fs/btrfs/block-group.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "misc.h"
   4 #include "ctree.h"
   5 #include "block-group.h"
   6 #include "space-info.h"
   7 #include "disk-io.h"
   8 #include "free-space-cache.h"
   9 #include "free-space-tree.h"
  10 #include "disk-io.h"
  11 #include "volumes.h"
  12 #include "transaction.h"
  13 #include "ref-verify.h"
  14 #include "sysfs.h"
  15 #include "tree-log.h"
  16 #include "delalloc-space.h"
  17 #include "discard.h"
  18 #include "raid56.h"
  19
  20 /*
  21  * Return target flags in extended format or 0 if restripe for this chunk_type
  22  * is not in progress
  23  *
  24  * Should be called with balance_lock held
  25  */
  26 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
  27 {
  28         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
  29         u64 target = 0;
  30
  31         if (!bctl)
  32                 return 0;
  33
  34         if (flags & BTRFS_BLOCK_GROUP_DATA &&
  35             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  36                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
  37         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
  38                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  39                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
  40         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
  41                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  42                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
  43         }
  44
  45         return target;
  46 }
  47
  48 /*
  49  * @flags: available profiles in extended format (see ctree.h)
  50  *
  51  * Return reduced profile in chunk format.  If profile changing is in progress
  52  * (either running or paused) picks the target profile (if it's already
  53  * available), otherwise falls back to plain reducing.
  54  */
  55 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
  56 {
  57         u64 num_devices = fs_info->fs_devices->rw_devices;
  58         u64 target;
  59         u64 raid_type;
  60         u64 allowed = 0;
  61
  62         /*
  63          * See if restripe for this chunk_type is in progress, if so try to
  64          * reduce to the target profile
  65          */
  66         spin_lock(&fs_info->balance_lock);
  67         target = get_restripe_target(fs_info, flags);
  68         if (target) {
  69                 /* Pick target profile only if it's already available */
  70                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
  71                         spin_unlock(&fs_info->balance_lock);
  72                         return extended_to_chunk(target);
  73                 }
  74         }
  75         spin_unlock(&fs_info->balance_lock);
  76
  77         /* First, mask out the RAID levels which aren't possible */
  78         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
  79                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
  80                         allowed |= btrfs_raid_array[raid_type].bg_flag;
  81         }
  82         allowed &= flags;
  83
  84         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
  85                 allowed = BTRFS_BLOCK_GROUP_RAID6;
  86         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
  87                 allowed = BTRFS_BLOCK_GROUP_RAID5;
  88         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
  89                 allowed = BTRFS_BLOCK_GROUP_RAID10;
  90         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
  91                 allowed = BTRFS_BLOCK_GROUP_RAID1;
  92         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
  93                 allowed = BTRFS_BLOCK_GROUP_RAID0;
  94
  95         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
  96
  97         return extended_to_chunk(flags | allowed);
  98 }
  99
 100 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
 101 {
 102         unsigned seq;
 103         u64 flags;
 104
 105         do {
 106                 flags = orig_flags;
 107                 seq = read_seqbegin(&fs_info->profiles_lock);
 108
 109                 if (flags & BTRFS_BLOCK_GROUP_DATA)
 110                         flags |= fs_info->avail_data_alloc_bits;
 111                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 112                         flags |= fs_info->avail_system_alloc_bits;
 113                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
 114                         flags |= fs_info->avail_metadata_alloc_bits;
 115         } while (read_seqretry(&fs_info->profiles_lock, seq));
 116
 117         return btrfs_reduce_alloc_profile(fs_info, flags);
 118 }
 119
 120 void btrfs_get_block_group(struct btrfs_block_group *cache)
 121 {
 122         atomic_inc(&cache->count);
 123 }
 124
 125 void btrfs_put_block_group(struct btrfs_block_group *cache)
 126 {
 127         if (atomic_dec_and_test(&cache->count)) {
 128                 WARN_ON(cache->pinned > 0);
 129                 WARN_ON(cache->reserved > 0);
 130
 131                 /*
 132                  * A block_group shouldn't be on the discard_list anymore.
 133                  * Remove the block_group from the discard_list to prevent us
 134                  * from causing a panic due to NULL pointer dereference.
 135                  */
 136                 if (WARN_ON(!list_empty(&cache->discard_list)))
 137                         btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
 138                                                   cache);
 139
 140                 /*
 141                  * If not empty, someone is still holding mutex of
 142                  * full_stripe_lock, which can only be released by caller.
 143                  * And it will definitely cause use-after-free when caller
 144                  * tries to release full stripe lock.
 145                  *
 146                  * No better way to resolve, but only to warn.
 147                  */
 148                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 149                 kfree(cache->free_space_ctl);
 150                 kfree(cache);
 151         }
 152 }
 153
 154 /*
 155  * This adds the block group to the fs_info rb tree for the block group cache
 156  */
 157 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 158                                        struct btrfs_block_group *block_group)
 159 {
 160         struct rb_node **p;
 161         struct rb_node *parent = NULL;
 162         struct btrfs_block_group *cache;
 163
 164         spin_lock(&info->block_group_cache_lock);
 165         p = &info->block_group_cache_tree.rb_node;
 166
 167         while (*p) {
 168                 parent = *p;
 169                 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
 170                 if (block_group->start < cache->start) {
 171                         p = &(*p)->rb_left;
 172                 } else if (block_group->start > cache->start) {
 173                         p = &(*p)->rb_right;
 174                 } else {
 175                         spin_unlock(&info->block_group_cache_lock);
 176                         return -EEXIST;
 177                 }
 178         }
 179
 180         rb_link_node(&block_group->cache_node, parent, p);
 181         rb_insert_color(&block_group->cache_node,
 182                         &info->block_group_cache_tree);
 183
 184         if (info->first_logical_byte > block_group->start)
 185                 info->first_logical_byte = block_group->start;
 186
 187         spin_unlock(&info->block_group_cache_lock);
 188
 189         return 0;
 190 }
 191
 192 /*
 193  * This will return the block group at or after bytenr if contains is 0, else
 194  * it will return the block group that contains the bytenr
 195  */
 196 static struct btrfs_block_group *block_group_cache_tree_search(
 197                 struct btrfs_fs_info *info, u64 bytenr, int contains)
 198 {
 199         struct btrfs_block_group *cache, *ret = NULL;
 200         struct rb_node *n;
 201         u64 end, start;
 202
 203         spin_lock(&info->block_group_cache_lock);
 204         n = info->block_group_cache_tree.rb_node;
 205
 206         while (n) {
 207                 cache = rb_entry(n, struct btrfs_block_group, cache_node);
 208                 end = cache->start + cache->length - 1;
 209                 start = cache->start;
 210
 211                 if (bytenr < start) {
 212                         if (!contains && (!ret || start < ret->start))
 213                                 ret = cache;
 214                         n = n->rb_left;
 215                 } else if (bytenr > start) {
 216                         if (contains && bytenr <= end) {
 217                                 ret = cache;
 218                                 break;
 219                         }
 220                         n = n->rb_right;
 221                 } else {
 222                         ret = cache;
 223                         break;
 224                 }
 225         }
 226         if (ret) {
 227                 btrfs_get_block_group(ret);
 228                 if (bytenr == 0 && info->first_logical_byte > ret->start)
 229                         info->first_logical_byte = ret->start;
 230         }
 231         spin_unlock(&info->block_group_cache_lock);
 232
 233         return ret;
 234 }
 235
 236 /*
 237  * Return the block group that starts at or after bytenr
 238  */
 239 struct btrfs_block_group *btrfs_lookup_first_block_group(
 240                 struct btrfs_fs_info *info, u64 bytenr)
 241 {
 242         return block_group_cache_tree_search(info, bytenr, 0);
 243 }
 244
 245 /*
 246  * Return the block group that contains the given bytenr
 247  */
 248 struct btrfs_block_group *btrfs_lookup_block_group(
 249                 struct btrfs_fs_info *info, u64 bytenr)
 250 {
 251         return block_group_cache_tree_search(info, bytenr, 1);
 252 }
 253
 254 struct btrfs_block_group *btrfs_next_block_group(
 255                 struct btrfs_block_group *cache)
 256 {
 257         struct btrfs_fs_info *fs_info = cache->fs_info;
 258         struct rb_node *node;
 259
 260         spin_lock(&fs_info->block_group_cache_lock);
 261
 262         /* If our block group was removed, we need a full search. */
 263         if (RB_EMPTY_NODE(&cache->cache_node)) {
 264                 const u64 next_bytenr = cache->start + cache->length;
 265
 266                 spin_unlock(&fs_info->block_group_cache_lock);
 267                 btrfs_put_block_group(cache);
 268                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
 269         }
 270         node = rb_next(&cache->cache_node);
 271         btrfs_put_block_group(cache);
 272         if (node) {
 273                 cache = rb_entry(node, struct btrfs_block_group, cache_node);
 274                 btrfs_get_block_group(cache);
 275         } else
 276                 cache = NULL;
 277         spin_unlock(&fs_info->block_group_cache_lock);
 278         return cache;
 279 }
 280
 281 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
 282 {
 283         struct btrfs_block_group *bg;
 284         bool ret = true;
 285
 286         bg = btrfs_lookup_block_group(fs_info, bytenr);
 287         if (!bg)
 288                 return false;
 289
 290         spin_lock(&bg->lock);
 291         if (bg->ro)
 292                 ret = false;
 293         else
 294                 atomic_inc(&bg->nocow_writers);
 295         spin_unlock(&bg->lock);
 296
 297         /* No put on block group, done by btrfs_dec_nocow_writers */
 298         if (!ret)
 299                 btrfs_put_block_group(bg);
 300
 301         return ret;
 302 }
 303
 304 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
 305 {
 306         struct btrfs_block_group *bg;
 307
 308         bg = btrfs_lookup_block_group(fs_info, bytenr);
 309         ASSERT(bg);
 310         if (atomic_dec_and_test(&bg->nocow_writers))
 311                 wake_up_var(&bg->nocow_writers);
 312         /*
 313          * Once for our lookup and once for the lookup done by a previous call
 314          * to btrfs_inc_nocow_writers()
 315          */
 316         btrfs_put_block_group(bg);
 317         btrfs_put_block_group(bg);
 318 }
 319
 320 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
 321 {
 322         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
 323 }
 324
 325 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
 326                                         const u64 start)
 327 {
 328         struct btrfs_block_group *bg;
 329
 330         bg = btrfs_lookup_block_group(fs_info, start);
 331         ASSERT(bg);
 332         if (atomic_dec_and_test(&bg->reservations))
 333                 wake_up_var(&bg->reservations);
 334         btrfs_put_block_group(bg);
 335 }
 336
 337 void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
 338 {
 339         struct btrfs_space_info *space_info = bg->space_info;
 340
 341         ASSERT(bg->ro);
 342
 343         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
 344                 return;
 345
 346         /*
 347          * Our block group is read only but before we set it to read only,
 348          * some task might have had allocated an extent from it already, but it
 349          * has not yet created a respective ordered extent (and added it to a
 350          * root's list of ordered extents).
 351          * Therefore wait for any task currently allocating extents, since the
 352          * block group's reservations counter is incremented while a read lock
 353          * on the groups' semaphore is held and decremented after releasing
 354          * the read access on that semaphore and creating the ordered extent.
 355          */
 356         down_write(&space_info->groups_sem);
 357         up_write(&space_info->groups_sem);
 358
 359         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
 360 }
 361
 362 struct btrfs_caching_control *btrfs_get_caching_control(
 363                 struct btrfs_block_group *cache)
 364 {
 365         struct btrfs_caching_control *ctl;
 366
 367         spin_lock(&cache->lock);
 368         if (!cache->caching_ctl) {
 369                 spin_unlock(&cache->lock);
 370                 return NULL;
 371         }
 372
 373         ctl = cache->caching_ctl;
 374         refcount_inc(&ctl->count);
 375         spin_unlock(&cache->lock);
 376         return ctl;
 377 }
 378
 379 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
 380 {
 381         if (refcount_dec_and_test(&ctl->count))
 382                 kfree(ctl);
 383 }
 384
 385 /*
 386  * When we wait for progress in the block group caching, its because our
 387  * allocation attempt failed at least once.  So, we must sleep and let some
 388  * progress happen before we try again.
 389  *
 390  * This function will sleep at least once waiting for new free space to show
 391  * up, and then it will check the block group free space numbers for our min
 392  * num_bytes.  Another option is to have it go ahead and look in the rbtree for
 393  * a free extent of a given size, but this is a good start.
 394  *
 395  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
 396  * any of the information in this block group.
 397  */
 398 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
 399                                            u64 num_bytes)
 400 {
 401         struct btrfs_caching_control *caching_ctl;
 402
 403         caching_ctl = btrfs_get_caching_control(cache);
 404         if (!caching_ctl)
 405                 return;
 406
 407         wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
 408                    (cache->free_space_ctl->free_space >= num_bytes));
 409
 410         btrfs_put_caching_control(caching_ctl);
 411 }
 412
 413 int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
 414 {
 415         struct btrfs_caching_control *caching_ctl;
 416         int ret = 0;
 417
 418         caching_ctl = btrfs_get_caching_control(cache);
 419         if (!caching_ctl)
 420                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
 421
 422         wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
 423         if (cache->cached == BTRFS_CACHE_ERROR)
 424                 ret = -EIO;
 425         btrfs_put_caching_control(caching_ctl);
 426         return ret;
 427 }
 428
 429 #ifdef CONFIG_BTRFS_DEBUG
 430 static void fragment_free_space(struct btrfs_block_group *block_group)
 431 {
 432         struct btrfs_fs_info *fs_info = block_group->fs_info;
 433         u64 start = block_group->start;
 434         u64 len = block_group->length;
 435         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 436                 fs_info->nodesize : fs_info->sectorsize;
 437         u64 step = chunk << 1;
 438
 439         while (len > chunk) {
 440                 btrfs_remove_free_space(block_group, start, chunk);
 441                 start += step;
 442                 if (len < step)
 443                         len = 0;
 444                 else
 445                         len -= step;
 446         }
 447 }
 448 #endif
 449
 450 /*
 451  * This is only called by btrfs_cache_block_group, since we could have freed
 452  * extents we need to check the pinned_extents for any extents that can't be
 453  * used yet since their free space will be released as soon as the transaction
 454  * commits.
