fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  55                                struct btrfs_fs_info *fs_info,
  56                                 struct btrfs_delayed_ref_node *node, u64 parent,
  57                                 u64 root_objectid, u64 owner_objectid,
  58                                 u64 owner_offset, int refs_to_drop,
  59                                 struct btrfs_delayed_extent_op *extra_op);
  60 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  61                                     struct extent_buffer *leaf,
  62                                     struct btrfs_extent_item *ei);
  63 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  64                                       struct btrfs_fs_info *fs_info,
  65                                       u64 parent, u64 root_objectid,
  66                                       u64 flags, u64 owner, u64 offset,
  67                                       struct btrfs_key *ins, int ref_mod);
  68 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  69                                      struct btrfs_delayed_ref_node *node,
  70                                      struct btrfs_delayed_extent_op *extent_op);
  71 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  72                           struct btrfs_fs_info *fs_info, u64 flags,
  73                           int force);
  74 static int find_next_key(struct btrfs_path *path, int level,
  75                          struct btrfs_key *key);
  76 static void dump_space_info(struct btrfs_fs_info *fs_info,
  77                             struct btrfs_space_info *info, u64 bytes,
  78                             int dump_block_groups);
  79 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  80                                u64 num_bytes);
  81 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  82                                      struct btrfs_space_info *space_info,
  83                                      u64 num_bytes);
  84 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  85                                      struct btrfs_space_info *space_info,
  86                                      u64 num_bytes);
  87
  88 static noinline int
  89 block_group_cache_done(struct btrfs_block_group_cache *cache)
  90 {
  91         smp_mb();
  92         return cache->cached == BTRFS_CACHE_FINISHED ||
  93                 cache->cached == BTRFS_CACHE_ERROR;
  94 }
  95
  96 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  97 {
  98         return (cache->flags & bits) == bits;
  99 }
 100
 101 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 102 {
 103         atomic_inc(&cache->count);
 104 }
 105
 106 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 107 {
 108         if (atomic_dec_and_test(&cache->count)) {
 109                 WARN_ON(cache->pinned > 0);
 110                 WARN_ON(cache->reserved > 0);
 111
 112                 /*
 113                  * If not empty, someone is still holding mutex of
 114                  * full_stripe_lock, which can only be released by caller.
 115                  * And it will definitely cause use-after-free when caller
 116                  * tries to release full stripe lock.
 117                  *
 118                  * No better way to resolve, but only to warn.
 119                  */
 120                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 121                 kfree(cache->free_space_ctl);
 122                 kfree(cache);
 123         }
 124 }
 125
 126 /*
 127  * this adds the block group to the fs_info rb tree for the block group
 128  * cache
 129  */
 130 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 131                                 struct btrfs_block_group_cache *block_group)
 132 {
 133         struct rb_node **p;
 134         struct rb_node *parent = NULL;
 135         struct btrfs_block_group_cache *cache;
 136
 137         spin_lock(&info->block_group_cache_lock);
 138         p = &info->block_group_cache_tree.rb_node;
 139
 140         while (*p) {
 141                 parent = *p;
 142                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 143                                  cache_node);
 144                 if (block_group->key.objectid < cache->key.objectid) {
 145                         p = &(*p)->rb_left;
 146                 } else if (block_group->key.objectid > cache->key.objectid) {
 147                         p = &(*p)->rb_right;
 148                 } else {
 149                         spin_unlock(&info->block_group_cache_lock);
 150                         return -EEXIST;
 151                 }
 152         }
 153
 154         rb_link_node(&block_group->cache_node, parent, p);
 155         rb_insert_color(&block_group->cache_node,
 156                         &info->block_group_cache_tree);
 157
 158         if (info->first_logical_byte > block_group->key.objectid)
 159                 info->first_logical_byte = block_group->key.objectid;
 160
 161         spin_unlock(&info->block_group_cache_lock);
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * This will return the block group at or after bytenr if contains is 0, else
 168  * it will return the block group that contains the bytenr
 169  */
 170 static struct btrfs_block_group_cache *
 171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 172                               int contains)
 173 {
 174         struct btrfs_block_group_cache *cache, *ret = NULL;
 175         struct rb_node *n;
 176         u64 end, start;
 177
 178         spin_lock(&info->block_group_cache_lock);
 179         n = info->block_group_cache_tree.rb_node;
 180
 181         while (n) {
 182                 cache = rb_entry(n, struct btrfs_block_group_cache,
 183                                  cache_node);
 184                 end = cache->key.objectid + cache->key.offset - 1;
 185                 start = cache->key.objectid;
 186
 187                 if (bytenr < start) {
 188                         if (!contains && (!ret || start < ret->key.objectid))
 189                                 ret = cache;
 190                         n = n->rb_left;
 191                 } else if (bytenr > start) {
 192                         if (contains && bytenr <= end) {
 193                                 ret = cache;
 194                                 break;
 195                         }
 196                         n = n->rb_right;
 197                 } else {
 198                         ret = cache;
 199                         break;
 200                 }
 201         }
 202         if (ret) {
 203                 btrfs_get_block_group(ret);
 204                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 205                         info->first_logical_byte = ret->key.objectid;
 206         }
 207         spin_unlock(&info->block_group_cache_lock);
 208
 209         return ret;
 210 }
 211
 212 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 213                                u64 start, u64 num_bytes)
 214 {
 215         u64 end = start + num_bytes - 1;
 216         set_extent_bits(&fs_info->freed_extents[0],
 217                         start, end, EXTENT_UPTODATE);
 218         set_extent_bits(&fs_info->freed_extents[1],
 219                         start, end, EXTENT_UPTODATE);
 220         return 0;
 221 }
 222
 223 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 224                                   struct btrfs_block_group_cache *cache)
 225 {
 226         u64 start, end;
 227
 228         start = cache->key.objectid;
 229         end = start + cache->key.offset - 1;
 230
 231         clear_extent_bits(&fs_info->freed_extents[0],
 232                           start, end, EXTENT_UPTODATE);
 233         clear_extent_bits(&fs_info->freed_extents[1],
 234                           start, end, EXTENT_UPTODATE);
 235 }
 236
 237 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 238                                  struct btrfs_block_group_cache *cache)
 239 {
 240         u64 bytenr;
 241         u64 *logical;
 242         int stripe_len;
 243         int i, nr, ret;
 244
 245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 247                 cache->bytes_super += stripe_len;
 248                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 249                                           stripe_len);
 250                 if (ret)
 251                         return ret;
 252         }
 253
 254         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 255                 bytenr = btrfs_sb_offset(i);
 256                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 257                                        bytenr, &logical, &nr, &stripe_len);
 258                 if (ret)
 259                         return ret;
 260
 261                 while (nr--) {
 262                         u64 start, len;
 263
 264                         if (logical[nr] > cache->key.objectid +
 265                             cache->key.offset)
 266                                 continue;
 267
 268                         if (logical[nr] + stripe_len <= cache->key.objectid)
 269                                 continue;
 270
 271                         start = logical[nr];
 272                         if (start < cache->key.objectid) {
 273                                 start = cache->key.objectid;
 274                                 len = (logical[nr] + stripe_len) - start;
 275                         } else {
 276                                 len = min_t(u64, stripe_len,
 277                                             cache->key.objectid +
 278                                             cache->key.offset - start);
 279                         }
 280
 281                         cache->bytes_super += len;
 282                         ret = add_excluded_extent(fs_info, start, len);
 283                         if (ret) {
 284                                 kfree(logical);
 285                                 return ret;
 286                         }
 287                 }
 288
 289                 kfree(logical);
 290         }
 291         return 0;
 292 }
 293
 294 static struct btrfs_caching_control *
 295 get_caching_control(struct btrfs_block_group_cache *cache)
 296 {
 297         struct btrfs_caching_control *ctl;
 298
 299         spin_lock(&cache->lock);
 300         if (!cache->caching_ctl) {
 301                 spin_unlock(&cache->lock);
 302                 return NULL;
 303         }
 304
 305         ctl = cache->caching_ctl;
 306         refcount_inc(&ctl->count);
 307         spin_unlock(&cache->lock);
 308         return ctl;
 309 }
 310
 311 static void put_caching_control(struct btrfs_caching_control *ctl)
 312 {
 313         if (refcount_dec_and_test(&ctl->count))
 314                 kfree(ctl);
 315 }
 316
 317 #ifdef CONFIG_BTRFS_DEBUG
 318 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 319 {
 320         struct btrfs_fs_info *fs_info = block_group->fs_info;
 321         u64 start = block_group->key.objectid;
 322         u64 len = block_group->key.offset;
 323         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 324                 fs_info->nodesize : fs_info->sectorsize;
 325         u64 step = chunk << 1;
 326
 327         while (len > chunk) {
 328                 btrfs_remove_free_space(block_group, start, chunk);
 329                 start += step;
 330                 if (len < step)
 331                         len = 0;
 332                 else
 333                         len -= step;
 334         }
 335 }
 336 #endif
 337
 338 /*
 339  * this is only called by cache_block_group, since we could have freed extents
 340  * we need to check the pinned_extents for any extents that can't be used yet
 341  * since their free space will be released as soon as the transaction commits.
 342  */
 343 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 344                        u64 start, u64 end)
 345 {
 346         struct btrfs_fs_info *info = block_group->fs_info;
 347         u64 extent_start, extent_end, size, total_added = 0;
 348         int ret;
 349
 350         while (start < end) {
 351                 ret = find_first_extent_bit(info->pinned_extents, start,
 352                                             &extent_start, &extent_end,
 353                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 354                                             NULL);
 355                 if (ret)
 356                         break;
 357
 358                 if (extent_start <= start) {
 359                         start = extent_end + 1;
 360                 } else if (extent_start > start && extent_start < end) {
 361                         size = extent_start - start;
 362                         total_added += size;
 363                         ret = btrfs_add_free_space(block_group, start,
 364                                                    size);
 365                         BUG_ON(ret); /* -ENOMEM or logic error */
 366                         start = extent_end + 1;
 367                 } else {
 368                         break;
 369                 }
 370         }
 371
 372         if (start < end) {
 373                 size = end - start;
 374                 total_added += size;
 375                 ret = btrfs_add_free_space(block_group, start, size);
 376                 BUG_ON(ret); /* -ENOMEM or logic error */
 377         }
 378
 379         return total_added;
 380 }
 381
 382 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 383 {
 384         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 385         struct btrfs_fs_info *fs_info = block_group->fs_info;
 386         struct btrfs_root *extent_root = fs_info->extent_root;
 387         struct btrfs_path *path;
 388         struct extent_buffer *leaf;
 389         struct btrfs_key key;
 390         u64 total_found = 0;
 391         u64 last = 0;
 392         u32 nritems;
 393         int ret;
 394         bool wakeup = true;
 395
 396         path = btrfs_alloc_path();
 397         if (!path)
 398                 return -ENOMEM;
 399
 400         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 401
 402 #ifdef CONFIG_BTRFS_DEBUG
 403         /*
 404          * If we're fragmenting we don't want to make anybody think we can
 405          * allocate from this block group until we've had a chance to fragment
 406          * the free space.
 407          */
 408         if (btrfs_should_fragment_free_space(block_group))
 409                 wakeup = false;
 410 #endif
 411         /*
 412          * We don't want to deadlock with somebody trying to allocate a new
 413          * extent for the extent root while also trying to search the extent
 414          * root to add free space.  So we skip locking and search the commit
 415          * root, since its read-only
 416          */
 417         path->skip_locking = 1;
 418         path->search_commit_root = 1;
 419         path->reada = READA_FORWARD;
 420
 421         key.objectid = last;
 422         key.offset = 0;
 423         key.type = BTRFS_EXTENT_ITEM_KEY;
 424
 425 next:
 426         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 427         if (ret < 0)
 428                 goto out;
 429
 430         leaf = path->nodes[0];
 431         nritems = btrfs_header_nritems(leaf);
 432
 433         while (1) {
 434                 if (btrfs_fs_closing(fs_info) > 1) {
 435                         last = (u64)-1;
 436                         break;
 437                 }
 438
 439                 if (path->slots[0] < nritems) {
 440                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 441                 } else {
 442                         ret = find_next_key(path, 0, &key);
 443                         if (ret)
 444                                 break;
 445
 446                         if (need_resched() ||
 447                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 448                                 if (wakeup)
 449                                         caching_ctl->progress = last;
 450                                 btrfs_release_path(path);
 451                                 up_read(&fs_info->commit_root_sem);
 452                                 mutex_unlock(&caching_ctl->mutex);
 453                                 cond_resched();
 454                                 mutex_lock(&caching_ctl->mutex);
 455                                 down_read(&fs_info->commit_root_sem);
 456                                 goto next;
 457                         }
 458
 459                         ret = btrfs_next_leaf(extent_root, path);
 460                         if (ret < 0)
 461                                 goto out;
 462                         if (ret)
 463                                 break;
 464                         leaf = path->nodes[0];
 465                         nritems = btrfs_header_nritems(leaf);
 466                         continue;
 467                 }
 468
 469                 if (key.objectid < last) {
 470                         key.objectid = last;
 471                         key.offset = 0;
 472                         key.type = BTRFS_EXTENT_ITEM_KEY;
 473
 474                         if (wakeup)
 475                                 caching_ctl->progress = last;
 476                         btrfs_release_path(path);
 477                         goto next;
 478                 }
 479
 480                 if (key.objectid < block_group->key.objectid) {
 481                         path->slots[0]++;
 482                         continue;
 483                 }
 484
 485                 if (key.objectid >= block_group->key.objectid +
 486                     block_group->key.offset)
 487                         break;
 488
 489                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 490                     key.type == BTRFS_METADATA_ITEM_KEY) {
 491                         total_found += add_new_free_space(block_group, last,
 492                                                           key.objectid);
 493                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 494                                 last = key.objectid +
 495                                         fs_info->nodesize;
 496                         else
 497                                 last = key.objectid + key.offset;
 498
 499                         if (total_found > CACHING_CTL_WAKE_UP) {
 500                                 total_found = 0;
 501                                 if (wakeup)
 502                                         wake_up(&caching_ctl->wait);
 503                         }
 504                 }
 505                 path->slots[0]++;
 506         }
 507         ret = 0;
 508
 509         total_found += add_new_free_space(block_group, last,
 510                                           block_group->key.objectid +
 511                                           block_group->key.offset);
 512         caching_ctl->progress = (u64)-1;
 513
 514 out:
 515         btrfs_free_path(path);
 516         return ret;
 517 }
 518
 519 static noinline void caching_thread(struct btrfs_work *work)
 520 {
 521         struct btrfs_block_group_cache *block_group;
 522         struct btrfs_fs_info *fs_info;
 523         struct btrfs_caching_control *caching_ctl;
 524         int ret;
 525
 526         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 527         block_group = caching_ctl->block_group;
 528         fs_info = block_group->fs_info;
 529
 530         mutex_lock(&caching_ctl->mutex);
 531         down_read(&fs_info->commit_root_sem);
 532
 533         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 534                 ret = load_free_space_tree(caching_ctl);
 535         else
 536                 ret = load_extent_tree_free(caching_ctl);
 537
 538         spin_lock(&block_group->lock);
 539         block_group->caching_ctl = NULL;
 540         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 541         spin_unlock(&block_group->lock);
 542
 543 #ifdef CONFIG_BTRFS_DEBUG
 544         if (btrfs_should_fragment_free_space(block_group)) {
 545                 u64 bytes_used;
 546
 547                 spin_lock(&block_group->space_info->lock);
 548                 spin_lock(&block_group->lock);
 549                 bytes_used = block_group->key.offset -
 550                         btrfs_block_group_used(&block_group->item);
 551                 block_group->space_info->bytes_used += bytes_used >> 1;
 552                 spin_unlock(&block_group->lock);
 553                 spin_unlock(&block_group->space_info->lock);
 554                 fragment_free_space(block_group);
 555         }
 556 #endif
 557
 558         caching_ctl->progress = (u64)-1;
 559
 560         up_read(&fs_info->commit_root_sem);
 561         free_excluded_extents(fs_info, block_group);
 562         mutex_unlock(&caching_ctl->mutex);
 563
 564         wake_up(&caching_ctl->wait);
 565
 566         put_caching_control(caching_ctl);
 567         btrfs_put_block_group(block_group);
 568 }
 569
 570 static int cache_block_group(struct btrfs_block_group_cache *cache,
 571                              int load_cache_only)
 572 {
 573         DEFINE_WAIT(wait);
 574         struct btrfs_fs_info *fs_info = cache->fs_info;
 575         struct btrfs_caching_control *caching_ctl;
 576         int ret = 0;
 577
 578         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 579         if (!caching_ctl)
 580                 return -ENOMEM;
 581
 582         INIT_LIST_HEAD(&caching_ctl->list);
 583         mutex_init(&caching_ctl->mutex);
 584         init_waitqueue_head(&caching_ctl->wait);
 585         caching_ctl->block_group = cache;
 586         caching_ctl->progress = cache->key.objectid;
 587         refcount_set(&caching_ctl->count, 1);
 588         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 589                         caching_thread, NULL, NULL);
 590
 591         spin_lock(&cache->lock);
 592         /*
 593          * This should be a rare occasion, but this could happen I think in the
 594          * case where one thread starts to load the space cache info, and then
 595          * some other thread starts a transaction commit which tries to do an
 596          * allocation while the other thread is still loading the space cache
 597          * info.  The previous loop should have kept us from choosing this block
 598          * group, but if we've moved to the state where we will wait on caching
 599          * block groups we need to first check if we're doing a fast load here,
 600          * so we can wait for it to finish, otherwise we could end up allocating
 601          * from a block group who's cache gets evicted for one reason or
 602          * another.
 603          */
 604         while (cache->cached == BTRFS_CACHE_FAST) {
 605                 struct btrfs_caching_control *ctl;
 606
 607                 ctl = cache->caching_ctl;
 608                 refcount_inc(&ctl->count);
 609                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 610                 spin_unlock(&cache->lock);
 611
 612                 schedule();
 613
 614                 finish_wait(&ctl->wait, &wait);
 615                 put_caching_control(ctl);
 616                 spin_lock(&cache->lock);
 617         }
 618
 619         if (cache->cached != BTRFS_CACHE_NO) {
 620                 spin_unlock(&cache->lock);
 621                 kfree(caching_ctl);
 622                 return 0;
 623         }
 624         WARN_ON(cache->caching_ctl);
 625         cache->caching_ctl = caching_ctl;
 626         cache->cached = BTRFS_CACHE_FAST;
 627         spin_unlock(&cache->lock);
 628
 629         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 630                 mutex_lock(&caching_ctl->mutex);
 631                 ret = load_free_space_cache(fs_info, cache);
 632
 633                 spin_lock(&cache->lock);
 634                 if (ret == 1) {
 635                         cache->caching_ctl = NULL;
 636                         cache->cached = BTRFS_CACHE_FINISHED;
 637                         cache->last_byte_to_unpin = (u64)-1;
 638                         caching_ctl->progress = (u64)-1;
 639                 } else {
 640                         if (load_cache_only) {
 641                                 cache->caching_ctl = NULL;
 642                                 cache->cached = BTRFS_CACHE_NO;
 643                         } else {
 644                                 cache->cached = BTRFS_CACHE_STARTED;
 645                                 cache->has_caching_ctl = 1;
 646                         }
 647                 }
 648                 spin_unlock(&cache->lock);
 649 #ifdef CONFIG_BTRFS_DEBUG
 650                 if (ret == 1 &&
 651                     btrfs_should_fragment_free_space(cache)) {
 652                         u64 bytes_used;
 653
 654                         spin_lock(&cache->space_info->lock);
 655                         spin_lock(&cache->lock);
 656                         bytes_used = cache->key.offset -
 657                                 btrfs_block_group_used(&cache->item);
 658                         cache->space_info->bytes_used += bytes_used >> 1;
 659                         spin_unlock(&cache->lock);
 660                         spin_unlock(&cache->space_info->lock);
 661                         fragment_free_space(cache);
 662                 }
 663 #endif
 664                 mutex_unlock(&caching_ctl->mutex);
 665
 666                 wake_up(&caching_ctl->wait);
 667                 if (ret == 1) {
 668                         put_caching_control(caching_ctl);
 669                         free_excluded_extents(fs_info, cache);
 670                         return 0;
 671                 }
 672         } else {
 673                 /*
 674                  * We're either using the free space tree or no caching at all.
 675                  * Set cached to the appropriate value and wakeup any waiters.
 676                  */
 677                 spin_lock(&cache->lock);
 678                 if (load_cache_only) {
 679                         cache->caching_ctl = NULL;
 680                         cache->cached = BTRFS_CACHE_NO;
 681                 } else {
 682                         cache->cached = BTRFS_CACHE_STARTED;
 683                         cache->has_caching_ctl = 1;
 684                 }
 685                 spin_unlock(&cache->lock);
 686                 wake_up(&caching_ctl->wait);
 687         }
 688
 689         if (load_cache_only) {
 690                 put_caching_control(caching_ctl);
 691                 return 0;
 692         }
 693
 694         down_write(&fs_info->commit_root_sem);
 695         refcount_inc(&caching_ctl->count);
 696         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 697         up_write(&fs_info->commit_root_sem);
 698
 699         btrfs_get_block_group(cache);
 700
 701         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 702
 703         return ret;
 704 }
 705
 706 /*
 707  * return the block group that starts at or after bytenr
 708  */
 709 static struct btrfs_block_group_cache *
 710 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 711 {
 712         return block_group_cache_tree_search(info, bytenr, 0);
 713 }
 714
 715 /*
 716  * return the block group that contains the given bytenr
 717  */
 718 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 719                                                  struct btrfs_fs_info *info,
 720                                                  u64 bytenr)
 721 {
 722         return block_group_cache_tree_search(info, bytenr, 1);
 723 }
 724
 725 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 726                                                   u64 flags)
 727 {
 728         struct list_head *head = &info->space_info;
 729         struct btrfs_space_info *found;
 730
 731         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 732
 733         rcu_read_lock();
 734         list_for_each_entry_rcu(found, head, list) {
 735                 if (found->flags & flags) {
 736                         rcu_read_unlock();
 737                         return found;
 738                 }
 739         }
 740         rcu_read_unlock();
 741         return NULL;
 742 }
 743
 744 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 745                              bool metadata, u64 root_objectid)
 746 {
 747         struct btrfs_space_info *space_info;
 748         u64 flags;
 749
 750         if (metadata) {
 751                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 752                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 753                 else
 754                         flags = BTRFS_BLOCK_GROUP_METADATA;
 755         } else {
 756                 flags = BTRFS_BLOCK_GROUP_DATA;
 757         }
 758
 759         space_info = __find_space_info(fs_info, flags);
 760         ASSERT(space_info);
 761         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 762 }
 763
 764 /*
 765  * after adding space to the filesystem, we need to clear the full flags
 766  * on all the space infos.
 767  */
 768 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 769 {
 770         struct list_head *head = &info->space_info;
 771         struct btrfs_space_info *found;
 772
 773         rcu_read_lock();
 774         list_for_each_entry_rcu(found, head, list)
 775                 found->full = 0;
 776         rcu_read_unlock();
 777 }
 778
 779 /* simple helper to search for an existing data extent at a given offset */
 780 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 781 {
 782         int ret;
 783         struct btrfs_key key;
 784         struct btrfs_path *path;
 785
 786         path = btrfs_alloc_path();
 787         if (!path)
 788                 return -ENOMEM;
 789
 790         key.objectid = start;
 791         key.offset = len;
 792         key.type = BTRFS_EXTENT_ITEM_KEY;
 793         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 794         btrfs_free_path(path);
 795         return ret;
 796 }
 797
 798 /*
 799  * helper function to lookup reference count and flags of a tree block.
 800  *
 801  * the head node for delayed ref is used to store the sum of all the
 802  * reference count modifications queued up in the rbtree. the head
 803  * node may also store the extent flags to set. This way you can check
 804  * to see what the reference count and extent flags would be if all of
 805  * the delayed refs are not processed.
 806  */
 807 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 808                              struct btrfs_fs_info *fs_info, u64 bytenr,
 809                              u64 offset, int metadata, u64 *refs, u64 *flags)
 810 {
 811         struct btrfs_delayed_ref_head *head;
 812         struct btrfs_delayed_ref_root *delayed_refs;
 813         struct btrfs_path *path;
 814         struct btrfs_extent_item *ei;
 815         struct extent_buffer *leaf;
 816         struct btrfs_key key;
 817         u32 item_size;
 818         u64 num_refs;
 819         u64 extent_flags;
 820         int ret;
 821
 822         /*
 823          * If we don't have skinny metadata, don't bother doing anything
 824          * different
 825          */
 826         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 827                 offset = fs_info->nodesize;
 828                 metadata = 0;
 829         }
 830
 831         path = btrfs_alloc_path();
 832         if (!path)
 833                 return -ENOMEM;
 834
 835         if (!trans) {
 836                 path->skip_locking = 1;
 837                 path->search_commit_root = 1;
 838         }
 839
 840 search_again:
 841         key.objectid = bytenr;
 842         key.offset = offset;
 843         if (metadata)
 844                 key.type = BTRFS_METADATA_ITEM_KEY;
 845         else
 846                 key.type = BTRFS_EXTENT_ITEM_KEY;
 847
 848         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 849         if (ret < 0)
 850                 goto out_free;
 851
 852         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 853                 if (path->slots[0]) {
 854                         path->slots[0]--;
 855                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 856                                               path->slots[0]);
 857                         if (key.objectid == bytenr &&
 858                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 859                             key.offset == fs_info->nodesize)
 860                                 ret = 0;
 861                 }
 862         }
 863
 864         if (ret == 0) {
 865                 leaf = path->nodes[0];
 866                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 867                 if (item_size >= sizeof(*ei)) {
 868                         ei = btrfs_item_ptr(leaf, path->slots[0],
 869                                             struct btrfs_extent_item);
 870                         num_refs = btrfs_extent_refs(leaf, ei);
 871                         extent_flags = btrfs_extent_flags(leaf, ei);
 872                 } else {
 873 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 874                         struct btrfs_extent_item_v0 *ei0;
 875                         BUG_ON(item_size != sizeof(*ei0));
 876                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 877                                              struct btrfs_extent_item_v0);
 878                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 879                         /* FIXME: this isn't correct for data */
 880                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 881 #else
 882                         BUG();
 883 #endif
 884                 }
 885                 BUG_ON(num_refs == 0);
 886         } else {
 887                 num_refs = 0;
 888                 extent_flags = 0;
 889                 ret = 0;
 890         }
 891
 892         if (!trans)
 893                 goto out;
 894
 895         delayed_refs = &trans->transaction->delayed_refs;
 896         spin_lock(&delayed_refs->lock);
 897         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 898         if (head) {
 899                 if (!mutex_trylock(&head->mutex)) {
 900                         refcount_inc(&head->refs);
 901                         spin_unlock(&delayed_refs->lock);
 902
 903                         btrfs_release_path(path);
 904
 905                         /*
 906                          * Mutex was contended, block until it's released and try
 907                          * again
 908                          */
 909                         mutex_lock(&head->mutex);
 910                         mutex_unlock(&head->mutex);
 911                         btrfs_put_delayed_ref_head(head);
 912                         goto search_again;
 913                 }
 914                 spin_lock(&head->lock);
 915                 if (head->extent_op && head->extent_op->update_flags)
 916                         extent_flags |= head->extent_op->flags_to_set;
 917                 else
 918                         BUG_ON(num_refs == 0);
 919
 920                 num_refs += head->ref_mod;
 921                 spin_unlock(&head->lock);
 922                 mutex_unlock(&head->mutex);
 923         }
 924         spin_unlock(&delayed_refs->lock);
 925 out:
 926         WARN_ON(num_refs == 0);
 927         if (refs)
 928                 *refs = num_refs;
 929         if (flags)
 930                 *flags = extent_flags;
 931 out_free:
 932         btrfs_free_path(path);
 933         return ret;
 934 }
 935
 936 /*
 937  * Back reference rules.  Back refs have three main goals:
 938  *
 939  * 1) differentiate between all holders of references to an extent so that
 940  *    when a reference is dropped we can make sure it was a valid reference
 941  *    before freeing the extent.
 942  *
 943  * 2) Provide enough information to quickly find the holders of an extent
 944  *    if we notice a given block is corrupted or bad.
 945  *
 946  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 947  *    maintenance.  This is actually the same as #2, but with a slightly
 948  *    different use case.
 949  *
 950  * There are two kinds of back refs. The implicit back refs is optimized
 951  * for pointers in non-shared tree blocks. For a given pointer in a block,
 952  * back refs of this kind provide information about the block's owner tree
 953  * and the pointer's key. These information allow us to find the block by
 954  * b-tree searching. The full back refs is for pointers in tree blocks not
 955  * referenced by their owner trees. The location of tree block is recorded
 956  * in the back refs. Actually the full back refs is generic, and can be
 957  * used in all cases the implicit back refs is used. The major shortcoming
 958  * of the full back refs is its overhead. Every time a tree block gets
 959  * COWed, we have to update back refs entry for all pointers in it.
 960  *
 961  * For a newly allocated tree block, we use implicit back refs for
 962  * pointers in it. This means most tree related operations only involve
 963  * implicit back refs. For a tree block created in old transaction, the
 964  * only way to drop a reference to it is COW it. So we can detect the
 965  * event that tree block loses its owner tree's reference and do the
 966  * back refs conversion.
 967  *
 968  * When a tree block is COWed through a tree, there are four cases:
 969  *
 970  * The reference count of the block is one and the tree is the block's
 971  * owner tree. Nothing to do in this case.
 972  *
 973  * The reference count of the block is one and the tree is not the
 974  * block's owner tree. In this case, full back refs is used for pointers
 975  * in the block. Remove these full back refs, add implicit back refs for
 976  * every pointers in the new block.
 977  *
 978  * The reference count of the block is greater than one and the tree is
 979  * the block's owner tree. In this case, implicit back refs is used for
 980  * pointers in the block. Add full back refs for every pointers in the
 981  * block, increase lower level extents' reference counts. The original
 982  * implicit back refs are entailed to the new block.
 983  *
 984  * The reference count of the block is greater than one and the tree is
 985  * not the block's owner tree. Add implicit back refs for every pointer in
 986  * the new block, increase lower level extents' reference count.
 987  *
 988  * Back Reference Key composing:
 989  *
 990  * The key objectid corresponds to the first byte in the extent,
 991  * The key type is used to differentiate between types of back refs.
 992  * There are different meanings of the key offset for different types
 993  * of back refs.
 994  *
 995  * File extents can be referenced by:
 996  *
 997  * - multiple snapshots, subvolumes, or different generations in one subvol
 998  * - different files inside a single subvolume
 999  * - different offsets inside a file (bookend extents in file.c)
1000  *
1001  * The extent ref structure for the implicit back refs has fields for:
1002  *
1003  * - Objectid of the subvolume root
1004  * - objectid of the file holding the reference
1005  * - original offset in the file
1006  * - how many bookend extents
1007  *
1008  * The key offset for the implicit back refs is hash of the first
1009  * three fields.
1010  *
1011  * The extent ref structure for the full back refs has field for:
1012  *
1013  * - number of pointers in the tree leaf
1014  *
1015  * The key offset for the implicit back refs is the first byte of
1016  * the tree leaf
1017  *
1018  * When a file extent is allocated, The implicit back refs is used.
1019  * the fields are filled in:
1020  *
1021  *     (root_key.objectid, inode objectid, offset in file, 1)
1022  *
1023  * When a file extent is removed file truncation, we find the
1024  * corresponding implicit back refs and check the following fields:
1025  *
1026  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1027  *
1028  * Btree extents can be referenced by:
1029  *
1030  * - Different subvolumes
1031  *
1032  * Both the implicit back refs and the full back refs for tree blocks
1033  * only consist of key. The key offset for the implicit back refs is
1034  * objectid of block's owner tree. The key offset for the full back refs
1035  * is the first byte of parent block.
1036  *
1037  * When implicit back refs is used, information about the lowest key and
1038  * level of the tree block are required. These information are stored in
1039  * tree block info structure.
1040  */
1041
1042 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1043 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1044                                   struct btrfs_fs_info *fs_info,
1045                                   struct btrfs_path *path,
1046                                   u64 owner, u32 extra_size)
1047 {
1048         struct btrfs_root *root = fs_info->extent_root;
1049         struct btrfs_extent_item *item;
1050         struct btrfs_extent_item_v0 *ei0;
1051         struct btrfs_extent_ref_v0 *ref0;
1052         struct btrfs_tree_block_info *bi;
1053         struct extent_buffer *leaf;
1054         struct btrfs_key key;
1055         struct btrfs_key found_key;
1056         u32 new_size = sizeof(*item);
1057         u64 refs;
1058         int ret;
1059
1060         leaf = path->nodes[0];
1061         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1062
1063         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1064         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1065                              struct btrfs_extent_item_v0);
1066         refs = btrfs_extent_refs_v0(leaf, ei0);
1067
1068         if (owner == (u64)-1) {
1069                 while (1) {
1070                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1071                                 ret = btrfs_next_leaf(root, path);
1072                                 if (ret < 0)
1073                                         return ret;
1074                                 BUG_ON(ret > 0); /* Corruption */
1075                                 leaf = path->nodes[0];
1076                         }
1077                         btrfs_item_key_to_cpu(leaf, &found_key,
1078                                               path->slots[0]);
1079                         BUG_ON(key.objectid != found_key.objectid);
1080                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1081                                 path->slots[0]++;
1082                                 continue;
1083                         }
1084                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1085                                               struct btrfs_extent_ref_v0);
1086                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1087                         break;
1088                 }
1089         }
1090         btrfs_release_path(path);
1091
1092         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1093                 new_size += sizeof(*bi);
1094
1095         new_size -= sizeof(*ei0);
1096         ret = btrfs_search_slot(trans, root, &key, path,
1097                                 new_size + extra_size, 1);
1098         if (ret < 0)
1099                 return ret;
1100         BUG_ON(ret); /* Corruption */
1101
1102         btrfs_extend_item(fs_info, path, new_size);
1103
1104         leaf = path->nodes[0];
1105         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1106         btrfs_set_extent_refs(leaf, item, refs);
1107         /* FIXME: get real generation */
1108         btrfs_set_extent_generation(leaf, item, 0);
1109         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1110                 btrfs_set_extent_flags(leaf, item,
1111                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1112                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1113                 bi = (struct btrfs_tree_block_info *)(item + 1);
1114                 /* FIXME: get first key of the block */
1115                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1116                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1117         } else {
1118                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1119         }
1120         btrfs_mark_buffer_dirty(leaf);
1121         return 0;
1122 }
1123 #endif
1124
1125 /*
1126  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1127  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1128  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1129  */
1130 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1131                                      struct btrfs_extent_inline_ref *iref,
1132                                      enum btrfs_inline_ref_type is_data)
1133 {
1134         int type = btrfs_extent_inline_ref_type(eb, iref);
1135         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1136
1137         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1138             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1139             type == BTRFS_SHARED_DATA_REF_KEY ||
1140             type == BTRFS_EXTENT_DATA_REF_KEY) {
1141                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1142                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1143                                 return type;
1144                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1145                                 ASSERT(eb->fs_info);
1146                                 /*
1147                                  * Every shared one has parent tree
1148                                  * block, which must be aligned to
1149                                  * nodesize.
1150                                  */
1151                                 if (offset &&
1152                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1153                                         return type;
1154                         }
1155                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1156                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1157                                 return type;
1158                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1159                                 ASSERT(eb->fs_info);
1160                                 /*
1161                                  * Every shared one has parent tree
1162                                  * block, which must be aligned to
1163                                  * nodesize.
1164                                  */
1165                                 if (offset &&
1166                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1167                                         return type;
1168                         }
1169                 } else {
1170                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1171                         return type;
1172                 }
1173         }
1174
1175         btrfs_print_leaf((struct extent_buffer *)eb);
1176         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1177                   eb->start, type);
1178         WARN_ON(1);
1179
1180         return BTRFS_REF_TYPE_INVALID;
1181 }
1182
1183 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1184 {
1185         u32 high_crc = ~(u32)0;
1186         u32 low_crc = ~(u32)0;
1187         __le64 lenum;
1188
1189         lenum = cpu_to_le64(root_objectid);
1190         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1191         lenum = cpu_to_le64(owner);
1192         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1193         lenum = cpu_to_le64(offset);
1194         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1195
1196         return ((u64)high_crc << 31) ^ (u64)low_crc;
1197 }
1198
1199 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1200                                      struct btrfs_extent_data_ref *ref)
1201 {
1202         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1203                                     btrfs_extent_data_ref_objectid(leaf, ref),
1204                                     btrfs_extent_data_ref_offset(leaf, ref));
1205 }
1206
1207 static int match_extent_data_ref(struct extent_buffer *leaf,
1208                                  struct btrfs_extent_data_ref *ref,
1209                                  u64 root_objectid, u64 owner, u64 offset)
1210 {
1211         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1212             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1213             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1214                 return 0;
1215         return 1;
1216 }
1217
1218 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1219                                            struct btrfs_fs_info *fs_info,
1220                                            struct btrfs_path *path,
1221                                            u64 bytenr, u64 parent,
1222                                            u64 root_objectid,
1223                                            u64 owner, u64 offset)
1224 {
1225         struct btrfs_root *root = fs_info->extent_root;
1226         struct btrfs_key key;
1227         struct btrfs_extent_data_ref *ref;
1228         struct extent_buffer *leaf;
1229         u32 nritems;
1230         int ret;
1231         int recow;
1232         int err = -ENOENT;
1233
1234         key.objectid = bytenr;
1235         if (parent) {
1236                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1237                 key.offset = parent;
1238         } else {
1239                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1240                 key.offset = hash_extent_data_ref(root_objectid,
1241                                                   owner, offset);
1242         }
1243 again:
1244         recow = 0;
1245         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1246         if (ret < 0) {
1247                 err = ret;
1248                 goto fail;
1249         }
1250
1251         if (parent) {
1252                 if (!ret)
1253                         return 0;
1254 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1255                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1256                 btrfs_release_path(path);
1257                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1258                 if (ret < 0) {
1259                         err = ret;
1260                         goto fail;
1261                 }
1262                 if (!ret)
1263                         return 0;
1264 #endif
1265                 goto fail;
1266         }
1267
1268         leaf = path->nodes[0];
1269         nritems = btrfs_header_nritems(leaf);
1270         while (1) {
1271                 if (path->slots[0] >= nritems) {
1272                         ret = btrfs_next_leaf(root, path);
1273                         if (ret < 0)
1274                                 err = ret;
1275                         if (ret)
1276                                 goto fail;
1277
1278                         leaf = path->nodes[0];
1279                         nritems = btrfs_header_nritems(leaf);
1280                         recow = 1;
1281                 }
1282
1283                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1284                 if (key.objectid != bytenr ||
1285                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1286                         goto fail;
1287
1288                 ref = btrfs_item_ptr(leaf, path->slots[0],
1289                                      struct btrfs_extent_data_ref);
1290
1291                 if (match_extent_data_ref(leaf, ref, root_objectid,
1292                                           owner, offset)) {
1293                         if (recow) {
1294                                 btrfs_release_path(path);
1295                                 goto again;
1296                         }
1297                         err = 0;
1298                         break;
1299                 }
1300                 path->slots[0]++;
1301         }
1302 fail:
1303         return err;
1304 }
1305
1306 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1307                                            struct btrfs_fs_info *fs_info,
1308                                            struct btrfs_path *path,
1309                                            u64 bytenr, u64 parent,
1310                                            u64 root_objectid, u64 owner,
1311                                            u64 offset, int refs_to_add)
1312 {
1313         struct btrfs_root *root = fs_info->extent_root;
1314         struct btrfs_key key;
1315         struct extent_buffer *leaf;
1316         u32 size;
1317         u32 num_refs;
1318         int ret;
1319
1320         key.objectid = bytenr;
1321         if (parent) {
1322                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1323                 key.offset = parent;
1324                 size = sizeof(struct btrfs_shared_data_ref);
1325         } else {
1326                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1327                 key.offset = hash_extent_data_ref(root_objectid,
1328                                                   owner, offset);
1329                 size = sizeof(struct btrfs_extent_data_ref);
1330         }
1331
1332         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1333         if (ret && ret != -EEXIST)
1334                 goto fail;
1335
1336         leaf = path->nodes[0];
1337         if (parent) {
1338                 struct btrfs_shared_data_ref *ref;
1339                 ref = btrfs_item_ptr(leaf, path->slots[0],
1340                                      struct btrfs_shared_data_ref);
1341                 if (ret == 0) {
1342                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1343                 } else {
1344                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1345                         num_refs += refs_to_add;
1346                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1347                 }
1348         } else {
1349                 struct btrfs_extent_data_ref *ref;
1350                 while (ret == -EEXIST) {
1351                         ref = btrfs_item_ptr(leaf, path->slots[0],
1352                                              struct btrfs_extent_data_ref);
1353                         if (match_extent_data_ref(leaf, ref, root_objectid,
1354                                                   owner, offset))
1355                                 break;
1356                         btrfs_release_path(path);
1357                         key.offset++;
1358                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1359                                                       size);
1360                         if (ret && ret != -EEXIST)
1361                                 goto fail;
1362
1363                         leaf = path->nodes[0];
1364                 }
1365                 ref = btrfs_item_ptr(leaf, path->slots[0],
1366                                      struct btrfs_extent_data_ref);
1367                 if (ret == 0) {
1368                         btrfs_set_extent_data_ref_root(leaf, ref,
1369                                                        root_objectid);
1370                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1371                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1372                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1373                 } else {
1374                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1375                         num_refs += refs_to_add;
1376                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1377                 }
1378         }
1379         btrfs_mark_buffer_dirty(leaf);
1380         ret = 0;
1381 fail:
1382         btrfs_release_path(path);
1383         return ret;
1384 }
1385
1386 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1387                                            struct btrfs_fs_info *fs_info,
1388                                            struct btrfs_path *path,
1389                                            int refs_to_drop, int *last_ref)
1390 {
1391         struct btrfs_key key;
1392         struct btrfs_extent_data_ref *ref1 = NULL;
1393         struct btrfs_shared_data_ref *ref2 = NULL;
1394         struct extent_buffer *leaf;
1395         u32 num_refs = 0;
1396         int ret = 0;
1397
1398         leaf = path->nodes[0];
1399         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1400
1401         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1402                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1403                                       struct btrfs_extent_data_ref);
1404                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1405         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1406                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1407                                       struct btrfs_shared_data_ref);
1408                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1409 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1410         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1411                 struct btrfs_extent_ref_v0 *ref0;
1412                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1413                                       struct btrfs_extent_ref_v0);
1414                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1415 #endif
1416         } else {
1417                 BUG();
1418         }
1419
1420         BUG_ON(num_refs < refs_to_drop);
1421         num_refs -= refs_to_drop;
1422
1423         if (num_refs == 0) {
1424                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1425                 *last_ref = 1;
1426         } else {
1427                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1428                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1429                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1430                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1431 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1432                 else {
1433                         struct btrfs_extent_ref_v0 *ref0;
1434                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1435                                         struct btrfs_extent_ref_v0);
1436                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1437                 }
1438 #endif
1439                 btrfs_mark_buffer_dirty(leaf);
1440         }
1441         return ret;
1442 }
1443
1444 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1445                                           struct btrfs_extent_inline_ref *iref)
1446 {
1447         struct btrfs_key key;
1448         struct extent_buffer *leaf;
1449         struct btrfs_extent_data_ref *ref1;
1450         struct btrfs_shared_data_ref *ref2;
1451         u32 num_refs = 0;
1452         int type;
1453
1454         leaf = path->nodes[0];
1455         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1456         if (iref) {
1457                 /*
1458                  * If type is invalid, we should have bailed out earlier than
1459                  * this call.
