
10413                 clear_nlink(inode);
10414                 /* One for the block groups ref */
10415                 spin_lock(&block_group->lock);
10416                 if (block_group->iref) {
10417                         block_group->iref = 0;
10418                         block_group->inode = NULL;
10419                         spin_unlock(&block_group->lock);
10420                         iput(inode);
10421                 } else {
10422                         spin_unlock(&block_group->lock);
10423                 }
10424                 /* One for our lookup ref */
10425                 btrfs_add_delayed_iput(inode);
10426         }
10427
10428         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10429         key.offset = block_group->key.objectid;
10430         key.type = 0;
10431
10432         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10433         if (ret < 0)
10434                 goto out;
10435         if (ret > 0)
10436                 btrfs_release_path(path);
10437         if (ret == 0) {
10438                 ret = btrfs_del_item(trans, tree_root, path);
10439                 if (ret)
10440                         goto out;
10441                 btrfs_release_path(path);
10442         }
10443
10444         spin_lock(&fs_info->block_group_cache_lock);
10445         rb_erase(&block_group->cache_node,
10446                  &fs_info->block_group_cache_tree);
10447         RB_CLEAR_NODE(&block_group->cache_node);
10448
10449         if (fs_info->first_logical_byte == block_group->key.objectid)
10450                 fs_info->first_logical_byte = (u64)-1;
10451         spin_unlock(&fs_info->block_group_cache_lock);
10452
10453         down_write(&block_group->space_info->groups_sem);
10454         /*
10455          * we must use list_del_init so people can check to see if they
10456          * are still on the list after taking the semaphore
10457          */
10458         list_del_init(&block_group->list);
10459         if (list_empty(&block_group->space_info->block_groups[index])) {
10460                 kobj = block_group->space_info->block_group_kobjs[index];
10461                 block_group->space_info->block_group_kobjs[index] = NULL;
10462                 clear_avail_alloc_bits(fs_info, block_group->flags);
10463         }
10464         up_write(&block_group->space_info->groups_sem);
10465         if (kobj) {
10466                 kobject_del(kobj);
10467                 kobject_put(kobj);
10468         }
10469
10470         if (block_group->has_caching_ctl)
10471                 caching_ctl = get_caching_control(block_group);
10472         if (block_group->cached == BTRFS_CACHE_STARTED)
10473                 wait_block_group_cache_done(block_group);
10474         if (block_group->has_caching_ctl) {
10475                 down_write(&fs_info->commit_root_sem);
10476                 if (!caching_ctl) {
10477                         struct btrfs_caching_control *ctl;
10478
10479                         list_for_each_entry(ctl,
10480                                     &fs_info->caching_block_groups, list)
10481                                 if (ctl->block_group == block_group) {
10482                                         caching_ctl = ctl;
10483                                         refcount_inc(&caching_ctl->count);
10484                                         break;
10485                                 }
10486                 }
10487                 if (caching_ctl)
10488                         list_del_init(&caching_ctl->list);
10489                 up_write(&fs_info->commit_root_sem);
10490                 if (caching_ctl) {
10491                         /* Once for the caching bgs list and once for us. */
10492                         put_caching_control(caching_ctl);
10493                         put_caching_control(caching_ctl);
10494                 }
10495         }
10496
10497         spin_lock(&trans->transaction->dirty_bgs_lock);
10498         if (!list_empty(&block_group->dirty_list)) {
10499                 WARN_ON(1);
10500         }
10501         if (!list_empty(&block_group->io_list)) {
10502                 WARN_ON(1);
10503         }
10504         spin_unlock(&trans->transaction->dirty_bgs_lock);
10505         btrfs_remove_free_space_cache(block_group);
10506
10507         spin_lock(&block_group->space_info->lock);
10508         list_del_init(&block_group->ro_list);
10509
10510         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10511                 WARN_ON(block_group->space_info->total_bytes
10512                         < block_group->key.offset);
10513                 WARN_ON(block_group->space_info->bytes_readonly
10514                         < block_group->key.offset);
10515                 WARN_ON(block_group->space_info->disk_total
10516                         < block_group->key.offset * factor);
10517         }
10518         block_group->space_info->total_bytes -= block_group->key.offset;
10519         block_group->space_info->bytes_readonly -= block_group->key.offset;
10520         block_group->space_info->disk_total -= block_group->key.offset * factor;
10521
10522         spin_unlock(&block_group->space_info->lock);
10523
10524         memcpy(&key, &block_group->key, sizeof(key));
10525
10526         mutex_lock(&fs_info->chunk_mutex);
10527         if (!list_empty(&em->list)) {
10528                 /* We're in the transaction->pending_chunks list. */
10529                 free_extent_map(em);
10530         }
10531         spin_lock(&block_group->lock);
10532         block_group->removed = 1;
10533         /*
10534          * At this point trimming can't start on this block group, because we
10535          * removed the block group from the tree fs_info->block_group_cache_tree
10536          * so no one can't find it anymore and even if someone already got this
10537          * block group before we removed it from the rbtree, they have already
10538          * incremented block_group->trimming - if they didn't, they won't find
10539          * any free space entries because we already removed them all when we
10540          * called btrfs_remove_free_space_cache().
