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 "volumes.h"
  11 #include "transaction.h"
  12 #include "ref-verify.h"
  13 #include "sysfs.h"
  14 #include "tree-log.h"
  15 #include "delalloc-space.h"
  16 #include "discard.h"
  17 #include "raid56.h"
  18 #include "zoned.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                 spin_unlock(&fs_info->balance_lock);
  70                 return extended_to_chunk(target);
  71         }
  72         spin_unlock(&fs_info->balance_lock);
  73
  74         /* First, mask out the RAID levels which aren't possible */
  75         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
  76                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
  77                         allowed |= btrfs_raid_array[raid_type].bg_flag;
  78         }
  79         allowed &= flags;
  80
  81         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
  82                 allowed = BTRFS_BLOCK_GROUP_RAID6;
  83         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
  84                 allowed = BTRFS_BLOCK_GROUP_RAID5;
  85         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
  86                 allowed = BTRFS_BLOCK_GROUP_RAID10;
  87         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
  88                 allowed = BTRFS_BLOCK_GROUP_RAID1;
  89         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
  90                 allowed = BTRFS_BLOCK_GROUP_RAID0;
  91
  92         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
  93
  94         return extended_to_chunk(flags | allowed);
  95 }
  96
  97 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
  98 {
  99         unsigned seq;
 100         u64 flags;
 101
 102         do {
 103                 flags = orig_flags;
 104                 seq = read_seqbegin(&fs_info->profiles_lock);
 105
 106                 if (flags & BTRFS_BLOCK_GROUP_DATA)
 107                         flags |= fs_info->avail_data_alloc_bits;
 108                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 109                         flags |= fs_info->avail_system_alloc_bits;
 110                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
 111                         flags |= fs_info->avail_metadata_alloc_bits;
 112         } while (read_seqretry(&fs_info->profiles_lock, seq));
 113
 114         return btrfs_reduce_alloc_profile(fs_info, flags);
 115 }
 116
 117 void btrfs_get_block_group(struct btrfs_block_group *cache)
 118 {
 119         refcount_inc(&cache->refs);
 120 }
 121
 122 void btrfs_put_block_group(struct btrfs_block_group *cache)
 123 {
 124         if (refcount_dec_and_test(&cache->refs)) {
 125                 WARN_ON(cache->pinned > 0);
 126                 WARN_ON(cache->reserved > 0);
 127
 128                 /*
 129                  * A block_group shouldn't be on the discard_list anymore.
 130                  * Remove the block_group from the discard_list to prevent us
 131                  * from causing a panic due to NULL pointer dereference.
 132                  */
 133                 if (WARN_ON(!list_empty(&cache->discard_list)))
 134                         btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
 135                                                   cache);
 136
 137                 /*
 138                  * If not empty, someone is still holding mutex of
 139                  * full_stripe_lock, which can only be released by caller.
 140                  * And it will definitely cause use-after-free when caller
 141                  * tries to release full stripe lock.
 142                  *
 143                  * No better way to resolve, but only to warn.
 144                  */
 145                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 146                 kfree(cache->free_space_ctl);
 147                 kfree(cache);
 148         }
 149 }
 150
 151 /*
 152  * This adds the block group to the fs_info rb tree for the block group cache
 153  */
 154 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 155                                        struct btrfs_block_group *block_group)
 156 {
 157         struct rb_node **p;
 158         struct rb_node *parent = NULL;
 159         struct btrfs_block_group *cache;
 160
 161         ASSERT(block_group->length != 0);
 162
 163         spin_lock(&info->block_group_cache_lock);
 164         p = &info->block_group_cache_tree.rb_node;
 165
 166         while (*p) {
 167                 parent = *p;
 168                 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
 169                 if (block_group->start < cache->start) {
 170                         p = &(*p)->rb_left;
 171                 } else if (block_group->start > cache->start) {
 172                         p = &(*p)->rb_right;
 173                 } else {
 174                         spin_unlock(&info->block_group_cache_lock);
 175                         return -EEXIST;
 176                 }
 177         }
 178
 179         rb_link_node(&block_group->cache_node, parent, p);
 180         rb_insert_color(&block_group->cache_node,
 181                         &info->block_group_cache_tree);
 182
 183         if (info->first_logical_byte > block_group->start)
 184                 info->first_logical_byte = block_group->start;
 185
 186         spin_unlock(&info->block_group_cache_lock);
 187
 188         return 0;
 189 }
 190
 191 /*
 192  * This will return the block group at or after bytenr if contains is 0, else
 193  * it will return the block group that contains the bytenr
 194  */
 195 static struct btrfs_block_group *block_group_cache_tree_search(
 196                 struct btrfs_fs_info *info, u64 bytenr, int contains)
 197 {
 198         struct btrfs_block_group *cache, *ret = NULL;
 199         struct rb_node *n;
 200         u64 end, start;
 201
 202         spin_lock(&info->block_group_cache_lock);
 203         n = info->block_group_cache_tree.rb_node;
 204
 205         while (n) {
 206                 cache = rb_entry(n, struct btrfs_block_group, cache_node);
 207                 end = cache->start + cache->length - 1;
 208                 start = cache->start;
 209
 210                 if (bytenr < start) {
 211                         if (!contains && (!ret || start < ret->start))
 212                                 ret = cache;
 213                         n = n->rb_left;
 214                 } else if (bytenr > start) {
 215                         if (contains && bytenr <= end) {
 216                                 ret = cache;
 217                                 break;
 218                         }
 219                         n = n->rb_right;
 220                 } else {
 221                         ret = cache;
 222                         break;
 223                 }
 224         }
 225         if (ret) {
 226                 btrfs_get_block_group(ret);
 227                 if (bytenr == 0 && info->first_logical_byte > ret->start)
 228                         info->first_logical_byte = ret->start;
 229         }
 230         spin_unlock(&info->block_group_cache_lock);
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * Return the block group that starts at or after bytenr
 237  */
 238 struct btrfs_block_group *btrfs_lookup_first_block_group(
 239                 struct btrfs_fs_info *info, u64 bytenr)
 240 {
 241         return block_group_cache_tree_search(info, bytenr, 0);
 242 }
 243
 244 /*
 245  * Return the block group that contains the given bytenr
 246  */
 247 struct btrfs_block_group *btrfs_lookup_block_group(
 248                 struct btrfs_fs_info *info, u64 bytenr)
 249 {
 250         return block_group_cache_tree_search(info, bytenr, 1);
 251 }
 252
 253 struct btrfs_block_group *btrfs_next_block_group(
 254                 struct btrfs_block_group *cache)
 255 {
 256         struct btrfs_fs_info *fs_info = cache->fs_info;
 257         struct rb_node *node;
 258
 259         spin_lock(&fs_info->block_group_cache_lock);
 260
 261         /* If our block group was removed, we need a full search. */
 262         if (RB_EMPTY_NODE(&cache->cache_node)) {
 263                 const u64 next_bytenr = cache->start + cache->length;
 264
 265                 spin_unlock(&fs_info->block_group_cache_lock);
 266                 btrfs_put_block_group(cache);
 267                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
 268         }
 269         node = rb_next(&cache->cache_node);
 270         btrfs_put_block_group(cache);
 271         if (node) {
 272                 cache = rb_entry(node, struct btrfs_block_group, cache_node);
 273                 btrfs_get_block_group(cache);
 274         } else
 275                 cache = NULL;
 276         spin_unlock(&fs_info->block_group_cache_lock);
 277         return cache;
 278 }
 279
 280 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
 281 {
 282         struct btrfs_block_group *bg;
 283         bool ret = true;
 284
 285         bg = btrfs_lookup_block_group(fs_info, bytenr);
 286         if (!bg)
 287                 return false;
 288
 289         spin_lock(&bg->lock);
 290         if (bg->ro)
 291                 ret = false;
 292         else
 293                 atomic_inc(&bg->nocow_writers);
 294         spin_unlock(&bg->lock);
 295
 296         /* No put on block group, done by btrfs_dec_nocow_writers */
 297         if (!ret)
 298                 btrfs_put_block_group(bg);
 299
 300         return ret;
 301 }
 302
 303 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
 304 {
 305         struct btrfs_block_group *bg;
 306
 307         bg = btrfs_lookup_block_group(fs_info, bytenr);
 308         ASSERT(bg);
 309         if (atomic_dec_and_test(&bg->nocow_writers))
 310                 wake_up_var(&bg->nocow_writers);
 311         /*
 312          * Once for our lookup and once for the lookup done by a previous call
 313          * to btrfs_inc_nocow_writers()
 314          */
 315         btrfs_put_block_group(bg);
 316         btrfs_put_block_group(bg);
 317 }
 318
 319 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
 320 {
 321         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
 322 }
 323
 324 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
 325                                         const u64 start)
 326 {
 327         struct btrfs_block_group *bg;
 328
 329         bg = btrfs_lookup_block_group(fs_info, start);
 330         ASSERT(bg);
 331         if (atomic_dec_and_test(&bg->reservations))
 332                 wake_up_var(&bg->reservations);
 333         btrfs_put_block_group(bg);
 334 }
 335
 336 void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
 337 {
 338         struct btrfs_space_info *space_info = bg->space_info;
 339
 340         ASSERT(bg->ro);
 341
 342         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
 343                 return;
 344
 345         /*
 346          * Our block group is read only but before we set it to read only,
 347          * some task might have had allocated an extent from it already, but it
 348          * has not yet created a respective ordered extent (and added it to a
 349          * root's list of ordered extents).
 350          * Therefore wait for any task currently allocating extents, since the
 351          * block group's reservations counter is incremented while a read lock
 352          * on the groups' semaphore is held and decremented after releasing
 353          * the read access on that semaphore and creating the ordered extent.
 354          */
 355         down_write(&space_info->groups_sem);
 356         up_write(&space_info->groups_sem);
 357
 358         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
 359 }
 360
 361 struct btrfs_caching_control *btrfs_get_caching_control(
 362                 struct btrfs_block_group *cache)
 363 {
 364         struct btrfs_caching_control *ctl;
 365
 366         spin_lock(&cache->lock);
 367         if (!cache->caching_ctl) {
 368                 spin_unlock(&cache->lock);
 369                 return NULL;
 370         }
 371
 372         ctl = cache->caching_ctl;
 373         refcount_inc(&ctl->count);
 374         spin_unlock(&cache->lock);
 375         return ctl;
 376 }
 377
 378 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
 379 {
 380         if (refcount_dec_and_test(&ctl->count))
 381                 kfree(ctl);
 382 }
 383
 384 /*
 385  * When we wait for progress in the block group caching, its because our
 386  * allocation attempt failed at least once.  So, we must sleep and let some
 387  * progress happen before we try again.
 388  *
 389  * This function will sleep at least once waiting for new free space to show
 390  * up, and then it will check the block group free space numbers for our min
 391  * num_bytes.  Another option is to have it go ahead and look in the rbtree for
 392  * a free extent of a given size, but this is a good start.
 393  *
 394  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
 395  * any of the information in this block group.
 396  */
 397 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
 398                                            u64 num_bytes)
 399 {
 400         struct btrfs_caching_control *caching_ctl;
 401
 402         caching_ctl = btrfs_get_caching_control(cache);
 403         if (!caching_ctl)
 404                 return;
 405
 406         wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
 407                    (cache->free_space_ctl->free_space >= num_bytes));
 408
 409         btrfs_put_caching_control(caching_ctl);
 410 }
 411
 412 int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
 413 {
 414         struct btrfs_caching_control *caching_ctl;
 415         int ret = 0;
 416
 417         caching_ctl = btrfs_get_caching_control(cache);
 418         if (!caching_ctl)
 419                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
 420
 421         wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
 422         if (cache->cached == BTRFS_CACHE_ERROR)
 423                 ret = -EIO;
 424         btrfs_put_caching_control(caching_ctl);
 425         return ret;
 426 }
 427
 428 static bool space_cache_v1_done(struct btrfs_block_group *cache)
 429 {
 430         bool ret;
 431
 432         spin_lock(&cache->lock);
 433         ret = cache->cached != BTRFS_CACHE_FAST;
 434         spin_unlock(&cache->lock);
 435
 436         return ret;
 437 }
 438
 439 void btrfs_wait_space_cache_v1_finished(struct btrfs_block_group *cache,
 440                                 struct btrfs_caching_control *caching_ctl)
 441 {
 442         wait_event(caching_ctl->wait, space_cache_v1_done(cache));
 443 }
 444
 445 #ifdef CONFIG_BTRFS_DEBUG
 446 static void fragment_free_space(struct btrfs_block_group *block_group)
 447 {
 448         struct btrfs_fs_info *fs_info = block_group->fs_info;
 449         u64 start = block_group->start;
 450         u64 len = block_group->length;
 451         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 452                 fs_info->nodesize : fs_info->sectorsize;
 453         u64 step = chunk << 1;
 454
 455         while (len > chunk) {
 456                 btrfs_remove_free_space(block_group, start, chunk);
 457                 start += step;
 458                 if (len < step)
 459                         len = 0;
 460                 else
 461                         len -= step;
 462         }
 463 }
 464 #endif
 465
 466 /*
 467  * This is only called by btrfs_cache_block_group, since we could have freed
 468  * extents we need to check the pinned_extents for any extents that can't be
 469  * used yet since their free space will be released as soon as the transaction
 470  * commits.
 471  */
 472 u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
 473 {
 474         struct btrfs_fs_info *info = block_group->fs_info;
 475         u64 extent_start, extent_end, size, total_added = 0;
 476         int ret;
 477
 478         while (start < end) {
 479                 ret = find_first_extent_bit(&info->excluded_extents, start,
 480                                             &extent_start, &extent_end,
 481                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 482                                             NULL);
 483                 if (ret)
 484                         break;
 485
 486                 if (extent_start <= start) {
 487                         start = extent_end + 1;
 488                 } else if (extent_start > start && extent_start < end) {
 489                         size = extent_start - start;
 490                         total_added += size;
 491                         ret = btrfs_add_free_space_async_trimmed(block_group,
 492                                                                  start, size);
 493                         BUG_ON(ret); /* -ENOMEM or logic error */
 494                         start = extent_end + 1;
 495                 } else {
 496                         break;
 497                 }
 498         }
 499
 500         if (start < end) {
 501                 size = end - start;
 502                 total_added += size;
 503                 ret = btrfs_add_free_space_async_trimmed(block_group, start,
 504                                                          size);
 505                 BUG_ON(ret); /* -ENOMEM or logic error */
 506         }
 507
 508         return total_added;
 509 }
 510
 511 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 512 {
 513         struct btrfs_block_group *block_group = caching_ctl->block_group;
 514         struct btrfs_fs_info *fs_info = block_group->fs_info;
 515         struct btrfs_root *extent_root = fs_info->extent_root;
 516         struct btrfs_path *path;
 517         struct extent_buffer *leaf;
 518         struct btrfs_key key;
 519         u64 total_found = 0;
 520         u64 last = 0;
 521         u32 nritems;
 522         int ret;
 523         bool wakeup = true;
 524
 525         path = btrfs_alloc_path();
 526         if (!path)
 527                 return -ENOMEM;
 528
 529         last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
 530
 531 #ifdef CONFIG_BTRFS_DEBUG
 532         /*
 533          * If we're fragmenting we don't want to make anybody think we can
 534          * allocate from this block group until we've had a chance to fragment
 535          * the free space.
