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_delayed_ref_node *node, u64 parent,
  56                                u64 root_objectid, u64 owner_objectid,
  57                                u64 owner_offset, int refs_to_drop,
  58                                struct btrfs_delayed_extent_op *extra_op);
  59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  60                                     struct extent_buffer *leaf,
  61                                     struct btrfs_extent_item *ei);
  62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  63                                       u64 parent, u64 root_objectid,
  64                                       u64 flags, u64 owner, u64 offset,
  65                                       struct btrfs_key *ins, int ref_mod);
  66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  67                                      struct btrfs_delayed_ref_node *node,
  68                                      struct btrfs_delayed_extent_op *extent_op);
  69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  70                           int force);
  71 static int find_next_key(struct btrfs_path *path, int level,
  72                          struct btrfs_key *key);
  73 static void dump_space_info(struct btrfs_fs_info *fs_info,
  74                             struct btrfs_space_info *info, u64 bytes,
  75                             int dump_block_groups);
  76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  77                                u64 num_bytes);
  78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  79                                      struct btrfs_space_info *space_info,
  80                                      u64 num_bytes);
  81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  82                                      struct btrfs_space_info *space_info,
  83                                      u64 num_bytes);
  84
  85 static noinline int
  86 block_group_cache_done(struct btrfs_block_group_cache *cache)
  87 {
  88         smp_mb();
  89         return cache->cached == BTRFS_CACHE_FINISHED ||
  90                 cache->cached == BTRFS_CACHE_ERROR;
  91 }
  92
  93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  94 {
  95         return (cache->flags & bits) == bits;
  96 }
  97
  98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
  99 {
 100         atomic_inc(&cache->count);
 101 }
 102
 103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 104 {
 105         if (atomic_dec_and_test(&cache->count)) {
 106                 WARN_ON(cache->pinned > 0);
 107                 WARN_ON(cache->reserved > 0);
 108
 109                 /*
 110                  * If not empty, someone is still holding mutex of
 111                  * full_stripe_lock, which can only be released by caller.
 112                  * And it will definitely cause use-after-free when caller
 113                  * tries to release full stripe lock.
 114                  *
 115                  * No better way to resolve, but only to warn.
 116                  */
 117                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 118                 kfree(cache->free_space_ctl);
 119                 kfree(cache);
 120         }
 121 }
 122
 123 /*
 124  * this adds the block group to the fs_info rb tree for the block group
 125  * cache
 126  */
 127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 128                                 struct btrfs_block_group_cache *block_group)
 129 {
 130         struct rb_node **p;
 131         struct rb_node *parent = NULL;
 132         struct btrfs_block_group_cache *cache;
 133
 134         spin_lock(&info->block_group_cache_lock);
 135         p = &info->block_group_cache_tree.rb_node;
 136
 137         while (*p) {
 138                 parent = *p;
 139                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 140                                  cache_node);
 141                 if (block_group->key.objectid < cache->key.objectid) {
 142                         p = &(*p)->rb_left;
 143                 } else if (block_group->key.objectid > cache->key.objectid) {
 144                         p = &(*p)->rb_right;
 145                 } else {
 146                         spin_unlock(&info->block_group_cache_lock);
 147                         return -EEXIST;
 148                 }
 149         }
 150
 151         rb_link_node(&block_group->cache_node, parent, p);
 152         rb_insert_color(&block_group->cache_node,
 153                         &info->block_group_cache_tree);
 154
 155         if (info->first_logical_byte > block_group->key.objectid)
 156                 info->first_logical_byte = block_group->key.objectid;
 157
 158         spin_unlock(&info->block_group_cache_lock);
 159
 160         return 0;
 161 }
 162
 163 /*
 164  * This will return the block group at or after bytenr if contains is 0, else
 165  * it will return the block group that contains the bytenr
 166  */
 167 static struct btrfs_block_group_cache *
 168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 169                               int contains)
 170 {
 171         struct btrfs_block_group_cache *cache, *ret = NULL;
 172         struct rb_node *n;
 173         u64 end, start;
 174
 175         spin_lock(&info->block_group_cache_lock);
 176         n = info->block_group_cache_tree.rb_node;
 177
 178         while (n) {
 179                 cache = rb_entry(n, struct btrfs_block_group_cache,
 180                                  cache_node);
 181                 end = cache->key.objectid + cache->key.offset - 1;
 182                 start = cache->key.objectid;
 183
 184                 if (bytenr < start) {
 185                         if (!contains && (!ret || start < ret->key.objectid))
 186                                 ret = cache;
 187                         n = n->rb_left;
 188                 } else if (bytenr > start) {
 189                         if (contains && bytenr <= end) {
 190                                 ret = cache;
 191                                 break;
 192                         }
 193                         n = n->rb_right;
 194                 } else {
 195                         ret = cache;
 196                         break;
 197                 }
 198         }
 199         if (ret) {
 200                 btrfs_get_block_group(ret);
 201                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 202                         info->first_logical_byte = ret->key.objectid;
 203         }
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE);
 215         set_extent_bits(&fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 221 {
 222         struct btrfs_fs_info *fs_info = cache->fs_info;
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE);
 230         clear_extent_bits(&fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 235 {
 236         struct btrfs_fs_info *fs_info = cache->fs_info;
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 246                                           stripe_len);
 247                 if (ret)
 248                         return ret;
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 254                                        bytenr, &logical, &nr, &stripe_len);
 255                 if (ret)
 256                         return ret;
 257
 258                 while (nr--) {
 259                         u64 start, len;
 260
 261                         if (logical[nr] > cache->key.objectid +
 262                             cache->key.offset)
 263                                 continue;
 264
 265                         if (logical[nr] + stripe_len <= cache->key.objectid)
 266                                 continue;
 267
 268                         start = logical[nr];
 269                         if (start < cache->key.objectid) {
 270                                 start = cache->key.objectid;
 271                                 len = (logical[nr] + stripe_len) - start;
 272                         } else {
 273                                 len = min_t(u64, stripe_len,
 274                                             cache->key.objectid +
 275                                             cache->key.offset - start);
 276                         }
 277
 278                         cache->bytes_super += len;
 279                         ret = add_excluded_extent(fs_info, start, len);
 280                         if (ret) {
 281                                 kfree(logical);
 282                                 return ret;
 283                         }
 284                 }
 285
 286                 kfree(logical);
 287         }
 288         return 0;
 289 }
 290
 291 static struct btrfs_caching_control *
 292 get_caching_control(struct btrfs_block_group_cache *cache)
 293 {
 294         struct btrfs_caching_control *ctl;
 295
 296         spin_lock(&cache->lock);
 297         if (!cache->caching_ctl) {
 298                 spin_unlock(&cache->lock);
 299                 return NULL;
 300         }
 301
 302         ctl = cache->caching_ctl;
 303         refcount_inc(&ctl->count);
 304         spin_unlock(&cache->lock);
 305         return ctl;
 306 }
 307
 308 static void put_caching_control(struct btrfs_caching_control *ctl)
 309 {
 310         if (refcount_dec_and_test(&ctl->count))
 311                 kfree(ctl);
 312 }
 313
 314 #ifdef CONFIG_BTRFS_DEBUG
 315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 316 {
 317         struct btrfs_fs_info *fs_info = block_group->fs_info;
 318         u64 start = block_group->key.objectid;
 319         u64 len = block_group->key.offset;
 320         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 321                 fs_info->nodesize : fs_info->sectorsize;
 322         u64 step = chunk << 1;
 323
 324         while (len > chunk) {
 325                 btrfs_remove_free_space(block_group, start, chunk);
 326                 start += step;
 327                 if (len < step)
 328                         len = 0;
 329                 else
 330                         len -= step;
 331         }
 332 }
 333 #endif
 334
 335 /*
 336  * this is only called by cache_block_group, since we could have freed extents
 337  * we need to check the pinned_extents for any extents that can't be used yet
 338  * since their free space will be released as soon as the transaction commits.
 339  */
 340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 341                        u64 start, u64 end)
 342 {
 343         struct btrfs_fs_info *info = block_group->fs_info;
 344         u64 extent_start, extent_end, size, total_added = 0;
 345         int ret;
 346
 347         while (start < end) {
 348                 ret = find_first_extent_bit(info->pinned_extents, start,
 349                                             &extent_start, &extent_end,
 350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 351                                             NULL);
 352                 if (ret)
 353                         break;
 354
 355                 if (extent_start <= start) {
 356                         start = extent_end + 1;
 357                 } else if (extent_start > start && extent_start < end) {
 358                         size = extent_start - start;
 359                         total_added += size;
 360                         ret = btrfs_add_free_space(block_group, start,
 361                                                    size);
 362                         BUG_ON(ret); /* -ENOMEM or logic error */
 363                         start = extent_end + 1;
 364                 } else {
 365                         break;
 366                 }
 367         }
 368
 369         if (start < end) {
 370                 size = end - start;
 371                 total_added += size;
 372                 ret = btrfs_add_free_space(block_group, start, size);
 373                 BUG_ON(ret); /* -ENOMEM or logic error */
 374         }
 375
 376         return total_added;
 377 }
 378
 379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 380 {
 381         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 382         struct btrfs_fs_info *fs_info = block_group->fs_info;
 383         struct btrfs_root *extent_root = fs_info->extent_root;
 384         struct btrfs_path *path;
 385         struct extent_buffer *leaf;
 386         struct btrfs_key key;
 387         u64 total_found = 0;
 388         u64 last = 0;
 389         u32 nritems;
 390         int ret;
 391         bool wakeup = true;
 392
 393         path = btrfs_alloc_path();
 394         if (!path)
 395                 return -ENOMEM;
 396
 397         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 398
 399 #ifdef CONFIG_BTRFS_DEBUG
 400         /*
 401          * If we're fragmenting we don't want to make anybody think we can
 402          * allocate from this block group until we've had a chance to fragment
 403          * the free space.
 404          */
 405         if (btrfs_should_fragment_free_space(block_group))
 406                 wakeup = false;
 407 #endif
 408         /*
 409          * We don't want to deadlock with somebody trying to allocate a new
 410          * extent for the extent root while also trying to search the extent
 411          * root to add free space.  So we skip locking and search the commit
 412          * root, since its read-only
 413          */
 414         path->skip_locking = 1;
 415         path->search_commit_root = 1;
 416         path->reada = READA_FORWARD;
 417
 418         key.objectid = last;
 419         key.offset = 0;
 420         key.type = BTRFS_EXTENT_ITEM_KEY;
 421
 422 next:
 423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 424         if (ret < 0)
 425                 goto out;
 426
 427         leaf = path->nodes[0];
 428         nritems = btrfs_header_nritems(leaf);
 429
 430         while (1) {
 431                 if (btrfs_fs_closing(fs_info) > 1) {
 432                         last = (u64)-1;
 433                         break;
 434                 }
 435
 436                 if (path->slots[0] < nritems) {
 437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 438                 } else {
 439                         ret = find_next_key(path, 0, &key);
 440                         if (ret)
 441                                 break;
 442
 443                         if (need_resched() ||
 444                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 445                                 if (wakeup)
 446                                         caching_ctl->progress = last;
 447                                 btrfs_release_path(path);
 448                                 up_read(&fs_info->commit_root_sem);
 449                                 mutex_unlock(&caching_ctl->mutex);
 450                                 cond_resched();
 451                                 mutex_lock(&caching_ctl->mutex);
 452                                 down_read(&fs_info->commit_root_sem);
 453                                 goto next;
 454                         }
 455
 456                         ret = btrfs_next_leaf(extent_root, path);
 457                         if (ret < 0)
 458                                 goto out;
 459                         if (ret)
 460                                 break;
 461                         leaf = path->nodes[0];
 462                         nritems = btrfs_header_nritems(leaf);
 463                         continue;
 464                 }
 465
 466                 if (key.objectid < last) {
 467                         key.objectid = last;
 468                         key.offset = 0;
 469                         key.type = BTRFS_EXTENT_ITEM_KEY;
 470
 471                         if (wakeup)
 472                                 caching_ctl->progress = last;
 473                         btrfs_release_path(path);
 474                         goto next;
 475                 }
 476
 477                 if (key.objectid < block_group->key.objectid) {
 478                         path->slots[0]++;
 479                         continue;
 480                 }
 481
 482                 if (key.objectid >= block_group->key.objectid +
 483                     block_group->key.offset)
 484                         break;
 485
 486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 487                     key.type == BTRFS_METADATA_ITEM_KEY) {
 488                         total_found += add_new_free_space(block_group, last,
 489                                                           key.objectid);
 490                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 491                                 last = key.objectid +
 492                                         fs_info->nodesize;
 493                         else
 494                                 last = key.objectid + key.offset;
 495
 496                         if (total_found > CACHING_CTL_WAKE_UP) {
 497                                 total_found = 0;
 498                                 if (wakeup)
 499                                         wake_up(&caching_ctl->wait);
 500                         }
 501                 }
 502                 path->slots[0]++;
 503         }
 504         ret = 0;
 505
 506         total_found += add_new_free_space(block_group, last,
 507                                           block_group->key.objectid +
 508                                           block_group->key.offset);
 509         caching_ctl->progress = (u64)-1;
 510
 511 out:
 512         btrfs_free_path(path);
 513         return ret;
 514 }
 515
 516 static noinline void caching_thread(struct btrfs_work *work)
 517 {
 518         struct btrfs_block_group_cache *block_group;
 519         struct btrfs_fs_info *fs_info;
 520         struct btrfs_caching_control *caching_ctl;
 521         int ret;
 522
 523         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 524         block_group = caching_ctl->block_group;
 525         fs_info = block_group->fs_info;
 526
 527         mutex_lock(&caching_ctl->mutex);
 528         down_read(&fs_info->commit_root_sem);
 529
 530         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 531                 ret = load_free_space_tree(caching_ctl);
 532         else
 533                 ret = load_extent_tree_free(caching_ctl);
 534
 535         spin_lock(&block_group->lock);
 536         block_group->caching_ctl = NULL;
 537         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 538         spin_unlock(&block_group->lock);
 539
 540 #ifdef CONFIG_BTRFS_DEBUG
 541         if (btrfs_should_fragment_free_space(block_group)) {
 542                 u64 bytes_used;
 543
 544                 spin_lock(&block_group->space_info->lock);
 545                 spin_lock(&block_group->lock);
 546                 bytes_used = block_group->key.offset -
 547                         btrfs_block_group_used(&block_group->item);
 548                 block_group->space_info->bytes_used += bytes_used >> 1;
 549                 spin_unlock(&block_group->lock);
 550                 spin_unlock(&block_group->space_info->lock);
 551                 fragment_free_space(block_group);
 552         }
 553 #endif
 554
 555         caching_ctl->progress = (u64)-1;
 556
 557         up_read(&fs_info->commit_root_sem);
 558         free_excluded_extents(block_group);
 559         mutex_unlock(&caching_ctl->mutex);
 560
 561         wake_up(&caching_ctl->wait);
 562
 563         put_caching_control(caching_ctl);
 564         btrfs_put_block_group(block_group);
 565 }
 566
 567 static int cache_block_group(struct btrfs_block_group_cache *cache,
 568                              int load_cache_only)
 569 {
 570         DEFINE_WAIT(wait);
 571         struct btrfs_fs_info *fs_info = cache->fs_info;
 572         struct btrfs_caching_control *caching_ctl;
 573         int ret = 0;
 574
 575         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 576         if (!caching_ctl)
 577                 return -ENOMEM;
 578
 579         INIT_LIST_HEAD(&caching_ctl->list);
 580         mutex_init(&caching_ctl->mutex);
 581         init_waitqueue_head(&caching_ctl->wait);
 582         caching_ctl->block_group = cache;
 583         caching_ctl->progress = cache->key.objectid;
 584         refcount_set(&caching_ctl->count, 1);
 585         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 586                         caching_thread, NULL, NULL);
 587
 588         spin_lock(&cache->lock);
 589         /*
 590          * This should be a rare occasion, but this could happen I think in the
 591          * case where one thread starts to load the space cache info, and then
 592          * some other thread starts a transaction commit which tries to do an
 593          * allocation while the other thread is still loading the space cache
 594          * info.  The previous loop should have kept us from choosing this block
 595          * group, but if we've moved to the state where we will wait on caching
 596          * block groups we need to first check if we're doing a fast load here,
 597          * so we can wait for it to finish, otherwise we could end up allocating
 598          * from a block group who's cache gets evicted for one reason or
 599          * another.