1460                  */
1461                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1462                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1463                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1464                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1465                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1466                 } else {
1467                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1468                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1469                 }
1470         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1471                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1472                                       struct btrfs_extent_data_ref);
1473                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1474         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1475                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1476                                       struct btrfs_shared_data_ref);
1477                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1478 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1479         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1480                 struct btrfs_extent_ref_v0 *ref0;
1481                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1482                                       struct btrfs_extent_ref_v0);
1483                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1484 #endif
1485         } else {
1486                 WARN_ON(1);
1487         }
1488         return num_refs;
1489 }
1490
1491 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1492                                           struct btrfs_fs_info *fs_info,
1493                                           struct btrfs_path *path,
1494                                           u64 bytenr, u64 parent,
1495                                           u64 root_objectid)
1496 {
1497         struct btrfs_root *root = fs_info->extent_root;
1498         struct btrfs_key key;
1499         int ret;
1500
1501         key.objectid = bytenr;
1502         if (parent) {
1503                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1504                 key.offset = parent;
1505         } else {
1506                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1507                 key.offset = root_objectid;
1508         }
1509
1510         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1511         if (ret > 0)
1512                 ret = -ENOENT;
1513 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1514         if (ret == -ENOENT && parent) {
1515                 btrfs_release_path(path);
1516                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1517                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1518                 if (ret > 0)
1519                         ret = -ENOENT;
1520         }
1521 #endif
1522         return ret;
1523 }
1524
1525 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1526                                           struct btrfs_fs_info *fs_info,
1527                                           struct btrfs_path *path,
1528                                           u64 bytenr, u64 parent,
1529                                           u64 root_objectid)
1530 {
1531         struct btrfs_key key;
1532         int ret;
1533
1534         key.objectid = bytenr;
1535         if (parent) {
1536                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1537                 key.offset = parent;
1538         } else {
1539                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1540                 key.offset = root_objectid;
1541         }
1542
1543         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1544                                       path, &key, 0);
1545         btrfs_release_path(path);
1546         return ret;
1547 }
1548
1549 static inline int extent_ref_type(u64 parent, u64 owner)
1550 {
1551         int type;
1552         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1553                 if (parent > 0)
1554                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1555                 else
1556                         type = BTRFS_TREE_BLOCK_REF_KEY;
1557         } else {
1558                 if (parent > 0)
1559                         type = BTRFS_SHARED_DATA_REF_KEY;
1560                 else
1561                         type = BTRFS_EXTENT_DATA_REF_KEY;
1562         }
1563         return type;
1564 }
1565
1566 static int find_next_key(struct btrfs_path *path, int level,
1567                          struct btrfs_key *key)
1568
1569 {
1570         for (; level < BTRFS_MAX_LEVEL; level++) {
1571                 if (!path->nodes[level])
1572                         break;
1573                 if (path->slots[level] + 1 >=
1574                     btrfs_header_nritems(path->nodes[level]))
1575                         continue;
1576                 if (level == 0)
1577                         btrfs_item_key_to_cpu(path->nodes[level], key,
1578                                               path->slots[level] + 1);
1579                 else
1580                         btrfs_node_key_to_cpu(path->nodes[level], key,
1581                                               path->slots[level] + 1);
1582                 return 0;
1583         }
1584         return 1;
1585 }
1586
1587 /*
1588  * look for inline back ref. if back ref is found, *ref_ret is set
1589  * to the address of inline back ref, and 0 is returned.
1590  *
1591  * if back ref isn't found, *ref_ret is set to the address where it
1592  * should be inserted, and -ENOENT is returned.
1593  *
1594  * if insert is true and there are too many inline back refs, the path
1595  * points to the extent item, and -EAGAIN is returned.
1596  *
1597  * NOTE: inline back refs are ordered in the same way that back ref
1598  *       items in the tree are ordered.
1599  */
1600 static noinline_for_stack
1601 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1602                                  struct btrfs_fs_info *fs_info,
1603                                  struct btrfs_path *path,
1604                                  struct btrfs_extent_inline_ref **ref_ret,
1605                                  u64 bytenr, u64 num_bytes,
1606                                  u64 parent, u64 root_objectid,
1607                                  u64 owner, u64 offset, int insert)
1608 {
1609         struct btrfs_root *root = fs_info->extent_root;
1610         struct btrfs_key key;
1611         struct extent_buffer *leaf;
1612         struct btrfs_extent_item *ei;
1613         struct btrfs_extent_inline_ref *iref;
1614         u64 flags;
1615         u64 item_size;
1616         unsigned long ptr;
1617         unsigned long end;
1618         int extra_size;
1619         int type;
1620         int want;
1621         int ret;
1622         int err = 0;
1623         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1624         int needed;
1625
1626         key.objectid = bytenr;
1627         key.type = BTRFS_EXTENT_ITEM_KEY;
1628         key.offset = num_bytes;
1629
1630         want = extent_ref_type(parent, owner);
1631         if (insert) {
1632                 extra_size = btrfs_extent_inline_ref_size(want);
1633                 path->keep_locks = 1;
1634         } else
1635                 extra_size = -1;
1636
1637         /*
1638          * Owner is our level, so we can just add one to get the level for the
1639          * block we are interested in.
1640          */
1641         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1642                 key.type = BTRFS_METADATA_ITEM_KEY;
1643                 key.offset = owner;
1644         }
1645
1646 again:
1647         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1648         if (ret < 0) {
1649                 err = ret;
1650                 goto out;
1651         }
1652
1653         /*
1654          * We may be a newly converted file system which still has the old fat
1655          * extent entries for metadata, so try and see if we have one of those.
1656          */
1657         if (ret > 0 && skinny_metadata) {
1658                 skinny_metadata = false;
1659                 if (path->slots[0]) {
1660                         path->slots[0]--;
1661                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1662                                               path->slots[0]);
1663                         if (key.objectid == bytenr &&
1664                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1665                             key.offset == num_bytes)
1666                                 ret = 0;
1667                 }
1668                 if (ret) {
1669                         key.objectid = bytenr;
1670                         key.type = BTRFS_EXTENT_ITEM_KEY;
1671                         key.offset = num_bytes;
1672                         btrfs_release_path(path);
1673                         goto again;
1674                 }
1675         }
1676
1677         if (ret && !insert) {
1678                 err = -ENOENT;
1679                 goto out;
1680         } else if (WARN_ON(ret)) {
1681                 err = -EIO;
1682                 goto out;
1683         }
1684
1685         leaf = path->nodes[0];
1686         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1687 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1688         if (item_size < sizeof(*ei)) {
1689                 if (!insert) {
1690                         err = -ENOENT;
1691                         goto out;
1692                 }
1693                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1694                                              extra_size);
1695                 if (ret < 0) {
1696                         err = ret;
1697                         goto out;
1698                 }
1699                 leaf = path->nodes[0];
1700                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701         }
1702 #endif
1703         BUG_ON(item_size < sizeof(*ei));
1704
1705         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1706         flags = btrfs_extent_flags(leaf, ei);
1707
1708         ptr = (unsigned long)(ei + 1);
1709         end = (unsigned long)ei + item_size;
1710
1711         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1712                 ptr += sizeof(struct btrfs_tree_block_info);
1713                 BUG_ON(ptr > end);
1714         }
1715
1716         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1717                 needed = BTRFS_REF_TYPE_DATA;
1718         else
1719                 needed = BTRFS_REF_TYPE_BLOCK;
1720
1721         err = -ENOENT;
1722         while (1) {
1723                 if (ptr >= end) {
1724                         WARN_ON(ptr > end);
1725                         break;
1726                 }
1727                 iref = (struct btrfs_extent_inline_ref *)ptr;
1728                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1729                 if (type == BTRFS_REF_TYPE_INVALID) {
1730                         err = -EINVAL;
1731                         goto out;
1732                 }
1733
1734                 if (want < type)
1735                         break;
1736                 if (want > type) {
1737                         ptr += btrfs_extent_inline_ref_size(type);
1738                         continue;
1739                 }
1740
1741                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1742                         struct btrfs_extent_data_ref *dref;
1743                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1744                         if (match_extent_data_ref(leaf, dref, root_objectid,
1745                                                   owner, offset)) {
1746                                 err = 0;
1747                                 break;
1748                         }
1749                         if (hash_extent_data_ref_item(leaf, dref) <
1750                             hash_extent_data_ref(root_objectid, owner, offset))
1751                                 break;
1752                 } else {
1753                         u64 ref_offset;
1754                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1755                         if (parent > 0) {
1756                                 if (parent == ref_offset) {
1757                                         err = 0;
1758                                         break;
1759                                 }
1760                                 if (ref_offset < parent)
1761                                         break;
1762                         } else {
1763                                 if (root_objectid == ref_offset) {
1764                                         err = 0;
1765                                         break;
1766                                 }
1767                                 if (ref_offset < root_objectid)
1768                                         break;
1769                         }
1770                 }
1771                 ptr += btrfs_extent_inline_ref_size(type);
1772         }
1773         if (err == -ENOENT && insert) {
1774                 if (item_size + extra_size >=
1775                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1776                         err = -EAGAIN;
1777                         goto out;
1778                 }
1779                 /*
1780                  * To add new inline back ref, we have to make sure
1781                  * there is no corresponding back ref item.
1782                  * For simplicity, we just do not add new inline back
1783                  * ref if there is any kind of item for this block
1784                  */
1785                 if (find_next_key(path, 0, &key) == 0 &&
1786                     key.objectid == bytenr &&
1787                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1788                         err = -EAGAIN;
1789                         goto out;
1790                 }
1791         }
1792         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1793 out:
1794         if (insert) {
1795                 path->keep_locks = 0;
1796                 btrfs_unlock_up_safe(path, 1);
1797         }
1798         return err;
1799 }
1800
1801 /*
1802  * helper to add new inline back ref
1803  */
1804 static noinline_for_stack
1805 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1806                                  struct btrfs_path *path,
1807                                  struct btrfs_extent_inline_ref *iref,
1808                                  u64 parent, u64 root_objectid,
1809                                  u64 owner, u64 offset, int refs_to_add,
1810                                  struct btrfs_delayed_extent_op *extent_op)
1811 {
1812         struct extent_buffer *leaf;
1813         struct btrfs_extent_item *ei;
1814         unsigned long ptr;
1815         unsigned long end;
1816         unsigned long item_offset;
1817         u64 refs;
1818         int size;
1819         int type;
1820
1821         leaf = path->nodes[0];
1822         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1823         item_offset = (unsigned long)iref - (unsigned long)ei;
1824
1825         type = extent_ref_type(parent, owner);
1826         size = btrfs_extent_inline_ref_size(type);
1827
1828         btrfs_extend_item(fs_info, path, size);
1829
1830         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1831         refs = btrfs_extent_refs(leaf, ei);
1832         refs += refs_to_add;
1833         btrfs_set_extent_refs(leaf, ei, refs);
1834         if (extent_op)
1835                 __run_delayed_extent_op(extent_op, leaf, ei);
1836
1837         ptr = (unsigned long)ei + item_offset;
1838         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1839         if (ptr < end - size)
1840                 memmove_extent_buffer(leaf, ptr + size, ptr,
1841                                       end - size - ptr);
1842
1843         iref = (struct btrfs_extent_inline_ref *)ptr;
1844         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1845         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1846                 struct btrfs_extent_data_ref *dref;
1847                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1848                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1849                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1850                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1851                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1852         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1853                 struct btrfs_shared_data_ref *sref;
1854                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1855                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1856                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1857         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1858                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1859         } else {
1860                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1861         }
1862         btrfs_mark_buffer_dirty(leaf);
1863 }
1864
1865 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1866                                  struct btrfs_fs_info *fs_info,
1867                                  struct btrfs_path *path,
1868                                  struct btrfs_extent_inline_ref **ref_ret,
1869                                  u64 bytenr, u64 num_bytes, u64 parent,
1870                                  u64 root_objectid, u64 owner, u64 offset)
1871 {
1872         int ret;
1873
1874         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1875                                            bytenr, num_bytes, parent,
1876                                            root_objectid, owner, offset, 0);
1877         if (ret != -ENOENT)
1878                 return ret;
1879
1880         btrfs_release_path(path);
1881         *ref_ret = NULL;
1882
1883         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1884                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1885                                             parent, root_objectid);
1886         } else {
1887                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1888                                              parent, root_objectid, owner,
1889                                              offset);
1890         }
1891         return ret;
1892 }
1893
1894 /*
1895  * helper to update/remove inline back ref
1896  */
1897 static noinline_for_stack
1898 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1899                                   struct btrfs_path *path,
1900                                   struct btrfs_extent_inline_ref *iref,
1901                                   int refs_to_mod,
1902                                   struct btrfs_delayed_extent_op *extent_op,
1903                                   int *last_ref)
1904 {
1905         struct extent_buffer *leaf;
1906         struct btrfs_extent_item *ei;
1907         struct btrfs_extent_data_ref *dref = NULL;
1908         struct btrfs_shared_data_ref *sref = NULL;
1909         unsigned long ptr;
1910         unsigned long end;
1911         u32 item_size;
1912         int size;
1913         int type;
1914         u64 refs;
1915
1916         leaf = path->nodes[0];
1917         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1918         refs = btrfs_extent_refs(leaf, ei);
1919         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1920         refs += refs_to_mod;
1921         btrfs_set_extent_refs(leaf, ei, refs);
1922         if (extent_op)
1923                 __run_delayed_extent_op(extent_op, leaf, ei);
1924
1925         /*
1926          * If type is invalid, we should have bailed out after
1927          * lookup_inline_extent_backref().
1928          */
1929         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1930         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1931
1932         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1933                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1934                 refs = btrfs_extent_data_ref_count(leaf, dref);
1935         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1936                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1937                 refs = btrfs_shared_data_ref_count(leaf, sref);
1938         } else {
1939                 refs = 1;
1940                 BUG_ON(refs_to_mod != -1);
1941         }
1942
1943         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1944         refs += refs_to_mod;
1945
1946         if (refs > 0) {
1947                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1948                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1949                 else
1950                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1951         } else {
1952                 *last_ref = 1;
1953                 size =  btrfs_extent_inline_ref_size(type);
1954                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1955                 ptr = (unsigned long)iref;
1956                 end = (unsigned long)ei + item_size;
1957                 if (ptr + size < end)
1958                         memmove_extent_buffer(leaf, ptr, ptr + size,
1959                                               end - ptr - size);
1960                 item_size -= size;
1961                 btrfs_truncate_item(fs_info, path, item_size, 1);
1962         }
1963         btrfs_mark_buffer_dirty(leaf);
1964 }
1965
1966 static noinline_for_stack
1967 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1968                                  struct btrfs_fs_info *fs_info,
1969                                  struct btrfs_path *path,
1970                                  u64 bytenr, u64 num_bytes, u64 parent,
1971                                  u64 root_objectid, u64 owner,
1972                                  u64 offset, int refs_to_add,
1973                                  struct btrfs_delayed_extent_op *extent_op)
1974 {
1975         struct btrfs_extent_inline_ref *iref;
1976         int ret;
1977
1978         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1979                                            bytenr, num_bytes, parent,
1980                                            root_objectid, owner, offset, 1);
1981         if (ret == 0) {
1982                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1983                 update_inline_extent_backref(fs_info, path, iref,
1984                                              refs_to_add, extent_op, NULL);
1985         } else if (ret == -ENOENT) {
1986                 setup_inline_extent_backref(fs_info, path, iref, parent,
1987                                             root_objectid, owner, offset,
1988                                             refs_to_add, extent_op);
1989                 ret = 0;
1990         }
1991         return ret;
1992 }
1993
1994 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1995                                  struct btrfs_fs_info *fs_info,
1996                                  struct btrfs_path *path,
1997                                  u64 bytenr, u64 parent, u64 root_objectid,
1998                                  u64 owner, u64 offset, int refs_to_add)
1999 {
2000         int ret;
2001         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2002                 BUG_ON(refs_to_add != 1);
2003                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
2004                                             parent, root_objectid);
2005         } else {
2006                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2007                                              parent, root_objectid,
2008                                              owner, offset, refs_to_add);
2009         }
2010         return ret;
2011 }
2012
2013 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2014                                  struct btrfs_fs_info *fs_info,
2015                                  struct btrfs_path *path,
2016                                  struct btrfs_extent_inline_ref *iref,
2017                                  int refs_to_drop, int is_data, int *last_ref)
2018 {
2019         int ret = 0;
2020
2021         BUG_ON(!is_data && refs_to_drop != 1);
2022         if (iref) {
2023                 update_inline_extent_backref(fs_info, path, iref,
2024                                              -refs_to_drop, NULL, last_ref);
2025         } else if (is_data) {
2026                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2027                                              last_ref);
2028         } else {
2029                 *last_ref = 1;
2030                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2031         }
2032         return ret;
2033 }
2034
2035 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2036 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2037                                u64 *discarded_bytes)
2038 {
2039         int j, ret = 0;
2040         u64 bytes_left, end;
2041         u64 aligned_start = ALIGN(start, 1 << 9);
2042
2043         if (WARN_ON(start != aligned_start)) {
2044                 len -= aligned_start - start;
2045                 len = round_down(len, 1 << 9);
2046                 start = aligned_start;
2047         }
2048
2049         *discarded_bytes = 0;
2050
2051         if (!len)
2052                 return 0;
2053
2054         end = start + len;
2055         bytes_left = len;
2056
2057         /* Skip any superblocks on this device. */
2058         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2059                 u64 sb_start = btrfs_sb_offset(j);
2060                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2061                 u64 size = sb_start - start;
2062
2063                 if (!in_range(sb_start, start, bytes_left) &&
2064                     !in_range(sb_end, start, bytes_left) &&
2065                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2066                         continue;
2067
2068                 /*
2069                  * Superblock spans beginning of range.  Adjust start and
2070                  * try again.
2071                  */
2072                 if (sb_start <= start) {
2073                         start += sb_end - start;
2074                         if (start > end) {
2075                                 bytes_left = 0;
2076                                 break;
2077                         }
2078                         bytes_left = end - start;
2079                         continue;
2080                 }
2081
2082                 if (size) {
2083                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2084                                                    GFP_NOFS, 0);
2085                         if (!ret)
2086                                 *discarded_bytes += size;
2087                         else if (ret != -EOPNOTSUPP)
2088                                 return ret;
2089                 }
2090
2091                 start = sb_end;
2092                 if (start > end) {
2093                         bytes_left = 0;
2094                         break;
2095                 }
2096                 bytes_left = end - start;
2097         }
2098
2099         if (bytes_left) {
2100                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2101                                            GFP_NOFS, 0);
2102                 if (!ret)
2103                         *discarded_bytes += bytes_left;
2104         }
2105         return ret;
2106 }
2107
2108 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2109                          u64 num_bytes, u64 *actual_bytes)
2110 {
2111         int ret;
2112         u64 discarded_bytes = 0;
2113         struct btrfs_bio *bbio = NULL;
2114
2115
2116         /*
2117          * Avoid races with device replace and make sure our bbio has devices
2118          * associated to its stripes that don't go away while we are discarding.
2119          */
2120         btrfs_bio_counter_inc_blocked(fs_info);
2121         /* Tell the block device(s) that the sectors can be discarded */
2122         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2123                               &bbio, 0);
2124         /* Error condition is -ENOMEM */
2125         if (!ret) {
2126                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2127                 int i;
2128
2129
2130                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2131                         u64 bytes;
2132                         struct request_queue *req_q;
2133
2134                         if (!stripe->dev->bdev) {
2135                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2136                                 continue;
2137                         }
2138                         req_q = bdev_get_queue(stripe->dev->bdev);
2139                         if (!blk_queue_discard(req_q))
2140                                 continue;
2141
2142                         ret = btrfs_issue_discard(stripe->dev->bdev,
2143                                                   stripe->physical,
2144                                                   stripe->length,
2145                                                   &bytes);
2146                         if (!ret)
2147                                 discarded_bytes += bytes;
2148                         else if (ret != -EOPNOTSUPP)
2149                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2150
2151                         /*
2152                          * Just in case we get back EOPNOTSUPP for some reason,
2153                          * just ignore the return value so we don't screw up
2154                          * people calling discard_extent.
2155                          */
2156                         ret = 0;
2157                 }
2158                 btrfs_put_bbio(bbio);
2159         }
2160         btrfs_bio_counter_dec(fs_info);
2161
2162         if (actual_bytes)
2163                 *actual_bytes = discarded_bytes;
2164
2165
2166         if (ret == -EOPNOTSUPP)
2167                 ret = 0;
2168         return ret;
2169 }
2170
2171 /* Can return -ENOMEM */
2172 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2173                          struct btrfs_root *root,
2174                          u64 bytenr, u64 num_bytes, u64 parent,
2175                          u64 root_objectid, u64 owner, u64 offset)
2176 {
2177         struct btrfs_fs_info *fs_info = root->fs_info;
2178         int old_ref_mod, new_ref_mod;
2179         int ret;
2180
2181         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2182                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2183
2184         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2185                            owner, offset, BTRFS_ADD_DELAYED_REF);
2186
2187         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2188                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2189                                                  num_bytes, parent,
2190                                                  root_objectid, (int)owner,
2191                                                  BTRFS_ADD_DELAYED_REF, NULL,
2192                                                  &old_ref_mod, &new_ref_mod);
2193         } else {
2194                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2195                                                  num_bytes, parent,
2196                                                  root_objectid, owner, offset,
2197                                                  0, BTRFS_ADD_DELAYED_REF,
2198                                                  &old_ref_mod, &new_ref_mod);
2199         }
2200
2201         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2202                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2203
2204                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2205         }
2206
2207         return ret;
2208 }
2209
2210 /*
2211  * __btrfs_inc_extent_ref - insert backreference for a given extent
2212  *
2213  * @trans:          Handle of transaction
2214  *
2215  * @node:           The delayed ref node used to get the bytenr/length for
2216  *                  extent whose references are incremented.
2217  *
2218  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2219  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2220  *                  bytenr of the parent block. Since new extents are always
2221  *                  created with indirect references, this will only be the case
2222  *                  when relocating a shared extent. In that case, root_objectid
2223  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2224  *                  be 0
2225  *
2226  * @root_objectid:  The id of the root where this modification has originated,
2227  *                  this can be either one of the well-known metadata trees or
2228  *                  the subvolume id which references this extent.
2229  *
2230  * @owner:          For data extents it is the inode number of the owning file.
2231  *                  For metadata extents this parameter holds the level in the
2232  *                  tree of the extent.
2233  *
2234  * @offset:         For metadata extents the offset is ignored and is currently
2235  *                  always passed as 0. For data extents it is the fileoffset
2236  *                  this extent belongs to.
2237  *
2238  * @refs_to_add     Number of references to add
2239  *
2240  * @extent_op       Pointer to a structure, holding information necessary when
2241  *                  updating a tree block's flags
2242  *
2243  */
2244 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2245                                   struct btrfs_fs_info *fs_info,
2246                                   struct btrfs_delayed_ref_node *node,
2247                                   u64 parent, u64 root_objectid,
2248                                   u64 owner, u64 offset, int refs_to_add,
2249                                   struct btrfs_delayed_extent_op *extent_op)
2250 {
2251         struct btrfs_path *path;
2252         struct extent_buffer *leaf;
2253         struct btrfs_extent_item *item;
2254         struct btrfs_key key;
2255         u64 bytenr = node->bytenr;
2256         u64 num_bytes = node->num_bytes;
2257         u64 refs;
2258         int ret;
2259
2260         path = btrfs_alloc_path();
2261         if (!path)
2262                 return -ENOMEM;
2263
2264         path->reada = READA_FORWARD;
2265         path->leave_spinning = 1;
2266         /* this will setup the path even if it fails to insert the back ref */
2267         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2268                                            num_bytes, parent, root_objectid,
2269                                            owner, offset,
2270                                            refs_to_add, extent_op);
2271         if ((ret < 0 && ret != -EAGAIN) || !ret)
2272                 goto out;
2273
2274         /*
2275          * Ok we had -EAGAIN which means we didn't have space to insert and
2276          * inline extent ref, so just update the reference count and add a
2277          * normal backref.
2278          */
2279         leaf = path->nodes[0];
2280         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2281         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2282         refs = btrfs_extent_refs(leaf, item);
2283         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2284         if (extent_op)
2285                 __run_delayed_extent_op(extent_op, leaf, item);
2286
2287         btrfs_mark_buffer_dirty(leaf);
2288         btrfs_release_path(path);
2289
2290         path->reada = READA_FORWARD;
2291         path->leave_spinning = 1;
2292         /* now insert the actual backref */
2293         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2294                                     root_objectid, owner, offset, refs_to_add);
2295         if (ret)
2296                 btrfs_abort_transaction(trans, ret);
2297 out:
2298         btrfs_free_path(path);
2299         return ret;
2300 }
2301
2302 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2303                                 struct btrfs_fs_info *fs_info,
2304                                 struct btrfs_delayed_ref_node *node,
2305                                 struct btrfs_delayed_extent_op *extent_op,
2306                                 int insert_reserved)
2307 {
2308         int ret = 0;
2309         struct btrfs_delayed_data_ref *ref;
2310         struct btrfs_key ins;
2311         u64 parent = 0;
2312         u64 ref_root = 0;
2313         u64 flags = 0;
2314
2315         ins.objectid = node->bytenr;
2316         ins.offset = node->num_bytes;
2317         ins.type = BTRFS_EXTENT_ITEM_KEY;
2318
2319         ref = btrfs_delayed_node_to_data_ref(node);
2320         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2321
2322         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2323                 parent = ref->parent;
2324         ref_root = ref->root;
2325
2326         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2327                 if (extent_op)
2328                         flags |= extent_op->flags_to_set;
2329                 ret = alloc_reserved_file_extent(trans, fs_info,
2330                                                  parent, ref_root, flags,
2331                                                  ref->objectid, ref->offset,
2332                                                  &ins, node->ref_mod);
2333         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2334                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2335                                              ref_root, ref->objectid,
2336                                              ref->offset, node->ref_mod,
2337                                              extent_op);
2338         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2339                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2340                                           ref_root, ref->objectid,
2341                                           ref->offset, node->ref_mod,
2342                                           extent_op);
2343         } else {
2344                 BUG();
2345         }
2346         return ret;
2347 }
2348
2349 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2350                                     struct extent_buffer *leaf,
2351                                     struct btrfs_extent_item *ei)
2352 {
2353         u64 flags = btrfs_extent_flags(leaf, ei);
2354         if (extent_op->update_flags) {
2355                 flags |= extent_op->flags_to_set;
2356                 btrfs_set_extent_flags(leaf, ei, flags);
2357         }
2358
2359         if (extent_op->update_key) {
2360                 struct btrfs_tree_block_info *bi;
2361                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2362                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2363                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2364         }
2365 }
2366
2367 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2368                                  struct btrfs_fs_info *fs_info,
2369                                  struct btrfs_delayed_ref_head *head,
2370                                  struct btrfs_delayed_extent_op *extent_op)
2371 {
2372         struct btrfs_key key;
2373         struct btrfs_path *path;
2374         struct btrfs_extent_item *ei;
2375         struct extent_buffer *leaf;
2376         u32 item_size;
2377         int ret;
2378         int err = 0;
2379         int metadata = !extent_op->is_data;
2380
2381         if (trans->aborted)
2382                 return 0;
2383
2384         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2385                 metadata = 0;
2386
2387         path = btrfs_alloc_path();
2388         if (!path)
2389                 return -ENOMEM;
2390
2391         key.objectid = head->bytenr;
2392
2393         if (metadata) {
2394                 key.type = BTRFS_METADATA_ITEM_KEY;
2395                 key.offset = extent_op->level;
2396         } else {
2397                 key.type = BTRFS_EXTENT_ITEM_KEY;
2398                 key.offset = head->num_bytes;
2399         }
2400
2401 again:
2402         path->reada = READA_FORWARD;
2403         path->leave_spinning = 1;
2404         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2405         if (ret < 0) {
2406                 err = ret;
2407                 goto out;
2408         }
2409         if (ret > 0) {
2410                 if (metadata) {
2411                         if (path->slots[0] > 0) {
2412                                 path->slots[0]--;
2413                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2414                                                       path->slots[0]);
2415                                 if (key.objectid == head->bytenr &&
2416                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2417                                     key.offset == head->num_bytes)
2418                                         ret = 0;
2419                         }
2420                         if (ret > 0) {
2421                                 btrfs_release_path(path);
2422                                 metadata = 0;
2423
2424                                 key.objectid = head->bytenr;
2425                                 key.offset = head->num_bytes;
2426                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2427                                 goto again;
2428                         }
2429                 } else {
2430                         err = -EIO;
2431                         goto out;
2432                 }
2433         }
2434
2435         leaf = path->nodes[0];
2436         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2437 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2438         if (item_size < sizeof(*ei)) {
2439                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2440                 if (ret < 0) {
2441                         err = ret;
2442                         goto out;
2443                 }
2444                 leaf = path->nodes[0];
2445                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2446         }
2447 #endif
2448         BUG_ON(item_size < sizeof(*ei));
2449         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2450         __run_delayed_extent_op(extent_op, leaf, ei);
2451
2452         btrfs_mark_buffer_dirty(leaf);
2453 out:
2454         btrfs_free_path(path);
2455         return err;
2456 }
2457
2458 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2459                                 struct btrfs_fs_info *fs_info,
2460                                 struct btrfs_delayed_ref_node *node,
2461                                 struct btrfs_delayed_extent_op *extent_op,
2462                                 int insert_reserved)
2463 {
2464         int ret = 0;
2465         struct btrfs_delayed_tree_ref *ref;
2466         u64 parent = 0;
2467         u64 ref_root = 0;
2468
2469         ref = btrfs_delayed_node_to_tree_ref(node);
2470         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2471
2472         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2473                 parent = ref->parent;
2474         ref_root = ref->root;
2475
2476         if (node->ref_mod != 1) {
2477                 btrfs_err(fs_info,
2478         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2479                           node->bytenr, node->ref_mod, node->action, ref_root,
2480                           parent);
2481                 return -EIO;
2482         }
2483         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2484                 BUG_ON(!extent_op || !extent_op->update_flags);
2485                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2486         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2487                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2488                                              parent, ref_root,
2489                                              ref->level, 0, 1,
2490                                              extent_op);
2491         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2492                 ret = __btrfs_free_extent(trans, fs_info, node,
2493                                           parent, ref_root,
2494                                           ref->level, 0, 1, extent_op);
2495         } else {
2496                 BUG();
2497         }
2498         return ret;
2499 }
2500
2501 /* helper function to actually process a single delayed ref entry */
2502 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2503                                struct btrfs_fs_info *fs_info,
2504                                struct btrfs_delayed_ref_node *node,
2505                                struct btrfs_delayed_extent_op *extent_op,
2506                                int insert_reserved)
2507 {
2508         int ret = 0;
2509
2510         if (trans->aborted) {
2511                 if (insert_reserved)
2512                         btrfs_pin_extent(fs_info, node->bytenr,
2513                                          node->num_bytes, 1);
2514                 return 0;
2515         }
2516
2517         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2518             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2519                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2520                                            insert_reserved);
2521         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2522                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2523                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2524                                            insert_reserved);
2525         else
2526                 BUG();
2527         return ret;
2528 }
2529
2530 static inline struct btrfs_delayed_ref_node *
2531 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2532 {
2533         struct btrfs_delayed_ref_node *ref;
2534
2535         if (RB_EMPTY_ROOT(&head->ref_tree))
2536                 return NULL;
2537
2538         /*
2539          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2540          * This is to prevent a ref count from going down to zero, which deletes
2541          * the extent item from the extent tree, when there still are references
2542          * to add, which would fail because they would not find the extent item.
2543          */
2544         if (!list_empty(&head->ref_add_list))
2545                 return list_first_entry(&head->ref_add_list,
2546                                 struct btrfs_delayed_ref_node, add_list);
2547
2548         ref = rb_entry(rb_first(&head->ref_tree),
2549                        struct btrfs_delayed_ref_node, ref_node);
2550         ASSERT(list_empty(&ref->add_list));
2551         return ref;
2552 }
2553
2554 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2555                                       struct btrfs_delayed_ref_head *head)
2556 {
2557         spin_lock(&delayed_refs->lock);
2558         head->processing = 0;
2559         delayed_refs->num_heads_ready++;
2560         spin_unlock(&delayed_refs->lock);
2561         btrfs_delayed_ref_unlock(head);
2562 }
2563
2564 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2565                              struct btrfs_fs_info *fs_info,
2566                              struct btrfs_delayed_ref_head *head)
2567 {
2568         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2569         int ret;
2570
2571         if (!extent_op)
2572                 return 0;
2573         head->extent_op = NULL;
2574         if (head->must_insert_reserved) {
2575                 btrfs_free_delayed_extent_op(extent_op);
2576                 return 0;
2577         }
2578         spin_unlock(&head->lock);
2579         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2580         btrfs_free_delayed_extent_op(extent_op);
2581         return ret ? ret : 1;
2582 }
2583
2584 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2585                             struct btrfs_fs_info *fs_info,
2586                             struct btrfs_delayed_ref_head *head)
2587 {
2588         struct btrfs_delayed_ref_root *delayed_refs;
2589         int ret;
2590
2591         delayed_refs = &trans->transaction->delayed_refs;
2592
2593         ret = cleanup_extent_op(trans, fs_info, head);
2594         if (ret < 0) {
2595                 unselect_delayed_ref_head(delayed_refs, head);
2596                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2597                 return ret;
2598         } else if (ret) {
2599                 return ret;
2600         }
2601
2602         /*
2603          * Need to drop our head ref lock and re-acquire the delayed ref lock
2604          * and then re-check to make sure nobody got added.
2605          */
2606         spin_unlock(&head->lock);
2607         spin_lock(&delayed_refs->lock);
2608         spin_lock(&head->lock);
2609         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2610                 spin_unlock(&head->lock);
2611                 spin_unlock(&delayed_refs->lock);
2612                 return 1;
2613         }
2614         delayed_refs->num_heads--;
2615         rb_erase(&head->href_node, &delayed_refs->href_root);
2616         RB_CLEAR_NODE(&head->href_node);
2617         spin_unlock(&head->lock);
2618         spin_unlock(&delayed_refs->lock);
2619         atomic_dec(&delayed_refs->num_entries);
2620
2621         trace_run_delayed_ref_head(fs_info, head, 0);
2622
2623         if (head->total_ref_mod < 0) {
2624                 struct btrfs_space_info *space_info;
2625                 u64 flags;
2626
2627                 if (head->is_data)
2628                         flags = BTRFS_BLOCK_GROUP_DATA;
2629                 else if (head->is_system)
2630                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2631                 else
2632                         flags = BTRFS_BLOCK_GROUP_METADATA;
2633                 space_info = __find_space_info(fs_info, flags);
2634                 ASSERT(space_info);
2635                 percpu_counter_add(&space_info->total_bytes_pinned,
2636                                    -head->num_bytes);
2637
2638                 if (head->is_data) {
2639                         spin_lock(&delayed_refs->lock);
2640                         delayed_refs->pending_csums -= head->num_bytes;
2641                         spin_unlock(&delayed_refs->lock);
2642                 }
2643         }
2644
2645         if (head->must_insert_reserved) {
2646                 btrfs_pin_extent(fs_info, head->bytenr,
2647                                  head->num_bytes, 1);
2648                 if (head->is_data) {
2649                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2650                                               head->num_bytes);
2651                 }
2652         }
2653
2654         /* Also free its reserved qgroup space */
2655         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2656                                       head->qgroup_reserved);
2657         btrfs_delayed_ref_unlock(head);
2658         btrfs_put_delayed_ref_head(head);
2659         return 0;
2660 }
2661
2662 /*
2663  * Returns 0 on success or if called with an already aborted transaction.
2664  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2665  */
2666 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2667                                              unsigned long nr)
2668 {
2669         struct btrfs_fs_info *fs_info = trans->fs_info;
2670         struct btrfs_delayed_ref_root *delayed_refs;
2671         struct btrfs_delayed_ref_node *ref;
2672         struct btrfs_delayed_ref_head *locked_ref = NULL;
2673         struct btrfs_delayed_extent_op *extent_op;
2674         ktime_t start = ktime_get();
2675         int ret;
2676         unsigned long count = 0;
2677         unsigned long actual_count = 0;
2678         int must_insert_reserved = 0;
2679
2680         delayed_refs = &trans->transaction->delayed_refs;
2681         while (1) {
2682                 if (!locked_ref) {
2683                         if (count >= nr)
2684                                 break;
2685
2686                         spin_lock(&delayed_refs->lock);
2687                         locked_ref = btrfs_select_ref_head(trans);
2688                         if (!locked_ref) {
2689                                 spin_unlock(&delayed_refs->lock);
2690                                 break;
2691                         }
2692
2693                         /* grab the lock that says we are going to process
2694                          * all the refs for this head */
2695                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2696                         spin_unlock(&delayed_refs->lock);
2697                         /*
2698                          * we may have dropped the spin lock to get the head
2699                          * mutex lock, and that might have given someone else
2700                          * time to free the head.  If that's true, it has been
2701                          * removed from our list and we can move on.
2702                          */
2703                         if (ret == -EAGAIN) {
2704                                 locked_ref = NULL;
2705                                 count++;
2706                                 continue;
2707                         }
2708                 }
2709
2710                 /*
2711                  * We need to try and merge add/drops of the same ref since we
2712                  * can run into issues with relocate dropping the implicit ref
2713                  * and then it being added back again before the drop can
2714                  * finish.  If we merged anything we need to re-loop so we can
2715                  * get a good ref.
2716                  * Or we can get node references of the same type that weren't
2717                  * merged when created due to bumps in the tree mod seq, and
2718                  * we need to merge them to prevent adding an inline extent
2719                  * backref before dropping it (triggering a BUG_ON at
2720                  * insert_inline_extent_backref()).
2721                  */
2722                 spin_lock(&locked_ref->lock);
2723                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2724
2725                 ref = select_delayed_ref(locked_ref);
2726
2727                 if (ref && ref->seq &&
2728                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2729                         spin_unlock(&locked_ref->lock);
2730                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2731                         locked_ref = NULL;
2732                         cond_resched();
2733                         count++;
2734                         continue;
2735                 }
2736
2737                 /*
2738                  * We're done processing refs in this ref_head, clean everything
2739                  * up and move on to the next ref_head.
2740                  */
2741                 if (!ref) {
2742                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2743                         if (ret > 0 ) {
2744                                 /* We dropped our lock, we need to loop. */
2745                                 ret = 0;
2746                                 continue;
2747                         } else if (ret) {
2748                                 return ret;
2749                         }
2750                         locked_ref = NULL;
2751                         count++;
2752                         continue;
2753                 }
2754
2755                 actual_count++;
2756                 ref->in_tree = 0;
2757                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2758                 RB_CLEAR_NODE(&ref->ref_node);
2759                 if (!list_empty(&ref->add_list))
2760                         list_del(&ref->add_list);
2761                 /*
2762                  * When we play the delayed ref, also correct the ref_mod on
2763                  * head
2764                  */
2765                 switch (ref->action) {
2766                 case BTRFS_ADD_DELAYED_REF:
2767                 case BTRFS_ADD_DELAYED_EXTENT:
2768                         locked_ref->ref_mod -= ref->ref_mod;
2769                         break;
2770                 case BTRFS_DROP_DELAYED_REF:
2771                         locked_ref->ref_mod += ref->ref_mod;
2772                         break;
2773                 default:
2774                         WARN_ON(1);
2775                 }
2776                 atomic_dec(&delayed_refs->num_entries);
2777
2778                 /*
2779                  * Record the must-insert_reserved flag before we drop the spin
2780                  * lock.
2781                  */
2782                 must_insert_reserved = locked_ref->must_insert_reserved;
2783                 locked_ref->must_insert_reserved = 0;
2784
2785                 extent_op = locked_ref->extent_op;
2786                 locked_ref->extent_op = NULL;
2787                 spin_unlock(&locked_ref->lock);
2788
2789                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2790                                           must_insert_reserved);
2791
2792                 btrfs_free_delayed_extent_op(extent_op);
2793                 if (ret) {
2794                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2795                         btrfs_put_delayed_ref(ref);
2796                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2797                                     ret);
2798                         return ret;
2799                 }
2800
2801                 btrfs_put_delayed_ref(ref);
2802                 count++;
2803                 cond_resched();
2804         }
2805
2806         /*
2807          * We don't want to include ref heads since we can have empty ref heads
2808          * and those will drastically skew our runtime down since we just do
2809          * accounting, no actual extent tree updates.
2810          */
2811         if (actual_count > 0) {
2812                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2813                 u64 avg;
2814
2815                 /*
2816                  * We weigh the current average higher than our current runtime
2817                  * to avoid large swings in the average.
2818                  */
2819                 spin_lock(&delayed_refs->lock);
2820                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2821                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2822                 spin_unlock(&delayed_refs->lock);
2823         }
2824         return 0;
2825 }
2826
2827 #ifdef SCRAMBLE_DELAYED_REFS
2828 /*
2829  * Normally delayed refs get processed in ascending bytenr order. This
2830  * correlates in most cases to the order added. To expose dependencies on this
2831  * order, we start to process the tree in the middle instead of the beginning
2832  */
2833 static u64 find_middle(struct rb_root *root)
2834 {
2835         struct rb_node *n = root->rb_node;
2836         struct btrfs_delayed_ref_node *entry;
2837         int alt = 1;
2838         u64 middle;
2839         u64 first = 0, last = 0;
2840
2841         n = rb_first(root);
2842         if (n) {
2843                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2844                 first = entry->bytenr;
2845         }
2846         n = rb_last(root);
2847         if (n) {
2848                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2849                 last = entry->bytenr;
2850         }
2851         n = root->rb_node;
2852
2853         while (n) {
2854                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2855                 WARN_ON(!entry->in_tree);
2856
2857                 middle = entry->bytenr;
2858
2859                 if (alt)
2860                         n = n->rb_left;
2861                 else
2862                         n = n->rb_right;
2863
2864                 alt = 1 - alt;
2865         }
2866         return middle;
2867 }
2868 #endif
2869
2870 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2871 {
2872         u64 num_bytes;
2873
2874         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2875                              sizeof(struct btrfs_extent_inline_ref));
2876         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2877                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2878
2879         /*
2880          * We don't ever fill up leaves all the way so multiply by 2 just to be
2881          * closer to what we're really going to want to use.
2882          */
2883         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2884 }
2885
2886 /*
2887  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2888  * would require to store the csums for that many bytes.