 536          */
 537         if (btrfs_should_fragment_free_space(block_group))
 538                 wakeup = false;
 539 #endif
 540         /*
 541          * We don't want to deadlock with somebody trying to allocate a new
 542          * extent for the extent root while also trying to search the extent
 543          * root to add free space.  So we skip locking and search the commit
 544          * root, since its read-only
 545          */
 546         path->skip_locking = 1;
 547         path->search_commit_root = 1;
 548         path->reada = READA_FORWARD;
 549
 550         key.objectid = last;
 551         key.offset = 0;
 552         key.type = BTRFS_EXTENT_ITEM_KEY;
 553
 554 next:
 555         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 556         if (ret < 0)
 557                 goto out;
 558
 559         leaf = path->nodes[0];
 560         nritems = btrfs_header_nritems(leaf);
 561
 562         while (1) {
 563                 if (btrfs_fs_closing(fs_info) > 1) {
 564                         last = (u64)-1;
 565                         break;
 566                 }
 567
 568                 if (path->slots[0] < nritems) {
 569                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 570                 } else {
 571                         ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
 572                         if (ret)
 573                                 break;
 574
 575                         if (need_resched() ||
 576                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 577                                 if (wakeup)
 578                                         caching_ctl->progress = last;
 579                                 btrfs_release_path(path);
 580                                 up_read(&fs_info->commit_root_sem);
 581                                 mutex_unlock(&caching_ctl->mutex);
 582                                 cond_resched();
 583                                 mutex_lock(&caching_ctl->mutex);
 584                                 down_read(&fs_info->commit_root_sem);
 585                                 goto next;
 586                         }
 587
 588                         ret = btrfs_next_leaf(extent_root, path);
 589                         if (ret < 0)
 590                                 goto out;
 591                         if (ret)
 592                                 break;
 593                         leaf = path->nodes[0];
 594                         nritems = btrfs_header_nritems(leaf);
 595                         continue;
 596                 }
 597
 598                 if (key.objectid < last) {
 599                         key.objectid = last;
 600                         key.offset = 0;
 601                         key.type = BTRFS_EXTENT_ITEM_KEY;
 602
 603                         if (wakeup)
 604                                 caching_ctl->progress = last;
 605                         btrfs_release_path(path);
 606                         goto next;
 607                 }
 608
 609                 if (key.objectid < block_group->start) {
 610                         path->slots[0]++;
 611                         continue;
 612                 }
 613
 614                 if (key.objectid >= block_group->start + block_group->length)
 615                         break;
 616
 617                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 618                     key.type == BTRFS_METADATA_ITEM_KEY) {
 619                         total_found += add_new_free_space(block_group, last,
 620                                                           key.objectid);
 621                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 622                                 last = key.objectid +
 623                                         fs_info->nodesize;
 624                         else
 625                                 last = key.objectid + key.offset;
 626
 627                         if (total_found > CACHING_CTL_WAKE_UP) {
 628                                 total_found = 0;
 629                                 if (wakeup)
 630                                         wake_up(&caching_ctl->wait);
 631                         }
 632                 }
 633                 path->slots[0]++;
 634         }
 635         ret = 0;
 636
 637         total_found += add_new_free_space(block_group, last,
 638                                 block_group->start + block_group->length);
 639         caching_ctl->progress = (u64)-1;
 640
 641 out:
 642         btrfs_free_path(path);
 643         return ret;
 644 }
 645
 646 static noinline void caching_thread(struct btrfs_work *work)
 647 {
 648         struct btrfs_block_group *block_group;
 649         struct btrfs_fs_info *fs_info;
 650         struct btrfs_caching_control *caching_ctl;
 651         int ret;
 652
 653         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 654         block_group = caching_ctl->block_group;
 655         fs_info = block_group->fs_info;
 656
 657         mutex_lock(&caching_ctl->mutex);
 658         down_read(&fs_info->commit_root_sem);
 659
 660         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 661                 ret = load_free_space_cache(block_group);
 662                 if (ret == 1) {
 663                         ret = 0;
 664                         goto done;
 665                 }
 666
 667                 /*
 668                  * We failed to load the space cache, set ourselves to
 669                  * CACHE_STARTED and carry on.
 670                  */
 671                 spin_lock(&block_group->lock);
 672                 block_group->cached = BTRFS_CACHE_STARTED;
 673                 spin_unlock(&block_group->lock);
 674                 wake_up(&caching_ctl->wait);
 675         }
 676
 677         /*
 678          * If we are in the transaction that populated the free space tree we
 679          * can't actually cache from the free space tree as our commit root and
 680          * real root are the same, so we could change the contents of the blocks
 681          * while caching.  Instead do the slow caching in this case, and after
 682          * the transaction has committed we will be safe.
 683          */
 684         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
 685             !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
 686                 ret = load_free_space_tree(caching_ctl);
 687         else
 688                 ret = load_extent_tree_free(caching_ctl);
 689 done:
 690         spin_lock(&block_group->lock);
 691         block_group->caching_ctl = NULL;
 692         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 693         spin_unlock(&block_group->lock);
 694
 695 #ifdef CONFIG_BTRFS_DEBUG
 696         if (btrfs_should_fragment_free_space(block_group)) {
 697                 u64 bytes_used;
 698
 699                 spin_lock(&block_group->space_info->lock);
 700                 spin_lock(&block_group->lock);
 701                 bytes_used = block_group->length - block_group->used;
 702                 block_group->space_info->bytes_used += bytes_used >> 1;
 703                 spin_unlock(&block_group->lock);
 704                 spin_unlock(&block_group->space_info->lock);
 705                 fragment_free_space(block_group);
 706         }
 707 #endif
 708
 709         caching_ctl->progress = (u64)-1;
 710
 711         up_read(&fs_info->commit_root_sem);
 712         btrfs_free_excluded_extents(block_group);
 713         mutex_unlock(&caching_ctl->mutex);
 714
 715         wake_up(&caching_ctl->wait);
 716
 717         btrfs_put_caching_control(caching_ctl);
 718         btrfs_put_block_group(block_group);
 719 }
 720
 721 int btrfs_cache_block_group(struct btrfs_block_group *cache, int load_cache_only)
 722 {
 723         DEFINE_WAIT(wait);
 724         struct btrfs_fs_info *fs_info = cache->fs_info;
 725         struct btrfs_caching_control *caching_ctl = NULL;
 726         int ret = 0;
 727
 728         /* Allocator for zoned filesystems does not use the cache at all */
 729         if (btrfs_is_zoned(fs_info))
 730                 return 0;
 731
 732         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 733         if (!caching_ctl)
 734                 return -ENOMEM;
 735
 736         INIT_LIST_HEAD(&caching_ctl->list);
 737         mutex_init(&caching_ctl->mutex);
 738         init_waitqueue_head(&caching_ctl->wait);
 739         caching_ctl->block_group = cache;
 740         caching_ctl->progress = cache->start;
 741         refcount_set(&caching_ctl->count, 2);
 742         btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 743
 744         spin_lock(&cache->lock);
 745         if (cache->cached != BTRFS_CACHE_NO) {
 746                 kfree(caching_ctl);
 747
 748                 caching_ctl = cache->caching_ctl;
 749                 if (caching_ctl)
 750                         refcount_inc(&caching_ctl->count);
 751                 spin_unlock(&cache->lock);
 752                 goto out;
 753         }
 754         WARN_ON(cache->caching_ctl);
 755         cache->caching_ctl = caching_ctl;
 756         if (btrfs_test_opt(fs_info, SPACE_CACHE))
 757                 cache->cached = BTRFS_CACHE_FAST;
 758         else
 759                 cache->cached = BTRFS_CACHE_STARTED;
 760         cache->has_caching_ctl = 1;
 761         spin_unlock(&cache->lock);
 762
 763         spin_lock(&fs_info->block_group_cache_lock);
 764         refcount_inc(&caching_ctl->count);
 765         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 766         spin_unlock(&fs_info->block_group_cache_lock);
 767
 768         btrfs_get_block_group(cache);
 769
 770         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 771 out:
 772         if (load_cache_only && caching_ctl)
 773                 btrfs_wait_space_cache_v1_finished(cache, caching_ctl);
 774         if (caching_ctl)
 775                 btrfs_put_caching_control(caching_ctl);
 776
 777         return ret;
 778 }
 779
 780 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
 781 {
 782         u64 extra_flags = chunk_to_extended(flags) &
 783                                 BTRFS_EXTENDED_PROFILE_MASK;
 784
 785         write_seqlock(&fs_info->profiles_lock);
 786         if (flags & BTRFS_BLOCK_GROUP_DATA)
 787                 fs_info->avail_data_alloc_bits &= ~extra_flags;
 788         if (flags & BTRFS_BLOCK_GROUP_METADATA)
 789                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
 790         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 791                 fs_info->avail_system_alloc_bits &= ~extra_flags;
 792         write_sequnlock(&fs_info->profiles_lock);
 793 }
 794
 795 /*
 796  * Clear incompat bits for the following feature(s):
 797  *
 798  * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
 799  *            in the whole filesystem
 800  *
 801  * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
 802  */
 803 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
 804 {
 805         bool found_raid56 = false;
 806         bool found_raid1c34 = false;
 807
 808         if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
 809             (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
 810             (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
 811                 struct list_head *head = &fs_info->space_info;
 812                 struct btrfs_space_info *sinfo;
 813
 814                 list_for_each_entry_rcu(sinfo, head, list) {
 815                         down_read(&sinfo->groups_sem);
 816                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
 817                                 found_raid56 = true;
 818                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
 819                                 found_raid56 = true;
 820                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
 821                                 found_raid1c34 = true;
 822                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
 823                                 found_raid1c34 = true;
 824                         up_read(&sinfo->groups_sem);
 825                 }
 826                 if (!found_raid56)
 827                         btrfs_clear_fs_incompat(fs_info, RAID56);
 828                 if (!found_raid1c34)
 829                         btrfs_clear_fs_incompat(fs_info, RAID1C34);
 830         }
 831 }
 832
 833 static int remove_block_group_item(struct btrfs_trans_handle *trans,
 834                                    struct btrfs_path *path,
 835                                    struct btrfs_block_group *block_group)
 836 {
 837         struct btrfs_fs_info *fs_info = trans->fs_info;
 838         struct btrfs_root *root;
 839         struct btrfs_key key;
 840         int ret;
 841
 842         root = fs_info->extent_root;
 843         key.objectid = block_group->start;
 844         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
 845         key.offset = block_group->length;
 846
 847         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 848         if (ret > 0)
 849                 ret = -ENOENT;
 850         if (ret < 0)
 851                 return ret;
 852
 853         ret = btrfs_del_item(trans, root, path);
 854         return ret;
 855 }
 856
 857 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 858                              u64 group_start, struct extent_map *em)
 859 {
 860         struct btrfs_fs_info *fs_info = trans->fs_info;
 861         struct btrfs_path *path;
 862         struct btrfs_block_group *block_group;
 863         struct btrfs_free_cluster *cluster;
 864         struct inode *inode;
 865         struct kobject *kobj = NULL;
 866         int ret;
 867         int index;
 868         int factor;
 869         struct btrfs_caching_control *caching_ctl = NULL;
 870         bool remove_em;
 871         bool remove_rsv = false;
 872
 873         block_group = btrfs_lookup_block_group(fs_info, group_start);
 874         BUG_ON(!block_group);
 875         BUG_ON(!block_group->ro);
 876
 877         trace_btrfs_remove_block_group(block_group);
 878         /*
 879          * Free the reserved super bytes from this block group before
 880          * remove it.