 600          */
 601         while (cache->cached == BTRFS_CACHE_FAST) {
 602                 struct btrfs_caching_control *ctl;
 603
 604                 ctl = cache->caching_ctl;
 605                 refcount_inc(&ctl->count);
 606                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 607                 spin_unlock(&cache->lock);
 608
 609                 schedule();
 610
 611                 finish_wait(&ctl->wait, &wait);
 612                 put_caching_control(ctl);
 613                 spin_lock(&cache->lock);
 614         }
 615
 616         if (cache->cached != BTRFS_CACHE_NO) {
 617                 spin_unlock(&cache->lock);
 618                 kfree(caching_ctl);
 619                 return 0;
 620         }
 621         WARN_ON(cache->caching_ctl);
 622         cache->caching_ctl = caching_ctl;
 623         cache->cached = BTRFS_CACHE_FAST;
 624         spin_unlock(&cache->lock);
 625
 626         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 627                 mutex_lock(&caching_ctl->mutex);
 628                 ret = load_free_space_cache(fs_info, cache);
 629
 630                 spin_lock(&cache->lock);
 631                 if (ret == 1) {
 632                         cache->caching_ctl = NULL;
 633                         cache->cached = BTRFS_CACHE_FINISHED;
 634                         cache->last_byte_to_unpin = (u64)-1;
 635                         caching_ctl->progress = (u64)-1;
 636                 } else {
 637                         if (load_cache_only) {
 638                                 cache->caching_ctl = NULL;
 639                                 cache->cached = BTRFS_CACHE_NO;
 640                         } else {
 641                                 cache->cached = BTRFS_CACHE_STARTED;
 642                                 cache->has_caching_ctl = 1;
 643                         }
 644                 }
 645                 spin_unlock(&cache->lock);
 646 #ifdef CONFIG_BTRFS_DEBUG
 647                 if (ret == 1 &&
 648                     btrfs_should_fragment_free_space(cache)) {
 649                         u64 bytes_used;
 650
 651                         spin_lock(&cache->space_info->lock);
 652                         spin_lock(&cache->lock);
 653                         bytes_used = cache->key.offset -
 654                                 btrfs_block_group_used(&cache->item);
 655                         cache->space_info->bytes_used += bytes_used >> 1;
 656                         spin_unlock(&cache->lock);
 657                         spin_unlock(&cache->space_info->lock);
 658                         fragment_free_space(cache);
 659                 }
 660 #endif
 661                 mutex_unlock(&caching_ctl->mutex);
 662
 663                 wake_up(&caching_ctl->wait);
 664                 if (ret == 1) {
 665                         put_caching_control(caching_ctl);
 666                         free_excluded_extents(cache);
 667                         return 0;
 668                 }
 669         } else {
 670                 /*
 671                  * We're either using the free space tree or no caching at all.
 672                  * Set cached to the appropriate value and wakeup any waiters.
 673                  */
 674                 spin_lock(&cache->lock);
 675                 if (load_cache_only) {
 676                         cache->caching_ctl = NULL;
 677                         cache->cached = BTRFS_CACHE_NO;
 678                 } else {
 679                         cache->cached = BTRFS_CACHE_STARTED;
 680                         cache->has_caching_ctl = 1;
 681                 }
 682                 spin_unlock(&cache->lock);
 683                 wake_up(&caching_ctl->wait);
 684         }
 685
 686         if (load_cache_only) {
 687                 put_caching_control(caching_ctl);
 688                 return 0;
 689         }
 690
 691         down_write(&fs_info->commit_root_sem);
 692         refcount_inc(&caching_ctl->count);
 693         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 694         up_write(&fs_info->commit_root_sem);
 695
 696         btrfs_get_block_group(cache);
 697
 698         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 699
 700         return ret;
 701 }
 702
 703 /*
 704  * return the block group that starts at or after bytenr
 705  */
 706 static struct btrfs_block_group_cache *
 707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 708 {
 709         return block_group_cache_tree_search(info, bytenr, 0);
 710 }
 711
 712 /*
 713  * return the block group that contains the given bytenr
 714  */
 715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 716                                                  struct btrfs_fs_info *info,
 717                                                  u64 bytenr)
 718 {
 719         return block_group_cache_tree_search(info, bytenr, 1);
 720 }
 721
 722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 723                                                   u64 flags)
 724 {
 725         struct list_head *head = &info->space_info;
 726         struct btrfs_space_info *found;
 727
 728         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 729
 730         rcu_read_lock();
 731         list_for_each_entry_rcu(found, head, list) {
 732                 if (found->flags & flags) {
 733                         rcu_read_unlock();
 734                         return found;
 735                 }
 736         }
 737         rcu_read_unlock();
 738         return NULL;
 739 }
 740
 741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 742                              bool metadata, u64 root_objectid)
 743 {
 744         struct btrfs_space_info *space_info;
 745         u64 flags;
 746
 747         if (metadata) {
 748                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 749                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 750                 else
 751                         flags = BTRFS_BLOCK_GROUP_METADATA;
 752         } else {
 753                 flags = BTRFS_BLOCK_GROUP_DATA;
 754         }
 755
 756         space_info = __find_space_info(fs_info, flags);
 757         ASSERT(space_info);
 758         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 759                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 760 }
 761
 762 /*
 763  * after adding space to the filesystem, we need to clear the full flags
 764  * on all the space infos.
 765  */
 766 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 767 {
 768         struct list_head *head = &info->space_info;
 769         struct btrfs_space_info *found;
 770
 771         rcu_read_lock();
 772         list_for_each_entry_rcu(found, head, list)
 773                 found->full = 0;
 774         rcu_read_unlock();
 775 }
 776
 777 /* simple helper to search for an existing data extent at a given offset */
 778 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 779 {
 780         int ret;
 781         struct btrfs_key key;
 782         struct btrfs_path *path;
 783
 784         path = btrfs_alloc_path();
 785         if (!path)
 786                 return -ENOMEM;
 787
 788         key.objectid = start;
 789         key.offset = len;
 790         key.type = BTRFS_EXTENT_ITEM_KEY;
 791         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 792         btrfs_free_path(path);
 793         return ret;
 794 }
 795
 796 /*
 797  * helper function to lookup reference count and flags of a tree block.
 798  *
 799  * the head node for delayed ref is used to store the sum of all the
 800  * reference count modifications queued up in the rbtree. the head
 801  * node may also store the extent flags to set. This way you can check
 802  * to see what the reference count and extent flags would be if all of
 803  * the delayed refs are not processed.
 804  */
 805 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 806                              struct btrfs_fs_info *fs_info, u64 bytenr,
 807                              u64 offset, int metadata, u64 *refs, u64 *flags)
 808 {
 809         struct btrfs_delayed_ref_head *head;
 810         struct btrfs_delayed_ref_root *delayed_refs;
 811         struct btrfs_path *path;
 812         struct btrfs_extent_item *ei;
 813         struct extent_buffer *leaf;
 814         struct btrfs_key key;
 815         u32 item_size;
 816         u64 num_refs;
 817         u64 extent_flags;
 818         int ret;
 819
 820         /*
 821          * If we don't have skinny metadata, don't bother doing anything
 822          * different
 823          */
 824         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 825                 offset = fs_info->nodesize;
 826                 metadata = 0;
 827         }
 828
 829         path = btrfs_alloc_path();
 830         if (!path)
 831                 return -ENOMEM;
 832
 833         if (!trans) {
 834                 path->skip_locking = 1;
 835                 path->search_commit_root = 1;
 836         }
 837
 838 search_again:
 839         key.objectid = bytenr;
 840         key.offset = offset;
 841         if (metadata)
 842                 key.type = BTRFS_METADATA_ITEM_KEY;
 843         else
 844                 key.type = BTRFS_EXTENT_ITEM_KEY;
 845
 846         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 847         if (ret < 0)
 848                 goto out_free;
 849
 850         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 851                 if (path->slots[0]) {
 852                         path->slots[0]--;
 853                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 854                                               path->slots[0]);
 855                         if (key.objectid == bytenr &&
 856                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 857                             key.offset == fs_info->nodesize)
 858                                 ret = 0;
 859                 }
 860         }
 861
 862         if (ret == 0) {
 863                 leaf = path->nodes[0];
 864                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 865                 if (item_size >= sizeof(*ei)) {
 866                         ei = btrfs_item_ptr(leaf, path->slots[0],
 867                                             struct btrfs_extent_item);
 868                         num_refs = btrfs_extent_refs(leaf, ei);
 869                         extent_flags = btrfs_extent_flags(leaf, ei);
 870                 } else {
 871                         ret = -EINVAL;
 872                         btrfs_print_v0_err(fs_info);
 873                         if (trans)
 874                                 btrfs_abort_transaction(trans, ret);
 875                         else
 876                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 877
 878                         goto out_free;
 879                 }
 880
 881                 BUG_ON(num_refs == 0);
 882         } else {
 883                 num_refs = 0;
 884                 extent_flags = 0;
 885                 ret = 0;
 886         }
 887
 888         if (!trans)
 889                 goto out;
 890
 891         delayed_refs = &trans->transaction->delayed_refs;
 892         spin_lock(&delayed_refs->lock);
 893         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 894         if (head) {
 895                 if (!mutex_trylock(&head->mutex)) {
 896                         refcount_inc(&head->refs);
 897                         spin_unlock(&delayed_refs->lock);
 898
 899                         btrfs_release_path(path);
 900
 901                         /*
 902                          * Mutex was contended, block until it's released and try
 903                          * again
 904                          */
 905                         mutex_lock(&head->mutex);
 906                         mutex_unlock(&head->mutex);
 907                         btrfs_put_delayed_ref_head(head);
 908                         goto search_again;
 909                 }
 910                 spin_lock(&head->lock);
 911                 if (head->extent_op && head->extent_op->update_flags)
 912                         extent_flags |= head->extent_op->flags_to_set;
 913                 else
 914                         BUG_ON(num_refs == 0);
 915
 916                 num_refs += head->ref_mod;
 917                 spin_unlock(&head->lock);
 918                 mutex_unlock(&head->mutex);
 919         }
 920         spin_unlock(&delayed_refs->lock);
 921 out:
 922         WARN_ON(num_refs == 0);
 923         if (refs)
 924                 *refs = num_refs;
 925         if (flags)
 926                 *flags = extent_flags;
 927 out_free:
 928         btrfs_free_path(path);
 929         return ret;
 930 }
 931
 932 /*
 933  * Back reference rules.  Back refs have three main goals:
 934  *
 935  * 1) differentiate between all holders of references to an extent so that
 936  *    when a reference is dropped we can make sure it was a valid reference
 937  *    before freeing the extent.
 938  *
 939  * 2) Provide enough information to quickly find the holders of an extent
 940  *    if we notice a given block is corrupted or bad.
 941  *
 942  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 943  *    maintenance.  This is actually the same as #2, but with a slightly
 944  *    different use case.
 945  *
 946  * There are two kinds of back refs. The implicit back refs is optimized
 947  * for pointers in non-shared tree blocks. For a given pointer in a block,
 948  * back refs of this kind provide information about the block's owner tree
 949  * and the pointer's key. These information allow us to find the block by
 950  * b-tree searching. The full back refs is for pointers in tree blocks not
 951  * referenced by their owner trees. The location of tree block is recorded
 952  * in the back refs. Actually the full back refs is generic, and can be
 953  * used in all cases the implicit back refs is used. The major shortcoming
 954  * of the full back refs is its overhead. Every time a tree block gets
 955  * COWed, we have to update back refs entry for all pointers in it.
 956  *
 957  * For a newly allocated tree block, we use implicit back refs for
 958  * pointers in it. This means most tree related operations only involve
 959  * implicit back refs. For a tree block created in old transaction, the
 960  * only way to drop a reference to it is COW it. So we can detect the
 961  * event that tree block loses its owner tree's reference and do the
 962  * back refs conversion.
 963  *
 964  * When a tree block is COWed through a tree, there are four cases:
 965  *
 966  * The reference count of the block is one and the tree is the block's
 967  * owner tree. Nothing to do in this case.
 968  *
 969  * The reference count of the block is one and the tree is not the
 970  * block's owner tree. In this case, full back refs is used for pointers
 971  * in the block. Remove these full back refs, add implicit back refs for
 972  * every pointers in the new block.
 973  *
 974  * The reference count of the block is greater than one and the tree is
 975  * the block's owner tree. In this case, implicit back refs is used for
 976  * pointers in the block. Add full back refs for every pointers in the
 977  * block, increase lower level extents' reference counts. The original
 978  * implicit back refs are entailed to the new block.
 979  *
 980  * The reference count of the block is greater than one and the tree is
 981  * not the block's owner tree. Add implicit back refs for every pointer in
 982  * the new block, increase lower level extents' reference count.
 983  *
 984  * Back Reference Key composing:
 985  *
 986  * The key objectid corresponds to the first byte in the extent,
 987  * The key type is used to differentiate between types of back refs.
 988  * There are different meanings of the key offset for different types
 989  * of back refs.
 990  *
 991  * File extents can be referenced by:
 992  *
 993  * - multiple snapshots, subvolumes, or different generations in one subvol
 994  * - different files inside a single subvolume
 995  * - different offsets inside a file (bookend extents in file.c)
 996  *
 997  * The extent ref structure for the implicit back refs has fields for:
 998  *
 999  * - Objectid of the subvolume root
1000  * - objectid of the file holding the reference
1001  * - original offset in the file
1002  * - how many bookend extents
1003  *
1004  * The key offset for the implicit back refs is hash of the first
1005  * three fields.
1006  *
1007  * The extent ref structure for the full back refs has field for:
1008  *
1009  * - number of pointers in the tree leaf
1010  *
1011  * The key offset for the implicit back refs is the first byte of
1012  * the tree leaf
1013  *
1014  * When a file extent is allocated, The implicit back refs is used.
1015  * the fields are filled in:
1016  *
1017  *     (root_key.objectid, inode objectid, offset in file, 1)
1018  *
1019  * When a file extent is removed file truncation, we find the
1020  * corresponding implicit back refs and check the following fields:
1021  *
1022  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1023  *
1024  * Btree extents can be referenced by:
1025  *
1026  * - Different subvolumes
1027  *
1028  * Both the implicit back refs and the full back refs for tree blocks
1029  * only consist of key. The key offset for the implicit back refs is
1030  * objectid of block's owner tree. The key offset for the full back refs
1031  * is the first byte of parent block.
1032  *
1033  * When implicit back refs is used, information about the lowest key and
1034  * level of the tree block are required. These information are stored in
1035  * tree block info structure.
1036  */
1037
1038 /*
1039  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1040  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1041  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1042  */
1043 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1044                                      struct btrfs_extent_inline_ref *iref,
1045                                      enum btrfs_inline_ref_type is_data)
1046 {
1047         int type = btrfs_extent_inline_ref_type(eb, iref);
1048         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1049
1050         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1051             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1052             type == BTRFS_SHARED_DATA_REF_KEY ||
1053             type == BTRFS_EXTENT_DATA_REF_KEY) {
1054                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1055                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1056                                 return type;
1057                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1058                                 ASSERT(eb->fs_info);
1059                                 /*
1060                                  * Every shared one has parent tree block,
1061                                  * which must be aligned to sector size.
1062                                  */
1063                                 if (offset &&
1064                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
1065                                         return type;
1066                         }
1067                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1068                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1069                                 return type;
1070                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1071                                 ASSERT(eb->fs_info);
1072                                 /*
1073                                  * Every shared one has parent tree block,
1074                                  * which must be aligned to sector size.