2889  */
2890 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2891 {
2892         u64 csum_size;
2893         u64 num_csums_per_leaf;
2894         u64 num_csums;
2895
2896         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2897         num_csums_per_leaf = div64_u64(csum_size,
2898                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2899         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2900         num_csums += num_csums_per_leaf - 1;
2901         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2902         return num_csums;
2903 }
2904
2905 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2906                                        struct btrfs_fs_info *fs_info)
2907 {
2908         struct btrfs_block_rsv *global_rsv;
2909         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2910         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2911         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2912         u64 num_bytes, num_dirty_bgs_bytes;
2913         int ret = 0;
2914
2915         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2916         num_heads = heads_to_leaves(fs_info, num_heads);
2917         if (num_heads > 1)
2918                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2919         num_bytes <<= 1;
2920         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2921                                                         fs_info->nodesize;
2922         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2923                                                              num_dirty_bgs);
2924         global_rsv = &fs_info->global_block_rsv;
2925
2926         /*
2927          * If we can't allocate any more chunks lets make sure we have _lots_ of
2928          * wiggle room since running delayed refs can create more delayed refs.
2929          */
2930         if (global_rsv->space_info->full) {
2931                 num_dirty_bgs_bytes <<= 1;
2932                 num_bytes <<= 1;
2933         }
2934
2935         spin_lock(&global_rsv->lock);
2936         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2937                 ret = 1;
2938         spin_unlock(&global_rsv->lock);
2939         return ret;
2940 }
2941
2942 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2943                                        struct btrfs_fs_info *fs_info)
2944 {
2945         u64 num_entries =
2946                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2947         u64 avg_runtime;
2948         u64 val;
2949
2950         smp_mb();
2951         avg_runtime = fs_info->avg_delayed_ref_runtime;
2952         val = num_entries * avg_runtime;
2953         if (val >= NSEC_PER_SEC)
2954                 return 1;
2955         if (val >= NSEC_PER_SEC / 2)
2956                 return 2;
2957
2958         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2959 }
2960
2961 struct async_delayed_refs {
2962         struct btrfs_root *root;
2963         u64 transid;
2964         int count;
2965         int error;
2966         int sync;
2967         struct completion wait;
2968         struct btrfs_work work;
2969 };
2970
2971 static inline struct async_delayed_refs *
2972 to_async_delayed_refs(struct btrfs_work *work)
2973 {
2974         return container_of(work, struct async_delayed_refs, work);
2975 }
2976
2977 static void delayed_ref_async_start(struct btrfs_work *work)
2978 {
2979         struct async_delayed_refs *async = to_async_delayed_refs(work);
2980         struct btrfs_trans_handle *trans;
2981         struct btrfs_fs_info *fs_info = async->root->fs_info;
2982         int ret;
2983
2984         /* if the commit is already started, we don't need to wait here */
2985         if (btrfs_transaction_blocked(fs_info))
2986                 goto done;
2987
2988         trans = btrfs_join_transaction(async->root);
2989         if (IS_ERR(trans)) {
2990                 async->error = PTR_ERR(trans);
2991                 goto done;
2992         }
2993
2994         /*
2995          * trans->sync means that when we call end_transaction, we won't
2996          * wait on delayed refs
2997          */
2998         trans->sync = true;
2999
3000         /* Don't bother flushing if we got into a different transaction */
3001         if (trans->transid > async->transid)
3002                 goto end;
3003
3004         ret = btrfs_run_delayed_refs(trans, async->count);
3005         if (ret)
3006                 async->error = ret;
3007 end:
3008         ret = btrfs_end_transaction(trans);
3009         if (ret && !async->error)
3010                 async->error = ret;
3011 done:
3012         if (async->sync)
3013                 complete(&async->wait);
3014         else
3015                 kfree(async);
3016 }
3017
3018 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3019                                  unsigned long count, u64 transid, int wait)
3020 {
3021         struct async_delayed_refs *async;
3022         int ret;
3023
3024         async = kmalloc(sizeof(*async), GFP_NOFS);
3025         if (!async)
3026                 return -ENOMEM;
3027
3028         async->root = fs_info->tree_root;
3029         async->count = count;
3030         async->error = 0;
3031         async->transid = transid;
3032         if (wait)
3033                 async->sync = 1;
3034         else
3035                 async->sync = 0;
3036         init_completion(&async->wait);
3037
3038         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3039                         delayed_ref_async_start, NULL, NULL);
3040
3041         btrfs_queue_work(fs_info->extent_workers, &async->work);
3042
3043         if (wait) {
3044                 wait_for_completion(&async->wait);
3045                 ret = async->error;
3046                 kfree(async);
3047                 return ret;
3048         }
3049         return 0;
3050 }
3051
3052 /*
3053  * this starts processing the delayed reference count updates and
3054  * extent insertions we have queued up so far.  count can be
3055  * 0, which means to process everything in the tree at the start
3056  * of the run (but not newly added entries), or it can be some target
3057  * number you'd like to process.
3058  *
3059  * Returns 0 on success or if called with an aborted transaction
3060  * Returns <0 on error and aborts the transaction
3061  */
3062 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3063                            unsigned long count)
3064 {
3065         struct btrfs_fs_info *fs_info = trans->fs_info;
3066         struct rb_node *node;
3067         struct btrfs_delayed_ref_root *delayed_refs;
3068         struct btrfs_delayed_ref_head *head;
3069         int ret;
3070         int run_all = count == (unsigned long)-1;
3071         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3072
3073         /* We'll clean this up in btrfs_cleanup_transaction */
3074         if (trans->aborted)
3075                 return 0;
3076
3077         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3078                 return 0;
3079
3080         delayed_refs = &trans->transaction->delayed_refs;
3081         if (count == 0)
3082                 count = atomic_read(&delayed_refs->num_entries) * 2;
3083
3084 again:
3085 #ifdef SCRAMBLE_DELAYED_REFS
3086         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3087 #endif
3088         trans->can_flush_pending_bgs = false;
3089         ret = __btrfs_run_delayed_refs(trans, count);
3090         if (ret < 0) {
3091                 btrfs_abort_transaction(trans, ret);
3092                 return ret;
3093         }
3094
3095         if (run_all) {
3096                 if (!list_empty(&trans->new_bgs))
3097                         btrfs_create_pending_block_groups(trans);
3098
3099                 spin_lock(&delayed_refs->lock);
3100                 node = rb_first(&delayed_refs->href_root);
3101                 if (!node) {
3102                         spin_unlock(&delayed_refs->lock);
3103                         goto out;
3104                 }
3105                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3106                                 href_node);
3107                 refcount_inc(&head->refs);
3108                 spin_unlock(&delayed_refs->lock);
3109
3110                 /* Mutex was contended, block until it's released and retry. */
3111                 mutex_lock(&head->mutex);
3112                 mutex_unlock(&head->mutex);
3113
3114                 btrfs_put_delayed_ref_head(head);
3115                 cond_resched();
3116                 goto again;
3117         }
3118 out:
3119         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3120         return 0;
3121 }
3122
3123 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3124                                 struct btrfs_fs_info *fs_info,
3125                                 u64 bytenr, u64 num_bytes, u64 flags,
3126                                 int level, int is_data)
3127 {
3128         struct btrfs_delayed_extent_op *extent_op;
3129         int ret;
3130
3131         extent_op = btrfs_alloc_delayed_extent_op();
3132         if (!extent_op)
3133                 return -ENOMEM;
3134
3135         extent_op->flags_to_set = flags;
3136         extent_op->update_flags = true;
3137         extent_op->update_key = false;
3138         extent_op->is_data = is_data ? true : false;
3139         extent_op->level = level;
3140
3141         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3142                                           num_bytes, extent_op);
3143         if (ret)
3144                 btrfs_free_delayed_extent_op(extent_op);
3145         return ret;
3146 }
3147
3148 static noinline int check_delayed_ref(struct btrfs_root *root,
3149                                       struct btrfs_path *path,
3150                                       u64 objectid, u64 offset, u64 bytenr)
3151 {
3152         struct btrfs_delayed_ref_head *head;
3153         struct btrfs_delayed_ref_node *ref;
3154         struct btrfs_delayed_data_ref *data_ref;
3155         struct btrfs_delayed_ref_root *delayed_refs;
3156         struct btrfs_transaction *cur_trans;
3157         struct rb_node *node;
3158         int ret = 0;
3159
3160         spin_lock(&root->fs_info->trans_lock);
3161         cur_trans = root->fs_info->running_transaction;
3162         if (cur_trans)
3163                 refcount_inc(&cur_trans->use_count);
3164         spin_unlock(&root->fs_info->trans_lock);
3165         if (!cur_trans)
3166                 return 0;
3167
3168         delayed_refs = &cur_trans->delayed_refs;
3169         spin_lock(&delayed_refs->lock);
3170         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3171         if (!head) {
3172                 spin_unlock(&delayed_refs->lock);
3173                 btrfs_put_transaction(cur_trans);
3174                 return 0;
3175         }
3176
3177         if (!mutex_trylock(&head->mutex)) {
3178                 refcount_inc(&head->refs);
3179                 spin_unlock(&delayed_refs->lock);
3180
3181                 btrfs_release_path(path);
3182
3183                 /*
3184                  * Mutex was contended, block until it's released and let
3185                  * caller try again
3186                  */
3187                 mutex_lock(&head->mutex);
3188                 mutex_unlock(&head->mutex);
3189                 btrfs_put_delayed_ref_head(head);
3190                 btrfs_put_transaction(cur_trans);
3191                 return -EAGAIN;
3192         }
3193         spin_unlock(&delayed_refs->lock);
3194
3195         spin_lock(&head->lock);
3196         /*
3197          * XXX: We should replace this with a proper search function in the
3198          * future.
3199          */
3200         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3201                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3202                 /* If it's a shared ref we know a cross reference exists */
3203                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3204                         ret = 1;
3205                         break;
3206                 }
3207
3208                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3209
3210                 /*
3211                  * If our ref doesn't match the one we're currently looking at
3212                  * then we have a cross reference.
3213                  */
3214                 if (data_ref->root != root->root_key.objectid ||
3215                     data_ref->objectid != objectid ||
3216                     data_ref->offset != offset) {
3217                         ret = 1;
3218                         break;
3219                 }
3220         }
3221         spin_unlock(&head->lock);
3222         mutex_unlock(&head->mutex);
3223         btrfs_put_transaction(cur_trans);
3224         return ret;
3225 }
3226
3227 static noinline int check_committed_ref(struct btrfs_root *root,
3228                                         struct btrfs_path *path,
3229                                         u64 objectid, u64 offset, u64 bytenr)
3230 {
3231         struct btrfs_fs_info *fs_info = root->fs_info;
3232         struct btrfs_root *extent_root = fs_info->extent_root;
3233         struct extent_buffer *leaf;
3234         struct btrfs_extent_data_ref *ref;
3235         struct btrfs_extent_inline_ref *iref;
3236         struct btrfs_extent_item *ei;
3237         struct btrfs_key key;
3238         u32 item_size;
3239         int type;
3240         int ret;
3241
3242         key.objectid = bytenr;
3243         key.offset = (u64)-1;
3244         key.type = BTRFS_EXTENT_ITEM_KEY;
3245
3246         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3247         if (ret < 0)
3248                 goto out;
3249         BUG_ON(ret == 0); /* Corruption */
3250
3251         ret = -ENOENT;
3252         if (path->slots[0] == 0)
3253                 goto out;
3254
3255         path->slots[0]--;
3256         leaf = path->nodes[0];
3257         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3258
3259         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3260                 goto out;
3261
3262         ret = 1;
3263         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3264 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3265         if (item_size < sizeof(*ei)) {
3266                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3267                 goto out;
3268         }
3269 #endif
3270         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3271
3272         if (item_size != sizeof(*ei) +
3273             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3274                 goto out;
3275
3276         if (btrfs_extent_generation(leaf, ei) <=
3277             btrfs_root_last_snapshot(&root->root_item))
3278                 goto out;
3279
3280         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3281
3282         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3283         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3284                 goto out;
3285
3286         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3287         if (btrfs_extent_refs(leaf, ei) !=
3288             btrfs_extent_data_ref_count(leaf, ref) ||
3289             btrfs_extent_data_ref_root(leaf, ref) !=
3290             root->root_key.objectid ||
3291             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3292             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3293                 goto out;
3294
3295         ret = 0;
3296 out:
3297         return ret;
3298 }
3299
3300 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3301                           u64 bytenr)
3302 {
3303         struct btrfs_path *path;
3304         int ret;
3305         int ret2;
3306
3307         path = btrfs_alloc_path();
3308         if (!path)
3309                 return -ENOMEM;
3310
3311         do {
3312                 ret = check_committed_ref(root, path, objectid,
3313                                           offset, bytenr);
3314                 if (ret && ret != -ENOENT)
3315                         goto out;
3316
3317                 ret2 = check_delayed_ref(root, path, objectid,
3318                                          offset, bytenr);
3319         } while (ret2 == -EAGAIN);
3320
3321         if (ret2 && ret2 != -ENOENT) {
3322                 ret = ret2;
3323                 goto out;
3324         }
3325
3326         if (ret != -ENOENT || ret2 != -ENOENT)
3327                 ret = 0;
3328 out:
3329         btrfs_free_path(path);
3330         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3331                 WARN_ON(ret > 0);
3332         return ret;
3333 }
3334
3335 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3336                            struct btrfs_root *root,
3337                            struct extent_buffer *buf,
3338                            int full_backref, int inc)
3339 {
3340         struct btrfs_fs_info *fs_info = root->fs_info;
3341         u64 bytenr;
3342         u64 num_bytes;
3343         u64 parent;
3344         u64 ref_root;
3345         u32 nritems;
3346         struct btrfs_key key;
3347         struct btrfs_file_extent_item *fi;
3348         int i;
3349         int level;
3350         int ret = 0;
3351         int (*process_func)(struct btrfs_trans_handle *,
3352                             struct btrfs_root *,
3353                             u64, u64, u64, u64, u64, u64);
3354
3355
3356         if (btrfs_is_testing(fs_info))
3357                 return 0;
3358
3359         ref_root = btrfs_header_owner(buf);
3360         nritems = btrfs_header_nritems(buf);
3361         level = btrfs_header_level(buf);
3362
3363         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3364                 return 0;
3365
3366         if (inc)
3367                 process_func = btrfs_inc_extent_ref;
3368         else
3369                 process_func = btrfs_free_extent;
3370
3371         if (full_backref)
3372                 parent = buf->start;
3373         else
3374                 parent = 0;
3375
3376         for (i = 0; i < nritems; i++) {
3377                 if (level == 0) {
3378                         btrfs_item_key_to_cpu(buf, &key, i);
3379                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3380                                 continue;
3381                         fi = btrfs_item_ptr(buf, i,
3382                                             struct btrfs_file_extent_item);
3383                         if (btrfs_file_extent_type(buf, fi) ==
3384                             BTRFS_FILE_EXTENT_INLINE)
3385                                 continue;
3386                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3387                         if (bytenr == 0)
3388                                 continue;
3389
3390                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3391                         key.offset -= btrfs_file_extent_offset(buf, fi);
3392                         ret = process_func(trans, root, bytenr, num_bytes,
3393                                            parent, ref_root, key.objectid,
3394                                            key.offset);
3395                         if (ret)
3396                                 goto fail;
3397                 } else {
3398                         bytenr = btrfs_node_blockptr(buf, i);
3399                         num_bytes = fs_info->nodesize;
3400                         ret = process_func(trans, root, bytenr, num_bytes,
3401                                            parent, ref_root, level - 1, 0);
3402                         if (ret)
3403                                 goto fail;
3404                 }
3405         }
3406         return 0;
3407 fail:
3408         return ret;
3409 }
3410
3411 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3412                   struct extent_buffer *buf, int full_backref)
3413 {
3414         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3415 }
3416
3417 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3418                   struct extent_buffer *buf, int full_backref)
3419 {
3420         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3421 }
3422
3423 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3424                                  struct btrfs_fs_info *fs_info,
3425                                  struct btrfs_path *path,
3426                                  struct btrfs_block_group_cache *cache)
3427 {
3428         int ret;
3429         struct btrfs_root *extent_root = fs_info->extent_root;
3430         unsigned long bi;
3431         struct extent_buffer *leaf;
3432
3433         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3434         if (ret) {
3435                 if (ret > 0)
3436                         ret = -ENOENT;
3437                 goto fail;
3438         }
3439
3440         leaf = path->nodes[0];
3441         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3442         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3443         btrfs_mark_buffer_dirty(leaf);
3444 fail:
3445         btrfs_release_path(path);
3446         return ret;
3447
3448 }
3449
3450 static struct btrfs_block_group_cache *
3451 next_block_group(struct btrfs_fs_info *fs_info,
3452                  struct btrfs_block_group_cache *cache)
3453 {
3454         struct rb_node *node;
3455
3456         spin_lock(&fs_info->block_group_cache_lock);
3457
3458         /* If our block group was removed, we need a full search. */
3459         if (RB_EMPTY_NODE(&cache->cache_node)) {
3460                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3461
3462                 spin_unlock(&fs_info->block_group_cache_lock);
3463                 btrfs_put_block_group(cache);
3464                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3465         }
3466         node = rb_next(&cache->cache_node);
3467         btrfs_put_block_group(cache);
3468         if (node) {
3469                 cache = rb_entry(node, struct btrfs_block_group_cache,
3470                                  cache_node);
3471                 btrfs_get_block_group(cache);
3472         } else
3473                 cache = NULL;
3474         spin_unlock(&fs_info->block_group_cache_lock);
3475         return cache;
3476 }
3477
3478 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3479                             struct btrfs_trans_handle *trans,
3480                             struct btrfs_path *path)
3481 {
3482         struct btrfs_fs_info *fs_info = block_group->fs_info;
3483         struct btrfs_root *root = fs_info->tree_root;
3484         struct inode *inode = NULL;
3485         struct extent_changeset *data_reserved = NULL;
3486         u64 alloc_hint = 0;
3487         int dcs = BTRFS_DC_ERROR;
3488         u64 num_pages = 0;
3489         int retries = 0;
3490         int ret = 0;
3491
3492         /*
3493          * If this block group is smaller than 100 megs don't bother caching the
3494          * block group.
3495          */
3496         if (block_group->key.offset < (100 * SZ_1M)) {
3497                 spin_lock(&block_group->lock);
3498                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3499                 spin_unlock(&block_group->lock);
3500                 return 0;
3501         }
3502
3503         if (trans->aborted)
3504                 return 0;
3505 again:
3506         inode = lookup_free_space_inode(fs_info, block_group, path);
3507         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3508                 ret = PTR_ERR(inode);
3509                 btrfs_release_path(path);
3510                 goto out;
3511         }
3512
3513         if (IS_ERR(inode)) {
3514                 BUG_ON(retries);
3515                 retries++;
3516
3517                 if (block_group->ro)
3518                         goto out_free;
3519
3520                 ret = create_free_space_inode(fs_info, trans, block_group,
3521                                               path);
3522                 if (ret)
3523                         goto out_free;
3524                 goto again;
3525         }
3526
3527         /*
3528          * We want to set the generation to 0, that way if anything goes wrong
3529          * from here on out we know not to trust this cache when we load up next
3530          * time.
3531          */
3532         BTRFS_I(inode)->generation = 0;
3533         ret = btrfs_update_inode(trans, root, inode);
3534         if (ret) {
3535                 /*
3536                  * So theoretically we could recover from this, simply set the
3537                  * super cache generation to 0 so we know to invalidate the
3538                  * cache, but then we'd have to keep track of the block groups
3539                  * that fail this way so we know we _have_ to reset this cache
3540                  * before the next commit or risk reading stale cache.  So to
3541                  * limit our exposure to horrible edge cases lets just abort the
3542                  * transaction, this only happens in really bad situations
3543                  * anyway.
3544                  */
3545                 btrfs_abort_transaction(trans, ret);
3546                 goto out_put;
3547         }
3548         WARN_ON(ret);
3549
3550         /* We've already setup this transaction, go ahead and exit */
3551         if (block_group->cache_generation == trans->transid &&
3552             i_size_read(inode)) {
3553                 dcs = BTRFS_DC_SETUP;
3554                 goto out_put;
3555         }
3556
3557         if (i_size_read(inode) > 0) {
3558                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3559                                         &fs_info->global_block_rsv);
3560                 if (ret)
3561                         goto out_put;
3562
3563                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3564                 if (ret)
3565                         goto out_put;
3566         }
3567
3568         spin_lock(&block_group->lock);
3569         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3570             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3571                 /*
3572                  * don't bother trying to write stuff out _if_
3573                  * a) we're not cached,
3574                  * b) we're with nospace_cache mount option,
3575                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3576                  */
3577                 dcs = BTRFS_DC_WRITTEN;
3578                 spin_unlock(&block_group->lock);
3579                 goto out_put;
3580         }
3581         spin_unlock(&block_group->lock);
3582
3583         /*
3584          * We hit an ENOSPC when setting up the cache in this transaction, just
3585          * skip doing the setup, we've already cleared the cache so we're safe.
3586          */
3587         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3588                 ret = -ENOSPC;
3589                 goto out_put;
3590         }
3591
3592         /*
3593          * Try to preallocate enough space based on how big the block group is.
3594          * Keep in mind this has to include any pinned space which could end up
3595          * taking up quite a bit since it's not folded into the other space
3596          * cache.
3597          */
3598         num_pages = div_u64(block_group->key.offset, SZ_256M);
3599         if (!num_pages)
3600                 num_pages = 1;
3601
3602         num_pages *= 16;
3603         num_pages *= PAGE_SIZE;
3604
3605         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3606         if (ret)
3607                 goto out_put;
3608
3609         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3610                                               num_pages, num_pages,
3611                                               &alloc_hint);
3612         /*
3613          * Our cache requires contiguous chunks so that we don't modify a bunch
3614          * of metadata or split extents when writing the cache out, which means
3615          * we can enospc if we are heavily fragmented in addition to just normal
3616          * out of space conditions.  So if we hit this just skip setting up any
3617          * other block groups for this transaction, maybe we'll unpin enough
3618          * space the next time around.
3619          */
3620         if (!ret)
3621                 dcs = BTRFS_DC_SETUP;
3622         else if (ret == -ENOSPC)
3623                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3624
3625 out_put:
3626         iput(inode);
3627 out_free:
3628         btrfs_release_path(path);
3629 out:
3630         spin_lock(&block_group->lock);
3631         if (!ret && dcs == BTRFS_DC_SETUP)
3632                 block_group->cache_generation = trans->transid;
3633         block_group->disk_cache_state = dcs;
3634         spin_unlock(&block_group->lock);
3635
3636         extent_changeset_free(data_reserved);
3637         return ret;
3638 }
3639
3640 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3641                             struct btrfs_fs_info *fs_info)
3642 {
3643         struct btrfs_block_group_cache *cache, *tmp;
3644         struct btrfs_transaction *cur_trans = trans->transaction;
3645         struct btrfs_path *path;
3646
3647         if (list_empty(&cur_trans->dirty_bgs) ||
3648             !btrfs_test_opt(fs_info, SPACE_CACHE))
3649                 return 0;
3650
3651         path = btrfs_alloc_path();
3652         if (!path)
3653                 return -ENOMEM;
3654
3655         /* Could add new block groups, use _safe just in case */
3656         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3657                                  dirty_list) {
3658                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3659                         cache_save_setup(cache, trans, path);
3660         }
3661
3662         btrfs_free_path(path);
3663         return 0;
3664 }
3665
3666 /*
3667  * transaction commit does final block group cache writeback during a
3668  * critical section where nothing is allowed to change the FS.  This is
3669  * required in order for the cache to actually match the block group,
3670  * but can introduce a lot of latency into the commit.
3671  *
3672  * So, btrfs_start_dirty_block_groups is here to kick off block group
3673  * cache IO.  There's a chance we'll have to redo some of it if the
3674  * block group changes again during the commit, but it greatly reduces
3675  * the commit latency by getting rid of the easy block groups while
3676  * we're still allowing others to join the commit.
3677  */
3678 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3679 {
3680         struct btrfs_fs_info *fs_info = trans->fs_info;
3681         struct btrfs_block_group_cache *cache;
3682         struct btrfs_transaction *cur_trans = trans->transaction;
3683         int ret = 0;
3684         int should_put;
3685         struct btrfs_path *path = NULL;
3686         LIST_HEAD(dirty);
3687         struct list_head *io = &cur_trans->io_bgs;
3688         int num_started = 0;
3689         int loops = 0;
3690
3691         spin_lock(&cur_trans->dirty_bgs_lock);
3692         if (list_empty(&cur_trans->dirty_bgs)) {
3693                 spin_unlock(&cur_trans->dirty_bgs_lock);
3694                 return 0;
3695         }
3696         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3697         spin_unlock(&cur_trans->dirty_bgs_lock);
3698
3699 again:
3700         /*
3701          * make sure all the block groups on our dirty list actually
3702          * exist
3703          */
3704         btrfs_create_pending_block_groups(trans);
3705
3706         if (!path) {
3707                 path = btrfs_alloc_path();
3708                 if (!path)
3709                         return -ENOMEM;
3710         }
3711
3712         /*
3713          * cache_write_mutex is here only to save us from balance or automatic
3714          * removal of empty block groups deleting this block group while we are
3715          * writing out the cache
3716          */
3717         mutex_lock(&trans->transaction->cache_write_mutex);
3718         while (!list_empty(&dirty)) {
3719                 cache = list_first_entry(&dirty,
3720                                          struct btrfs_block_group_cache,
3721                                          dirty_list);
3722                 /*
3723                  * this can happen if something re-dirties a block
3724                  * group that is already under IO.  Just wait for it to
3725                  * finish and then do it all again
3726                  */
3727                 if (!list_empty(&cache->io_list)) {
3728                         list_del_init(&cache->io_list);
3729                         btrfs_wait_cache_io(trans, cache, path);
3730                         btrfs_put_block_group(cache);
3731                 }
3732
3733
3734                 /*
3735                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3736                  * if it should update the cache_state.  Don't delete
3737                  * until after we wait.
3738                  *
3739                  * Since we're not running in the commit critical section
3740                  * we need the dirty_bgs_lock to protect from update_block_group
3741                  */
3742                 spin_lock(&cur_trans->dirty_bgs_lock);
3743                 list_del_init(&cache->dirty_list);
3744                 spin_unlock(&cur_trans->dirty_bgs_lock);
3745
3746                 should_put = 1;
3747
3748                 cache_save_setup(cache, trans, path);
3749
3750                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3751                         cache->io_ctl.inode = NULL;
3752                         ret = btrfs_write_out_cache(fs_info, trans,
3753                                                     cache, path);
3754                         if (ret == 0 && cache->io_ctl.inode) {
3755                                 num_started++;
3756                                 should_put = 0;
3757
3758                                 /*
3759                                  * The cache_write_mutex is protecting the
3760                                  * io_list, also refer to the definition of
3761                                  * btrfs_transaction::io_bgs for more details
3762                                  */
3763                                 list_add_tail(&cache->io_list, io);
3764                         } else {
3765                                 /*
3766                                  * if we failed to write the cache, the
3767                                  * generation will be bad and life goes on
3768                                  */
3769                                 ret = 0;
3770                         }
3771                 }
3772                 if (!ret) {
3773                         ret = write_one_cache_group(trans, fs_info,
3774                                                     path, cache);
3775                         /*
3776                          * Our block group might still be attached to the list
3777                          * of new block groups in the transaction handle of some
3778                          * other task (struct btrfs_trans_handle->new_bgs). This
3779                          * means its block group item isn't yet in the extent
3780                          * tree. If this happens ignore the error, as we will
3781                          * try again later in the critical section of the
3782                          * transaction commit.
3783                          */
3784                         if (ret == -ENOENT) {
3785                                 ret = 0;
3786                                 spin_lock(&cur_trans->dirty_bgs_lock);
3787                                 if (list_empty(&cache->dirty_list)) {
3788                                         list_add_tail(&cache->dirty_list,
3789                                                       &cur_trans->dirty_bgs);
3790                                         btrfs_get_block_group(cache);
3791                                 }
3792                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3793                         } else if (ret) {
3794                                 btrfs_abort_transaction(trans, ret);
3795                         }
3796                 }
3797
3798                 /* if its not on the io list, we need to put the block group */
3799                 if (should_put)
3800                         btrfs_put_block_group(cache);
3801
3802                 if (ret)
3803                         break;
3804
3805                 /*
3806                  * Avoid blocking other tasks for too long. It might even save
3807                  * us from writing caches for block groups that are going to be
3808                  * removed.
3809                  */
3810                 mutex_unlock(&trans->transaction->cache_write_mutex);
3811                 mutex_lock(&trans->transaction->cache_write_mutex);
3812         }
3813         mutex_unlock(&trans->transaction->cache_write_mutex);
3814
3815         /*
3816          * go through delayed refs for all the stuff we've just kicked off
3817          * and then loop back (just once)
3818          */
3819         ret = btrfs_run_delayed_refs(trans, 0);
3820         if (!ret && loops == 0) {
3821                 loops++;
3822                 spin_lock(&cur_trans->dirty_bgs_lock);
3823                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3824                 /*
3825                  * dirty_bgs_lock protects us from concurrent block group
3826                  * deletes too (not just cache_write_mutex).
3827                  */
3828                 if (!list_empty(&dirty)) {
3829                         spin_unlock(&cur_trans->dirty_bgs_lock);
3830                         goto again;
3831                 }
3832                 spin_unlock(&cur_trans->dirty_bgs_lock);
3833         } else if (ret < 0) {
3834                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3835         }
3836
3837         btrfs_free_path(path);
3838         return ret;
3839 }
3840
3841 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3842                                    struct btrfs_fs_info *fs_info)
3843 {
3844         struct btrfs_block_group_cache *cache;
3845         struct btrfs_transaction *cur_trans = trans->transaction;
3846         int ret = 0;
3847         int should_put;
3848         struct btrfs_path *path;
3849         struct list_head *io = &cur_trans->io_bgs;
3850         int num_started = 0;
3851
3852         path = btrfs_alloc_path();
3853         if (!path)
3854                 return -ENOMEM;
3855
3856         /*
3857          * Even though we are in the critical section of the transaction commit,
3858          * we can still have concurrent tasks adding elements to this
3859          * transaction's list of dirty block groups. These tasks correspond to
3860          * endio free space workers started when writeback finishes for a
3861          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3862          * allocate new block groups as a result of COWing nodes of the root
3863          * tree when updating the free space inode. The writeback for the space
3864          * caches is triggered by an earlier call to
3865          * btrfs_start_dirty_block_groups() and iterations of the following
3866          * loop.
3867          * Also we want to do the cache_save_setup first and then run the
3868          * delayed refs to make sure we have the best chance at doing this all
3869          * in one shot.
3870          */
3871         spin_lock(&cur_trans->dirty_bgs_lock);
3872         while (!list_empty(&cur_trans->dirty_bgs)) {
3873                 cache = list_first_entry(&cur_trans->dirty_bgs,
3874                                          struct btrfs_block_group_cache,
3875                                          dirty_list);
3876
3877                 /*
3878                  * this can happen if cache_save_setup re-dirties a block
3879                  * group that is already under IO.  Just wait for it to
3880                  * finish and then do it all again
3881                  */
3882                 if (!list_empty(&cache->io_list)) {
3883                         spin_unlock(&cur_trans->dirty_bgs_lock);
3884                         list_del_init(&cache->io_list);
3885                         btrfs_wait_cache_io(trans, cache, path);
3886                         btrfs_put_block_group(cache);
3887                         spin_lock(&cur_trans->dirty_bgs_lock);
3888                 }
3889
3890                 /*
3891                  * don't remove from the dirty list until after we've waited
3892                  * on any pending IO
3893                  */
3894                 list_del_init(&cache->dirty_list);
3895                 spin_unlock(&cur_trans->dirty_bgs_lock);
3896                 should_put = 1;
3897
3898                 cache_save_setup(cache, trans, path);
3899
3900                 if (!ret)
3901                         ret = btrfs_run_delayed_refs(trans,
3902                                                      (unsigned long) -1);
3903
3904                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3905                         cache->io_ctl.inode = NULL;
3906                         ret = btrfs_write_out_cache(fs_info, trans,
3907                                                     cache, path);
3908                         if (ret == 0 && cache->io_ctl.inode) {
3909                                 num_started++;
3910                                 should_put = 0;
3911                                 list_add_tail(&cache->io_list, io);
3912                         } else {
3913                                 /*
3914                                  * if we failed to write the cache, the
3915                                  * generation will be bad and life goes on
3916                                  */
3917                                 ret = 0;
3918                         }
3919                 }
3920                 if (!ret) {
3921                         ret = write_one_cache_group(trans, fs_info,
3922                                                     path, cache);
3923                         /*
3924                          * One of the free space endio workers might have
3925                          * created a new block group while updating a free space
3926                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3927                          * and hasn't released its transaction handle yet, in
3928                          * which case the new block group is still attached to
3929                          * its transaction handle and its creation has not
3930                          * finished yet (no block group item in the extent tree
3931                          * yet, etc). If this is the case, wait for all free
3932                          * space endio workers to finish and retry. This is a
3933                          * a very rare case so no need for a more efficient and
3934                          * complex approach.
3935                          */
3936                         if (ret == -ENOENT) {
3937                                 wait_event(cur_trans->writer_wait,
3938                                    atomic_read(&cur_trans->num_writers) == 1);
3939                                 ret = write_one_cache_group(trans, fs_info,
3940                                                             path, cache);
3941                         }
3942                         if (ret)
3943                                 btrfs_abort_transaction(trans, ret);
3944                 }
3945
3946                 /* if its not on the io list, we need to put the block group */
3947                 if (should_put)
3948                         btrfs_put_block_group(cache);
3949                 spin_lock(&cur_trans->dirty_bgs_lock);
3950         }
3951         spin_unlock(&cur_trans->dirty_bgs_lock);
3952
3953         /*
3954          * Refer to the definition of io_bgs member for details why it's safe
3955          * to use it without any locking
3956          */
3957         while (!list_empty(io)) {
3958                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3959                                          io_list);
3960                 list_del_init(&cache->io_list);
3961                 btrfs_wait_cache_io(trans, cache, path);
3962                 btrfs_put_block_group(cache);
3963         }
3964
3965         btrfs_free_path(path);
3966         return ret;
3967 }
3968
3969 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3970 {
3971         struct btrfs_block_group_cache *block_group;
3972         int readonly = 0;
3973
3974         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3975         if (!block_group || block_group->ro)
3976                 readonly = 1;
3977         if (block_group)
3978                 btrfs_put_block_group(block_group);
3979         return readonly;
3980 }
3981
3982 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3983 {
3984         struct btrfs_block_group_cache *bg;
3985         bool ret = true;
3986
3987         bg = btrfs_lookup_block_group(fs_info, bytenr);
3988         if (!bg)
3989                 return false;
3990
3991         spin_lock(&bg->lock);
3992         if (bg->ro)
3993                 ret = false;
3994         else
3995                 atomic_inc(&bg->nocow_writers);
3996         spin_unlock(&bg->lock);
3997
3998         /* no put on block group, done by btrfs_dec_nocow_writers */
3999         if (!ret)
4000                 btrfs_put_block_group(bg);
4001
4002         return ret;
4003
4004 }
4005
4006 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
4007 {
4008         struct btrfs_block_group_cache *bg;
4009
4010         bg = btrfs_lookup_block_group(fs_info, bytenr);
4011         ASSERT(bg);
4012         if (atomic_dec_and_test(&bg->nocow_writers))
4013                 wake_up_var(&bg->nocow_writers);
4014         /*
4015          * Once for our lookup and once for the lookup done by a previous call
4016          * to btrfs_inc_nocow_writers()
4017          */
4018         btrfs_put_block_group(bg);
4019         btrfs_put_block_group(bg);
4020 }
4021
4022 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
4023 {
4024         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
4025 }
4026
4027 static const char *alloc_name(u64 flags)
4028 {
4029         switch (flags) {
4030         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4031                 return "mixed";
4032         case BTRFS_BLOCK_GROUP_METADATA:
4033                 return "metadata";
4034         case BTRFS_BLOCK_GROUP_DATA:
4035                 return "data";
4036         case BTRFS_BLOCK_GROUP_SYSTEM:
4037                 return "system";
4038         default:
4039                 WARN_ON(1);
4040                 return "invalid-combination";
4041         };
4042 }
4043
4044 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
4045 {
4046
4047         struct btrfs_space_info *space_info;
4048         int i;
4049         int ret;
4050
4051         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4052         if (!space_info)
4053                 return -ENOMEM;
4054
4055         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4056                                  GFP_KERNEL);
4057         if (ret) {
4058                 kfree(space_info);
4059                 return ret;
4060         }
4061
4062         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4063                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4064         init_rwsem(&space_info->groups_sem);
4065         spin_lock_init(&space_info->lock);
4066         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4067         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4068         init_waitqueue_head(&space_info->wait);
4069         INIT_LIST_HEAD(&space_info->ro_bgs);
4070         INIT_LIST_HEAD(&space_info->tickets);
4071         INIT_LIST_HEAD(&space_info->priority_tickets);
4072
4073         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4074                                     info->space_info_kobj, "%s",
4075                                     alloc_name(space_info->flags));
4076         if (ret) {
4077                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4078                 kfree(space_info);
4079                 return ret;
4080         }
4081
4082         list_add_rcu(&space_info->list, &info->space_info);
4083         if (flags & BTRFS_BLOCK_GROUP_DATA)
4084                 info->data_sinfo = space_info;
4085
4086         return ret;
4087 }
4088
4089 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4090                              u64 total_bytes, u64 bytes_used,
4091                              u64 bytes_readonly,
4092                              struct btrfs_space_info **space_info)
4093 {
4094         struct btrfs_space_info *found;
4095         int factor;
4096
4097         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4098                      BTRFS_BLOCK_GROUP_RAID10))
4099                 factor = 2;
4100         else
4101                 factor = 1;
4102
4103         found = __find_space_info(info, flags);
4104         ASSERT(found);
4105         spin_lock(&found->lock);
4106         found->total_bytes += total_bytes;
4107         found->disk_total += total_bytes * factor;
4108         found->bytes_used += bytes_used;
4109         found->disk_used += bytes_used * factor;
4110         found->bytes_readonly += bytes_readonly;
4111         if (total_bytes > 0)
4112                 found->full = 0;
4113         space_info_add_new_bytes(info, found, total_bytes -
4114                                  bytes_used - bytes_readonly);
4115         spin_unlock(&found->lock);
4116         *space_info = found;
4117 }
4118
4119 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4120 {
4121         u64 extra_flags = chunk_to_extended(flags) &
4122                                 BTRFS_EXTENDED_PROFILE_MASK;
4123
4124         write_seqlock(&fs_info->profiles_lock);
4125         if (flags & BTRFS_BLOCK_GROUP_DATA)
4126                 fs_info->avail_data_alloc_bits |= extra_flags;
4127         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4128                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4129         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4130                 fs_info->avail_system_alloc_bits |= extra_flags;
4131         write_sequnlock(&fs_info->profiles_lock);
4132 }
4133
4134 /*
4135  * returns target flags in extended format or 0 if restripe for this
4136  * chunk_type is not in progress
4137  *
4138  * should be called with balance_lock held
4139  */
4140 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4141 {
4142         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4143         u64 target = 0;
4144
4145         if (!bctl)
4146                 return 0;
4147
4148         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4149             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4150                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4151         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4152                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4153                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4154         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4155                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4156                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4157         }
4158
4159         return target;
4160 }
4161
4162 /*
4163  * @flags: available profiles in extended format (see ctree.h)
4164  *
4165  * Returns reduced profile in chunk format.  If profile changing is in
4166  * progress (either running or paused) picks the target profile (if it's
4167  * already available), otherwise falls back to plain reducing.
4168  */
4169 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4170 {
4171         u64 num_devices = fs_info->fs_devices->rw_devices;
4172         u64 target;
4173         u64 raid_type;
4174         u64 allowed = 0;
4175
4176         /*
4177          * see if restripe for this chunk_type is in progress, if so
4178          * try to reduce to the target profile
4179          */
4180         spin_lock(&fs_info->balance_lock);
4181         target = get_restripe_target(fs_info, flags);
4182         if (target) {
4183                 /* pick target profile only if it's already available */
4184                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4185                         spin_unlock(&fs_info->balance_lock);
4186                         return extended_to_chunk(target);
4187                 }
4188         }
4189         spin_unlock(&fs_info->balance_lock);
4190
4191         /* First, mask out the RAID levels which aren't possible */
4192         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4193                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4194                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4195         }
4196         allowed &= flags;
4197
4198         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4199                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4200         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4201                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4202         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4203                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4204         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4205                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4206         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4207                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4208
4209         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4210
4211         return extended_to_chunk(flags | allowed);
4212 }
4213
4214 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4215 {
4216         unsigned seq;
4217         u64 flags;
4218
4219         do {
4220                 flags = orig_flags;
4221                 seq = read_seqbegin(&fs_info->profiles_lock);
4222
4223                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4224                         flags |= fs_info->avail_data_alloc_bits;
4225                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4226                         flags |= fs_info->avail_system_alloc_bits;
4227                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4228                         flags |= fs_info->avail_metadata_alloc_bits;
4229         } while (read_seqretry(&fs_info->profiles_lock, seq));
4230
4231         return btrfs_reduce_alloc_profile(fs_info, flags);
4232 }
4233
4234 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4235 {
4236         struct btrfs_fs_info *fs_info = root->fs_info;
4237         u64 flags;
4238         u64 ret;
4239
4240         if (data)
4241                 flags = BTRFS_BLOCK_GROUP_DATA;
4242         else if (root == fs_info->chunk_root)
4243                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4244         else
4245                 flags = BTRFS_BLOCK_GROUP_METADATA;
4246
4247         ret = get_alloc_profile(fs_info, flags);
4248         return ret;
4249 }
4250
4251 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4252 {
4253         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4254 }
4255
4256 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4257 {
4258         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4259 }
4260
4261 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4262 {
4263         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4264 }
4265
4266 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4267                                  bool may_use_included)
4268 {
4269         ASSERT(s_info);
4270         return s_info->bytes_used + s_info->bytes_reserved +
4271                 s_info->bytes_pinned + s_info->bytes_readonly +
4272                 (may_use_included ? s_info->bytes_may_use : 0);
4273 }
4274
4275 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4276 {
4277         struct btrfs_root *root = inode->root;
4278         struct btrfs_fs_info *fs_info = root->fs_info;
4279         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4280         u64 used;
4281         int ret = 0;
4282         int need_commit = 2;
4283         int have_pinned_space;
4284
4285         /* make sure bytes are sectorsize aligned */
4286         bytes = ALIGN(bytes, fs_info->sectorsize);
4287
4288         if (btrfs_is_free_space_inode(inode)) {
4289                 need_commit = 0;
4290                 ASSERT(current->journal_info);
4291         }
4292
4293 again:
4294         /* make sure we have enough space to handle the data first */
4295         spin_lock(&data_sinfo->lock);
4296         used = btrfs_space_info_used(data_sinfo, true);
4297
4298         if (used + bytes > data_sinfo->total_bytes) {
4299                 struct btrfs_trans_handle *trans;
4300
4301                 /*
4302                  * if we don't have enough free bytes in this space then we need
4303                  * to alloc a new chunk.