 455  */
 456 u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
 457 {
 458         struct btrfs_fs_info *info = block_group->fs_info;
 459         u64 extent_start, extent_end, size, total_added = 0;
 460         int ret;
 461
 462         while (start < end) {
 463                 ret = find_first_extent_bit(info->pinned_extents, start,
 464                                             &extent_start, &extent_end,
 465                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 466                                             NULL);
 467                 if (ret)
 468                         break;
 469
 470                 if (extent_start <= start) {
 471                         start = extent_end + 1;
 472                 } else if (extent_start > start && extent_start < end) {
 473                         size = extent_start - start;
 474                         total_added += size;
 475                         ret = btrfs_add_free_space_async_trimmed(block_group,
 476                                                                  start, size);
 477                         BUG_ON(ret); /* -ENOMEM or logic error */
 478                         start = extent_end + 1;
 479                 } else {
 480                         break;
 481                 }
 482         }
 483
 484         if (start < end) {
 485                 size = end - start;
 486                 total_added += size;
 487                 ret = btrfs_add_free_space_async_trimmed(block_group, start,
 488                                                          size);
 489                 BUG_ON(ret); /* -ENOMEM or logic error */
 490         }
 491
 492         return total_added;
 493 }
 494
 495 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 496 {
 497         struct btrfs_block_group *block_group = caching_ctl->block_group;
 498         struct btrfs_fs_info *fs_info = block_group->fs_info;
 499         struct btrfs_root *extent_root = fs_info->extent_root;
 500         struct btrfs_path *path;
 501         struct extent_buffer *leaf;
 502         struct btrfs_key key;
 503         u64 total_found = 0;
 504         u64 last = 0;
 505         u32 nritems;
 506         int ret;
 507         bool wakeup = true;
 508
 509         path = btrfs_alloc_path();
 510         if (!path)
 511                 return -ENOMEM;
 512
 513         last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
 514
 515 #ifdef CONFIG_BTRFS_DEBUG
 516         /*
 517          * If we're fragmenting we don't want to make anybody think we can
 518          * allocate from this block group until we've had a chance to fragment
 519          * the free space.
 520          */
 521         if (btrfs_should_fragment_free_space(block_group))
 522                 wakeup = false;
 523 #endif
 524         /*
 525          * We don't want to deadlock with somebody trying to allocate a new
 526          * extent for the extent root while also trying to search the extent
 527          * root to add free space.  So we skip locking and search the commit
 528          * root, since its read-only
 529          */
 530         path->skip_locking = 1;
 531         path->search_commit_root = 1;
 532         path->reada = READA_FORWARD;
 533
 534         key.objectid = last;
 535         key.offset = 0;
 536         key.type = BTRFS_EXTENT_ITEM_KEY;
 537
 538 next:
 539         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 540         if (ret < 0)
 541                 goto out;
 542
 543         leaf = path->nodes[0];
 544         nritems = btrfs_header_nritems(leaf);
 545
 546         while (1) {
 547                 if (btrfs_fs_closing(fs_info) > 1) {
 548                         last = (u64)-1;
 549                         break;
 550                 }
 551
 552                 if (path->slots[0] < nritems) {
 553                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 554                 } else {
 555                         ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
 556                         if (ret)
 557                                 break;
 558
 559                         if (need_resched() ||
 560                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 561                                 if (wakeup)
 562                                         caching_ctl->progress = last;
 563                                 btrfs_release_path(path);
 564                                 up_read(&fs_info->commit_root_sem);
 565                                 mutex_unlock(&caching_ctl->mutex);
 566                                 cond_resched();
 567                                 mutex_lock(&caching_ctl->mutex);
 568                                 down_read(&fs_info->commit_root_sem);
 569                                 goto next;
 570                         }
 571
 572                         ret = btrfs_next_leaf(extent_root, path);
 573                         if (ret < 0)
 574                                 goto out;
 575                         if (ret)
 576                                 break;
 577                         leaf = path->nodes[0];
 578                         nritems = btrfs_header_nritems(leaf);
 579                         continue;
 580                 }
 581
 582                 if (key.objectid < last) {
 583                         key.objectid = last;
 584                         key.offset = 0;
 585                         key.type = BTRFS_EXTENT_ITEM_KEY;
 586
 587                         if (wakeup)
 588                                 caching_ctl->progress = last;
 589                         btrfs_release_path(path);
 590                         goto next;
 591                 }
 592
 593                 if (key.objectid < block_group->start) {
 594                         path->slots[0]++;
 595                         continue;
 596                 }
 597
 598                 if (key.objectid >= block_group->start + block_group->length)
 599                         break;
 600
 601                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 602                     key.type == BTRFS_METADATA_ITEM_KEY) {
 603                         total_found += add_new_free_space(block_group, last,
 604                                                           key.objectid);
 605                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 606                                 last = key.objectid +
 607                                         fs_info->nodesize;
 608                         else
 609                                 last = key.objectid + key.offset;
 610
 611                         if (total_found > CACHING_CTL_WAKE_UP) {
 612                                 total_found = 0;
 613                                 if (wakeup)
 614                                         wake_up(&caching_ctl->wait);
 615                         }
 616                 }
 617                 path->slots[0]++;
 618         }
 619         ret = 0;
 620
 621         total_found += add_new_free_space(block_group, last,
 622                                 block_group->start + block_group->length);
 623         caching_ctl->progress = (u64)-1;
 624
 625 out:
 626         btrfs_free_path(path);
 627         return ret;
 628 }
 629
 630 static noinline void caching_thread(struct btrfs_work *work)
 631 {
 632         struct btrfs_block_group *block_group;
 633         struct btrfs_fs_info *fs_info;
 634         struct btrfs_caching_control *caching_ctl;
 635         int ret;
 636
 637         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 638         block_group = caching_ctl->block_group;
 639         fs_info = block_group->fs_info;
 640
 641         mutex_lock(&caching_ctl->mutex);
 642         down_read(&fs_info->commit_root_sem);
 643
 644         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 645                 ret = load_free_space_tree(caching_ctl);
 646         else
 647                 ret = load_extent_tree_free(caching_ctl);
 648
 649         spin_lock(&block_group->lock);
 650         block_group->caching_ctl = NULL;
 651         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 652         spin_unlock(&block_group->lock);
 653
 654 #ifdef CONFIG_BTRFS_DEBUG
 655         if (btrfs_should_fragment_free_space(block_group)) {
 656                 u64 bytes_used;
 657
 658                 spin_lock(&block_group->space_info->lock);
 659                 spin_lock(&block_group->lock);
 660                 bytes_used = block_group->length - block_group->used;
 661                 block_group->space_info->bytes_used += bytes_used >> 1;
 662                 spin_unlock(&block_group->lock);
 663                 spin_unlock(&block_group->space_info->lock);
 664                 fragment_free_space(block_group);
 665         }
 666 #endif
 667
 668         caching_ctl->progress = (u64)-1;
 669
 670         up_read(&fs_info->commit_root_sem);
 671         btrfs_free_excluded_extents(block_group);
 672         mutex_unlock(&caching_ctl->mutex);
 673
 674         wake_up(&caching_ctl->wait);
 675
 676         btrfs_put_caching_control(caching_ctl);
 677         btrfs_put_block_group(block_group);
 678 }
 679
 680 int btrfs_cache_block_group(struct btrfs_block_group *cache, int load_cache_only)
 681 {
 682         DEFINE_WAIT(wait);
 683         struct btrfs_fs_info *fs_info = cache->fs_info;
 684         struct btrfs_caching_control *caching_ctl;
 685         int ret = 0;
 686
 687         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 688         if (!caching_ctl)
 689                 return -ENOMEM;
 690
 691         INIT_LIST_HEAD(&caching_ctl->list);
 692         mutex_init(&caching_ctl->mutex);
 693         init_waitqueue_head(&caching_ctl->wait);
 694         caching_ctl->block_group = cache;
 695         caching_ctl->progress = cache->start;
 696         refcount_set(&caching_ctl->count, 1);
 697         btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 698
 699         spin_lock(&cache->lock);
 700         /*
 701          * This should be a rare occasion, but this could happen I think in the
 702          * case where one thread starts to load the space cache info, and then
 703          * some other thread starts a transaction commit which tries to do an
 704          * allocation while the other thread is still loading the space cache
 705          * info.  The previous loop should have kept us from choosing this block
 706          * group, but if we've moved to the state where we will wait on caching
 707          * block groups we need to first check if we're doing a fast load here,
 708          * so we can wait for it to finish, otherwise we could end up allocating
 709          * from a block group who's cache gets evicted for one reason or
 710          * another.
 711          */
 712         while (cache->cached == BTRFS_CACHE_FAST) {
 713                 struct btrfs_caching_control *ctl;
 714
 715                 ctl = cache->caching_ctl;
 716                 refcount_inc(&ctl->count);
 717                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 718                 spin_unlock(&cache->lock);
 719
 720                 schedule();
 721
 722                 finish_wait(&ctl->wait, &wait);
 723                 btrfs_put_caching_control(ctl);
 724                 spin_lock(&cache->lock);
 725         }
 726
 727         if (cache->cached != BTRFS_CACHE_NO) {
 728                 spin_unlock(&cache->lock);
 729                 kfree(caching_ctl);
 730                 return 0;
 731         }
 732         WARN_ON(cache->caching_ctl);
 733         cache->caching_ctl = caching_ctl;
 734         cache->cached = BTRFS_CACHE_FAST;
 735         spin_unlock(&cache->lock);
 736
 737         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 738                 mutex_lock(&caching_ctl->mutex);
 739                 ret = load_free_space_cache(cache);
 740
 741                 spin_lock(&cache->lock);
 742                 if (ret == 1) {
 743                         cache->caching_ctl = NULL;
 744                         cache->cached = BTRFS_CACHE_FINISHED;
 745                         cache->last_byte_to_unpin = (u64)-1;
 746                         caching_ctl->progress = (u64)-1;
 747                 } else {
 748                         if (load_cache_only) {
 749                                 cache->caching_ctl = NULL;
 750                                 cache->cached = BTRFS_CACHE_NO;
 751                         } else {
 752                                 cache->cached = BTRFS_CACHE_STARTED;
 753                                 cache->has_caching_ctl = 1;
 754                         }
 755                 }
 756                 spin_unlock(&cache->lock);
 757 #ifdef CONFIG_BTRFS_DEBUG
 758                 if (ret == 1 &&
 759                     btrfs_should_fragment_free_space(cache)) {
 760                         u64 bytes_used;
 761
 762                         spin_lock(&cache->space_info->lock);
 763                         spin_lock(&cache->lock);
 764                         bytes_used = cache->length - cache->used;
 765                         cache->space_info->bytes_used += bytes_used >> 1;
 766                         spin_unlock(&cache->lock);
 767                         spin_unlock(&cache->space_info->lock);
 768                         fragment_free_space(cache);
 769                 }
 770 #endif
 771                 mutex_unlock(&caching_ctl->mutex);
 772
 773                 wake_up(&caching_ctl->wait);
 774                 if (ret == 1) {
 775                         btrfs_put_caching_control(caching_ctl);
 776                         btrfs_free_excluded_extents(cache);
 777                         return 0;
 778                 }
 779         } else {
 780                 /*
 781                  * We're either using the free space tree or no caching at all.
 782                  * Set cached to the appropriate value and wakeup any waiters.
 783                  */
 784                 spin_lock(&cache->lock);
 785                 if (load_cache_only) {
 786                         cache->caching_ctl = NULL;
 787                         cache->cached = BTRFS_CACHE_NO;
 788                 } else {
 789                         cache->cached = BTRFS_CACHE_STARTED;
 790                         cache->has_caching_ctl = 1;
 791                 }
 792                 spin_unlock(&cache->lock);
 793                 wake_up(&caching_ctl->wait);
 794         }
 795
 796         if (load_cache_only) {
 797                 btrfs_put_caching_control(caching_ctl);
 798                 return 0;
 799         }
 800
 801         down_write(&fs_info->commit_root_sem);
 802         refcount_inc(&caching_ctl->count);
 803         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 804         up_write(&fs_info->commit_root_sem);
 805
 806         btrfs_get_block_group(cache);
 807
 808         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 809
 810         return ret;
 811 }
 812
 813 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
 814 {
 815         u64 extra_flags = chunk_to_extended(flags) &
 816                                 BTRFS_EXTENDED_PROFILE_MASK;
 817
 818         write_seqlock(&fs_info->profiles_lock);
 819         if (flags & BTRFS_BLOCK_GROUP_DATA)
 820                 fs_info->avail_data_alloc_bits &= ~extra_flags;
 821         if (flags & BTRFS_BLOCK_GROUP_METADATA)
 822                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
 823         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 824                 fs_info->avail_system_alloc_bits &= ~extra_flags;
 825         write_sequnlock(&fs_info->profiles_lock);
 826 }
 827
 828 /*
 829  * Clear incompat bits for the following feature(s):
 830  *
 831  * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
 832  *            in the whole filesystem
 833  *
 834  * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
 835  */
 836 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
 837 {
 838         bool found_raid56 = false;
 839         bool found_raid1c34 = false;
 840
 841         if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
 842             (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
 843             (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
 844                 struct list_head *head = &fs_info->space_info;
 845                 struct btrfs_space_info *sinfo;
 846
 847                 list_for_each_entry_rcu(sinfo, head, list) {
 848                         down_read(&sinfo->groups_sem);
 849                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
 850                                 found_raid56 = true;
 851                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
 852                                 found_raid56 = true;
 853                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
 854                                 found_raid1c34 = true;
 855                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
 856                                 found_raid1c34 = true;
 857                         up_read(&sinfo->groups_sem);
 858                 }
 859                 if (found_raid56)
 860                         btrfs_clear_fs_incompat(fs_info, RAID56);
 861                 if (found_raid1c34)
 862                         btrfs_clear_fs_incompat(fs_info, RAID1C34);
 863         }
 864 }
 865
 866 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 867                              u64 group_start, struct extent_map *em)
 868 {
 869         struct btrfs_fs_info *fs_info = trans->fs_info;
 870         struct btrfs_root *root = fs_info->extent_root;
 871         struct btrfs_path *path;
 872         struct btrfs_block_group *block_group;
 873         struct btrfs_free_cluster *cluster;
 874         struct btrfs_root *tree_root = fs_info->tree_root;
 875         struct btrfs_key key;
 876         struct inode *inode;
 877         struct kobject *kobj = NULL;
 878         int ret;
 879         int index;
 880         int factor;
 881         struct btrfs_caching_control *caching_ctl = NULL;
 882         bool remove_em;
 883         bool remove_rsv = false;
 884
 885         block_group = btrfs_lookup_block_group(fs_info, group_start);
 886         BUG_ON(!block_group);
 887         BUG_ON(!block_group->ro);
 888
 889         trace_btrfs_remove_block_group(block_group);
 890         /*
 891          * Free the reserved super bytes from this block group before
 892          * remove it.