1075                                  */
1076                                 if (offset &&
1077                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
1078                                         return type;
1079                         }
1080                 } else {
1081                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1082                         return type;
1083                 }
1084         }
1085
1086         btrfs_print_leaf((struct extent_buffer *)eb);
1087         btrfs_err(eb->fs_info,
1088                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
1089                   eb->start, (unsigned long)iref, type);
1090         WARN_ON(1);
1091
1092         return BTRFS_REF_TYPE_INVALID;
1093 }
1094
1095 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1096 {
1097         u32 high_crc = ~(u32)0;
1098         u32 low_crc = ~(u32)0;
1099         __le64 lenum;
1100
1101         lenum = cpu_to_le64(root_objectid);
1102         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1103         lenum = cpu_to_le64(owner);
1104         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1105         lenum = cpu_to_le64(offset);
1106         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1107
1108         return ((u64)high_crc << 31) ^ (u64)low_crc;
1109 }
1110
1111 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1112                                      struct btrfs_extent_data_ref *ref)
1113 {
1114         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1115                                     btrfs_extent_data_ref_objectid(leaf, ref),
1116                                     btrfs_extent_data_ref_offset(leaf, ref));
1117 }
1118
1119 static int match_extent_data_ref(struct extent_buffer *leaf,
1120                                  struct btrfs_extent_data_ref *ref,
1121                                  u64 root_objectid, u64 owner, u64 offset)
1122 {
1123         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1124             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1125             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1126                 return 0;
1127         return 1;
1128 }
1129
1130 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1131                                            struct btrfs_path *path,
1132                                            u64 bytenr, u64 parent,
1133                                            u64 root_objectid,
1134                                            u64 owner, u64 offset)
1135 {
1136         struct btrfs_root *root = trans->fs_info->extent_root;
1137         struct btrfs_key key;
1138         struct btrfs_extent_data_ref *ref;
1139         struct extent_buffer *leaf;
1140         u32 nritems;
1141         int ret;
1142         int recow;
1143         int err = -ENOENT;
1144
1145         key.objectid = bytenr;
1146         if (parent) {
1147                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1148                 key.offset = parent;
1149         } else {
1150                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1151                 key.offset = hash_extent_data_ref(root_objectid,
1152                                                   owner, offset);
1153         }
1154 again:
1155         recow = 0;
1156         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1157         if (ret < 0) {
1158                 err = ret;
1159                 goto fail;
1160         }
1161
1162         if (parent) {
1163                 if (!ret)
1164                         return 0;
1165                 goto fail;
1166         }
1167
1168         leaf = path->nodes[0];
1169         nritems = btrfs_header_nritems(leaf);
1170         while (1) {
1171                 if (path->slots[0] >= nritems) {
1172                         ret = btrfs_next_leaf(root, path);
1173                         if (ret < 0)
1174                                 err = ret;
1175                         if (ret)
1176                                 goto fail;
1177
1178                         leaf = path->nodes[0];
1179                         nritems = btrfs_header_nritems(leaf);
1180                         recow = 1;
1181                 }
1182
1183                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1184                 if (key.objectid != bytenr ||
1185                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1186                         goto fail;
1187
1188                 ref = btrfs_item_ptr(leaf, path->slots[0],
1189                                      struct btrfs_extent_data_ref);
1190
1191                 if (match_extent_data_ref(leaf, ref, root_objectid,
1192                                           owner, offset)) {
1193                         if (recow) {
1194                                 btrfs_release_path(path);
1195                                 goto again;
1196                         }
1197                         err = 0;
1198                         break;
1199                 }
1200                 path->slots[0]++;
1201         }
1202 fail:
1203         return err;
1204 }
1205
1206 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1207                                            struct btrfs_path *path,
1208                                            u64 bytenr, u64 parent,
1209                                            u64 root_objectid, u64 owner,
1210                                            u64 offset, int refs_to_add)
1211 {
1212         struct btrfs_root *root = trans->fs_info->extent_root;
1213         struct btrfs_key key;
1214         struct extent_buffer *leaf;
1215         u32 size;
1216         u32 num_refs;
1217         int ret;
1218
1219         key.objectid = bytenr;
1220         if (parent) {
1221                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1222                 key.offset = parent;
1223                 size = sizeof(struct btrfs_shared_data_ref);
1224         } else {
1225                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1226                 key.offset = hash_extent_data_ref(root_objectid,
1227                                                   owner, offset);
1228                 size = sizeof(struct btrfs_extent_data_ref);
1229         }
1230
1231         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1232         if (ret && ret != -EEXIST)
1233                 goto fail;
1234
1235         leaf = path->nodes[0];
1236         if (parent) {
1237                 struct btrfs_shared_data_ref *ref;
1238                 ref = btrfs_item_ptr(leaf, path->slots[0],
1239                                      struct btrfs_shared_data_ref);
1240                 if (ret == 0) {
1241                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1242                 } else {
1243                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1244                         num_refs += refs_to_add;
1245                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1246                 }
1247         } else {
1248                 struct btrfs_extent_data_ref *ref;
1249                 while (ret == -EEXIST) {
1250                         ref = btrfs_item_ptr(leaf, path->slots[0],
1251                                              struct btrfs_extent_data_ref);
1252                         if (match_extent_data_ref(leaf, ref, root_objectid,
1253                                                   owner, offset))
1254                                 break;
1255                         btrfs_release_path(path);
1256                         key.offset++;
1257                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1258                                                       size);
1259                         if (ret && ret != -EEXIST)
1260                                 goto fail;
1261
1262                         leaf = path->nodes[0];
1263                 }
1264                 ref = btrfs_item_ptr(leaf, path->slots[0],
1265                                      struct btrfs_extent_data_ref);
1266                 if (ret == 0) {
1267                         btrfs_set_extent_data_ref_root(leaf, ref,
1268                                                        root_objectid);
1269                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1270                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1271                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1272                 } else {
1273                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1274                         num_refs += refs_to_add;
1275                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1276                 }
1277         }
1278         btrfs_mark_buffer_dirty(leaf);
1279         ret = 0;
1280 fail:
1281         btrfs_release_path(path);
1282         return ret;
1283 }
1284
1285 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1286                                            struct btrfs_path *path,
1287                                            int refs_to_drop, int *last_ref)
1288 {
1289         struct btrfs_key key;
1290         struct btrfs_extent_data_ref *ref1 = NULL;
1291         struct btrfs_shared_data_ref *ref2 = NULL;
1292         struct extent_buffer *leaf;
1293         u32 num_refs = 0;
1294         int ret = 0;
1295
1296         leaf = path->nodes[0];
1297         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1298
1299         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1300                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1301                                       struct btrfs_extent_data_ref);
1302                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1303         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1304                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1305                                       struct btrfs_shared_data_ref);
1306                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1307         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1308                 btrfs_print_v0_err(trans->fs_info);
1309                 btrfs_abort_transaction(trans, -EINVAL);
1310                 return -EINVAL;
1311         } else {
1312                 BUG();
1313         }
1314
1315         BUG_ON(num_refs < refs_to_drop);
1316         num_refs -= refs_to_drop;
1317
1318         if (num_refs == 0) {
1319                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1320                 *last_ref = 1;
1321         } else {
1322                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1323                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1324                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1325                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1326                 btrfs_mark_buffer_dirty(leaf);
1327         }
1328         return ret;
1329 }
1330
1331 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1332                                           struct btrfs_extent_inline_ref *iref)
1333 {
1334         struct btrfs_key key;
1335         struct extent_buffer *leaf;
1336         struct btrfs_extent_data_ref *ref1;
1337         struct btrfs_shared_data_ref *ref2;
1338         u32 num_refs = 0;
1339         int type;
1340
1341         leaf = path->nodes[0];
1342         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1343
1344         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1345         if (iref) {
1346                 /*
1347                  * If type is invalid, we should have bailed out earlier than
1348                  * this call.
1349                  */
1350                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1351                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1352                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1353                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1354                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1355                 } else {
1356                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1357                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1358                 }
1359         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1360                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1361                                       struct btrfs_extent_data_ref);
1362                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1363         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1364                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1365                                       struct btrfs_shared_data_ref);
1366                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1367         } else {
1368                 WARN_ON(1);
1369         }
1370         return num_refs;
1371 }
1372
1373 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1374                                           struct btrfs_path *path,
1375                                           u64 bytenr, u64 parent,
1376                                           u64 root_objectid)
1377 {
1378         struct btrfs_root *root = trans->fs_info->extent_root;
1379         struct btrfs_key key;
1380         int ret;
1381
1382         key.objectid = bytenr;
1383         if (parent) {
1384                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1385                 key.offset = parent;
1386         } else {
1387                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1388                 key.offset = root_objectid;
1389         }
1390
1391         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1392         if (ret > 0)
1393                 ret = -ENOENT;
1394         return ret;
1395 }
1396
1397 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1398                                           struct btrfs_path *path,
1399                                           u64 bytenr, u64 parent,
1400                                           u64 root_objectid)
1401 {
1402         struct btrfs_key key;
1403         int ret;
1404
1405         key.objectid = bytenr;
1406         if (parent) {
1407                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1408                 key.offset = parent;
1409         } else {
1410                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1411                 key.offset = root_objectid;
1412         }
1413
1414         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1415                                       path, &key, 0);
1416         btrfs_release_path(path);
1417         return ret;
1418 }
1419
1420 static inline int extent_ref_type(u64 parent, u64 owner)
1421 {
1422         int type;
1423         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1424                 if (parent > 0)
1425                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1426                 else
1427                         type = BTRFS_TREE_BLOCK_REF_KEY;
1428         } else {
1429                 if (parent > 0)
1430                         type = BTRFS_SHARED_DATA_REF_KEY;
1431                 else
1432                         type = BTRFS_EXTENT_DATA_REF_KEY;
1433         }
1434         return type;
1435 }
1436
1437 static int find_next_key(struct btrfs_path *path, int level,
1438                          struct btrfs_key *key)
1439
1440 {
1441         for (; level < BTRFS_MAX_LEVEL; level++) {
1442                 if (!path->nodes[level])
1443                         break;
1444                 if (path->slots[level] + 1 >=
1445                     btrfs_header_nritems(path->nodes[level]))
1446                         continue;
1447                 if (level == 0)
1448                         btrfs_item_key_to_cpu(path->nodes[level], key,
1449                                               path->slots[level] + 1);
1450                 else
1451                         btrfs_node_key_to_cpu(path->nodes[level], key,
1452                                               path->slots[level] + 1);
1453                 return 0;
1454         }
1455         return 1;
1456 }
1457
1458 /*
1459  * look for inline back ref. if back ref is found, *ref_ret is set
1460  * to the address of inline back ref, and 0 is returned.
1461  *
1462  * if back ref isn't found, *ref_ret is set to the address where it
1463  * should be inserted, and -ENOENT is returned.
1464  *
1465  * if insert is true and there are too many inline back refs, the path
1466  * points to the extent item, and -EAGAIN is returned.
1467  *
1468  * NOTE: inline back refs are ordered in the same way that back ref
1469  *       items in the tree are ordered.
1470  */
1471 static noinline_for_stack
1472 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1473                                  struct btrfs_path *path,
1474                                  struct btrfs_extent_inline_ref **ref_ret,
1475                                  u64 bytenr, u64 num_bytes,
1476                                  u64 parent, u64 root_objectid,
1477                                  u64 owner, u64 offset, int insert)
1478 {
1479         struct btrfs_fs_info *fs_info = trans->fs_info;
1480         struct btrfs_root *root = fs_info->extent_root;
1481         struct btrfs_key key;
1482         struct extent_buffer *leaf;
1483         struct btrfs_extent_item *ei;
1484         struct btrfs_extent_inline_ref *iref;
1485         u64 flags;
1486         u64 item_size;
1487         unsigned long ptr;
1488         unsigned long end;
1489         int extra_size;
1490         int type;
1491         int want;
1492         int ret;
1493         int err = 0;
1494         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1495         int needed;
1496
1497         key.objectid = bytenr;
1498         key.type = BTRFS_EXTENT_ITEM_KEY;
1499         key.offset = num_bytes;
1500
1501         want = extent_ref_type(parent, owner);
1502         if (insert) {
1503                 extra_size = btrfs_extent_inline_ref_size(want);
1504                 path->keep_locks = 1;
1505         } else
1506                 extra_size = -1;
1507
1508         /*
1509          * Owner is our level, so we can just add one to get the level for the
1510          * block we are interested in.
1511          */
1512         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1513                 key.type = BTRFS_METADATA_ITEM_KEY;
1514                 key.offset = owner;
1515         }
1516
1517 again:
1518         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1519         if (ret < 0) {
1520                 err = ret;
1521                 goto out;
1522         }
1523
1524         /*
1525          * We may be a newly converted file system which still has the old fat
1526          * extent entries for metadata, so try and see if we have one of those.
1527          */
1528         if (ret > 0 && skinny_metadata) {
1529                 skinny_metadata = false;
1530                 if (path->slots[0]) {
1531                         path->slots[0]--;
1532                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1533                                               path->slots[0]);
1534                         if (key.objectid == bytenr &&
1535                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1536                             key.offset == num_bytes)
1537                                 ret = 0;
1538                 }
1539                 if (ret) {
1540                         key.objectid = bytenr;
1541                         key.type = BTRFS_EXTENT_ITEM_KEY;
1542                         key.offset = num_bytes;
1543                         btrfs_release_path(path);
1544                         goto again;
1545                 }
1546         }
1547
1548         if (ret && !insert) {
1549                 err = -ENOENT;
1550                 goto out;
1551         } else if (WARN_ON(ret)) {
1552                 err = -EIO;
1553                 goto out;
1554         }
1555
1556         leaf = path->nodes[0];
1557         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1558         if (unlikely(item_size < sizeof(*ei))) {
1559                 err = -EINVAL;
1560                 btrfs_print_v0_err(fs_info);
1561                 btrfs_abort_transaction(trans, err);
1562                 goto out;
1563         }
1564
1565         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1566         flags = btrfs_extent_flags(leaf, ei);
1567
1568         ptr = (unsigned long)(ei + 1);
1569         end = (unsigned long)ei + item_size;
1570
1571         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1572                 ptr += sizeof(struct btrfs_tree_block_info);
1573                 BUG_ON(ptr > end);
1574         }
1575
1576         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1577                 needed = BTRFS_REF_TYPE_DATA;
1578         else
1579                 needed = BTRFS_REF_TYPE_BLOCK;
1580
1581         err = -ENOENT;
1582         while (1) {
1583                 if (ptr >= end) {
1584                         WARN_ON(ptr > end);
1585                         break;
1586                 }
1587                 iref = (struct btrfs_extent_inline_ref *)ptr;
1588                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1589                 if (type == BTRFS_REF_TYPE_INVALID) {
1590                         err = -EUCLEAN;
1591                         goto out;
1592                 }
1593
1594                 if (want < type)
1595                         break;
1596                 if (want > type) {
1597                         ptr += btrfs_extent_inline_ref_size(type);
1598                         continue;
1599                 }
1600
1601                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1602                         struct btrfs_extent_data_ref *dref;
1603                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1604                         if (match_extent_data_ref(leaf, dref, root_objectid,
1605                                                   owner, offset)) {
1606                                 err = 0;
1607                                 break;
1608                         }
1609                         if (hash_extent_data_ref_item(leaf, dref) <
1610                             hash_extent_data_ref(root_objectid, owner, offset))
1611                                 break;
1612                 } else {
1613                         u64 ref_offset;
1614                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1615                         if (parent > 0) {
1616                                 if (parent == ref_offset) {
1617                                         err = 0;
1618                                         break;
1619                                 }
1620                                 if (ref_offset < parent)
1621                                         break;
1622                         } else {
1623                                 if (root_objectid == ref_offset) {
1624                                         err = 0;
1625                                         break;
1626                                 }
1627                                 if (ref_offset < root_objectid)
1628                                         break;
1629                         }
1630                 }
1631                 ptr += btrfs_extent_inline_ref_size(type);
1632         }
1633         if (err == -ENOENT && insert) {
1634                 if (item_size + extra_size >=
1635                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1636                         err = -EAGAIN;
1637                         goto out;
1638                 }
1639                 /*
1640                  * To add new inline back ref, we have to make sure
1641                  * there is no corresponding back ref item.