4304                  */
4305                 if (!data_sinfo->full) {
4306                         u64 alloc_target;
4307
4308                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4309                         spin_unlock(&data_sinfo->lock);
4310
4311                         alloc_target = btrfs_data_alloc_profile(fs_info);
4312                         /*
4313                          * It is ugly that we don't call nolock join
4314                          * transaction for the free space inode case here.
4315                          * But it is safe because we only do the data space
4316                          * reservation for the free space cache in the
4317                          * transaction context, the common join transaction
4318                          * just increase the counter of the current transaction
4319                          * handler, doesn't try to acquire the trans_lock of
4320                          * the fs.
4321                          */
4322                         trans = btrfs_join_transaction(root);
4323                         if (IS_ERR(trans))
4324                                 return PTR_ERR(trans);
4325
4326                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4327                                              CHUNK_ALLOC_NO_FORCE);
4328                         btrfs_end_transaction(trans);
4329                         if (ret < 0) {
4330                                 if (ret != -ENOSPC)
4331                                         return ret;
4332                                 else {
4333                                         have_pinned_space = 1;
4334                                         goto commit_trans;
4335                                 }
4336                         }
4337
4338                         goto again;
4339                 }
4340
4341                 /*
4342                  * If we don't have enough pinned space to deal with this
4343                  * allocation, and no removed chunk in current transaction,
4344                  * don't bother committing the transaction.
4345                  */
4346                 have_pinned_space = percpu_counter_compare(
4347                         &data_sinfo->total_bytes_pinned,
4348                         used + bytes - data_sinfo->total_bytes);
4349                 spin_unlock(&data_sinfo->lock);
4350
4351                 /* commit the current transaction and try again */
4352 commit_trans:
4353                 if (need_commit) {
4354                         need_commit--;
4355
4356                         if (need_commit > 0) {
4357                                 btrfs_start_delalloc_roots(fs_info, -1);
4358                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4359                                                          (u64)-1);
4360                         }
4361
4362                         trans = btrfs_join_transaction(root);
4363                         if (IS_ERR(trans))
4364                                 return PTR_ERR(trans);
4365                         if (have_pinned_space >= 0 ||
4366                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4367                                      &trans->transaction->flags) ||
4368                             need_commit > 0) {
4369                                 ret = btrfs_commit_transaction(trans);
4370                                 if (ret)
4371                                         return ret;
4372                                 /*
4373                                  * The cleaner kthread might still be doing iput
4374                                  * operations. Wait for it to finish so that
4375                                  * more space is released.
4376                                  */
4377                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4378                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4379                                 goto again;
4380                         } else {
4381                                 btrfs_end_transaction(trans);
4382                         }
4383                 }
4384
4385                 trace_btrfs_space_reservation(fs_info,
4386                                               "space_info:enospc",
4387                                               data_sinfo->flags, bytes, 1);
4388                 return -ENOSPC;
4389         }
4390         data_sinfo->bytes_may_use += bytes;
4391         trace_btrfs_space_reservation(fs_info, "space_info",
4392                                       data_sinfo->flags, bytes, 1);
4393         spin_unlock(&data_sinfo->lock);
4394
4395         return ret;
4396 }
4397
4398 int btrfs_check_data_free_space(struct inode *inode,
4399                         struct extent_changeset **reserved, u64 start, u64 len)
4400 {
4401         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4402         int ret;
4403
4404         /* align the range */
4405         len = round_up(start + len, fs_info->sectorsize) -
4406               round_down(start, fs_info->sectorsize);
4407         start = round_down(start, fs_info->sectorsize);
4408
4409         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4410         if (ret < 0)
4411                 return ret;
4412
4413         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4414         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4415         if (ret < 0)
4416                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4417         else
4418                 ret = 0;
4419         return ret;
4420 }
4421
4422 /*
4423  * Called if we need to clear a data reservation for this inode
4424  * Normally in a error case.
4425  *
4426  * This one will *NOT* use accurate qgroup reserved space API, just for case
4427  * which we can't sleep and is sure it won't affect qgroup reserved space.
4428  * Like clear_bit_hook().
4429  */
4430 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4431                                             u64 len)
4432 {
4433         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4434         struct btrfs_space_info *data_sinfo;
4435
4436         /* Make sure the range is aligned to sectorsize */
4437         len = round_up(start + len, fs_info->sectorsize) -
4438               round_down(start, fs_info->sectorsize);
4439         start = round_down(start, fs_info->sectorsize);
4440
4441         data_sinfo = fs_info->data_sinfo;
4442         spin_lock(&data_sinfo->lock);
4443         if (WARN_ON(data_sinfo->bytes_may_use < len))
4444                 data_sinfo->bytes_may_use = 0;
4445         else
4446                 data_sinfo->bytes_may_use -= len;
4447         trace_btrfs_space_reservation(fs_info, "space_info",
4448                                       data_sinfo->flags, len, 0);
4449         spin_unlock(&data_sinfo->lock);
4450 }
4451
4452 /*
4453  * Called if we need to clear a data reservation for this inode
4454  * Normally in a error case.
4455  *
4456  * This one will handle the per-inode data rsv map for accurate reserved
4457  * space framework.
4458  */
4459 void btrfs_free_reserved_data_space(struct inode *inode,
4460                         struct extent_changeset *reserved, u64 start, u64 len)
4461 {
4462         struct btrfs_root *root = BTRFS_I(inode)->root;
4463
4464         /* Make sure the range is aligned to sectorsize */
4465         len = round_up(start + len, root->fs_info->sectorsize) -
4466               round_down(start, root->fs_info->sectorsize);
4467         start = round_down(start, root->fs_info->sectorsize);
4468
4469         btrfs_free_reserved_data_space_noquota(inode, start, len);
4470         btrfs_qgroup_free_data(inode, reserved, start, len);
4471 }
4472
4473 static void force_metadata_allocation(struct btrfs_fs_info *info)
4474 {
4475         struct list_head *head = &info->space_info;
4476         struct btrfs_space_info *found;
4477
4478         rcu_read_lock();
4479         list_for_each_entry_rcu(found, head, list) {
4480                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4481                         found->force_alloc = CHUNK_ALLOC_FORCE;
4482         }
4483         rcu_read_unlock();
4484 }
4485
4486 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4487 {
4488         return (global->size << 1);
4489 }
4490
4491 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4492                               struct btrfs_space_info *sinfo, int force)
4493 {
4494         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4495         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4496         u64 thresh;
4497
4498         if (force == CHUNK_ALLOC_FORCE)
4499                 return 1;
4500
4501         /*
4502          * We need to take into account the global rsv because for all intents
4503          * and purposes it's used space.  Don't worry about locking the
4504          * global_rsv, it doesn't change except when the transaction commits.
4505          */
4506         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4507                 bytes_used += calc_global_rsv_need_space(global_rsv);
4508
4509         /*
4510          * in limited mode, we want to have some free space up to
4511          * about 1% of the FS size.
4512          */
4513         if (force == CHUNK_ALLOC_LIMITED) {
4514                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4515                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4516
4517                 if (sinfo->total_bytes - bytes_used < thresh)
4518                         return 1;
4519         }
4520
4521         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4522                 return 0;
4523         return 1;
4524 }
4525
4526 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4527 {
4528         u64 num_dev;
4529
4530         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4531                     BTRFS_BLOCK_GROUP_RAID0 |
4532                     BTRFS_BLOCK_GROUP_RAID5 |
4533                     BTRFS_BLOCK_GROUP_RAID6))
4534                 num_dev = fs_info->fs_devices->rw_devices;
4535         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4536                 num_dev = 2;
4537         else
4538                 num_dev = 1;    /* DUP or single */
4539
4540         return num_dev;
4541 }
4542
4543 /*
4544  * If @is_allocation is true, reserve space in the system space info necessary
4545  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4546  * removing a chunk.
4547  */
4548 void check_system_chunk(struct btrfs_trans_handle *trans,
4549                         struct btrfs_fs_info *fs_info, u64 type)
4550 {
4551         struct btrfs_space_info *info;
4552         u64 left;
4553         u64 thresh;
4554         int ret = 0;
4555         u64 num_devs;
4556
4557         /*
4558          * Needed because we can end up allocating a system chunk and for an
4559          * atomic and race free space reservation in the chunk block reserve.
4560          */
4561         lockdep_assert_held(&fs_info->chunk_mutex);
4562
4563         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4564         spin_lock(&info->lock);
4565         left = info->total_bytes - btrfs_space_info_used(info, true);
4566         spin_unlock(&info->lock);
4567
4568         num_devs = get_profile_num_devs(fs_info, type);
4569
4570         /* num_devs device items to update and 1 chunk item to add or remove */
4571         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4572                 btrfs_calc_trans_metadata_size(fs_info, 1);
4573
4574         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4575                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4576                            left, thresh, type);
4577                 dump_space_info(fs_info, info, 0, 0);
4578         }
4579
4580         if (left < thresh) {
4581                 u64 flags = btrfs_system_alloc_profile(fs_info);
4582
4583                 /*
4584                  * Ignore failure to create system chunk. We might end up not
4585                  * needing it, as we might not need to COW all nodes/leafs from
4586                  * the paths we visit in the chunk tree (they were already COWed
4587                  * or created in the current transaction for example).
4588                  */
4589                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4590         }
4591
4592         if (!ret) {
4593                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4594                                           &fs_info->chunk_block_rsv,
4595                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4596                 if (!ret)
4597                         trans->chunk_bytes_reserved += thresh;
4598         }
4599 }
4600
4601 /*
4602  * If force is CHUNK_ALLOC_FORCE:
4603  *    - return 1 if it successfully allocates a chunk,
4604  *    - return errors including -ENOSPC otherwise.
4605  * If force is NOT CHUNK_ALLOC_FORCE:
4606  *    - return 0 if it doesn't need to allocate a new chunk,
4607  *    - return 1 if it successfully allocates a chunk,
4608  *    - return errors including -ENOSPC otherwise.
4609  */
4610 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4611                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4612 {
4613         struct btrfs_space_info *space_info;
4614         int wait_for_alloc = 0;
4615         int ret = 0;
4616
4617         /* Don't re-enter if we're already allocating a chunk */
4618         if (trans->allocating_chunk)
4619                 return -ENOSPC;
4620
4621         space_info = __find_space_info(fs_info, flags);
4622         ASSERT(space_info);
4623
4624 again:
4625         spin_lock(&space_info->lock);
4626         if (force < space_info->force_alloc)
4627                 force = space_info->force_alloc;
4628         if (space_info->full) {
4629                 if (should_alloc_chunk(fs_info, space_info, force))
4630                         ret = -ENOSPC;
4631                 else
4632                         ret = 0;
4633                 spin_unlock(&space_info->lock);
4634                 return ret;
4635         }
4636
4637         if (!should_alloc_chunk(fs_info, space_info, force)) {
4638                 spin_unlock(&space_info->lock);
4639                 return 0;
4640         } else if (space_info->chunk_alloc) {
4641                 wait_for_alloc = 1;
4642         } else {
4643                 space_info->chunk_alloc = 1;
4644         }
4645
4646         spin_unlock(&space_info->lock);
4647
4648         mutex_lock(&fs_info->chunk_mutex);
4649
4650         /*
4651          * The chunk_mutex is held throughout the entirety of a chunk
4652          * allocation, so once we've acquired the chunk_mutex we know that the
4653          * other guy is done and we need to recheck and see if we should
4654          * allocate.
4655          */
4656         if (wait_for_alloc) {
4657                 mutex_unlock(&fs_info->chunk_mutex);
4658                 wait_for_alloc = 0;
4659                 cond_resched();
4660                 goto again;
4661         }
4662
4663         trans->allocating_chunk = true;
4664
4665         /*
4666          * If we have mixed data/metadata chunks we want to make sure we keep
4667          * allocating mixed chunks instead of individual chunks.
4668          */
4669         if (btrfs_mixed_space_info(space_info))
4670                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4671
4672         /*
4673          * if we're doing a data chunk, go ahead and make sure that
4674          * we keep a reasonable number of metadata chunks allocated in the
4675          * FS as well.
4676          */
4677         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4678                 fs_info->data_chunk_allocations++;
4679                 if (!(fs_info->data_chunk_allocations %
4680                       fs_info->metadata_ratio))
4681                         force_metadata_allocation(fs_info);
4682         }
4683
4684         /*
4685          * Check if we have enough space in SYSTEM chunk because we may need
4686          * to update devices.
4687          */
4688         check_system_chunk(trans, fs_info, flags);
4689
4690         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4691         trans->allocating_chunk = false;
4692
4693         spin_lock(&space_info->lock);
4694         if (ret < 0) {
4695                 if (ret == -ENOSPC)
4696                         space_info->full = 1;
4697                 else
4698                         goto out;
4699         } else {
4700                 ret = 1;
4701         }
4702
4703         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4704 out:
4705         space_info->chunk_alloc = 0;
4706         spin_unlock(&space_info->lock);
4707         mutex_unlock(&fs_info->chunk_mutex);
4708         /*
4709          * When we allocate a new chunk we reserve space in the chunk block
4710          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4711          * add new nodes/leafs to it if we end up needing to do it when
4712          * inserting the chunk item and updating device items as part of the
4713          * second phase of chunk allocation, performed by
4714          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4715          * large number of new block groups to create in our transaction
4716          * handle's new_bgs list to avoid exhausting the chunk block reserve
4717          * in extreme cases - like having a single transaction create many new
4718          * block groups when starting to write out the free space caches of all
4719          * the block groups that were made dirty during the lifetime of the
4720          * transaction.
4721          */
4722         if (trans->can_flush_pending_bgs &&
4723             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4724                 btrfs_create_pending_block_groups(trans);
4725                 btrfs_trans_release_chunk_metadata(trans);
4726         }
4727         return ret;
4728 }
4729
4730 static int can_overcommit(struct btrfs_fs_info *fs_info,
4731                           struct btrfs_space_info *space_info, u64 bytes,
4732                           enum btrfs_reserve_flush_enum flush,
4733                           bool system_chunk)
4734 {
4735         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4736         u64 profile;
4737         u64 space_size;
4738         u64 avail;
4739         u64 used;
4740
4741         /* Don't overcommit when in mixed mode. */
4742         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4743                 return 0;
4744
4745         if (system_chunk)
4746                 profile = btrfs_system_alloc_profile(fs_info);
4747         else
4748                 profile = btrfs_metadata_alloc_profile(fs_info);
4749
4750         used = btrfs_space_info_used(space_info, false);
4751
4752         /*
4753          * We only want to allow over committing if we have lots of actual space
4754          * free, but if we don't have enough space to handle the global reserve
4755          * space then we could end up having a real enospc problem when trying
4756          * to allocate a chunk or some other such important allocation.
4757          */
4758         spin_lock(&global_rsv->lock);
4759         space_size = calc_global_rsv_need_space(global_rsv);
4760         spin_unlock(&global_rsv->lock);
4761         if (used + space_size >= space_info->total_bytes)
4762                 return 0;
4763
4764         used += space_info->bytes_may_use;
4765
4766         avail = atomic64_read(&fs_info->free_chunk_space);
4767
4768         /*
4769          * If we have dup, raid1 or raid10 then only half of the free
4770          * space is actually useable.  For raid56, the space info used
4771          * doesn't include the parity drive, so we don't have to
4772          * change the math
4773          */
4774         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4775                        BTRFS_BLOCK_GROUP_RAID1 |
4776                        BTRFS_BLOCK_GROUP_RAID10))
4777                 avail >>= 1;
4778
4779         /*
4780          * If we aren't flushing all things, let us overcommit up to
4781          * 1/2th of the space. If we can flush, don't let us overcommit
4782          * too much, let it overcommit up to 1/8 of the space.
4783          */
4784         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4785                 avail >>= 3;
4786         else
4787                 avail >>= 1;
4788
4789         if (used + bytes < space_info->total_bytes + avail)
4790                 return 1;
4791         return 0;
4792 }
4793
4794 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4795                                          unsigned long nr_pages, int nr_items)
4796 {
4797         struct super_block *sb = fs_info->sb;
4798
4799         if (down_read_trylock(&sb->s_umount)) {
4800                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4801                 up_read(&sb->s_umount);
4802         } else {
4803                 /*
4804                  * We needn't worry the filesystem going from r/w to r/o though
4805                  * we don't acquire ->s_umount mutex, because the filesystem
4806                  * should guarantee the delalloc inodes list be empty after
4807                  * the filesystem is readonly(all dirty pages are written to
4808                  * the disk).
4809                  */
4810                 btrfs_start_delalloc_roots(fs_info, nr_items);
4811                 if (!current->journal_info)
4812                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4813         }
4814 }
4815
4816 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4817                                         u64 to_reclaim)
4818 {
4819         u64 bytes;
4820         u64 nr;
4821
4822         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4823         nr = div64_u64(to_reclaim, bytes);
4824         if (!nr)
4825                 nr = 1;
4826         return nr;
4827 }
4828
4829 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4830
4831 /*
4832  * shrink metadata reservation for delalloc
4833  */
4834 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4835                             u64 orig, bool wait_ordered)
4836 {
4837         struct btrfs_space_info *space_info;
4838         struct btrfs_trans_handle *trans;
4839         u64 delalloc_bytes;
4840         u64 max_reclaim;
4841         u64 items;
4842         long time_left;
4843         unsigned long nr_pages;
4844         int loops;
4845
4846         /* Calc the number of the pages we need flush for space reservation */
4847         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4848         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4849
4850         trans = (struct btrfs_trans_handle *)current->journal_info;
4851         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4852
4853         delalloc_bytes = percpu_counter_sum_positive(
4854                                                 &fs_info->delalloc_bytes);
4855         if (delalloc_bytes == 0) {
4856                 if (trans)
4857                         return;
4858                 if (wait_ordered)
4859                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4860                 return;
4861         }
4862
4863         loops = 0;
4864         while (delalloc_bytes && loops < 3) {
4865                 max_reclaim = min(delalloc_bytes, to_reclaim);
4866                 nr_pages = max_reclaim >> PAGE_SHIFT;
4867                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4868                 /*
4869                  * We need to wait for the async pages to actually start before
4870                  * we do anything.
4871                  */
4872                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4873                 if (!max_reclaim)
4874                         goto skip_async;
4875
4876                 if (max_reclaim <= nr_pages)
4877                         max_reclaim = 0;
4878                 else
4879                         max_reclaim -= nr_pages;
4880
4881                 wait_event(fs_info->async_submit_wait,
4882                            atomic_read(&fs_info->async_delalloc_pages) <=
4883                            (int)max_reclaim);
4884 skip_async:
4885                 spin_lock(&space_info->lock);
4886                 if (list_empty(&space_info->tickets) &&
4887                     list_empty(&space_info->priority_tickets)) {
4888                         spin_unlock(&space_info->lock);
4889                         break;
4890                 }
4891                 spin_unlock(&space_info->lock);
4892
4893                 loops++;
4894                 if (wait_ordered && !trans) {
4895                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4896                 } else {
4897                         time_left = schedule_timeout_killable(1);
4898                         if (time_left)
4899                                 break;
4900                 }
4901                 delalloc_bytes = percpu_counter_sum_positive(
4902                                                 &fs_info->delalloc_bytes);
4903         }
4904 }
4905
4906 struct reserve_ticket {
4907         u64 bytes;
4908         int error;
4909         struct list_head list;
4910         wait_queue_head_t wait;
4911 };
4912
4913 /**
4914  * maybe_commit_transaction - possibly commit the transaction if its ok to
4915  * @root - the root we're allocating for
4916  * @bytes - the number of bytes we want to reserve
4917  * @force - force the commit
4918  *
4919  * This will check to make sure that committing the transaction will actually
4920  * get us somewhere and then commit the transaction if it does.  Otherwise it
4921  * will return -ENOSPC.
4922  */
4923 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4924                                   struct btrfs_space_info *space_info)
4925 {
4926         struct reserve_ticket *ticket = NULL;
4927         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4928         struct btrfs_trans_handle *trans;
4929         u64 bytes;
4930
4931         trans = (struct btrfs_trans_handle *)current->journal_info;
4932         if (trans)
4933                 return -EAGAIN;
4934
4935         spin_lock(&space_info->lock);
4936         if (!list_empty(&space_info->priority_tickets))
4937                 ticket = list_first_entry(&space_info->priority_tickets,
4938                                           struct reserve_ticket, list);
4939         else if (!list_empty(&space_info->tickets))
4940                 ticket = list_first_entry(&space_info->tickets,
4941                                           struct reserve_ticket, list);
4942         bytes = (ticket) ? ticket->bytes : 0;
4943         spin_unlock(&space_info->lock);
4944
4945         if (!bytes)
4946                 return 0;
4947
4948         /* See if there is enough pinned space to make this reservation */
4949         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4950                                    bytes) >= 0)
4951                 goto commit;
4952
4953         /*
4954          * See if there is some space in the delayed insertion reservation for
4955          * this reservation.
4956          */
4957         if (space_info != delayed_rsv->space_info)
4958                 return -ENOSPC;
4959
4960         spin_lock(&delayed_rsv->lock);
4961         if (delayed_rsv->size > bytes)
4962                 bytes = 0;
4963         else
4964                 bytes -= delayed_rsv->size;
4965         spin_unlock(&delayed_rsv->lock);
4966
4967         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4968                                    bytes) < 0) {
4969                 return -ENOSPC;
4970         }
4971
4972 commit:
4973         trans = btrfs_join_transaction(fs_info->extent_root);
4974         if (IS_ERR(trans))
4975                 return -ENOSPC;
4976
4977         return btrfs_commit_transaction(trans);
4978 }
4979
4980 /*
4981  * Try to flush some data based on policy set by @state. This is only advisory
4982  * and may fail for various reasons. The caller is supposed to examine the
4983  * state of @space_info to detect the outcome.
4984  */
4985 static void flush_space(struct btrfs_fs_info *fs_info,
4986                        struct btrfs_space_info *space_info, u64 num_bytes,
4987                        int state)
4988 {
4989         struct btrfs_root *root = fs_info->extent_root;
4990         struct btrfs_trans_handle *trans;
4991         int nr;
4992         int ret = 0;
4993
4994         switch (state) {
4995         case FLUSH_DELAYED_ITEMS_NR:
4996         case FLUSH_DELAYED_ITEMS:
4997                 if (state == FLUSH_DELAYED_ITEMS_NR)
4998                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4999                 else
5000                         nr = -1;
5001
5002                 trans = btrfs_join_transaction(root);
5003                 if (IS_ERR(trans)) {
5004                         ret = PTR_ERR(trans);
5005                         break;
5006                 }
5007                 ret = btrfs_run_delayed_items_nr(trans, nr);
5008                 btrfs_end_transaction(trans);
5009                 break;
5010         case FLUSH_DELALLOC:
5011         case FLUSH_DELALLOC_WAIT:
5012                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
5013                                 state == FLUSH_DELALLOC_WAIT);
5014                 break;
5015         case ALLOC_CHUNK:
5016                 trans = btrfs_join_transaction(root);
5017                 if (IS_ERR(trans)) {
5018                         ret = PTR_ERR(trans);
5019                         break;
5020                 }
5021                 ret = do_chunk_alloc(trans, fs_info,
5022                                      btrfs_metadata_alloc_profile(fs_info),
5023                                      CHUNK_ALLOC_NO_FORCE);
5024                 btrfs_end_transaction(trans);
5025                 if (ret > 0 || ret == -ENOSPC)
5026                         ret = 0;
5027                 break;
5028         case COMMIT_TRANS:
5029                 ret = may_commit_transaction(fs_info, space_info);
5030                 break;
5031         default:
5032                 ret = -ENOSPC;
5033                 break;
5034         }
5035
5036         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5037                                 ret);
5038         return;
5039 }
5040
5041 static inline u64
5042 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5043                                  struct btrfs_space_info *space_info,
5044                                  bool system_chunk)
5045 {
5046         struct reserve_ticket *ticket;
5047         u64 used;
5048         u64 expected;
5049         u64 to_reclaim = 0;
5050
5051         list_for_each_entry(ticket, &space_info->tickets, list)
5052                 to_reclaim += ticket->bytes;
5053         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5054                 to_reclaim += ticket->bytes;
5055         if (to_reclaim)
5056                 return to_reclaim;
5057
5058         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5059         if (can_overcommit(fs_info, space_info, to_reclaim,
5060                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5061                 return 0;
5062
5063         used = btrfs_space_info_used(space_info, true);
5064
5065         if (can_overcommit(fs_info, space_info, SZ_1M,
5066                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5067                 expected = div_factor_fine(space_info->total_bytes, 95);
5068         else
5069                 expected = div_factor_fine(space_info->total_bytes, 90);
5070
5071         if (used > expected)
5072                 to_reclaim = used - expected;
5073         else
5074                 to_reclaim = 0;
5075         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5076                                      space_info->bytes_reserved);
5077         return to_reclaim;
5078 }
5079
5080 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5081                                         struct btrfs_space_info *space_info,
5082                                         u64 used, bool system_chunk)
5083 {
5084         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5085
5086         /* If we're just plain full then async reclaim just slows us down. */
5087         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5088                 return 0;
5089
5090         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5091                                               system_chunk))
5092                 return 0;
5093
5094         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5095                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5096 }
5097
5098 static void wake_all_tickets(struct list_head *head)
5099 {
5100         struct reserve_ticket *ticket;
5101
5102         while (!list_empty(head)) {
5103                 ticket = list_first_entry(head, struct reserve_ticket, list);
5104                 list_del_init(&ticket->list);
5105                 ticket->error = -ENOSPC;
5106                 wake_up(&ticket->wait);
5107         }
5108 }
5109
5110 /*
5111  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5112  * will loop and continuously try to flush as long as we are making progress.
5113  * We count progress as clearing off tickets each time we have to loop.
5114  */
5115 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5116 {
5117         struct btrfs_fs_info *fs_info;
5118         struct btrfs_space_info *space_info;
5119         u64 to_reclaim;
5120         int flush_state;
5121         int commit_cycles = 0;
5122         u64 last_tickets_id;
5123
5124         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5125         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5126
5127         spin_lock(&space_info->lock);
5128         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5129                                                       false);
5130         if (!to_reclaim) {
5131                 space_info->flush = 0;
5132                 spin_unlock(&space_info->lock);
5133                 return;
5134         }
5135         last_tickets_id = space_info->tickets_id;
5136         spin_unlock(&space_info->lock);
5137
5138         flush_state = FLUSH_DELAYED_ITEMS_NR;
5139         do {
5140                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5141                 spin_lock(&space_info->lock);
5142                 if (list_empty(&space_info->tickets)) {
5143                         space_info->flush = 0;
5144                         spin_unlock(&space_info->lock);
5145                         return;
5146                 }
5147                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5148                                                               space_info,
5149                                                               false);
5150                 if (last_tickets_id == space_info->tickets_id) {
5151                         flush_state++;
5152                 } else {
5153                         last_tickets_id = space_info->tickets_id;
5154                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5155                         if (commit_cycles)
5156                                 commit_cycles--;
5157                 }
5158
5159                 if (flush_state > COMMIT_TRANS) {
5160                         commit_cycles++;
5161                         if (commit_cycles > 2) {
5162                                 wake_all_tickets(&space_info->tickets);
5163                                 space_info->flush = 0;
5164                         } else {
5165                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5166                         }
5167                 }
5168                 spin_unlock(&space_info->lock);
5169         } while (flush_state <= COMMIT_TRANS);
5170 }
5171
5172 void btrfs_init_async_reclaim_work(struct work_struct *work)
5173 {
5174         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5175 }
5176
5177 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5178                                             struct btrfs_space_info *space_info,
5179                                             struct reserve_ticket *ticket)
5180 {
5181         u64 to_reclaim;
5182         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5183
5184         spin_lock(&space_info->lock);
5185         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5186                                                       false);
5187         if (!to_reclaim) {
5188                 spin_unlock(&space_info->lock);
5189                 return;
5190         }
5191         spin_unlock(&space_info->lock);
5192
5193         do {
5194                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5195                 flush_state++;
5196                 spin_lock(&space_info->lock);
5197                 if (ticket->bytes == 0) {
5198                         spin_unlock(&space_info->lock);
5199                         return;
5200                 }
5201                 spin_unlock(&space_info->lock);
5202
5203                 /*
5204                  * Priority flushers can't wait on delalloc without
5205                  * deadlocking.
5206                  */
5207                 if (flush_state == FLUSH_DELALLOC ||
5208                     flush_state == FLUSH_DELALLOC_WAIT)
5209                         flush_state = ALLOC_CHUNK;
5210         } while (flush_state < COMMIT_TRANS);
5211 }
5212
5213 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5214                                struct btrfs_space_info *space_info,
5215                                struct reserve_ticket *ticket, u64 orig_bytes)
5216
5217 {
5218         DEFINE_WAIT(wait);
5219         int ret = 0;
5220
5221         spin_lock(&space_info->lock);
5222         while (ticket->bytes > 0 && ticket->error == 0) {
5223                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5224                 if (ret) {
5225                         ret = -EINTR;
5226                         break;
5227                 }
5228                 spin_unlock(&space_info->lock);
5229
5230                 schedule();
5231
5232                 finish_wait(&ticket->wait, &wait);
5233                 spin_lock(&space_info->lock);
5234         }
5235         if (!ret)
5236                 ret = ticket->error;
5237         if (!list_empty(&ticket->list))
5238                 list_del_init(&ticket->list);
5239         if (ticket->bytes && ticket->bytes < orig_bytes) {
5240                 u64 num_bytes = orig_bytes - ticket->bytes;
5241                 space_info->bytes_may_use -= num_bytes;
5242                 trace_btrfs_space_reservation(fs_info, "space_info",
5243                                               space_info->flags, num_bytes, 0);
5244         }
5245         spin_unlock(&space_info->lock);
5246
5247         return ret;
5248 }
5249
5250 /**
5251  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5252  * @root - the root we're allocating for
5253  * @space_info - the space info we want to allocate from
5254  * @orig_bytes - the number of bytes we want
5255  * @flush - whether or not we can flush to make our reservation
5256  *
5257  * This will reserve orig_bytes number of bytes from the space info associated
5258  * with the block_rsv.  If there is not enough space it will make an attempt to
5259  * flush out space to make room.  It will do this by flushing delalloc if
5260  * possible or committing the transaction.  If flush is 0 then no attempts to
5261  * regain reservations will be made and this will fail if there is not enough
5262  * space already.
5263  */
5264 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5265                                     struct btrfs_space_info *space_info,
5266                                     u64 orig_bytes,
5267                                     enum btrfs_reserve_flush_enum flush,
5268                                     bool system_chunk)
5269 {
5270         struct reserve_ticket ticket;
5271         u64 used;
5272         int ret = 0;
5273
5274         ASSERT(orig_bytes);
5275         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5276
5277         spin_lock(&space_info->lock);
5278         ret = -ENOSPC;
5279         used = btrfs_space_info_used(space_info, true);
5280
5281         /*
5282          * If we have enough space then hooray, make our reservation and carry
5283          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5284          * If not things get more complicated.
5285          */
5286         if (used + orig_bytes <= space_info->total_bytes) {
5287                 space_info->bytes_may_use += orig_bytes;
5288                 trace_btrfs_space_reservation(fs_info, "space_info",
5289                                               space_info->flags, orig_bytes, 1);
5290                 ret = 0;
5291         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5292                                   system_chunk)) {
5293                 space_info->bytes_may_use += orig_bytes;
5294                 trace_btrfs_space_reservation(fs_info, "space_info",
5295                                               space_info->flags, orig_bytes, 1);
5296                 ret = 0;
5297         }
5298
5299         /*
5300          * If we couldn't make a reservation then setup our reservation ticket
5301          * and kick the async worker if it's not already running.
5302          *
5303          * If we are a priority flusher then we just need to add our ticket to
5304          * the list and we will do our own flushing further down.
5305          */
5306         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5307                 ticket.bytes = orig_bytes;
5308                 ticket.error = 0;
5309                 init_waitqueue_head(&ticket.wait);
5310                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5311                         list_add_tail(&ticket.list, &space_info->tickets);
5312                         if (!space_info->flush) {
5313                                 space_info->flush = 1;
5314                                 trace_btrfs_trigger_flush(fs_info,
5315                                                           space_info->flags,
5316                                                           orig_bytes, flush,
5317                                                           "enospc");
5318                                 queue_work(system_unbound_wq,
5319                                            &fs_info->async_reclaim_work);
5320                         }
5321                 } else {
5322                         list_add_tail(&ticket.list,
5323                                       &space_info->priority_tickets);
5324                 }
5325         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5326                 used += orig_bytes;
5327                 /*
5328                  * We will do the space reservation dance during log replay,
5329                  * which means we won't have fs_info->fs_root set, so don't do
5330                  * the async reclaim as we will panic.
5331                  */
5332                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5333                     need_do_async_reclaim(fs_info, space_info,
5334                                           used, system_chunk) &&
5335                     !work_busy(&fs_info->async_reclaim_work)) {
5336                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5337                                                   orig_bytes, flush, "preempt");
5338                         queue_work(system_unbound_wq,
5339                                    &fs_info->async_reclaim_work);
5340                 }
5341         }
5342         spin_unlock(&space_info->lock);
5343         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5344                 return ret;
5345
5346         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5347                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5348                                            orig_bytes);
5349
5350         ret = 0;
5351         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5352         spin_lock(&space_info->lock);
5353         if (ticket.bytes) {
5354                 if (ticket.bytes < orig_bytes) {
5355                         u64 num_bytes = orig_bytes - ticket.bytes;
5356                         space_info->bytes_may_use -= num_bytes;
5357                         trace_btrfs_space_reservation(fs_info, "space_info",
5358                                                       space_info->flags,
5359                                                       num_bytes, 0);
5360
5361                 }
5362                 list_del_init(&ticket.list);
5363                 ret = -ENOSPC;
5364         }
5365         spin_unlock(&space_info->lock);
5366         ASSERT(list_empty(&ticket.list));
5367         return ret;
5368 }
5369
5370 /**
5371  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5372  * @root - the root we're allocating for
5373  * @block_rsv - the block_rsv we're allocating for
5374  * @orig_bytes - the number of bytes we want
5375  * @flush - whether or not we can flush to make our reservation
5376  *
5377  * This will reserve orgi_bytes number of bytes from the space info associated
5378  * with the block_rsv.  If there is not enough space it will make an attempt to
5379  * flush out space to make room.  It will do this by flushing delalloc if
5380  * possible or committing the transaction.  If flush is 0 then no attempts to
5381  * regain reservations will be made and this will fail if there is not enough
5382  * space already.
5383  */
5384 static int reserve_metadata_bytes(struct btrfs_root *root,
5385                                   struct btrfs_block_rsv *block_rsv,
5386                                   u64 orig_bytes,
5387                                   enum btrfs_reserve_flush_enum flush)
5388 {
5389         struct btrfs_fs_info *fs_info = root->fs_info;
5390         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5391         int ret;
5392         bool system_chunk = (root == fs_info->chunk_root);
5393
5394         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5395                                        orig_bytes, flush, system_chunk);
5396         if (ret == -ENOSPC &&
5397             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5398                 if (block_rsv != global_rsv &&
5399                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5400                         ret = 0;
5401         }
5402         if (ret == -ENOSPC) {
5403                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5404                                               block_rsv->space_info->flags,
5405                                               orig_bytes, 1);
5406
5407                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5408                         dump_space_info(fs_info, block_rsv->space_info,
5409                                         orig_bytes, 0);
5410         }
5411         return ret;
5412 }
5413
5414 static struct btrfs_block_rsv *get_block_rsv(
5415                                         const struct btrfs_trans_handle *trans,
5416                                         const struct btrfs_root *root)
5417 {
5418         struct btrfs_fs_info *fs_info = root->fs_info;
5419         struct btrfs_block_rsv *block_rsv = NULL;
5420
5421         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5422             (root == fs_info->csum_root && trans->adding_csums) ||
5423             (root == fs_info->uuid_root))
5424                 block_rsv = trans->block_rsv;
5425
5426         if (!block_rsv)
5427                 block_rsv = root->block_rsv;
5428
5429         if (!block_rsv)
5430                 block_rsv = &fs_info->empty_block_rsv;
5431
5432         return block_rsv;
5433 }
5434
5435 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5436                                u64 num_bytes)
5437 {
5438         int ret = -ENOSPC;
5439         spin_lock(&block_rsv->lock);
5440         if (block_rsv->reserved >= num_bytes) {
5441                 block_rsv->reserved -= num_bytes;
5442                 if (block_rsv->reserved < block_rsv->size)
5443                         block_rsv->full = 0;
5444                 ret = 0;
5445         }
5446         spin_unlock(&block_rsv->lock);
5447         return ret;
5448 }
5449
5450 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5451                                 u64 num_bytes, int update_size)
5452 {
5453         spin_lock(&block_rsv->lock);
5454         block_rsv->reserved += num_bytes;
5455         if (update_size)
5456                 block_rsv->size += num_bytes;
5457         else if (block_rsv->reserved >= block_rsv->size)
5458                 block_rsv->full = 1;
5459         spin_unlock(&block_rsv->lock);
5460 }
5461
5462 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5463                              struct btrfs_block_rsv *dest, u64 num_bytes,
5464                              int min_factor)
5465 {
5466         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5467         u64 min_bytes;
5468
5469         if (global_rsv->space_info != dest->space_info)
5470                 return -ENOSPC;
5471
5472         spin_lock(&global_rsv->lock);
5473         min_bytes = div_factor(global_rsv->size, min_factor);
5474         if (global_rsv->reserved < min_bytes + num_bytes) {
5475                 spin_unlock(&global_rsv->lock);
5476                 return -ENOSPC;
5477         }
5478         global_rsv->reserved -= num_bytes;
5479         if (global_rsv->reserved < global_rsv->size)
5480                 global_rsv->full = 0;
5481         spin_unlock(&global_rsv->lock);
5482
5483         block_rsv_add_bytes(dest, num_bytes, 1);
5484         return 0;
5485 }
5486
5487 /*
5488  * This is for space we already have accounted in space_info->bytes_may_use, so
5489  * basically when we're returning space from block_rsv's.
5490  */
5491 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5492                                      struct btrfs_space_info *space_info,
5493                                      u64 num_bytes)
5494 {
5495         struct reserve_ticket *ticket;
5496         struct list_head *head;
5497         u64 used;
5498         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5499         bool check_overcommit = false;
5500
5501         spin_lock(&space_info->lock);
5502         head = &space_info->priority_tickets;
5503
5504         /*
5505          * If we are over our limit then we need to check and see if we can
5506          * overcommit, and if we can't then we just need to free up our space
5507          * and not satisfy any requests.
5508          */
5509         used = btrfs_space_info_used(space_info, true);
5510         if (used - num_bytes >= space_info->total_bytes)
5511                 check_overcommit = true;
5512 again:
5513         while (!list_empty(head) && num_bytes) {
5514                 ticket = list_first_entry(head, struct reserve_ticket,
5515                                           list);
5516                 /*
5517                  * We use 0 bytes because this space is already reserved, so
5518                  * adding the ticket space would be a double count.
5519                  */
5520                 if (check_overcommit &&
5521                     !can_overcommit(fs_info, space_info, 0, flush, false))
5522                         break;
5523                 if (num_bytes >= ticket->bytes) {
5524                         list_del_init(&ticket->list);
5525                         num_bytes -= ticket->bytes;
5526                         ticket->bytes = 0;
5527                         space_info->tickets_id++;
5528                         wake_up(&ticket->wait);
5529                 } else {
5530                         ticket->bytes -= num_bytes;
5531                         num_bytes = 0;
5532                 }
5533         }
5534
5535         if (num_bytes && head == &space_info->priority_tickets) {
5536                 head = &space_info->tickets;
5537                 flush = BTRFS_RESERVE_FLUSH_ALL;
5538                 goto again;
5539         }
5540         space_info->bytes_may_use -= num_bytes;
5541         trace_btrfs_space_reservation(fs_info, "space_info",
5542                                       space_info->flags, num_bytes, 0);
5543         spin_unlock(&space_info->lock);
5544 }
5545
5546 /*
5547  * This is for newly allocated space that isn't accounted in
5548  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5549  * we use this helper.