10541          *
10542          * And we must not remove the extent map from the fs_info->mapping_tree
10543          * to prevent the same logical address range and physical device space
10544          * ranges from being reused for a new block group. This is because our
10545          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10546          * completely transactionless, so while it is trimming a range the
10547          * currently running transaction might finish and a new one start,
10548          * allowing for new block groups to be created that can reuse the same
10549          * physical device locations unless we take this special care.
10550          *
10551          * There may also be an implicit trim operation if the file system
10552          * is mounted with -odiscard. The same protections must remain
10553          * in place until the extents have been discarded completely when
10554          * the transaction commit has completed.
10555          */
10556         remove_em = (atomic_read(&block_group->trimming) == 0);
10557         /*
10558          * Make sure a trimmer task always sees the em in the pinned_chunks list
10559          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10560          * before checking block_group->removed).
10561          */
10562         if (!remove_em) {
10563                 /*
10564                  * Our em might be in trans->transaction->pending_chunks which
10565                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10566                  * and so is the fs_info->pinned_chunks list.
10567                  *
10568                  * So at this point we must be holding the chunk_mutex to avoid
10569                  * any races with chunk allocation (more specifically at
10570                  * volumes.c:contains_pending_extent()), to ensure it always
10571                  * sees the em, either in the pending_chunks list or in the
10572                  * pinned_chunks list.
10573                  */
10574                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10575         }
10576         spin_unlock(&block_group->lock);
10577
10578         if (remove_em) {
10579                 struct extent_map_tree *em_tree;
10580
10581                 em_tree = &fs_info->mapping_tree.map_tree;
10582                 write_lock(&em_tree->lock);
10583                 /*
10584                  * The em might be in the pending_chunks list, so make sure the
10585                  * chunk mutex is locked, since remove_extent_mapping() will
10586                  * delete us from that list.
10587                  */
10588                 remove_extent_mapping(em_tree, em);
10589                 write_unlock(&em_tree->lock);
10590                 /* once for the tree */
10591                 free_extent_map(em);
10592         }
10593
10594         mutex_unlock(&fs_info->chunk_mutex);
10595
10596         ret = remove_block_group_free_space(trans, fs_info, block_group);
10597         if (ret)
10598                 goto out;
10599
10600         btrfs_put_block_group(block_group);
10601         btrfs_put_block_group(block_group);
10602
10603         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10604         if (ret > 0)
10605                 ret = -EIO;
10606         if (ret < 0)
10607                 goto out;
10608
10609         ret = btrfs_del_item(trans, root, path);
10610 out:
10611         btrfs_free_path(path);
10612         return ret;
10613 }
10614
10615 struct btrfs_trans_handle *
10616 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10617                                      const u64 chunk_offset)
10618 {
10619         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10620         struct extent_map *em;
10621         struct map_lookup *map;
10622         unsigned int num_items;
10623
10624         read_lock(&em_tree->lock);
10625         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10626         read_unlock(&em_tree->lock);
10627         ASSERT(em && em->start == chunk_offset);
10628
10629         /*
10630          * We need to reserve 3 + N units from the metadata space info in order
10631          * to remove a block group (done at btrfs_remove_chunk() and at
10632          * btrfs_remove_block_group()), which are used for:
10633          *
10634          * 1 unit for adding the free space inode's orphan (located in the tree
10635          * of tree roots).
10636          * 1 unit for deleting the block group item (located in the extent
10637          * tree).
10638          * 1 unit for deleting the free space item (located in tree of tree
10639          * roots).
10640          * N units for deleting N device extent items corresponding to each
10641          * stripe (located in the device tree).
10642          *
10643          * In order to remove a block group we also need to reserve units in the
10644          * system space info in order to update the chunk tree (update one or
10645          * more device items and remove one chunk item), but this is done at
10646          * btrfs_remove_chunk() through a call to check_system_chunk().
10647          */
10648         map = em->map_lookup;
10649         num_items = 3 + map->num_stripes;
10650         free_extent_map(em);
10651
10652         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10653                                                            num_items, 1);
10654 }
10655
10656 /*
10657  * Process the unused_bgs list and remove any that don't have any allocated
10658  * space inside of them.