 881          */
 882         btrfs_free_excluded_extents(block_group);
 883         btrfs_free_ref_tree_range(fs_info, block_group->start,
 884                                   block_group->length);
 885
 886         index = btrfs_bg_flags_to_raid_index(block_group->flags);
 887         factor = btrfs_bg_type_to_factor(block_group->flags);
 888
 889         /* make sure this block group isn't part of an allocation cluster */
 890         cluster = &fs_info->data_alloc_cluster;
 891         spin_lock(&cluster->refill_lock);
 892         btrfs_return_cluster_to_free_space(block_group, cluster);
 893         spin_unlock(&cluster->refill_lock);
 894
 895         /*
 896          * make sure this block group isn't part of a metadata
 897          * allocation cluster
 898          */
 899         cluster = &fs_info->meta_alloc_cluster;
 900         spin_lock(&cluster->refill_lock);
 901         btrfs_return_cluster_to_free_space(block_group, cluster);
 902         spin_unlock(&cluster->refill_lock);
 903
 904         btrfs_clear_treelog_bg(block_group);
 905
 906         path = btrfs_alloc_path();
 907         if (!path) {
 908                 ret = -ENOMEM;
 909                 goto out;
 910         }
 911
 912         /*
 913          * get the inode first so any iput calls done for the io_list
 914          * aren't the final iput (no unlinks allowed now)
 915          */
 916         inode = lookup_free_space_inode(block_group, path);
 917
 918         mutex_lock(&trans->transaction->cache_write_mutex);
 919         /*
 920          * Make sure our free space cache IO is done before removing the
 921          * free space inode
 922          */
 923         spin_lock(&trans->transaction->dirty_bgs_lock);
 924         if (!list_empty(&block_group->io_list)) {
 925                 list_del_init(&block_group->io_list);
 926
 927                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
 928
 929                 spin_unlock(&trans->transaction->dirty_bgs_lock);
 930                 btrfs_wait_cache_io(trans, block_group, path);
 931                 btrfs_put_block_group(block_group);
 932                 spin_lock(&trans->transaction->dirty_bgs_lock);
 933         }
 934
 935         if (!list_empty(&block_group->dirty_list)) {
 936                 list_del_init(&block_group->dirty_list);
 937                 remove_rsv = true;
 938                 btrfs_put_block_group(block_group);
 939         }
 940         spin_unlock(&trans->transaction->dirty_bgs_lock);
 941         mutex_unlock(&trans->transaction->cache_write_mutex);
 942
 943         ret = btrfs_remove_free_space_inode(trans, inode, block_group);
 944         if (ret)
 945                 goto out;
 946
 947         spin_lock(&fs_info->block_group_cache_lock);
 948         rb_erase(&block_group->cache_node,
 949                  &fs_info->block_group_cache_tree);
 950         RB_CLEAR_NODE(&block_group->cache_node);
 951
 952         /* Once for the block groups rbtree */
 953         btrfs_put_block_group(block_group);
 954
 955         if (fs_info->first_logical_byte == block_group->start)
 956                 fs_info->first_logical_byte = (u64)-1;
 957         spin_unlock(&fs_info->block_group_cache_lock);
 958
 959         down_write(&block_group->space_info->groups_sem);
 960         /*
 961          * we must use list_del_init so people can check to see if they
 962          * are still on the list after taking the semaphore
 963          */
 964         list_del_init(&block_group->list);
 965         if (list_empty(&block_group->space_info->block_groups[index])) {
 966                 kobj = block_group->space_info->block_group_kobjs[index];
 967                 block_group->space_info->block_group_kobjs[index] = NULL;
 968                 clear_avail_alloc_bits(fs_info, block_group->flags);
 969         }
 970         up_write(&block_group->space_info->groups_sem);
 971         clear_incompat_bg_bits(fs_info, block_group->flags);
 972         if (kobj) {
 973                 kobject_del(kobj);
 974                 kobject_put(kobj);
 975         }
 976
 977         if (block_group->has_caching_ctl)
 978                 caching_ctl = btrfs_get_caching_control(block_group);
 979         if (block_group->cached == BTRFS_CACHE_STARTED)
 980                 btrfs_wait_block_group_cache_done(block_group);
 981         if (block_group->has_caching_ctl) {
 982                 spin_lock(&fs_info->block_group_cache_lock);
 983                 if (!caching_ctl) {
 984                         struct btrfs_caching_control *ctl;
 985
 986                         list_for_each_entry(ctl,
 987                                     &fs_info->caching_block_groups, list)
 988                                 if (ctl->block_group == block_group) {
 989                                         caching_ctl = ctl;
 990                                         refcount_inc(&caching_ctl->count);
 991                                         break;
 992                                 }
 993                 }
 994                 if (caching_ctl)
 995                         list_del_init(&caching_ctl->list);
 996                 spin_unlock(&fs_info->block_group_cache_lock);
 997                 if (caching_ctl) {
 998                         /* Once for the caching bgs list and once for us. */
 999                         btrfs_put_caching_control(caching_ctl);
1000                         btrfs_put_caching_control(caching_ctl);
1001                 }
1002         }
1003
1004         spin_lock(&trans->transaction->dirty_bgs_lock);
1005         WARN_ON(!list_empty(&block_group->dirty_list));
1006         WARN_ON(!list_empty(&block_group->io_list));
1007         spin_unlock(&trans->transaction->dirty_bgs_lock);
1008
1009         btrfs_remove_free_space_cache(block_group);
1010
1011         spin_lock(&block_group->space_info->lock);
1012         list_del_init(&block_group->ro_list);
1013
1014         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1015                 WARN_ON(block_group->space_info->total_bytes
1016                         < block_group->length);
1017                 WARN_ON(block_group->space_info->bytes_readonly
1018                         < block_group->length - block_group->zone_unusable);
1019                 WARN_ON(block_group->space_info->bytes_zone_unusable
1020                         < block_group->zone_unusable);
1021                 WARN_ON(block_group->space_info->disk_total
1022                         < block_group->length * factor);
1023         }
1024         block_group->space_info->total_bytes -= block_group->length;
1025         block_group->space_info->bytes_readonly -=
1026                 (block_group->length - block_group->zone_unusable);
1027         block_group->space_info->bytes_zone_unusable -=
1028                 block_group->zone_unusable;
1029         block_group->space_info->disk_total -= block_group->length * factor;
1030
1031         spin_unlock(&block_group->space_info->lock);
1032
1033         /*
1034          * Remove the free space for the block group from the free space tree
1035          * and the block group's item from the extent tree before marking the
1036          * block group as removed. This is to prevent races with tasks that
1037          * freeze and unfreeze a block group, this task and another task
1038          * allocating a new block group - the unfreeze task ends up removing
1039          * the block group's extent map before the task calling this function
1040          * deletes the block group item from the extent tree, allowing for
1041          * another task to attempt to create another block group with the same
1042          * item key (and failing with -EEXIST and a transaction abort).
1043          */
1044         ret = remove_block_group_free_space(trans, block_group);
1045         if (ret)
1046                 goto out;
1047
1048         ret = remove_block_group_item(trans, path, block_group);
1049         if (ret < 0)
1050                 goto out;
1051
1052         spin_lock(&block_group->lock);
1053         block_group->removed = 1;
1054         /*
1055          * At this point trimming or scrub can't start on this block group,
1056          * because we removed the block group from the rbtree
1057          * fs_info->block_group_cache_tree so no one can't find it anymore and
1058          * even if someone already got this block group before we removed it
1059          * from the rbtree, they have already incremented block_group->frozen -
1060          * if they didn't, for the trimming case they won't find any free space
1061          * entries because we already removed them all when we called
1062          * btrfs_remove_free_space_cache().
1063          *
1064          * And we must not remove the extent map from the fs_info->mapping_tree
1065          * to prevent the same logical address range and physical device space
1066          * ranges from being reused for a new block group. This is needed to
1067          * avoid races with trimming and scrub.
1068          *
1069          * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1070          * completely transactionless, so while it is trimming a range the
1071          * currently running transaction might finish and a new one start,
1072          * allowing for new block groups to be created that can reuse the same
1073          * physical device locations unless we take this special care.
1074          *
1075          * There may also be an implicit trim operation if the file system
1076          * is mounted with -odiscard. The same protections must remain
1077          * in place until the extents have been discarded completely when
1078          * the transaction commit has completed.
1079          */
1080         remove_em = (atomic_read(&block_group->frozen) == 0);
1081         spin_unlock(&block_group->lock);
1082
1083         if (remove_em) {
1084                 struct extent_map_tree *em_tree;
1085
1086                 em_tree = &fs_info->mapping_tree;
1087                 write_lock(&em_tree->lock);
1088                 remove_extent_mapping(em_tree, em);
1089                 write_unlock(&em_tree->lock);
1090                 /* once for the tree */
1091                 free_extent_map(em);
1092         }
1093
1094 out:
1095         /* Once for the lookup reference */
1096         btrfs_put_block_group(block_group);
1097         if (remove_rsv)
1098                 btrfs_delayed_refs_rsv_release(fs_info, 1);
1099         btrfs_free_path(path);
1100         return ret;
1101 }
1102
1103 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1104                 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1105 {
1106         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1107         struct extent_map *em;
1108         struct map_lookup *map;
1109         unsigned int num_items;
1110
1111         read_lock(&em_tree->lock);
1112         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1113         read_unlock(&em_tree->lock);
1114         ASSERT(em && em->start == chunk_offset);
1115
1116         /*
1117          * We need to reserve 3 + N units from the metadata space info in order
1118          * to remove a block group (done at btrfs_remove_chunk() and at
1119          * btrfs_remove_block_group()), which are used for:
1120          *
1121          * 1 unit for adding the free space inode's orphan (located in the tree
1122          * of tree roots).
1123          * 1 unit for deleting the block group item (located in the extent
1124          * tree).
1125          * 1 unit for deleting the free space item (located in tree of tree
1126          * roots).
1127          * N units for deleting N device extent items corresponding to each
1128          * stripe (located in the device tree).
1129          *
1130          * In order to remove a block group we also need to reserve units in the
1131          * system space info in order to update the chunk tree (update one or
1132          * more device items and remove one chunk item), but this is done at
1133          * btrfs_remove_chunk() through a call to check_system_chunk().
1134          */
1135         map = em->map_lookup;
1136         num_items = 3 + map->num_stripes;
1137         free_extent_map(em);
1138
1139         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
1140                                                            num_items);
1141 }
1142
1143 /*
1144  * Mark block group @cache read-only, so later write won't happen to block
1145  * group @cache.
1146  *
1147  * If @force is not set, this function will only mark the block group readonly
1148  * if we have enough free space (1M) in other metadata/system block groups.
1149  * If @force is not set, this function will mark the block group readonly
1150  * without checking free space.
1151  *
1152  * NOTE: This function doesn't care if other block groups can contain all the
1153  * data in this block group. That check should be done by relocation routine,
1154  * not this function.
1155  */
1156 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1157 {
1158         struct btrfs_space_info *sinfo = cache->space_info;
1159         u64 num_bytes;
1160         int ret = -ENOSPC;
1161
1162         spin_lock(&sinfo->lock);
1163         spin_lock(&cache->lock);
1164
1165         if (cache->swap_extents) {
1166                 ret = -ETXTBSY;
1167                 goto out;
1168         }
1169
1170         if (cache->ro) {
1171                 cache->ro++;
1172                 ret = 0;
1173                 goto out;
1174         }
1175
1176         num_bytes = cache->length - cache->reserved - cache->pinned -
1177                     cache->bytes_super - cache->zone_unusable - cache->used;
1178
1179         /*
1180          * Data never overcommits, even in mixed mode, so do just the straight
1181          * check of left over space in how much we have allocated.
1182          */
1183         if (force) {
1184                 ret = 0;
1185         } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1186                 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1187
1188                 /*
1189                  * Here we make sure if we mark this bg RO, we still have enough
1190                  * free space as buffer.
1191                  */
1192                 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1193                         ret = 0;
1194         } else {
1195                 /*
1196                  * We overcommit metadata, so we need to do the
1197                  * btrfs_can_overcommit check here, and we need to pass in
1198                  * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1199                  * leeway to allow us to mark this block group as read only.
1200                  */
1201                 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1202                                          BTRFS_RESERVE_NO_FLUSH))
1203                         ret = 0;
1204         }
1205
1206         if (!ret) {
1207                 sinfo->bytes_readonly += num_bytes;
1208                 if (btrfs_is_zoned(cache->fs_info)) {
1209                         /* Migrate zone_unusable bytes to readonly */
1210                         sinfo->bytes_readonly += cache->zone_unusable;
1211                         sinfo->bytes_zone_unusable -= cache->zone_unusable;
1212                         cache->zone_unusable = 0;
1213                 }
1214                 cache->ro++;
1215                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1216         }
1217 out:
1218         spin_unlock(&cache->lock);
1219         spin_unlock(&sinfo->lock);
1220         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1221                 btrfs_info(cache->fs_info,
1222                         "unable to make block group %llu ro", cache->start);
1223                 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1224         }
1225         return ret;
1226 }
1227
1228 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1229                                  struct btrfs_block_group *bg)
1230 {
1231         struct btrfs_fs_info *fs_info = bg->fs_info;
1232         struct btrfs_transaction *prev_trans = NULL;
1233         const u64 start = bg->start;
1234         const u64 end = start + bg->length - 1;
1235         int ret;
1236
1237         spin_lock(&fs_info->trans_lock);
1238         if (trans->transaction->list.prev != &fs_info->trans_list) {
1239                 prev_trans = list_last_entry(&trans->transaction->list,
1240                                              struct btrfs_transaction, list);
1241                 refcount_inc(&prev_trans->use_count);
1242         }
1243         spin_unlock(&fs_info->trans_lock);
1244
1245         /*
1246          * Hold the unused_bg_unpin_mutex lock to avoid racing with
1247          * btrfs_finish_extent_commit(). If we are at transaction N, another
1248          * task might be running finish_extent_commit() for the previous
1249          * transaction N - 1, and have seen a range belonging to the block
1250          * group in pinned_extents before we were able to clear the whole block
1251          * group range from pinned_extents. This means that task can lookup for
1252          * the block group after we unpinned it from pinned_extents and removed
1253          * it, leading to a BUG_ON() at unpin_extent_range().
1254          */
1255         mutex_lock(&fs_info->unused_bg_unpin_mutex);
1256         if (prev_trans) {
1257                 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1258                                         EXTENT_DIRTY);
1259                 if (ret)
1260                         goto out;
1261         }
1262
1263         ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1264                                 EXTENT_DIRTY);
1265 out:
1266         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1267         if (prev_trans)
1268                 btrfs_put_transaction(prev_trans);
1269
1270         return ret == 0;
1271 }
1272
1273 /*
1274  * Process the unused_bgs list and remove any that don't have any allocated
1275  * space inside of them.
1276  */
1277 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1278 {
1279         struct btrfs_block_group *block_group;
1280         struct btrfs_space_info *space_info;
1281         struct btrfs_trans_handle *trans;
1282         const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1283         int ret = 0;
1284
1285         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1286                 return;
1287
1288         /*
1289          * Long running balances can keep us blocked here for eternity, so
1290          * simply skip deletion if we're unable to get the mutex.
1291          */
1292         if (!mutex_trylock(&fs_info->delete_unused_bgs_mutex))
1293                 return;
1294
1295         spin_lock(&fs_info->unused_bgs_lock);
1296         while (!list_empty(&fs_info->unused_bgs)) {
1297                 int trimming;
1298
1299                 block_group = list_first_entry(&fs_info->unused_bgs,
1300                                                struct btrfs_block_group,
1301                                                bg_list);
1302                 list_del_init(&block_group->bg_list);
1303
1304                 space_info = block_group->space_info;
1305
1306                 if (ret || btrfs_mixed_space_info(space_info)) {
1307                         btrfs_put_block_group(block_group);
1308                         continue;
1309                 }
1310                 spin_unlock(&fs_info->unused_bgs_lock);
1311
1312                 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1313
1314                 /* Don't want to race with allocators so take the groups_sem */
1315                 down_write(&space_info->groups_sem);
1316
1317                 /*
1318                  * Async discard moves the final block group discard to be prior
1319                  * to the unused_bgs code path.  Therefore, if it's not fully
1320                  * trimmed, punt it back to the async discard lists.
1321                  */
1322                 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1323                     !btrfs_is_free_space_trimmed(block_group)) {
1324                         trace_btrfs_skip_unused_block_group(block_group);
1325                         up_write(&space_info->groups_sem);
1326                         /* Requeue if we failed because of async discard */
1327                         btrfs_discard_queue_work(&fs_info->discard_ctl,
1328                                                  block_group);
1329                         goto next;
1330                 }
1331
1332                 spin_lock(&block_group->lock);
1333                 if (block_group->reserved || block_group->pinned ||
1334                     block_group->used || block_group->ro ||
1335                     list_is_singular(&block_group->list)) {
1336                         /*
1337                          * We want to bail if we made new allocations or have
1338                          * outstanding allocations in this block group.  We do
1339                          * the ro check in case balance is currently acting on
1340                          * this block group.
1341                          */
1342                         trace_btrfs_skip_unused_block_group(block_group);
1343                         spin_unlock(&block_group->lock);
1344                         up_write(&space_info->groups_sem);
1345                         goto next;
1346                 }
1347                 spin_unlock(&block_group->lock);
1348
1349                 /* We don't want to force the issue, only flip if it's ok. */
1350                 ret = inc_block_group_ro(block_group, 0);
1351                 up_write(&space_info->groups_sem);
1352                 if (ret < 0) {
1353                         ret = 0;
1354                         goto next;
1355                 }
1356
1357                 /*
1358                  * Want to do this before we do anything else so we can recover
1359                  * properly if we fail to join the transaction.