5550  */
5551 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5552                                      struct btrfs_space_info *space_info,
5553                                      u64 num_bytes)
5554 {
5555         struct reserve_ticket *ticket;
5556         struct list_head *head = &space_info->priority_tickets;
5557
5558 again:
5559         while (!list_empty(head) && num_bytes) {
5560                 ticket = list_first_entry(head, struct reserve_ticket,
5561                                           list);
5562                 if (num_bytes >= ticket->bytes) {
5563                         trace_btrfs_space_reservation(fs_info, "space_info",
5564                                                       space_info->flags,
5565                                                       ticket->bytes, 1);
5566                         list_del_init(&ticket->list);
5567                         num_bytes -= ticket->bytes;
5568                         space_info->bytes_may_use += ticket->bytes;
5569                         ticket->bytes = 0;
5570                         space_info->tickets_id++;
5571                         wake_up(&ticket->wait);
5572                 } else {
5573                         trace_btrfs_space_reservation(fs_info, "space_info",
5574                                                       space_info->flags,
5575                                                       num_bytes, 1);
5576                         space_info->bytes_may_use += num_bytes;
5577                         ticket->bytes -= num_bytes;
5578                         num_bytes = 0;
5579                 }
5580         }
5581
5582         if (num_bytes && head == &space_info->priority_tickets) {
5583                 head = &space_info->tickets;
5584                 goto again;
5585         }
5586 }
5587
5588 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5589                                     struct btrfs_block_rsv *block_rsv,
5590                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5591                                     u64 *qgroup_to_release_ret)
5592 {
5593         struct btrfs_space_info *space_info = block_rsv->space_info;
5594         u64 qgroup_to_release = 0;
5595         u64 ret;
5596
5597         spin_lock(&block_rsv->lock);
5598         if (num_bytes == (u64)-1) {
5599                 num_bytes = block_rsv->size;
5600                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5601         }
5602         block_rsv->size -= num_bytes;
5603         if (block_rsv->reserved >= block_rsv->size) {
5604                 num_bytes = block_rsv->reserved - block_rsv->size;
5605                 block_rsv->reserved = block_rsv->size;
5606                 block_rsv->full = 1;
5607         } else {
5608                 num_bytes = 0;
5609         }
5610         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5611                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5612                                     block_rsv->qgroup_rsv_size;
5613                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5614         } else {
5615                 qgroup_to_release = 0;
5616         }
5617         spin_unlock(&block_rsv->lock);
5618
5619         ret = num_bytes;
5620         if (num_bytes > 0) {
5621                 if (dest) {
5622                         spin_lock(&dest->lock);
5623                         if (!dest->full) {
5624                                 u64 bytes_to_add;
5625
5626                                 bytes_to_add = dest->size - dest->reserved;
5627                                 bytes_to_add = min(num_bytes, bytes_to_add);
5628                                 dest->reserved += bytes_to_add;
5629                                 if (dest->reserved >= dest->size)
5630                                         dest->full = 1;
5631                                 num_bytes -= bytes_to_add;
5632                         }
5633                         spin_unlock(&dest->lock);
5634                 }
5635                 if (num_bytes)
5636                         space_info_add_old_bytes(fs_info, space_info,
5637                                                  num_bytes);
5638         }
5639         if (qgroup_to_release_ret)
5640                 *qgroup_to_release_ret = qgroup_to_release;
5641         return ret;
5642 }
5643
5644 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5645                             struct btrfs_block_rsv *dst, u64 num_bytes,
5646                             int update_size)
5647 {
5648         int ret;
5649
5650         ret = block_rsv_use_bytes(src, num_bytes);
5651         if (ret)
5652                 return ret;
5653
5654         block_rsv_add_bytes(dst, num_bytes, update_size);
5655         return 0;
5656 }
5657
5658 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5659 {
5660         memset(rsv, 0, sizeof(*rsv));
5661         spin_lock_init(&rsv->lock);
5662         rsv->type = type;
5663 }
5664
5665 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5666                                    struct btrfs_block_rsv *rsv,
5667                                    unsigned short type)
5668 {
5669         btrfs_init_block_rsv(rsv, type);
5670         rsv->space_info = __find_space_info(fs_info,
5671                                             BTRFS_BLOCK_GROUP_METADATA);
5672 }
5673
5674 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5675                                               unsigned short type)
5676 {
5677         struct btrfs_block_rsv *block_rsv;
5678
5679         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5680         if (!block_rsv)
5681                 return NULL;
5682
5683         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5684         return block_rsv;
5685 }
5686
5687 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5688                           struct btrfs_block_rsv *rsv)
5689 {
5690         if (!rsv)
5691                 return;
5692         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5693         kfree(rsv);
5694 }
5695
5696 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5697 {
5698         kfree(rsv);
5699 }
5700
5701 int btrfs_block_rsv_add(struct btrfs_root *root,
5702                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5703                         enum btrfs_reserve_flush_enum flush)
5704 {
5705         int ret;
5706
5707         if (num_bytes == 0)
5708                 return 0;
5709
5710         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5711         if (!ret) {
5712                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5713                 return 0;
5714         }
5715
5716         return ret;
5717 }
5718
5719 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5720 {
5721         u64 num_bytes = 0;
5722         int ret = -ENOSPC;
5723
5724         if (!block_rsv)
5725                 return 0;
5726
5727         spin_lock(&block_rsv->lock);
5728         num_bytes = div_factor(block_rsv->size, min_factor);
5729         if (block_rsv->reserved >= num_bytes)
5730                 ret = 0;
5731         spin_unlock(&block_rsv->lock);
5732
5733         return ret;
5734 }
5735
5736 int btrfs_block_rsv_refill(struct btrfs_root *root,
5737                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5738                            enum btrfs_reserve_flush_enum flush)
5739 {
5740         u64 num_bytes = 0;
5741         int ret = -ENOSPC;
5742
5743         if (!block_rsv)
5744                 return 0;
5745
5746         spin_lock(&block_rsv->lock);
5747         num_bytes = min_reserved;
5748         if (block_rsv->reserved >= num_bytes)
5749                 ret = 0;
5750         else
5751                 num_bytes -= block_rsv->reserved;
5752         spin_unlock(&block_rsv->lock);
5753
5754         if (!ret)
5755                 return 0;
5756
5757         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5758         if (!ret) {
5759                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5760                 return 0;
5761         }
5762
5763         return ret;
5764 }
5765
5766 /**
5767  * btrfs_inode_rsv_refill - refill the inode block rsv.
5768  * @inode - the inode we are refilling.
5769  * @flush - the flusing restriction.
5770  *
5771  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5772  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5773  * or return if we already have enough space.  This will also handle the resreve
5774  * tracepoint for the reserved amount.
5775  */
5776 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5777                                   enum btrfs_reserve_flush_enum flush)
5778 {
5779         struct btrfs_root *root = inode->root;
5780         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5781         u64 num_bytes = 0;
5782         u64 qgroup_num_bytes = 0;
5783         int ret = -ENOSPC;
5784
5785         spin_lock(&block_rsv->lock);
5786         if (block_rsv->reserved < block_rsv->size)
5787                 num_bytes = block_rsv->size - block_rsv->reserved;
5788         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5789                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5790                                    block_rsv->qgroup_rsv_reserved;
5791         spin_unlock(&block_rsv->lock);
5792
5793         if (num_bytes == 0)
5794                 return 0;
5795
5796         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5797         if (ret)
5798                 return ret;
5799         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5800         if (!ret) {
5801                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5802                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5803                                               btrfs_ino(inode), num_bytes, 1);
5804
5805                 /* Don't forget to increase qgroup_rsv_reserved */
5806                 spin_lock(&block_rsv->lock);
5807                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5808                 spin_unlock(&block_rsv->lock);
5809         } else
5810                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5811         return ret;
5812 }
5813
5814 /**
5815  * btrfs_inode_rsv_release - release any excessive reservation.
5816  * @inode - the inode we need to release from.
5817  * @qgroup_free - free or convert qgroup meta.
5818  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5819  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5820  *   @qgroup_free is true for error handling, and false for normal release.
5821  *
5822  * This is the same as btrfs_block_rsv_release, except that it handles the
5823  * tracepoint for the reservation.
5824  */
5825 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5826 {
5827         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5828         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5829         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5830         u64 released = 0;
5831         u64 qgroup_to_release = 0;
5832
5833         /*
5834          * Since we statically set the block_rsv->size we just want to say we
5835          * are releasing 0 bytes, and then we'll just get the reservation over
5836          * the size free'd.
5837          */
5838         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5839                                            &qgroup_to_release);
5840         if (released > 0)
5841                 trace_btrfs_space_reservation(fs_info, "delalloc",
5842                                               btrfs_ino(inode), released, 0);
5843         if (qgroup_free)
5844                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5845         else
5846                 btrfs_qgroup_convert_reserved_meta(inode->root,
5847                                                    qgroup_to_release);
5848 }
5849
5850 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5851                              struct btrfs_block_rsv *block_rsv,
5852                              u64 num_bytes)
5853 {
5854         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5855
5856         if (global_rsv == block_rsv ||
5857             block_rsv->space_info != global_rsv->space_info)
5858                 global_rsv = NULL;
5859         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5860 }
5861
5862 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5863 {
5864         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5865         struct btrfs_space_info *sinfo = block_rsv->space_info;
5866         u64 num_bytes;
5867
5868         /*
5869          * The global block rsv is based on the size of the extent tree, the
5870          * checksum tree and the root tree.  If the fs is empty we want to set
5871          * it to a minimal amount for safety.
5872          */
5873         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5874                 btrfs_root_used(&fs_info->csum_root->root_item) +
5875                 btrfs_root_used(&fs_info->tree_root->root_item);
5876         num_bytes = max_t(u64, num_bytes, SZ_16M);
5877
5878         spin_lock(&sinfo->lock);
5879         spin_lock(&block_rsv->lock);
5880
5881         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5882
5883         if (block_rsv->reserved < block_rsv->size) {
5884                 num_bytes = btrfs_space_info_used(sinfo, true);
5885                 if (sinfo->total_bytes > num_bytes) {
5886                         num_bytes = sinfo->total_bytes - num_bytes;
5887                         num_bytes = min(num_bytes,
5888                                         block_rsv->size - block_rsv->reserved);
5889                         block_rsv->reserved += num_bytes;
5890                         sinfo->bytes_may_use += num_bytes;
5891                         trace_btrfs_space_reservation(fs_info, "space_info",
5892                                                       sinfo->flags, num_bytes,
5893                                                       1);
5894                 }
5895         } else if (block_rsv->reserved > block_rsv->size) {
5896                 num_bytes = block_rsv->reserved - block_rsv->size;
5897                 sinfo->bytes_may_use -= num_bytes;
5898                 trace_btrfs_space_reservation(fs_info, "space_info",
5899                                       sinfo->flags, num_bytes, 0);
5900                 block_rsv->reserved = block_rsv->size;
5901         }
5902
5903         if (block_rsv->reserved == block_rsv->size)
5904                 block_rsv->full = 1;
5905         else
5906                 block_rsv->full = 0;
5907
5908         spin_unlock(&block_rsv->lock);
5909         spin_unlock(&sinfo->lock);
5910 }
5911
5912 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5913 {
5914         struct btrfs_space_info *space_info;
5915
5916         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5917         fs_info->chunk_block_rsv.space_info = space_info;
5918
5919         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5920         fs_info->global_block_rsv.space_info = space_info;
5921         fs_info->trans_block_rsv.space_info = space_info;
5922         fs_info->empty_block_rsv.space_info = space_info;
5923         fs_info->delayed_block_rsv.space_info = space_info;
5924
5925         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5926         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5927         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5928         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5929         if (fs_info->quota_root)
5930                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5931         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5932
5933         update_global_block_rsv(fs_info);
5934 }
5935
5936 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5937 {
5938         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5939                                 (u64)-1, NULL);
5940         WARN_ON(fs_info->trans_block_rsv.size > 0);
5941         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5942         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5943         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5944         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5945         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5946 }
5947
5948
5949 /*
5950  * To be called after all the new block groups attached to the transaction
5951  * handle have been created (btrfs_create_pending_block_groups()).
5952  */
5953 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5954 {
5955         struct btrfs_fs_info *fs_info = trans->fs_info;
5956
5957         if (!trans->chunk_bytes_reserved)
5958                 return;
5959
5960         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5961
5962         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5963                                 trans->chunk_bytes_reserved, NULL);
5964         trans->chunk_bytes_reserved = 0;
5965 }
5966
5967 /*
5968  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5969  * root: the root of the parent directory
5970  * rsv: block reservation
5971  * items: the number of items that we need do reservation
5972  * qgroup_reserved: used to return the reserved size in qgroup
5973  *
5974  * This function is used to reserve the space for snapshot/subvolume
5975  * creation and deletion. Those operations are different with the
5976  * common file/directory operations, they change two fs/file trees
5977  * and root tree, the number of items that the qgroup reserves is
5978  * different with the free space reservation. So we can not use
5979  * the space reservation mechanism in start_transaction().
5980  */
5981 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5982                                      struct btrfs_block_rsv *rsv,
5983                                      int items,
5984                                      bool use_global_rsv)
5985 {
5986         u64 num_bytes;
5987         int ret;
5988         struct btrfs_fs_info *fs_info = root->fs_info;
5989         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5990
5991         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5992                 /* One for parent inode, two for dir entries */
5993                 num_bytes = 3 * fs_info->nodesize;
5994                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
5995                 if (ret)
5996                         return ret;
5997         } else {
5998                 num_bytes = 0;
5999         }
6000
6001         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6002         rsv->space_info = __find_space_info(fs_info,
6003                                             BTRFS_BLOCK_GROUP_METADATA);
6004         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6005                                   BTRFS_RESERVE_FLUSH_ALL);
6006
6007         if (ret == -ENOSPC && use_global_rsv)
6008                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6009
6010         if (ret && num_bytes)
6011                 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
6012
6013         return ret;
6014 }
6015
6016 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6017                                       struct btrfs_block_rsv *rsv)
6018 {
6019         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6020 }
6021
6022 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6023                                                  struct btrfs_inode *inode)
6024 {
6025         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6026         u64 reserve_size = 0;
6027         u64 qgroup_rsv_size = 0;
6028         u64 csum_leaves;
6029         unsigned outstanding_extents;
6030
6031         lockdep_assert_held(&inode->lock);
6032         outstanding_extents = inode->outstanding_extents;
6033         if (outstanding_extents)
6034                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6035                                                 outstanding_extents + 1);
6036         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6037                                                  inode->csum_bytes);
6038         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6039                                                        csum_leaves);
6040         /*
6041          * For qgroup rsv, the calculation is very simple:
6042          * account one nodesize for each outstanding extent
6043          *
6044          * This is overestimating in most cases.
6045          */
6046         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6047
6048         spin_lock(&block_rsv->lock);
6049         block_rsv->size = reserve_size;
6050         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6051         spin_unlock(&block_rsv->lock);
6052 }
6053
6054 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6055 {
6056         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6057         unsigned nr_extents;
6058         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6059         int ret = 0;
6060         bool delalloc_lock = true;
6061
6062         /* If we are a free space inode we need to not flush since we will be in
6063          * the middle of a transaction commit.  We also don't need the delalloc
6064          * mutex since we won't race with anybody.  We need this mostly to make
6065          * lockdep shut its filthy mouth.
6066          *
6067          * If we have a transaction open (can happen if we call truncate_block
6068          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6069          */
6070         if (btrfs_is_free_space_inode(inode)) {
6071                 flush = BTRFS_RESERVE_NO_FLUSH;
6072                 delalloc_lock = false;
6073         } else {
6074                 if (current->journal_info)
6075                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6076
6077                 if (btrfs_transaction_in_commit(fs_info))
6078                         schedule_timeout(1);
6079         }
6080
6081         if (delalloc_lock)
6082                 mutex_lock(&inode->delalloc_mutex);
6083
6084         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6085
6086         /* Add our new extents and calculate the new rsv size. */
6087         spin_lock(&inode->lock);
6088         nr_extents = count_max_extents(num_bytes);
6089         btrfs_mod_outstanding_extents(inode, nr_extents);
6090         inode->csum_bytes += num_bytes;
6091         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6092         spin_unlock(&inode->lock);
6093
6094         ret = btrfs_inode_rsv_refill(inode, flush);
6095         if (unlikely(ret))
6096                 goto out_fail;
6097
6098         if (delalloc_lock)
6099                 mutex_unlock(&inode->delalloc_mutex);
6100         return 0;
6101
6102 out_fail:
6103         spin_lock(&inode->lock);
6104         nr_extents = count_max_extents(num_bytes);
6105         btrfs_mod_outstanding_extents(inode, -nr_extents);
6106         inode->csum_bytes -= num_bytes;
6107         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6108         spin_unlock(&inode->lock);
6109
6110         btrfs_inode_rsv_release(inode, true);
6111         if (delalloc_lock)
6112                 mutex_unlock(&inode->delalloc_mutex);
6113         return ret;
6114 }
6115
6116 /**
6117  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6118  * @inode: the inode to release the reservation for.
6119  * @num_bytes: the number of bytes we are releasing.
6120  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6121  *
6122  * This will release the metadata reservation for an inode.  This can be called
6123  * once we complete IO for a given set of bytes to release their metadata
6124  * reservations, or on error for the same reason.
6125  */
6126 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6127                                      bool qgroup_free)
6128 {
6129         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6130
6131         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6132         spin_lock(&inode->lock);
6133         inode->csum_bytes -= num_bytes;
6134         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6135         spin_unlock(&inode->lock);
6136
6137         if (btrfs_is_testing(fs_info))
6138                 return;
6139
6140         btrfs_inode_rsv_release(inode, qgroup_free);
6141 }
6142
6143 /**
6144  * btrfs_delalloc_release_extents - release our outstanding_extents
6145  * @inode: the inode to balance the reservation for.
6146  * @num_bytes: the number of bytes we originally reserved with
6147  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6148  *
6149  * When we reserve space we increase outstanding_extents for the extents we may
6150  * add.  Once we've set the range as delalloc or created our ordered extents we
6151  * have outstanding_extents to track the real usage, so we use this to free our
6152  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6153  * with btrfs_delalloc_reserve_metadata.
6154  */
6155 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6156                                     bool qgroup_free)
6157 {
6158         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6159         unsigned num_extents;
6160
6161         spin_lock(&inode->lock);
6162         num_extents = count_max_extents(num_bytes);
6163         btrfs_mod_outstanding_extents(inode, -num_extents);
6164         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6165         spin_unlock(&inode->lock);
6166
6167         if (btrfs_is_testing(fs_info))
6168                 return;
6169
6170         btrfs_inode_rsv_release(inode, qgroup_free);
6171 }
6172
6173 /**
6174  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6175  * delalloc
6176  * @inode: inode we're writing to
6177  * @start: start range we are writing to
6178  * @len: how long the range we are writing to
6179  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6180  *            current reservation.
6181  *
6182  * This will do the following things
6183  *
6184  * o reserve space in data space info for num bytes
6185  *   and reserve precious corresponding qgroup space
6186  *   (Done in check_data_free_space)
6187  *
6188  * o reserve space for metadata space, based on the number of outstanding
6189  *   extents and how much csums will be needed
6190  *   also reserve metadata space in a per root over-reserve method.
6191  * o add to the inodes->delalloc_bytes
6192  * o add it to the fs_info's delalloc inodes list.
6193  *   (Above 3 all done in delalloc_reserve_metadata)
6194  *
6195  * Return 0 for success
6196  * Return <0 for error(-ENOSPC or -EQUOT)
6197  */
6198 int btrfs_delalloc_reserve_space(struct inode *inode,
6199                         struct extent_changeset **reserved, u64 start, u64 len)
6200 {
6201         int ret;
6202
6203         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6204         if (ret < 0)
6205                 return ret;
6206         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6207         if (ret < 0)
6208                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6209         return ret;
6210 }
6211
6212 /**
6213  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6214  * @inode: inode we're releasing space for
6215  * @start: start position of the space already reserved
6216  * @len: the len of the space already reserved
6217  * @release_bytes: the len of the space we consumed or didn't use
6218  *
6219  * This function will release the metadata space that was not used and will
6220  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6221  * list if there are no delalloc bytes left.
6222  * Also it will handle the qgroup reserved space.
6223  */
6224 void btrfs_delalloc_release_space(struct inode *inode,
6225                                   struct extent_changeset *reserved,
6226                                   u64 start, u64 len, bool qgroup_free)
6227 {
6228         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6229         btrfs_free_reserved_data_space(inode, reserved, start, len);
6230 }
6231
6232 static int update_block_group(struct btrfs_trans_handle *trans,
6233                               struct btrfs_fs_info *info, u64 bytenr,
6234                               u64 num_bytes, int alloc)
6235 {
6236         struct btrfs_block_group_cache *cache = NULL;
6237         u64 total = num_bytes;
6238         u64 old_val;
6239         u64 byte_in_group;
6240         int factor;
6241
6242         /* block accounting for super block */
6243         spin_lock(&info->delalloc_root_lock);
6244         old_val = btrfs_super_bytes_used(info->super_copy);
6245         if (alloc)
6246                 old_val += num_bytes;
6247         else
6248                 old_val -= num_bytes;
6249         btrfs_set_super_bytes_used(info->super_copy, old_val);
6250         spin_unlock(&info->delalloc_root_lock);
6251
6252         while (total) {
6253                 cache = btrfs_lookup_block_group(info, bytenr);
6254                 if (!cache)
6255                         return -ENOENT;
6256                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6257                                     BTRFS_BLOCK_GROUP_RAID1 |
6258                                     BTRFS_BLOCK_GROUP_RAID10))
6259                         factor = 2;
6260                 else
6261                         factor = 1;
6262                 /*
6263                  * If this block group has free space cache written out, we
6264                  * need to make sure to load it if we are removing space.  This
6265                  * is because we need the unpinning stage to actually add the
6266                  * space back to the block group, otherwise we will leak space.
6267                  */
6268                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6269                         cache_block_group(cache, 1);
6270
6271                 byte_in_group = bytenr - cache->key.objectid;
6272                 WARN_ON(byte_in_group > cache->key.offset);
6273
6274                 spin_lock(&cache->space_info->lock);
6275                 spin_lock(&cache->lock);
6276
6277                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6278                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6279                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6280
6281                 old_val = btrfs_block_group_used(&cache->item);
6282                 num_bytes = min(total, cache->key.offset - byte_in_group);
6283                 if (alloc) {
6284                         old_val += num_bytes;
6285                         btrfs_set_block_group_used(&cache->item, old_val);
6286                         cache->reserved -= num_bytes;
6287                         cache->space_info->bytes_reserved -= num_bytes;
6288                         cache->space_info->bytes_used += num_bytes;
6289                         cache->space_info->disk_used += num_bytes * factor;
6290                         spin_unlock(&cache->lock);
6291                         spin_unlock(&cache->space_info->lock);
6292                 } else {
6293                         old_val -= num_bytes;
6294                         btrfs_set_block_group_used(&cache->item, old_val);
6295                         cache->pinned += num_bytes;
6296                         cache->space_info->bytes_pinned += num_bytes;
6297                         cache->space_info->bytes_used -= num_bytes;
6298                         cache->space_info->disk_used -= num_bytes * factor;
6299                         spin_unlock(&cache->lock);
6300                         spin_unlock(&cache->space_info->lock);
6301
6302                         trace_btrfs_space_reservation(info, "pinned",
6303                                                       cache->space_info->flags,
6304                                                       num_bytes, 1);
6305                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6306                                            num_bytes);
6307                         set_extent_dirty(info->pinned_extents,
6308                                          bytenr, bytenr + num_bytes - 1,
6309                                          GFP_NOFS | __GFP_NOFAIL);
6310                 }
6311
6312                 spin_lock(&trans->transaction->dirty_bgs_lock);
6313                 if (list_empty(&cache->dirty_list)) {
6314                         list_add_tail(&cache->dirty_list,
6315                                       &trans->transaction->dirty_bgs);
6316                         trans->transaction->num_dirty_bgs++;
6317                         btrfs_get_block_group(cache);
6318                 }
6319                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6320
6321                 /*
6322                  * No longer have used bytes in this block group, queue it for
6323                  * deletion. We do this after adding the block group to the
6324                  * dirty list to avoid races between cleaner kthread and space
6325                  * cache writeout.
6326                  */
6327                 if (!alloc && old_val == 0) {
6328                         spin_lock(&info->unused_bgs_lock);
6329                         if (list_empty(&cache->bg_list)) {
6330                                 btrfs_get_block_group(cache);
6331                                 trace_btrfs_add_unused_block_group(cache);
6332                                 list_add_tail(&cache->bg_list,
6333                                               &info->unused_bgs);
6334                         }
6335                         spin_unlock(&info->unused_bgs_lock);
6336                 }
6337
6338                 btrfs_put_block_group(cache);
6339                 total -= num_bytes;
6340                 bytenr += num_bytes;
6341         }
6342         return 0;
6343 }
6344
6345 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6346 {
6347         struct btrfs_block_group_cache *cache;
6348         u64 bytenr;
6349
6350         spin_lock(&fs_info->block_group_cache_lock);
6351         bytenr = fs_info->first_logical_byte;
6352         spin_unlock(&fs_info->block_group_cache_lock);
6353
6354         if (bytenr < (u64)-1)
6355                 return bytenr;
6356
6357         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6358         if (!cache)
6359                 return 0;
6360
6361         bytenr = cache->key.objectid;
6362         btrfs_put_block_group(cache);
6363
6364         return bytenr;
6365 }
6366
6367 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6368                            struct btrfs_block_group_cache *cache,
6369                            u64 bytenr, u64 num_bytes, int reserved)
6370 {
6371         spin_lock(&cache->space_info->lock);
6372         spin_lock(&cache->lock);
6373         cache->pinned += num_bytes;
6374         cache->space_info->bytes_pinned += num_bytes;
6375         if (reserved) {
6376                 cache->reserved -= num_bytes;
6377                 cache->space_info->bytes_reserved -= num_bytes;
6378         }
6379         spin_unlock(&cache->lock);
6380         spin_unlock(&cache->space_info->lock);
6381
6382         trace_btrfs_space_reservation(fs_info, "pinned",
6383                                       cache->space_info->flags, num_bytes, 1);
6384         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6385         set_extent_dirty(fs_info->pinned_extents, bytenr,
6386                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6387         return 0;
6388 }
6389
6390 /*
6391  * this function must be called within transaction
6392  */
6393 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6394                      u64 bytenr, u64 num_bytes, int reserved)
6395 {
6396         struct btrfs_block_group_cache *cache;
6397
6398         cache = btrfs_lookup_block_group(fs_info, bytenr);
6399         BUG_ON(!cache); /* Logic error */
6400
6401         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6402
6403         btrfs_put_block_group(cache);
6404         return 0;
6405 }
6406
6407 /*
6408  * this function must be called within transaction
6409  */
6410 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6411                                     u64 bytenr, u64 num_bytes)
6412 {
6413         struct btrfs_block_group_cache *cache;
6414         int ret;
6415
6416         cache = btrfs_lookup_block_group(fs_info, bytenr);
6417         if (!cache)
6418                 return -EINVAL;
6419
6420         /*
6421          * pull in the free space cache (if any) so that our pin
6422          * removes the free space from the cache.  We have load_only set
6423          * to one because the slow code to read in the free extents does check
6424          * the pinned extents.
6425          */
6426         cache_block_group(cache, 1);
6427
6428         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6429
6430         /* remove us from the free space cache (if we're there at all) */
6431         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6432         btrfs_put_block_group(cache);
6433         return ret;
6434 }
6435
6436 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6437                                    u64 start, u64 num_bytes)
6438 {
6439         int ret;
6440         struct btrfs_block_group_cache *block_group;
6441         struct btrfs_caching_control *caching_ctl;
6442
6443         block_group = btrfs_lookup_block_group(fs_info, start);
6444         if (!block_group)
6445                 return -EINVAL;
6446
6447         cache_block_group(block_group, 0);
6448         caching_ctl = get_caching_control(block_group);
6449
6450         if (!caching_ctl) {
6451                 /* Logic error */
6452                 BUG_ON(!block_group_cache_done(block_group));
6453                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6454         } else {
6455                 mutex_lock(&caching_ctl->mutex);
6456
6457                 if (start >= caching_ctl->progress) {
6458                         ret = add_excluded_extent(fs_info, start, num_bytes);
6459                 } else if (start + num_bytes <= caching_ctl->progress) {
6460                         ret = btrfs_remove_free_space(block_group,
6461                                                       start, num_bytes);
6462                 } else {
6463                         num_bytes = caching_ctl->progress - start;
6464                         ret = btrfs_remove_free_space(block_group,
6465                                                       start, num_bytes);
6466                         if (ret)
6467                                 goto out_lock;
6468
6469                         num_bytes = (start + num_bytes) -
6470                                 caching_ctl->progress;
6471                         start = caching_ctl->progress;
6472                         ret = add_excluded_extent(fs_info, start, num_bytes);
6473                 }
6474 out_lock:
6475                 mutex_unlock(&caching_ctl->mutex);
6476                 put_caching_control(caching_ctl);
6477         }
6478         btrfs_put_block_group(block_group);
6479         return ret;
6480 }
6481
6482 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6483                                  struct extent_buffer *eb)
6484 {
6485         struct btrfs_file_extent_item *item;
6486         struct btrfs_key key;
6487         int found_type;
6488         int i;
6489         int ret = 0;
6490
6491         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6492                 return 0;
6493
6494         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6495                 btrfs_item_key_to_cpu(eb, &key, i);
6496                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6497                         continue;
6498                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6499                 found_type = btrfs_file_extent_type(eb, item);
6500                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6501                         continue;
6502                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6503                         continue;
6504                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6505                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6506                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6507                 if (ret)
6508                         break;
6509         }
6510
6511         return ret;
6512 }
6513
6514 static void
6515 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6516 {
6517         atomic_inc(&bg->reservations);
6518 }
6519
6520 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6521                                         const u64 start)
6522 {
6523         struct btrfs_block_group_cache *bg;
6524
6525         bg = btrfs_lookup_block_group(fs_info, start);
6526         ASSERT(bg);
6527         if (atomic_dec_and_test(&bg->reservations))
6528                 wake_up_var(&bg->reservations);
6529         btrfs_put_block_group(bg);
6530 }
6531
6532 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6533 {
6534         struct btrfs_space_info *space_info = bg->space_info;
6535
6536         ASSERT(bg->ro);
6537
6538         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6539                 return;
6540
6541         /*
6542          * Our block group is read only but before we set it to read only,
6543          * some task might have had allocated an extent from it already, but it
6544          * has not yet created a respective ordered extent (and added it to a
6545          * root's list of ordered extents).
6546          * Therefore wait for any task currently allocating extents, since the
6547          * block group's reservations counter is incremented while a read lock
6548          * on the groups' semaphore is held and decremented after releasing
6549          * the read access on that semaphore and creating the ordered extent.
6550          */
6551         down_write(&space_info->groups_sem);
6552         up_write(&space_info->groups_sem);
6553
6554         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6555 }
6556
6557 /**
6558  * btrfs_add_reserved_bytes - update the block_group and space info counters
6559  * @cache:      The cache we are manipulating
6560  * @ram_bytes:  The number of bytes of file content, and will be same to
6561  *              @num_bytes except for the compress path.
6562  * @num_bytes:  The number of bytes in question
6563  * @delalloc:   The blocks are allocated for the delalloc write
6564  *
6565  * This is called by the allocator when it reserves space. If this is a
6566  * reservation and the block group has become read only we cannot make the
6567  * reservation and return -EAGAIN, otherwise this function always succeeds.
6568  */
6569 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6570                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6571 {
6572         struct btrfs_space_info *space_info = cache->space_info;
6573         int ret = 0;
6574
6575         spin_lock(&space_info->lock);
6576         spin_lock(&cache->lock);
6577         if (cache->ro) {
6578                 ret = -EAGAIN;
6579         } else {
6580                 cache->reserved += num_bytes;
6581                 space_info->bytes_reserved += num_bytes;
6582
6583                 trace_btrfs_space_reservation(cache->fs_info,
6584                                 "space_info", space_info->flags,
6585                                 ram_bytes, 0);
6586                 space_info->bytes_may_use -= ram_bytes;
6587                 if (delalloc)
6588                         cache->delalloc_bytes += num_bytes;
6589         }
6590         spin_unlock(&cache->lock);
6591         spin_unlock(&space_info->lock);
6592         return ret;
6593 }
6594
6595 /**
6596  * btrfs_free_reserved_bytes - update the block_group and space info counters
6597  * @cache:      The cache we are manipulating
6598  * @num_bytes:  The number of bytes in question
6599  * @delalloc:   The blocks are allocated for the delalloc write
6600  *
6601  * This is called by somebody who is freeing space that was never actually used
6602  * on disk.  For example if you reserve some space for a new leaf in transaction
6603  * A and before transaction A commits you free that leaf, you call this with
6604  * reserve set to 0 in order to clear the reservation.
6605  */
6606
6607 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6608                                      u64 num_bytes, int delalloc)
6609 {
6610         struct btrfs_space_info *space_info = cache->space_info;
6611         int ret = 0;
6612
6613         spin_lock(&space_info->lock);
6614         spin_lock(&cache->lock);
6615         if (cache->ro)
6616                 space_info->bytes_readonly += num_bytes;
6617         cache->reserved -= num_bytes;
6618         space_info->bytes_reserved -= num_bytes;
6619
6620         if (delalloc)
6621                 cache->delalloc_bytes -= num_bytes;
6622         spin_unlock(&cache->lock);
6623         spin_unlock(&space_info->lock);
6624         return ret;
6625 }
6626 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6627 {
6628         struct btrfs_caching_control *next;
6629         struct btrfs_caching_control *caching_ctl;
6630         struct btrfs_block_group_cache *cache;
6631
6632         down_write(&fs_info->commit_root_sem);
6633
6634         list_for_each_entry_safe(caching_ctl, next,
6635                                  &fs_info->caching_block_groups, list) {
6636                 cache = caching_ctl->block_group;
6637                 if (block_group_cache_done(cache)) {
6638                         cache->last_byte_to_unpin = (u64)-1;
6639                         list_del_init(&caching_ctl->list);
6640                         put_caching_control(caching_ctl);
6641                 } else {
6642                         cache->last_byte_to_unpin = caching_ctl->progress;
6643                 }
6644         }
6645
6646         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6647                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6648         else
6649                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6650
6651         up_write(&fs_info->commit_root_sem);
6652
6653         update_global_block_rsv(fs_info);
6654 }
6655
6656 /*
6657  * Returns the free cluster for the given space info and sets empty_cluster to
6658  * what it should be based on the mount options.
6659  */
6660 static struct btrfs_free_cluster *
6661 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6662                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6663 {
6664         struct btrfs_free_cluster *ret = NULL;
6665
6666         *empty_cluster = 0;
6667         if (btrfs_mixed_space_info(space_info))
6668                 return ret;
6669
6670         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6671                 ret = &fs_info->meta_alloc_cluster;
6672                 if (btrfs_test_opt(fs_info, SSD))
6673                         *empty_cluster = SZ_2M;
6674                 else
6675                         *empty_cluster = SZ_64K;
6676         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6677                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6678                 *empty_cluster = SZ_2M;
6679                 ret = &fs_info->data_alloc_cluster;
6680         }
6681
6682         return ret;
6683 }
6684
6685 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6686                               u64 start, u64 end,
6687                               const bool return_free_space)
6688 {
6689         struct btrfs_block_group_cache *cache = NULL;
6690         struct btrfs_space_info *space_info;
6691         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6692         struct btrfs_free_cluster *cluster = NULL;
6693         u64 len;
6694         u64 total_unpinned = 0;
6695         u64 empty_cluster = 0;
6696         bool readonly;
6697
6698         while (start <= end) {
6699                 readonly = false;
6700                 if (!cache ||
6701                     start >= cache->key.objectid + cache->key.offset) {
6702                         if (cache)
6703                                 btrfs_put_block_group(cache);
6704                         total_unpinned = 0;
6705                         cache = btrfs_lookup_block_group(fs_info, start);
6706                         BUG_ON(!cache); /* Logic error */
6707
6708                         cluster = fetch_cluster_info(fs_info,
6709                                                      cache->space_info,
6710                                                      &empty_cluster);
6711                         empty_cluster <<= 1;
6712                 }
6713
6714                 len = cache->key.objectid + cache->key.offset - start;
6715                 len = min(len, end + 1 - start);
6716
6717                 if (start < cache->last_byte_to_unpin) {
6718                         len = min(len, cache->last_byte_to_unpin - start);
6719                         if (return_free_space)
6720                                 btrfs_add_free_space(cache, start, len);
6721                 }
6722
6723                 start += len;
6724                 total_unpinned += len;
6725                 space_info = cache->space_info;
6726
6727                 /*
6728                  * If this space cluster has been marked as fragmented and we've
6729                  * unpinned enough in this block group to potentially allow a
6730                  * cluster to be created inside of it go ahead and clear the
6731                  * fragmented check.
6732                  */
6733                 if (cluster && cluster->fragmented &&
6734                     total_unpinned > empty_cluster) {
6735                         spin_lock(&cluster->lock);
6736                         cluster->fragmented = 0;
6737                         spin_unlock(&cluster->lock);
6738                 }
6739
6740                 spin_lock(&space_info->lock);
6741                 spin_lock(&cache->lock);
6742                 cache->pinned -= len;
6743                 space_info->bytes_pinned -= len;
6744
6745                 trace_btrfs_space_reservation(fs_info, "pinned",
6746                                               space_info->flags, len, 0);
6747                 space_info->max_extent_size = 0;
6748                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6749                 if (cache->ro) {
6750                         space_info->bytes_readonly += len;
6751                         readonly = true;
6752                 }
6753                 spin_unlock(&cache->lock);
6754                 if (!readonly && return_free_space &&
6755                     global_rsv->space_info == space_info) {
6756                         u64 to_add = len;
6757
6758                         spin_lock(&global_rsv->lock);
6759                         if (!global_rsv->full) {
6760                                 to_add = min(len, global_rsv->size -
6761                                              global_rsv->reserved);
6762                                 global_rsv->reserved += to_add;
6763                                 space_info->bytes_may_use += to_add;
6764                                 if (global_rsv->reserved >= global_rsv->size)
6765                                         global_rsv->full = 1;
6766                                 trace_btrfs_space_reservation(fs_info,
6767                                                               "space_info",
6768                                                               space_info->flags,
6769                                                               to_add, 1);
6770                                 len -= to_add;
6771                         }
6772                         spin_unlock(&global_rsv->lock);
6773                         /* Add to any tickets we may have */
6774                         if (len)
6775                                 space_info_add_new_bytes(fs_info, space_info,
6776                                                          len);
6777                 }
6778                 spin_unlock(&space_info->lock);
6779         }
6780
6781         if (cache)
6782                 btrfs_put_block_group(cache);
6783         return 0;
6784 }
6785
6786 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6787 {
6788         struct btrfs_fs_info *fs_info = trans->fs_info;
6789         struct btrfs_block_group_cache *block_group, *tmp;
6790         struct list_head *deleted_bgs;
6791         struct extent_io_tree *unpin;
6792         u64 start;
6793         u64 end;
6794         int ret;
6795
6796         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6797                 unpin = &fs_info->freed_extents[1];
6798         else
6799                 unpin = &fs_info->freed_extents[0];
6800
6801         while (!trans->aborted) {
6802                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6803                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6804                                             EXTENT_DIRTY, NULL);
6805                 if (ret) {
6806                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6807                         break;
6808                 }
6809
6810                 if (btrfs_test_opt(fs_info, DISCARD))
6811                         ret = btrfs_discard_extent(fs_info, start,
6812                                                    end + 1 - start, NULL);
6813
6814                 clear_extent_dirty(unpin, start, end);
6815                 unpin_extent_range(fs_info, start, end, true);
6816                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6817                 cond_resched();
6818         }
6819
6820         /*
6821          * Transaction is finished.  We don't need the lock anymore.  We
6822          * do need to clean up the block groups in case of a transaction
6823          * abort.
6824          */
6825         deleted_bgs = &trans->transaction->deleted_bgs;
6826         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6827                 u64 trimmed = 0;
6828
6829                 ret = -EROFS;
6830                 if (!trans->aborted)
6831                         ret = btrfs_discard_extent(fs_info,
6832                                                    block_group->key.objectid,
6833                                                    block_group->key.offset,
6834                                                    &trimmed);
6835
6836                 list_del_init(&block_group->bg_list);
6837                 btrfs_put_block_group_trimming(block_group);
6838                 btrfs_put_block_group(block_group);
6839
6840                 if (ret) {
6841                         const char *errstr = btrfs_decode_error(ret);
6842                         btrfs_warn(fs_info,
6843                            "discard failed while removing blockgroup: errno=%d %s",
6844                                    ret, errstr);
6845                 }
6846         }
6847
6848         return 0;
6849 }
6850
6851 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6852                                 struct btrfs_fs_info *info,
6853                                 struct btrfs_delayed_ref_node *node, u64 parent,
6854                                 u64 root_objectid, u64 owner_objectid,
6855                                 u64 owner_offset, int refs_to_drop,
6856                                 struct btrfs_delayed_extent_op *extent_op)
6857 {
6858         struct btrfs_key key;
6859         struct btrfs_path *path;
6860         struct btrfs_root *extent_root = info->extent_root;
6861         struct extent_buffer *leaf;
6862         struct btrfs_extent_item *ei;
6863         struct btrfs_extent_inline_ref *iref;
6864         int ret;
6865         int is_data;
6866         int extent_slot = 0;
6867         int found_extent = 0;
6868         int num_to_del = 1;
6869         u32 item_size;
6870         u64 refs;
6871         u64 bytenr = node->bytenr;
6872         u64 num_bytes = node->num_bytes;
6873         int last_ref = 0;
6874         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6875
6876         path = btrfs_alloc_path();
6877         if (!path)
6878                 return -ENOMEM;
6879
6880         path->reada = READA_FORWARD;
6881         path->leave_spinning = 1;
6882
6883         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6884         BUG_ON(!is_data && refs_to_drop != 1);
6885
6886         if (is_data)
6887                 skinny_metadata = false;
6888
6889         ret = lookup_extent_backref(trans, info, path, &iref,
6890                                     bytenr, num_bytes, parent,
6891                                     root_objectid, owner_objectid,
6892                                     owner_offset);
6893         if (ret == 0) {
6894                 extent_slot = path->slots[0];
6895                 while (extent_slot >= 0) {
6896                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6897                                               extent_slot);
6898                         if (key.objectid != bytenr)
6899                                 break;
6900                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6901                             key.offset == num_bytes) {
6902                                 found_extent = 1;
6903                                 break;
6904                         }
6905                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6906                             key.offset == owner_objectid) {
6907                                 found_extent = 1;
6908                                 break;
6909                         }
6910                         if (path->slots[0] - extent_slot > 5)
6911                                 break;
6912                         extent_slot--;
6913                 }
6914 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6915                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6916                 if (found_extent && item_size < sizeof(*ei))
6917                         found_extent = 0;
6918 #endif
6919                 if (!found_extent) {
6920                         BUG_ON(iref);
6921                         ret = remove_extent_backref(trans, info, path, NULL,
6922                                                     refs_to_drop,
6923                                                     is_data, &last_ref);
6924                         if (ret) {
6925                                 btrfs_abort_transaction(trans, ret);
6926                                 goto out;
6927                         }
6928                         btrfs_release_path(path);
6929                         path->leave_spinning = 1;
6930
6931                         key.objectid = bytenr;
6932                         key.type = BTRFS_EXTENT_ITEM_KEY;
6933                         key.offset = num_bytes;
6934
6935                         if (!is_data && skinny_metadata) {
6936                                 key.type = BTRFS_METADATA_ITEM_KEY;
6937                                 key.offset = owner_objectid;
6938                         }
6939
6940                         ret = btrfs_search_slot(trans, extent_root,
6941                                                 &key, path, -1, 1);
6942                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6943                                 /*
6944                                  * Couldn't find our skinny metadata item,
6945                                  * see if we have ye olde extent item.