1642                  * For simplicity, we just do not add new inline back
1643                  * ref if there is any kind of item for this block
1644                  */
1645                 if (find_next_key(path, 0, &key) == 0 &&
1646                     key.objectid == bytenr &&
1647                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1648                         err = -EAGAIN;
1649                         goto out;
1650                 }
1651         }
1652         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1653 out:
1654         if (insert) {
1655                 path->keep_locks = 0;
1656                 btrfs_unlock_up_safe(path, 1);
1657         }
1658         return err;
1659 }
1660
1661 /*
1662  * helper to add new inline back ref
1663  */
1664 static noinline_for_stack
1665 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1666                                  struct btrfs_path *path,
1667                                  struct btrfs_extent_inline_ref *iref,
1668                                  u64 parent, u64 root_objectid,
1669                                  u64 owner, u64 offset, int refs_to_add,
1670                                  struct btrfs_delayed_extent_op *extent_op)
1671 {
1672         struct extent_buffer *leaf;
1673         struct btrfs_extent_item *ei;
1674         unsigned long ptr;
1675         unsigned long end;
1676         unsigned long item_offset;
1677         u64 refs;
1678         int size;
1679         int type;
1680
1681         leaf = path->nodes[0];
1682         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1683         item_offset = (unsigned long)iref - (unsigned long)ei;
1684
1685         type = extent_ref_type(parent, owner);
1686         size = btrfs_extent_inline_ref_size(type);
1687
1688         btrfs_extend_item(fs_info, path, size);
1689
1690         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1691         refs = btrfs_extent_refs(leaf, ei);
1692         refs += refs_to_add;
1693         btrfs_set_extent_refs(leaf, ei, refs);
1694         if (extent_op)
1695                 __run_delayed_extent_op(extent_op, leaf, ei);
1696
1697         ptr = (unsigned long)ei + item_offset;
1698         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1699         if (ptr < end - size)
1700                 memmove_extent_buffer(leaf, ptr + size, ptr,
1701                                       end - size - ptr);
1702
1703         iref = (struct btrfs_extent_inline_ref *)ptr;
1704         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1705         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1706                 struct btrfs_extent_data_ref *dref;
1707                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1708                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1709                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1710                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1711                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1712         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1713                 struct btrfs_shared_data_ref *sref;
1714                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1715                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1716                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1717         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1718                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1719         } else {
1720                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1721         }
1722         btrfs_mark_buffer_dirty(leaf);
1723 }
1724
1725 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1726                                  struct btrfs_path *path,
1727                                  struct btrfs_extent_inline_ref **ref_ret,
1728                                  u64 bytenr, u64 num_bytes, u64 parent,
1729                                  u64 root_objectid, u64 owner, u64 offset)
1730 {
1731         int ret;
1732
1733         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1734                                            num_bytes, parent, root_objectid,
1735                                            owner, offset, 0);
1736         if (ret != -ENOENT)
1737                 return ret;
1738
1739         btrfs_release_path(path);
1740         *ref_ret = NULL;
1741
1742         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1743                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1744                                             root_objectid);
1745         } else {
1746                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1747                                              root_objectid, owner, offset);
1748         }
1749         return ret;
1750 }
1751
1752 /*
1753  * helper to update/remove inline back ref
1754  */
1755 static noinline_for_stack
1756 void update_inline_extent_backref(struct btrfs_path *path,
1757                                   struct btrfs_extent_inline_ref *iref,
1758                                   int refs_to_mod,
1759                                   struct btrfs_delayed_extent_op *extent_op,
1760                                   int *last_ref)
1761 {
1762         struct extent_buffer *leaf = path->nodes[0];
1763         struct btrfs_fs_info *fs_info = leaf->fs_info;
1764         struct btrfs_extent_item *ei;
1765         struct btrfs_extent_data_ref *dref = NULL;
1766         struct btrfs_shared_data_ref *sref = NULL;
1767         unsigned long ptr;
1768         unsigned long end;
1769         u32 item_size;
1770         int size;
1771         int type;
1772         u64 refs;
1773
1774         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1775         refs = btrfs_extent_refs(leaf, ei);
1776         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1777         refs += refs_to_mod;
1778         btrfs_set_extent_refs(leaf, ei, refs);
1779         if (extent_op)
1780                 __run_delayed_extent_op(extent_op, leaf, ei);
1781
1782         /*
1783          * If type is invalid, we should have bailed out after
1784          * lookup_inline_extent_backref().
1785          */
1786         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1787         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1788
1789         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1790                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1791                 refs = btrfs_extent_data_ref_count(leaf, dref);
1792         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1793                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1794                 refs = btrfs_shared_data_ref_count(leaf, sref);
1795         } else {
1796                 refs = 1;
1797                 BUG_ON(refs_to_mod != -1);
1798         }
1799
1800         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1801         refs += refs_to_mod;
1802
1803         if (refs > 0) {
1804                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1805                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1806                 else
1807                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1808         } else {
1809                 *last_ref = 1;
1810                 size =  btrfs_extent_inline_ref_size(type);
1811                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1812                 ptr = (unsigned long)iref;
1813                 end = (unsigned long)ei + item_size;
1814                 if (ptr + size < end)
1815                         memmove_extent_buffer(leaf, ptr, ptr + size,
1816                                               end - ptr - size);
1817                 item_size -= size;
1818                 btrfs_truncate_item(fs_info, path, item_size, 1);
1819         }
1820         btrfs_mark_buffer_dirty(leaf);
1821 }
1822
1823 static noinline_for_stack
1824 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1825                                  struct btrfs_path *path,
1826                                  u64 bytenr, u64 num_bytes, u64 parent,
1827                                  u64 root_objectid, u64 owner,
1828                                  u64 offset, int refs_to_add,
1829                                  struct btrfs_delayed_extent_op *extent_op)
1830 {
1831         struct btrfs_extent_inline_ref *iref;
1832         int ret;
1833
1834         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1835                                            num_bytes, parent, root_objectid,
1836                                            owner, offset, 1);
1837         if (ret == 0) {
1838                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1839                 update_inline_extent_backref(path, iref, refs_to_add,
1840                                              extent_op, NULL);
1841         } else if (ret == -ENOENT) {
1842                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1843                                             root_objectid, owner, offset,
1844                                             refs_to_add, extent_op);
1845                 ret = 0;
1846         }
1847         return ret;
1848 }
1849
1850 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1851                                  struct btrfs_path *path,
1852                                  u64 bytenr, u64 parent, u64 root_objectid,
1853                                  u64 owner, u64 offset, int refs_to_add)
1854 {
1855         int ret;
1856         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1857                 BUG_ON(refs_to_add != 1);
1858                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1859                                             root_objectid);
1860         } else {
1861                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1862                                              root_objectid, owner, offset,
1863                                              refs_to_add);
1864         }
1865         return ret;
1866 }
1867
1868 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1869                                  struct btrfs_path *path,
1870                                  struct btrfs_extent_inline_ref *iref,
1871                                  int refs_to_drop, int is_data, int *last_ref)
1872 {
1873         int ret = 0;
1874
1875         BUG_ON(!is_data && refs_to_drop != 1);
1876         if (iref) {
1877                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1878                                              last_ref);
1879         } else if (is_data) {
1880                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1881                                              last_ref);
1882         } else {
1883                 *last_ref = 1;
1884                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1885         }
1886         return ret;
1887 }
1888
1889 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1890 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1891                                u64 *discarded_bytes)
1892 {
1893         int j, ret = 0;
1894         u64 bytes_left, end;
1895         u64 aligned_start = ALIGN(start, 1 << 9);
1896
1897         if (WARN_ON(start != aligned_start)) {
1898                 len -= aligned_start - start;
1899                 len = round_down(len, 1 << 9);
1900                 start = aligned_start;
1901         }
1902
1903         *discarded_bytes = 0;
1904
1905         if (!len)
1906                 return 0;
1907
1908         end = start + len;
1909         bytes_left = len;
1910
1911         /* Skip any superblocks on this device. */
1912         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1913                 u64 sb_start = btrfs_sb_offset(j);
1914                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1915                 u64 size = sb_start - start;
1916
1917                 if (!in_range(sb_start, start, bytes_left) &&
1918                     !in_range(sb_end, start, bytes_left) &&
1919                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1920                         continue;
1921
1922                 /*
1923                  * Superblock spans beginning of range.  Adjust start and
1924                  * try again.
1925                  */
1926                 if (sb_start <= start) {
1927                         start += sb_end - start;
1928                         if (start > end) {
1929                                 bytes_left = 0;
1930                                 break;
1931                         }
1932                         bytes_left = end - start;
1933                         continue;
1934                 }
1935
1936                 if (size) {
1937                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1938                                                    GFP_NOFS, 0);
1939                         if (!ret)
1940                                 *discarded_bytes += size;
1941                         else if (ret != -EOPNOTSUPP)
1942                                 return ret;
1943                 }
1944
1945                 start = sb_end;
1946                 if (start > end) {
1947                         bytes_left = 0;
1948                         break;
1949                 }
1950                 bytes_left = end - start;
1951         }
1952
1953         if (bytes_left) {
1954                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1955                                            GFP_NOFS, 0);
1956                 if (!ret)
1957                         *discarded_bytes += bytes_left;
1958         }
1959         return ret;
1960 }
1961
1962 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1963                          u64 num_bytes, u64 *actual_bytes)
1964 {
1965         int ret;
1966         u64 discarded_bytes = 0;
1967         struct btrfs_bio *bbio = NULL;
1968
1969
1970         /*
1971          * Avoid races with device replace and make sure our bbio has devices
1972          * associated to its stripes that don't go away while we are discarding.
1973          */
1974         btrfs_bio_counter_inc_blocked(fs_info);
1975         /* Tell the block device(s) that the sectors can be discarded */
1976         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1977                               &bbio, 0);
1978         /* Error condition is -ENOMEM */
1979         if (!ret) {
1980                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1981                 int i;
1982
1983
1984                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1985                         u64 bytes;
1986                         struct request_queue *req_q;
1987                         struct btrfs_device *device = stripe->dev;
1988
1989                         if (!device->bdev) {
1990                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1991                                 continue;
1992                         }
1993                         req_q = bdev_get_queue(device->bdev);
1994                         if (!blk_queue_discard(req_q))
1995                                 continue;
1996
1997                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
1998                                 continue;
1999
2000                         ret = btrfs_issue_discard(device->bdev,
2001                                                   stripe->physical,
2002                                                   stripe->length,
2003                                                   &bytes);
2004                         if (!ret)
2005                                 discarded_bytes += bytes;
2006                         else if (ret != -EOPNOTSUPP)
2007                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2008
2009                         /*
2010                          * Just in case we get back EOPNOTSUPP for some reason,
2011                          * just ignore the return value so we don't screw up
2012                          * people calling discard_extent.
2013                          */
2014                         ret = 0;
2015                 }
2016                 btrfs_put_bbio(bbio);
2017         }
2018         btrfs_bio_counter_dec(fs_info);
2019
2020         if (actual_bytes)
2021                 *actual_bytes = discarded_bytes;
2022
2023
2024         if (ret == -EOPNOTSUPP)
2025                 ret = 0;
2026         return ret;
2027 }
2028
2029 /* Can return -ENOMEM */
2030 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2031                          struct btrfs_root *root,
2032                          u64 bytenr, u64 num_bytes, u64 parent,
2033                          u64 root_objectid, u64 owner, u64 offset)
2034 {
2035         struct btrfs_fs_info *fs_info = root->fs_info;
2036         int old_ref_mod, new_ref_mod;
2037         int ret;
2038
2039         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2040                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2041
2042         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2043                            owner, offset, BTRFS_ADD_DELAYED_REF);
2044
2045         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2046                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2047                                                  num_bytes, parent,
2048                                                  root_objectid, (int)owner,
2049                                                  BTRFS_ADD_DELAYED_REF, NULL,
2050                                                  &old_ref_mod, &new_ref_mod);
2051         } else {
2052                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2053                                                  num_bytes, parent,
2054                                                  root_objectid, owner, offset,
2055                                                  0, BTRFS_ADD_DELAYED_REF,
2056                                                  &old_ref_mod, &new_ref_mod);
2057         }
2058
2059         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2060                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2061
2062                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2063         }
2064
2065         return ret;
2066 }
2067
2068 /*
2069  * __btrfs_inc_extent_ref - insert backreference for a given extent
2070  *
2071  * @trans:          Handle of transaction
2072  *
2073  * @node:           The delayed ref node used to get the bytenr/length for
2074  *                  extent whose references are incremented.
2075  *
2076  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2077  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2078  *                  bytenr of the parent block. Since new extents are always
2079  *                  created with indirect references, this will only be the case
2080  *                  when relocating a shared extent. In that case, root_objectid
2081  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2082  *                  be 0
2083  *
2084  * @root_objectid:  The id of the root where this modification has originated,
2085  *                  this can be either one of the well-known metadata trees or
2086  *                  the subvolume id which references this extent.
2087  *
2088  * @owner:          For data extents it is the inode number of the owning file.
2089  *                  For metadata extents this parameter holds the level in the
2090  *                  tree of the extent.
2091  *
2092  * @offset:         For metadata extents the offset is ignored and is currently
2093  *                  always passed as 0. For data extents it is the fileoffset
2094  *                  this extent belongs to.
2095  *
2096  * @refs_to_add     Number of references to add
2097  *
2098  * @extent_op       Pointer to a structure, holding information necessary when
2099  *                  updating a tree block's flags
2100  *
2101  */
2102 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2103                                   struct btrfs_delayed_ref_node *node,
2104                                   u64 parent, u64 root_objectid,
2105                                   u64 owner, u64 offset, int refs_to_add,
2106                                   struct btrfs_delayed_extent_op *extent_op)
2107 {
2108         struct btrfs_path *path;
2109         struct extent_buffer *leaf;
2110         struct btrfs_extent_item *item;
2111         struct btrfs_key key;
2112         u64 bytenr = node->bytenr;
2113         u64 num_bytes = node->num_bytes;
2114         u64 refs;
2115         int ret;
2116
2117         path = btrfs_alloc_path();
2118         if (!path)
2119                 return -ENOMEM;
2120
2121         path->reada = READA_FORWARD;
2122         path->leave_spinning = 1;
2123         /* this will setup the path even if it fails to insert the back ref */
2124         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2125                                            parent, root_objectid, owner,
2126                                            offset, refs_to_add, extent_op);
2127         if ((ret < 0 && ret != -EAGAIN) || !ret)
2128                 goto out;
2129
2130         /*
2131          * Ok we had -EAGAIN which means we didn't have space to insert and
2132          * inline extent ref, so just update the reference count and add a
2133          * normal backref.