10659  */
10660 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10661 {
10662         struct btrfs_block_group_cache *block_group;
10663         struct btrfs_space_info *space_info;
10664         struct btrfs_trans_handle *trans;
10665         int ret = 0;
10666
10667         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10668                 return;
10669
10670         spin_lock(&fs_info->unused_bgs_lock);
10671         while (!list_empty(&fs_info->unused_bgs)) {
10672                 u64 start, end;
10673                 int trimming;
10674
10675                 block_group = list_first_entry(&fs_info->unused_bgs,
10676                                                struct btrfs_block_group_cache,
10677                                                bg_list);
10678                 list_del_init(&block_group->bg_list);
10679
10680                 space_info = block_group->space_info;
10681
10682                 if (ret || btrfs_mixed_space_info(space_info)) {
10683                         btrfs_put_block_group(block_group);
10684                         continue;
10685                 }
10686                 spin_unlock(&fs_info->unused_bgs_lock);
10687
10688                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10689
10690                 /* Don't want to race with allocators so take the groups_sem */
10691                 down_write(&space_info->groups_sem);
10692                 spin_lock(&block_group->lock);
10693                 if (block_group->reserved ||
10694                     btrfs_block_group_used(&block_group->item) ||
10695                     block_group->ro ||
10696                     list_is_singular(&block_group->list)) {
10697                         /*
10698                          * We want to bail if we made new allocations or have
10699                          * outstanding allocations in this block group.  We do
10700                          * the ro check in case balance is currently acting on
10701                          * this block group.
10702                          */
10703                         spin_unlock(&block_group->lock);
10704                         up_write(&space_info->groups_sem);
10705                         goto next;
10706                 }
10707                 spin_unlock(&block_group->lock);
10708
10709                 /* We don't want to force the issue, only flip if it's ok. */
10710                 ret = inc_block_group_ro(block_group, 0);
10711                 up_write(&space_info->groups_sem);
10712                 if (ret < 0) {
10713                         ret = 0;
10714                         goto next;
10715                 }
10716
10717                 /*
10718                  * Want to do this before we do anything else so we can recover
10719                  * properly if we fail to join the transaction.
10720                  */
10721                 trans = btrfs_start_trans_remove_block_group(fs_info,
10722                                                      block_group->key.objectid);
10723                 if (IS_ERR(trans)) {
10724                         btrfs_dec_block_group_ro(block_group);
10725                         ret = PTR_ERR(trans);
10726                         goto next;
10727                 }
10728
10729                 /*
10730                  * We could have pending pinned extents for this block group,
10731                  * just delete them, we don't care about them anymore.
10732                  */
10733                 start = block_group->key.objectid;
10734                 end = start + block_group->key.offset - 1;
10735                 /*
10736                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10737                  * btrfs_finish_extent_commit(). If we are at transaction N,
10738                  * another task might be running finish_extent_commit() for the
10739                  * previous transaction N - 1, and have seen a range belonging
10740                  * to the block group in freed_extents[] before we were able to
10741                  * clear the whole block group range from freed_extents[]. This
10742                  * means that task can lookup for the block group after we
10743                  * unpinned it from freed_extents[] and removed it, leading to
10744                  * a BUG_ON() at btrfs_unpin_extent_range().
10745                  */
10746                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10747                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10748                                   EXTENT_DIRTY);
10749                 if (ret) {
10750                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10751                         btrfs_dec_block_group_ro(block_group);
10752                         goto end_trans;
10753                 }
10754                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10755                                   EXTENT_DIRTY);
10756                 if (ret) {
10757                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10758                         btrfs_dec_block_group_ro(block_group);
10759                         goto end_trans;
10760                 }
10761                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10762
10763                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10764                 spin_lock(&space_info->lock);
10765                 spin_lock(&block_group->lock);
10766
10767                 space_info->bytes_pinned -= block_group->pinned;
10768                 space_info->bytes_readonly += block_group->pinned;
10769                 percpu_counter_add(&space_info->total_bytes_pinned,
10770                                    -block_group->pinned);
10771                 block_group->pinned = 0;
10772
10773                 spin_unlock(&block_group->lock);
10774                 spin_unlock(&space_info->lock);
10775
10776                 /* DISCARD can flip during remount */
10777                 trimming = btrfs_test_opt(fs_info, DISCARD);
10778
10779                 /* Implicit trim during transaction commit. */
10780                 if (trimming)
10781                         btrfs_get_block_group_trimming(block_group);
10782
10783                 /*
10784                  * Btrfs_remove_chunk will abort the transaction if things go
10785                  * horribly wrong.