 893          */
 894         btrfs_free_excluded_extents(block_group);
 895         btrfs_free_ref_tree_range(fs_info, block_group->start,
 896                                   block_group->length);
 897
 898         index = btrfs_bg_flags_to_raid_index(block_group->flags);
 899         factor = btrfs_bg_type_to_factor(block_group->flags);
 900
 901         /* make sure this block group isn't part of an allocation cluster */
 902         cluster = &fs_info->data_alloc_cluster;
 903         spin_lock(&cluster->refill_lock);
 904         btrfs_return_cluster_to_free_space(block_group, cluster);
 905         spin_unlock(&cluster->refill_lock);
 906
 907         /*
 908          * make sure this block group isn't part of a metadata
 909          * allocation cluster
 910          */
 911         cluster = &fs_info->meta_alloc_cluster;
 912         spin_lock(&cluster->refill_lock);
 913         btrfs_return_cluster_to_free_space(block_group, cluster);
 914         spin_unlock(&cluster->refill_lock);
 915
 916         path = btrfs_alloc_path();
 917         if (!path) {
 918                 ret = -ENOMEM;
 919                 goto out;
 920         }
 921
 922         /*
 923          * get the inode first so any iput calls done for the io_list
 924          * aren't the final iput (no unlinks allowed now)
 925          */
 926         inode = lookup_free_space_inode(block_group, path);
 927
 928         mutex_lock(&trans->transaction->cache_write_mutex);
 929         /*
 930          * Make sure our free space cache IO is done before removing the
 931          * free space inode
 932          */
 933         spin_lock(&trans->transaction->dirty_bgs_lock);
 934         if (!list_empty(&block_group->io_list)) {
 935                 list_del_init(&block_group->io_list);
 936
 937                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
 938
 939                 spin_unlock(&trans->transaction->dirty_bgs_lock);
 940                 btrfs_wait_cache_io(trans, block_group, path);
 941                 btrfs_put_block_group(block_group);
 942                 spin_lock(&trans->transaction->dirty_bgs_lock);
 943         }
 944
 945         if (!list_empty(&block_group->dirty_list)) {
 946                 list_del_init(&block_group->dirty_list);
 947                 remove_rsv = true;
 948                 btrfs_put_block_group(block_group);
 949         }
 950         spin_unlock(&trans->transaction->dirty_bgs_lock);
 951         mutex_unlock(&trans->transaction->cache_write_mutex);
 952
 953         if (!IS_ERR(inode)) {
 954                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 955                 if (ret) {
 956                         btrfs_add_delayed_iput(inode);
 957                         goto out;
 958                 }
 959                 clear_nlink(inode);
 960                 /* One for the block groups ref */
 961                 spin_lock(&block_group->lock);
 962                 if (block_group->iref) {
 963                         block_group->iref = 0;
 964                         block_group->inode = NULL;
 965                         spin_unlock(&block_group->lock);
 966                         iput(inode);
 967                 } else {
 968                         spin_unlock(&block_group->lock);
 969                 }
 970                 /* One for our lookup ref */
 971                 btrfs_add_delayed_iput(inode);
 972         }
 973
 974         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 975         key.type = 0;
 976         key.offset = block_group->start;
 977
 978         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
 979         if (ret < 0)
 980                 goto out;
 981         if (ret > 0)
 982                 btrfs_release_path(path);
 983         if (ret == 0) {
 984                 ret = btrfs_del_item(trans, tree_root, path);
 985                 if (ret)
 986                         goto out;
 987                 btrfs_release_path(path);
 988         }
 989
 990         spin_lock(&fs_info->block_group_cache_lock);
 991         rb_erase(&block_group->cache_node,
 992                  &fs_info->block_group_cache_tree);
 993         RB_CLEAR_NODE(&block_group->cache_node);
 994
 995         if (fs_info->first_logical_byte == block_group->start)
 996                 fs_info->first_logical_byte = (u64)-1;
 997         spin_unlock(&fs_info->block_group_cache_lock);
 998
 999         down_write(&block_group->space_info->groups_sem);
1000         /*
1001          * we must use list_del_init so people can check to see if they
1002          * are still on the list after taking the semaphore
1003          */
1004         list_del_init(&block_group->list);
1005         if (list_empty(&block_group->space_info->block_groups[index])) {
1006                 kobj = block_group->space_info->block_group_kobjs[index];
1007                 block_group->space_info->block_group_kobjs[index] = NULL;
1008                 clear_avail_alloc_bits(fs_info, block_group->flags);
1009         }
1010         up_write(&block_group->space_info->groups_sem);
1011         clear_incompat_bg_bits(fs_info, block_group->flags);
1012         if (kobj) {
1013                 kobject_del(kobj);
1014                 kobject_put(kobj);
1015         }
1016
1017         if (block_group->has_caching_ctl)
1018                 caching_ctl = btrfs_get_caching_control(block_group);
1019         if (block_group->cached == BTRFS_CACHE_STARTED)
1020                 btrfs_wait_block_group_cache_done(block_group);
1021         if (block_group->has_caching_ctl) {
1022                 down_write(&fs_info->commit_root_sem);
1023                 if (!caching_ctl) {
1024                         struct btrfs_caching_control *ctl;
1025
1026                         list_for_each_entry(ctl,
1027                                     &fs_info->caching_block_groups, list)
1028                                 if (ctl->block_group == block_group) {
1029                                         caching_ctl = ctl;
1030                                         refcount_inc(&caching_ctl->count);
1031                                         break;
1032                                 }
1033                 }
1034                 if (caching_ctl)
1035                         list_del_init(&caching_ctl->list);
1036                 up_write(&fs_info->commit_root_sem);
1037                 if (caching_ctl) {
1038                         /* Once for the caching bgs list and once for us. */
1039                         btrfs_put_caching_control(caching_ctl);
1040                         btrfs_put_caching_control(caching_ctl);
1041                 }
1042         }
1043
1044         spin_lock(&trans->transaction->dirty_bgs_lock);
1045         WARN_ON(!list_empty(&block_group->dirty_list));
1046         WARN_ON(!list_empty(&block_group->io_list));
1047         spin_unlock(&trans->transaction->dirty_bgs_lock);
1048
1049         btrfs_remove_free_space_cache(block_group);
1050
1051         spin_lock(&block_group->space_info->lock);
1052         list_del_init(&block_group->ro_list);
1053
1054         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1055                 WARN_ON(block_group->space_info->total_bytes
1056                         < block_group->length);
1057                 WARN_ON(block_group->space_info->bytes_readonly
1058                         < block_group->length);
1059                 WARN_ON(block_group->space_info->disk_total
1060                         < block_group->length * factor);
1061         }
1062         block_group->space_info->total_bytes -= block_group->length;
1063         block_group->space_info->bytes_readonly -= block_group->length;
1064         block_group->space_info->disk_total -= block_group->length * factor;
1065
1066         spin_unlock(&block_group->space_info->lock);
1067
1068         key.objectid = block_group->start;
1069         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1070         key.offset = block_group->length;
1071
1072         mutex_lock(&fs_info->chunk_mutex);
1073         spin_lock(&block_group->lock);
1074         block_group->removed = 1;
1075         /*
1076          * At this point trimming can't start on this block group, because we
1077          * removed the block group from the tree fs_info->block_group_cache_tree
1078          * so no one can't find it anymore and even if someone already got this
1079          * block group before we removed it from the rbtree, they have already
1080          * incremented block_group->trimming - if they didn't, they won't find
1081          * any free space entries because we already removed them all when we
1082          * called btrfs_remove_free_space_cache().
1083          *
1084          * And we must not remove the extent map from the fs_info->mapping_tree
1085          * to prevent the same logical address range and physical device space
1086          * ranges from being reused for a new block group. This is because our
1087          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1088          * completely transactionless, so while it is trimming a range the
1089          * currently running transaction might finish and a new one start,
1090          * allowing for new block groups to be created that can reuse the same
1091          * physical device locations unless we take this special care.
1092          *
1093          * There may also be an implicit trim operation if the file system
1094          * is mounted with -odiscard. The same protections must remain
1095          * in place until the extents have been discarded completely when
1096          * the transaction commit has completed.
1097          */
1098         remove_em = (atomic_read(&block_group->trimming) == 0);
1099         spin_unlock(&block_group->lock);
1100
1101         mutex_unlock(&fs_info->chunk_mutex);
1102
1103         ret = remove_block_group_free_space(trans, block_group);
1104         if (ret)
1105                 goto out;
1106
1107         btrfs_put_block_group(block_group);
1108         btrfs_put_block_group(block_group);
1109
1110         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111         if (ret > 0)
1112                 ret = -EIO;
1113         if (ret < 0)
1114                 goto out;
1115
1116         ret = btrfs_del_item(trans, root, path);
1117         if (ret)
1118                 goto out;
1119
1120         if (remove_em) {
1121                 struct extent_map_tree *em_tree;
1122
1123                 em_tree = &fs_info->mapping_tree;
1124                 write_lock(&em_tree->lock);
1125                 remove_extent_mapping(em_tree, em);
1126                 write_unlock(&em_tree->lock);
1127                 /* once for the tree */
1128                 free_extent_map(em);
1129         }
1130 out:
1131         if (remove_rsv)
1132                 btrfs_delayed_refs_rsv_release(fs_info, 1);
1133         btrfs_free_path(path);
1134         return ret;
1135 }
1136
1137 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1138                 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1139 {
1140         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1141         struct extent_map *em;
1142         struct map_lookup *map;
1143         unsigned int num_items;
1144
1145         read_lock(&em_tree->lock);
1146         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1147         read_unlock(&em_tree->lock);
1148         ASSERT(em && em->start == chunk_offset);
1149
1150         /*
1151          * We need to reserve 3 + N units from the metadata space info in order
1152          * to remove a block group (done at btrfs_remove_chunk() and at
1153          * btrfs_remove_block_group()), which are used for:
1154          *
1155          * 1 unit for adding the free space inode's orphan (located in the tree
1156          * of tree roots).
1157          * 1 unit for deleting the block group item (located in the extent
1158          * tree).
1159          * 1 unit for deleting the free space item (located in tree of tree
1160          * roots).
1161          * N units for deleting N device extent items corresponding to each
1162          * stripe (located in the device tree).
1163          *
1164          * In order to remove a block group we also need to reserve units in the
1165          * system space info in order to update the chunk tree (update one or
1166          * more device items and remove one chunk item), but this is done at
1167          * btrfs_remove_chunk() through a call to check_system_chunk().
1168          */
1169         map = em->map_lookup;
1170         num_items = 3 + map->num_stripes;
1171         free_extent_map(em);
1172
1173         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
1174                                                            num_items, 1);
1175 }
1176
1177 /*
1178  * Mark block group @cache read-only, so later write won't happen to block
1179  * group @cache.
1180  *
1181  * If @force is not set, this function will only mark the block group readonly
1182  * if we have enough free space (1M) in other metadata/system block groups.
1183  * If @force is not set, this function will mark the block group readonly
1184  * without checking free space.
1185  *
1186  * NOTE: This function doesn't care if other block groups can contain all the
1187  * data in this block group. That check should be done by relocation routine,
1188  * not this function.
1189  */
1190 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1191 {
1192         struct btrfs_space_info *sinfo = cache->space_info;
1193         u64 num_bytes;
1194         int ret = -ENOSPC;
1195
1196         spin_lock(&sinfo->lock);
1197         spin_lock(&cache->lock);
1198
1199         if (cache->ro) {
1200                 cache->ro++;
1201                 ret = 0;
1202                 goto out;
1203         }
1204
1205         num_bytes = cache->length - cache->reserved - cache->pinned -
1206                     cache->bytes_super - cache->used;
1207
1208         /*
1209          * Data never overcommits, even in mixed mode, so do just the straight
1210          * check of left over space in how much we have allocated.
1211          */
1212         if (force) {
1213                 ret = 0;
1214         } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1215                 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1216
1217                 /*
1218                  * Here we make sure if we mark this bg RO, we still have enough
1219                  * free space as buffer.
1220                  */
1221                 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1222                         ret = 0;
1223         } else {
1224                 /*
1225                  * We overcommit metadata, so we need to do the
1226                  * btrfs_can_overcommit check here, and we need to pass in
1227                  * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1228                  * leeway to allow us to mark this block group as read only.
1229                  */
1230                 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1231                                          BTRFS_RESERVE_NO_FLUSH))
1232                         ret = 0;
1233         }
1234
1235         if (!ret) {
1236                 sinfo->bytes_readonly += num_bytes;
1237                 cache->ro++;
1238                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1239         }
1240 out:
1241         spin_unlock(&cache->lock);
1242         spin_unlock(&sinfo->lock);
1243         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1244                 btrfs_info(cache->fs_info,
1245                         "unable to make block group %llu ro", cache->start);
1246                 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1247         }
1248         return ret;
1249 }
1250
1251 /*
1252  * Process the unused_bgs list and remove any that don't have any allocated
1253  * space inside of them.
1254  */
1255 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1256 {
1257         struct btrfs_block_group *block_group;
1258         struct btrfs_space_info *space_info;
1259         struct btrfs_trans_handle *trans;
1260         const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1261         int ret = 0;
1262
1263         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1264                 return;
1265
1266         spin_lock(&fs_info->unused_bgs_lock);
1267         while (!list_empty(&fs_info->unused_bgs)) {
1268                 u64 start, end;
1269                 int trimming;
1270
1271                 block_group = list_first_entry(&fs_info->unused_bgs,
1272                                                struct btrfs_block_group,
1273                                                bg_list);
1274                 list_del_init(&block_group->bg_list);
1275
1276                 space_info = block_group->space_info;
1277
1278                 if (ret || btrfs_mixed_space_info(space_info)) {
1279                         btrfs_put_block_group(block_group);
1280                         continue;
1281                 }
1282                 spin_unlock(&fs_info->unused_bgs_lock);
1283
1284                 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1285
1286                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
1287
1288                 /* Don't want to race with allocators so take the groups_sem */
1289                 down_write(&space_info->groups_sem);
1290
1291                 /*
1292                  * Async discard moves the final block group discard to be prior
1293                  * to the unused_bgs code path.  Therefore, if it's not fully
1294                  * trimmed, punt it back to the async discard lists.