2134          */
2135         leaf = path->nodes[0];
2136         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2137         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2138         refs = btrfs_extent_refs(leaf, item);
2139         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2140         if (extent_op)
2141                 __run_delayed_extent_op(extent_op, leaf, item);
2142
2143         btrfs_mark_buffer_dirty(leaf);
2144         btrfs_release_path(path);
2145
2146         path->reada = READA_FORWARD;
2147         path->leave_spinning = 1;
2148         /* now insert the actual backref */
2149         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2150                                     owner, offset, refs_to_add);
2151         if (ret)
2152                 btrfs_abort_transaction(trans, ret);
2153 out:
2154         btrfs_free_path(path);
2155         return ret;
2156 }
2157
2158 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2159                                 struct btrfs_delayed_ref_node *node,
2160                                 struct btrfs_delayed_extent_op *extent_op,
2161                                 int insert_reserved)
2162 {
2163         int ret = 0;
2164         struct btrfs_delayed_data_ref *ref;
2165         struct btrfs_key ins;
2166         u64 parent = 0;
2167         u64 ref_root = 0;
2168         u64 flags = 0;
2169
2170         ins.objectid = node->bytenr;
2171         ins.offset = node->num_bytes;
2172         ins.type = BTRFS_EXTENT_ITEM_KEY;
2173
2174         ref = btrfs_delayed_node_to_data_ref(node);
2175         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2176
2177         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2178                 parent = ref->parent;
2179         ref_root = ref->root;
2180
2181         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2182                 if (extent_op)
2183                         flags |= extent_op->flags_to_set;
2184                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2185                                                  flags, ref->objectid,
2186                                                  ref->offset, &ins,
2187                                                  node->ref_mod);
2188         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2189                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2190                                              ref->objectid, ref->offset,
2191                                              node->ref_mod, extent_op);
2192         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2193                 ret = __btrfs_free_extent(trans, node, parent,
2194                                           ref_root, ref->objectid,
2195                                           ref->offset, node->ref_mod,
2196                                           extent_op);
2197         } else {
2198                 BUG();
2199         }
2200         return ret;
2201 }
2202
2203 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2204                                     struct extent_buffer *leaf,
2205                                     struct btrfs_extent_item *ei)
2206 {
2207         u64 flags = btrfs_extent_flags(leaf, ei);
2208         if (extent_op->update_flags) {
2209                 flags |= extent_op->flags_to_set;
2210                 btrfs_set_extent_flags(leaf, ei, flags);
2211         }
2212
2213         if (extent_op->update_key) {
2214                 struct btrfs_tree_block_info *bi;
2215                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2216                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2217                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2218         }
2219 }
2220
2221 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2222                                  struct btrfs_delayed_ref_head *head,
2223                                  struct btrfs_delayed_extent_op *extent_op)
2224 {
2225         struct btrfs_fs_info *fs_info = trans->fs_info;
2226         struct btrfs_key key;
2227         struct btrfs_path *path;
2228         struct btrfs_extent_item *ei;
2229         struct extent_buffer *leaf;
2230         u32 item_size;
2231         int ret;
2232         int err = 0;
2233         int metadata = !extent_op->is_data;
2234
2235         if (trans->aborted)
2236                 return 0;
2237
2238         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2239                 metadata = 0;
2240
2241         path = btrfs_alloc_path();
2242         if (!path)
2243                 return -ENOMEM;
2244
2245         key.objectid = head->bytenr;
2246
2247         if (metadata) {
2248                 key.type = BTRFS_METADATA_ITEM_KEY;
2249                 key.offset = extent_op->level;
2250         } else {
2251                 key.type = BTRFS_EXTENT_ITEM_KEY;
2252                 key.offset = head->num_bytes;
2253         }
2254
2255 again:
2256         path->reada = READA_FORWARD;
2257         path->leave_spinning = 1;
2258         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2259         if (ret < 0) {
2260                 err = ret;
2261                 goto out;
2262         }
2263         if (ret > 0) {
2264                 if (metadata) {
2265                         if (path->slots[0] > 0) {
2266                                 path->slots[0]--;
2267                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2268                                                       path->slots[0]);
2269                                 if (key.objectid == head->bytenr &&
2270                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2271                                     key.offset == head->num_bytes)
2272                                         ret = 0;
2273                         }
2274                         if (ret > 0) {
2275                                 btrfs_release_path(path);
2276                                 metadata = 0;
2277
2278                                 key.objectid = head->bytenr;
2279                                 key.offset = head->num_bytes;
2280                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2281                                 goto again;
2282                         }
2283                 } else {
2284                         err = -EIO;
2285                         goto out;
2286                 }
2287         }
2288
2289         leaf = path->nodes[0];
2290         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2291
2292         if (unlikely(item_size < sizeof(*ei))) {
2293                 err = -EINVAL;
2294                 btrfs_print_v0_err(fs_info);
2295                 btrfs_abort_transaction(trans, err);
2296                 goto out;
2297         }
2298
2299         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2300         __run_delayed_extent_op(extent_op, leaf, ei);
2301
2302         btrfs_mark_buffer_dirty(leaf);
2303 out:
2304         btrfs_free_path(path);
2305         return err;
2306 }
2307
2308 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2309                                 struct btrfs_delayed_ref_node *node,
2310                                 struct btrfs_delayed_extent_op *extent_op,
2311                                 int insert_reserved)
2312 {
2313         int ret = 0;
2314         struct btrfs_delayed_tree_ref *ref;
2315         u64 parent = 0;
2316         u64 ref_root = 0;
2317
2318         ref = btrfs_delayed_node_to_tree_ref(node);
2319         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2320
2321         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2322                 parent = ref->parent;
2323         ref_root = ref->root;
2324
2325         if (node->ref_mod != 1) {
2326                 btrfs_err(trans->fs_info,
2327         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2328                           node->bytenr, node->ref_mod, node->action, ref_root,
2329                           parent);
2330                 return -EIO;
2331         }
2332         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2333                 BUG_ON(!extent_op || !extent_op->update_flags);
2334                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2335         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2336                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2337                                              ref->level, 0, 1, extent_op);
2338         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2339                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2340                                           ref->level, 0, 1, extent_op);
2341         } else {
2342                 BUG();
2343         }
2344         return ret;
2345 }
2346
2347 /* helper function to actually process a single delayed ref entry */
2348 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2349                                struct btrfs_delayed_ref_node *node,
2350                                struct btrfs_delayed_extent_op *extent_op,
2351                                int insert_reserved)
2352 {
2353         int ret = 0;
2354
2355         if (trans->aborted) {
2356                 if (insert_reserved)
2357                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2358                                          node->num_bytes, 1);
2359                 return 0;
2360         }
2361
2362         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2363             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2364                 ret = run_delayed_tree_ref(trans, node, extent_op,
2365                                            insert_reserved);
2366         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2367                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2368                 ret = run_delayed_data_ref(trans, node, extent_op,
2369                                            insert_reserved);
2370         else
2371                 BUG();
2372         if (ret && insert_reserved)
2373                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2374                                  node->num_bytes, 1);
2375         return ret;
2376 }
2377
2378 static inline struct btrfs_delayed_ref_node *
2379 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2380 {
2381         struct btrfs_delayed_ref_node *ref;
2382
2383         if (RB_EMPTY_ROOT(&head->ref_tree))
2384                 return NULL;
2385
2386         /*
2387          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2388          * This is to prevent a ref count from going down to zero, which deletes
2389          * the extent item from the extent tree, when there still are references
2390          * to add, which would fail because they would not find the extent item.
2391          */
2392         if (!list_empty(&head->ref_add_list))
2393                 return list_first_entry(&head->ref_add_list,
2394                                 struct btrfs_delayed_ref_node, add_list);
2395
2396         ref = rb_entry(rb_first(&head->ref_tree),
2397                        struct btrfs_delayed_ref_node, ref_node);
2398         ASSERT(list_empty(&ref->add_list));
2399         return ref;
2400 }
2401
2402 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2403                                       struct btrfs_delayed_ref_head *head)
2404 {
2405         spin_lock(&delayed_refs->lock);
2406         head->processing = 0;
2407         delayed_refs->num_heads_ready++;
2408         spin_unlock(&delayed_refs->lock);
2409         btrfs_delayed_ref_unlock(head);
2410 }
2411
2412 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2413                              struct btrfs_delayed_ref_head *head)
2414 {
2415         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2416         int ret;
2417
2418         if (!extent_op)
2419                 return 0;
2420         head->extent_op = NULL;
2421         if (head->must_insert_reserved) {
2422                 btrfs_free_delayed_extent_op(extent_op);
2423                 return 0;
2424         }
2425         spin_unlock(&head->lock);
2426         ret = run_delayed_extent_op(trans, head, extent_op);
2427         btrfs_free_delayed_extent_op(extent_op);
2428         return ret ? ret : 1;
2429 }
2430
2431 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2432                             struct btrfs_delayed_ref_head *head)
2433 {
2434
2435         struct btrfs_fs_info *fs_info = trans->fs_info;
2436         struct btrfs_delayed_ref_root *delayed_refs;
2437         int ret;
2438
2439         delayed_refs = &trans->transaction->delayed_refs;
2440
2441         ret = cleanup_extent_op(trans, head);
2442         if (ret < 0) {
2443                 unselect_delayed_ref_head(delayed_refs, head);
2444                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2445                 return ret;
2446         } else if (ret) {
2447                 return ret;
2448         }
2449
2450         /*
2451          * Need to drop our head ref lock and re-acquire the delayed ref lock
2452          * and then re-check to make sure nobody got added.
2453          */
2454         spin_unlock(&head->lock);
2455         spin_lock(&delayed_refs->lock);
2456         spin_lock(&head->lock);
2457         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2458                 spin_unlock(&head->lock);
2459                 spin_unlock(&delayed_refs->lock);
2460                 return 1;
2461         }
2462         delayed_refs->num_heads--;
2463         rb_erase(&head->href_node, &delayed_refs->href_root);
2464         RB_CLEAR_NODE(&head->href_node);
2465         spin_unlock(&head->lock);
2466         spin_unlock(&delayed_refs->lock);
2467         atomic_dec(&delayed_refs->num_entries);
2468
2469         trace_run_delayed_ref_head(fs_info, head, 0);
2470
2471         if (head->total_ref_mod < 0) {
2472                 struct btrfs_space_info *space_info;
2473                 u64 flags;
2474
2475                 if (head->is_data)
2476                         flags = BTRFS_BLOCK_GROUP_DATA;
2477                 else if (head->is_system)
2478                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2479                 else
2480                         flags = BTRFS_BLOCK_GROUP_METADATA;
2481                 space_info = __find_space_info(fs_info, flags);
2482                 ASSERT(space_info);
2483                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2484                                    -head->num_bytes,
2485                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2486
2487                 if (head->is_data) {
2488                         spin_lock(&delayed_refs->lock);
2489                         delayed_refs->pending_csums -= head->num_bytes;
2490                         spin_unlock(&delayed_refs->lock);
2491                 }
2492         }
2493
2494         if (head->must_insert_reserved) {
2495                 btrfs_pin_extent(fs_info, head->bytenr,
2496                                  head->num_bytes, 1);
2497                 if (head->is_data) {
2498                         ret = btrfs_del_csums(trans, fs_info->csum_root,
2499                                               head->bytenr, head->num_bytes);
2500                 }
2501         }
2502
2503         /* Also free its reserved qgroup space */
2504         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2505                                       head->qgroup_reserved);
2506         btrfs_delayed_ref_unlock(head);
2507         btrfs_put_delayed_ref_head(head);
2508         return ret;
2509 }
2510
2511 /*
2512  * Returns 0 on success or if called with an already aborted transaction.
2513  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2514  */
2515 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2516                                              unsigned long nr)
2517 {
2518         struct btrfs_fs_info *fs_info = trans->fs_info;
2519         struct btrfs_delayed_ref_root *delayed_refs;
2520         struct btrfs_delayed_ref_node *ref;
2521         struct btrfs_delayed_ref_head *locked_ref = NULL;
2522         struct btrfs_delayed_extent_op *extent_op;
2523         ktime_t start = ktime_get();
2524         int ret;
2525         unsigned long count = 0;
2526         unsigned long actual_count = 0;
2527         int must_insert_reserved = 0;
2528
2529         delayed_refs = &trans->transaction->delayed_refs;
2530         while (1) {
2531                 if (!locked_ref) {
2532                         if (count >= nr)
2533                                 break;
2534
2535                         spin_lock(&delayed_refs->lock);
2536                         locked_ref = btrfs_select_ref_head(trans);
2537                         if (!locked_ref) {
2538                                 spin_unlock(&delayed_refs->lock);
2539                                 break;
2540                         }
2541
2542                         /* grab the lock that says we are going to process
2543                          * all the refs for this head */
2544                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2545                         spin_unlock(&delayed_refs->lock);
2546                         /*
2547                          * we may have dropped the spin lock to get the head
2548                          * mutex lock, and that might have given someone else
2549                          * time to free the head.  If that's true, it has been
2550                          * removed from our list and we can move on.
2551                          */
2552                         if (ret == -EAGAIN) {
2553                                 locked_ref = NULL;
2554                                 count++;
2555                                 continue;
2556                         }
2557                 }
2558
2559                 /*
2560                  * We need to try and merge add/drops of the same ref since we
2561                  * can run into issues with relocate dropping the implicit ref
2562                  * and then it being added back again before the drop can
2563                  * finish.  If we merged anything we need to re-loop so we can
2564                  * get a good ref.
2565                  * Or we can get node references of the same type that weren't
2566                  * merged when created due to bumps in the tree mod seq, and
2567                  * we need to merge them to prevent adding an inline extent
2568                  * backref before dropping it (triggering a BUG_ON at
2569                  * insert_inline_extent_backref()).
2570                  */
2571                 spin_lock(&locked_ref->lock);
2572                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2573
2574                 ref = select_delayed_ref(locked_ref);
2575
2576                 if (ref && ref->seq &&
2577                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2578                         spin_unlock(&locked_ref->lock);
2579                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2580                         locked_ref = NULL;
2581                         cond_resched();
2582                         count++;
2583                         continue;
2584                 }
2585
2586                 /*
2587                  * We're done processing refs in this ref_head, clean everything
2588                  * up and move on to the next ref_head.
2589                  */
2590                 if (!ref) {
2591                         ret = cleanup_ref_head(trans, locked_ref);
2592                         if (ret > 0 ) {
2593                                 /* We dropped our lock, we need to loop. */
2594                                 ret = 0;
2595                                 continue;
2596                         } else if (ret) {
2597                                 return ret;
2598                         }
2599                         locked_ref = NULL;
2600                         count++;
2601                         continue;
2602                 }
2603
2604                 actual_count++;
2605                 ref->in_tree = 0;
2606                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2607                 RB_CLEAR_NODE(&ref->ref_node);
2608                 if (!list_empty(&ref->add_list))
2609                         list_del(&ref->add_list);
2610                 /*
2611                  * When we play the delayed ref, also correct the ref_mod on
2612                  * head
2613                  */
2614                 switch (ref->action) {
2615                 case BTRFS_ADD_DELAYED_REF:
2616                 case BTRFS_ADD_DELAYED_EXTENT:
2617                         locked_ref->ref_mod -= ref->ref_mod;
2618                         break;
2619                 case BTRFS_DROP_DELAYED_REF:
2620                         locked_ref->ref_mod += ref->ref_mod;
2621                         break;
2622                 default:
2623                         WARN_ON(1);
2624                 }
2625                 atomic_dec(&delayed_refs->num_entries);
2626
2627                 /*
2628                  * Record the must-insert_reserved flag before we drop the spin
2629                  * lock.
2630                  */
2631                 must_insert_reserved = locked_ref->must_insert_reserved;
2632                 locked_ref->must_insert_reserved = 0;
2633
2634                 extent_op = locked_ref->extent_op;
2635                 locked_ref->extent_op = NULL;
2636                 spin_unlock(&locked_ref->lock);
2637
2638                 ret = run_one_delayed_ref(trans, ref, extent_op,
2639                                           must_insert_reserved);
2640
2641                 btrfs_free_delayed_extent_op(extent_op);
2642                 if (ret) {
2643                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2644                         btrfs_put_delayed_ref(ref);
2645                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2646                                     ret);
2647                         return ret;
2648                 }
2649
2650                 btrfs_put_delayed_ref(ref);
2651                 count++;
2652                 cond_resched();
2653         }
2654
2655         /*
2656          * We don't want to include ref heads since we can have empty ref heads
2657          * and those will drastically skew our runtime down since we just do
2658          * accounting, no actual extent tree updates.
2659          */
2660         if (actual_count > 0) {
2661                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2662                 u64 avg;
2663
2664                 /*
2665                  * We weigh the current average higher than our current runtime
2666                  * to avoid large swings in the average.
2667                  */
2668                 spin_lock(&delayed_refs->lock);
2669                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2670                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2671                 spin_unlock(&delayed_refs->lock);
2672         }
2673         return 0;
2674 }
2675
2676 #ifdef SCRAMBLE_DELAYED_REFS
2677 /*
2678  * Normally delayed refs get processed in ascending bytenr order. This
2679  * correlates in most cases to the order added. To expose dependencies on this
2680  * order, we start to process the tree in the middle instead of the beginning
2681  */
2682 static u64 find_middle(struct rb_root *root)
2683 {
2684         struct rb_node *n = root->rb_node;
2685         struct btrfs_delayed_ref_node *entry;
2686         int alt = 1;
2687         u64 middle;
2688         u64 first = 0, last = 0;
2689
2690         n = rb_first(root);
2691         if (n) {
2692                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2693                 first = entry->bytenr;
2694         }
2695         n = rb_last(root);
2696         if (n) {
2697                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2698                 last = entry->bytenr;
2699         }
2700         n = root->rb_node;
2701
2702         while (n) {
2703                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2704                 WARN_ON(!entry->in_tree);
2705
2706                 middle = entry->bytenr;
2707
2708                 if (alt)
2709                         n = n->rb_left;
2710                 else
2711                         n = n->rb_right;
2712
2713                 alt = 1 - alt;
2714         }
2715         return middle;
2716 }
2717 #endif
2718
2719 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2720 {
2721         u64 num_bytes;
2722
2723         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2724                              sizeof(struct btrfs_extent_inline_ref));
2725         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2726                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2727
2728         /*
2729          * We don't ever fill up leaves all the way so multiply by 2 just to be
2730          * closer to what we're really going to want to use.
2731          */
2732         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2733 }
2734
2735 /*
2736  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2737  * would require to store the csums for that many bytes.