10786                  */
10787                 ret = btrfs_remove_chunk(trans, fs_info,
10788                                          block_group->key.objectid);
10789
10790                 if (ret) {
10791                         if (trimming)
10792                                 btrfs_put_block_group_trimming(block_group);
10793                         goto end_trans;
10794                 }
10795
10796                 /*
10797                  * If we're not mounted with -odiscard, we can just forget
10798                  * about this block group. Otherwise we'll need to wait
10799                  * until transaction commit to do the actual discard.
10800                  */
10801                 if (trimming) {
10802                         spin_lock(&fs_info->unused_bgs_lock);
10803                         /*
10804                          * A concurrent scrub might have added us to the list
10805                          * fs_info->unused_bgs, so use a list_move operation
10806                          * to add the block group to the deleted_bgs list.
10807                          */
10808                         list_move(&block_group->bg_list,
10809                                   &trans->transaction->deleted_bgs);
10810                         spin_unlock(&fs_info->unused_bgs_lock);
10811                         btrfs_get_block_group(block_group);
10812                 }
10813 end_trans:
10814                 btrfs_end_transaction(trans);
10815 next:
10816                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10817                 btrfs_put_block_group(block_group);
10818                 spin_lock(&fs_info->unused_bgs_lock);
10819         }
10820         spin_unlock(&fs_info->unused_bgs_lock);
10821 }
10822
10823 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10824 {
10825         struct btrfs_space_info *space_info;
10826         struct btrfs_super_block *disk_super;
10827         u64 features;
10828         u64 flags;
10829         int mixed = 0;
10830         int ret;
10831
10832         disk_super = fs_info->super_copy;
10833         if (!btrfs_super_root(disk_super))
10834                 return -EINVAL;
10835
10836         features = btrfs_super_incompat_flags(disk_super);
10837         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10838                 mixed = 1;
10839
10840         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10841         ret = create_space_info(fs_info, flags, &space_info);
10842         if (ret)
10843                 goto out;
10844
10845         if (mixed) {
10846                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10847                 ret = create_space_info(fs_info, flags, &space_info);
10848         } else {
10849                 flags = BTRFS_BLOCK_GROUP_METADATA;
10850                 ret = create_space_info(fs_info, flags, &space_info);
10851                 if (ret)
10852                         goto out;
10853
10854                 flags = BTRFS_BLOCK_GROUP_DATA;
10855                 ret = create_space_info(fs_info, flags, &space_info);
10856         }
10857 out:
10858         return ret;
10859 }
10860
10861 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10862                                    u64 start, u64 end)
10863 {
10864         return unpin_extent_range(fs_info, start, end, false);
10865 }
10866
10867 /*
10868  * It used to be that old block groups would be left around forever.
10869  * Iterating over them would be enough to trim unused space.  Since we
10870  * now automatically remove them, we also need to iterate over unallocated
10871  * space.
10872  *
10873  * We don't want a transaction for this since the discard may take a
10874  * substantial amount of time.  We don't require that a transaction be
10875  * running, but we do need to take a running transaction into account
10876  * to ensure that we're not discarding chunks that were released in
10877  * the current transaction.
10878  *
10879  * Holding the chunks lock will prevent other threads from allocating
10880  * or releasing chunks, but it won't prevent a running transaction
10881  * from committing and releasing the memory that the pending chunks
10882  * list head uses.  For that, we need to take a reference to the
10883  * transaction.