1360                  */
1361                 trans = btrfs_start_trans_remove_block_group(fs_info,
1362                                                      block_group->start);
1363                 if (IS_ERR(trans)) {
1364                         btrfs_dec_block_group_ro(block_group);
1365                         ret = PTR_ERR(trans);
1366                         goto next;
1367                 }
1368
1369                 /*
1370                  * We could have pending pinned extents for this block group,
1371                  * just delete them, we don't care about them anymore.
1372                  */
1373                 if (!clean_pinned_extents(trans, block_group)) {
1374                         btrfs_dec_block_group_ro(block_group);
1375                         goto end_trans;
1376                 }
1377
1378                 /*
1379                  * At this point, the block_group is read only and should fail
1380                  * new allocations.  However, btrfs_finish_extent_commit() can
1381                  * cause this block_group to be placed back on the discard
1382                  * lists because now the block_group isn't fully discarded.
1383                  * Bail here and try again later after discarding everything.
1384                  */
1385                 spin_lock(&fs_info->discard_ctl.lock);
1386                 if (!list_empty(&block_group->discard_list)) {
1387                         spin_unlock(&fs_info->discard_ctl.lock);
1388                         btrfs_dec_block_group_ro(block_group);
1389                         btrfs_discard_queue_work(&fs_info->discard_ctl,
1390                                                  block_group);
1391                         goto end_trans;
1392                 }
1393                 spin_unlock(&fs_info->discard_ctl.lock);
1394
1395                 /* Reset pinned so btrfs_put_block_group doesn't complain */
1396                 spin_lock(&space_info->lock);
1397                 spin_lock(&block_group->lock);
1398
1399                 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1400                                                      -block_group->pinned);
1401                 space_info->bytes_readonly += block_group->pinned;
1402                 __btrfs_mod_total_bytes_pinned(space_info, -block_group->pinned);
1403                 block_group->pinned = 0;
1404
1405                 spin_unlock(&block_group->lock);
1406                 spin_unlock(&space_info->lock);
1407
1408                 /*
1409                  * The normal path here is an unused block group is passed here,
1410                  * then trimming is handled in the transaction commit path.
1411                  * Async discard interposes before this to do the trimming
1412                  * before coming down the unused block group path as trimming
1413                  * will no longer be done later in the transaction commit path.
1414                  */
1415                 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1416                         goto flip_async;
1417
1418                 /*
1419                  * DISCARD can flip during remount. On zoned filesystems, we
1420                  * need to reset sequential-required zones.
1421                  */
1422                 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1423                                 btrfs_is_zoned(fs_info);
1424
1425                 /* Implicit trim during transaction commit. */
1426                 if (trimming)
1427                         btrfs_freeze_block_group(block_group);
1428
1429                 /*
1430                  * Btrfs_remove_chunk will abort the transaction if things go
1431                  * horribly wrong.
1432                  */
1433                 ret = btrfs_remove_chunk(trans, block_group->start);
1434
1435                 if (ret) {
1436                         if (trimming)
1437                                 btrfs_unfreeze_block_group(block_group);
1438                         goto end_trans;
1439                 }
1440
1441                 /*
1442                  * If we're not mounted with -odiscard, we can just forget
1443                  * about this block group. Otherwise we'll need to wait
1444                  * until transaction commit to do the actual discard.
1445                  */
1446                 if (trimming) {
1447                         spin_lock(&fs_info->unused_bgs_lock);
1448                         /*
1449                          * A concurrent scrub might have added us to the list
1450                          * fs_info->unused_bgs, so use a list_move operation
1451                          * to add the block group to the deleted_bgs list.
1452                          */
1453                         list_move(&block_group->bg_list,
1454                                   &trans->transaction->deleted_bgs);
1455                         spin_unlock(&fs_info->unused_bgs_lock);
1456                         btrfs_get_block_group(block_group);
1457                 }
1458 end_trans:
1459                 btrfs_end_transaction(trans);
1460 next:
1461                 btrfs_put_block_group(block_group);
1462                 spin_lock(&fs_info->unused_bgs_lock);
1463         }
1464         spin_unlock(&fs_info->unused_bgs_lock);
1465         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1466         return;
1467
1468 flip_async:
1469         btrfs_end_transaction(trans);
1470         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1471         btrfs_put_block_group(block_group);
1472         btrfs_discard_punt_unused_bgs_list(fs_info);
1473 }
1474
1475 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1476 {
1477         struct btrfs_fs_info *fs_info = bg->fs_info;
1478
1479         spin_lock(&fs_info->unused_bgs_lock);
1480         if (list_empty(&bg->bg_list)) {
1481                 btrfs_get_block_group(bg);
1482                 trace_btrfs_add_unused_block_group(bg);
1483                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1484         }
1485         spin_unlock(&fs_info->unused_bgs_lock);
1486 }
1487
1488 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1489                            struct btrfs_path *path)
1490 {
1491         struct extent_map_tree *em_tree;
1492         struct extent_map *em;
1493         struct btrfs_block_group_item bg;
1494         struct extent_buffer *leaf;
1495         int slot;
1496         u64 flags;
1497         int ret = 0;
1498
1499         slot = path->slots[0];
1500         leaf = path->nodes[0];
1501
1502         em_tree = &fs_info->mapping_tree;
1503         read_lock(&em_tree->lock);
1504         em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1505         read_unlock(&em_tree->lock);
1506         if (!em) {
1507                 btrfs_err(fs_info,
1508                           "logical %llu len %llu found bg but no related chunk",
1509                           key->objectid, key->offset);
1510                 return -ENOENT;
1511         }
1512
1513         if (em->start != key->objectid || em->len != key->offset) {
1514                 btrfs_err(fs_info,
1515                         "block group %llu len %llu mismatch with chunk %llu len %llu",
1516                         key->objectid, key->offset, em->start, em->len);
1517                 ret = -EUCLEAN;
1518                 goto out_free_em;
1519         }
1520
1521         read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1522                            sizeof(bg));
1523         flags = btrfs_stack_block_group_flags(&bg) &
1524                 BTRFS_BLOCK_GROUP_TYPE_MASK;
1525
1526         if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1527                 btrfs_err(fs_info,
1528 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1529                           key->objectid, key->offset, flags,
1530                           (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1531                 ret = -EUCLEAN;
1532         }
1533
1534 out_free_em:
1535         free_extent_map(em);
1536         return ret;
1537 }
1538
1539 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1540                                   struct btrfs_path *path,
1541                                   struct btrfs_key *key)
1542 {
1543         struct btrfs_root *root = fs_info->extent_root;
1544         int ret;
1545         struct btrfs_key found_key;
1546         struct extent_buffer *leaf;
1547         int slot;
1548
1549         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1550         if (ret < 0)
1551                 return ret;
1552
1553         while (1) {
1554                 slot = path->slots[0];
1555                 leaf = path->nodes[0];
1556                 if (slot >= btrfs_header_nritems(leaf)) {
1557                         ret = btrfs_next_leaf(root, path);
1558                         if (ret == 0)
1559                                 continue;
1560                         if (ret < 0)
1561                                 goto out;
1562                         break;
1563                 }
1564                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1565
1566                 if (found_key.objectid >= key->objectid &&
1567                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1568                         ret = read_bg_from_eb(fs_info, &found_key, path);
1569                         break;
1570                 }
1571
1572                 path->slots[0]++;
1573         }
1574 out:
1575         return ret;
1576 }
1577
1578 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1579 {
1580         u64 extra_flags = chunk_to_extended(flags) &
1581                                 BTRFS_EXTENDED_PROFILE_MASK;
1582
1583         write_seqlock(&fs_info->profiles_lock);
1584         if (flags & BTRFS_BLOCK_GROUP_DATA)
1585                 fs_info->avail_data_alloc_bits |= extra_flags;
1586         if (flags & BTRFS_BLOCK_GROUP_METADATA)
1587                 fs_info->avail_metadata_alloc_bits |= extra_flags;
1588         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1589                 fs_info->avail_system_alloc_bits |= extra_flags;
1590         write_sequnlock(&fs_info->profiles_lock);
1591 }
1592
1593 /**
1594  * Map a physical disk address to a list of logical addresses
1595  *
1596  * @fs_info:       the filesystem
1597  * @chunk_start:   logical address of block group
1598  * @bdev:          physical device to resolve, can be NULL to indicate any device
1599  * @physical:      physical address to map to logical addresses
1600  * @logical:       return array of logical addresses which map to @physical
1601  * @naddrs:        length of @logical
1602  * @stripe_len:    size of IO stripe for the given block group
1603  *
1604  * Maps a particular @physical disk address to a list of @logical addresses.
1605  * Used primarily to exclude those portions of a block group that contain super
1606  * block copies.
1607  */
1608 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1609                      struct block_device *bdev, u64 physical, u64 **logical,
1610                      int *naddrs, int *stripe_len)
1611 {
1612         struct extent_map *em;
1613         struct map_lookup *map;
1614         u64 *buf;
1615         u64 bytenr;
1616         u64 data_stripe_length;
1617         u64 io_stripe_size;
1618         int i, nr = 0;
1619         int ret = 0;
1620
1621         em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1622         if (IS_ERR(em))
1623                 return -EIO;
1624
1625         map = em->map_lookup;
1626         data_stripe_length = em->orig_block_len;
1627         io_stripe_size = map->stripe_len;
1628         chunk_start = em->start;
1629
1630         /* For RAID5/6 adjust to a full IO stripe length */
1631         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1632                 io_stripe_size = map->stripe_len * nr_data_stripes(map);
1633
1634         buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1635         if (!buf) {
1636                 ret = -ENOMEM;
1637                 goto out;
1638         }
1639
1640         for (i = 0; i < map->num_stripes; i++) {
1641                 bool already_inserted = false;
1642                 u64 stripe_nr;
1643                 u64 offset;
1644                 int j;
1645
1646                 if (!in_range(physical, map->stripes[i].physical,
1647                               data_stripe_length))
1648                         continue;
1649
1650                 if (bdev && map->stripes[i].dev->bdev != bdev)
1651                         continue;
1652
1653                 stripe_nr = physical - map->stripes[i].physical;
1654                 stripe_nr = div64_u64_rem(stripe_nr, map->stripe_len, &offset);
1655
1656                 if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1657                         stripe_nr = stripe_nr * map->num_stripes + i;
1658                         stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1659                 } else if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1660                         stripe_nr = stripe_nr * map->num_stripes + i;
1661                 }
1662                 /*
1663                  * The remaining case would be for RAID56, multiply by
1664                  * nr_data_stripes().  Alternatively, just use rmap_len below
1665                  * instead of map->stripe_len
1666                  */
1667
1668                 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
1669
1670                 /* Ensure we don't add duplicate addresses */
1671                 for (j = 0; j < nr; j++) {
1672                         if (buf[j] == bytenr) {
1673                                 already_inserted = true;
1674                                 break;
1675                         }
1676                 }
1677
1678                 if (!already_inserted)
1679                         buf[nr++] = bytenr;
1680         }
1681
1682         *logical = buf;
1683         *naddrs = nr;
1684         *stripe_len = io_stripe_size;
1685 out:
1686         free_extent_map(em);
1687         return ret;
1688 }
1689
1690 static int exclude_super_stripes(struct btrfs_block_group *cache)
1691 {
1692         struct btrfs_fs_info *fs_info = cache->fs_info;
1693         const bool zoned = btrfs_is_zoned(fs_info);
1694         u64 bytenr;
1695         u64 *logical;
1696         int stripe_len;
1697         int i, nr, ret;
1698
1699         if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1700                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1701                 cache->bytes_super += stripe_len;
1702                 ret = btrfs_add_excluded_extent(fs_info, cache->start,
1703                                                 stripe_len);
1704                 if (ret)
1705                         return ret;
1706         }
1707
1708         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1709                 bytenr = btrfs_sb_offset(i);
1710                 ret = btrfs_rmap_block(fs_info, cache->start, NULL,
1711                                        bytenr, &logical, &nr, &stripe_len);
1712                 if (ret)
1713                         return ret;
1714
1715                 /* Shouldn't have super stripes in sequential zones */
1716                 if (zoned && nr) {
1717                         btrfs_err(fs_info,
1718                         "zoned: block group %llu must not contain super block",
1719                                   cache->start);
1720                         return -EUCLEAN;
1721                 }
1722
1723                 while (nr--) {
1724                         u64 len = min_t(u64, stripe_len,
1725                                 cache->start + cache->length - logical[nr]);
1726
1727                         cache->bytes_super += len;
1728                         ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1729                                                         len);
1730                         if (ret) {
1731                                 kfree(logical);
1732                                 return ret;
1733                         }
1734                 }
1735
1736                 kfree(logical);
1737         }
1738         return 0;
1739 }
1740
1741 static void link_block_group(struct btrfs_block_group *cache)
1742 {
1743         struct btrfs_space_info *space_info = cache->space_info;
1744         int index = btrfs_bg_flags_to_raid_index(cache->flags);
1745
1746         down_write(&space_info->groups_sem);
1747         list_add_tail(&cache->list, &space_info->block_groups[index]);
1748         up_write(&space_info->groups_sem);
1749 }
1750
1751 static struct btrfs_block_group *btrfs_create_block_group_cache(
1752                 struct btrfs_fs_info *fs_info, u64 start)
1753 {
1754         struct btrfs_block_group *cache;
1755
1756         cache = kzalloc(sizeof(*cache), GFP_NOFS);
1757         if (!cache)
1758                 return NULL;
1759
1760         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1761                                         GFP_NOFS);
1762         if (!cache->free_space_ctl) {
1763                 kfree(cache);
1764                 return NULL;
1765         }
1766
1767         cache->start = start;
1768
1769         cache->fs_info = fs_info;
1770         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1771
1772         cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1773
1774         refcount_set(&cache->refs, 1);
1775         spin_lock_init(&cache->lock);
1776         init_rwsem(&cache->data_rwsem);
1777         INIT_LIST_HEAD(&cache->list);
1778         INIT_LIST_HEAD(&cache->cluster_list);
1779         INIT_LIST_HEAD(&cache->bg_list);
1780         INIT_LIST_HEAD(&cache->ro_list);
1781         INIT_LIST_HEAD(&cache->discard_list);
1782         INIT_LIST_HEAD(&cache->dirty_list);
1783         INIT_LIST_HEAD(&cache->io_list);
1784         btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1785         atomic_set(&cache->frozen, 0);
1786         mutex_init(&cache->free_space_lock);
1787         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
1788
1789         return cache;
1790 }
1791
1792 /*
1793  * Iterate all chunks and verify that each of them has the corresponding block
1794  * group
1795  */
1796 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
1797 {
1798         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
1799         struct extent_map *em;
1800         struct btrfs_block_group *bg;
1801         u64 start = 0;
1802         int ret = 0;
1803
1804         while (1) {
1805                 read_lock(&map_tree->lock);
1806                 /*
1807                  * lookup_extent_mapping will return the first extent map
1808                  * intersecting the range, so setting @len to 1 is enough to
1809                  * get the first chunk.