2738  */
2739 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2740 {
2741         u64 csum_size;
2742         u64 num_csums_per_leaf;
2743         u64 num_csums;
2744
2745         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2746         num_csums_per_leaf = div64_u64(csum_size,
2747                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2748         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2749         num_csums += num_csums_per_leaf - 1;
2750         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2751         return num_csums;
2752 }
2753
2754 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2755                                        struct btrfs_fs_info *fs_info)
2756 {
2757         struct btrfs_block_rsv *global_rsv;
2758         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2759         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2760         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2761         u64 num_bytes, num_dirty_bgs_bytes;
2762         int ret = 0;
2763
2764         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2765         num_heads = heads_to_leaves(fs_info, num_heads);
2766         if (num_heads > 1)
2767                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2768         num_bytes <<= 1;
2769         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2770                                                         fs_info->nodesize;
2771         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2772                                                              num_dirty_bgs);
2773         global_rsv = &fs_info->global_block_rsv;
2774
2775         /*
2776          * If we can't allocate any more chunks lets make sure we have _lots_ of
2777          * wiggle room since running delayed refs can create more delayed refs.
2778          */
2779         if (global_rsv->space_info->full) {
2780                 num_dirty_bgs_bytes <<= 1;
2781                 num_bytes <<= 1;
2782         }
2783
2784         spin_lock(&global_rsv->lock);
2785         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2786                 ret = 1;
2787         spin_unlock(&global_rsv->lock);
2788         return ret;
2789 }
2790
2791 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2792                                        struct btrfs_fs_info *fs_info)
2793 {
2794         u64 num_entries =
2795                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2796         u64 avg_runtime;
2797         u64 val;
2798
2799         smp_mb();
2800         avg_runtime = fs_info->avg_delayed_ref_runtime;
2801         val = num_entries * avg_runtime;
2802         if (val >= NSEC_PER_SEC)
2803                 return 1;
2804         if (val >= NSEC_PER_SEC / 2)
2805                 return 2;
2806
2807         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2808 }
2809
2810 struct async_delayed_refs {
2811         struct btrfs_root *root;
2812         u64 transid;
2813         int count;
2814         int error;
2815         int sync;
2816         struct completion wait;
2817         struct btrfs_work work;
2818 };
2819
2820 static inline struct async_delayed_refs *
2821 to_async_delayed_refs(struct btrfs_work *work)
2822 {
2823         return container_of(work, struct async_delayed_refs, work);
2824 }
2825
2826 static void delayed_ref_async_start(struct btrfs_work *work)
2827 {
2828         struct async_delayed_refs *async = to_async_delayed_refs(work);
2829         struct btrfs_trans_handle *trans;
2830         struct btrfs_fs_info *fs_info = async->root->fs_info;
2831         int ret;
2832
2833         /* if the commit is already started, we don't need to wait here */
2834         if (btrfs_transaction_blocked(fs_info))
2835                 goto done;
2836
2837         trans = btrfs_join_transaction(async->root);
2838         if (IS_ERR(trans)) {
2839                 async->error = PTR_ERR(trans);
2840                 goto done;
2841         }
2842
2843         /*
2844          * trans->sync means that when we call end_transaction, we won't
2845          * wait on delayed refs
2846          */
2847         trans->sync = true;
2848
2849         /* Don't bother flushing if we got into a different transaction */
2850         if (trans->transid > async->transid)
2851                 goto end;
2852
2853         ret = btrfs_run_delayed_refs(trans, async->count);
2854         if (ret)
2855                 async->error = ret;
2856 end:
2857         ret = btrfs_end_transaction(trans);
2858         if (ret && !async->error)
2859                 async->error = ret;
2860 done:
2861         if (async->sync)
2862                 complete(&async->wait);
2863         else
2864                 kfree(async);
2865 }
2866
2867 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2868                                  unsigned long count, u64 transid, int wait)
2869 {
2870         struct async_delayed_refs *async;
2871         int ret;
2872
2873         async = kmalloc(sizeof(*async), GFP_NOFS);
2874         if (!async)
2875                 return -ENOMEM;
2876
2877         async->root = fs_info->tree_root;
2878         async->count = count;
2879         async->error = 0;
2880         async->transid = transid;
2881         if (wait)
2882                 async->sync = 1;
2883         else
2884                 async->sync = 0;
2885         init_completion(&async->wait);
2886
2887         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2888                         delayed_ref_async_start, NULL, NULL);
2889
2890         btrfs_queue_work(fs_info->extent_workers, &async->work);
2891
2892         if (wait) {
2893                 wait_for_completion(&async->wait);
2894                 ret = async->error;
2895                 kfree(async);
2896                 return ret;
2897         }
2898         return 0;
2899 }
2900
2901 /*
2902  * this starts processing the delayed reference count updates and
2903  * extent insertions we have queued up so far.  count can be
2904  * 0, which means to process everything in the tree at the start
2905  * of the run (but not newly added entries), or it can be some target
2906  * number you'd like to process.
2907  *
2908  * Returns 0 on success or if called with an aborted transaction
2909  * Returns <0 on error and aborts the transaction
2910  */
2911 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2912                            unsigned long count)
2913 {
2914         struct btrfs_fs_info *fs_info = trans->fs_info;
2915         struct rb_node *node;
2916         struct btrfs_delayed_ref_root *delayed_refs;
2917         struct btrfs_delayed_ref_head *head;
2918         int ret;
2919         int run_all = count == (unsigned long)-1;
2920
2921         /* We'll clean this up in btrfs_cleanup_transaction */
2922         if (trans->aborted)
2923                 return 0;
2924
2925         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2926                 return 0;
2927
2928         delayed_refs = &trans->transaction->delayed_refs;
2929         if (count == 0)
2930                 count = atomic_read(&delayed_refs->num_entries) * 2;
2931
2932 again:
2933 #ifdef SCRAMBLE_DELAYED_REFS
2934         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2935 #endif
2936         ret = __btrfs_run_delayed_refs(trans, count);
2937         if (ret < 0) {
2938                 btrfs_abort_transaction(trans, ret);
2939                 return ret;
2940         }
2941
2942         if (run_all) {
2943                 if (!list_empty(&trans->new_bgs))
2944                         btrfs_create_pending_block_groups(trans);
2945
2946                 spin_lock(&delayed_refs->lock);
2947                 node = rb_first(&delayed_refs->href_root);
2948                 if (!node) {
2949                         spin_unlock(&delayed_refs->lock);
2950                         goto out;
2951                 }
2952                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2953                                 href_node);
2954                 refcount_inc(&head->refs);
2955                 spin_unlock(&delayed_refs->lock);
2956
2957                 /* Mutex was contended, block until it's released and retry. */
2958                 mutex_lock(&head->mutex);
2959                 mutex_unlock(&head->mutex);
2960
2961                 btrfs_put_delayed_ref_head(head);
2962                 cond_resched();
2963                 goto again;
2964         }
2965 out:
2966         return 0;
2967 }
2968
2969 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2970                                 struct btrfs_fs_info *fs_info,
2971                                 u64 bytenr, u64 num_bytes, u64 flags,
2972                                 int level, int is_data)
2973 {
2974         struct btrfs_delayed_extent_op *extent_op;
2975         int ret;
2976
2977         extent_op = btrfs_alloc_delayed_extent_op();
2978         if (!extent_op)
2979                 return -ENOMEM;
2980
2981         extent_op->flags_to_set = flags;
2982         extent_op->update_flags = true;
2983         extent_op->update_key = false;
2984         extent_op->is_data = is_data ? true : false;
2985         extent_op->level = level;
2986
2987         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
2988                                           num_bytes, extent_op);
2989         if (ret)
2990                 btrfs_free_delayed_extent_op(extent_op);
2991         return ret;
2992 }
2993
2994 static noinline int check_delayed_ref(struct btrfs_root *root,
2995                                       struct btrfs_path *path,
2996                                       u64 objectid, u64 offset, u64 bytenr)
2997 {
2998         struct btrfs_delayed_ref_head *head;
2999         struct btrfs_delayed_ref_node *ref;
3000         struct btrfs_delayed_data_ref *data_ref;
3001         struct btrfs_delayed_ref_root *delayed_refs;
3002         struct btrfs_transaction *cur_trans;
3003         struct rb_node *node;
3004         int ret = 0;
3005
3006         spin_lock(&root->fs_info->trans_lock);
3007         cur_trans = root->fs_info->running_transaction;
3008         if (cur_trans)
3009                 refcount_inc(&cur_trans->use_count);
3010         spin_unlock(&root->fs_info->trans_lock);
3011         if (!cur_trans)
3012                 return 0;
3013
3014         delayed_refs = &cur_trans->delayed_refs;
3015         spin_lock(&delayed_refs->lock);
3016         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3017         if (!head) {
3018                 spin_unlock(&delayed_refs->lock);
3019                 btrfs_put_transaction(cur_trans);
3020                 return 0;
3021         }
3022
3023         if (!mutex_trylock(&head->mutex)) {
3024                 refcount_inc(&head->refs);
3025                 spin_unlock(&delayed_refs->lock);
3026
3027                 btrfs_release_path(path);
3028
3029                 /*
3030                  * Mutex was contended, block until it's released and let
3031                  * caller try again
3032                  */
3033                 mutex_lock(&head->mutex);
3034                 mutex_unlock(&head->mutex);
3035                 btrfs_put_delayed_ref_head(head);
3036                 btrfs_put_transaction(cur_trans);
3037                 return -EAGAIN;
3038         }
3039         spin_unlock(&delayed_refs->lock);
3040
3041         spin_lock(&head->lock);
3042         /*
3043          * XXX: We should replace this with a proper search function in the
3044          * future.
3045          */
3046         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3047                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3048                 /* If it's a shared ref we know a cross reference exists */
3049                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3050                         ret = 1;
3051                         break;
3052                 }
3053
3054                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3055
3056                 /*
3057                  * If our ref doesn't match the one we're currently looking at
3058                  * then we have a cross reference.
3059                  */
3060                 if (data_ref->root != root->root_key.objectid ||
3061                     data_ref->objectid != objectid ||
3062                     data_ref->offset != offset) {
3063                         ret = 1;
3064                         break;
3065                 }
3066         }
3067         spin_unlock(&head->lock);
3068         mutex_unlock(&head->mutex);
3069         btrfs_put_transaction(cur_trans);
3070         return ret;
3071 }
3072
3073 static noinline int check_committed_ref(struct btrfs_root *root,
3074                                         struct btrfs_path *path,
3075                                         u64 objectid, u64 offset, u64 bytenr)
3076 {
3077         struct btrfs_fs_info *fs_info = root->fs_info;
3078         struct btrfs_root *extent_root = fs_info->extent_root;
3079         struct extent_buffer *leaf;
3080         struct btrfs_extent_data_ref *ref;
3081         struct btrfs_extent_inline_ref *iref;
3082         struct btrfs_extent_item *ei;
3083         struct btrfs_key key;
3084         u32 item_size;
3085         int type;
3086         int ret;
3087
3088         key.objectid = bytenr;
3089         key.offset = (u64)-1;
3090         key.type = BTRFS_EXTENT_ITEM_KEY;
3091
3092         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3093         if (ret < 0)
3094                 goto out;
3095         BUG_ON(ret == 0); /* Corruption */
3096
3097         ret = -ENOENT;
3098         if (path->slots[0] == 0)
3099                 goto out;
3100
3101         path->slots[0]--;
3102         leaf = path->nodes[0];
3103         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3104
3105         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3106                 goto out;
3107
3108         ret = 1;
3109         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3110         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3111
3112         if (item_size != sizeof(*ei) +
3113             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3114                 goto out;
3115
3116         if (btrfs_extent_generation(leaf, ei) <=
3117             btrfs_root_last_snapshot(&root->root_item))
3118                 goto out;
3119
3120         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3121
3122         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3123         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3124                 goto out;
3125
3126         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3127         if (btrfs_extent_refs(leaf, ei) !=
3128             btrfs_extent_data_ref_count(leaf, ref) ||
3129             btrfs_extent_data_ref_root(leaf, ref) !=
3130             root->root_key.objectid ||
3131             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3132             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3133                 goto out;
3134
3135         ret = 0;
3136 out:
3137         return ret;
3138 }
3139
3140 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3141                           u64 bytenr)
3142 {
3143         struct btrfs_path *path;
3144         int ret;
3145         int ret2;
3146
3147         path = btrfs_alloc_path();
3148         if (!path)
3149                 return -ENOMEM;
3150
3151         do {
3152                 ret = check_committed_ref(root, path, objectid,
3153                                           offset, bytenr);
3154                 if (ret && ret != -ENOENT)
3155                         goto out;
3156
3157                 ret2 = check_delayed_ref(root, path, objectid,
3158                                          offset, bytenr);
3159         } while (ret2 == -EAGAIN);
3160
3161         if (ret2 && ret2 != -ENOENT) {
3162                 ret = ret2;
3163                 goto out;
3164         }
3165
3166         if (ret != -ENOENT || ret2 != -ENOENT)
3167                 ret = 0;
3168 out:
3169         btrfs_free_path(path);
3170         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3171                 WARN_ON(ret > 0);
3172         return ret;
3173 }
3174
3175 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3176                            struct btrfs_root *root,
3177                            struct extent_buffer *buf,
3178                            int full_backref, int inc)
3179 {
3180         struct btrfs_fs_info *fs_info = root->fs_info;
3181         u64 bytenr;
3182         u64 num_bytes;
3183         u64 parent;
3184         u64 ref_root;
3185         u32 nritems;
3186         struct btrfs_key key;
3187         struct btrfs_file_extent_item *fi;
3188         int i;
3189         int level;
3190         int ret = 0;
3191         int (*process_func)(struct btrfs_trans_handle *,
3192                             struct btrfs_root *,
3193                             u64, u64, u64, u64, u64, u64);
3194
3195
3196         if (btrfs_is_testing(fs_info))
3197                 return 0;
3198
3199         ref_root = btrfs_header_owner(buf);
3200         nritems = btrfs_header_nritems(buf);
3201         level = btrfs_header_level(buf);
3202
3203         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3204                 return 0;
3205
3206         if (inc)
3207                 process_func = btrfs_inc_extent_ref;
3208         else
3209                 process_func = btrfs_free_extent;
3210
3211         if (full_backref)
3212                 parent = buf->start;
3213         else
3214                 parent = 0;
3215
3216         for (i = 0; i < nritems; i++) {
3217                 if (level == 0) {
3218                         btrfs_item_key_to_cpu(buf, &key, i);
3219                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3220                                 continue;
3221                         fi = btrfs_item_ptr(buf, i,
3222                                             struct btrfs_file_extent_item);
3223                         if (btrfs_file_extent_type(buf, fi) ==
3224                             BTRFS_FILE_EXTENT_INLINE)
3225                                 continue;
3226                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3227                         if (bytenr == 0)
3228                                 continue;
3229
3230                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3231                         key.offset -= btrfs_file_extent_offset(buf, fi);
3232                         ret = process_func(trans, root, bytenr, num_bytes,
3233                                            parent, ref_root, key.objectid,
3234                                            key.offset);
3235                         if (ret)
3236                                 goto fail;
3237                 } else {
3238                         bytenr = btrfs_node_blockptr(buf, i);
3239                         num_bytes = fs_info->nodesize;
3240                         ret = process_func(trans, root, bytenr, num_bytes,
3241                                            parent, ref_root, level - 1, 0);
3242                         if (ret)
3243                                 goto fail;
3244                 }
3245         }
3246         return 0;
3247 fail:
3248         return ret;
3249 }
3250
3251 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3252                   struct extent_buffer *buf, int full_backref)
3253 {
3254         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3255 }
3256
3257 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3258                   struct extent_buffer *buf, int full_backref)
3259 {
3260         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3261 }
3262
3263 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3264                                  struct btrfs_fs_info *fs_info,
3265                                  struct btrfs_path *path,
3266                                  struct btrfs_block_group_cache *cache)
3267 {
3268         int ret;
3269         struct btrfs_root *extent_root = fs_info->extent_root;
3270         unsigned long bi;
3271         struct extent_buffer *leaf;
3272
3273         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3274         if (ret) {
3275                 if (ret > 0)
3276                         ret = -ENOENT;
3277                 goto fail;
3278         }
3279
3280         leaf = path->nodes[0];
3281         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3282         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3283         btrfs_mark_buffer_dirty(leaf);
3284 fail:
3285         btrfs_release_path(path);
3286         return ret;
3287
3288 }
3289
3290 static struct btrfs_block_group_cache *
3291 next_block_group(struct btrfs_fs_info *fs_info,
3292                  struct btrfs_block_group_cache *cache)
3293 {
3294         struct rb_node *node;
3295
3296         spin_lock(&fs_info->block_group_cache_lock);
3297
3298         /* If our block group was removed, we need a full search. */
3299         if (RB_EMPTY_NODE(&cache->cache_node)) {
3300                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3301
3302                 spin_unlock(&fs_info->block_group_cache_lock);
3303                 btrfs_put_block_group(cache);
3304                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3305         }
3306         node = rb_next(&cache->cache_node);
3307         btrfs_put_block_group(cache);
3308         if (node) {
3309                 cache = rb_entry(node, struct btrfs_block_group_cache,
3310                                  cache_node);
3311                 btrfs_get_block_group(cache);
3312         } else
3313                 cache = NULL;
3314         spin_unlock(&fs_info->block_group_cache_lock);
3315         return cache;
3316 }
3317
3318 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3319                             struct btrfs_trans_handle *trans,
3320                             struct btrfs_path *path)
3321 {
3322         struct btrfs_fs_info *fs_info = block_group->fs_info;
3323         struct btrfs_root *root = fs_info->tree_root;
3324         struct inode *inode = NULL;
3325         struct extent_changeset *data_reserved = NULL;
3326         u64 alloc_hint = 0;
3327         int dcs = BTRFS_DC_ERROR;
3328         u64 num_pages = 0;
3329         int retries = 0;
3330         int ret = 0;
3331
3332         /*
3333          * If this block group is smaller than 100 megs don't bother caching the
3334          * block group.