1295                  */
1296                 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1297                     !btrfs_is_free_space_trimmed(block_group)) {
1298                         trace_btrfs_skip_unused_block_group(block_group);
1299                         up_write(&space_info->groups_sem);
1300                         /* Requeue if we failed because of async discard */
1301                         btrfs_discard_queue_work(&fs_info->discard_ctl,
1302                                                  block_group);
1303                         goto next;
1304                 }
1305
1306                 spin_lock(&block_group->lock);
1307                 if (block_group->reserved || block_group->pinned ||
1308                     block_group->used || block_group->ro ||
1309                     list_is_singular(&block_group->list)) {
1310                         /*
1311                          * We want to bail if we made new allocations or have
1312                          * outstanding allocations in this block group.  We do
1313                          * the ro check in case balance is currently acting on
1314                          * this block group.
1315                          */
1316                         trace_btrfs_skip_unused_block_group(block_group);
1317                         spin_unlock(&block_group->lock);
1318                         up_write(&space_info->groups_sem);
1319                         goto next;
1320                 }
1321                 spin_unlock(&block_group->lock);
1322
1323                 /* We don't want to force the issue, only flip if it's ok. */
1324                 ret = inc_block_group_ro(block_group, 0);
1325                 up_write(&space_info->groups_sem);
1326                 if (ret < 0) {
1327                         ret = 0;
1328                         goto next;
1329                 }
1330
1331                 /*
1332                  * Want to do this before we do anything else so we can recover
1333                  * properly if we fail to join the transaction.
1334                  */
1335                 trans = btrfs_start_trans_remove_block_group(fs_info,
1336                                                      block_group->start);
1337                 if (IS_ERR(trans)) {
1338                         btrfs_dec_block_group_ro(block_group);
1339                         ret = PTR_ERR(trans);
1340                         goto next;
1341                 }
1342
1343                 /*
1344                  * We could have pending pinned extents for this block group,
1345                  * just delete them, we don't care about them anymore.
1346                  */
1347                 start = block_group->start;
1348                 end = start + block_group->length - 1;
1349                 /*
1350                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
1351                  * btrfs_finish_extent_commit(). If we are at transaction N,
1352                  * another task might be running finish_extent_commit() for the
1353                  * previous transaction N - 1, and have seen a range belonging
1354                  * to the block group in freed_extents[] before we were able to
1355                  * clear the whole block group range from freed_extents[]. This
1356                  * means that task can lookup for the block group after we
1357                  * unpinned it from freed_extents[] and removed it, leading to
1358                  * a BUG_ON() at btrfs_unpin_extent_range().
1359                  */
1360                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1361                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
1362                                   EXTENT_DIRTY);
1363                 if (ret) {
1364                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1365                         btrfs_dec_block_group_ro(block_group);
1366                         goto end_trans;
1367                 }
1368                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
1369                                   EXTENT_DIRTY);
1370                 if (ret) {
1371                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1372                         btrfs_dec_block_group_ro(block_group);
1373                         goto end_trans;
1374                 }
1375                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1376
1377                 /*
1378                  * At this point, the block_group is read only and should fail
1379                  * new allocations.  However, btrfs_finish_extent_commit() can
1380                  * cause this block_group to be placed back on the discard
1381                  * lists because now the block_group isn't fully discarded.
1382                  * Bail here and try again later after discarding everything.
1383                  */
1384                 spin_lock(&fs_info->discard_ctl.lock);
1385                 if (!list_empty(&block_group->discard_list)) {
1386                         spin_unlock(&fs_info->discard_ctl.lock);
1387                         btrfs_dec_block_group_ro(block_group);
1388                         btrfs_discard_queue_work(&fs_info->discard_ctl,
1389                                                  block_group);
1390                         goto end_trans;
1391                 }
1392                 spin_unlock(&fs_info->discard_ctl.lock);
1393
1394                 /* Reset pinned so btrfs_put_block_group doesn't complain */
1395                 spin_lock(&space_info->lock);
1396                 spin_lock(&block_group->lock);
1397
1398                 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1399                                                      -block_group->pinned);
1400                 space_info->bytes_readonly += block_group->pinned;
1401                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
1402                                    -block_group->pinned,
1403                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
1404                 block_group->pinned = 0;
1405
1406                 spin_unlock(&block_group->lock);
1407                 spin_unlock(&space_info->lock);
1408
1409                 /*
1410                  * The normal path here is an unused block group is passed here,
1411                  * then trimming is handled in the transaction commit path.
1412                  * Async discard interposes before this to do the trimming
1413                  * before coming down the unused block group path as trimming
1414                  * will no longer be done later in the transaction commit path.
1415                  */
1416                 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1417                         goto flip_async;
1418
1419                 /* DISCARD can flip during remount */
1420                 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC);
1421
1422                 /* Implicit trim during transaction commit. */
1423                 if (trimming)
1424                         btrfs_get_block_group_trimming(block_group);
1425
1426                 /*
1427                  * Btrfs_remove_chunk will abort the transaction if things go
1428                  * horribly wrong.
1429                  */
1430                 ret = btrfs_remove_chunk(trans, block_group->start);
1431
1432                 if (ret) {
1433                         if (trimming)
1434                                 btrfs_put_block_group_trimming(block_group);
1435                         goto end_trans;
1436                 }
1437
1438                 /*
1439                  * If we're not mounted with -odiscard, we can just forget
1440                  * about this block group. Otherwise we'll need to wait
1441                  * until transaction commit to do the actual discard.
1442                  */
1443                 if (trimming) {
1444                         spin_lock(&fs_info->unused_bgs_lock);
1445                         /*
1446                          * A concurrent scrub might have added us to the list
1447                          * fs_info->unused_bgs, so use a list_move operation
1448                          * to add the block group to the deleted_bgs list.
1449                          */
1450                         list_move(&block_group->bg_list,
1451                                   &trans->transaction->deleted_bgs);
1452                         spin_unlock(&fs_info->unused_bgs_lock);
1453                         btrfs_get_block_group(block_group);
1454                 }
1455 end_trans:
1456                 btrfs_end_transaction(trans);
1457 next:
1458                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1459                 btrfs_put_block_group(block_group);
1460                 spin_lock(&fs_info->unused_bgs_lock);
1461         }
1462         spin_unlock(&fs_info->unused_bgs_lock);
1463         return;
1464
1465 flip_async:
1466         btrfs_end_transaction(trans);
1467         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1468         btrfs_put_block_group(block_group);
1469         btrfs_discard_punt_unused_bgs_list(fs_info);
1470 }
1471
1472 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1473 {
1474         struct btrfs_fs_info *fs_info = bg->fs_info;
1475
1476         spin_lock(&fs_info->unused_bgs_lock);
1477         if (list_empty(&bg->bg_list)) {
1478                 btrfs_get_block_group(bg);
1479                 trace_btrfs_add_unused_block_group(bg);
1480                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1481         }
1482         spin_unlock(&fs_info->unused_bgs_lock);
1483 }
1484
1485 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1486                                   struct btrfs_path *path,
1487                                   struct btrfs_key *key)
1488 {
1489         struct btrfs_root *root = fs_info->extent_root;
1490         int ret = 0;
1491         struct btrfs_key found_key;
1492         struct extent_buffer *leaf;
1493         struct btrfs_block_group_item bg;
1494         u64 flags;
1495         int slot;
1496
1497         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1498         if (ret < 0)
1499                 goto out;
1500
1501         while (1) {
1502                 slot = path->slots[0];
1503                 leaf = path->nodes[0];
1504                 if (slot >= btrfs_header_nritems(leaf)) {
1505                         ret = btrfs_next_leaf(root, path);
1506                         if (ret == 0)
1507                                 continue;
1508                         if (ret < 0)
1509                                 goto out;
1510                         break;
1511                 }
1512                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1513
1514                 if (found_key.objectid >= key->objectid &&
1515                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1516                         struct extent_map_tree *em_tree;
1517                         struct extent_map *em;
1518
1519                         em_tree = &root->fs_info->mapping_tree;
1520                         read_lock(&em_tree->lock);
1521                         em = lookup_extent_mapping(em_tree, found_key.objectid,
1522                                                    found_key.offset);
1523                         read_unlock(&em_tree->lock);
1524                         if (!em) {
1525                                 btrfs_err(fs_info,
1526                         "logical %llu len %llu found bg but no related chunk",
1527                                           found_key.objectid, found_key.offset);
1528                                 ret = -ENOENT;
1529                         } else if (em->start != found_key.objectid ||
1530                                    em->len != found_key.offset) {
1531                                 btrfs_err(fs_info,
1532                 "block group %llu len %llu mismatch with chunk %llu len %llu",
1533                                           found_key.objectid, found_key.offset,
1534                                           em->start, em->len);
1535                                 ret = -EUCLEAN;
1536                         } else {
1537                                 read_extent_buffer(leaf, &bg,
1538                                         btrfs_item_ptr_offset(leaf, slot),
1539                                         sizeof(bg));
1540                                 flags = btrfs_stack_block_group_flags(&bg) &
1541                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
1542
1543                                 if (flags != (em->map_lookup->type &
1544                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1545                                         btrfs_err(fs_info,
1546 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1547                                                 found_key.objectid,
1548                                                 found_key.offset, flags,
1549                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
1550                                                  em->map_lookup->type));
1551                                         ret = -EUCLEAN;
1552                                 } else {
1553                                         ret = 0;
1554                                 }
1555                         }
1556                         free_extent_map(em);
1557                         goto out;
1558                 }
1559                 path->slots[0]++;
1560         }
1561 out:
1562         return ret;
1563 }
1564
1565 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1566 {
1567         u64 extra_flags = chunk_to_extended(flags) &
1568                                 BTRFS_EXTENDED_PROFILE_MASK;
1569
1570         write_seqlock(&fs_info->profiles_lock);
1571         if (flags & BTRFS_BLOCK_GROUP_DATA)
1572                 fs_info->avail_data_alloc_bits |= extra_flags;
1573         if (flags & BTRFS_BLOCK_GROUP_METADATA)
1574                 fs_info->avail_metadata_alloc_bits |= extra_flags;
1575         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1576                 fs_info->avail_system_alloc_bits |= extra_flags;
1577         write_sequnlock(&fs_info->profiles_lock);
1578 }
1579
1580 /**
1581  * btrfs_rmap_block - Map a physical disk address to a list of logical addresses
1582  * @chunk_start:   logical address of block group
1583  * @physical:      physical address to map to logical addresses
1584  * @logical:       return array of logical addresses which map to @physical
1585  * @naddrs:        length of @logical
1586  * @stripe_len:    size of IO stripe for the given block group
1587  *
1588  * Maps a particular @physical disk address to a list of @logical addresses.
1589  * Used primarily to exclude those portions of a block group that contain super
1590  * block copies.
1591  */
1592 EXPORT_FOR_TESTS
1593 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1594                      u64 physical, u64 **logical, int *naddrs, int *stripe_len)
1595 {
1596         struct extent_map *em;
1597         struct map_lookup *map;
1598         u64 *buf;
1599         u64 bytenr;
1600         u64 data_stripe_length;
1601         u64 io_stripe_size;
1602         int i, nr = 0;
1603         int ret = 0;
1604
1605         em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1606         if (IS_ERR(em))
1607                 return -EIO;
1608
1609         map = em->map_lookup;
1610         data_stripe_length = em->len;
1611         io_stripe_size = map->stripe_len;
1612
1613         if (map->type & BTRFS_BLOCK_GROUP_RAID10)
1614                 data_stripe_length = div_u64(data_stripe_length,
1615                                              map->num_stripes / map->sub_stripes);
1616         else if (map->type & BTRFS_BLOCK_GROUP_RAID0)
1617                 data_stripe_length = div_u64(data_stripe_length, map->num_stripes);
1618         else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1619                 data_stripe_length = div_u64(data_stripe_length,
1620                                              nr_data_stripes(map));
1621                 io_stripe_size = map->stripe_len * nr_data_stripes(map);
1622         }
1623
1624         buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1625         if (!buf) {
1626                 ret = -ENOMEM;
1627                 goto out;
1628         }
1629
1630         for (i = 0; i < map->num_stripes; i++) {
1631                 bool already_inserted = false;
1632                 u64 stripe_nr;
1633                 int j;
1634
1635                 if (!in_range(physical, map->stripes[i].physical,
1636                               data_stripe_length))
1637                         continue;
1638
1639                 stripe_nr = physical - map->stripes[i].physical;
1640                 stripe_nr = div64_u64(stripe_nr, map->stripe_len);
1641
1642                 if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1643                         stripe_nr = stripe_nr * map->num_stripes + i;
1644                         stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1645                 } else if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1646                         stripe_nr = stripe_nr * map->num_stripes + i;
1647                 }
1648                 /*
1649                  * The remaining case would be for RAID56, multiply by
1650                  * nr_data_stripes().  Alternatively, just use rmap_len below
1651                  * instead of map->stripe_len
1652                  */
1653
1654                 bytenr = chunk_start + stripe_nr * io_stripe_size;
1655
1656                 /* Ensure we don't add duplicate addresses */
1657                 for (j = 0; j < nr; j++) {
1658                         if (buf[j] == bytenr) {
1659                                 already_inserted = true;
1660                                 break;
1661                         }
1662                 }
1663
1664                 if (!already_inserted)
1665                         buf[nr++] = bytenr;
1666         }
1667
1668         *logical = buf;
1669         *naddrs = nr;
1670         *stripe_len = io_stripe_size;
1671 out:
1672         free_extent_map(em);
1673         return ret;
1674 }
1675
1676 static int exclude_super_stripes(struct btrfs_block_group *cache)
1677 {
1678         struct btrfs_fs_info *fs_info = cache->fs_info;
1679         u64 bytenr;
1680         u64 *logical;
1681         int stripe_len;
1682         int i, nr, ret;
1683
1684         if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1685                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1686                 cache->bytes_super += stripe_len;
1687                 ret = btrfs_add_excluded_extent(fs_info, cache->start,
1688                                                 stripe_len);
1689                 if (ret)
1690                         return ret;
1691         }
1692
1693         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1694                 bytenr = btrfs_sb_offset(i);
1695                 ret = btrfs_rmap_block(fs_info, cache->start,
1696                                        bytenr, &logical, &nr, &stripe_len);
1697                 if (ret)
1698                         return ret;
1699
1700                 while (nr--) {
1701                         u64 start, len;
1702
1703                         if (logical[nr] > cache->start + cache->length)
1704                                 continue;
1705
1706                         if (logical[nr] + stripe_len <= cache->start)
1707                                 continue;
1708
1709                         start = logical[nr];
1710                         if (start < cache->start) {
1711                                 start = cache->start;
1712                                 len = (logical[nr] + stripe_len) - start;
1713                         } else {
1714                                 len = min_t(u64, stripe_len,
1715                                             cache->start + cache->length - start);
1716                         }
1717
1718                         cache->bytes_super += len;
1719                         ret = btrfs_add_excluded_extent(fs_info, start, len);
1720                         if (ret) {
1721                                 kfree(logical);
1722                                 return ret;
1723                         }
1724                 }
1725
1726                 kfree(logical);
1727         }
1728         return 0;
1729 }
1730
1731 static void link_block_group(struct btrfs_block_group *cache)
1732 {
1733         struct btrfs_space_info *space_info = cache->space_info;
1734         int index = btrfs_bg_flags_to_raid_index(cache->flags);
1735         bool first = false;
1736
1737         down_write(&space_info->groups_sem);
1738         if (list_empty(&space_info->block_groups[index]))
1739                 first = true;
1740         list_add_tail(&cache->list, &space_info->block_groups[index]);
1741         up_write(&space_info->groups_sem);
1742
1743         if (first)
1744                 btrfs_sysfs_add_block_group_type(cache);
1745 }
1746
1747 static struct btrfs_block_group *btrfs_create_block_group_cache(
1748                 struct btrfs_fs_info *fs_info, u64 start, u64 size)
1749 {
1750         struct btrfs_block_group *cache;
1751
1752         cache = kzalloc(sizeof(*cache), GFP_NOFS);
1753         if (!cache)
1754                 return NULL;
1755
1756         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1757                                         GFP_NOFS);
1758         if (!cache->free_space_ctl) {
1759                 kfree(cache);
1760                 return NULL;
1761         }
1762
1763         cache->start = start;
1764         cache->length = size;
1765
1766         cache->fs_info = fs_info;
1767         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1768         set_free_space_tree_thresholds(cache);
1769
1770         cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1771
1772         atomic_set(&cache->count, 1);
1773         spin_lock_init(&cache->lock);
1774         init_rwsem(&cache->data_rwsem);
1775         INIT_LIST_HEAD(&cache->list);
1776         INIT_LIST_HEAD(&cache->cluster_list);
1777         INIT_LIST_HEAD(&cache->bg_list);
1778         INIT_LIST_HEAD(&cache->ro_list);
1779         INIT_LIST_HEAD(&cache->discard_list);
1780         INIT_LIST_HEAD(&cache->dirty_list);
1781         INIT_LIST_HEAD(&cache->io_list);
1782         btrfs_init_free_space_ctl(cache);
1783         atomic_set(&cache->trimming, 0);
1784         mutex_init(&cache->free_space_lock);
1785         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
1786
1787         return cache;
1788 }
1789
1790 /*
1791  * Iterate all chunks and verify that each of them has the corresponding block
1792  * group
1793  */
1794 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
1795 {
1796         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
1797         struct extent_map *em;
1798         struct btrfs_block_group *bg;
1799         u64 start = 0;
1800         int ret = 0;
1801
1802         while (1) {
1803                 read_lock(&map_tree->lock);
1804                 /*
1805                  * lookup_extent_mapping will return the first extent map
1806                  * intersecting the range, so setting @len to 1 is enough to
1807                  * get the first chunk.