1810                  */
1811                 em = lookup_extent_mapping(map_tree, start, 1);
1812                 read_unlock(&map_tree->lock);
1813                 if (!em)
1814                         break;
1815
1816                 bg = btrfs_lookup_block_group(fs_info, em->start);
1817                 if (!bg) {
1818                         btrfs_err(fs_info,
1819         "chunk start=%llu len=%llu doesn't have corresponding block group",
1820                                      em->start, em->len);
1821                         ret = -EUCLEAN;
1822                         free_extent_map(em);
1823                         break;
1824                 }
1825                 if (bg->start != em->start || bg->length != em->len ||
1826                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
1827                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1828                         btrfs_err(fs_info,
1829 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1830                                 em->start, em->len,
1831                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
1832                                 bg->start, bg->length,
1833                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
1834                         ret = -EUCLEAN;
1835                         free_extent_map(em);
1836                         btrfs_put_block_group(bg);
1837                         break;
1838                 }
1839                 start = em->start + em->len;
1840                 free_extent_map(em);
1841                 btrfs_put_block_group(bg);
1842         }
1843         return ret;
1844 }
1845
1846 static int read_one_block_group(struct btrfs_fs_info *info,
1847                                 struct btrfs_block_group_item *bgi,
1848                                 const struct btrfs_key *key,
1849                                 int need_clear)
1850 {
1851         struct btrfs_block_group *cache;
1852         struct btrfs_space_info *space_info;
1853         const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
1854         int ret;
1855
1856         ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
1857
1858         cache = btrfs_create_block_group_cache(info, key->objectid);
1859         if (!cache)
1860                 return -ENOMEM;
1861
1862         cache->length = key->offset;
1863         cache->used = btrfs_stack_block_group_used(bgi);
1864         cache->flags = btrfs_stack_block_group_flags(bgi);
1865
1866         set_free_space_tree_thresholds(cache);
1867
1868         if (need_clear) {
1869                 /*
1870                  * When we mount with old space cache, we need to
1871                  * set BTRFS_DC_CLEAR and set dirty flag.
1872                  *
1873                  * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
1874                  *    truncate the old free space cache inode and
1875                  *    setup a new one.
1876                  * b) Setting 'dirty flag' makes sure that we flush
1877                  *    the new space cache info onto disk.
1878                  */
1879                 if (btrfs_test_opt(info, SPACE_CACHE))
1880                         cache->disk_cache_state = BTRFS_DC_CLEAR;
1881         }
1882         if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
1883             (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
1884                         btrfs_err(info,
1885 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
1886                                   cache->start);
1887                         ret = -EINVAL;
1888                         goto error;
1889         }
1890
1891         ret = btrfs_load_block_group_zone_info(cache, false);
1892         if (ret) {
1893                 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
1894                           cache->start);
1895                 goto error;
1896         }
1897
1898         /*
1899          * We need to exclude the super stripes now so that the space info has
1900          * super bytes accounted for, otherwise we'll think we have more space
1901          * than we actually do.
1902          */
1903         ret = exclude_super_stripes(cache);
1904         if (ret) {
1905                 /* We may have excluded something, so call this just in case. */
1906                 btrfs_free_excluded_extents(cache);
1907                 goto error;
1908         }
1909
1910         /*
1911          * For zoned filesystem, space after the allocation offset is the only
1912          * free space for a block group. So, we don't need any caching work.
1913          * btrfs_calc_zone_unusable() will set the amount of free space and
1914          * zone_unusable space.
1915          *
1916          * For regular filesystem, check for two cases, either we are full, and
1917          * therefore don't need to bother with the caching work since we won't
1918          * find any space, or we are empty, and we can just add all the space
1919          * in and be done with it.  This saves us _a_lot_ of time, particularly
1920          * in the full case.
1921          */
1922         if (btrfs_is_zoned(info)) {
1923                 btrfs_calc_zone_unusable(cache);
1924         } else if (cache->length == cache->used) {
1925                 cache->last_byte_to_unpin = (u64)-1;
1926                 cache->cached = BTRFS_CACHE_FINISHED;
1927                 btrfs_free_excluded_extents(cache);
1928         } else if (cache->used == 0) {
1929                 cache->last_byte_to_unpin = (u64)-1;
1930                 cache->cached = BTRFS_CACHE_FINISHED;
1931                 add_new_free_space(cache, cache->start,
1932                                    cache->start + cache->length);
1933                 btrfs_free_excluded_extents(cache);
1934         }
1935
1936         ret = btrfs_add_block_group_cache(info, cache);
1937         if (ret) {
1938                 btrfs_remove_free_space_cache(cache);
1939                 goto error;
1940         }
1941         trace_btrfs_add_block_group(info, cache, 0);
1942         btrfs_update_space_info(info, cache->flags, cache->length,
1943                                 cache->used, cache->bytes_super,
1944                                 cache->zone_unusable, &space_info);
1945
1946         cache->space_info = space_info;
1947
1948         link_block_group(cache);
1949
1950         set_avail_alloc_bits(info, cache->flags);
1951         if (btrfs_chunk_readonly(info, cache->start)) {
1952                 inc_block_group_ro(cache, 1);
1953         } else if (cache->used == 0) {
1954                 ASSERT(list_empty(&cache->bg_list));
1955                 if (btrfs_test_opt(info, DISCARD_ASYNC))
1956                         btrfs_discard_queue_work(&info->discard_ctl, cache);
1957                 else
1958                         btrfs_mark_bg_unused(cache);
1959         }
1960         return 0;
1961 error:
1962         btrfs_put_block_group(cache);
1963         return ret;
1964 }
1965
1966 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
1967 {
1968         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1969         struct btrfs_space_info *space_info;
1970         struct rb_node *node;
1971         int ret = 0;
1972
1973         for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
1974                 struct extent_map *em;
1975                 struct map_lookup *map;
1976                 struct btrfs_block_group *bg;
1977
1978                 em = rb_entry(node, struct extent_map, rb_node);
1979                 map = em->map_lookup;
1980                 bg = btrfs_create_block_group_cache(fs_info, em->start);
1981                 if (!bg) {
1982                         ret = -ENOMEM;
1983                         break;
1984                 }
1985
1986                 /* Fill dummy cache as FULL */
1987                 bg->length = em->len;
1988                 bg->flags = map->type;
1989                 bg->last_byte_to_unpin = (u64)-1;
1990                 bg->cached = BTRFS_CACHE_FINISHED;
1991                 bg->used = em->len;
1992                 bg->flags = map->type;
1993                 ret = btrfs_add_block_group_cache(fs_info, bg);
1994                 if (ret) {
1995                         btrfs_remove_free_space_cache(bg);
1996                         btrfs_put_block_group(bg);
1997                         break;
1998                 }
1999                 btrfs_update_space_info(fs_info, bg->flags, em->len, em->len,
2000                                         0, 0, &space_info);
2001                 bg->space_info = space_info;
2002                 link_block_group(bg);
2003
2004                 set_avail_alloc_bits(fs_info, bg->flags);
2005         }
2006         if (!ret)
2007                 btrfs_init_global_block_rsv(fs_info);
2008         return ret;
2009 }
2010
2011 int btrfs_read_block_groups(struct btrfs_fs_info *info)
2012 {
2013         struct btrfs_path *path;
2014         int ret;
2015         struct btrfs_block_group *cache;
2016         struct btrfs_space_info *space_info;
2017         struct btrfs_key key;
2018         int need_clear = 0;
2019         u64 cache_gen;
2020
2021         if (!info->extent_root)
2022                 return fill_dummy_bgs(info);
2023
2024         key.objectid = 0;
2025         key.offset = 0;
2026         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2027         path = btrfs_alloc_path();
2028         if (!path)
2029                 return -ENOMEM;
2030
2031         cache_gen = btrfs_super_cache_generation(info->super_copy);
2032         if (btrfs_test_opt(info, SPACE_CACHE) &&
2033             btrfs_super_generation(info->super_copy) != cache_gen)
2034                 need_clear = 1;
2035         if (btrfs_test_opt(info, CLEAR_CACHE))
2036                 need_clear = 1;
2037
2038         while (1) {
2039                 struct btrfs_block_group_item bgi;
2040                 struct extent_buffer *leaf;
2041                 int slot;
2042
2043                 ret = find_first_block_group(info, path, &key);
2044                 if (ret > 0)
2045                         break;
2046                 if (ret != 0)
2047                         goto error;
2048
2049                 leaf = path->nodes[0];
2050                 slot = path->slots[0];
2051
2052                 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2053                                    sizeof(bgi));
2054
2055                 btrfs_item_key_to_cpu(leaf, &key, slot);
2056                 btrfs_release_path(path);
2057                 ret = read_one_block_group(info, &bgi, &key, need_clear);
2058                 if (ret < 0)
2059                         goto error;
2060                 key.objectid += key.offset;
2061                 key.offset = 0;
2062         }
2063         btrfs_release_path(path);
2064
2065         list_for_each_entry(space_info, &info->space_info, list) {
2066                 int i;
2067
2068                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2069                         if (list_empty(&space_info->block_groups[i]))
2070                                 continue;
2071                         cache = list_first_entry(&space_info->block_groups[i],
2072                                                  struct btrfs_block_group,
2073                                                  list);
2074                         btrfs_sysfs_add_block_group_type(cache);
2075                 }
2076
2077                 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2078                       (BTRFS_BLOCK_GROUP_RAID10 |
2079                        BTRFS_BLOCK_GROUP_RAID1_MASK |
2080                        BTRFS_BLOCK_GROUP_RAID56_MASK |
2081                        BTRFS_BLOCK_GROUP_DUP)))
2082                         continue;
2083                 /*
2084                  * Avoid allocating from un-mirrored block group if there are
2085                  * mirrored block groups.
2086                  */
2087                 list_for_each_entry(cache,
2088                                 &space_info->block_groups[BTRFS_RAID_RAID0],
2089                                 list)
2090                         inc_block_group_ro(cache, 1);
2091                 list_for_each_entry(cache,
2092                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
2093                                 list)
2094                         inc_block_group_ro(cache, 1);
2095         }
2096
2097         btrfs_init_global_block_rsv(info);
2098         ret = check_chunk_block_group_mappings(info);
2099 error:
2100         btrfs_free_path(path);
2101         return ret;
2102 }
2103
2104 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2105                                    struct btrfs_block_group *block_group)
2106 {
2107         struct btrfs_fs_info *fs_info = trans->fs_info;
2108         struct btrfs_block_group_item bgi;
2109         struct btrfs_root *root;
2110         struct btrfs_key key;
2111
2112         spin_lock(&block_group->lock);
2113         btrfs_set_stack_block_group_used(&bgi, block_group->used);
2114         btrfs_set_stack_block_group_chunk_objectid(&bgi,
2115                                 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2116         btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2117         key.objectid = block_group->start;
2118         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2119         key.offset = block_group->length;
2120         spin_unlock(&block_group->lock);
2121
2122         root = fs_info->extent_root;
2123         return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2124 }
2125
2126 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2127 {
2128         struct btrfs_fs_info *fs_info = trans->fs_info;
2129         struct btrfs_block_group *block_group;
2130         int ret = 0;
2131
2132         if (!trans->can_flush_pending_bgs)
2133                 return;
2134
2135         while (!list_empty(&trans->new_bgs)) {
2136                 int index;
2137
2138                 block_group = list_first_entry(&trans->new_bgs,
2139                                                struct btrfs_block_group,
2140                                                bg_list);
2141                 if (ret)
2142                         goto next;
2143
2144                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2145
2146                 ret = insert_block_group_item(trans, block_group);
2147                 if (ret)
2148                         btrfs_abort_transaction(trans, ret);
2149                 ret = btrfs_finish_chunk_alloc(trans, block_group->start,
2150                                         block_group->length);
2151                 if (ret)
2152                         btrfs_abort_transaction(trans, ret);
2153                 add_block_group_free_space(trans, block_group);
2154
2155                 /*
2156                  * If we restriped during balance, we may have added a new raid
2157                  * type, so now add the sysfs entries when it is safe to do so.
2158                  * We don't have to worry about locking here as it's handled in
2159                  * btrfs_sysfs_add_block_group_type.
2160                  */
2161                 if (block_group->space_info->block_group_kobjs[index] == NULL)
2162                         btrfs_sysfs_add_block_group_type(block_group);
2163
2164                 /* Already aborted the transaction if it failed. */
2165 next:
2166                 btrfs_delayed_refs_rsv_release(fs_info, 1);
2167                 list_del_init(&block_group->bg_list);
2168         }
2169         btrfs_trans_release_chunk_metadata(trans);
2170 }
2171
2172 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
2173                            u64 type, u64 chunk_offset, u64 size)
2174 {
2175         struct btrfs_fs_info *fs_info = trans->fs_info;
2176         struct btrfs_block_group *cache;
2177         int ret;
2178
2179         btrfs_set_log_full_commit(trans);
2180
2181         cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2182         if (!cache)
2183                 return -ENOMEM;
2184
2185         cache->length = size;
2186         set_free_space_tree_thresholds(cache);
2187         cache->used = bytes_used;
2188         cache->flags = type;
2189         cache->last_byte_to_unpin = (u64)-1;
2190         cache->cached = BTRFS_CACHE_FINISHED;
2191         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2192                 cache->needs_free_space = 1;
2193
2194         ret = btrfs_load_block_group_zone_info(cache, true);
2195         if (ret) {
2196                 btrfs_put_block_group(cache);
2197                 return ret;
2198         }
2199
2200         ret = exclude_super_stripes(cache);
2201         if (ret) {
2202                 /* We may have excluded something, so call this just in case */
2203                 btrfs_free_excluded_extents(cache);
2204                 btrfs_put_block_group(cache);
2205                 return ret;
2206         }
2207
2208         add_new_free_space(cache, chunk_offset, chunk_offset + size);
2209
2210         btrfs_free_excluded_extents(cache);
2211
2212 #ifdef CONFIG_BTRFS_DEBUG
2213         if (btrfs_should_fragment_free_space(cache)) {
2214                 u64 new_bytes_used = size - bytes_used;
2215
2216                 bytes_used += new_bytes_used >> 1;
2217                 fragment_free_space(cache);
2218         }
2219 #endif
2220         /*
2221          * Ensure the corresponding space_info object is created and
2222          * assigned to our block group. We want our bg to be added to the rbtree
2223          * with its ->space_info set.
2224          */
2225         cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2226         ASSERT(cache->space_info);
2227
2228         ret = btrfs_add_block_group_cache(fs_info, cache);
2229         if (ret) {
2230                 btrfs_remove_free_space_cache(cache);
2231                 btrfs_put_block_group(cache);
2232                 return ret;
2233         }
2234
2235         /*
2236          * Now that our block group has its ->space_info set and is inserted in
2237          * the rbtree, update the space info's counters.