3335          */
3336         if (block_group->key.offset < (100 * SZ_1M)) {
3337                 spin_lock(&block_group->lock);
3338                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3339                 spin_unlock(&block_group->lock);
3340                 return 0;
3341         }
3342
3343         if (trans->aborted)
3344                 return 0;
3345 again:
3346         inode = lookup_free_space_inode(fs_info, block_group, path);
3347         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3348                 ret = PTR_ERR(inode);
3349                 btrfs_release_path(path);
3350                 goto out;
3351         }
3352
3353         if (IS_ERR(inode)) {
3354                 BUG_ON(retries);
3355                 retries++;
3356
3357                 if (block_group->ro)
3358                         goto out_free;
3359
3360                 ret = create_free_space_inode(fs_info, trans, block_group,
3361                                               path);
3362                 if (ret)
3363                         goto out_free;
3364                 goto again;
3365         }
3366
3367         /*
3368          * We want to set the generation to 0, that way if anything goes wrong
3369          * from here on out we know not to trust this cache when we load up next
3370          * time.
3371          */
3372         BTRFS_I(inode)->generation = 0;
3373         ret = btrfs_update_inode(trans, root, inode);
3374         if (ret) {
3375                 /*
3376                  * So theoretically we could recover from this, simply set the
3377                  * super cache generation to 0 so we know to invalidate the
3378                  * cache, but then we'd have to keep track of the block groups
3379                  * that fail this way so we know we _have_ to reset this cache
3380                  * before the next commit or risk reading stale cache.  So to
3381                  * limit our exposure to horrible edge cases lets just abort the
3382                  * transaction, this only happens in really bad situations
3383                  * anyway.
3384                  */
3385                 btrfs_abort_transaction(trans, ret);
3386                 goto out_put;
3387         }
3388         WARN_ON(ret);
3389
3390         /* We've already setup this transaction, go ahead and exit */
3391         if (block_group->cache_generation == trans->transid &&
3392             i_size_read(inode)) {
3393                 dcs = BTRFS_DC_SETUP;
3394                 goto out_put;
3395         }
3396
3397         if (i_size_read(inode) > 0) {
3398                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3399                                         &fs_info->global_block_rsv);
3400                 if (ret)
3401                         goto out_put;
3402
3403                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3404                 if (ret)
3405                         goto out_put;
3406         }
3407
3408         spin_lock(&block_group->lock);
3409         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3410             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3411                 /*
3412                  * don't bother trying to write stuff out _if_
3413                  * a) we're not cached,
3414                  * b) we're with nospace_cache mount option,
3415                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3416                  */
3417                 dcs = BTRFS_DC_WRITTEN;
3418                 spin_unlock(&block_group->lock);
3419                 goto out_put;
3420         }
3421         spin_unlock(&block_group->lock);
3422
3423         /*
3424          * We hit an ENOSPC when setting up the cache in this transaction, just
3425          * skip doing the setup, we've already cleared the cache so we're safe.
3426          */
3427         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3428                 ret = -ENOSPC;
3429                 goto out_put;
3430         }
3431
3432         /*
3433          * Try to preallocate enough space based on how big the block group is.
3434          * Keep in mind this has to include any pinned space which could end up
3435          * taking up quite a bit since it's not folded into the other space
3436          * cache.
3437          */
3438         num_pages = div_u64(block_group->key.offset, SZ_256M);
3439         if (!num_pages)
3440                 num_pages = 1;
3441
3442         num_pages *= 16;
3443         num_pages *= PAGE_SIZE;
3444
3445         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3446         if (ret)
3447                 goto out_put;
3448
3449         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3450                                               num_pages, num_pages,
3451                                               &alloc_hint);
3452         /*
3453          * Our cache requires contiguous chunks so that we don't modify a bunch
3454          * of metadata or split extents when writing the cache out, which means
3455          * we can enospc if we are heavily fragmented in addition to just normal
3456          * out of space conditions.  So if we hit this just skip setting up any
3457          * other block groups for this transaction, maybe we'll unpin enough
3458          * space the next time around.
3459          */
3460         if (!ret)
3461                 dcs = BTRFS_DC_SETUP;
3462         else if (ret == -ENOSPC)
3463                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3464
3465 out_put:
3466         iput(inode);
3467 out_free:
3468         btrfs_release_path(path);
3469 out:
3470         spin_lock(&block_group->lock);
3471         if (!ret && dcs == BTRFS_DC_SETUP)
3472                 block_group->cache_generation = trans->transid;
3473         block_group->disk_cache_state = dcs;
3474         spin_unlock(&block_group->lock);
3475
3476         extent_changeset_free(data_reserved);
3477         return ret;
3478 }
3479
3480 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3481                             struct btrfs_fs_info *fs_info)
3482 {
3483         struct btrfs_block_group_cache *cache, *tmp;
3484         struct btrfs_transaction *cur_trans = trans->transaction;
3485         struct btrfs_path *path;
3486
3487         if (list_empty(&cur_trans->dirty_bgs) ||
3488             !btrfs_test_opt(fs_info, SPACE_CACHE))
3489                 return 0;
3490
3491         path = btrfs_alloc_path();
3492         if (!path)
3493                 return -ENOMEM;
3494
3495         /* Could add new block groups, use _safe just in case */
3496         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3497                                  dirty_list) {
3498                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3499                         cache_save_setup(cache, trans, path);
3500         }
3501
3502         btrfs_free_path(path);
3503         return 0;
3504 }
3505
3506 /*
3507  * transaction commit does final block group cache writeback during a
3508  * critical section where nothing is allowed to change the FS.  This is
3509  * required in order for the cache to actually match the block group,
3510  * but can introduce a lot of latency into the commit.
3511  *
3512  * So, btrfs_start_dirty_block_groups is here to kick off block group
3513  * cache IO.  There's a chance we'll have to redo some of it if the
3514  * block group changes again during the commit, but it greatly reduces
3515  * the commit latency by getting rid of the easy block groups while
3516  * we're still allowing others to join the commit.
3517  */
3518 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3519 {
3520         struct btrfs_fs_info *fs_info = trans->fs_info;
3521         struct btrfs_block_group_cache *cache;
3522         struct btrfs_transaction *cur_trans = trans->transaction;
3523         int ret = 0;
3524         int should_put;
3525         struct btrfs_path *path = NULL;
3526         LIST_HEAD(dirty);
3527         struct list_head *io = &cur_trans->io_bgs;
3528         int num_started = 0;
3529         int loops = 0;
3530
3531         spin_lock(&cur_trans->dirty_bgs_lock);
3532         if (list_empty(&cur_trans->dirty_bgs)) {
3533                 spin_unlock(&cur_trans->dirty_bgs_lock);
3534                 return 0;
3535         }
3536         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3537         spin_unlock(&cur_trans->dirty_bgs_lock);
3538
3539 again:
3540         /*
3541          * make sure all the block groups on our dirty list actually
3542          * exist
3543          */
3544         btrfs_create_pending_block_groups(trans);
3545
3546         if (!path) {
3547                 path = btrfs_alloc_path();
3548                 if (!path)
3549                         return -ENOMEM;
3550         }
3551
3552         /*
3553          * cache_write_mutex is here only to save us from balance or automatic
3554          * removal of empty block groups deleting this block group while we are
3555          * writing out the cache
3556          */
3557         mutex_lock(&trans->transaction->cache_write_mutex);
3558         while (!list_empty(&dirty)) {
3559                 cache = list_first_entry(&dirty,
3560                                          struct btrfs_block_group_cache,
3561                                          dirty_list);
3562                 /*
3563                  * this can happen if something re-dirties a block
3564                  * group that is already under IO.  Just wait for it to
3565                  * finish and then do it all again
3566                  */
3567                 if (!list_empty(&cache->io_list)) {
3568                         list_del_init(&cache->io_list);
3569                         btrfs_wait_cache_io(trans, cache, path);
3570                         btrfs_put_block_group(cache);
3571                 }
3572
3573
3574                 /*
3575                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3576                  * if it should update the cache_state.  Don't delete
3577                  * until after we wait.
3578                  *
3579                  * Since we're not running in the commit critical section
3580                  * we need the dirty_bgs_lock to protect from update_block_group
3581                  */
3582                 spin_lock(&cur_trans->dirty_bgs_lock);
3583                 list_del_init(&cache->dirty_list);
3584                 spin_unlock(&cur_trans->dirty_bgs_lock);
3585
3586                 should_put = 1;
3587
3588                 cache_save_setup(cache, trans, path);
3589
3590                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3591                         cache->io_ctl.inode = NULL;
3592                         ret = btrfs_write_out_cache(fs_info, trans,
3593                                                     cache, path);
3594                         if (ret == 0 && cache->io_ctl.inode) {
3595                                 num_started++;
3596                                 should_put = 0;
3597
3598                                 /*
3599                                  * The cache_write_mutex is protecting the
3600                                  * io_list, also refer to the definition of
3601                                  * btrfs_transaction::io_bgs for more details
3602                                  */
3603                                 list_add_tail(&cache->io_list, io);
3604                         } else {
3605                                 /*
3606                                  * if we failed to write the cache, the
3607                                  * generation will be bad and life goes on
3608                                  */
3609                                 ret = 0;
3610                         }
3611                 }
3612                 if (!ret) {
3613                         ret = write_one_cache_group(trans, fs_info,
3614                                                     path, cache);
3615                         /*
3616                          * Our block group might still be attached to the list
3617                          * of new block groups in the transaction handle of some
3618                          * other task (struct btrfs_trans_handle->new_bgs). This
3619                          * means its block group item isn't yet in the extent
3620                          * tree. If this happens ignore the error, as we will
3621                          * try again later in the critical section of the
3622                          * transaction commit.
3623                          */
3624                         if (ret == -ENOENT) {
3625                                 ret = 0;
3626                                 spin_lock(&cur_trans->dirty_bgs_lock);
3627                                 if (list_empty(&cache->dirty_list)) {
3628                                         list_add_tail(&cache->dirty_list,
3629                                                       &cur_trans->dirty_bgs);
3630                                         btrfs_get_block_group(cache);
3631                                 }
3632                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3633                         } else if (ret) {
3634                                 btrfs_abort_transaction(trans, ret);
3635                         }
3636                 }
3637
3638                 /* if its not on the io list, we need to put the block group */
3639                 if (should_put)
3640                         btrfs_put_block_group(cache);
3641
3642                 if (ret)
3643                         break;
3644
3645                 /*
3646                  * Avoid blocking other tasks for too long. It might even save
3647                  * us from writing caches for block groups that are going to be
3648                  * removed.
3649                  */
3650                 mutex_unlock(&trans->transaction->cache_write_mutex);
3651                 mutex_lock(&trans->transaction->cache_write_mutex);
3652         }
3653         mutex_unlock(&trans->transaction->cache_write_mutex);
3654
3655         /*
3656          * go through delayed refs for all the stuff we've just kicked off
3657          * and then loop back (just once)
3658          */
3659         ret = btrfs_run_delayed_refs(trans, 0);
3660         if (!ret && loops == 0) {
3661                 loops++;
3662                 spin_lock(&cur_trans->dirty_bgs_lock);
3663                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3664                 /*
3665                  * dirty_bgs_lock protects us from concurrent block group
3666                  * deletes too (not just cache_write_mutex).
3667                  */
3668                 if (!list_empty(&dirty)) {
3669                         spin_unlock(&cur_trans->dirty_bgs_lock);
3670                         goto again;
3671                 }
3672                 spin_unlock(&cur_trans->dirty_bgs_lock);
3673         } else if (ret < 0) {
3674                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3675         }
3676
3677         btrfs_free_path(path);
3678         return ret;
3679 }
3680
3681 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3682                                    struct btrfs_fs_info *fs_info)
3683 {
3684         struct btrfs_block_group_cache *cache;
3685         struct btrfs_transaction *cur_trans = trans->transaction;
3686         int ret = 0;
3687         int should_put;
3688         struct btrfs_path *path;
3689         struct list_head *io = &cur_trans->io_bgs;
3690         int num_started = 0;
3691
3692         path = btrfs_alloc_path();
3693         if (!path)
3694                 return -ENOMEM;
3695
3696         /*
3697          * Even though we are in the critical section of the transaction commit,
3698          * we can still have concurrent tasks adding elements to this
3699          * transaction's list of dirty block groups. These tasks correspond to
3700          * endio free space workers started when writeback finishes for a
3701          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3702          * allocate new block groups as a result of COWing nodes of the root
3703          * tree when updating the free space inode. The writeback for the space
3704          * caches is triggered by an earlier call to
3705          * btrfs_start_dirty_block_groups() and iterations of the following
3706          * loop.
3707          * Also we want to do the cache_save_setup first and then run the
3708          * delayed refs to make sure we have the best chance at doing this all
3709          * in one shot.
3710          */
3711         spin_lock(&cur_trans->dirty_bgs_lock);
3712         while (!list_empty(&cur_trans->dirty_bgs)) {
3713                 cache = list_first_entry(&cur_trans->dirty_bgs,
3714                                          struct btrfs_block_group_cache,
3715                                          dirty_list);
3716
3717                 /*
3718                  * this can happen if cache_save_setup re-dirties a block
3719                  * group that is already under IO.  Just wait for it to
3720                  * finish and then do it all again
3721                  */
3722                 if (!list_empty(&cache->io_list)) {
3723                         spin_unlock(&cur_trans->dirty_bgs_lock);
3724                         list_del_init(&cache->io_list);
3725                         btrfs_wait_cache_io(trans, cache, path);
3726                         btrfs_put_block_group(cache);
3727                         spin_lock(&cur_trans->dirty_bgs_lock);
3728                 }
3729
3730                 /*
3731                  * don't remove from the dirty list until after we've waited
3732                  * on any pending IO
3733                  */
3734                 list_del_init(&cache->dirty_list);
3735                 spin_unlock(&cur_trans->dirty_bgs_lock);
3736                 should_put = 1;
3737
3738                 cache_save_setup(cache, trans, path);
3739
3740                 if (!ret)
3741                         ret = btrfs_run_delayed_refs(trans,
3742                                                      (unsigned long) -1);
3743
3744                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3745                         cache->io_ctl.inode = NULL;
3746                         ret = btrfs_write_out_cache(fs_info, trans,
3747                                                     cache, path);
3748                         if (ret == 0 && cache->io_ctl.inode) {
3749                                 num_started++;
3750                                 should_put = 0;
3751                                 list_add_tail(&cache->io_list, io);
3752                         } else {
3753                                 /*
3754                                  * if we failed to write the cache, the
3755                                  * generation will be bad and life goes on
3756                                  */
3757                                 ret = 0;
3758                         }
3759                 }
3760                 if (!ret) {
3761                         ret = write_one_cache_group(trans, fs_info,
3762                                                     path, cache);
3763                         /*
3764                          * One of the free space endio workers might have
3765                          * created a new block group while updating a free space
3766                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3767                          * and hasn't released its transaction handle yet, in
3768                          * which case the new block group is still attached to
3769                          * its transaction handle and its creation has not
3770                          * finished yet (no block group item in the extent tree
3771                          * yet, etc). If this is the case, wait for all free
3772                          * space endio workers to finish and retry. This is a
3773                          * a very rare case so no need for a more efficient and
3774                          * complex approach.