1808                  */
1809                 em = lookup_extent_mapping(map_tree, start, 1);
1810                 read_unlock(&map_tree->lock);
1811                 if (!em)
1812                         break;
1813
1814                 bg = btrfs_lookup_block_group(fs_info, em->start);
1815                 if (!bg) {
1816                         btrfs_err(fs_info,
1817         "chunk start=%llu len=%llu doesn't have corresponding block group",
1818                                      em->start, em->len);
1819                         ret = -EUCLEAN;
1820                         free_extent_map(em);
1821                         break;
1822                 }
1823                 if (bg->start != em->start || bg->length != em->len ||
1824                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
1825                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1826                         btrfs_err(fs_info,
1827 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1828                                 em->start, em->len,
1829                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
1830                                 bg->start, bg->length,
1831                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
1832                         ret = -EUCLEAN;
1833                         free_extent_map(em);
1834                         btrfs_put_block_group(bg);
1835                         break;
1836                 }
1837                 start = em->start + em->len;
1838                 free_extent_map(em);
1839                 btrfs_put_block_group(bg);
1840         }
1841         return ret;
1842 }
1843
1844 static int read_one_block_group(struct btrfs_fs_info *info,
1845                                 struct btrfs_path *path,
1846                                 const struct btrfs_key *key,
1847                                 int need_clear)
1848 {
1849         struct extent_buffer *leaf = path->nodes[0];
1850         struct btrfs_block_group *cache;
1851         struct btrfs_space_info *space_info;
1852         struct btrfs_block_group_item bgi;
1853         const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
1854         int slot = path->slots[0];
1855         int ret;
1856
1857         ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
1858
1859         cache = btrfs_create_block_group_cache(info, key->objectid, key->offset);
1860         if (!cache)
1861                 return -ENOMEM;
1862
1863         if (need_clear) {
1864                 /*
1865                  * When we mount with old space cache, we need to
1866                  * set BTRFS_DC_CLEAR and set dirty flag.
1867                  *
1868                  * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
1869                  *    truncate the old free space cache inode and
1870                  *    setup a new one.
1871                  * b) Setting 'dirty flag' makes sure that we flush
1872                  *    the new space cache info onto disk.
1873                  */
1874                 if (btrfs_test_opt(info, SPACE_CACHE))
1875                         cache->disk_cache_state = BTRFS_DC_CLEAR;
1876         }
1877         read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
1878                            sizeof(bgi));
1879         cache->used = btrfs_stack_block_group_used(&bgi);
1880         cache->flags = btrfs_stack_block_group_flags(&bgi);
1881         if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
1882             (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
1883                         btrfs_err(info,
1884 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
1885                                   cache->start);
1886                         ret = -EINVAL;
1887                         goto error;
1888         }
1889
1890         /*
1891          * We need to exclude the super stripes now so that the space info has
1892          * super bytes accounted for, otherwise we'll think we have more space
1893          * than we actually do.
1894          */
1895         ret = exclude_super_stripes(cache);
1896         if (ret) {
1897                 /* We may have excluded something, so call this just in case. */
1898                 btrfs_free_excluded_extents(cache);
1899                 goto error;
1900         }
1901
1902         /*
1903          * Check for two cases, either we are full, and therefore don't need
1904          * to bother with the caching work since we won't find any space, or we
1905          * are empty, and we can just add all the space in and be done with it.
1906          * This saves us _a_lot_ of time, particularly in the full case.
1907          */
1908         if (key->offset == cache->used) {
1909                 cache->last_byte_to_unpin = (u64)-1;
1910                 cache->cached = BTRFS_CACHE_FINISHED;
1911                 btrfs_free_excluded_extents(cache);
1912         } else if (cache->used == 0) {
1913                 cache->last_byte_to_unpin = (u64)-1;
1914                 cache->cached = BTRFS_CACHE_FINISHED;
1915                 add_new_free_space(cache, key->objectid,
1916                                    key->objectid + key->offset);
1917                 btrfs_free_excluded_extents(cache);
1918         }
1919
1920         ret = btrfs_add_block_group_cache(info, cache);
1921         if (ret) {
1922                 btrfs_remove_free_space_cache(cache);
1923                 goto error;
1924         }
1925         trace_btrfs_add_block_group(info, cache, 0);
1926         btrfs_update_space_info(info, cache->flags, key->offset,
1927                                 cache->used, cache->bytes_super, &space_info);
1928
1929         cache->space_info = space_info;
1930
1931         link_block_group(cache);
1932
1933         set_avail_alloc_bits(info, cache->flags);
1934         if (btrfs_chunk_readonly(info, cache->start)) {
1935                 inc_block_group_ro(cache, 1);
1936         } else if (cache->used == 0) {
1937                 ASSERT(list_empty(&cache->bg_list));
1938                 if (btrfs_test_opt(info, DISCARD_ASYNC))
1939                         btrfs_discard_queue_work(&info->discard_ctl, cache);
1940                 else
1941                         btrfs_mark_bg_unused(cache);
1942         }
1943         return 0;
1944 error:
1945         btrfs_put_block_group(cache);
1946         return ret;
1947 }
1948
1949 int btrfs_read_block_groups(struct btrfs_fs_info *info)
1950 {
1951         struct btrfs_path *path;
1952         int ret;
1953         struct btrfs_block_group *cache;
1954         struct btrfs_space_info *space_info;
1955         struct btrfs_key key;
1956         int need_clear = 0;
1957         u64 cache_gen;
1958
1959         key.objectid = 0;
1960         key.offset = 0;
1961         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1962         path = btrfs_alloc_path();
1963         if (!path)
1964                 return -ENOMEM;
1965         path->reada = READA_FORWARD;
1966
1967         cache_gen = btrfs_super_cache_generation(info->super_copy);
1968         if (btrfs_test_opt(info, SPACE_CACHE) &&
1969             btrfs_super_generation(info->super_copy) != cache_gen)
1970                 need_clear = 1;
1971         if (btrfs_test_opt(info, CLEAR_CACHE))
1972                 need_clear = 1;
1973
1974         while (1) {
1975                 ret = find_first_block_group(info, path, &key);
1976                 if (ret > 0)
1977                         break;
1978                 if (ret != 0)
1979                         goto error;
1980
1981                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1982                 ret = read_one_block_group(info, path, &key, need_clear);
1983                 if (ret < 0)
1984                         goto error;
1985                 key.objectid += key.offset;
1986                 key.offset = 0;
1987                 btrfs_release_path(path);
1988         }
1989
1990         list_for_each_entry_rcu(space_info, &info->space_info, list) {
1991                 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
1992                       (BTRFS_BLOCK_GROUP_RAID10 |
1993                        BTRFS_BLOCK_GROUP_RAID1_MASK |
1994                        BTRFS_BLOCK_GROUP_RAID56_MASK |
1995                        BTRFS_BLOCK_GROUP_DUP)))
1996                         continue;
1997                 /*
1998                  * Avoid allocating from un-mirrored block group if there are
1999                  * mirrored block groups.
2000                  */
2001                 list_for_each_entry(cache,
2002                                 &space_info->block_groups[BTRFS_RAID_RAID0],
2003                                 list)
2004                         inc_block_group_ro(cache, 1);
2005                 list_for_each_entry(cache,
2006                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
2007                                 list)
2008                         inc_block_group_ro(cache, 1);
2009         }
2010
2011         btrfs_init_global_block_rsv(info);
2012         ret = check_chunk_block_group_mappings(info);
2013 error:
2014         btrfs_free_path(path);
2015         return ret;
2016 }
2017
2018 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2019 {
2020         struct btrfs_fs_info *fs_info = trans->fs_info;
2021         struct btrfs_block_group *block_group;
2022         struct btrfs_root *extent_root = fs_info->extent_root;
2023         struct btrfs_block_group_item item;
2024         struct btrfs_key key;
2025         int ret = 0;
2026
2027         if (!trans->can_flush_pending_bgs)
2028                 return;
2029
2030         while (!list_empty(&trans->new_bgs)) {
2031                 block_group = list_first_entry(&trans->new_bgs,
2032                                                struct btrfs_block_group,
2033                                                bg_list);
2034                 if (ret)
2035                         goto next;
2036
2037                 spin_lock(&block_group->lock);
2038                 btrfs_set_stack_block_group_used(&item, block_group->used);
2039                 btrfs_set_stack_block_group_chunk_objectid(&item,
2040                                 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2041                 btrfs_set_stack_block_group_flags(&item, block_group->flags);
2042                 key.objectid = block_group->start;
2043                 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2044                 key.offset = block_group->length;
2045                 spin_unlock(&block_group->lock);
2046
2047                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
2048                                         sizeof(item));
2049                 if (ret)
2050                         btrfs_abort_transaction(trans, ret);
2051                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
2052                 if (ret)
2053                         btrfs_abort_transaction(trans, ret);
2054                 add_block_group_free_space(trans, block_group);
2055                 /* Already aborted the transaction if it failed. */
2056 next:
2057                 btrfs_delayed_refs_rsv_release(fs_info, 1);
2058                 list_del_init(&block_group->bg_list);
2059         }
2060         btrfs_trans_release_chunk_metadata(trans);
2061 }
2062
2063 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
2064                            u64 type, u64 chunk_offset, u64 size)
2065 {
2066         struct btrfs_fs_info *fs_info = trans->fs_info;
2067         struct btrfs_block_group *cache;
2068         int ret;
2069
2070         btrfs_set_log_full_commit(trans);
2071
2072         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
2073         if (!cache)
2074                 return -ENOMEM;
2075
2076         cache->used = bytes_used;
2077         cache->flags = type;
2078         cache->last_byte_to_unpin = (u64)-1;
2079         cache->cached = BTRFS_CACHE_FINISHED;
2080         cache->needs_free_space = 1;
2081         ret = exclude_super_stripes(cache);
2082         if (ret) {
2083                 /* We may have excluded something, so call this just in case */
2084                 btrfs_free_excluded_extents(cache);
2085                 btrfs_put_block_group(cache);
2086                 return ret;
2087         }
2088
2089         add_new_free_space(cache, chunk_offset, chunk_offset + size);
2090
2091         btrfs_free_excluded_extents(cache);
2092
2093 #ifdef CONFIG_BTRFS_DEBUG
2094         if (btrfs_should_fragment_free_space(cache)) {
2095                 u64 new_bytes_used = size - bytes_used;
2096
2097                 bytes_used += new_bytes_used >> 1;
2098                 fragment_free_space(cache);
2099         }
2100 #endif
2101         /*
2102          * Ensure the corresponding space_info object is created and
2103          * assigned to our block group. We want our bg to be added to the rbtree
2104          * with its ->space_info set.
2105          */
2106         cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2107         ASSERT(cache->space_info);
2108
2109         ret = btrfs_add_block_group_cache(fs_info, cache);
2110         if (ret) {
2111                 btrfs_remove_free_space_cache(cache);
2112                 btrfs_put_block_group(cache);
2113                 return ret;
2114         }
2115
2116         /*
2117          * Now that our block group has its ->space_info set and is inserted in
2118          * the rbtree, update the space info's counters.