10884  */
10885 static int btrfs_trim_free_extents(struct btrfs_device *device,
10886                                    u64 minlen, u64 *trimmed)
10887 {
10888         u64 start = 0, len = 0;
10889         int ret;
10890
10891         *trimmed = 0;
10892
10893         /* Not writeable = nothing to do. */
10894         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10895                 return 0;
10896
10897         /* No free space = nothing to do. */
10898         if (device->total_bytes <= device->bytes_used)
10899                 return 0;
10900
10901         ret = 0;
10902
10903         while (1) {
10904                 struct btrfs_fs_info *fs_info = device->fs_info;
10905                 struct btrfs_transaction *trans;
10906                 u64 bytes;
10907
10908                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10909                 if (ret)
10910                         return ret;
10911
10912                 down_read(&fs_info->commit_root_sem);
10913
10914                 spin_lock(&fs_info->trans_lock);
10915                 trans = fs_info->running_transaction;
10916                 if (trans)
10917                         refcount_inc(&trans->use_count);
10918                 spin_unlock(&fs_info->trans_lock);
10919
10920                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10921                                                  &start, &len);
10922                 if (trans)
10923                         btrfs_put_transaction(trans);
10924
10925                 if (ret) {
10926                         up_read(&fs_info->commit_root_sem);
10927                         mutex_unlock(&fs_info->chunk_mutex);
10928                         if (ret == -ENOSPC)
10929                                 ret = 0;
10930                         break;
10931                 }
10932
10933                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10934                 up_read(&fs_info->commit_root_sem);
10935                 mutex_unlock(&fs_info->chunk_mutex);
10936
10937                 if (ret)
10938                         break;
10939
10940                 start += len;
10941                 *trimmed += bytes;
10942
10943                 if (fatal_signal_pending(current)) {
10944                         ret = -ERESTARTSYS;
10945                         break;
10946                 }
10947
10948                 cond_resched();
10949         }
10950
10951         return ret;
10952 }
10953
10954 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10955 {
10956         struct btrfs_block_group_cache *cache = NULL;
10957         struct btrfs_device *device;
10958         struct list_head *devices;
10959         u64 group_trimmed;
10960         u64 start;
10961         u64 end;
10962         u64 trimmed = 0;
10963         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10964         int ret = 0;
10965
10966         /*
10967          * try to trim all FS space, our block group may start from non-zero.
10968          */
10969         if (range->len == total_bytes)
10970                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10971         else
10972                 cache = btrfs_lookup_block_group(fs_info, range->start);
10973
10974         while (cache) {
10975                 if (cache->key.objectid >= (range->start + range->len)) {
10976                         btrfs_put_block_group(cache);
10977                         break;
10978                 }
10979
10980                 start = max(range->start, cache->key.objectid);
10981                 end = min(range->start + range->len,
10982                                 cache->key.objectid + cache->key.offset);
10983
10984                 if (end - start >= range->minlen) {
10985                         if (!block_group_cache_done(cache)) {
10986                                 ret = cache_block_group(cache, 0);
10987                                 if (ret) {
10988                                         btrfs_put_block_group(cache);
10989                                         break;
10990                                 }
10991                                 ret = wait_block_group_cache_done(cache);
10992                                 if (ret) {
10993                                         btrfs_put_block_group(cache);
10994                                         break;
10995                                 }
10996                         }
10997                         ret = btrfs_trim_block_group(cache,
10998                                                      &group_trimmed,
10999                                                      start,
11000                                                      end,
11001                                                      range->minlen);
11002
11003                         trimmed += group_trimmed;
11004                         if (ret) {
11005                                 btrfs_put_block_group(cache);
11006                                 break;
11007                         }
11008                 }
11009
11010                 cache = next_block_group(fs_info, cache);
11011         }
11012
11013         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11014         devices = &fs_info->fs_devices->alloc_list;
11015         list_for_each_entry(device, devices, dev_alloc_list) {
11016                 ret = btrfs_trim_free_extents(device, range->minlen,
11017                                               &group_trimmed);
11018                 if (ret)
11019                         break;
11020
11021                 trimmed += group_trimmed;
11022         }
11023         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11024
11025         range->len = trimmed;
11026         return ret;
11027 }
11028
11029 /*
11030  * btrfs_{start,end}_write_no_snapshotting() are similar to
11031  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11032  * data into the page cache through nocow before the subvolume is snapshoted,
11033  * but flush the data into disk after the snapshot creation, or to prevent
11034  * operations while snapshotting is ongoing and that cause the snapshot to be
11035  * inconsistent (writes followed by expanding truncates for example).
11036  */
11037 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11038 {
11039         percpu_counter_dec(&root->subv_writers->counter);
11040         /*
11041          * Make sure counter is updated before we wake up waiters.
11042          */
11043         smp_mb();
11044         if (waitqueue_active(&root->subv_writers->wait))
11045                 wake_up(&root->subv_writers->wait);
11046 }
11047
11048 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11049 {
11050         if (atomic_read(&root->will_be_snapshotted))
11051                 return 0;
11052
11053         percpu_counter_inc(&root->subv_writers->counter);
11054         /*
11055          * Make sure counter is updated before we check for snapshot creation.
11056          */
11057         smp_mb();
11058         if (atomic_read(&root->will_be_snapshotted)) {
11059                 btrfs_end_write_no_snapshotting(root);
11060                 return 0;
11061         }
11062         return 1;
11063 }
11064
11065 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11066 {
11067         while (true) {
11068                 int ret;
11069
11070                 ret = btrfs_start_write_no_snapshotting(root);
11071                 if (ret)
11072                         break;
11073                 wait_on_atomic_t(&root->will_be_snapshotted, atomic_t_wait,
11074                                  TASK_UNINTERRUPTIBLE);
11075         }
11076 }