2238          */
2239         trace_btrfs_add_block_group(fs_info, cache, 1);
2240         btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
2241                                 cache->bytes_super, 0, &cache->space_info);
2242         btrfs_update_global_block_rsv(fs_info);
2243
2244         link_block_group(cache);
2245
2246         list_add_tail(&cache->bg_list, &trans->new_bgs);
2247         trans->delayed_ref_updates++;
2248         btrfs_update_delayed_refs_rsv(trans);
2249
2250         set_avail_alloc_bits(fs_info, type);
2251         return 0;
2252 }
2253
2254 /*
2255  * Mark one block group RO, can be called several times for the same block
2256  * group.
2257  *
2258  * @cache:              the destination block group
2259  * @do_chunk_alloc:     whether need to do chunk pre-allocation, this is to
2260  *                      ensure we still have some free space after marking this
2261  *                      block group RO.
2262  */
2263 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2264                              bool do_chunk_alloc)
2265 {
2266         struct btrfs_fs_info *fs_info = cache->fs_info;
2267         struct btrfs_trans_handle *trans;
2268         u64 alloc_flags;
2269         int ret;
2270
2271 again:
2272         trans = btrfs_join_transaction(fs_info->extent_root);
2273         if (IS_ERR(trans))
2274                 return PTR_ERR(trans);
2275
2276         /*
2277          * we're not allowed to set block groups readonly after the dirty
2278          * block groups cache has started writing.  If it already started,
2279          * back off and let this transaction commit
2280          */
2281         mutex_lock(&fs_info->ro_block_group_mutex);
2282         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2283                 u64 transid = trans->transid;
2284
2285                 mutex_unlock(&fs_info->ro_block_group_mutex);
2286                 btrfs_end_transaction(trans);
2287
2288                 ret = btrfs_wait_for_commit(fs_info, transid);
2289                 if (ret)
2290                         return ret;
2291                 goto again;
2292         }
2293
2294         if (do_chunk_alloc) {
2295                 /*
2296                  * If we are changing raid levels, try to allocate a
2297                  * corresponding block group with the new raid level.
2298                  */
2299                 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2300                 if (alloc_flags != cache->flags) {
2301                         ret = btrfs_chunk_alloc(trans, alloc_flags,
2302                                                 CHUNK_ALLOC_FORCE);
2303                         /*
2304                          * ENOSPC is allowed here, we may have enough space
2305                          * already allocated at the new raid level to carry on
2306                          */
2307                         if (ret == -ENOSPC)
2308                                 ret = 0;
2309                         if (ret < 0)
2310                                 goto out;
2311                 }
2312         }
2313
2314         ret = inc_block_group_ro(cache, 0);
2315         if (!do_chunk_alloc || ret == -ETXTBSY)
2316                 goto unlock_out;
2317         if (!ret)
2318                 goto out;
2319         alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2320         ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2321         if (ret < 0)
2322                 goto out;
2323         ret = inc_block_group_ro(cache, 0);
2324         if (ret == -ETXTBSY)
2325                 goto unlock_out;
2326 out:
2327         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2328                 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2329                 mutex_lock(&fs_info->chunk_mutex);
2330                 check_system_chunk(trans, alloc_flags);
2331                 mutex_unlock(&fs_info->chunk_mutex);
2332         }
2333 unlock_out:
2334         mutex_unlock(&fs_info->ro_block_group_mutex);
2335
2336         btrfs_end_transaction(trans);
2337         return ret;
2338 }
2339
2340 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2341 {
2342         struct btrfs_space_info *sinfo = cache->space_info;
2343         u64 num_bytes;
2344
2345         BUG_ON(!cache->ro);
2346
2347         spin_lock(&sinfo->lock);
2348         spin_lock(&cache->lock);
2349         if (!--cache->ro) {
2350                 num_bytes = cache->length - cache->reserved -
2351                             cache->pinned - cache->bytes_super -
2352                             cache->zone_unusable - cache->used;
2353                 sinfo->bytes_readonly -= num_bytes;
2354                 if (btrfs_is_zoned(cache->fs_info)) {
2355                         /* Migrate zone_unusable bytes back */
2356                         cache->zone_unusable = cache->alloc_offset - cache->used;
2357                         sinfo->bytes_zone_unusable += cache->zone_unusable;
2358                         sinfo->bytes_readonly -= cache->zone_unusable;
2359                 }
2360                 list_del_init(&cache->ro_list);
2361         }
2362         spin_unlock(&cache->lock);
2363         spin_unlock(&sinfo->lock);
2364 }
2365
2366 static int update_block_group_item(struct btrfs_trans_handle *trans,
2367                                    struct btrfs_path *path,
2368                                    struct btrfs_block_group *cache)
2369 {
2370         struct btrfs_fs_info *fs_info = trans->fs_info;
2371         int ret;
2372         struct btrfs_root *root = fs_info->extent_root;
2373         unsigned long bi;
2374         struct extent_buffer *leaf;
2375         struct btrfs_block_group_item bgi;
2376         struct btrfs_key key;
2377
2378         key.objectid = cache->start;
2379         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2380         key.offset = cache->length;
2381
2382         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2383         if (ret) {
2384                 if (ret > 0)
2385                         ret = -ENOENT;
2386                 goto fail;
2387         }
2388
2389         leaf = path->nodes[0];
2390         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2391         btrfs_set_stack_block_group_used(&bgi, cache->used);
2392         btrfs_set_stack_block_group_chunk_objectid(&bgi,
2393                         BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2394         btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2395         write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2396         btrfs_mark_buffer_dirty(leaf);
2397 fail:
2398         btrfs_release_path(path);
2399         return ret;
2400
2401 }
2402
2403 static int cache_save_setup(struct btrfs_block_group *block_group,
2404                             struct btrfs_trans_handle *trans,
2405                             struct btrfs_path *path)
2406 {
2407         struct btrfs_fs_info *fs_info = block_group->fs_info;
2408         struct btrfs_root *root = fs_info->tree_root;
2409         struct inode *inode = NULL;
2410         struct extent_changeset *data_reserved = NULL;
2411         u64 alloc_hint = 0;
2412         int dcs = BTRFS_DC_ERROR;
2413         u64 num_pages = 0;
2414         int retries = 0;
2415         int ret = 0;
2416
2417         if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2418                 return 0;
2419
2420         /*
2421          * If this block group is smaller than 100 megs don't bother caching the
2422          * block group.
2423          */
2424         if (block_group->length < (100 * SZ_1M)) {
2425                 spin_lock(&block_group->lock);
2426                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2427                 spin_unlock(&block_group->lock);
2428                 return 0;
2429         }
2430
2431         if (TRANS_ABORTED(trans))
2432                 return 0;
2433 again:
2434         inode = lookup_free_space_inode(block_group, path);
2435         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2436                 ret = PTR_ERR(inode);
2437                 btrfs_release_path(path);
2438                 goto out;
2439         }
2440
2441         if (IS_ERR(inode)) {
2442                 BUG_ON(retries);
2443                 retries++;
2444
2445                 if (block_group->ro)
2446                         goto out_free;
2447
2448                 ret = create_free_space_inode(trans, block_group, path);
2449                 if (ret)
2450                         goto out_free;
2451                 goto again;
2452         }
2453
2454         /*
2455          * We want to set the generation to 0, that way if anything goes wrong
2456          * from here on out we know not to trust this cache when we load up next
2457          * time.
2458          */
2459         BTRFS_I(inode)->generation = 0;
2460         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2461         if (ret) {
2462                 /*
2463                  * So theoretically we could recover from this, simply set the
2464                  * super cache generation to 0 so we know to invalidate the
2465                  * cache, but then we'd have to keep track of the block groups
2466                  * that fail this way so we know we _have_ to reset this cache
2467                  * before the next commit or risk reading stale cache.  So to
2468                  * limit our exposure to horrible edge cases lets just abort the
2469                  * transaction, this only happens in really bad situations
2470                  * anyway.
2471                  */
2472                 btrfs_abort_transaction(trans, ret);
2473                 goto out_put;
2474         }
2475         WARN_ON(ret);
2476
2477         /* We've already setup this transaction, go ahead and exit */
2478         if (block_group->cache_generation == trans->transid &&
2479             i_size_read(inode)) {
2480                 dcs = BTRFS_DC_SETUP;
2481                 goto out_put;
2482         }
2483
2484         if (i_size_read(inode) > 0) {
2485                 ret = btrfs_check_trunc_cache_free_space(fs_info,
2486                                         &fs_info->global_block_rsv);
2487                 if (ret)
2488                         goto out_put;
2489
2490                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2491                 if (ret)
2492                         goto out_put;
2493         }
2494
2495         spin_lock(&block_group->lock);
2496         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2497             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2498                 /*
2499                  * don't bother trying to write stuff out _if_
2500                  * a) we're not cached,
2501                  * b) we're with nospace_cache mount option,
2502                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
2503                  */
2504                 dcs = BTRFS_DC_WRITTEN;
2505                 spin_unlock(&block_group->lock);
2506                 goto out_put;
2507         }
2508         spin_unlock(&block_group->lock);
2509
2510         /*
2511          * We hit an ENOSPC when setting up the cache in this transaction, just
2512          * skip doing the setup, we've already cleared the cache so we're safe.
2513          */
2514         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2515                 ret = -ENOSPC;
2516                 goto out_put;
2517         }
2518
2519         /*
2520          * Try to preallocate enough space based on how big the block group is.
2521          * Keep in mind this has to include any pinned space which could end up
2522          * taking up quite a bit since it's not folded into the other space
2523          * cache.
2524          */
2525         num_pages = div_u64(block_group->length, SZ_256M);
2526         if (!num_pages)
2527                 num_pages = 1;
2528
2529         num_pages *= 16;
2530         num_pages *= PAGE_SIZE;
2531
2532         ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2533                                           num_pages);
2534         if (ret)
2535                 goto out_put;
2536
2537         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2538                                               num_pages, num_pages,
2539                                               &alloc_hint);
2540         /*
2541          * Our cache requires contiguous chunks so that we don't modify a bunch
2542          * of metadata or split extents when writing the cache out, which means
2543          * we can enospc if we are heavily fragmented in addition to just normal
2544          * out of space conditions.  So if we hit this just skip setting up any
2545          * other block groups for this transaction, maybe we'll unpin enough
2546          * space the next time around.
2547          */
2548         if (!ret)
2549                 dcs = BTRFS_DC_SETUP;
2550         else if (ret == -ENOSPC)
2551                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2552
2553 out_put:
2554         iput(inode);
2555 out_free:
2556         btrfs_release_path(path);
2557 out:
2558         spin_lock(&block_group->lock);
2559         if (!ret && dcs == BTRFS_DC_SETUP)
2560                 block_group->cache_generation = trans->transid;
2561         block_group->disk_cache_state = dcs;
2562         spin_unlock(&block_group->lock);
2563
2564         extent_changeset_free(data_reserved);
2565         return ret;
2566 }
2567
2568 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2569 {
2570         struct btrfs_fs_info *fs_info = trans->fs_info;
2571         struct btrfs_block_group *cache, *tmp;
2572         struct btrfs_transaction *cur_trans = trans->transaction;
2573         struct btrfs_path *path;
2574
2575         if (list_empty(&cur_trans->dirty_bgs) ||
2576             !btrfs_test_opt(fs_info, SPACE_CACHE))
2577                 return 0;
2578
2579         path = btrfs_alloc_path();
2580         if (!path)
2581                 return -ENOMEM;
2582
2583         /* Could add new block groups, use _safe just in case */
2584         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
2585                                  dirty_list) {
2586                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2587                         cache_save_setup(cache, trans, path);
2588         }
2589
2590         btrfs_free_path(path);
2591         return 0;
2592 }
2593
2594 /*
2595  * Transaction commit does final block group cache writeback during a critical
2596  * section where nothing is allowed to change the FS.  This is required in
2597  * order for the cache to actually match the block group, but can introduce a
2598  * lot of latency into the commit.
2599  *
2600  * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2601  * There's a chance we'll have to redo some of it if the block group changes
2602  * again during the commit, but it greatly reduces the commit latency by
2603  * getting rid of the easy block groups while we're still allowing others to
2604  * join the commit.
2605  */
2606 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
2607 {
2608         struct btrfs_fs_info *fs_info = trans->fs_info;
2609         struct btrfs_block_group *cache;
2610         struct btrfs_transaction *cur_trans = trans->transaction;
2611         int ret = 0;
2612         int should_put;
2613         struct btrfs_path *path = NULL;
2614         LIST_HEAD(dirty);
2615         struct list_head *io = &cur_trans->io_bgs;
2616         int num_started = 0;
2617         int loops = 0;
2618
2619         spin_lock(&cur_trans->dirty_bgs_lock);
2620         if (list_empty(&cur_trans->dirty_bgs)) {
2621                 spin_unlock(&cur_trans->dirty_bgs_lock);
2622                 return 0;
2623         }
2624         list_splice_init(&cur_trans->dirty_bgs, &dirty);
2625         spin_unlock(&cur_trans->dirty_bgs_lock);
2626
2627 again:
2628         /* Make sure all the block groups on our dirty list actually exist */
2629         btrfs_create_pending_block_groups(trans);
2630
2631         if (!path) {
2632                 path = btrfs_alloc_path();
2633                 if (!path) {
2634                         ret = -ENOMEM;
2635                         goto out;
2636                 }
2637         }
2638
2639         /*
2640          * cache_write_mutex is here only to save us from balance or automatic
2641          * removal of empty block groups deleting this block group while we are
2642          * writing out the cache
2643          */
2644         mutex_lock(&trans->transaction->cache_write_mutex);
2645         while (!list_empty(&dirty)) {
2646                 bool drop_reserve = true;
2647
2648                 cache = list_first_entry(&dirty, struct btrfs_block_group,
2649                                          dirty_list);
2650                 /*
2651                  * This can happen if something re-dirties a block group that
2652                  * is already under IO.  Just wait for it to finish and then do
2653                  * it all again
2654                  */
2655                 if (!list_empty(&cache->io_list)) {
2656                         list_del_init(&cache->io_list);
2657                         btrfs_wait_cache_io(trans, cache, path);
2658                         btrfs_put_block_group(cache);
2659                 }
2660
2661
2662                 /*
2663                  * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2664                  * it should update the cache_state.  Don't delete until after
2665                  * we wait.