2119          */
2120         trace_btrfs_add_block_group(fs_info, cache, 1);
2121         btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
2122                                 cache->bytes_super, &cache->space_info);
2123         btrfs_update_global_block_rsv(fs_info);
2124
2125         link_block_group(cache);
2126
2127         list_add_tail(&cache->bg_list, &trans->new_bgs);
2128         trans->delayed_ref_updates++;
2129         btrfs_update_delayed_refs_rsv(trans);
2130
2131         set_avail_alloc_bits(fs_info, type);
2132         return 0;
2133 }
2134
2135 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
2136 {
2137         u64 num_devices;
2138         u64 stripped;
2139
2140         /*
2141          * if restripe for this chunk_type is on pick target profile and
2142          * return, otherwise do the usual balance
2143          */
2144         stripped = get_restripe_target(fs_info, flags);
2145         if (stripped)
2146                 return extended_to_chunk(stripped);
2147
2148         num_devices = fs_info->fs_devices->rw_devices;
2149
2150         stripped = BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID56_MASK |
2151                 BTRFS_BLOCK_GROUP_RAID1_MASK | BTRFS_BLOCK_GROUP_RAID10;
2152
2153         if (num_devices == 1) {
2154                 stripped |= BTRFS_BLOCK_GROUP_DUP;
2155                 stripped = flags & ~stripped;
2156
2157                 /* turn raid0 into single device chunks */
2158                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
2159                         return stripped;
2160
2161                 /* turn mirroring into duplication */
2162                 if (flags & (BTRFS_BLOCK_GROUP_RAID1_MASK |
2163                              BTRFS_BLOCK_GROUP_RAID10))
2164                         return stripped | BTRFS_BLOCK_GROUP_DUP;
2165         } else {
2166                 /* they already had raid on here, just return */
2167                 if (flags & stripped)
2168                         return flags;
2169
2170                 stripped |= BTRFS_BLOCK_GROUP_DUP;
2171                 stripped = flags & ~stripped;
2172
2173                 /* switch duplicated blocks with raid1 */
2174                 if (flags & BTRFS_BLOCK_GROUP_DUP)
2175                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
2176
2177                 /* this is drive concat, leave it alone */
2178         }
2179
2180         return flags;
2181 }
2182
2183 /*
2184  * Mark one block group RO, can be called several times for the same block
2185  * group.
2186  *
2187  * @cache:              the destination block group
2188  * @do_chunk_alloc:     whether need to do chunk pre-allocation, this is to
2189  *                      ensure we still have some free space after marking this
2190  *                      block group RO.
2191  */
2192 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2193                              bool do_chunk_alloc)
2194 {
2195         struct btrfs_fs_info *fs_info = cache->fs_info;
2196         struct btrfs_trans_handle *trans;
2197         u64 alloc_flags;
2198         int ret;
2199
2200 again:
2201         trans = btrfs_join_transaction(fs_info->extent_root);
2202         if (IS_ERR(trans))
2203                 return PTR_ERR(trans);
2204
2205         /*
2206          * we're not allowed to set block groups readonly after the dirty
2207          * block groups cache has started writing.  If it already started,
2208          * back off and let this transaction commit
2209          */
2210         mutex_lock(&fs_info->ro_block_group_mutex);
2211         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2212                 u64 transid = trans->transid;
2213
2214                 mutex_unlock(&fs_info->ro_block_group_mutex);
2215                 btrfs_end_transaction(trans);
2216
2217                 ret = btrfs_wait_for_commit(fs_info, transid);
2218                 if (ret)
2219                         return ret;
2220                 goto again;
2221         }
2222
2223         if (do_chunk_alloc) {
2224                 /*
2225                  * If we are changing raid levels, try to allocate a
2226                  * corresponding block group with the new raid level.
2227                  */
2228                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
2229                 if (alloc_flags != cache->flags) {
2230                         ret = btrfs_chunk_alloc(trans, alloc_flags,
2231                                                 CHUNK_ALLOC_FORCE);
2232                         /*
2233                          * ENOSPC is allowed here, we may have enough space
2234                          * already allocated at the new raid level to carry on
2235                          */
2236                         if (ret == -ENOSPC)
2237                                 ret = 0;
2238                         if (ret < 0)
2239                                 goto out;
2240                 }
2241         }
2242
2243         ret = inc_block_group_ro(cache, 0);
2244         if (!do_chunk_alloc)
2245                 goto unlock_out;
2246         if (!ret)
2247                 goto out;
2248         alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2249         ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2250         if (ret < 0)
2251                 goto out;
2252         ret = inc_block_group_ro(cache, 0);
2253 out:
2254         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2255                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
2256                 mutex_lock(&fs_info->chunk_mutex);
2257                 check_system_chunk(trans, alloc_flags);
2258                 mutex_unlock(&fs_info->chunk_mutex);
2259         }
2260 unlock_out:
2261         mutex_unlock(&fs_info->ro_block_group_mutex);
2262
2263         btrfs_end_transaction(trans);
2264         return ret;
2265 }
2266
2267 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2268 {
2269         struct btrfs_space_info *sinfo = cache->space_info;
2270         u64 num_bytes;
2271
2272         BUG_ON(!cache->ro);
2273
2274         spin_lock(&sinfo->lock);
2275         spin_lock(&cache->lock);
2276         if (!--cache->ro) {
2277                 num_bytes = cache->length - cache->reserved -
2278                             cache->pinned - cache->bytes_super - cache->used;
2279                 sinfo->bytes_readonly -= num_bytes;
2280                 list_del_init(&cache->ro_list);
2281         }
2282         spin_unlock(&cache->lock);
2283         spin_unlock(&sinfo->lock);
2284 }
2285
2286 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2287                                  struct btrfs_path *path,
2288                                  struct btrfs_block_group *cache)
2289 {
2290         struct btrfs_fs_info *fs_info = trans->fs_info;
2291         int ret;
2292         struct btrfs_root *extent_root = fs_info->extent_root;
2293         unsigned long bi;
2294         struct extent_buffer *leaf;
2295         struct btrfs_block_group_item bgi;
2296         struct btrfs_key key;
2297
2298         key.objectid = cache->start;
2299         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2300         key.offset = cache->length;
2301
2302         ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 1);
2303         if (ret) {
2304                 if (ret > 0)
2305                         ret = -ENOENT;
2306                 goto fail;
2307         }
2308
2309         leaf = path->nodes[0];
2310         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2311         btrfs_set_stack_block_group_used(&bgi, cache->used);
2312         btrfs_set_stack_block_group_chunk_objectid(&bgi,
2313                         BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2314         btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2315         write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2316         btrfs_mark_buffer_dirty(leaf);
2317 fail:
2318         btrfs_release_path(path);
2319         return ret;
2320
2321 }
2322
2323 static int cache_save_setup(struct btrfs_block_group *block_group,
2324                             struct btrfs_trans_handle *trans,
2325                             struct btrfs_path *path)
2326 {
2327         struct btrfs_fs_info *fs_info = block_group->fs_info;
2328         struct btrfs_root *root = fs_info->tree_root;
2329         struct inode *inode = NULL;
2330         struct extent_changeset *data_reserved = NULL;
2331         u64 alloc_hint = 0;
2332         int dcs = BTRFS_DC_ERROR;
2333         u64 num_pages = 0;
2334         int retries = 0;
2335         int ret = 0;
2336
2337         /*
2338          * If this block group is smaller than 100 megs don't bother caching the
2339          * block group.
2340          */
2341         if (block_group->length < (100 * SZ_1M)) {
2342                 spin_lock(&block_group->lock);
2343                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2344                 spin_unlock(&block_group->lock);
2345                 return 0;
2346         }
2347
2348         if (trans->aborted)
2349                 return 0;
2350 again:
2351         inode = lookup_free_space_inode(block_group, path);
2352         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2353                 ret = PTR_ERR(inode);
2354                 btrfs_release_path(path);
2355                 goto out;
2356         }
2357
2358         if (IS_ERR(inode)) {
2359                 BUG_ON(retries);
2360                 retries++;
2361
2362                 if (block_group->ro)
2363                         goto out_free;
2364
2365                 ret = create_free_space_inode(trans, block_group, path);
2366                 if (ret)
2367                         goto out_free;
2368                 goto again;
2369         }
2370
2371         /*
2372          * We want to set the generation to 0, that way if anything goes wrong
2373          * from here on out we know not to trust this cache when we load up next
2374          * time.
2375          */
2376         BTRFS_I(inode)->generation = 0;
2377         ret = btrfs_update_inode(trans, root, inode);
2378         if (ret) {
2379                 /*
2380                  * So theoretically we could recover from this, simply set the
2381                  * super cache generation to 0 so we know to invalidate the
2382                  * cache, but then we'd have to keep track of the block groups
2383                  * that fail this way so we know we _have_ to reset this cache
2384                  * before the next commit or risk reading stale cache.  So to
2385                  * limit our exposure to horrible edge cases lets just abort the
2386                  * transaction, this only happens in really bad situations
2387                  * anyway.
2388                  */
2389                 btrfs_abort_transaction(trans, ret);
2390                 goto out_put;
2391         }
2392         WARN_ON(ret);
2393
2394         /* We've already setup this transaction, go ahead and exit */
2395         if (block_group->cache_generation == trans->transid &&
2396             i_size_read(inode)) {
2397                 dcs = BTRFS_DC_SETUP;
2398                 goto out_put;
2399         }
2400
2401         if (i_size_read(inode) > 0) {
2402                 ret = btrfs_check_trunc_cache_free_space(fs_info,
2403                                         &fs_info->global_block_rsv);
2404                 if (ret)
2405                         goto out_put;
2406
2407                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2408                 if (ret)
2409                         goto out_put;
2410         }
2411
2412         spin_lock(&block_group->lock);
2413         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2414             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2415                 /*
2416                  * don't bother trying to write stuff out _if_
2417                  * a) we're not cached,
2418                  * b) we're with nospace_cache mount option,
2419                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
2420                  */
2421                 dcs = BTRFS_DC_WRITTEN;
2422                 spin_unlock(&block_group->lock);
2423                 goto out_put;
2424         }
2425         spin_unlock(&block_group->lock);
2426
2427         /*
2428          * We hit an ENOSPC when setting up the cache in this transaction, just
2429          * skip doing the setup, we've already cleared the cache so we're safe.
2430          */
2431         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2432                 ret = -ENOSPC;
2433                 goto out_put;
2434         }
2435
2436         /*
2437          * Try to preallocate enough space based on how big the block group is.
2438          * Keep in mind this has to include any pinned space which could end up
2439          * taking up quite a bit since it's not folded into the other space
2440          * cache.
2441          */
2442         num_pages = div_u64(block_group->length, SZ_256M);
2443         if (!num_pages)
2444                 num_pages = 1;
2445
2446         num_pages *= 16;
2447         num_pages *= PAGE_SIZE;
2448
2449         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
2450         if (ret)
2451                 goto out_put;
2452
2453         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2454                                               num_pages, num_pages,
2455                                               &alloc_hint);
2456         /*
2457          * Our cache requires contiguous chunks so that we don't modify a bunch
2458          * of metadata or split extents when writing the cache out, which means
2459          * we can enospc if we are heavily fragmented in addition to just normal
2460          * out of space conditions.  So if we hit this just skip setting up any
2461          * other block groups for this transaction, maybe we'll unpin enough
2462          * space the next time around.
2463          */
2464         if (!ret)
2465                 dcs = BTRFS_DC_SETUP;
2466         else if (ret == -ENOSPC)
2467                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2468
2469 out_put:
2470         iput(inode);
2471 out_free:
2472         btrfs_release_path(path);
2473 out:
2474         spin_lock(&block_group->lock);
2475         if (!ret && dcs == BTRFS_DC_SETUP)
2476                 block_group->cache_generation = trans->transid;
2477         block_group->disk_cache_state = dcs;
2478         spin_unlock(&block_group->lock);
2479
2480         extent_changeset_free(data_reserved);
2481         return ret;
2482 }
2483
2484 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2485 {
2486         struct btrfs_fs_info *fs_info = trans->fs_info;
2487         struct btrfs_block_group *cache, *tmp;
2488         struct btrfs_transaction *cur_trans = trans->transaction;
2489         struct btrfs_path *path;
2490
2491         if (list_empty(&cur_trans->dirty_bgs) ||
2492             !btrfs_test_opt(fs_info, SPACE_CACHE))
2493                 return 0;
2494
2495         path = btrfs_alloc_path();
2496         if (!path)
2497                 return -ENOMEM;
2498
2499         /* Could add new block groups, use _safe just in case */
2500         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
2501                                  dirty_list) {
2502                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2503                         cache_save_setup(cache, trans, path);
2504         }
2505
2506         btrfs_free_path(path);
2507         return 0;
2508 }
2509
2510 /*
2511  * Transaction commit does final block group cache writeback during a critical
2512  * section where nothing is allowed to change the FS.  This is required in
2513  * order for the cache to actually match the block group, but can introduce a
2514  * lot of latency into the commit.
2515  *
2516  * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2517  * There's a chance we'll have to redo some of it if the block group changes
2518  * again during the commit, but it greatly reduces the commit latency by
2519  * getting rid of the easy block groups while we're still allowing others to
2520  * join the commit.
2521  */
2522 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
2523 {
2524         struct btrfs_fs_info *fs_info = trans->fs_info;
2525         struct btrfs_block_group *cache;
2526         struct btrfs_transaction *cur_trans = trans->transaction;
2527         int ret = 0;
2528         int should_put;
2529         struct btrfs_path *path = NULL;
2530         LIST_HEAD(dirty);
2531         struct list_head *io = &cur_trans->io_bgs;
2532         int num_started = 0;
2533         int loops = 0;
2534
2535         spin_lock(&cur_trans->dirty_bgs_lock);
2536         if (list_empty(&cur_trans->dirty_bgs)) {
2537                 spin_unlock(&cur_trans->dirty_bgs_lock);
2538                 return 0;
2539         }
2540         list_splice_init(&cur_trans->dirty_bgs, &dirty);
2541         spin_unlock(&cur_trans->dirty_bgs_lock);
2542
2543 again:
2544         /* Make sure all the block groups on our dirty list actually exist */
2545         btrfs_create_pending_block_groups(trans);
2546
2547         if (!path) {
2548                 path = btrfs_alloc_path();
2549                 if (!path)
2550                         return -ENOMEM;
2551         }
2552
2553         /*
2554          * cache_write_mutex is here only to save us from balance or automatic
2555          * removal of empty block groups deleting this block group while we are
2556          * writing out the cache
2557          */
2558         mutex_lock(&trans->transaction->cache_write_mutex);
2559         while (!list_empty(&dirty)) {
2560                 bool drop_reserve = true;
2561
2562                 cache = list_first_entry(&dirty, struct btrfs_block_group,
2563                                          dirty_list);
2564                 /*
2565                  * This can happen if something re-dirties a block group that
2566                  * is already under IO.  Just wait for it to finish and then do
2567                  * it all again
2568                  */
2569                 if (!list_empty(&cache->io_list)) {
2570                         list_del_init(&cache->io_list);
2571                         btrfs_wait_cache_io(trans, cache, path);
2572                         btrfs_put_block_group(cache);
2573                 }
2574
2575
2576                 /*
2577                  * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2578                  * it should update the cache_state.  Don't delete until after
2579                  * we wait.