6946                                  */
6947                                 path->slots[0]--;
6948                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6949                                                       path->slots[0]);
6950                                 if (key.objectid == bytenr &&
6951                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6952                                     key.offset == num_bytes)
6953                                         ret = 0;
6954                         }
6955
6956                         if (ret > 0 && skinny_metadata) {
6957                                 skinny_metadata = false;
6958                                 key.objectid = bytenr;
6959                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6960                                 key.offset = num_bytes;
6961                                 btrfs_release_path(path);
6962                                 ret = btrfs_search_slot(trans, extent_root,
6963                                                         &key, path, -1, 1);
6964                         }
6965
6966                         if (ret) {
6967                                 btrfs_err(info,
6968                                           "umm, got %d back from search, was looking for %llu",
6969                                           ret, bytenr);
6970                                 if (ret > 0)
6971                                         btrfs_print_leaf(path->nodes[0]);
6972                         }
6973                         if (ret < 0) {
6974                                 btrfs_abort_transaction(trans, ret);
6975                                 goto out;
6976                         }
6977                         extent_slot = path->slots[0];
6978                 }
6979         } else if (WARN_ON(ret == -ENOENT)) {
6980                 btrfs_print_leaf(path->nodes[0]);
6981                 btrfs_err(info,
6982                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6983                         bytenr, parent, root_objectid, owner_objectid,
6984                         owner_offset);
6985                 btrfs_abort_transaction(trans, ret);
6986                 goto out;
6987         } else {
6988                 btrfs_abort_transaction(trans, ret);
6989                 goto out;
6990         }
6991
6992         leaf = path->nodes[0];
6993         item_size = btrfs_item_size_nr(leaf, extent_slot);
6994 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6995         if (item_size < sizeof(*ei)) {
6996                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6997                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6998                                              0);
6999                 if (ret < 0) {
7000                         btrfs_abort_transaction(trans, ret);
7001                         goto out;
7002                 }
7003
7004                 btrfs_release_path(path);
7005                 path->leave_spinning = 1;
7006
7007                 key.objectid = bytenr;
7008                 key.type = BTRFS_EXTENT_ITEM_KEY;
7009                 key.offset = num_bytes;
7010
7011                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7012                                         -1, 1);
7013                 if (ret) {
7014                         btrfs_err(info,
7015                                   "umm, got %d back from search, was looking for %llu",
7016                                 ret, bytenr);
7017                         btrfs_print_leaf(path->nodes[0]);
7018                 }
7019                 if (ret < 0) {
7020                         btrfs_abort_transaction(trans, ret);
7021                         goto out;
7022                 }
7023
7024                 extent_slot = path->slots[0];
7025                 leaf = path->nodes[0];
7026                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7027         }
7028 #endif
7029         BUG_ON(item_size < sizeof(*ei));
7030         ei = btrfs_item_ptr(leaf, extent_slot,
7031                             struct btrfs_extent_item);
7032         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7033             key.type == BTRFS_EXTENT_ITEM_KEY) {
7034                 struct btrfs_tree_block_info *bi;
7035                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7036                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7037                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7038         }
7039
7040         refs = btrfs_extent_refs(leaf, ei);
7041         if (refs < refs_to_drop) {
7042                 btrfs_err(info,
7043                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7044                           refs_to_drop, refs, bytenr);
7045                 ret = -EINVAL;
7046                 btrfs_abort_transaction(trans, ret);
7047                 goto out;
7048         }
7049         refs -= refs_to_drop;
7050
7051         if (refs > 0) {
7052                 if (extent_op)
7053                         __run_delayed_extent_op(extent_op, leaf, ei);
7054                 /*
7055                  * In the case of inline back ref, reference count will
7056                  * be updated by remove_extent_backref
7057                  */
7058                 if (iref) {
7059                         BUG_ON(!found_extent);
7060                 } else {
7061                         btrfs_set_extent_refs(leaf, ei, refs);
7062                         btrfs_mark_buffer_dirty(leaf);
7063                 }
7064                 if (found_extent) {
7065                         ret = remove_extent_backref(trans, info, path,
7066                                                     iref, refs_to_drop,
7067                                                     is_data, &last_ref);
7068                         if (ret) {
7069                                 btrfs_abort_transaction(trans, ret);
7070                                 goto out;
7071                         }
7072                 }
7073         } else {
7074                 if (found_extent) {
7075                         BUG_ON(is_data && refs_to_drop !=
7076                                extent_data_ref_count(path, iref));
7077                         if (iref) {
7078                                 BUG_ON(path->slots[0] != extent_slot);
7079                         } else {
7080                                 BUG_ON(path->slots[0] != extent_slot + 1);
7081                                 path->slots[0] = extent_slot;
7082                                 num_to_del = 2;
7083                         }
7084                 }
7085
7086                 last_ref = 1;
7087                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7088                                       num_to_del);
7089                 if (ret) {
7090                         btrfs_abort_transaction(trans, ret);
7091                         goto out;
7092                 }
7093                 btrfs_release_path(path);
7094
7095                 if (is_data) {
7096                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7097                         if (ret) {
7098                                 btrfs_abort_transaction(trans, ret);
7099                                 goto out;
7100                         }
7101                 }
7102
7103                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7104                 if (ret) {
7105                         btrfs_abort_transaction(trans, ret);
7106                         goto out;
7107                 }
7108
7109                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7110                 if (ret) {
7111                         btrfs_abort_transaction(trans, ret);
7112                         goto out;
7113                 }
7114         }
7115         btrfs_release_path(path);
7116
7117 out:
7118         btrfs_free_path(path);
7119         return ret;
7120 }
7121
7122 /*
7123  * when we free an block, it is possible (and likely) that we free the last
7124  * delayed ref for that extent as well.  This searches the delayed ref tree for
7125  * a given extent, and if there are no other delayed refs to be processed, it
7126  * removes it from the tree.
7127  */
7128 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7129                                       u64 bytenr)
7130 {
7131         struct btrfs_delayed_ref_head *head;
7132         struct btrfs_delayed_ref_root *delayed_refs;
7133         int ret = 0;
7134
7135         delayed_refs = &trans->transaction->delayed_refs;
7136         spin_lock(&delayed_refs->lock);
7137         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7138         if (!head)
7139                 goto out_delayed_unlock;
7140
7141         spin_lock(&head->lock);
7142         if (!RB_EMPTY_ROOT(&head->ref_tree))
7143                 goto out;
7144
7145         if (head->extent_op) {
7146                 if (!head->must_insert_reserved)
7147                         goto out;
7148                 btrfs_free_delayed_extent_op(head->extent_op);
7149                 head->extent_op = NULL;
7150         }
7151
7152         /*
7153          * waiting for the lock here would deadlock.  If someone else has it
7154          * locked they are already in the process of dropping it anyway
7155          */
7156         if (!mutex_trylock(&head->mutex))
7157                 goto out;
7158
7159         /*
7160          * at this point we have a head with no other entries.  Go
7161          * ahead and process it.
7162          */
7163         rb_erase(&head->href_node, &delayed_refs->href_root);
7164         RB_CLEAR_NODE(&head->href_node);
7165         atomic_dec(&delayed_refs->num_entries);
7166
7167         /*
7168          * we don't take a ref on the node because we're removing it from the
7169          * tree, so we just steal the ref the tree was holding.
7170          */
7171         delayed_refs->num_heads--;
7172         if (head->processing == 0)
7173                 delayed_refs->num_heads_ready--;
7174         head->processing = 0;
7175         spin_unlock(&head->lock);
7176         spin_unlock(&delayed_refs->lock);
7177
7178         BUG_ON(head->extent_op);
7179         if (head->must_insert_reserved)
7180                 ret = 1;
7181
7182         mutex_unlock(&head->mutex);
7183         btrfs_put_delayed_ref_head(head);
7184         return ret;
7185 out:
7186         spin_unlock(&head->lock);
7187
7188 out_delayed_unlock:
7189         spin_unlock(&delayed_refs->lock);
7190         return 0;
7191 }
7192
7193 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7194                            struct btrfs_root *root,
7195                            struct extent_buffer *buf,
7196                            u64 parent, int last_ref)
7197 {
7198         struct btrfs_fs_info *fs_info = root->fs_info;
7199         int pin = 1;
7200         int ret;
7201
7202         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7203                 int old_ref_mod, new_ref_mod;
7204
7205                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7206                                    root->root_key.objectid,
7207                                    btrfs_header_level(buf), 0,
7208                                    BTRFS_DROP_DELAYED_REF);
7209                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7210                                                  buf->len, parent,
7211                                                  root->root_key.objectid,
7212                                                  btrfs_header_level(buf),
7213                                                  BTRFS_DROP_DELAYED_REF, NULL,
7214                                                  &old_ref_mod, &new_ref_mod);
7215                 BUG_ON(ret); /* -ENOMEM */
7216                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7217         }
7218
7219         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7220                 struct btrfs_block_group_cache *cache;
7221
7222                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7223                         ret = check_ref_cleanup(trans, buf->start);
7224                         if (!ret)
7225                                 goto out;
7226                 }
7227
7228                 pin = 0;
7229                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7230
7231                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7232                         pin_down_extent(fs_info, cache, buf->start,
7233                                         buf->len, 1);
7234                         btrfs_put_block_group(cache);
7235                         goto out;
7236                 }
7237
7238                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7239
7240                 btrfs_add_free_space(cache, buf->start, buf->len);
7241                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7242                 btrfs_put_block_group(cache);
7243                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7244         }
7245 out:
7246         if (pin)
7247                 add_pinned_bytes(fs_info, buf->len, true,
7248                                  root->root_key.objectid);
7249
7250         if (last_ref) {
7251                 /*
7252                  * Deleting the buffer, clear the corrupt flag since it doesn't
7253                  * matter anymore.
7254                  */
7255                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7256         }
7257 }
7258
7259 /* Can return -ENOMEM */
7260 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7261                       struct btrfs_root *root,
7262                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7263                       u64 owner, u64 offset)
7264 {
7265         struct btrfs_fs_info *fs_info = root->fs_info;
7266         int old_ref_mod, new_ref_mod;
7267         int ret;
7268
7269         if (btrfs_is_testing(fs_info))
7270                 return 0;
7271
7272         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7273                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7274                                    root_objectid, owner, offset,
7275                                    BTRFS_DROP_DELAYED_REF);
7276
7277         /*
7278          * tree log blocks never actually go into the extent allocation
7279          * tree, just update pinning info and exit early.
7280          */
7281         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7282                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7283                 /* unlocks the pinned mutex */
7284                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7285                 old_ref_mod = new_ref_mod = 0;
7286                 ret = 0;
7287         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7288                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7289                                                  num_bytes, parent,
7290                                                  root_objectid, (int)owner,
7291                                                  BTRFS_DROP_DELAYED_REF, NULL,
7292                                                  &old_ref_mod, &new_ref_mod);
7293         } else {
7294                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7295                                                  num_bytes, parent,
7296                                                  root_objectid, owner, offset,
7297                                                  0, BTRFS_DROP_DELAYED_REF,
7298                                                  &old_ref_mod, &new_ref_mod);
7299         }
7300
7301         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7302                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7303
7304                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7305         }
7306
7307         return ret;
7308 }
7309
7310 /*
7311  * when we wait for progress in the block group caching, its because
7312  * our allocation attempt failed at least once.  So, we must sleep
7313  * and let some progress happen before we try again.
7314  *
7315  * This function will sleep at least once waiting for new free space to
7316  * show up, and then it will check the block group free space numbers
7317  * for our min num_bytes.  Another option is to have it go ahead
7318  * and look in the rbtree for a free extent of a given size, but this
7319  * is a good start.
7320  *
7321  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7322  * any of the information in this block group.
7323  */
7324 static noinline void
7325 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7326                                 u64 num_bytes)
7327 {
7328         struct btrfs_caching_control *caching_ctl;
7329
7330         caching_ctl = get_caching_control(cache);
7331         if (!caching_ctl)
7332                 return;
7333
7334         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7335                    (cache->free_space_ctl->free_space >= num_bytes));
7336
7337         put_caching_control(caching_ctl);
7338 }
7339
7340 static noinline int
7341 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7342 {
7343         struct btrfs_caching_control *caching_ctl;
7344         int ret = 0;
7345
7346         caching_ctl = get_caching_control(cache);
7347         if (!caching_ctl)
7348                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7349
7350         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7351         if (cache->cached == BTRFS_CACHE_ERROR)
7352                 ret = -EIO;
7353         put_caching_control(caching_ctl);
7354         return ret;
7355 }
7356
7357 enum btrfs_loop_type {
7358         LOOP_CACHING_NOWAIT = 0,
7359         LOOP_CACHING_WAIT = 1,
7360         LOOP_ALLOC_CHUNK = 2,
7361         LOOP_NO_EMPTY_SIZE = 3,
7362 };
7363
7364 static inline void
7365 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7366                        int delalloc)
7367 {
7368         if (delalloc)
7369                 down_read(&cache->data_rwsem);
7370 }
7371
7372 static inline void
7373 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7374                        int delalloc)
7375 {
7376         btrfs_get_block_group(cache);
7377         if (delalloc)
7378                 down_read(&cache->data_rwsem);
7379 }
7380
7381 static struct btrfs_block_group_cache *
7382 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7383                    struct btrfs_free_cluster *cluster,
7384                    int delalloc)
7385 {
7386         struct btrfs_block_group_cache *used_bg = NULL;
7387
7388         spin_lock(&cluster->refill_lock);
7389         while (1) {
7390                 used_bg = cluster->block_group;
7391                 if (!used_bg)
7392                         return NULL;
7393
7394                 if (used_bg == block_group)
7395                         return used_bg;
7396
7397                 btrfs_get_block_group(used_bg);
7398
7399                 if (!delalloc)
7400                         return used_bg;
7401
7402                 if (down_read_trylock(&used_bg->data_rwsem))
7403                         return used_bg;
7404
7405                 spin_unlock(&cluster->refill_lock);
7406
7407                 /* We should only have one-level nested. */
7408                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7409
7410                 spin_lock(&cluster->refill_lock);
7411                 if (used_bg == cluster->block_group)
7412                         return used_bg;
7413
7414                 up_read(&used_bg->data_rwsem);
7415                 btrfs_put_block_group(used_bg);
7416         }
7417 }
7418
7419 static inline void
7420 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7421                          int delalloc)
7422 {
7423         if (delalloc)
7424                 up_read(&cache->data_rwsem);
7425         btrfs_put_block_group(cache);
7426 }
7427
7428 /*
7429  * walks the btree of allocated extents and find a hole of a given size.
7430  * The key ins is changed to record the hole:
7431  * ins->objectid == start position
7432  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7433  * ins->offset == the size of the hole.
7434  * Any available blocks before search_start are skipped.
7435  *
7436  * If there is no suitable free space, we will record the max size of
7437  * the free space extent currently.
7438  */
7439 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7440                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7441                                 u64 hint_byte, struct btrfs_key *ins,
7442                                 u64 flags, int delalloc)
7443 {
7444         int ret = 0;
7445         struct btrfs_root *root = fs_info->extent_root;
7446         struct btrfs_free_cluster *last_ptr = NULL;
7447         struct btrfs_block_group_cache *block_group = NULL;
7448         u64 search_start = 0;
7449         u64 max_extent_size = 0;
7450         u64 empty_cluster = 0;
7451         struct btrfs_space_info *space_info;
7452         int loop = 0;
7453         int index = btrfs_bg_flags_to_raid_index(flags);
7454         bool failed_cluster_refill = false;
7455         bool failed_alloc = false;
7456         bool use_cluster = true;
7457         bool have_caching_bg = false;
7458         bool orig_have_caching_bg = false;
7459         bool full_search = false;
7460
7461         WARN_ON(num_bytes < fs_info->sectorsize);
7462         ins->type = BTRFS_EXTENT_ITEM_KEY;
7463         ins->objectid = 0;
7464         ins->offset = 0;
7465
7466         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7467
7468         space_info = __find_space_info(fs_info, flags);
7469         if (!space_info) {
7470                 btrfs_err(fs_info, "No space info for %llu", flags);
7471                 return -ENOSPC;
7472         }
7473
7474         /*
7475          * If our free space is heavily fragmented we may not be able to make
7476          * big contiguous allocations, so instead of doing the expensive search
7477          * for free space, simply return ENOSPC with our max_extent_size so we
7478          * can go ahead and search for a more manageable chunk.
7479          *
7480          * If our max_extent_size is large enough for our allocation simply
7481          * disable clustering since we will likely not be able to find enough
7482          * space to create a cluster and induce latency trying.
7483          */
7484         if (unlikely(space_info->max_extent_size)) {
7485                 spin_lock(&space_info->lock);
7486                 if (space_info->max_extent_size &&
7487                     num_bytes > space_info->max_extent_size) {
7488                         ins->offset = space_info->max_extent_size;
7489                         spin_unlock(&space_info->lock);
7490                         return -ENOSPC;
7491                 } else if (space_info->max_extent_size) {
7492                         use_cluster = false;
7493                 }
7494                 spin_unlock(&space_info->lock);
7495         }
7496
7497         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7498         if (last_ptr) {
7499                 spin_lock(&last_ptr->lock);
7500                 if (last_ptr->block_group)
7501                         hint_byte = last_ptr->window_start;
7502                 if (last_ptr->fragmented) {
7503                         /*
7504                          * We still set window_start so we can keep track of the
7505                          * last place we found an allocation to try and save
7506                          * some time.
7507                          */
7508                         hint_byte = last_ptr->window_start;
7509                         use_cluster = false;
7510                 }
7511                 spin_unlock(&last_ptr->lock);
7512         }
7513
7514         search_start = max(search_start, first_logical_byte(fs_info, 0));
7515         search_start = max(search_start, hint_byte);
7516         if (search_start == hint_byte) {
7517                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7518                 /*
7519                  * we don't want to use the block group if it doesn't match our
7520                  * allocation bits, or if its not cached.
7521                  *
7522                  * However if we are re-searching with an ideal block group
7523                  * picked out then we don't care that the block group is cached.
7524                  */
7525                 if (block_group && block_group_bits(block_group, flags) &&
7526                     block_group->cached != BTRFS_CACHE_NO) {
7527                         down_read(&space_info->groups_sem);
7528                         if (list_empty(&block_group->list) ||
7529                             block_group->ro) {
7530                                 /*
7531                                  * someone is removing this block group,
7532                                  * we can't jump into the have_block_group
7533                                  * target because our list pointers are not
7534                                  * valid
7535                                  */
7536                                 btrfs_put_block_group(block_group);
7537                                 up_read(&space_info->groups_sem);
7538                         } else {
7539                                 index = btrfs_bg_flags_to_raid_index(
7540                                                 block_group->flags);
7541                                 btrfs_lock_block_group(block_group, delalloc);
7542                                 goto have_block_group;
7543                         }
7544                 } else if (block_group) {
7545                         btrfs_put_block_group(block_group);
7546                 }
7547         }
7548 search:
7549         have_caching_bg = false;
7550         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7551                 full_search = true;
7552         down_read(&space_info->groups_sem);
7553         list_for_each_entry(block_group, &space_info->block_groups[index],
7554                             list) {
7555                 u64 offset;
7556                 int cached;
7557
7558                 /* If the block group is read-only, we can skip it entirely. */
7559                 if (unlikely(block_group->ro))
7560                         continue;
7561
7562                 btrfs_grab_block_group(block_group, delalloc);
7563                 search_start = block_group->key.objectid;
7564
7565                 /*
7566                  * this can happen if we end up cycling through all the
7567                  * raid types, but we want to make sure we only allocate
7568                  * for the proper type.
7569                  */
7570                 if (!block_group_bits(block_group, flags)) {
7571                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7572                                 BTRFS_BLOCK_GROUP_RAID1 |
7573                                 BTRFS_BLOCK_GROUP_RAID5 |
7574                                 BTRFS_BLOCK_GROUP_RAID6 |
7575                                 BTRFS_BLOCK_GROUP_RAID10;
7576
7577                         /*
7578                          * if they asked for extra copies and this block group
7579                          * doesn't provide them, bail.  This does allow us to
7580                          * fill raid0 from raid1.
7581                          */
7582                         if ((flags & extra) && !(block_group->flags & extra))
7583                                 goto loop;
7584                 }
7585
7586 have_block_group:
7587                 cached = block_group_cache_done(block_group);
7588                 if (unlikely(!cached)) {
7589                         have_caching_bg = true;
7590                         ret = cache_block_group(block_group, 0);
7591                         BUG_ON(ret < 0);
7592                         ret = 0;
7593                 }
7594
7595                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7596                         goto loop;
7597
7598                 /*
7599                  * Ok we want to try and use the cluster allocator, so
7600                  * lets look there
7601                  */
7602                 if (last_ptr && use_cluster) {
7603                         struct btrfs_block_group_cache *used_block_group;
7604                         unsigned long aligned_cluster;
7605                         /*
7606                          * the refill lock keeps out other
7607                          * people trying to start a new cluster
7608                          */
7609                         used_block_group = btrfs_lock_cluster(block_group,
7610                                                               last_ptr,
7611                                                               delalloc);
7612                         if (!used_block_group)
7613                                 goto refill_cluster;
7614
7615                         if (used_block_group != block_group &&
7616                             (used_block_group->ro ||
7617                              !block_group_bits(used_block_group, flags)))
7618                                 goto release_cluster;
7619
7620                         offset = btrfs_alloc_from_cluster(used_block_group,
7621                                                 last_ptr,
7622                                                 num_bytes,
7623                                                 used_block_group->key.objectid,
7624                                                 &max_extent_size);
7625                         if (offset) {
7626                                 /* we have a block, we're done */
7627                                 spin_unlock(&last_ptr->refill_lock);
7628                                 trace_btrfs_reserve_extent_cluster(
7629                                                 used_block_group,
7630                                                 search_start, num_bytes);
7631                                 if (used_block_group != block_group) {
7632                                         btrfs_release_block_group(block_group,
7633                                                                   delalloc);
7634                                         block_group = used_block_group;
7635                                 }
7636                                 goto checks;
7637                         }
7638
7639                         WARN_ON(last_ptr->block_group != used_block_group);
7640 release_cluster:
7641                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7642                          * set up a new clusters, so lets just skip it
7643                          * and let the allocator find whatever block
7644                          * it can find.  If we reach this point, we
7645                          * will have tried the cluster allocator
7646                          * plenty of times and not have found
7647                          * anything, so we are likely way too
7648                          * fragmented for the clustering stuff to find
7649                          * anything.
7650                          *
7651                          * However, if the cluster is taken from the
7652                          * current block group, release the cluster
7653                          * first, so that we stand a better chance of
7654                          * succeeding in the unclustered
7655                          * allocation.  */
7656                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7657                             used_block_group != block_group) {
7658                                 spin_unlock(&last_ptr->refill_lock);
7659                                 btrfs_release_block_group(used_block_group,
7660                                                           delalloc);
7661                                 goto unclustered_alloc;
7662                         }
7663
7664                         /*
7665                          * this cluster didn't work out, free it and
7666                          * start over
7667                          */
7668                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7669
7670                         if (used_block_group != block_group)
7671                                 btrfs_release_block_group(used_block_group,
7672                                                           delalloc);
7673 refill_cluster:
7674                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7675                                 spin_unlock(&last_ptr->refill_lock);
7676                                 goto unclustered_alloc;
7677                         }
7678
7679                         aligned_cluster = max_t(unsigned long,
7680                                                 empty_cluster + empty_size,
7681                                               block_group->full_stripe_len);
7682
7683                         /* allocate a cluster in this block group */
7684                         ret = btrfs_find_space_cluster(fs_info, block_group,
7685                                                        last_ptr, search_start,
7686                                                        num_bytes,
7687                                                        aligned_cluster);
7688                         if (ret == 0) {
7689                                 /*
7690                                  * now pull our allocation out of this
7691                                  * cluster
7692                                  */
7693                                 offset = btrfs_alloc_from_cluster(block_group,
7694                                                         last_ptr,
7695                                                         num_bytes,
7696                                                         search_start,
7697                                                         &max_extent_size);
7698                                 if (offset) {
7699                                         /* we found one, proceed */
7700                                         spin_unlock(&last_ptr->refill_lock);
7701                                         trace_btrfs_reserve_extent_cluster(
7702                                                 block_group, search_start,
7703                                                 num_bytes);
7704                                         goto checks;
7705                                 }
7706                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7707                                    && !failed_cluster_refill) {
7708                                 spin_unlock(&last_ptr->refill_lock);
7709
7710                                 failed_cluster_refill = true;
7711                                 wait_block_group_cache_progress(block_group,
7712                                        num_bytes + empty_cluster + empty_size);
7713                                 goto have_block_group;
7714                         }
7715
7716                         /*
7717                          * at this point we either didn't find a cluster
7718                          * or we weren't able to allocate a block from our
7719                          * cluster.  Free the cluster we've been trying
7720                          * to use, and go to the next block group
7721                          */
7722                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7723                         spin_unlock(&last_ptr->refill_lock);
7724                         goto loop;
7725                 }
7726
7727 unclustered_alloc:
7728                 /*
7729                  * We are doing an unclustered alloc, set the fragmented flag so
7730                  * we don't bother trying to setup a cluster again until we get
7731                  * more space.
7732                  */
7733                 if (unlikely(last_ptr)) {
7734                         spin_lock(&last_ptr->lock);
7735                         last_ptr->fragmented = 1;
7736                         spin_unlock(&last_ptr->lock);
7737                 }
7738                 if (cached) {
7739                         struct btrfs_free_space_ctl *ctl =
7740                                 block_group->free_space_ctl;
7741
7742                         spin_lock(&ctl->tree_lock);
7743                         if (ctl->free_space <
7744                             num_bytes + empty_cluster + empty_size) {
7745                                 if (ctl->free_space > max_extent_size)
7746                                         max_extent_size = ctl->free_space;
7747                                 spin_unlock(&ctl->tree_lock);
7748                                 goto loop;
7749                         }
7750                         spin_unlock(&ctl->tree_lock);
7751                 }
7752
7753                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7754                                                     num_bytes, empty_size,
7755                                                     &max_extent_size);
7756                 /*
7757                  * If we didn't find a chunk, and we haven't failed on this
7758                  * block group before, and this block group is in the middle of
7759                  * caching and we are ok with waiting, then go ahead and wait
7760                  * for progress to be made, and set failed_alloc to true.
7761                  *
7762                  * If failed_alloc is true then we've already waited on this
7763                  * block group once and should move on to the next block group.
7764                  */
7765                 if (!offset && !failed_alloc && !cached &&
7766                     loop > LOOP_CACHING_NOWAIT) {
7767                         wait_block_group_cache_progress(block_group,
7768                                                 num_bytes + empty_size);
7769                         failed_alloc = true;
7770                         goto have_block_group;
7771                 } else if (!offset) {
7772                         goto loop;
7773                 }
7774 checks:
7775                 search_start = ALIGN(offset, fs_info->stripesize);
7776
7777                 /* move on to the next group */
7778                 if (search_start + num_bytes >
7779                     block_group->key.objectid + block_group->key.offset) {
7780                         btrfs_add_free_space(block_group, offset, num_bytes);
7781                         goto loop;
7782                 }
7783
7784                 if (offset < search_start)
7785                         btrfs_add_free_space(block_group, offset,
7786                                              search_start - offset);
7787                 BUG_ON(offset > search_start);
7788
7789                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7790                                 num_bytes, delalloc);
7791                 if (ret == -EAGAIN) {
7792                         btrfs_add_free_space(block_group, offset, num_bytes);
7793                         goto loop;
7794                 }
7795                 btrfs_inc_block_group_reservations(block_group);
7796
7797                 /* we are all good, lets return */
7798                 ins->objectid = search_start;
7799                 ins->offset = num_bytes;
7800
7801                 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7802                 btrfs_release_block_group(block_group, delalloc);
7803                 break;
7804 loop:
7805                 failed_cluster_refill = false;
7806                 failed_alloc = false;
7807                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7808                        index);
7809                 btrfs_release_block_group(block_group, delalloc);
7810                 cond_resched();
7811         }
7812         up_read(&space_info->groups_sem);
7813
7814         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7815                 && !orig_have_caching_bg)
7816                 orig_have_caching_bg = true;
7817
7818         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7819                 goto search;
7820
7821         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7822                 goto search;
7823
7824         /*
7825          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7826          *                      caching kthreads as we move along
7827          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7828          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7829          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7830          *                      again
7831          */
7832         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7833                 index = 0;
7834                 if (loop == LOOP_CACHING_NOWAIT) {
7835                         /*
7836                          * We want to skip the LOOP_CACHING_WAIT step if we
7837                          * don't have any uncached bgs and we've already done a
7838                          * full search through.
7839                          */
7840                         if (orig_have_caching_bg || !full_search)
7841                                 loop = LOOP_CACHING_WAIT;
7842                         else
7843                                 loop = LOOP_ALLOC_CHUNK;
7844                 } else {
7845                         loop++;
7846                 }
7847
7848                 if (loop == LOOP_ALLOC_CHUNK) {
7849                         struct btrfs_trans_handle *trans;
7850                         int exist = 0;
7851
7852                         trans = current->journal_info;
7853                         if (trans)
7854                                 exist = 1;
7855                         else
7856                                 trans = btrfs_join_transaction(root);
7857
7858                         if (IS_ERR(trans)) {
7859                                 ret = PTR_ERR(trans);
7860                                 goto out;
7861                         }
7862
7863                         ret = do_chunk_alloc(trans, fs_info, flags,
7864                                              CHUNK_ALLOC_FORCE);
7865
7866                         /*
7867                          * If we can't allocate a new chunk we've already looped
7868                          * through at least once, move on to the NO_EMPTY_SIZE
7869                          * case.
7870                          */
7871                         if (ret == -ENOSPC)
7872                                 loop = LOOP_NO_EMPTY_SIZE;
7873
7874                         /*
7875                          * Do not bail out on ENOSPC since we
7876                          * can do more things.
7877                          */
7878                         if (ret < 0 && ret != -ENOSPC)
7879                                 btrfs_abort_transaction(trans, ret);
7880                         else
7881                                 ret = 0;
7882                         if (!exist)
7883                                 btrfs_end_transaction(trans);
7884                         if (ret)
7885                                 goto out;
7886                 }
7887
7888                 if (loop == LOOP_NO_EMPTY_SIZE) {
7889                         /*
7890                          * Don't loop again if we already have no empty_size and
7891                          * no empty_cluster.
7892                          */
7893                         if (empty_size == 0 &&
7894                             empty_cluster == 0) {
7895                                 ret = -ENOSPC;
7896                                 goto out;
7897                         }
7898                         empty_size = 0;
7899                         empty_cluster = 0;
7900                 }
7901
7902                 goto search;
7903         } else if (!ins->objectid) {
7904                 ret = -ENOSPC;
7905         } else if (ins->objectid) {
7906                 if (!use_cluster && last_ptr) {
7907                         spin_lock(&last_ptr->lock);
7908                         last_ptr->window_start = ins->objectid;
7909                         spin_unlock(&last_ptr->lock);
7910                 }
7911                 ret = 0;
7912         }
7913 out:
7914         if (ret == -ENOSPC) {
7915                 spin_lock(&space_info->lock);
7916                 space_info->max_extent_size = max_extent_size;
7917                 spin_unlock(&space_info->lock);
7918                 ins->offset = max_extent_size;
7919         }
7920         return ret;
7921 }
7922
7923 static void dump_space_info(struct btrfs_fs_info *fs_info,
7924                             struct btrfs_space_info *info, u64 bytes,
7925                             int dump_block_groups)
7926 {
7927         struct btrfs_block_group_cache *cache;
7928         int index = 0;
7929
7930         spin_lock(&info->lock);
7931         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7932                    info->flags,
7933                    info->total_bytes - btrfs_space_info_used(info, true),
7934                    info->full ? "" : "not ");
7935         btrfs_info(fs_info,
7936                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7937                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7938                 info->bytes_reserved, info->bytes_may_use,
7939                 info->bytes_readonly);
7940         spin_unlock(&info->lock);
7941
7942         if (!dump_block_groups)
7943                 return;
7944
7945         down_read(&info->groups_sem);
7946 again:
7947         list_for_each_entry(cache, &info->block_groups[index], list) {
7948                 spin_lock(&cache->lock);
7949                 btrfs_info(fs_info,
7950                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7951                         cache->key.objectid, cache->key.offset,
7952                         btrfs_block_group_used(&cache->item), cache->pinned,
7953                         cache->reserved, cache->ro ? "[readonly]" : "");
7954                 btrfs_dump_free_space(cache, bytes);
7955                 spin_unlock(&cache->lock);
7956         }
7957         if (++index < BTRFS_NR_RAID_TYPES)
7958                 goto again;
7959         up_read(&info->groups_sem);
7960 }
7961
7962 /*
7963  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7964  *                        hole that is at least as big as @num_bytes.
7965  *
7966  * @root           -    The root that will contain this extent
7967  *
7968  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7969  *                      is used for accounting purposes. This value differs
7970  *                      from @num_bytes only in the case of compressed extents.
7971  *
7972  * @num_bytes      -    Number of bytes to allocate on-disk.
7973  *
7974  * @min_alloc_size -    Indicates the minimum amount of space that the
7975  *                      allocator should try to satisfy. In some cases
7976  *                      @num_bytes may be larger than what is required and if
7977  *                      the filesystem is fragmented then allocation fails.
7978  *                      However, the presence of @min_alloc_size gives a
7979  *                      chance to try and satisfy the smaller allocation.
7980  *
7981  * @empty_size     -    A hint that you plan on doing more COW. This is the
7982  *                      size in bytes the allocator should try to find free
7983  *                      next to the block it returns.  This is just a hint and
7984  *                      may be ignored by the allocator.
7985  *
7986  * @hint_byte      -    Hint to the allocator to start searching above the byte
7987  *                      address passed. It might be ignored.
7988  *
7989  * @ins            -    This key is modified to record the found hole. It will
7990  *                      have the following values:
7991  *                      ins->objectid == start position
7992  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7993  *                      ins->offset == the size of the hole.
7994  *
7995  * @is_data        -    Boolean flag indicating whether an extent is
7996  *                      allocated for data (true) or metadata (false)
7997  *
7998  * @delalloc       -    Boolean flag indicating whether this allocation is for
7999  *                      delalloc or not. If 'true' data_rwsem of block groups
8000  *                      is going to be acquired.
8001  *
8002  *
8003  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8004  * case -ENOSPC is returned then @ins->offset will contain the size of the
8005  * largest available hole the allocator managed to find.
8006  */
8007 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8008                          u64 num_bytes, u64 min_alloc_size,
8009                          u64 empty_size, u64 hint_byte,
8010                          struct btrfs_key *ins, int is_data, int delalloc)
8011 {
8012         struct btrfs_fs_info *fs_info = root->fs_info;
8013         bool final_tried = num_bytes == min_alloc_size;
8014         u64 flags;
8015         int ret;
8016
8017         flags = get_alloc_profile_by_root(root, is_data);
8018 again:
8019         WARN_ON(num_bytes < fs_info->sectorsize);
8020         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8021                                hint_byte, ins, flags, delalloc);
8022         if (!ret && !is_data) {
8023                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8024         } else if (ret == -ENOSPC) {
8025                 if (!final_tried && ins->offset) {
8026                         num_bytes = min(num_bytes >> 1, ins->offset);
8027                         num_bytes = round_down(num_bytes,
8028                                                fs_info->sectorsize);
8029                         num_bytes = max(num_bytes, min_alloc_size);
8030                         ram_bytes = num_bytes;
8031                         if (num_bytes == min_alloc_size)
8032                                 final_tried = true;
8033                         goto again;
8034                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8035                         struct btrfs_space_info *sinfo;
8036
8037                         sinfo = __find_space_info(fs_info, flags);
8038                         btrfs_err(fs_info,
8039                                   "allocation failed flags %llu, wanted %llu",
8040                                   flags, num_bytes);
8041                         if (sinfo)
8042                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8043                 }
8044         }
8045
8046         return ret;
8047 }
8048
8049 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8050                                         u64 start, u64 len,
8051                                         int pin, int delalloc)
8052 {
8053         struct btrfs_block_group_cache *cache;
8054         int ret = 0;
8055
8056         cache = btrfs_lookup_block_group(fs_info, start);
8057         if (!cache) {
8058                 btrfs_err(fs_info, "Unable to find block group for %llu",
8059                           start);
8060                 return -ENOSPC;
8061         }
8062
8063         if (pin)
8064                 pin_down_extent(fs_info, cache, start, len, 1);
8065         else {
8066                 if (btrfs_test_opt(fs_info, DISCARD))
8067                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8068                 btrfs_add_free_space(cache, start, len);
8069                 btrfs_free_reserved_bytes(cache, len, delalloc);
8070                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8071         }
8072
8073         btrfs_put_block_group(cache);
8074         return ret;
8075 }
8076
8077 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8078                                u64 start, u64 len, int delalloc)
8079 {
8080         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8081 }
8082
8083 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8084                                        u64 start, u64 len)
8085 {
8086         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8087 }
8088
8089 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8090                                       struct btrfs_fs_info *fs_info,
8091                                       u64 parent, u64 root_objectid,
8092                                       u64 flags, u64 owner, u64 offset,
8093                                       struct btrfs_key *ins, int ref_mod)
8094 {
8095         int ret;
8096         struct btrfs_extent_item *extent_item;
8097         struct btrfs_extent_inline_ref *iref;
8098         struct btrfs_path *path;
8099         struct extent_buffer *leaf;
8100         int type;
8101         u32 size;
8102
8103         if (parent > 0)
8104                 type = BTRFS_SHARED_DATA_REF_KEY;
8105         else
8106                 type = BTRFS_EXTENT_DATA_REF_KEY;
8107
8108         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8109
8110         path = btrfs_alloc_path();
8111         if (!path)
8112                 return -ENOMEM;
8113
8114         path->leave_spinning = 1;
8115         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8116                                       ins, size);
8117         if (ret) {
8118                 btrfs_free_path(path);
8119                 return ret;
8120         }
8121
8122         leaf = path->nodes[0];
8123         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8124                                      struct btrfs_extent_item);
8125         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8126         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8127         btrfs_set_extent_flags(leaf, extent_item,
8128                                flags | BTRFS_EXTENT_FLAG_DATA);
8129
8130         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8131         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8132         if (parent > 0) {
8133                 struct btrfs_shared_data_ref *ref;
8134                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8135                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8136                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8137         } else {
8138                 struct btrfs_extent_data_ref *ref;
8139                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8140                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8141                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8142                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8143                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8144         }
8145
8146         btrfs_mark_buffer_dirty(path->nodes[0]);
8147         btrfs_free_path(path);
8148
8149         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8150         if (ret)
8151                 return ret;
8152
8153         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8154         if (ret) { /* -ENOENT, logic error */
8155                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8156                         ins->objectid, ins->offset);
8157                 BUG();
8158         }
8159         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8160         return ret;
8161 }
8162
8163 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8164                                      struct btrfs_delayed_ref_node *node,
8165                                      struct btrfs_delayed_extent_op *extent_op)
8166 {
8167         struct btrfs_fs_info *fs_info = trans->fs_info;
8168         int ret;
8169         struct btrfs_extent_item *extent_item;
8170         struct btrfs_key extent_key;
8171         struct btrfs_tree_block_info *block_info;
8172         struct btrfs_extent_inline_ref *iref;
8173         struct btrfs_path *path;
8174         struct extent_buffer *leaf;
8175         struct btrfs_delayed_tree_ref *ref;
8176         u32 size = sizeof(*extent_item) + sizeof(*iref);
8177         u64 num_bytes;
8178         u64 flags = extent_op->flags_to_set;
8179         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8180
8181         ref = btrfs_delayed_node_to_tree_ref(node);
8182
8183         extent_key.objectid = node->bytenr;
8184         if (skinny_metadata) {
8185                 extent_key.offset = ref->level;
8186                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8187                 num_bytes = fs_info->nodesize;
8188         } else {
8189                 extent_key.offset = node->num_bytes;
8190                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8191                 size += sizeof(*block_info);
8192                 num_bytes = node->num_bytes;
8193         }
8194
8195         path = btrfs_alloc_path();
8196         if (!path) {
8197                 btrfs_free_and_pin_reserved_extent(fs_info,
8198                                                    extent_key.objectid,
8199                                                    fs_info->nodesize);
8200                 return -ENOMEM;
8201         }
8202
8203         path->leave_spinning = 1;
8204         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8205                                       &extent_key, size);
8206         if (ret) {
8207                 btrfs_free_path(path);
8208                 btrfs_free_and_pin_reserved_extent(fs_info,
8209                                                    extent_key.objectid,
8210                                                    fs_info->nodesize);
8211                 return ret;
8212         }
8213
8214         leaf = path->nodes[0];
8215         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8216                                      struct btrfs_extent_item);
8217         btrfs_set_extent_refs(leaf, extent_item, 1);
8218         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8219         btrfs_set_extent_flags(leaf, extent_item,
8220                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8221
8222         if (skinny_metadata) {
8223                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8224         } else {
8225                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8226                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8227                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8228                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8229         }
8230
8231         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8232                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8233                 btrfs_set_extent_inline_ref_type(leaf, iref,
8234                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8235                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8236         } else {
8237                 btrfs_set_extent_inline_ref_type(leaf, iref,
8238                                                  BTRFS_TREE_BLOCK_REF_KEY);
8239                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8240         }
8241
8242         btrfs_mark_buffer_dirty(leaf);
8243         btrfs_free_path(path);
8244
8245         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8246                                           num_bytes);
8247         if (ret)
8248                 return ret;
8249
8250         ret = update_block_group(trans, fs_info, extent_key.objectid,
8251                                  fs_info->nodesize, 1);
8252         if (ret) { /* -ENOENT, logic error */
8253                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8254                         extent_key.objectid, extent_key.offset);
8255                 BUG();
8256         }
8257
8258         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8259                                           fs_info->nodesize);
8260         return ret;
8261 }
8262
8263 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8264                                      struct btrfs_root *root, u64 owner,
8265                                      u64 offset, u64 ram_bytes,
8266                                      struct btrfs_key *ins)
8267 {
8268         struct btrfs_fs_info *fs_info = root->fs_info;
8269         int ret;
8270
8271         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8272
8273         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8274                            root->root_key.objectid, owner, offset,
8275                            BTRFS_ADD_DELAYED_EXTENT);
8276
8277         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8278                                          ins->offset, 0,
8279                                          root->root_key.objectid, owner,
8280                                          offset, ram_bytes,
8281                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8282         return ret;
8283 }
8284
8285 /*
8286  * this is used by the tree logging recovery code.  It records that
8287  * an extent has been allocated and makes sure to clear the free
8288  * space cache bits as well
8289  */
8290 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8291                                    struct btrfs_fs_info *fs_info,
8292                                    u64 root_objectid, u64 owner, u64 offset,
8293                                    struct btrfs_key *ins)
8294 {
8295         int ret;
8296         struct btrfs_block_group_cache *block_group;
8297         struct btrfs_space_info *space_info;
8298
8299         /*
8300          * Mixed block groups will exclude before processing the log so we only
8301          * need to do the exclude dance if this fs isn't mixed.
8302          */
8303         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8304                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8305                                               ins->offset);
8306                 if (ret)
8307                         return ret;
8308         }
8309
8310         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8311         if (!block_group)
8312                 return -EINVAL;
8313
8314         space_info = block_group->space_info;
8315         spin_lock(&space_info->lock);
8316         spin_lock(&block_group->lock);
8317         space_info->bytes_reserved += ins->offset;
8318         block_group->reserved += ins->offset;
8319         spin_unlock(&block_group->lock);
8320         spin_unlock(&space_info->lock);
8321
8322         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8323                                          0, owner, offset, ins, 1);
8324         btrfs_put_block_group(block_group);
8325         return ret;
8326 }
8327
8328 static struct extent_buffer *
8329 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8330                       u64 bytenr, int level)
8331 {
8332         struct btrfs_fs_info *fs_info = root->fs_info;
8333         struct extent_buffer *buf;
8334
8335         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8336         if (IS_ERR(buf))
8337                 return buf;
8338
8339         btrfs_set_header_generation(buf, trans->transid);
8340         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8341         btrfs_tree_lock(buf);
8342         clean_tree_block(fs_info, buf);
8343         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8344
8345         btrfs_set_lock_blocking(buf);
8346         set_extent_buffer_uptodate(buf);
8347
8348         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8349                 buf->log_index = root->log_transid % 2;
8350                 /*
8351                  * we allow two log transactions at a time, use different
8352                  * EXENT bit to differentiate dirty pages.