2666                  *
2667                  * Since we're not running in the commit critical section
2668                  * we need the dirty_bgs_lock to protect from update_block_group
2669                  */
2670                 spin_lock(&cur_trans->dirty_bgs_lock);
2671                 list_del_init(&cache->dirty_list);
2672                 spin_unlock(&cur_trans->dirty_bgs_lock);
2673
2674                 should_put = 1;
2675
2676                 cache_save_setup(cache, trans, path);
2677
2678                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
2679                         cache->io_ctl.inode = NULL;
2680                         ret = btrfs_write_out_cache(trans, cache, path);
2681                         if (ret == 0 && cache->io_ctl.inode) {
2682                                 num_started++;
2683                                 should_put = 0;
2684
2685                                 /*
2686                                  * The cache_write_mutex is protecting the
2687                                  * io_list, also refer to the definition of
2688                                  * btrfs_transaction::io_bgs for more details
2689                                  */
2690                                 list_add_tail(&cache->io_list, io);
2691                         } else {
2692                                 /*
2693                                  * If we failed to write the cache, the
2694                                  * generation will be bad and life goes on
2695                                  */
2696                                 ret = 0;
2697                         }
2698                 }
2699                 if (!ret) {
2700                         ret = update_block_group_item(trans, path, cache);
2701                         /*
2702                          * Our block group might still be attached to the list
2703                          * of new block groups in the transaction handle of some
2704                          * other task (struct btrfs_trans_handle->new_bgs). This
2705                          * means its block group item isn't yet in the extent
2706                          * tree. If this happens ignore the error, as we will
2707                          * try again later in the critical section of the
2708                          * transaction commit.
2709                          */
2710                         if (ret == -ENOENT) {
2711                                 ret = 0;
2712                                 spin_lock(&cur_trans->dirty_bgs_lock);
2713                                 if (list_empty(&cache->dirty_list)) {
2714                                         list_add_tail(&cache->dirty_list,
2715                                                       &cur_trans->dirty_bgs);
2716                                         btrfs_get_block_group(cache);
2717                                         drop_reserve = false;
2718                                 }
2719                                 spin_unlock(&cur_trans->dirty_bgs_lock);
2720                         } else if (ret) {
2721                                 btrfs_abort_transaction(trans, ret);
2722                         }
2723                 }
2724
2725                 /* If it's not on the io list, we need to put the block group */
2726                 if (should_put)
2727                         btrfs_put_block_group(cache);
2728                 if (drop_reserve)
2729                         btrfs_delayed_refs_rsv_release(fs_info, 1);
2730                 /*
2731                  * Avoid blocking other tasks for too long. It might even save
2732                  * us from writing caches for block groups that are going to be
2733                  * removed.
2734                  */
2735                 mutex_unlock(&trans->transaction->cache_write_mutex);
2736                 if (ret)
2737                         goto out;
2738                 mutex_lock(&trans->transaction->cache_write_mutex);
2739         }
2740         mutex_unlock(&trans->transaction->cache_write_mutex);
2741
2742         /*
2743          * Go through delayed refs for all the stuff we've just kicked off
2744          * and then loop back (just once)
2745          */
2746         if (!ret)
2747                 ret = btrfs_run_delayed_refs(trans, 0);
2748         if (!ret && loops == 0) {
2749                 loops++;
2750                 spin_lock(&cur_trans->dirty_bgs_lock);
2751                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
2752                 /*
2753                  * dirty_bgs_lock protects us from concurrent block group
2754                  * deletes too (not just cache_write_mutex).
2755                  */
2756                 if (!list_empty(&dirty)) {
2757                         spin_unlock(&cur_trans->dirty_bgs_lock);
2758                         goto again;
2759                 }
2760                 spin_unlock(&cur_trans->dirty_bgs_lock);
2761         }
2762 out:
2763         if (ret < 0) {
2764                 spin_lock(&cur_trans->dirty_bgs_lock);
2765                 list_splice_init(&dirty, &cur_trans->dirty_bgs);
2766                 spin_unlock(&cur_trans->dirty_bgs_lock);
2767                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
2768         }
2769
2770         btrfs_free_path(path);
2771         return ret;
2772 }
2773
2774 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
2775 {
2776         struct btrfs_fs_info *fs_info = trans->fs_info;
2777         struct btrfs_block_group *cache;
2778         struct btrfs_transaction *cur_trans = trans->transaction;
2779         int ret = 0;
2780         int should_put;
2781         struct btrfs_path *path;
2782         struct list_head *io = &cur_trans->io_bgs;
2783         int num_started = 0;
2784
2785         path = btrfs_alloc_path();
2786         if (!path)
2787                 return -ENOMEM;
2788
2789         /*
2790          * Even though we are in the critical section of the transaction commit,
2791          * we can still have concurrent tasks adding elements to this
2792          * transaction's list of dirty block groups. These tasks correspond to
2793          * endio free space workers started when writeback finishes for a
2794          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
2795          * allocate new block groups as a result of COWing nodes of the root
2796          * tree when updating the free space inode. The writeback for the space
2797          * caches is triggered by an earlier call to
2798          * btrfs_start_dirty_block_groups() and iterations of the following
2799          * loop.
2800          * Also we want to do the cache_save_setup first and then run the
2801          * delayed refs to make sure we have the best chance at doing this all
2802          * in one shot.
2803          */
2804         spin_lock(&cur_trans->dirty_bgs_lock);
2805         while (!list_empty(&cur_trans->dirty_bgs)) {
2806                 cache = list_first_entry(&cur_trans->dirty_bgs,
2807                                          struct btrfs_block_group,
2808                                          dirty_list);
2809
2810                 /*
2811                  * This can happen if cache_save_setup re-dirties a block group
2812                  * that is already under IO.  Just wait for it to finish and
2813                  * then do it all again
2814                  */
2815                 if (!list_empty(&cache->io_list)) {
2816                         spin_unlock(&cur_trans->dirty_bgs_lock);
2817                         list_del_init(&cache->io_list);
2818                         btrfs_wait_cache_io(trans, cache, path);
2819                         btrfs_put_block_group(cache);
2820                         spin_lock(&cur_trans->dirty_bgs_lock);
2821                 }
2822
2823                 /*
2824                  * Don't remove from the dirty list until after we've waited on
2825                  * any pending IO
2826                  */
2827                 list_del_init(&cache->dirty_list);
2828                 spin_unlock(&cur_trans->dirty_bgs_lock);
2829                 should_put = 1;
2830
2831                 cache_save_setup(cache, trans, path);
2832
2833                 if (!ret)
2834                         ret = btrfs_run_delayed_refs(trans,
2835                                                      (unsigned long) -1);
2836
2837                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
2838                         cache->io_ctl.inode = NULL;
2839                         ret = btrfs_write_out_cache(trans, cache, path);
2840                         if (ret == 0 && cache->io_ctl.inode) {
2841                                 num_started++;
2842                                 should_put = 0;
2843                                 list_add_tail(&cache->io_list, io);
2844                         } else {
2845                                 /*
2846                                  * If we failed to write the cache, the
2847                                  * generation will be bad and life goes on
2848                                  */
2849                                 ret = 0;
2850                         }
2851                 }
2852                 if (!ret) {
2853                         ret = update_block_group_item(trans, path, cache);
2854                         /*
2855                          * One of the free space endio workers might have
2856                          * created a new block group while updating a free space
2857                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
2858                          * and hasn't released its transaction handle yet, in
2859                          * which case the new block group is still attached to
2860                          * its transaction handle and its creation has not
2861                          * finished yet (no block group item in the extent tree
2862                          * yet, etc). If this is the case, wait for all free
2863                          * space endio workers to finish and retry. This is a
2864                          * very rare case so no need for a more efficient and
2865                          * complex approach.
2866                          */
2867                         if (ret == -ENOENT) {
2868                                 wait_event(cur_trans->writer_wait,
2869                                    atomic_read(&cur_trans->num_writers) == 1);
2870                                 ret = update_block_group_item(trans, path, cache);
2871                         }
2872                         if (ret)
2873                                 btrfs_abort_transaction(trans, ret);
2874                 }
2875
2876                 /* If its not on the io list, we need to put the block group */
2877                 if (should_put)
2878                         btrfs_put_block_group(cache);
2879                 btrfs_delayed_refs_rsv_release(fs_info, 1);
2880                 spin_lock(&cur_trans->dirty_bgs_lock);
2881         }
2882         spin_unlock(&cur_trans->dirty_bgs_lock);
2883
2884         /*
2885          * Refer to the definition of io_bgs member for details why it's safe
2886          * to use it without any locking
2887          */
2888         while (!list_empty(io)) {
2889                 cache = list_first_entry(io, struct btrfs_block_group,
2890                                          io_list);
2891                 list_del_init(&cache->io_list);
2892                 btrfs_wait_cache_io(trans, cache, path);
2893                 btrfs_put_block_group(cache);
2894         }
2895
2896         btrfs_free_path(path);
2897         return ret;
2898 }
2899
2900 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
2901                              u64 bytenr, u64 num_bytes, int alloc)
2902 {
2903         struct btrfs_fs_info *info = trans->fs_info;
2904         struct btrfs_block_group *cache = NULL;
2905         u64 total = num_bytes;
2906         u64 old_val;
2907         u64 byte_in_group;
2908         int factor;
2909         int ret = 0;
2910
2911         /* Block accounting for super block */
2912         spin_lock(&info->delalloc_root_lock);
2913         old_val = btrfs_super_bytes_used(info->super_copy);
2914         if (alloc)
2915                 old_val += num_bytes;
2916         else
2917                 old_val -= num_bytes;
2918         btrfs_set_super_bytes_used(info->super_copy, old_val);
2919         spin_unlock(&info->delalloc_root_lock);
2920
2921         while (total) {
2922                 cache = btrfs_lookup_block_group(info, bytenr);
2923                 if (!cache) {
2924                         ret = -ENOENT;
2925                         break;
2926                 }
2927                 factor = btrfs_bg_type_to_factor(cache->flags);
2928
2929                 /*
2930                  * If this block group has free space cache written out, we
2931                  * need to make sure to load it if we are removing space.  This
2932                  * is because we need the unpinning stage to actually add the
2933                  * space back to the block group, otherwise we will leak space.
2934                  */
2935                 if (!alloc && !btrfs_block_group_done(cache))
2936                         btrfs_cache_block_group(cache, 1);
2937
2938                 byte_in_group = bytenr - cache->start;
2939                 WARN_ON(byte_in_group > cache->length);
2940
2941                 spin_lock(&cache->space_info->lock);
2942                 spin_lock(&cache->lock);
2943
2944                 if (btrfs_test_opt(info, SPACE_CACHE) &&
2945                     cache->disk_cache_state < BTRFS_DC_CLEAR)
2946                         cache->disk_cache_state = BTRFS_DC_CLEAR;
2947
2948                 old_val = cache->used;
2949                 num_bytes = min(total, cache->length - byte_in_group);
2950                 if (alloc) {
2951                         old_val += num_bytes;
2952                         cache->used = old_val;
2953                         cache->reserved -= num_bytes;
2954                         cache->space_info->bytes_reserved -= num_bytes;
2955                         cache->space_info->bytes_used += num_bytes;
2956                         cache->space_info->disk_used += num_bytes * factor;
2957                         spin_unlock(&cache->lock);
2958                         spin_unlock(&cache->space_info->lock);
2959                 } else {
2960                         old_val -= num_bytes;
2961                         cache->used = old_val;
2962                         cache->pinned += num_bytes;
2963                         btrfs_space_info_update_bytes_pinned(info,
2964                                         cache->space_info, num_bytes);
2965                         cache->space_info->bytes_used -= num_bytes;
2966                         cache->space_info->disk_used -= num_bytes * factor;
2967                         spin_unlock(&cache->lock);
2968                         spin_unlock(&cache->space_info->lock);
2969
2970                         __btrfs_mod_total_bytes_pinned(cache->space_info,
2971                                                        num_bytes);
2972                         set_extent_dirty(&trans->transaction->pinned_extents,
2973                                          bytenr, bytenr + num_bytes - 1,
2974                                          GFP_NOFS | __GFP_NOFAIL);
2975                 }
2976
2977                 spin_lock(&trans->transaction->dirty_bgs_lock);
2978                 if (list_empty(&cache->dirty_list)) {
2979                         list_add_tail(&cache->dirty_list,
2980                                       &trans->transaction->dirty_bgs);
2981                         trans->delayed_ref_updates++;
2982                         btrfs_get_block_group(cache);
2983                 }
2984                 spin_unlock(&trans->transaction->dirty_bgs_lock);
2985
2986                 /*
2987                  * No longer have used bytes in this block group, queue it for
2988                  * deletion. We do this after adding the block group to the
2989                  * dirty list to avoid races between cleaner kthread and space
2990                  * cache writeout.
2991                  */
2992                 if (!alloc && old_val == 0) {
2993                         if (!btrfs_test_opt(info, DISCARD_ASYNC))
2994                                 btrfs_mark_bg_unused(cache);
2995                 }
2996
2997                 btrfs_put_block_group(cache);
2998                 total -= num_bytes;
2999                 bytenr += num_bytes;
3000         }
3001
3002         /* Modified block groups are accounted for in the delayed_refs_rsv. */
3003         btrfs_update_delayed_refs_rsv(trans);
3004         return ret;
3005 }
3006
3007 /**
3008  * btrfs_add_reserved_bytes - update the block_group and space info counters
3009  * @cache:      The cache we are manipulating
3010  * @ram_bytes:  The number of bytes of file content, and will be same to
3011  *              @num_bytes except for the compress path.
3012  * @num_bytes:  The number of bytes in question
3013  * @delalloc:   The blocks are allocated for the delalloc write
3014  *
3015  * This is called by the allocator when it reserves space. If this is a
3016  * reservation and the block group has become read only we cannot make the
3017  * reservation and return -EAGAIN, otherwise this function always succeeds.
3018  */
3019 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3020                              u64 ram_bytes, u64 num_bytes, int delalloc)
3021 {
3022         struct btrfs_space_info *space_info = cache->space_info;
3023         int ret = 0;
3024
3025         spin_lock(&space_info->lock);
3026         spin_lock(&cache->lock);
3027         if (cache->ro) {
3028                 ret = -EAGAIN;
3029         } else {
3030                 cache->reserved += num_bytes;
3031                 space_info->bytes_reserved += num_bytes;
3032                 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3033                                               space_info->flags, num_bytes, 1);
3034                 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3035                                                       space_info, -ram_bytes);
3036                 if (delalloc)
3037                         cache->delalloc_bytes += num_bytes;
3038
3039                 /*
3040                  * Compression can use less space than we reserved, so wake
3041                  * tickets if that happens
3042                  */
3043                 if (num_bytes < ram_bytes)
3044                         btrfs_try_granting_tickets(cache->fs_info, space_info);
3045         }
3046         spin_unlock(&cache->lock);
3047         spin_unlock(&space_info->lock);
3048         return ret;
3049 }
3050
3051 /**
3052  * btrfs_free_reserved_bytes - update the block_group and space info counters
3053  * @cache:      The cache we are manipulating
3054  * @num_bytes:  The number of bytes in question
3055  * @delalloc:   The blocks are allocated for the delalloc write
3056  *
3057  * This is called by somebody who is freeing space that was never actually used
3058  * on disk.  For example if you reserve some space for a new leaf in transaction
3059  * A and before transaction A commits you free that leaf, you call this with
3060  * reserve set to 0 in order to clear the reservation.