2580                  *
2581                  * Since we're not running in the commit critical section
2582                  * we need the dirty_bgs_lock to protect from update_block_group
2583                  */
2584                 spin_lock(&cur_trans->dirty_bgs_lock);
2585                 list_del_init(&cache->dirty_list);
2586                 spin_unlock(&cur_trans->dirty_bgs_lock);
2587
2588                 should_put = 1;
2589
2590                 cache_save_setup(cache, trans, path);
2591
2592                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
2593                         cache->io_ctl.inode = NULL;
2594                         ret = btrfs_write_out_cache(trans, cache, path);
2595                         if (ret == 0 && cache->io_ctl.inode) {
2596                                 num_started++;
2597                                 should_put = 0;
2598
2599                                 /*
2600                                  * The cache_write_mutex is protecting the
2601                                  * io_list, also refer to the definition of
2602                                  * btrfs_transaction::io_bgs for more details
2603                                  */
2604                                 list_add_tail(&cache->io_list, io);
2605                         } else {
2606                                 /*
2607                                  * If we failed to write the cache, the
2608                                  * generation will be bad and life goes on
2609                                  */
2610                                 ret = 0;
2611                         }
2612                 }
2613                 if (!ret) {
2614                         ret = write_one_cache_group(trans, path, cache);
2615                         /*
2616                          * Our block group might still be attached to the list
2617                          * of new block groups in the transaction handle of some
2618                          * other task (struct btrfs_trans_handle->new_bgs). This
2619                          * means its block group item isn't yet in the extent
2620                          * tree. If this happens ignore the error, as we will
2621                          * try again later in the critical section of the
2622                          * transaction commit.
2623                          */
2624                         if (ret == -ENOENT) {
2625                                 ret = 0;
2626                                 spin_lock(&cur_trans->dirty_bgs_lock);
2627                                 if (list_empty(&cache->dirty_list)) {
2628                                         list_add_tail(&cache->dirty_list,
2629                                                       &cur_trans->dirty_bgs);
2630                                         btrfs_get_block_group(cache);
2631                                         drop_reserve = false;
2632                                 }
2633                                 spin_unlock(&cur_trans->dirty_bgs_lock);
2634                         } else if (ret) {
2635                                 btrfs_abort_transaction(trans, ret);
2636                         }
2637                 }
2638
2639                 /* If it's not on the io list, we need to put the block group */
2640                 if (should_put)
2641                         btrfs_put_block_group(cache);
2642                 if (drop_reserve)
2643                         btrfs_delayed_refs_rsv_release(fs_info, 1);
2644
2645                 if (ret)
2646                         break;
2647
2648                 /*
2649                  * Avoid blocking other tasks for too long. It might even save
2650                  * us from writing caches for block groups that are going to be
2651                  * removed.
2652                  */
2653                 mutex_unlock(&trans->transaction->cache_write_mutex);
2654                 mutex_lock(&trans->transaction->cache_write_mutex);
2655         }
2656         mutex_unlock(&trans->transaction->cache_write_mutex);
2657
2658         /*
2659          * Go through delayed refs for all the stuff we've just kicked off
2660          * and then loop back (just once)
2661          */
2662         ret = btrfs_run_delayed_refs(trans, 0);
2663         if (!ret && loops == 0) {
2664                 loops++;
2665                 spin_lock(&cur_trans->dirty_bgs_lock);
2666                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
2667                 /*
2668                  * dirty_bgs_lock protects us from concurrent block group
2669                  * deletes too (not just cache_write_mutex).
2670                  */
2671                 if (!list_empty(&dirty)) {
2672                         spin_unlock(&cur_trans->dirty_bgs_lock);
2673                         goto again;
2674                 }
2675                 spin_unlock(&cur_trans->dirty_bgs_lock);
2676         } else if (ret < 0) {
2677                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
2678         }
2679
2680         btrfs_free_path(path);
2681         return ret;
2682 }
2683
2684 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
2685 {
2686         struct btrfs_fs_info *fs_info = trans->fs_info;
2687         struct btrfs_block_group *cache;
2688         struct btrfs_transaction *cur_trans = trans->transaction;
2689         int ret = 0;
2690         int should_put;
2691         struct btrfs_path *path;
2692         struct list_head *io = &cur_trans->io_bgs;
2693         int num_started = 0;
2694
2695         path = btrfs_alloc_path();
2696         if (!path)
2697                 return -ENOMEM;
2698
2699         /*
2700          * Even though we are in the critical section of the transaction commit,
2701          * we can still have concurrent tasks adding elements to this
2702          * transaction's list of dirty block groups. These tasks correspond to
2703          * endio free space workers started when writeback finishes for a
2704          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
2705          * allocate new block groups as a result of COWing nodes of the root
2706          * tree when updating the free space inode. The writeback for the space
2707          * caches is triggered by an earlier call to
2708          * btrfs_start_dirty_block_groups() and iterations of the following
2709          * loop.
2710          * Also we want to do the cache_save_setup first and then run the
2711          * delayed refs to make sure we have the best chance at doing this all
2712          * in one shot.
2713          */
2714         spin_lock(&cur_trans->dirty_bgs_lock);
2715         while (!list_empty(&cur_trans->dirty_bgs)) {
2716                 cache = list_first_entry(&cur_trans->dirty_bgs,
2717                                          struct btrfs_block_group,
2718                                          dirty_list);
2719
2720                 /*
2721                  * This can happen if cache_save_setup re-dirties a block group
2722                  * that is already under IO.  Just wait for it to finish and
2723                  * then do it all again
2724                  */
2725                 if (!list_empty(&cache->io_list)) {
2726                         spin_unlock(&cur_trans->dirty_bgs_lock);
2727                         list_del_init(&cache->io_list);
2728                         btrfs_wait_cache_io(trans, cache, path);
2729                         btrfs_put_block_group(cache);
2730                         spin_lock(&cur_trans->dirty_bgs_lock);
2731                 }
2732
2733                 /*
2734                  * Don't remove from the dirty list until after we've waited on
2735                  * any pending IO
2736                  */
2737                 list_del_init(&cache->dirty_list);
2738                 spin_unlock(&cur_trans->dirty_bgs_lock);
2739                 should_put = 1;
2740
2741                 cache_save_setup(cache, trans, path);
2742
2743                 if (!ret)
2744                         ret = btrfs_run_delayed_refs(trans,
2745                                                      (unsigned long) -1);
2746
2747                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
2748                         cache->io_ctl.inode = NULL;
2749                         ret = btrfs_write_out_cache(trans, cache, path);
2750                         if (ret == 0 && cache->io_ctl.inode) {
2751                                 num_started++;
2752                                 should_put = 0;
2753                                 list_add_tail(&cache->io_list, io);
2754                         } else {
2755                                 /*
2756                                  * If we failed to write the cache, the
2757                                  * generation will be bad and life goes on
2758                                  */
2759                                 ret = 0;
2760                         }
2761                 }
2762                 if (!ret) {
2763                         ret = write_one_cache_group(trans, path, cache);
2764                         /*
2765                          * One of the free space endio workers might have
2766                          * created a new block group while updating a free space
2767                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
2768                          * and hasn't released its transaction handle yet, in
2769                          * which case the new block group is still attached to
2770                          * its transaction handle and its creation has not
2771                          * finished yet (no block group item in the extent tree
2772                          * yet, etc). If this is the case, wait for all free
2773                          * space endio workers to finish and retry. This is a
2774                          * a very rare case so no need for a more efficient and
2775                          * complex approach.
2776                          */
2777                         if (ret == -ENOENT) {
2778                                 wait_event(cur_trans->writer_wait,
2779                                    atomic_read(&cur_trans->num_writers) == 1);
2780                                 ret = write_one_cache_group(trans, path, cache);
2781                         }
2782                         if (ret)
2783                                 btrfs_abort_transaction(trans, ret);
2784                 }
2785
2786                 /* If its not on the io list, we need to put the block group */
2787                 if (should_put)
2788                         btrfs_put_block_group(cache);
2789                 btrfs_delayed_refs_rsv_release(fs_info, 1);
2790                 spin_lock(&cur_trans->dirty_bgs_lock);
2791         }
2792         spin_unlock(&cur_trans->dirty_bgs_lock);
2793
2794         /*
2795          * Refer to the definition of io_bgs member for details why it's safe
2796          * to use it without any locking
2797          */
2798         while (!list_empty(io)) {
2799                 cache = list_first_entry(io, struct btrfs_block_group,
2800                                          io_list);
2801                 list_del_init(&cache->io_list);
2802                 btrfs_wait_cache_io(trans, cache, path);
2803                 btrfs_put_block_group(cache);
2804         }
2805
2806         btrfs_free_path(path);
2807         return ret;
2808 }
2809
2810 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
2811                              u64 bytenr, u64 num_bytes, int alloc)
2812 {
2813         struct btrfs_fs_info *info = trans->fs_info;
2814         struct btrfs_block_group *cache = NULL;
2815         u64 total = num_bytes;
2816         u64 old_val;
2817         u64 byte_in_group;
2818         int factor;
2819         int ret = 0;
2820
2821         /* Block accounting for super block */
2822         spin_lock(&info->delalloc_root_lock);
2823         old_val = btrfs_super_bytes_used(info->super_copy);
2824         if (alloc)
2825                 old_val += num_bytes;
2826         else
2827                 old_val -= num_bytes;
2828         btrfs_set_super_bytes_used(info->super_copy, old_val);
2829         spin_unlock(&info->delalloc_root_lock);
2830
2831         while (total) {
2832                 cache = btrfs_lookup_block_group(info, bytenr);
2833                 if (!cache) {
2834                         ret = -ENOENT;
2835                         break;
2836                 }
2837                 factor = btrfs_bg_type_to_factor(cache->flags);
2838
2839                 /*
2840                  * If this block group has free space cache written out, we
2841                  * need to make sure to load it if we are removing space.  This
2842                  * is because we need the unpinning stage to actually add the
2843                  * space back to the block group, otherwise we will leak space.
2844                  */
2845                 if (!alloc && !btrfs_block_group_done(cache))
2846                         btrfs_cache_block_group(cache, 1);
2847
2848                 byte_in_group = bytenr - cache->start;
2849                 WARN_ON(byte_in_group > cache->length);
2850
2851                 spin_lock(&cache->space_info->lock);
2852                 spin_lock(&cache->lock);
2853
2854                 if (btrfs_test_opt(info, SPACE_CACHE) &&
2855                     cache->disk_cache_state < BTRFS_DC_CLEAR)
2856                         cache->disk_cache_state = BTRFS_DC_CLEAR;
2857
2858                 old_val = cache->used;
2859                 num_bytes = min(total, cache->length - byte_in_group);
2860                 if (alloc) {
2861                         old_val += num_bytes;
2862                         cache->used = old_val;
2863                         cache->reserved -= num_bytes;
2864                         cache->space_info->bytes_reserved -= num_bytes;
2865                         cache->space_info->bytes_used += num_bytes;
2866                         cache->space_info->disk_used += num_bytes * factor;
2867                         spin_unlock(&cache->lock);
2868                         spin_unlock(&cache->space_info->lock);
2869                 } else {
2870                         old_val -= num_bytes;
2871                         cache->used = old_val;
2872                         cache->pinned += num_bytes;
2873                         btrfs_space_info_update_bytes_pinned(info,
2874                                         cache->space_info, num_bytes);
2875                         cache->space_info->bytes_used -= num_bytes;
2876                         cache->space_info->disk_used -= num_bytes * factor;
2877                         spin_unlock(&cache->lock);
2878                         spin_unlock(&cache->space_info->lock);
2879
2880                         percpu_counter_add_batch(
2881                                         &cache->space_info->total_bytes_pinned,
2882                                         num_bytes,
2883                                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
2884                         set_extent_dirty(info->pinned_extents,
2885                                          bytenr, bytenr + num_bytes - 1,
2886                                          GFP_NOFS | __GFP_NOFAIL);
2887                 }
2888
2889                 spin_lock(&trans->transaction->dirty_bgs_lock);
2890                 if (list_empty(&cache->dirty_list)) {
2891                         list_add_tail(&cache->dirty_list,
2892                                       &trans->transaction->dirty_bgs);
2893                         trans->delayed_ref_updates++;
2894                         btrfs_get_block_group(cache);
2895                 }
2896                 spin_unlock(&trans->transaction->dirty_bgs_lock);
2897
2898                 /*
2899                  * No longer have used bytes in this block group, queue it for
2900                  * deletion. We do this after adding the block group to the
2901                  * dirty list to avoid races between cleaner kthread and space
2902                  * cache writeout.
2903                  */
2904                 if (!alloc && old_val == 0) {
2905                         if (!btrfs_test_opt(info, DISCARD_ASYNC))
2906                                 btrfs_mark_bg_unused(cache);
2907                 }
2908
2909                 btrfs_put_block_group(cache);
2910                 total -= num_bytes;
2911                 bytenr += num_bytes;
2912         }
2913
2914         /* Modified block groups are accounted for in the delayed_refs_rsv. */
2915         btrfs_update_delayed_refs_rsv(trans);
2916         return ret;
2917 }
2918
2919 /**
2920  * btrfs_add_reserved_bytes - update the block_group and space info counters
2921  * @cache:      The cache we are manipulating
2922  * @ram_bytes:  The number of bytes of file content, and will be same to
2923  *              @num_bytes except for the compress path.
2924  * @num_bytes:  The number of bytes in question
2925  * @delalloc:   The blocks are allocated for the delalloc write
2926  *
2927  * This is called by the allocator when it reserves space. If this is a
2928  * reservation and the block group has become read only we cannot make the
2929  * reservation and return -EAGAIN, otherwise this function always succeeds.