8353                  */
8354                 if (buf->log_index == 0)
8355                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8356                                         buf->start + buf->len - 1, GFP_NOFS);
8357                 else
8358                         set_extent_new(&root->dirty_log_pages, buf->start,
8359                                         buf->start + buf->len - 1);
8360         } else {
8361                 buf->log_index = -1;
8362                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8363                          buf->start + buf->len - 1, GFP_NOFS);
8364         }
8365         trans->dirty = true;
8366         /* this returns a buffer locked for blocking */
8367         return buf;
8368 }
8369
8370 static struct btrfs_block_rsv *
8371 use_block_rsv(struct btrfs_trans_handle *trans,
8372               struct btrfs_root *root, u32 blocksize)
8373 {
8374         struct btrfs_fs_info *fs_info = root->fs_info;
8375         struct btrfs_block_rsv *block_rsv;
8376         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8377         int ret;
8378         bool global_updated = false;
8379
8380         block_rsv = get_block_rsv(trans, root);
8381
8382         if (unlikely(block_rsv->size == 0))
8383                 goto try_reserve;
8384 again:
8385         ret = block_rsv_use_bytes(block_rsv, blocksize);
8386         if (!ret)
8387                 return block_rsv;
8388
8389         if (block_rsv->failfast)
8390                 return ERR_PTR(ret);
8391
8392         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8393                 global_updated = true;
8394                 update_global_block_rsv(fs_info);
8395                 goto again;
8396         }
8397
8398         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8399                 static DEFINE_RATELIMIT_STATE(_rs,
8400                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8401                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8402                 if (__ratelimit(&_rs))
8403                         WARN(1, KERN_DEBUG
8404                                 "BTRFS: block rsv returned %d\n", ret);
8405         }
8406 try_reserve:
8407         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8408                                      BTRFS_RESERVE_NO_FLUSH);
8409         if (!ret)
8410                 return block_rsv;
8411         /*
8412          * If we couldn't reserve metadata bytes try and use some from
8413          * the global reserve if its space type is the same as the global
8414          * reservation.
8415          */
8416         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8417             block_rsv->space_info == global_rsv->space_info) {
8418                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8419                 if (!ret)
8420                         return global_rsv;
8421         }
8422         return ERR_PTR(ret);
8423 }
8424
8425 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8426                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8427 {
8428         block_rsv_add_bytes(block_rsv, blocksize, 0);
8429         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8430 }
8431
8432 /*
8433  * finds a free extent and does all the dirty work required for allocation
8434  * returns the tree buffer or an ERR_PTR on error.
8435  */
8436 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8437                                              struct btrfs_root *root,
8438                                              u64 parent, u64 root_objectid,
8439                                              const struct btrfs_disk_key *key,
8440                                              int level, u64 hint,
8441                                              u64 empty_size)
8442 {
8443         struct btrfs_fs_info *fs_info = root->fs_info;
8444         struct btrfs_key ins;
8445         struct btrfs_block_rsv *block_rsv;
8446         struct extent_buffer *buf;
8447         struct btrfs_delayed_extent_op *extent_op;
8448         u64 flags = 0;
8449         int ret;
8450         u32 blocksize = fs_info->nodesize;
8451         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8452
8453 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8454         if (btrfs_is_testing(fs_info)) {
8455                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8456                                             level);
8457                 if (!IS_ERR(buf))
8458                         root->alloc_bytenr += blocksize;
8459                 return buf;
8460         }
8461 #endif
8462
8463         block_rsv = use_block_rsv(trans, root, blocksize);
8464         if (IS_ERR(block_rsv))
8465                 return ERR_CAST(block_rsv);
8466
8467         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8468                                    empty_size, hint, &ins, 0, 0);
8469         if (ret)
8470                 goto out_unuse;
8471
8472         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8473         if (IS_ERR(buf)) {
8474                 ret = PTR_ERR(buf);
8475                 goto out_free_reserved;
8476         }
8477
8478         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8479                 if (parent == 0)
8480                         parent = ins.objectid;
8481                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8482         } else
8483                 BUG_ON(parent > 0);
8484
8485         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8486                 extent_op = btrfs_alloc_delayed_extent_op();
8487                 if (!extent_op) {
8488                         ret = -ENOMEM;
8489                         goto out_free_buf;
8490                 }
8491                 if (key)
8492                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8493                 else
8494                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8495                 extent_op->flags_to_set = flags;
8496                 extent_op->update_key = skinny_metadata ? false : true;
8497                 extent_op->update_flags = true;
8498                 extent_op->is_data = false;
8499                 extent_op->level = level;
8500
8501                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8502                                    root_objectid, level, 0,
8503                                    BTRFS_ADD_DELAYED_EXTENT);
8504                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8505                                                  ins.offset, parent,
8506                                                  root_objectid, level,
8507                                                  BTRFS_ADD_DELAYED_EXTENT,
8508                                                  extent_op, NULL, NULL);
8509                 if (ret)
8510                         goto out_free_delayed;
8511         }
8512         return buf;
8513
8514 out_free_delayed:
8515         btrfs_free_delayed_extent_op(extent_op);
8516 out_free_buf:
8517         free_extent_buffer(buf);
8518 out_free_reserved:
8519         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8520 out_unuse:
8521         unuse_block_rsv(fs_info, block_rsv, blocksize);
8522         return ERR_PTR(ret);
8523 }
8524
8525 struct walk_control {
8526         u64 refs[BTRFS_MAX_LEVEL];
8527         u64 flags[BTRFS_MAX_LEVEL];
8528         struct btrfs_key update_progress;
8529         int stage;
8530         int level;
8531         int shared_level;
8532         int update_ref;
8533         int keep_locks;
8534         int reada_slot;
8535         int reada_count;
8536         int for_reloc;
8537 };
8538
8539 #define DROP_REFERENCE  1
8540 #define UPDATE_BACKREF  2
8541
8542 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8543                                      struct btrfs_root *root,
8544                                      struct walk_control *wc,
8545                                      struct btrfs_path *path)
8546 {
8547         struct btrfs_fs_info *fs_info = root->fs_info;
8548         u64 bytenr;
8549         u64 generation;
8550         u64 refs;
8551         u64 flags;
8552         u32 nritems;
8553         struct btrfs_key key;
8554         struct extent_buffer *eb;
8555         int ret;
8556         int slot;
8557         int nread = 0;
8558
8559         if (path->slots[wc->level] < wc->reada_slot) {
8560                 wc->reada_count = wc->reada_count * 2 / 3;
8561                 wc->reada_count = max(wc->reada_count, 2);
8562         } else {
8563                 wc->reada_count = wc->reada_count * 3 / 2;
8564                 wc->reada_count = min_t(int, wc->reada_count,
8565                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8566         }
8567
8568         eb = path->nodes[wc->level];
8569         nritems = btrfs_header_nritems(eb);
8570
8571         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8572                 if (nread >= wc->reada_count)
8573                         break;
8574
8575                 cond_resched();
8576                 bytenr = btrfs_node_blockptr(eb, slot);
8577                 generation = btrfs_node_ptr_generation(eb, slot);
8578
8579                 if (slot == path->slots[wc->level])
8580                         goto reada;
8581
8582                 if (wc->stage == UPDATE_BACKREF &&
8583                     generation <= root->root_key.offset)
8584                         continue;
8585
8586                 /* We don't lock the tree block, it's OK to be racy here */
8587                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8588                                                wc->level - 1, 1, &refs,
8589                                                &flags);
8590                 /* We don't care about errors in readahead. */
8591                 if (ret < 0)
8592                         continue;
8593                 BUG_ON(refs == 0);
8594
8595                 if (wc->stage == DROP_REFERENCE) {
8596                         if (refs == 1)
8597                                 goto reada;
8598
8599                         if (wc->level == 1 &&
8600                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8601                                 continue;
8602                         if (!wc->update_ref ||
8603                             generation <= root->root_key.offset)
8604                                 continue;
8605                         btrfs_node_key_to_cpu(eb, &key, slot);
8606                         ret = btrfs_comp_cpu_keys(&key,
8607                                                   &wc->update_progress);
8608                         if (ret < 0)
8609                                 continue;
8610                 } else {
8611                         if (wc->level == 1 &&
8612                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8613                                 continue;
8614                 }
8615 reada:
8616                 readahead_tree_block(fs_info, bytenr);
8617                 nread++;
8618         }
8619         wc->reada_slot = slot;
8620 }
8621
8622 /*
8623  * helper to process tree block while walking down the tree.
8624  *
8625  * when wc->stage == UPDATE_BACKREF, this function updates
8626  * back refs for pointers in the block.
8627  *
8628  * NOTE: return value 1 means we should stop walking down.
8629  */
8630 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8631                                    struct btrfs_root *root,
8632                                    struct btrfs_path *path,
8633                                    struct walk_control *wc, int lookup_info)
8634 {
8635         struct btrfs_fs_info *fs_info = root->fs_info;
8636         int level = wc->level;
8637         struct extent_buffer *eb = path->nodes[level];
8638         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8639         int ret;
8640
8641         if (wc->stage == UPDATE_BACKREF &&
8642             btrfs_header_owner(eb) != root->root_key.objectid)
8643                 return 1;
8644
8645         /*
8646          * when reference count of tree block is 1, it won't increase
8647          * again. once full backref flag is set, we never clear it.
8648          */
8649         if (lookup_info &&
8650             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8651              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8652                 BUG_ON(!path->locks[level]);
8653                 ret = btrfs_lookup_extent_info(trans, fs_info,
8654                                                eb->start, level, 1,
8655                                                &wc->refs[level],
8656                                                &wc->flags[level]);
8657                 BUG_ON(ret == -ENOMEM);
8658                 if (ret)
8659                         return ret;
8660                 BUG_ON(wc->refs[level] == 0);
8661         }
8662
8663         if (wc->stage == DROP_REFERENCE) {
8664                 if (wc->refs[level] > 1)
8665                         return 1;
8666
8667                 if (path->locks[level] && !wc->keep_locks) {
8668                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8669                         path->locks[level] = 0;
8670                 }
8671                 return 0;
8672         }
8673
8674         /* wc->stage == UPDATE_BACKREF */
8675         if (!(wc->flags[level] & flag)) {
8676                 BUG_ON(!path->locks[level]);
8677                 ret = btrfs_inc_ref(trans, root, eb, 1);
8678                 BUG_ON(ret); /* -ENOMEM */
8679                 ret = btrfs_dec_ref(trans, root, eb, 0);
8680                 BUG_ON(ret); /* -ENOMEM */
8681                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8682                                                   eb->len, flag,
8683                                                   btrfs_header_level(eb), 0);
8684                 BUG_ON(ret); /* -ENOMEM */
8685                 wc->flags[level] |= flag;
8686         }
8687
8688         /*
8689          * the block is shared by multiple trees, so it's not good to
8690          * keep the tree lock
8691          */
8692         if (path->locks[level] && level > 0) {
8693                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8694                 path->locks[level] = 0;
8695         }
8696         return 0;
8697 }
8698
8699 /*
8700  * helper to process tree block pointer.
8701  *
8702  * when wc->stage == DROP_REFERENCE, this function checks
8703  * reference count of the block pointed to. if the block
8704  * is shared and we need update back refs for the subtree
8705  * rooted at the block, this function changes wc->stage to
8706  * UPDATE_BACKREF. if the block is shared and there is no
8707  * need to update back, this function drops the reference
8708  * to the block.
8709  *
8710  * NOTE: return value 1 means we should stop walking down.
8711  */
8712 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8713                                  struct btrfs_root *root,
8714                                  struct btrfs_path *path,
8715                                  struct walk_control *wc, int *lookup_info)
8716 {
8717         struct btrfs_fs_info *fs_info = root->fs_info;
8718         u64 bytenr;
8719         u64 generation;
8720         u64 parent;
8721         u32 blocksize;
8722         struct btrfs_key key;
8723         struct btrfs_key first_key;
8724         struct extent_buffer *next;
8725         int level = wc->level;
8726         int reada = 0;
8727         int ret = 0;
8728         bool need_account = false;
8729
8730         generation = btrfs_node_ptr_generation(path->nodes[level],
8731                                                path->slots[level]);
8732         /*
8733          * if the lower level block was created before the snapshot
8734          * was created, we know there is no need to update back refs
8735          * for the subtree
8736          */
8737         if (wc->stage == UPDATE_BACKREF &&
8738             generation <= root->root_key.offset) {
8739                 *lookup_info = 1;
8740                 return 1;
8741         }
8742
8743         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8744         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8745                               path->slots[level]);
8746         blocksize = fs_info->nodesize;
8747
8748         next = find_extent_buffer(fs_info, bytenr);
8749         if (!next) {
8750                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8751                 if (IS_ERR(next))
8752                         return PTR_ERR(next);
8753
8754                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8755                                                level - 1);
8756                 reada = 1;
8757         }
8758         btrfs_tree_lock(next);
8759         btrfs_set_lock_blocking(next);
8760
8761         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8762                                        &wc->refs[level - 1],
8763                                        &wc->flags[level - 1]);
8764         if (ret < 0)
8765                 goto out_unlock;
8766
8767         if (unlikely(wc->refs[level - 1] == 0)) {
8768                 btrfs_err(fs_info, "Missing references.");
8769                 ret = -EIO;
8770                 goto out_unlock;
8771         }
8772         *lookup_info = 0;
8773
8774         if (wc->stage == DROP_REFERENCE) {
8775                 if (wc->refs[level - 1] > 1) {
8776                         need_account = true;
8777                         if (level == 1 &&
8778                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8779                                 goto skip;
8780
8781                         if (!wc->update_ref ||
8782                             generation <= root->root_key.offset)
8783                                 goto skip;
8784
8785                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8786                                               path->slots[level]);
8787                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8788                         if (ret < 0)
8789                                 goto skip;
8790
8791                         wc->stage = UPDATE_BACKREF;
8792                         wc->shared_level = level - 1;
8793                 }
8794         } else {
8795                 if (level == 1 &&
8796                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8797                         goto skip;
8798         }
8799
8800         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8801                 btrfs_tree_unlock(next);
8802                 free_extent_buffer(next);
8803                 next = NULL;
8804                 *lookup_info = 1;
8805         }
8806
8807         if (!next) {
8808                 if (reada && level == 1)
8809                         reada_walk_down(trans, root, wc, path);
8810                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8811                                        &first_key);
8812                 if (IS_ERR(next)) {
8813                         return PTR_ERR(next);
8814                 } else if (!extent_buffer_uptodate(next)) {
8815                         free_extent_buffer(next);
8816                         return -EIO;
8817                 }
8818                 btrfs_tree_lock(next);
8819                 btrfs_set_lock_blocking(next);
8820         }
8821
8822         level--;
8823         ASSERT(level == btrfs_header_level(next));
8824         if (level != btrfs_header_level(next)) {
8825                 btrfs_err(root->fs_info, "mismatched level");
8826                 ret = -EIO;
8827                 goto out_unlock;
8828         }
8829         path->nodes[level] = next;
8830         path->slots[level] = 0;
8831         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8832         wc->level = level;
8833         if (wc->level == 1)
8834                 wc->reada_slot = 0;
8835         return 0;
8836 skip:
8837         wc->refs[level - 1] = 0;
8838         wc->flags[level - 1] = 0;
8839         if (wc->stage == DROP_REFERENCE) {
8840                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8841                         parent = path->nodes[level]->start;
8842                 } else {
8843                         ASSERT(root->root_key.objectid ==
8844                                btrfs_header_owner(path->nodes[level]));
8845                         if (root->root_key.objectid !=
8846                             btrfs_header_owner(path->nodes[level])) {
8847                                 btrfs_err(root->fs_info,
8848                                                 "mismatched block owner");
8849                                 ret = -EIO;
8850                                 goto out_unlock;
8851                         }
8852                         parent = 0;
8853                 }
8854
8855                 if (need_account) {
8856                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8857                                                          generation, level - 1);
8858                         if (ret) {
8859                                 btrfs_err_rl(fs_info,
8860                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8861                                              ret);
8862                         }
8863                 }
8864                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8865                                         parent, root->root_key.objectid,
8866                                         level - 1, 0);
8867                 if (ret)
8868                         goto out_unlock;
8869         }
8870
8871         *lookup_info = 1;
8872         ret = 1;
8873
8874 out_unlock:
8875         btrfs_tree_unlock(next);
8876         free_extent_buffer(next);
8877
8878         return ret;
8879 }
8880
8881 /*
8882  * helper to process tree block while walking up the tree.
8883  *
8884  * when wc->stage == DROP_REFERENCE, this function drops
8885  * reference count on the block.
8886  *
8887  * when wc->stage == UPDATE_BACKREF, this function changes
8888  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8889  * to UPDATE_BACKREF previously while processing the block.
8890  *
8891  * NOTE: return value 1 means we should stop walking up.
8892  */
8893 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8894                                  struct btrfs_root *root,
8895                                  struct btrfs_path *path,
8896                                  struct walk_control *wc)
8897 {
8898         struct btrfs_fs_info *fs_info = root->fs_info;
8899         int ret;
8900         int level = wc->level;
8901         struct extent_buffer *eb = path->nodes[level];
8902         u64 parent = 0;
8903
8904         if (wc->stage == UPDATE_BACKREF) {
8905                 BUG_ON(wc->shared_level < level);
8906                 if (level < wc->shared_level)
8907                         goto out;
8908
8909                 ret = find_next_key(path, level + 1, &wc->update_progress);
8910                 if (ret > 0)
8911                         wc->update_ref = 0;
8912
8913                 wc->stage = DROP_REFERENCE;
8914                 wc->shared_level = -1;
8915                 path->slots[level] = 0;
8916
8917                 /*
8918                  * check reference count again if the block isn't locked.
8919                  * we should start walking down the tree again if reference
8920                  * count is one.
8921                  */
8922                 if (!path->locks[level]) {
8923                         BUG_ON(level == 0);
8924                         btrfs_tree_lock(eb);
8925                         btrfs_set_lock_blocking(eb);
8926                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8927
8928                         ret = btrfs_lookup_extent_info(trans, fs_info,
8929                                                        eb->start, level, 1,
8930                                                        &wc->refs[level],
8931                                                        &wc->flags[level]);
8932                         if (ret < 0) {
8933                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8934                                 path->locks[level] = 0;
8935                                 return ret;
8936                         }
8937                         BUG_ON(wc->refs[level] == 0);
8938                         if (wc->refs[level] == 1) {
8939                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8940                                 path->locks[level] = 0;
8941                                 return 1;
8942                         }
8943                 }
8944         }
8945
8946         /* wc->stage == DROP_REFERENCE */
8947         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8948
8949         if (wc->refs[level] == 1) {
8950                 if (level == 0) {
8951                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8952                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8953                         else
8954                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8955                         BUG_ON(ret); /* -ENOMEM */
8956                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8957                         if (ret) {
8958                                 btrfs_err_rl(fs_info,
8959                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8960                                              ret);
8961                         }
8962                 }
8963                 /* make block locked assertion in clean_tree_block happy */
8964                 if (!path->locks[level] &&
8965                     btrfs_header_generation(eb) == trans->transid) {
8966                         btrfs_tree_lock(eb);
8967                         btrfs_set_lock_blocking(eb);
8968                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8969                 }
8970                 clean_tree_block(fs_info, eb);
8971         }
8972
8973         if (eb == root->node) {
8974                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8975                         parent = eb->start;
8976                 else
8977                         BUG_ON(root->root_key.objectid !=
8978                                btrfs_header_owner(eb));
8979         } else {
8980                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8981                         parent = path->nodes[level + 1]->start;
8982                 else
8983                         BUG_ON(root->root_key.objectid !=
8984                                btrfs_header_owner(path->nodes[level + 1]));
8985         }
8986
8987         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8988 out:
8989         wc->refs[level] = 0;
8990         wc->flags[level] = 0;
8991         return 0;
8992 }
8993
8994 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8995                                    struct btrfs_root *root,
8996                                    struct btrfs_path *path,
8997                                    struct walk_control *wc)
8998 {
8999         int level = wc->level;
9000         int lookup_info = 1;
9001         int ret;
9002
9003         while (level >= 0) {
9004                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9005                 if (ret > 0)
9006                         break;
9007
9008                 if (level == 0)
9009                         break;
9010
9011                 if (path->slots[level] >=
9012                     btrfs_header_nritems(path->nodes[level]))
9013                         break;
9014
9015                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9016                 if (ret > 0) {
9017                         path->slots[level]++;
9018                         continue;
9019                 } else if (ret < 0)
9020                         return ret;
9021                 level = wc->level;
9022         }
9023         return 0;
9024 }
9025
9026 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9027                                  struct btrfs_root *root,
9028                                  struct btrfs_path *path,
9029                                  struct walk_control *wc, int max_level)
9030 {
9031         int level = wc->level;
9032         int ret;
9033
9034         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9035         while (level < max_level && path->nodes[level]) {
9036                 wc->level = level;
9037                 if (path->slots[level] + 1 <
9038                     btrfs_header_nritems(path->nodes[level])) {
9039                         path->slots[level]++;
9040                         return 0;
9041                 } else {
9042                         ret = walk_up_proc(trans, root, path, wc);
9043                         if (ret > 0)
9044                                 return 0;
9045
9046                         if (path->locks[level]) {
9047                                 btrfs_tree_unlock_rw(path->nodes[level],
9048                                                      path->locks[level]);
9049                                 path->locks[level] = 0;
9050                         }
9051                         free_extent_buffer(path->nodes[level]);
9052                         path->nodes[level] = NULL;
9053                         level++;
9054                 }
9055         }
9056         return 1;
9057 }
9058
9059 /*
9060  * drop a subvolume tree.
9061  *
9062  * this function traverses the tree freeing any blocks that only
9063  * referenced by the tree.
9064  *
9065  * when a shared tree block is found. this function decreases its
9066  * reference count by one. if update_ref is true, this function
9067  * also make sure backrefs for the shared block and all lower level
9068  * blocks are properly updated.
9069  *
9070  * If called with for_reloc == 0, may exit early with -EAGAIN
9071  */
9072 int btrfs_drop_snapshot(struct btrfs_root *root,
9073                          struct btrfs_block_rsv *block_rsv, int update_ref,
9074                          int for_reloc)
9075 {
9076         struct btrfs_fs_info *fs_info = root->fs_info;
9077         struct btrfs_path *path;
9078         struct btrfs_trans_handle *trans;
9079         struct btrfs_root *tree_root = fs_info->tree_root;
9080         struct btrfs_root_item *root_item = &root->root_item;
9081         struct walk_control *wc;
9082         struct btrfs_key key;
9083         int err = 0;
9084         int ret;
9085         int level;
9086         bool root_dropped = false;
9087
9088         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9089
9090         path = btrfs_alloc_path();
9091         if (!path) {
9092                 err = -ENOMEM;
9093                 goto out;
9094         }
9095
9096         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9097         if (!wc) {
9098                 btrfs_free_path(path);
9099                 err = -ENOMEM;
9100                 goto out;
9101         }
9102
9103         trans = btrfs_start_transaction(tree_root, 0);
9104         if (IS_ERR(trans)) {
9105                 err = PTR_ERR(trans);
9106                 goto out_free;
9107         }
9108
9109         if (block_rsv)
9110                 trans->block_rsv = block_rsv;
9111
9112         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9113                 level = btrfs_header_level(root->node);
9114                 path->nodes[level] = btrfs_lock_root_node(root);
9115                 btrfs_set_lock_blocking(path->nodes[level]);
9116                 path->slots[level] = 0;
9117                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9118                 memset(&wc->update_progress, 0,
9119                        sizeof(wc->update_progress));
9120         } else {
9121                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9122                 memcpy(&wc->update_progress, &key,
9123                        sizeof(wc->update_progress));
9124
9125                 level = root_item->drop_level;
9126                 BUG_ON(level == 0);
9127                 path->lowest_level = level;
9128                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9129                 path->lowest_level = 0;
9130                 if (ret < 0) {
9131                         err = ret;
9132                         goto out_end_trans;
9133                 }
9134                 WARN_ON(ret > 0);
9135
9136                 /*
9137                  * unlock our path, this is safe because only this
9138                  * function is allowed to delete this snapshot
9139                  */
9140                 btrfs_unlock_up_safe(path, 0);
9141
9142                 level = btrfs_header_level(root->node);
9143                 while (1) {
9144                         btrfs_tree_lock(path->nodes[level]);
9145                         btrfs_set_lock_blocking(path->nodes[level]);
9146                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9147
9148                         ret = btrfs_lookup_extent_info(trans, fs_info,
9149                                                 path->nodes[level]->start,
9150                                                 level, 1, &wc->refs[level],
9151                                                 &wc->flags[level]);
9152                         if (ret < 0) {
9153                                 err = ret;
9154                                 goto out_end_trans;
9155                         }
9156                         BUG_ON(wc->refs[level] == 0);
9157
9158                         if (level == root_item->drop_level)
9159                                 break;
9160
9161                         btrfs_tree_unlock(path->nodes[level]);
9162                         path->locks[level] = 0;
9163                         WARN_ON(wc->refs[level] != 1);
9164                         level--;
9165                 }
9166         }
9167
9168         wc->level = level;
9169         wc->shared_level = -1;
9170         wc->stage = DROP_REFERENCE;
9171         wc->update_ref = update_ref;
9172         wc->keep_locks = 0;
9173         wc->for_reloc = for_reloc;
9174         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9175
9176         while (1) {
9177
9178                 ret = walk_down_tree(trans, root, path, wc);
9179                 if (ret < 0) {
9180                         err = ret;
9181                         break;
9182                 }
9183
9184                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9185                 if (ret < 0) {
9186                         err = ret;
9187                         break;
9188                 }
9189
9190                 if (ret > 0) {
9191                         BUG_ON(wc->stage != DROP_REFERENCE);
9192                         break;
9193                 }
9194
9195                 if (wc->stage == DROP_REFERENCE) {
9196                         level = wc->level;
9197                         btrfs_node_key(path->nodes[level],
9198                                        &root_item->drop_progress,
9199                                        path->slots[level]);
9200                         root_item->drop_level = level;
9201                 }
9202
9203                 BUG_ON(wc->level == 0);
9204                 if (btrfs_should_end_transaction(trans) ||
9205                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9206                         ret = btrfs_update_root(trans, tree_root,
9207                                                 &root->root_key,
9208                                                 root_item);
9209                         if (ret) {
9210                                 btrfs_abort_transaction(trans, ret);
9211                                 err = ret;
9212                                 goto out_end_trans;
9213                         }
9214
9215                         btrfs_end_transaction_throttle(trans);
9216                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9217                                 btrfs_debug(fs_info,
9218                                             "drop snapshot early exit");
9219                                 err = -EAGAIN;
9220                                 goto out_free;
9221                         }
9222
9223                         trans = btrfs_start_transaction(tree_root, 0);
9224                         if (IS_ERR(trans)) {
9225                                 err = PTR_ERR(trans);
9226                                 goto out_free;
9227                         }
9228                         if (block_rsv)
9229                                 trans->block_rsv = block_rsv;
9230                 }
9231         }
9232         btrfs_release_path(path);
9233         if (err)
9234                 goto out_end_trans;
9235
9236         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9237         if (ret) {
9238                 btrfs_abort_transaction(trans, ret);
9239                 err = ret;
9240                 goto out_end_trans;
9241         }
9242
9243         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9244                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9245                                       NULL, NULL);
9246                 if (ret < 0) {
9247                         btrfs_abort_transaction(trans, ret);
9248                         err = ret;
9249                         goto out_end_trans;
9250                 } else if (ret > 0) {
9251                         /* if we fail to delete the orphan item this time
9252                          * around, it'll get picked up the next time.
9253                          *
9254                          * The most common failure here is just -ENOENT.
9255                          */
9256                         btrfs_del_orphan_item(trans, tree_root,
9257                                               root->root_key.objectid);
9258                 }
9259         }
9260
9261         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9262                 btrfs_add_dropped_root(trans, root);
9263         } else {
9264                 free_extent_buffer(root->node);
9265                 free_extent_buffer(root->commit_root);
9266                 btrfs_put_fs_root(root);
9267         }
9268         root_dropped = true;
9269 out_end_trans:
9270         btrfs_end_transaction_throttle(trans);
9271 out_free:
9272         kfree(wc);
9273         btrfs_free_path(path);
9274 out:
9275         /*
9276          * So if we need to stop dropping the snapshot for whatever reason we
9277          * need to make sure to add it back to the dead root list so that we
9278          * keep trying to do the work later.  This also cleans up roots if we
9279          * don't have it in the radix (like when we recover after a power fail
9280          * or unmount) so we don't leak memory.
9281          */
9282         if (!for_reloc && !root_dropped)
9283                 btrfs_add_dead_root(root);
9284         if (err && err != -EAGAIN)
9285                 btrfs_handle_fs_error(fs_info, err, NULL);
9286         return err;
9287 }
9288
9289 /*
9290  * drop subtree rooted at tree block 'node'.
9291  *
9292  * NOTE: this function will unlock and release tree block 'node'
9293  * only used by relocation code
9294  */
9295 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9296                         struct btrfs_root *root,
9297                         struct extent_buffer *node,
9298                         struct extent_buffer *parent)
9299 {
9300         struct btrfs_fs_info *fs_info = root->fs_info;
9301         struct btrfs_path *path;
9302         struct walk_control *wc;
9303         int level;
9304         int parent_level;
9305         int ret = 0;
9306         int wret;
9307
9308         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9309
9310         path = btrfs_alloc_path();
9311         if (!path)
9312                 return -ENOMEM;
9313
9314         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9315         if (!wc) {
9316                 btrfs_free_path(path);
9317                 return -ENOMEM;
9318         }
9319
9320         btrfs_assert_tree_locked(parent);
9321         parent_level = btrfs_header_level(parent);
9322         extent_buffer_get(parent);
9323         path->nodes[parent_level] = parent;
9324         path->slots[parent_level] = btrfs_header_nritems(parent);
9325
9326         btrfs_assert_tree_locked(node);
9327         level = btrfs_header_level(node);
9328         path->nodes[level] = node;
9329         path->slots[level] = 0;
9330         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9331
9332         wc->refs[parent_level] = 1;
9333         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9334         wc->level = level;
9335         wc->shared_level = -1;
9336         wc->stage = DROP_REFERENCE;
9337         wc->update_ref = 0;
9338         wc->keep_locks = 1;
9339         wc->for_reloc = 1;
9340         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9341
9342         while (1) {
9343                 wret = walk_down_tree(trans, root, path, wc);
9344                 if (wret < 0) {
9345                         ret = wret;
9346                         break;
9347                 }
9348
9349                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9350                 if (wret < 0)
9351                         ret = wret;
9352                 if (wret != 0)
9353                         break;
9354         }
9355
9356         kfree(wc);
9357         btrfs_free_path(path);
9358         return ret;
9359 }
9360
9361 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9362 {
9363         u64 num_devices;
9364         u64 stripped;
9365
9366         /*
9367          * if restripe for this chunk_type is on pick target profile and
9368          * return, otherwise do the usual balance
9369          */
9370         stripped = get_restripe_target(fs_info, flags);
9371         if (stripped)
9372                 return extended_to_chunk(stripped);
9373
9374         num_devices = fs_info->fs_devices->rw_devices;
9375
9376         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9377                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9378                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9379
9380         if (num_devices == 1) {
9381                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9382                 stripped = flags & ~stripped;
9383
9384                 /* turn raid0 into single device chunks */
9385                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9386                         return stripped;
9387
9388                 /* turn mirroring into duplication */
9389                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9390                              BTRFS_BLOCK_GROUP_RAID10))
9391                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9392         } else {
9393                 /* they already had raid on here, just return */
9394                 if (flags & stripped)
9395                         return flags;
9396
9397                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9398                 stripped = flags & ~stripped;
9399
9400                 /* switch duplicated blocks with raid1 */
9401                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9402                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9403
9404                 /* this is drive concat, leave it alone */
9405         }
9406
9407         return flags;
9408 }
9409
9410 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9411 {
9412         struct btrfs_space_info *sinfo = cache->space_info;
9413         u64 num_bytes;
9414         u64 min_allocable_bytes;
9415         int ret = -ENOSPC;
9416
9417         /*
9418          * We need some metadata space and system metadata space for
9419          * allocating chunks in some corner cases until we force to set
9420          * it to be readonly.
9421          */
9422         if ((sinfo->flags &
9423              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9424             !force)
9425                 min_allocable_bytes = SZ_1M;
9426         else
9427                 min_allocable_bytes = 0;
9428
9429         spin_lock(&sinfo->lock);
9430         spin_lock(&cache->lock);
9431
9432         if (cache->ro) {
9433                 cache->ro++;
9434                 ret = 0;
9435                 goto out;
9436         }
9437
9438         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9439                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9440
9441         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9442             min_allocable_bytes <= sinfo->total_bytes) {
9443                 sinfo->bytes_readonly += num_bytes;
9444                 cache->ro++;
9445                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9446                 ret = 0;
9447         }
9448 out:
9449         spin_unlock(&cache->lock);
9450         spin_unlock(&sinfo->lock);
9451         return ret;
9452 }
9453
9454 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9455                              struct btrfs_block_group_cache *cache)
9456
9457 {
9458         struct btrfs_trans_handle *trans;
9459         u64 alloc_flags;
9460         int ret;
9461
9462 again:
9463         trans = btrfs_join_transaction(fs_info->extent_root);
9464         if (IS_ERR(trans))
9465                 return PTR_ERR(trans);
9466
9467         /*
9468          * we're not allowed to set block groups readonly after the dirty
9469          * block groups cache has started writing.  If it already started,
9470          * back off and let this transaction commit
9471          */
9472         mutex_lock(&fs_info->ro_block_group_mutex);
9473         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9474                 u64 transid = trans->transid;
9475
9476                 mutex_unlock(&fs_info->ro_block_group_mutex);
9477                 btrfs_end_transaction(trans);
9478
9479                 ret = btrfs_wait_for_commit(fs_info, transid);
9480                 if (ret)
9481                         return ret;
9482                 goto again;
9483         }
9484
9485         /*
9486          * if we are changing raid levels, try to allocate a corresponding
9487          * block group with the new raid level.
9488          */
9489         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9490         if (alloc_flags != cache->flags) {
9491                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9492                                      CHUNK_ALLOC_FORCE);
9493                 /*
9494                  * ENOSPC is allowed here, we may have enough space
9495                  * already allocated at the new raid level to
9496                  * carry on
9497                  */
9498                 if (ret == -ENOSPC)
9499                         ret = 0;
9500                 if (ret < 0)
9501                         goto out;
9502         }
9503
9504         ret = inc_block_group_ro(cache, 0);
9505         if (!ret)
9506                 goto out;
9507         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9508         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9509                              CHUNK_ALLOC_FORCE);
9510         if (ret < 0)
9511                 goto out;
9512         ret = inc_block_group_ro(cache, 0);
9513 out:
9514         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9515                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9516                 mutex_lock(&fs_info->chunk_mutex);
9517                 check_system_chunk(trans, fs_info, alloc_flags);
9518                 mutex_unlock(&fs_info->chunk_mutex);
9519         }
9520         mutex_unlock(&fs_info->ro_block_group_mutex);
9521
9522         btrfs_end_transaction(trans);
9523         return ret;
9524 }
9525
9526 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9527                             struct btrfs_fs_info *fs_info, u64 type)
9528 {
9529         u64 alloc_flags = get_alloc_profile(fs_info, type);
9530
9531         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9532 }
9533
9534 /*
9535  * helper to account the unused space of all the readonly block group in the
9536  * space_info. takes mirrors into account.
9537  */
9538 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9539 {
9540         struct btrfs_block_group_cache *block_group;
9541         u64 free_bytes = 0;
9542         int factor;
9543
9544         /* It's df, we don't care if it's racy */
9545         if (list_empty(&sinfo->ro_bgs))
9546                 return 0;
9547
9548         spin_lock(&sinfo->lock);
9549         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9550                 spin_lock(&block_group->lock);
9551
9552                 if (!block_group->ro) {
9553                         spin_unlock(&block_group->lock);
9554                         continue;
9555                 }
9556
9557                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9558                                           BTRFS_BLOCK_GROUP_RAID10 |
9559                                           BTRFS_BLOCK_GROUP_DUP))
9560                         factor = 2;
9561                 else
9562                         factor = 1;
9563
9564                 free_bytes += (block_group->key.offset -
9565                                btrfs_block_group_used(&block_group->item)) *
9566                                factor;
9567
9568                 spin_unlock(&block_group->lock);
9569         }
9570         spin_unlock(&sinfo->lock);
9571
9572         return free_bytes;
9573 }
9574
9575 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9576 {
9577         struct btrfs_space_info *sinfo = cache->space_info;
9578         u64 num_bytes;
9579
9580         BUG_ON(!cache->ro);
9581
9582         spin_lock(&sinfo->lock);
9583         spin_lock(&cache->lock);
9584         if (!--cache->ro) {
9585                 num_bytes = cache->key.offset - cache->reserved -
9586                             cache->pinned - cache->bytes_super -
9587                             btrfs_block_group_used(&cache->item);
9588                 sinfo->bytes_readonly -= num_bytes;
9589                 list_del_init(&cache->ro_list);
9590         }
9591         spin_unlock(&cache->lock);
9592         spin_unlock(&sinfo->lock);
9593 }
9594
9595 /*
9596  * checks to see if its even possible to relocate this block group.
9597  *
9598  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9599  * ok to go ahead and try.
9600  */
9601 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9602 {
9603         struct btrfs_root *root = fs_info->extent_root;
9604         struct btrfs_block_group_cache *block_group;
9605         struct btrfs_space_info *space_info;
9606         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9607         struct btrfs_device *device;
9608         struct btrfs_trans_handle *trans;
9609         u64 min_free;
9610         u64 dev_min = 1;
9611         u64 dev_nr = 0;
9612         u64 target;
9613         int debug;
9614         int index;
9615         int full = 0;
9616         int ret = 0;
9617
9618         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9619
9620         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9621
9622         /* odd, couldn't find the block group, leave it alone */
9623         if (!block_group) {
9624                 if (debug)
9625                         btrfs_warn(fs_info,
9626                                    "can't find block group for bytenr %llu",
9627                                    bytenr);
9628                 return -1;
9629         }
9630
9631         min_free = btrfs_block_group_used(&block_group->item);
9632
9633         /* no bytes used, we're good */
9634         if (!min_free)
9635                 goto out;
9636
9637         space_info = block_group->space_info;
9638         spin_lock(&space_info->lock);
9639
9640         full = space_info->full;
9641
9642         /*
9643          * if this is the last block group we have in this space, we can't
9644          * relocate it unless we're able to allocate a new chunk below.
9645          *
9646          * Otherwise, we need to make sure we have room in the space to handle
9647          * all of the extents from this block group.  If we can, we're good
9648          */
9649         if ((space_info->total_bytes != block_group->key.offset) &&
9650             (btrfs_space_info_used(space_info, false) + min_free <
9651              space_info->total_bytes)) {
9652                 spin_unlock(&space_info->lock);
9653                 goto out;
9654         }
9655         spin_unlock(&space_info->lock);
9656
9657         /*
9658          * ok we don't have enough space, but maybe we have free space on our
9659          * devices to allocate new chunks for relocation, so loop through our
9660          * alloc devices and guess if we have enough space.  if this block
9661          * group is going to be restriped, run checks against the target
9662          * profile instead of the current one.
9663          */
9664         ret = -1;
9665
9666         /*
9667          * index:
9668          *      0: raid10
9669          *      1: raid1
9670          *      2: dup
9671          *      3: raid0
9672          *      4: single
9673          */
9674         target = get_restripe_target(fs_info, block_group->flags);
9675         if (target) {
9676                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9677         } else {
9678                 /*
9679                  * this is just a balance, so if we were marked as full
9680                  * we know there is no space for a new chunk
9681                  */
9682                 if (full) {
9683                         if (debug)
9684                                 btrfs_warn(fs_info,
9685                                            "no space to alloc new chunk for block group %llu",
9686                                            block_group->key.objectid);
9687                         goto out;
9688                 }
9689
9690                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9691         }
9692
9693         if (index == BTRFS_RAID_RAID10) {
9694                 dev_min = 4;
9695                 /* Divide by 2 */
9696                 min_free >>= 1;
9697         } else if (index == BTRFS_RAID_RAID1) {
9698                 dev_min = 2;
9699         } else if (index == BTRFS_RAID_DUP) {
9700                 /* Multiply by 2 */
9701                 min_free <<= 1;
9702         } else if (index == BTRFS_RAID_RAID0) {
9703                 dev_min = fs_devices->rw_devices;
9704                 min_free = div64_u64(min_free, dev_min);
9705         }
9706
9707         /* We need to do this so that we can look at pending chunks */
9708         trans = btrfs_join_transaction(root);
9709         if (IS_ERR(trans)) {
9710                 ret = PTR_ERR(trans);
9711                 goto out;
9712         }
9713
9714         mutex_lock(&fs_info->chunk_mutex);
9715         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9716                 u64 dev_offset;
9717
9718                 /*
9719                  * check to make sure we can actually find a chunk with enough
9720                  * space to fit our block group in.
9721                  */
9722                 if (device->total_bytes > device->bytes_used + min_free &&
9723                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9724                         ret = find_free_dev_extent(trans, device, min_free,
9725                                                    &dev_offset, NULL);
9726                         if (!ret)
9727                                 dev_nr++;
9728
9729                         if (dev_nr >= dev_min)
9730                                 break;
9731
9732                         ret = -1;
9733                 }
9734         }
9735         if (debug && ret == -1)
9736                 btrfs_warn(fs_info,
9737                            "no space to allocate a new chunk for block group %llu",
9738                            block_group->key.objectid);
9739         mutex_unlock(&fs_info->chunk_mutex);
9740         btrfs_end_transaction(trans);
9741 out:
9742         btrfs_put_block_group(block_group);
9743         return ret;
9744 }
9745
9746 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9747                                   struct btrfs_path *path,
9748                                   struct btrfs_key *key)
9749 {
9750         struct btrfs_root *root = fs_info->extent_root;
9751         int ret = 0;
9752         struct btrfs_key found_key;
9753         struct extent_buffer *leaf;
9754         int slot;
9755
9756         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9757         if (ret < 0)
9758                 goto out;
9759
9760         while (1) {
9761                 slot = path->slots[0];
9762                 leaf = path->nodes[0];
9763                 if (slot >= btrfs_header_nritems(leaf)) {
9764                         ret = btrfs_next_leaf(root, path);
9765                         if (ret == 0)
9766                                 continue;
9767                         if (ret < 0)
9768                                 goto out;
9769                         break;
9770                 }
9771                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9772
9773                 if (found_key.objectid >= key->objectid &&
9774                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9775                         struct extent_map_tree *em_tree;
9776                         struct extent_map *em;
9777
9778                         em_tree = &root->fs_info->mapping_tree.map_tree;
9779                         read_lock(&em_tree->lock);
9780                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9781                                                    found_key.offset);
9782                         read_unlock(&em_tree->lock);
9783                         if (!em) {
9784                                 btrfs_err(fs_info,
9785                         "logical %llu len %llu found bg but no related chunk",
9786                                           found_key.objectid, found_key.offset);
9787                                 ret = -ENOENT;
9788                         } else {
9789                                 ret = 0;
9790                         }
9791                         free_extent_map(em);
9792                         goto out;
9793                 }
9794                 path->slots[0]++;
9795         }
9796 out:
9797         return ret;
9798 }
9799
9800 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9801 {
9802         struct btrfs_block_group_cache *block_group;
9803         u64 last = 0;
9804
9805         while (1) {
9806                 struct inode *inode;
9807
9808                 block_group = btrfs_lookup_first_block_group(info, last);
9809                 while (block_group) {
9810                         spin_lock(&block_group->lock);
9811                         if (block_group->iref)
9812                                 break;
9813                         spin_unlock(&block_group->lock);
9814                         block_group = next_block_group(info, block_group);
9815                 }
9816                 if (!block_group) {
9817                         if (last == 0)
9818                                 break;
9819                         last = 0;
9820                         continue;
9821                 }
9822
9823                 inode = block_group->inode;
9824                 block_group->iref = 0;
9825                 block_group->inode = NULL;
9826                 spin_unlock(&block_group->lock);
9827                 ASSERT(block_group->io_ctl.inode == NULL);
9828                 iput(inode);
9829                 last = block_group->key.objectid + block_group->key.offset;
9830                 btrfs_put_block_group(block_group);
9831         }
9832 }
9833
9834 /*
9835  * Must be called only after stopping all workers, since we could have block
9836  * group caching kthreads running, and therefore they could race with us if we
9837  * freed the block groups before stopping them.