3061  */
3062 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3063                                u64 num_bytes, int delalloc)
3064 {
3065         struct btrfs_space_info *space_info = cache->space_info;
3066
3067         spin_lock(&space_info->lock);
3068         spin_lock(&cache->lock);
3069         if (cache->ro)
3070                 space_info->bytes_readonly += num_bytes;
3071         cache->reserved -= num_bytes;
3072         space_info->bytes_reserved -= num_bytes;
3073         space_info->max_extent_size = 0;
3074
3075         if (delalloc)
3076                 cache->delalloc_bytes -= num_bytes;
3077         spin_unlock(&cache->lock);
3078
3079         btrfs_try_granting_tickets(cache->fs_info, space_info);
3080         spin_unlock(&space_info->lock);
3081 }
3082
3083 static void force_metadata_allocation(struct btrfs_fs_info *info)
3084 {
3085         struct list_head *head = &info->space_info;
3086         struct btrfs_space_info *found;
3087
3088         list_for_each_entry(found, head, list) {
3089                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3090                         found->force_alloc = CHUNK_ALLOC_FORCE;
3091         }
3092 }
3093
3094 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3095                               struct btrfs_space_info *sinfo, int force)
3096 {
3097         u64 bytes_used = btrfs_space_info_used(sinfo, false);
3098         u64 thresh;
3099
3100         if (force == CHUNK_ALLOC_FORCE)
3101                 return 1;
3102
3103         /*
3104          * in limited mode, we want to have some free space up to
3105          * about 1% of the FS size.
3106          */
3107         if (force == CHUNK_ALLOC_LIMITED) {
3108                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3109                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3110
3111                 if (sinfo->total_bytes - bytes_used < thresh)
3112                         return 1;
3113         }
3114
3115         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3116                 return 0;
3117         return 1;
3118 }
3119
3120 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3121 {
3122         u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3123
3124         return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3125 }
3126
3127 /*
3128  * If force is CHUNK_ALLOC_FORCE:
3129  *    - return 1 if it successfully allocates a chunk,
3130  *    - return errors including -ENOSPC otherwise.
3131  * If force is NOT CHUNK_ALLOC_FORCE:
3132  *    - return 0 if it doesn't need to allocate a new chunk,
3133  *    - return 1 if it successfully allocates a chunk,
3134  *    - return errors including -ENOSPC otherwise.
3135  */
3136 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3137                       enum btrfs_chunk_alloc_enum force)
3138 {
3139         struct btrfs_fs_info *fs_info = trans->fs_info;
3140         struct btrfs_space_info *space_info;
3141         bool wait_for_alloc = false;
3142         bool should_alloc = false;
3143         int ret = 0;
3144
3145         /* Don't re-enter if we're already allocating a chunk */
3146         if (trans->allocating_chunk)
3147                 return -ENOSPC;
3148
3149         space_info = btrfs_find_space_info(fs_info, flags);
3150         ASSERT(space_info);
3151
3152         do {
3153                 spin_lock(&space_info->lock);
3154                 if (force < space_info->force_alloc)
3155                         force = space_info->force_alloc;
3156                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
3157                 if (space_info->full) {
3158                         /* No more free physical space */
3159                         if (should_alloc)
3160                                 ret = -ENOSPC;
3161                         else
3162                                 ret = 0;
3163                         spin_unlock(&space_info->lock);
3164                         return ret;
3165                 } else if (!should_alloc) {
3166                         spin_unlock(&space_info->lock);
3167                         return 0;
3168                 } else if (space_info->chunk_alloc) {
3169                         /*
3170                          * Someone is already allocating, so we need to block
3171                          * until this someone is finished and then loop to
3172                          * recheck if we should continue with our allocation
3173                          * attempt.
3174                          */
3175                         wait_for_alloc = true;
3176                         spin_unlock(&space_info->lock);
3177                         mutex_lock(&fs_info->chunk_mutex);
3178                         mutex_unlock(&fs_info->chunk_mutex);
3179                 } else {
3180                         /* Proceed with allocation */
3181                         space_info->chunk_alloc = 1;
3182                         wait_for_alloc = false;
3183                         spin_unlock(&space_info->lock);
3184                 }
3185
3186                 cond_resched();
3187         } while (wait_for_alloc);
3188
3189         mutex_lock(&fs_info->chunk_mutex);
3190         trans->allocating_chunk = true;
3191
3192         /*
3193          * If we have mixed data/metadata chunks we want to make sure we keep
3194          * allocating mixed chunks instead of individual chunks.
3195          */
3196         if (btrfs_mixed_space_info(space_info))
3197                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3198
3199         /*
3200          * if we're doing a data chunk, go ahead and make sure that
3201          * we keep a reasonable number of metadata chunks allocated in the
3202          * FS as well.
3203          */
3204         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3205                 fs_info->data_chunk_allocations++;
3206                 if (!(fs_info->data_chunk_allocations %
3207                       fs_info->metadata_ratio))
3208                         force_metadata_allocation(fs_info);
3209         }
3210
3211         /*
3212          * Check if we have enough space in SYSTEM chunk because we may need
3213          * to update devices.
3214          */
3215         check_system_chunk(trans, flags);
3216
3217         ret = btrfs_alloc_chunk(trans, flags);
3218         trans->allocating_chunk = false;
3219
3220         spin_lock(&space_info->lock);
3221         if (ret < 0) {
3222                 if (ret == -ENOSPC)
3223                         space_info->full = 1;
3224                 else
3225                         goto out;
3226         } else {
3227                 ret = 1;
3228                 space_info->max_extent_size = 0;
3229         }
3230
3231         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3232 out:
3233         space_info->chunk_alloc = 0;
3234         spin_unlock(&space_info->lock);
3235         mutex_unlock(&fs_info->chunk_mutex);
3236         /*
3237          * When we allocate a new chunk we reserve space in the chunk block
3238          * reserve to make sure we can COW nodes/leafs in the chunk tree or
3239          * add new nodes/leafs to it if we end up needing to do it when
3240          * inserting the chunk item and updating device items as part of the
3241          * second phase of chunk allocation, performed by
3242          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
3243          * large number of new block groups to create in our transaction
3244          * handle's new_bgs list to avoid exhausting the chunk block reserve
3245          * in extreme cases - like having a single transaction create many new
3246          * block groups when starting to write out the free space caches of all
3247          * the block groups that were made dirty during the lifetime of the
3248          * transaction.
3249          */
3250         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
3251                 btrfs_create_pending_block_groups(trans);
3252
3253         return ret;
3254 }
3255
3256 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3257 {
3258         u64 num_dev;
3259
3260         num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3261         if (!num_dev)
3262                 num_dev = fs_info->fs_devices->rw_devices;
3263
3264         return num_dev;
3265 }
3266
3267 /*
3268  * Reserve space in the system space for allocating or removing a chunk
3269  */
3270 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3271 {
3272         struct btrfs_fs_info *fs_info = trans->fs_info;
3273         struct btrfs_space_info *info;
3274         u64 left;
3275         u64 thresh;
3276         int ret = 0;
3277         u64 num_devs;
3278
3279         /*
3280          * Needed because we can end up allocating a system chunk and for an
3281          * atomic and race free space reservation in the chunk block reserve.
3282          */
3283         lockdep_assert_held(&fs_info->chunk_mutex);
3284
3285         info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3286         spin_lock(&info->lock);
3287         left = info->total_bytes - btrfs_space_info_used(info, true);
3288         spin_unlock(&info->lock);
3289
3290         num_devs = get_profile_num_devs(fs_info, type);
3291
3292         /* num_devs device items to update and 1 chunk item to add or remove */
3293         thresh = btrfs_calc_metadata_size(fs_info, num_devs) +
3294                 btrfs_calc_insert_metadata_size(fs_info, 1);
3295
3296         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3297                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3298                            left, thresh, type);
3299                 btrfs_dump_space_info(fs_info, info, 0, 0);
3300         }
3301
3302         if (left < thresh) {
3303                 u64 flags = btrfs_system_alloc_profile(fs_info);
3304
3305                 /*
3306                  * Ignore failure to create system chunk. We might end up not
3307                  * needing it, as we might not need to COW all nodes/leafs from
3308                  * the paths we visit in the chunk tree (they were already COWed
3309                  * or created in the current transaction for example).
3310                  */
3311                 ret = btrfs_alloc_chunk(trans, flags);
3312         }
3313
3314         if (!ret) {
3315                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
3316                                           &fs_info->chunk_block_rsv,
3317                                           thresh, BTRFS_RESERVE_NO_FLUSH);
3318                 if (!ret)
3319                         trans->chunk_bytes_reserved += thresh;
3320         }
3321 }
3322
3323 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
3324 {
3325         struct btrfs_block_group *block_group;
3326         u64 last = 0;
3327
3328         while (1) {
3329                 struct inode *inode;
3330
3331                 block_group = btrfs_lookup_first_block_group(info, last);
3332                 while (block_group) {
3333                         btrfs_wait_block_group_cache_done(block_group);
3334                         spin_lock(&block_group->lock);
3335                         if (block_group->iref)
3336                                 break;
3337                         spin_unlock(&block_group->lock);
3338                         block_group = btrfs_next_block_group(block_group);
3339                 }
3340                 if (!block_group) {
3341                         if (last == 0)
3342                                 break;
3343                         last = 0;
3344                         continue;
3345                 }
3346
3347                 inode = block_group->inode;
3348                 block_group->iref = 0;
3349                 block_group->inode = NULL;
3350                 spin_unlock(&block_group->lock);
3351                 ASSERT(block_group->io_ctl.inode == NULL);
3352                 iput(inode);
3353                 last = block_group->start + block_group->length;
3354                 btrfs_put_block_group(block_group);
3355         }
3356 }
3357
3358 /*
3359  * Must be called only after stopping all workers, since we could have block
3360  * group caching kthreads running, and therefore they could race with us if we
3361  * freed the block groups before stopping them.
3362  */
3363 int btrfs_free_block_groups(struct btrfs_fs_info *info)
3364 {
3365         struct btrfs_block_group *block_group;
3366         struct btrfs_space_info *space_info;
3367         struct btrfs_caching_control *caching_ctl;
3368         struct rb_node *n;
3369
3370         spin_lock(&info->block_group_cache_lock);
3371         while (!list_empty(&info->caching_block_groups)) {
3372                 caching_ctl = list_entry(info->caching_block_groups.next,
3373                                          struct btrfs_caching_control, list);
3374                 list_del(&caching_ctl->list);
3375                 btrfs_put_caching_control(caching_ctl);
3376         }
3377         spin_unlock(&info->block_group_cache_lock);
3378
3379         spin_lock(&info->unused_bgs_lock);
3380         while (!list_empty(&info->unused_bgs)) {
3381                 block_group = list_first_entry(&info->unused_bgs,
3382                                                struct btrfs_block_group,
3383                                                bg_list);
3384                 list_del_init(&block_group->bg_list);
3385                 btrfs_put_block_group(block_group);
3386         }
3387         spin_unlock(&info->unused_bgs_lock);
3388
3389         spin_lock(&info->block_group_cache_lock);
3390         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
3391                 block_group = rb_entry(n, struct btrfs_block_group,
3392                                        cache_node);
3393                 rb_erase(&block_group->cache_node,
3394                          &info->block_group_cache_tree);
3395                 RB_CLEAR_NODE(&block_group->cache_node);
3396                 spin_unlock(&info->block_group_cache_lock);
3397
3398                 down_write(&block_group->space_info->groups_sem);
3399                 list_del(&block_group->list);
3400                 up_write(&block_group->space_info->groups_sem);
3401
3402                 /*
3403                  * We haven't cached this block group, which means we could
3404                  * possibly have excluded extents on this block group.
3405                  */
3406                 if (block_group->cached == BTRFS_CACHE_NO ||
3407                     block_group->cached == BTRFS_CACHE_ERROR)
3408                         btrfs_free_excluded_extents(block_group);
3409
3410                 btrfs_remove_free_space_cache(block_group);
3411                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
3412                 ASSERT(list_empty(&block_group->dirty_list));
3413                 ASSERT(list_empty(&block_group->io_list));
3414                 ASSERT(list_empty(&block_group->bg_list));
3415                 ASSERT(refcount_read(&block_group->refs) == 1);
3416                 ASSERT(block_group->swap_extents == 0);
3417                 btrfs_put_block_group(block_group);
3418
3419                 spin_lock(&info->block_group_cache_lock);
3420         }
3421         spin_unlock(&info->block_group_cache_lock);
3422
3423         btrfs_release_global_block_rsv(info);
3424
3425         while (!list_empty(&info->space_info)) {
3426                 space_info = list_entry(info->space_info.next,
3427                                         struct btrfs_space_info,
3428                                         list);
3429
3430                 /*
3431                  * Do not hide this behind enospc_debug, this is actually
3432                  * important and indicates a real bug if this happens.
3433                  */
3434                 if (WARN_ON(space_info->bytes_pinned > 0 ||
3435                             space_info->bytes_reserved > 0 ||
3436                             space_info->bytes_may_use > 0))
3437                         btrfs_dump_space_info(info, space_info, 0, 0);
3438                 WARN_ON(space_info->reclaim_size > 0);
3439                 list_del(&space_info->list);
3440                 btrfs_sysfs_remove_space_info(space_info);
3441         }
3442         return 0;
3443 }
3444
3445 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
3446 {
3447         atomic_inc(&cache->frozen);
3448 }
3449
3450 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
3451 {
3452         struct btrfs_fs_info *fs_info = block_group->fs_info;
3453         struct extent_map_tree *em_tree;
3454         struct extent_map *em;
3455         bool cleanup;
3456
3457         spin_lock(&block_group->lock);
3458         cleanup = (atomic_dec_and_test(&block_group->frozen) &&
3459                    block_group->removed);
3460         spin_unlock(&block_group->lock);
3461
3462         if (cleanup) {
3463                 em_tree = &fs_info->mapping_tree;
3464                 write_lock(&em_tree->lock);
3465                 em = lookup_extent_mapping(em_tree, block_group->start,
3466                                            1);
3467                 BUG_ON(!em); /* logic error, can't happen */
3468                 remove_extent_mapping(em_tree, em);
3469                 write_unlock(&em_tree->lock);
3470
3471                 /* once for us and once for the tree */
3472                 free_extent_map(em);
3473                 free_extent_map(em);
3474
3475                 /*
3476                  * We may have left one free space entry and other possible
3477                  * tasks trimming this block group have left 1 entry each one.
3478                  * Free them if any.
3479                  */
3480                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3481         }
3482 }
3483
3484 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
3485 {
3486         bool ret = true;
3487
3488         spin_lock(&bg->lock);
3489         if (bg->ro)
3490                 ret = false;
3491         else
3492                 bg->swap_extents++;
3493         spin_unlock(&bg->lock);
3494
3495         return ret;
3496 }
3497
3498 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
3499 {
3500         spin_lock(&bg->lock);
3501         ASSERT(!bg->ro);
3502         ASSERT(bg->swap_extents >= amount);
3503         bg->swap_extents -= amount;
3504         spin_unlock(&bg->lock);
3505 }