2930  */
2931 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
2932                              u64 ram_bytes, u64 num_bytes, int delalloc)
2933 {
2934         struct btrfs_space_info *space_info = cache->space_info;
2935         int ret = 0;
2936
2937         spin_lock(&space_info->lock);
2938         spin_lock(&cache->lock);
2939         if (cache->ro) {
2940                 ret = -EAGAIN;
2941         } else {
2942                 cache->reserved += num_bytes;
2943                 space_info->bytes_reserved += num_bytes;
2944                 trace_btrfs_space_reservation(cache->fs_info, "space_info",
2945                                               space_info->flags, num_bytes, 1);
2946                 btrfs_space_info_update_bytes_may_use(cache->fs_info,
2947                                                       space_info, -ram_bytes);
2948                 if (delalloc)
2949                         cache->delalloc_bytes += num_bytes;
2950         }
2951         spin_unlock(&cache->lock);
2952         spin_unlock(&space_info->lock);
2953         return ret;
2954 }
2955
2956 /**
2957  * btrfs_free_reserved_bytes - update the block_group and space info counters
2958  * @cache:      The cache we are manipulating
2959  * @num_bytes:  The number of bytes in question
2960  * @delalloc:   The blocks are allocated for the delalloc write
2961  *
2962  * This is called by somebody who is freeing space that was never actually used
2963  * on disk.  For example if you reserve some space for a new leaf in transaction
2964  * A and before transaction A commits you free that leaf, you call this with
2965  * reserve set to 0 in order to clear the reservation.
2966  */
2967 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
2968                                u64 num_bytes, int delalloc)
2969 {
2970         struct btrfs_space_info *space_info = cache->space_info;
2971
2972         spin_lock(&space_info->lock);
2973         spin_lock(&cache->lock);
2974         if (cache->ro)
2975                 space_info->bytes_readonly += num_bytes;
2976         cache->reserved -= num_bytes;
2977         space_info->bytes_reserved -= num_bytes;
2978         space_info->max_extent_size = 0;
2979
2980         if (delalloc)
2981                 cache->delalloc_bytes -= num_bytes;
2982         spin_unlock(&cache->lock);
2983         spin_unlock(&space_info->lock);
2984 }
2985
2986 static void force_metadata_allocation(struct btrfs_fs_info *info)
2987 {
2988         struct list_head *head = &info->space_info;
2989         struct btrfs_space_info *found;
2990
2991         rcu_read_lock();
2992         list_for_each_entry_rcu(found, head, list) {
2993                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
2994                         found->force_alloc = CHUNK_ALLOC_FORCE;
2995         }
2996         rcu_read_unlock();
2997 }
2998
2999 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3000                               struct btrfs_space_info *sinfo, int force)
3001 {
3002         u64 bytes_used = btrfs_space_info_used(sinfo, false);
3003         u64 thresh;
3004
3005         if (force == CHUNK_ALLOC_FORCE)
3006                 return 1;
3007
3008         /*
3009          * in limited mode, we want to have some free space up to
3010          * about 1% of the FS size.
3011          */
3012         if (force == CHUNK_ALLOC_LIMITED) {
3013                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3014                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3015
3016                 if (sinfo->total_bytes - bytes_used < thresh)
3017                         return 1;
3018         }
3019
3020         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3021                 return 0;
3022         return 1;
3023 }
3024
3025 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3026 {
3027         u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3028
3029         return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3030 }
3031
3032 /*
3033  * If force is CHUNK_ALLOC_FORCE:
3034  *    - return 1 if it successfully allocates a chunk,
3035  *    - return errors including -ENOSPC otherwise.
3036  * If force is NOT CHUNK_ALLOC_FORCE:
3037  *    - return 0 if it doesn't need to allocate a new chunk,
3038  *    - return 1 if it successfully allocates a chunk,
3039  *    - return errors including -ENOSPC otherwise.
3040  */
3041 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3042                       enum btrfs_chunk_alloc_enum force)
3043 {
3044         struct btrfs_fs_info *fs_info = trans->fs_info;
3045         struct btrfs_space_info *space_info;
3046         bool wait_for_alloc = false;
3047         bool should_alloc = false;
3048         int ret = 0;
3049
3050         /* Don't re-enter if we're already allocating a chunk */
3051         if (trans->allocating_chunk)
3052                 return -ENOSPC;
3053
3054         space_info = btrfs_find_space_info(fs_info, flags);
3055         ASSERT(space_info);
3056
3057         do {
3058                 spin_lock(&space_info->lock);
3059                 if (force < space_info->force_alloc)
3060                         force = space_info->force_alloc;
3061                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
3062                 if (space_info->full) {
3063                         /* No more free physical space */
3064                         if (should_alloc)
3065                                 ret = -ENOSPC;
3066                         else
3067                                 ret = 0;
3068                         spin_unlock(&space_info->lock);
3069                         return ret;
3070                 } else if (!should_alloc) {
3071                         spin_unlock(&space_info->lock);
3072                         return 0;
3073                 } else if (space_info->chunk_alloc) {
3074                         /*
3075                          * Someone is already allocating, so we need to block
3076                          * until this someone is finished and then loop to
3077                          * recheck if we should continue with our allocation
3078                          * attempt.
3079                          */
3080                         wait_for_alloc = true;
3081                         spin_unlock(&space_info->lock);
3082                         mutex_lock(&fs_info->chunk_mutex);
3083                         mutex_unlock(&fs_info->chunk_mutex);
3084                 } else {
3085                         /* Proceed with allocation */
3086                         space_info->chunk_alloc = 1;
3087                         wait_for_alloc = false;
3088                         spin_unlock(&space_info->lock);
3089                 }
3090
3091                 cond_resched();
3092         } while (wait_for_alloc);
3093
3094         mutex_lock(&fs_info->chunk_mutex);
3095         trans->allocating_chunk = true;
3096
3097         /*
3098          * If we have mixed data/metadata chunks we want to make sure we keep
3099          * allocating mixed chunks instead of individual chunks.
3100          */
3101         if (btrfs_mixed_space_info(space_info))
3102                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3103
3104         /*
3105          * if we're doing a data chunk, go ahead and make sure that
3106          * we keep a reasonable number of metadata chunks allocated in the
3107          * FS as well.
3108          */
3109         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3110                 fs_info->data_chunk_allocations++;
3111                 if (!(fs_info->data_chunk_allocations %
3112                       fs_info->metadata_ratio))
3113                         force_metadata_allocation(fs_info);
3114         }
3115
3116         /*
3117          * Check if we have enough space in SYSTEM chunk because we may need
3118          * to update devices.
3119          */
3120         check_system_chunk(trans, flags);
3121
3122         ret = btrfs_alloc_chunk(trans, flags);
3123         trans->allocating_chunk = false;
3124
3125         spin_lock(&space_info->lock);
3126         if (ret < 0) {
3127                 if (ret == -ENOSPC)
3128                         space_info->full = 1;
3129                 else
3130                         goto out;
3131         } else {
3132                 ret = 1;
3133                 space_info->max_extent_size = 0;
3134         }
3135
3136         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3137 out:
3138         space_info->chunk_alloc = 0;
3139         spin_unlock(&space_info->lock);
3140         mutex_unlock(&fs_info->chunk_mutex);
3141         /*
3142          * When we allocate a new chunk we reserve space in the chunk block
3143          * reserve to make sure we can COW nodes/leafs in the chunk tree or
3144          * add new nodes/leafs to it if we end up needing to do it when
3145          * inserting the chunk item and updating device items as part of the
3146          * second phase of chunk allocation, performed by
3147          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
3148          * large number of new block groups to create in our transaction
3149          * handle's new_bgs list to avoid exhausting the chunk block reserve
3150          * in extreme cases - like having a single transaction create many new
3151          * block groups when starting to write out the free space caches of all
3152          * the block groups that were made dirty during the lifetime of the
3153          * transaction.
3154          */
3155         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
3156                 btrfs_create_pending_block_groups(trans);
3157
3158         return ret;
3159 }
3160
3161 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3162 {
3163         u64 num_dev;
3164
3165         num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3166         if (!num_dev)
3167                 num_dev = fs_info->fs_devices->rw_devices;
3168
3169         return num_dev;
3170 }
3171
3172 /*
3173  * Reserve space in the system space for allocating or removing a chunk
3174  */
3175 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3176 {
3177         struct btrfs_fs_info *fs_info = trans->fs_info;
3178         struct btrfs_space_info *info;
3179         u64 left;
3180         u64 thresh;
3181         int ret = 0;
3182         u64 num_devs;
3183
3184         /*
3185          * Needed because we can end up allocating a system chunk and for an
3186          * atomic and race free space reservation in the chunk block reserve.
3187          */
3188         lockdep_assert_held(&fs_info->chunk_mutex);
3189
3190         info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3191         spin_lock(&info->lock);
3192         left = info->total_bytes - btrfs_space_info_used(info, true);
3193         spin_unlock(&info->lock);
3194
3195         num_devs = get_profile_num_devs(fs_info, type);
3196
3197         /* num_devs device items to update and 1 chunk item to add or remove */
3198         thresh = btrfs_calc_metadata_size(fs_info, num_devs) +
3199                 btrfs_calc_insert_metadata_size(fs_info, 1);
3200
3201         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3202                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3203                            left, thresh, type);
3204                 btrfs_dump_space_info(fs_info, info, 0, 0);
3205         }
3206
3207         if (left < thresh) {
3208                 u64 flags = btrfs_system_alloc_profile(fs_info);
3209
3210                 /*
3211                  * Ignore failure to create system chunk. We might end up not
3212                  * needing it, as we might not need to COW all nodes/leafs from
3213                  * the paths we visit in the chunk tree (they were already COWed
3214                  * or created in the current transaction for example).
3215                  */
3216                 ret = btrfs_alloc_chunk(trans, flags);
3217         }
3218
3219         if (!ret) {
3220                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
3221                                           &fs_info->chunk_block_rsv,
3222                                           thresh, BTRFS_RESERVE_NO_FLUSH);
3223                 if (!ret)
3224                         trans->chunk_bytes_reserved += thresh;
3225         }
3226 }
3227
3228 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
3229 {
3230         struct btrfs_block_group *block_group;
3231         u64 last = 0;
3232
3233         while (1) {
3234                 struct inode *inode;
3235
3236                 block_group = btrfs_lookup_first_block_group(info, last);
3237                 while (block_group) {
3238                         btrfs_wait_block_group_cache_done(block_group);
3239                         spin_lock(&block_group->lock);
3240                         if (block_group->iref)
3241                                 break;
3242                         spin_unlock(&block_group->lock);
3243                         block_group = btrfs_next_block_group(block_group);
3244                 }
3245                 if (!block_group) {
3246                         if (last == 0)
3247                                 break;
3248                         last = 0;
3249                         continue;
3250                 }
3251
3252                 inode = block_group->inode;
3253                 block_group->iref = 0;
3254                 block_group->inode = NULL;
3255                 spin_unlock(&block_group->lock);
3256                 ASSERT(block_group->io_ctl.inode == NULL);
3257                 iput(inode);
3258                 last = block_group->start + block_group->length;
3259                 btrfs_put_block_group(block_group);
3260         }
3261 }
3262
3263 /*
3264  * Must be called only after stopping all workers, since we could have block
3265  * group caching kthreads running, and therefore they could race with us if we
3266  * freed the block groups before stopping them.
3267  */
3268 int btrfs_free_block_groups(struct btrfs_fs_info *info)
3269 {
3270         struct btrfs_block_group *block_group;
3271         struct btrfs_space_info *space_info;
3272         struct btrfs_caching_control *caching_ctl;
3273         struct rb_node *n;
3274
3275         down_write(&info->commit_root_sem);
3276         while (!list_empty(&info->caching_block_groups)) {
3277                 caching_ctl = list_entry(info->caching_block_groups.next,
3278                                          struct btrfs_caching_control, list);
3279                 list_del(&caching_ctl->list);
3280                 btrfs_put_caching_control(caching_ctl);
3281         }
3282         up_write(&info->commit_root_sem);
3283
3284         spin_lock(&info->unused_bgs_lock);
3285         while (!list_empty(&info->unused_bgs)) {
3286                 block_group = list_first_entry(&info->unused_bgs,
3287                                                struct btrfs_block_group,
3288                                                bg_list);
3289                 list_del_init(&block_group->bg_list);
3290                 btrfs_put_block_group(block_group);
3291         }
3292         spin_unlock(&info->unused_bgs_lock);
3293
3294         spin_lock(&info->block_group_cache_lock);
3295         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
3296                 block_group = rb_entry(n, struct btrfs_block_group,
3297                                        cache_node);
3298                 rb_erase(&block_group->cache_node,
3299                          &info->block_group_cache_tree);
3300                 RB_CLEAR_NODE(&block_group->cache_node);
3301                 spin_unlock(&info->block_group_cache_lock);
3302
3303                 down_write(&block_group->space_info->groups_sem);
3304                 list_del(&block_group->list);
3305                 up_write(&block_group->space_info->groups_sem);
3306
3307                 /*
3308                  * We haven't cached this block group, which means we could
3309                  * possibly have excluded extents on this block group.
3310                  */
3311                 if (block_group->cached == BTRFS_CACHE_NO ||
3312                     block_group->cached == BTRFS_CACHE_ERROR)
3313                         btrfs_free_excluded_extents(block_group);
3314
3315                 btrfs_remove_free_space_cache(block_group);
3316                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
3317                 ASSERT(list_empty(&block_group->dirty_list));
3318                 ASSERT(list_empty(&block_group->io_list));
3319                 ASSERT(list_empty(&block_group->bg_list));
3320                 ASSERT(atomic_read(&block_group->count) == 1);
3321                 btrfs_put_block_group(block_group);
3322
3323                 spin_lock(&info->block_group_cache_lock);
3324         }
3325         spin_unlock(&info->block_group_cache_lock);
3326
3327         /*
3328          * Now that all the block groups are freed, go through and free all the
3329          * space_info structs.  This is only called during the final stages of
3330          * unmount, and so we know nobody is using them.  We call
3331          * synchronize_rcu() once before we start, just to be on the safe side.
3332          */
3333         synchronize_rcu();
3334
3335         btrfs_release_global_block_rsv(info);
3336
3337         while (!list_empty(&info->space_info)) {
3338                 space_info = list_entry(info->space_info.next,
3339                                         struct btrfs_space_info,
3340                                         list);
3341
3342                 /*
3343                  * Do not hide this behind enospc_debug, this is actually
3344                  * important and indicates a real bug if this happens.
3345                  */
3346                 if (WARN_ON(space_info->bytes_pinned > 0 ||
3347                             space_info->bytes_reserved > 0 ||
3348                             space_info->bytes_may_use > 0))
3349                         btrfs_dump_space_info(info, space_info, 0, 0);
3350                 list_del(&space_info->list);
3351                 btrfs_sysfs_remove_space_info(space_info);
3352         }
3353         return 0;
3354 }