9838  */
9839 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9840 {
9841         struct btrfs_block_group_cache *block_group;
9842         struct btrfs_space_info *space_info;
9843         struct btrfs_caching_control *caching_ctl;
9844         struct rb_node *n;
9845
9846         down_write(&info->commit_root_sem);
9847         while (!list_empty(&info->caching_block_groups)) {
9848                 caching_ctl = list_entry(info->caching_block_groups.next,
9849                                          struct btrfs_caching_control, list);
9850                 list_del(&caching_ctl->list);
9851                 put_caching_control(caching_ctl);
9852         }
9853         up_write(&info->commit_root_sem);
9854
9855         spin_lock(&info->unused_bgs_lock);
9856         while (!list_empty(&info->unused_bgs)) {
9857                 block_group = list_first_entry(&info->unused_bgs,
9858                                                struct btrfs_block_group_cache,
9859                                                bg_list);
9860                 list_del_init(&block_group->bg_list);
9861                 btrfs_put_block_group(block_group);
9862         }
9863         spin_unlock(&info->unused_bgs_lock);
9864
9865         spin_lock(&info->block_group_cache_lock);
9866         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9867                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9868                                        cache_node);
9869                 rb_erase(&block_group->cache_node,
9870                          &info->block_group_cache_tree);
9871                 RB_CLEAR_NODE(&block_group->cache_node);
9872                 spin_unlock(&info->block_group_cache_lock);
9873
9874                 down_write(&block_group->space_info->groups_sem);
9875                 list_del(&block_group->list);
9876                 up_write(&block_group->space_info->groups_sem);
9877
9878                 /*
9879                  * We haven't cached this block group, which means we could
9880                  * possibly have excluded extents on this block group.
9881                  */
9882                 if (block_group->cached == BTRFS_CACHE_NO ||
9883                     block_group->cached == BTRFS_CACHE_ERROR)
9884                         free_excluded_extents(info, block_group);
9885
9886                 btrfs_remove_free_space_cache(block_group);
9887                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9888                 ASSERT(list_empty(&block_group->dirty_list));
9889                 ASSERT(list_empty(&block_group->io_list));
9890                 ASSERT(list_empty(&block_group->bg_list));
9891                 ASSERT(atomic_read(&block_group->count) == 1);
9892                 btrfs_put_block_group(block_group);
9893
9894                 spin_lock(&info->block_group_cache_lock);
9895         }
9896         spin_unlock(&info->block_group_cache_lock);
9897
9898         /* now that all the block groups are freed, go through and
9899          * free all the space_info structs.  This is only called during
9900          * the final stages of unmount, and so we know nobody is
9901          * using them.  We call synchronize_rcu() once before we start,
9902          * just to be on the safe side.
9903          */
9904         synchronize_rcu();
9905
9906         release_global_block_rsv(info);
9907
9908         while (!list_empty(&info->space_info)) {
9909                 int i;
9910
9911                 space_info = list_entry(info->space_info.next,
9912                                         struct btrfs_space_info,
9913                                         list);
9914
9915                 /*
9916                  * Do not hide this behind enospc_debug, this is actually
9917                  * important and indicates a real bug if this happens.
9918                  */
9919                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9920                             space_info->bytes_reserved > 0 ||
9921                             space_info->bytes_may_use > 0))
9922                         dump_space_info(info, space_info, 0, 0);
9923                 list_del(&space_info->list);
9924                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9925                         struct kobject *kobj;
9926                         kobj = space_info->block_group_kobjs[i];
9927                         space_info->block_group_kobjs[i] = NULL;
9928                         if (kobj) {
9929                                 kobject_del(kobj);
9930                                 kobject_put(kobj);
9931                         }
9932                 }
9933                 kobject_del(&space_info->kobj);
9934                 kobject_put(&space_info->kobj);
9935         }
9936         return 0;
9937 }
9938
9939 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9940 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9941 {
9942         struct btrfs_space_info *space_info;
9943         struct raid_kobject *rkobj;
9944         LIST_HEAD(list);
9945         int index;
9946         int ret = 0;
9947
9948         spin_lock(&fs_info->pending_raid_kobjs_lock);
9949         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9950         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9951
9952         list_for_each_entry(rkobj, &list, list) {
9953                 space_info = __find_space_info(fs_info, rkobj->flags);
9954                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9955
9956                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9957                                   "%s", get_raid_name(index));
9958                 if (ret) {
9959                         kobject_put(&rkobj->kobj);
9960                         break;
9961                 }
9962         }
9963         if (ret)
9964                 btrfs_warn(fs_info,
9965                            "failed to add kobject for block cache, ignoring");
9966 }
9967
9968 static void link_block_group(struct btrfs_block_group_cache *cache)
9969 {
9970         struct btrfs_space_info *space_info = cache->space_info;
9971         struct btrfs_fs_info *fs_info = cache->fs_info;
9972         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9973         bool first = false;
9974
9975         down_write(&space_info->groups_sem);
9976         if (list_empty(&space_info->block_groups[index]))
9977                 first = true;
9978         list_add_tail(&cache->list, &space_info->block_groups[index]);
9979         up_write(&space_info->groups_sem);
9980
9981         if (first) {
9982                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9983                 if (!rkobj) {
9984                         btrfs_warn(cache->fs_info,
9985                                 "couldn't alloc memory for raid level kobject");
9986                         return;
9987                 }
9988                 rkobj->flags = cache->flags;
9989                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9990
9991                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9992                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9993                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9994                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9995         }
9996 }
9997
9998 static struct btrfs_block_group_cache *
9999 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10000                                u64 start, u64 size)
10001 {
10002         struct btrfs_block_group_cache *cache;
10003
10004         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10005         if (!cache)
10006                 return NULL;
10007
10008         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10009                                         GFP_NOFS);
10010         if (!cache->free_space_ctl) {
10011                 kfree(cache);
10012                 return NULL;
10013         }
10014
10015         cache->key.objectid = start;
10016         cache->key.offset = size;
10017         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10018
10019         cache->fs_info = fs_info;
10020         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10021         set_free_space_tree_thresholds(cache);
10022
10023         atomic_set(&cache->count, 1);
10024         spin_lock_init(&cache->lock);
10025         init_rwsem(&cache->data_rwsem);
10026         INIT_LIST_HEAD(&cache->list);
10027         INIT_LIST_HEAD(&cache->cluster_list);
10028         INIT_LIST_HEAD(&cache->bg_list);
10029         INIT_LIST_HEAD(&cache->ro_list);
10030         INIT_LIST_HEAD(&cache->dirty_list);
10031         INIT_LIST_HEAD(&cache->io_list);
10032         btrfs_init_free_space_ctl(cache);
10033         atomic_set(&cache->trimming, 0);
10034         mutex_init(&cache->free_space_lock);
10035         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10036
10037         return cache;
10038 }
10039
10040 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10041 {
10042         struct btrfs_path *path;
10043         int ret;
10044         struct btrfs_block_group_cache *cache;
10045         struct btrfs_space_info *space_info;
10046         struct btrfs_key key;
10047         struct btrfs_key found_key;
10048         struct extent_buffer *leaf;
10049         int need_clear = 0;
10050         u64 cache_gen;
10051         u64 feature;
10052         int mixed;
10053
10054         feature = btrfs_super_incompat_flags(info->super_copy);
10055         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10056
10057         key.objectid = 0;
10058         key.offset = 0;
10059         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10060         path = btrfs_alloc_path();
10061         if (!path)
10062                 return -ENOMEM;
10063         path->reada = READA_FORWARD;
10064
10065         cache_gen = btrfs_super_cache_generation(info->super_copy);
10066         if (btrfs_test_opt(info, SPACE_CACHE) &&
10067             btrfs_super_generation(info->super_copy) != cache_gen)
10068                 need_clear = 1;
10069         if (btrfs_test_opt(info, CLEAR_CACHE))
10070                 need_clear = 1;
10071
10072         while (1) {
10073                 ret = find_first_block_group(info, path, &key);
10074                 if (ret > 0)
10075                         break;
10076                 if (ret != 0)
10077                         goto error;
10078
10079                 leaf = path->nodes[0];
10080                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10081
10082                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10083                                                        found_key.offset);
10084                 if (!cache) {
10085                         ret = -ENOMEM;
10086                         goto error;
10087                 }
10088
10089                 if (need_clear) {
10090                         /*
10091                          * When we mount with old space cache, we need to
10092                          * set BTRFS_DC_CLEAR and set dirty flag.
10093                          *
10094                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10095                          *    truncate the old free space cache inode and
10096                          *    setup a new one.
10097                          * b) Setting 'dirty flag' makes sure that we flush
10098                          *    the new space cache info onto disk.
10099                          */
10100                         if (btrfs_test_opt(info, SPACE_CACHE))
10101                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10102                 }
10103
10104                 read_extent_buffer(leaf, &cache->item,
10105                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10106                                    sizeof(cache->item));
10107                 cache->flags = btrfs_block_group_flags(&cache->item);
10108                 if (!mixed &&
10109                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10110                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10111                         btrfs_err(info,
10112 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10113                                   cache->key.objectid);
10114                         ret = -EINVAL;
10115                         goto error;
10116                 }
10117
10118                 key.objectid = found_key.objectid + found_key.offset;
10119                 btrfs_release_path(path);
10120
10121                 /*
10122                  * We need to exclude the super stripes now so that the space
10123                  * info has super bytes accounted for, otherwise we'll think
10124                  * we have more space than we actually do.
10125                  */
10126                 ret = exclude_super_stripes(info, cache);
10127                 if (ret) {
10128                         /*
10129                          * We may have excluded something, so call this just in
10130                          * case.
10131                          */
10132                         free_excluded_extents(info, cache);
10133                         btrfs_put_block_group(cache);
10134                         goto error;
10135                 }
10136
10137                 /*
10138                  * check for two cases, either we are full, and therefore
10139                  * don't need to bother with the caching work since we won't
10140                  * find any space, or we are empty, and we can just add all
10141                  * the space in and be done with it.  This saves us _alot_ of
10142                  * time, particularly in the full case.
10143                  */
10144                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10145                         cache->last_byte_to_unpin = (u64)-1;
10146                         cache->cached = BTRFS_CACHE_FINISHED;
10147                         free_excluded_extents(info, cache);
10148                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10149                         cache->last_byte_to_unpin = (u64)-1;
10150                         cache->cached = BTRFS_CACHE_FINISHED;
10151                         add_new_free_space(cache, found_key.objectid,
10152                                            found_key.objectid +
10153                                            found_key.offset);
10154                         free_excluded_extents(info, cache);
10155                 }
10156
10157                 ret = btrfs_add_block_group_cache(info, cache);
10158                 if (ret) {
10159                         btrfs_remove_free_space_cache(cache);
10160                         btrfs_put_block_group(cache);
10161                         goto error;
10162                 }
10163
10164                 trace_btrfs_add_block_group(info, cache, 0);
10165                 update_space_info(info, cache->flags, found_key.offset,
10166                                   btrfs_block_group_used(&cache->item),
10167                                   cache->bytes_super, &space_info);
10168
10169                 cache->space_info = space_info;
10170
10171                 link_block_group(cache);
10172
10173                 set_avail_alloc_bits(info, cache->flags);
10174                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10175                         inc_block_group_ro(cache, 1);
10176                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10177                         spin_lock(&info->unused_bgs_lock);
10178                         /* Should always be true but just in case. */
10179                         if (list_empty(&cache->bg_list)) {
10180                                 btrfs_get_block_group(cache);
10181                                 trace_btrfs_add_unused_block_group(cache);
10182                                 list_add_tail(&cache->bg_list,
10183                                               &info->unused_bgs);
10184                         }
10185                         spin_unlock(&info->unused_bgs_lock);
10186                 }
10187         }
10188
10189         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10190                 if (!(get_alloc_profile(info, space_info->flags) &
10191                       (BTRFS_BLOCK_GROUP_RAID10 |
10192                        BTRFS_BLOCK_GROUP_RAID1 |
10193                        BTRFS_BLOCK_GROUP_RAID5 |
10194                        BTRFS_BLOCK_GROUP_RAID6 |
10195                        BTRFS_BLOCK_GROUP_DUP)))
10196                         continue;
10197                 /*
10198                  * avoid allocating from un-mirrored block group if there are
10199                  * mirrored block groups.
10200                  */
10201                 list_for_each_entry(cache,
10202                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10203                                 list)
10204                         inc_block_group_ro(cache, 1);
10205                 list_for_each_entry(cache,
10206                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10207                                 list)
10208                         inc_block_group_ro(cache, 1);
10209         }
10210
10211         btrfs_add_raid_kobjects(info);
10212         init_global_block_rsv(info);
10213         ret = 0;
10214 error:
10215         btrfs_free_path(path);
10216         return ret;
10217 }
10218
10219 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10220 {
10221         struct btrfs_fs_info *fs_info = trans->fs_info;
10222         struct btrfs_block_group_cache *block_group, *tmp;
10223         struct btrfs_root *extent_root = fs_info->extent_root;
10224         struct btrfs_block_group_item item;
10225         struct btrfs_key key;
10226         int ret = 0;
10227         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10228
10229         trans->can_flush_pending_bgs = false;
10230         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10231                 if (ret)
10232                         goto next;
10233
10234                 spin_lock(&block_group->lock);
10235                 memcpy(&item, &block_group->item, sizeof(item));
10236                 memcpy(&key, &block_group->key, sizeof(key));
10237                 spin_unlock(&block_group->lock);
10238
10239                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10240                                         sizeof(item));
10241                 if (ret)
10242                         btrfs_abort_transaction(trans, ret);
10243                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10244                                                key.offset);
10245                 if (ret)
10246                         btrfs_abort_transaction(trans, ret);
10247                 add_block_group_free_space(trans, block_group);
10248                 /* already aborted the transaction if it failed. */
10249 next:
10250                 list_del_init(&block_group->bg_list);
10251         }
10252         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10253 }
10254
10255 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10256                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10257                            u64 type, u64 chunk_offset, u64 size)
10258 {
10259         struct btrfs_block_group_cache *cache;
10260         int ret;
10261
10262         btrfs_set_log_full_commit(fs_info, trans);
10263
10264         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10265         if (!cache)
10266                 return -ENOMEM;
10267
10268         btrfs_set_block_group_used(&cache->item, bytes_used);
10269         btrfs_set_block_group_chunk_objectid(&cache->item,
10270                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10271         btrfs_set_block_group_flags(&cache->item, type);
10272
10273         cache->flags = type;
10274         cache->last_byte_to_unpin = (u64)-1;
10275         cache->cached = BTRFS_CACHE_FINISHED;
10276         cache->needs_free_space = 1;
10277         ret = exclude_super_stripes(fs_info, cache);
10278         if (ret) {
10279                 /*
10280                  * We may have excluded something, so call this just in
10281                  * case.
10282                  */
10283                 free_excluded_extents(fs_info, cache);
10284                 btrfs_put_block_group(cache);
10285                 return ret;
10286         }
10287
10288         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10289
10290         free_excluded_extents(fs_info, cache);
10291
10292 #ifdef CONFIG_BTRFS_DEBUG
10293         if (btrfs_should_fragment_free_space(cache)) {
10294                 u64 new_bytes_used = size - bytes_used;
10295
10296                 bytes_used += new_bytes_used >> 1;
10297                 fragment_free_space(cache);
10298         }
10299 #endif
10300         /*
10301          * Ensure the corresponding space_info object is created and
10302          * assigned to our block group. We want our bg to be added to the rbtree
10303          * with its ->space_info set.
10304          */
10305         cache->space_info = __find_space_info(fs_info, cache->flags);
10306         ASSERT(cache->space_info);
10307
10308         ret = btrfs_add_block_group_cache(fs_info, cache);
10309         if (ret) {
10310                 btrfs_remove_free_space_cache(cache);
10311                 btrfs_put_block_group(cache);
10312                 return ret;
10313         }
10314
10315         /*
10316          * Now that our block group has its ->space_info set and is inserted in
10317          * the rbtree, update the space info's counters.
10318          */
10319         trace_btrfs_add_block_group(fs_info, cache, 1);
10320         update_space_info(fs_info, cache->flags, size, bytes_used,
10321                                 cache->bytes_super, &cache->space_info);
10322         update_global_block_rsv(fs_info);
10323
10324         link_block_group(cache);
10325
10326         list_add_tail(&cache->bg_list, &trans->new_bgs);
10327
10328         set_avail_alloc_bits(fs_info, type);
10329         return 0;
10330 }
10331
10332 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10333 {
10334         u64 extra_flags = chunk_to_extended(flags) &
10335                                 BTRFS_EXTENDED_PROFILE_MASK;
10336
10337         write_seqlock(&fs_info->profiles_lock);
10338         if (flags & BTRFS_BLOCK_GROUP_DATA)
10339                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10340         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10341                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10342         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10343                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10344         write_sequnlock(&fs_info->profiles_lock);
10345 }
10346
10347 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10348                              struct btrfs_fs_info *fs_info, u64 group_start,
10349                              struct extent_map *em)
10350 {
10351         struct btrfs_root *root = fs_info->extent_root;
10352         struct btrfs_path *path;
10353         struct btrfs_block_group_cache *block_group;
10354         struct btrfs_free_cluster *cluster;
10355         struct btrfs_root *tree_root = fs_info->tree_root;
10356         struct btrfs_key key;
10357         struct inode *inode;
10358         struct kobject *kobj = NULL;
10359         int ret;
10360         int index;
10361         int factor;
10362         struct btrfs_caching_control *caching_ctl = NULL;
10363         bool remove_em;
10364
10365         block_group = btrfs_lookup_block_group(fs_info, group_start);
10366         BUG_ON(!block_group);
10367         BUG_ON(!block_group->ro);
10368
10369         trace_btrfs_remove_block_group(block_group);
10370         /*
10371          * Free the reserved super bytes from this block group before
10372          * remove it.
10373          */
10374         free_excluded_extents(fs_info, block_group);
10375         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10376                                   block_group->key.offset);
10377
10378         memcpy(&key, &block_group->key, sizeof(key));
10379         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10380         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10381                                   BTRFS_BLOCK_GROUP_RAID1 |
10382                                   BTRFS_BLOCK_GROUP_RAID10))
10383                 factor = 2;
10384         else
10385                 factor = 1;
10386
10387         /* make sure this block group isn't part of an allocation cluster */
10388         cluster = &fs_info->data_alloc_cluster;
10389         spin_lock(&cluster->refill_lock);
10390         btrfs_return_cluster_to_free_space(block_group, cluster);
10391         spin_unlock(&cluster->refill_lock);
10392
10393         /*
10394          * make sure this block group isn't part of a metadata
10395          * allocation cluster
10396          */
10397         cluster = &fs_info->meta_alloc_cluster;
10398         spin_lock(&cluster->refill_lock);
10399         btrfs_return_cluster_to_free_space(block_group, cluster);
10400         spin_unlock(&cluster->refill_lock);
10401
10402         path = btrfs_alloc_path();
10403         if (!path) {
10404                 ret = -ENOMEM;
10405                 goto out;
10406         }
10407
10408         /*
10409          * get the inode first so any iput calls done for the io_list
10410          * aren't the final iput (no unlinks allowed now)
10411          */
10412         inode = lookup_free_space_inode(fs_info, block_group, path);
10413
10414         mutex_lock(&trans->transaction->cache_write_mutex);
10415         /*
10416          * make sure our free spache cache IO is done before remove the
10417          * free space inode
10418          */
10419         spin_lock(&trans->transaction->dirty_bgs_lock);
10420         if (!list_empty(&block_group->io_list)) {
10421                 list_del_init(&block_group->io_list);
10422
10423                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10424
10425                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10426                 btrfs_wait_cache_io(trans, block_group, path);
10427                 btrfs_put_block_group(block_group);
10428                 spin_lock(&trans->transaction->dirty_bgs_lock);
10429         }
10430
10431         if (!list_empty(&block_group->dirty_list)) {
10432                 list_del_init(&block_group->dirty_list);
10433                 btrfs_put_block_group(block_group);
10434         }
10435         spin_unlock(&trans->transaction->dirty_bgs_lock);
10436         mutex_unlock(&trans->transaction->cache_write_mutex);
10437
10438         if (!IS_ERR(inode)) {
10439                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10440                 if (ret) {
10441                         btrfs_add_delayed_iput(inode);
10442                         goto out;
10443                 }
10444                 clear_nlink(inode);
10445                 /* One for the block groups ref */
10446                 spin_lock(&block_group->lock);
10447                 if (block_group->iref) {
10448                         block_group->iref = 0;
10449                         block_group->inode = NULL;
10450                         spin_unlock(&block_group->lock);
10451                         iput(inode);
10452                 } else {
10453                         spin_unlock(&block_group->lock);
10454                 }
10455                 /* One for our lookup ref */
10456                 btrfs_add_delayed_iput(inode);
10457         }
10458
10459         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10460         key.offset = block_group->key.objectid;
10461         key.type = 0;
10462
10463         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10464         if (ret < 0)
10465                 goto out;
10466         if (ret > 0)
10467                 btrfs_release_path(path);
10468         if (ret == 0) {
10469                 ret = btrfs_del_item(trans, tree_root, path);
10470                 if (ret)
10471                         goto out;
10472                 btrfs_release_path(path);
10473         }
10474
10475         spin_lock(&fs_info->block_group_cache_lock);
10476         rb_erase(&block_group->cache_node,
10477                  &fs_info->block_group_cache_tree);
10478         RB_CLEAR_NODE(&block_group->cache_node);
10479
10480         if (fs_info->first_logical_byte == block_group->key.objectid)
10481                 fs_info->first_logical_byte = (u64)-1;
10482         spin_unlock(&fs_info->block_group_cache_lock);
10483
10484         down_write(&block_group->space_info->groups_sem);
10485         /*
10486          * we must use list_del_init so people can check to see if they
10487          * are still on the list after taking the semaphore
10488          */
10489         list_del_init(&block_group->list);
10490         if (list_empty(&block_group->space_info->block_groups[index])) {
10491                 kobj = block_group->space_info->block_group_kobjs[index];
10492                 block_group->space_info->block_group_kobjs[index] = NULL;
10493                 clear_avail_alloc_bits(fs_info, block_group->flags);
10494         }
10495         up_write(&block_group->space_info->groups_sem);
10496         if (kobj) {
10497                 kobject_del(kobj);
10498                 kobject_put(kobj);
10499         }
10500
10501         if (block_group->has_caching_ctl)
10502                 caching_ctl = get_caching_control(block_group);
10503         if (block_group->cached == BTRFS_CACHE_STARTED)
10504                 wait_block_group_cache_done(block_group);
10505         if (block_group->has_caching_ctl) {
10506                 down_write(&fs_info->commit_root_sem);
10507                 if (!caching_ctl) {
10508                         struct btrfs_caching_control *ctl;
10509
10510                         list_for_each_entry(ctl,
10511                                     &fs_info->caching_block_groups, list)
10512                                 if (ctl->block_group == block_group) {
10513                                         caching_ctl = ctl;
10514                                         refcount_inc(&caching_ctl->count);
10515                                         break;
10516                                 }
10517                 }
10518                 if (caching_ctl)
10519                         list_del_init(&caching_ctl->list);
10520                 up_write(&fs_info->commit_root_sem);
10521                 if (caching_ctl) {
10522                         /* Once for the caching bgs list and once for us. */
10523                         put_caching_control(caching_ctl);
10524                         put_caching_control(caching_ctl);
10525                 }
10526         }
10527
10528         spin_lock(&trans->transaction->dirty_bgs_lock);
10529         if (!list_empty(&block_group->dirty_list)) {
10530                 WARN_ON(1);
10531         }
10532         if (!list_empty(&block_group->io_list)) {
10533                 WARN_ON(1);
10534         }
10535         spin_unlock(&trans->transaction->dirty_bgs_lock);
10536         btrfs_remove_free_space_cache(block_group);
10537
10538         spin_lock(&block_group->space_info->lock);
10539         list_del_init(&block_group->ro_list);
10540
10541         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10542                 WARN_ON(block_group->space_info->total_bytes
10543                         < block_group->key.offset);
10544                 WARN_ON(block_group->space_info->bytes_readonly
10545                         < block_group->key.offset);
10546                 WARN_ON(block_group->space_info->disk_total
10547                         < block_group->key.offset * factor);
10548         }
10549         block_group->space_info->total_bytes -= block_group->key.offset;
10550         block_group->space_info->bytes_readonly -= block_group->key.offset;
10551         block_group->space_info->disk_total -= block_group->key.offset * factor;
10552
10553         spin_unlock(&block_group->space_info->lock);
10554
10555         memcpy(&key, &block_group->key, sizeof(key));
10556
10557         mutex_lock(&fs_info->chunk_mutex);
10558         if (!list_empty(&em->list)) {
10559                 /* We're in the transaction->pending_chunks list. */
10560                 free_extent_map(em);
10561         }
10562         spin_lock(&block_group->lock);
10563         block_group->removed = 1;
10564         /*
10565          * At this point trimming can't start on this block group, because we
10566          * removed the block group from the tree fs_info->block_group_cache_tree
10567          * so no one can't find it anymore and even if someone already got this
10568          * block group before we removed it from the rbtree, they have already
10569          * incremented block_group->trimming - if they didn't, they won't find
10570          * any free space entries because we already removed them all when we
10571          * called btrfs_remove_free_space_cache().
10572          *
10573          * And we must not remove the extent map from the fs_info->mapping_tree
10574          * to prevent the same logical address range and physical device space
10575          * ranges from being reused for a new block group. This is because our
10576          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10577          * completely transactionless, so while it is trimming a range the
10578          * currently running transaction might finish and a new one start,
10579          * allowing for new block groups to be created that can reuse the same
10580          * physical device locations unless we take this special care.
10581          *
10582          * There may also be an implicit trim operation if the file system
10583          * is mounted with -odiscard. The same protections must remain
10584          * in place until the extents have been discarded completely when
10585          * the transaction commit has completed.
10586          */
10587         remove_em = (atomic_read(&block_group->trimming) == 0);
10588         /*
10589          * Make sure a trimmer task always sees the em in the pinned_chunks list
10590          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10591          * before checking block_group->removed).
10592          */
10593         if (!remove_em) {
10594                 /*
10595                  * Our em might be in trans->transaction->pending_chunks which
10596                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10597                  * and so is the fs_info->pinned_chunks list.
10598                  *
10599                  * So at this point we must be holding the chunk_mutex to avoid
10600                  * any races with chunk allocation (more specifically at
10601                  * volumes.c:contains_pending_extent()), to ensure it always
10602                  * sees the em, either in the pending_chunks list or in the
10603                  * pinned_chunks list.
10604                  */
10605                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10606         }
10607         spin_unlock(&block_group->lock);
10608
10609         if (remove_em) {
10610                 struct extent_map_tree *em_tree;
10611
10612                 em_tree = &fs_info->mapping_tree.map_tree;
10613                 write_lock(&em_tree->lock);
10614                 /*
10615                  * The em might be in the pending_chunks list, so make sure the
10616                  * chunk mutex is locked, since remove_extent_mapping() will
10617                  * delete us from that list.
10618                  */
10619                 remove_extent_mapping(em_tree, em);
10620                 write_unlock(&em_tree->lock);
10621                 /* once for the tree */
10622                 free_extent_map(em);
10623         }
10624
10625         mutex_unlock(&fs_info->chunk_mutex);
10626
10627         ret = remove_block_group_free_space(trans, block_group);
10628         if (ret)
10629                 goto out;
10630
10631         btrfs_put_block_group(block_group);
10632         btrfs_put_block_group(block_group);
10633
10634         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10635         if (ret > 0)
10636                 ret = -EIO;
10637         if (ret < 0)
10638                 goto out;
10639
10640         ret = btrfs_del_item(trans, root, path);
10641 out:
10642         btrfs_free_path(path);
10643         return ret;
10644 }
10645
10646 struct btrfs_trans_handle *
10647 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10648                                      const u64 chunk_offset)
10649 {
10650         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10651         struct extent_map *em;
10652         struct map_lookup *map;
10653         unsigned int num_items;
10654
10655         read_lock(&em_tree->lock);
10656         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10657         read_unlock(&em_tree->lock);
10658         ASSERT(em && em->start == chunk_offset);
10659
10660         /*
10661          * We need to reserve 3 + N units from the metadata space info in order
10662          * to remove a block group (done at btrfs_remove_chunk() and at
10663          * btrfs_remove_block_group()), which are used for:
10664          *
10665          * 1 unit for adding the free space inode's orphan (located in the tree
10666          * of tree roots).
10667          * 1 unit for deleting the block group item (located in the extent
10668          * tree).
10669          * 1 unit for deleting the free space item (located in tree of tree
10670          * roots).
10671          * N units for deleting N device extent items corresponding to each
10672          * stripe (located in the device tree).
10673          *
10674          * In order to remove a block group we also need to reserve units in the
10675          * system space info in order to update the chunk tree (update one or
10676          * more device items and remove one chunk item), but this is done at
10677          * btrfs_remove_chunk() through a call to check_system_chunk().
10678          */
10679         map = em->map_lookup;
10680         num_items = 3 + map->num_stripes;
10681         free_extent_map(em);
10682
10683         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10684                                                            num_items, 1);
10685 }
10686
10687 /*
10688  * Process the unused_bgs list and remove any that don't have any allocated
10689  * space inside of them.
10690  */
10691 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10692 {
10693         struct btrfs_block_group_cache *block_group;
10694         struct btrfs_space_info *space_info;
10695         struct btrfs_trans_handle *trans;
10696         int ret = 0;
10697
10698         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10699                 return;
10700
10701         spin_lock(&fs_info->unused_bgs_lock);
10702         while (!list_empty(&fs_info->unused_bgs)) {
10703                 u64 start, end;
10704                 int trimming;
10705
10706                 block_group = list_first_entry(&fs_info->unused_bgs,
10707                                                struct btrfs_block_group_cache,
10708                                                bg_list);
10709                 list_del_init(&block_group->bg_list);
10710
10711                 space_info = block_group->space_info;
10712
10713                 if (ret || btrfs_mixed_space_info(space_info)) {
10714                         btrfs_put_block_group(block_group);
10715                         continue;
10716                 }
10717                 spin_unlock(&fs_info->unused_bgs_lock);
10718
10719                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10720
10721                 /* Don't want to race with allocators so take the groups_sem */
10722                 down_write(&space_info->groups_sem);
10723                 spin_lock(&block_group->lock);
10724                 if (block_group->reserved ||
10725                     btrfs_block_group_used(&block_group->item) ||
10726                     block_group->ro ||
10727                     list_is_singular(&block_group->list)) {
10728                         /*
10729                          * We want to bail if we made new allocations or have
10730                          * outstanding allocations in this block group.  We do
10731                          * the ro check in case balance is currently acting on
10732                          * this block group.
10733                          */
10734                         trace_btrfs_skip_unused_block_group(block_group);
10735                         spin_unlock(&block_group->lock);
10736                         up_write(&space_info->groups_sem);
10737                         goto next;
10738                 }
10739                 spin_unlock(&block_group->lock);
10740
10741                 /* We don't want to force the issue, only flip if it's ok. */
10742                 ret = inc_block_group_ro(block_group, 0);
10743                 up_write(&space_info->groups_sem);
10744                 if (ret < 0) {
10745                         ret = 0;
10746                         goto next;
10747                 }
10748
10749                 /*
10750                  * Want to do this before we do anything else so we can recover
10751                  * properly if we fail to join the transaction.
10752                  */
10753                 trans = btrfs_start_trans_remove_block_group(fs_info,
10754                                                      block_group->key.objectid);
10755                 if (IS_ERR(trans)) {
10756                         btrfs_dec_block_group_ro(block_group);
10757                         ret = PTR_ERR(trans);
10758                         goto next;
10759                 }
10760
10761                 /*
10762                  * We could have pending pinned extents for this block group,
10763                  * just delete them, we don't care about them anymore.
10764                  */
10765                 start = block_group->key.objectid;
10766                 end = start + block_group->key.offset - 1;
10767                 /*
10768                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10769                  * btrfs_finish_extent_commit(). If we are at transaction N,
10770                  * another task might be running finish_extent_commit() for the
10771                  * previous transaction N - 1, and have seen a range belonging
10772                  * to the block group in freed_extents[] before we were able to
10773                  * clear the whole block group range from freed_extents[]. This
10774                  * means that task can lookup for the block group after we
10775                  * unpinned it from freed_extents[] and removed it, leading to
10776                  * a BUG_ON() at btrfs_unpin_extent_range().
10777                  */
10778                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10779                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10780                                   EXTENT_DIRTY);
10781                 if (ret) {
10782                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10783                         btrfs_dec_block_group_ro(block_group);
10784                         goto end_trans;
10785                 }
10786                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10787                                   EXTENT_DIRTY);
10788                 if (ret) {
10789                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10790                         btrfs_dec_block_group_ro(block_group);
10791                         goto end_trans;
10792                 }
10793                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10794
10795                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10796                 spin_lock(&space_info->lock);
10797                 spin_lock(&block_group->lock);
10798
10799                 space_info->bytes_pinned -= block_group->pinned;
10800                 space_info->bytes_readonly += block_group->pinned;
10801                 percpu_counter_add(&space_info->total_bytes_pinned,
10802                                    -block_group->pinned);
10803                 block_group->pinned = 0;
10804
10805                 spin_unlock(&block_group->lock);
10806                 spin_unlock(&space_info->lock);
10807
10808                 /* DISCARD can flip during remount */
10809                 trimming = btrfs_test_opt(fs_info, DISCARD);
10810
10811                 /* Implicit trim during transaction commit. */
10812                 if (trimming)
10813                         btrfs_get_block_group_trimming(block_group);
10814
10815                 /*
10816                  * Btrfs_remove_chunk will abort the transaction if things go
10817                  * horribly wrong.
10818                  */
10819                 ret = btrfs_remove_chunk(trans, fs_info,
10820                                          block_group->key.objectid);
10821
10822                 if (ret) {
10823                         if (trimming)
10824                                 btrfs_put_block_group_trimming(block_group);
10825                         goto end_trans;
10826                 }
10827
10828                 /*
10829                  * If we're not mounted with -odiscard, we can just forget
10830                  * about this block group. Otherwise we'll need to wait
10831                  * until transaction commit to do the actual discard.
10832                  */
10833                 if (trimming) {
10834                         spin_lock(&fs_info->unused_bgs_lock);
10835                         /*
10836                          * A concurrent scrub might have added us to the list
10837                          * fs_info->unused_bgs, so use a list_move operation
10838                          * to add the block group to the deleted_bgs list.
10839                          */
10840                         list_move(&block_group->bg_list,
10841                                   &trans->transaction->deleted_bgs);
10842                         spin_unlock(&fs_info->unused_bgs_lock);
10843                         btrfs_get_block_group(block_group);
10844                 }
10845 end_trans:
10846                 btrfs_end_transaction(trans);
10847 next:
10848                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10849                 btrfs_put_block_group(block_group);
10850                 spin_lock(&fs_info->unused_bgs_lock);
10851         }
10852         spin_unlock(&fs_info->unused_bgs_lock);
10853 }
10854
10855 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10856 {
10857         struct btrfs_super_block *disk_super;
10858         u64 features;
10859         u64 flags;
10860         int mixed = 0;
10861         int ret;
10862
10863         disk_super = fs_info->super_copy;
10864         if (!btrfs_super_root(disk_super))
10865                 return -EINVAL;
10866
10867         features = btrfs_super_incompat_flags(disk_super);
10868         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10869                 mixed = 1;
10870
10871         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10872         ret = create_space_info(fs_info, flags);
10873         if (ret)
10874                 goto out;
10875
10876         if (mixed) {
10877                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10878                 ret = create_space_info(fs_info, flags);
10879         } else {
10880                 flags = BTRFS_BLOCK_GROUP_METADATA;
10881                 ret = create_space_info(fs_info, flags);
10882                 if (ret)
10883                         goto out;
10884
10885                 flags = BTRFS_BLOCK_GROUP_DATA;
10886                 ret = create_space_info(fs_info, flags);
10887         }
10888 out:
10889         return ret;
10890 }
10891
10892 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10893                                    u64 start, u64 end)
10894 {
10895         return unpin_extent_range(fs_info, start, end, false);
10896 }
10897
10898 /*
10899  * It used to be that old block groups would be left around forever.
10900  * Iterating over them would be enough to trim unused space.  Since we
10901  * now automatically remove them, we also need to iterate over unallocated
10902  * space.
10903  *
10904  * We don't want a transaction for this since the discard may take a
10905  * substantial amount of time.  We don't require that a transaction be
10906  * running, but we do need to take a running transaction into account
10907  * to ensure that we're not discarding chunks that were released in
10908  * the current transaction.
10909  *
10910  * Holding the chunks lock will prevent other threads from allocating
10911  * or releasing chunks, but it won't prevent a running transaction
10912  * from committing and releasing the memory that the pending chunks
10913  * list head uses.  For that, we need to take a reference to the
10914  * transaction.
10915  */
10916 static int btrfs_trim_free_extents(struct btrfs_device *device,
10917                                    u64 minlen, u64 *trimmed)
10918 {
10919         u64 start = 0, len = 0;
10920         int ret;
10921
10922         *trimmed = 0;
10923
10924         /* Not writeable = nothing to do. */
10925         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10926                 return 0;
10927
10928         /* No free space = nothing to do. */
10929         if (device->total_bytes <= device->bytes_used)
10930                 return 0;
10931
10932         ret = 0;
10933
10934         while (1) {
10935                 struct btrfs_fs_info *fs_info = device->fs_info;
10936                 struct btrfs_transaction *trans;
10937                 u64 bytes;
10938
10939                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10940                 if (ret)
10941                         return ret;
10942
10943                 down_read(&fs_info->commit_root_sem);
10944
10945                 spin_lock(&fs_info->trans_lock);
10946                 trans = fs_info->running_transaction;
10947                 if (trans)
10948                         refcount_inc(&trans->use_count);
10949                 spin_unlock(&fs_info->trans_lock);
10950
10951                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10952                                                  &start, &len);
10953                 if (trans)
10954                         btrfs_put_transaction(trans);
10955
10956                 if (ret) {
10957                         up_read(&fs_info->commit_root_sem);
10958                         mutex_unlock(&fs_info->chunk_mutex);
10959                         if (ret == -ENOSPC)
10960                                 ret = 0;
10961                         break;
10962                 }
10963
10964                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10965                 up_read(&fs_info->commit_root_sem);
10966                 mutex_unlock(&fs_info->chunk_mutex);
10967
10968                 if (ret)
10969                         break;
10970
10971                 start += len;
10972                 *trimmed += bytes;
10973
10974                 if (fatal_signal_pending(current)) {
10975                         ret = -ERESTARTSYS;
10976                         break;
10977                 }
10978
10979                 cond_resched();
10980         }
10981
10982         return ret;
10983 }
10984
10985 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10986 {
10987         struct btrfs_block_group_cache *cache = NULL;
10988         struct btrfs_device *device;
10989         struct list_head *devices;
10990         u64 group_trimmed;
10991         u64 start;
10992         u64 end;
10993         u64 trimmed = 0;
10994         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10995         int ret = 0;
10996
10997         /*
10998          * try to trim all FS space, our block group may start from non-zero.
10999          */
11000         if (range->len == total_bytes)
11001                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11002         else
11003                 cache = btrfs_lookup_block_group(fs_info, range->start);
11004
11005         while (cache) {
11006                 if (cache->key.objectid >= (range->start + range->len)) {
11007                         btrfs_put_block_group(cache);
11008                         break;
11009                 }
11010
11011                 start = max(range->start, cache->key.objectid);
11012                 end = min(range->start + range->len,
11013                                 cache->key.objectid + cache->key.offset);
11014
11015                 if (end - start >= range->minlen) {
11016                         if (!block_group_cache_done(cache)) {
11017                                 ret = cache_block_group(cache, 0);
11018                                 if (ret) {
11019                                         btrfs_put_block_group(cache);
11020                                         break;
11021                                 }
11022                                 ret = wait_block_group_cache_done(cache);
11023                                 if (ret) {
11024                                         btrfs_put_block_group(cache);
11025                                         break;
11026                                 }
11027                         }
11028                         ret = btrfs_trim_block_group(cache,
11029                                                      &group_trimmed,
11030                                                      start,
11031                                                      end,
11032                                                      range->minlen);
11033
11034                         trimmed += group_trimmed;
11035                         if (ret) {
11036                                 btrfs_put_block_group(cache);
11037                                 break;
11038                         }
11039                 }
11040
11041                 cache = next_block_group(fs_info, cache);
11042         }
11043
11044         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11045         devices = &fs_info->fs_devices->alloc_list;
11046         list_for_each_entry(device, devices, dev_alloc_list) {
11047                 ret = btrfs_trim_free_extents(device, range->minlen,
11048                                               &group_trimmed);
11049                 if (ret)
11050                         break;
11051
11052                 trimmed += group_trimmed;
11053         }
11054         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11055
11056         range->len = trimmed;
11057         return ret;
11058 }
11059
11060 /*
11061  * btrfs_{start,end}_write_no_snapshotting() are similar to
11062  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11063  * data into the page cache through nocow before the subvolume is snapshoted,
11064  * but flush the data into disk after the snapshot creation, or to prevent
11065  * operations while snapshotting is ongoing and that cause the snapshot to be
11066  * inconsistent (writes followed by expanding truncates for example).
11067  */
11068 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11069 {
11070         percpu_counter_dec(&root->subv_writers->counter);
11071         cond_wake_up(&root->subv_writers->wait);
11072 }
11073
11074 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11075 {
11076         if (atomic_read(&root->will_be_snapshotted))
11077                 return 0;
11078
11079         percpu_counter_inc(&root->subv_writers->counter);
11080         /*
11081          * Make sure counter is updated before we check for snapshot creation.
11082          */
11083         smp_mb();
11084         if (atomic_read(&root->will_be_snapshotted)) {
11085                 btrfs_end_write_no_snapshotting(root);
11086                 return 0;
11087         }
11088         return 1;
11089 }
11090
11091 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11092 {
11093         while (true) {
11094                 int ret;
11095
11096                 ret = btrfs_start_write_no_snapshotting(root);
11097                 if (ret)
11098                         break;
11099                 wait_var_event(&root->will_be_snapshotted,
11100                                !atomic_read(&root->will_be_snapshotted));
11101         }
11102 }