3775                          */
3776                         if (ret == -ENOENT) {
3777                                 wait_event(cur_trans->writer_wait,
3778                                    atomic_read(&cur_trans->num_writers) == 1);
3779                                 ret = write_one_cache_group(trans, fs_info,
3780                                                             path, cache);
3781                         }
3782                         if (ret)
3783                                 btrfs_abort_transaction(trans, ret);
3784                 }
3785
3786                 /* if its not on the io list, we need to put the block group */
3787                 if (should_put)
3788                         btrfs_put_block_group(cache);
3789                 spin_lock(&cur_trans->dirty_bgs_lock);
3790         }
3791         spin_unlock(&cur_trans->dirty_bgs_lock);
3792
3793         /*
3794          * Refer to the definition of io_bgs member for details why it's safe
3795          * to use it without any locking
3796          */
3797         while (!list_empty(io)) {
3798                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3799                                          io_list);
3800                 list_del_init(&cache->io_list);
3801                 btrfs_wait_cache_io(trans, cache, path);
3802                 btrfs_put_block_group(cache);
3803         }
3804
3805         btrfs_free_path(path);
3806         return ret;
3807 }
3808
3809 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3810 {
3811         struct btrfs_block_group_cache *block_group;
3812         int readonly = 0;
3813
3814         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3815         if (!block_group || block_group->ro)
3816                 readonly = 1;
3817         if (block_group)
3818                 btrfs_put_block_group(block_group);
3819         return readonly;
3820 }
3821
3822 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3823 {
3824         struct btrfs_block_group_cache *bg;
3825         bool ret = true;
3826
3827         bg = btrfs_lookup_block_group(fs_info, bytenr);
3828         if (!bg)
3829                 return false;
3830
3831         spin_lock(&bg->lock);
3832         if (bg->ro)
3833                 ret = false;
3834         else
3835                 atomic_inc(&bg->nocow_writers);
3836         spin_unlock(&bg->lock);
3837
3838         /* no put on block group, done by btrfs_dec_nocow_writers */
3839         if (!ret)
3840                 btrfs_put_block_group(bg);
3841
3842         return ret;
3843
3844 }
3845
3846 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3847 {
3848         struct btrfs_block_group_cache *bg;
3849
3850         bg = btrfs_lookup_block_group(fs_info, bytenr);
3851         ASSERT(bg);
3852         if (atomic_dec_and_test(&bg->nocow_writers))
3853                 wake_up_var(&bg->nocow_writers);
3854         /*
3855          * Once for our lookup and once for the lookup done by a previous call
3856          * to btrfs_inc_nocow_writers()
3857          */
3858         btrfs_put_block_group(bg);
3859         btrfs_put_block_group(bg);
3860 }
3861
3862 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3863 {
3864         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3865 }
3866
3867 static const char *alloc_name(u64 flags)
3868 {
3869         switch (flags) {
3870         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3871                 return "mixed";
3872         case BTRFS_BLOCK_GROUP_METADATA:
3873                 return "metadata";
3874         case BTRFS_BLOCK_GROUP_DATA:
3875                 return "data";
3876         case BTRFS_BLOCK_GROUP_SYSTEM:
3877                 return "system";
3878         default:
3879                 WARN_ON(1);
3880                 return "invalid-combination";
3881         };
3882 }
3883
3884 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3885 {
3886
3887         struct btrfs_space_info *space_info;
3888         int i;
3889         int ret;
3890
3891         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3892         if (!space_info)
3893                 return -ENOMEM;
3894
3895         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3896                                  GFP_KERNEL);
3897         if (ret) {
3898                 kfree(space_info);
3899                 return ret;
3900         }
3901
3902         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3903                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3904         init_rwsem(&space_info->groups_sem);
3905         spin_lock_init(&space_info->lock);
3906         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3907         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3908         init_waitqueue_head(&space_info->wait);
3909         INIT_LIST_HEAD(&space_info->ro_bgs);
3910         INIT_LIST_HEAD(&space_info->tickets);
3911         INIT_LIST_HEAD(&space_info->priority_tickets);
3912
3913         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3914                                     info->space_info_kobj, "%s",
3915                                     alloc_name(space_info->flags));
3916         if (ret) {
3917                 kobject_put(&space_info->kobj);
3918                 return ret;
3919         }
3920
3921         list_add_rcu(&space_info->list, &info->space_info);
3922         if (flags & BTRFS_BLOCK_GROUP_DATA)
3923                 info->data_sinfo = space_info;
3924
3925         return ret;
3926 }
3927
3928 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3929                              u64 total_bytes, u64 bytes_used,
3930                              u64 bytes_readonly,
3931                              struct btrfs_space_info **space_info)
3932 {
3933         struct btrfs_space_info *found;
3934         int factor;
3935
3936         factor = btrfs_bg_type_to_factor(flags);
3937
3938         found = __find_space_info(info, flags);
3939         ASSERT(found);
3940         spin_lock(&found->lock);
3941         found->total_bytes += total_bytes;
3942         found->disk_total += total_bytes * factor;
3943         found->bytes_used += bytes_used;
3944         found->disk_used += bytes_used * factor;
3945         found->bytes_readonly += bytes_readonly;
3946         if (total_bytes > 0)
3947                 found->full = 0;
3948         space_info_add_new_bytes(info, found, total_bytes -
3949                                  bytes_used - bytes_readonly);
3950         spin_unlock(&found->lock);
3951         *space_info = found;
3952 }
3953
3954 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3955 {
3956         u64 extra_flags = chunk_to_extended(flags) &
3957                                 BTRFS_EXTENDED_PROFILE_MASK;
3958
3959         write_seqlock(&fs_info->profiles_lock);
3960         if (flags & BTRFS_BLOCK_GROUP_DATA)
3961                 fs_info->avail_data_alloc_bits |= extra_flags;
3962         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3963                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3964         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3965                 fs_info->avail_system_alloc_bits |= extra_flags;
3966         write_sequnlock(&fs_info->profiles_lock);
3967 }
3968
3969 /*
3970  * returns target flags in extended format or 0 if restripe for this
3971  * chunk_type is not in progress
3972  *
3973  * should be called with balance_lock held
3974  */
3975 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3976 {
3977         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3978         u64 target = 0;
3979
3980         if (!bctl)
3981                 return 0;
3982
3983         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3984             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3985                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3986         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3987                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3988                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3989         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3990                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3991                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3992         }
3993
3994         return target;
3995 }
3996
3997 /*
3998  * @flags: available profiles in extended format (see ctree.h)
3999  *
4000  * Returns reduced profile in chunk format.  If profile changing is in
4001  * progress (either running or paused) picks the target profile (if it's
4002  * already available), otherwise falls back to plain reducing.
4003  */
4004 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4005 {
4006         u64 num_devices = fs_info->fs_devices->rw_devices;
4007         u64 target;
4008         u64 raid_type;
4009         u64 allowed = 0;
4010
4011         /*
4012          * see if restripe for this chunk_type is in progress, if so
4013          * try to reduce to the target profile
4014          */
4015         spin_lock(&fs_info->balance_lock);
4016         target = get_restripe_target(fs_info, flags);
4017         if (target) {
4018                 /* pick target profile only if it's already available */
4019                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4020                         spin_unlock(&fs_info->balance_lock);
4021                         return extended_to_chunk(target);
4022                 }
4023         }
4024         spin_unlock(&fs_info->balance_lock);
4025
4026         /* First, mask out the RAID levels which aren't possible */
4027         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4028                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4029                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4030         }
4031         allowed &= flags;
4032
4033         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4034                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4035         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4036                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4037         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4038                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4039         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4040                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4041         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4042                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4043
4044         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4045
4046         return extended_to_chunk(flags | allowed);
4047 }
4048
4049 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4050 {
4051         unsigned seq;
4052         u64 flags;
4053
4054         do {
4055                 flags = orig_flags;
4056                 seq = read_seqbegin(&fs_info->profiles_lock);
4057
4058                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4059                         flags |= fs_info->avail_data_alloc_bits;
4060                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4061                         flags |= fs_info->avail_system_alloc_bits;
4062                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4063                         flags |= fs_info->avail_metadata_alloc_bits;
4064         } while (read_seqretry(&fs_info->profiles_lock, seq));
4065
4066         return btrfs_reduce_alloc_profile(fs_info, flags);
4067 }
4068
4069 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4070 {
4071         struct btrfs_fs_info *fs_info = root->fs_info;
4072         u64 flags;
4073         u64 ret;
4074
4075         if (data)
4076                 flags = BTRFS_BLOCK_GROUP_DATA;
4077         else if (root == fs_info->chunk_root)
4078                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4079         else
4080                 flags = BTRFS_BLOCK_GROUP_METADATA;
4081
4082         ret = get_alloc_profile(fs_info, flags);
4083         return ret;
4084 }
4085
4086 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4087 {
4088         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4089 }
4090
4091 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4092 {
4093         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4094 }
4095
4096 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4097 {
4098         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4099 }
4100
4101 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4102                                  bool may_use_included)
4103 {
4104         ASSERT(s_info);
4105         return s_info->bytes_used + s_info->bytes_reserved +
4106                 s_info->bytes_pinned + s_info->bytes_readonly +
4107                 (may_use_included ? s_info->bytes_may_use : 0);
4108 }
4109
4110 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4111 {
4112         struct btrfs_root *root = inode->root;
4113         struct btrfs_fs_info *fs_info = root->fs_info;
4114         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4115         u64 used;
4116         int ret = 0;
4117         int need_commit = 2;
4118         int have_pinned_space;
4119
4120         /* make sure bytes are sectorsize aligned */
4121         bytes = ALIGN(bytes, fs_info->sectorsize);
4122
4123         if (btrfs_is_free_space_inode(inode)) {
4124                 need_commit = 0;
4125                 ASSERT(current->journal_info);
4126         }
4127
4128 again:
4129         /* make sure we have enough space to handle the data first */
4130         spin_lock(&data_sinfo->lock);
4131         used = btrfs_space_info_used(data_sinfo, true);
4132
4133         if (used + bytes > data_sinfo->total_bytes) {
4134                 struct btrfs_trans_handle *trans;
4135
4136                 /*
4137                  * if we don't have enough free bytes in this space then we need
4138                  * to alloc a new chunk.
4139                  */
4140                 if (!data_sinfo->full) {
4141                         u64 alloc_target;
4142
4143                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4144                         spin_unlock(&data_sinfo->lock);
4145
4146                         alloc_target = btrfs_data_alloc_profile(fs_info);
4147                         /*
4148                          * It is ugly that we don't call nolock join
4149                          * transaction for the free space inode case here.
4150                          * But it is safe because we only do the data space
4151                          * reservation for the free space cache in the
4152                          * transaction context, the common join transaction
4153                          * just increase the counter of the current transaction
4154                          * handler, doesn't try to acquire the trans_lock of
4155                          * the fs.
4156                          */
4157                         trans = btrfs_join_transaction(root);
4158                         if (IS_ERR(trans))
4159                                 return PTR_ERR(trans);
4160
4161                         ret = do_chunk_alloc(trans, alloc_target,
4162                                              CHUNK_ALLOC_NO_FORCE);
4163                         btrfs_end_transaction(trans);
4164                         if (ret < 0) {
4165                                 if (ret != -ENOSPC)
4166                                         return ret;
4167                                 else {
4168                                         have_pinned_space = 1;
4169                                         goto commit_trans;
4170                                 }
4171                         }
4172
4173                         goto again;
4174                 }
4175
4176                 /*
4177                  * If we don't have enough pinned space to deal with this
4178                  * allocation, and no removed chunk in current transaction,
4179                  * don't bother committing the transaction.
4180                  */
4181                 have_pinned_space = __percpu_counter_compare(
4182                         &data_sinfo->total_bytes_pinned,
4183                         used + bytes - data_sinfo->total_bytes,
4184                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4185                 spin_unlock(&data_sinfo->lock);
4186
4187                 /* commit the current transaction and try again */
4188 commit_trans:
4189                 if (need_commit) {
4190                         need_commit--;
4191
4192                         if (need_commit > 0) {
4193                                 btrfs_start_delalloc_roots(fs_info, -1);
4194                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4195                                                          (u64)-1);
4196                         }
4197
4198                         trans = btrfs_join_transaction(root);
4199                         if (IS_ERR(trans))
4200                                 return PTR_ERR(trans);
4201                         if (have_pinned_space >= 0 ||
4202                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4203                                      &trans->transaction->flags) ||
4204                             need_commit > 0) {
4205                                 ret = btrfs_commit_transaction(trans);
4206                                 if (ret)
4207                                         return ret;
4208                                 /*
4209                                  * The cleaner kthread might still be doing iput
4210                                  * operations. Wait for it to finish so that
4211                                  * more space is released.
4212                                  */
4213                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4214                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4215                                 goto again;
4216                         } else {
4217                                 btrfs_end_transaction(trans);
4218                         }
4219                 }
4220
4221                 trace_btrfs_space_reservation(fs_info,
4222                                               "space_info:enospc",
4223                                               data_sinfo->flags, bytes, 1);
4224                 return -ENOSPC;
4225         }
4226         data_sinfo->bytes_may_use += bytes;
4227         trace_btrfs_space_reservation(fs_info, "space_info",
4228                                       data_sinfo->flags, bytes, 1);
4229         spin_unlock(&data_sinfo->lock);
4230
4231         return 0;
4232 }
4233
4234 int btrfs_check_data_free_space(struct inode *inode,
4235                         struct extent_changeset **reserved, u64 start, u64 len)
4236 {
4237         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4238         int ret;
4239
4240         /* align the range */
4241         len = round_up(start + len, fs_info->sectorsize) -
4242               round_down(start, fs_info->sectorsize);
4243         start = round_down(start, fs_info->sectorsize);
4244
4245         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4246         if (ret < 0)
4247                 return ret;
4248
4249         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4250         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4251         if (ret < 0)
4252                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4253         else
4254                 ret = 0;
4255         return ret;
4256 }
4257
4258 /*
4259  * Called if we need to clear a data reservation for this inode
4260  * Normally in a error case.
4261  *
4262  * This one will *NOT* use accurate qgroup reserved space API, just for case
4263  * which we can't sleep and is sure it won't affect qgroup reserved space.
4264  * Like clear_bit_hook().
4265  */
4266 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4267                                             u64 len)
4268 {
4269         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4270         struct btrfs_space_info *data_sinfo;
4271
4272         /* Make sure the range is aligned to sectorsize */
4273         len = round_up(start + len, fs_info->sectorsize) -
4274               round_down(start, fs_info->sectorsize);
4275         start = round_down(start, fs_info->sectorsize);
4276
4277         data_sinfo = fs_info->data_sinfo;
4278         spin_lock(&data_sinfo->lock);
4279         if (WARN_ON(data_sinfo->bytes_may_use < len))
4280                 data_sinfo->bytes_may_use = 0;
4281         else
4282                 data_sinfo->bytes_may_use -= len;
4283         trace_btrfs_space_reservation(fs_info, "space_info",
4284                                       data_sinfo->flags, len, 0);
4285         spin_unlock(&data_sinfo->lock);
4286 }
4287
4288 /*
4289  * Called if we need to clear a data reservation for this inode
4290  * Normally in a error case.
4291  *
4292  * This one will handle the per-inode data rsv map for accurate reserved
4293  * space framework.
4294  */
4295 void btrfs_free_reserved_data_space(struct inode *inode,
4296                         struct extent_changeset *reserved, u64 start, u64 len)
4297 {
4298         struct btrfs_root *root = BTRFS_I(inode)->root;
4299
4300         /* Make sure the range is aligned to sectorsize */
4301         len = round_up(start + len, root->fs_info->sectorsize) -
4302               round_down(start, root->fs_info->sectorsize);
4303         start = round_down(start, root->fs_info->sectorsize);
4304
4305         btrfs_free_reserved_data_space_noquota(inode, start, len);
4306         btrfs_qgroup_free_data(inode, reserved, start, len);
4307 }
4308
4309 static void force_metadata_allocation(struct btrfs_fs_info *info)
4310 {
4311         struct list_head *head = &info->space_info;
4312         struct btrfs_space_info *found;
4313
4314         rcu_read_lock();
4315         list_for_each_entry_rcu(found, head, list) {
4316                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4317                         found->force_alloc = CHUNK_ALLOC_FORCE;
4318         }
4319         rcu_read_unlock();
4320 }
4321
4322 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4323 {
4324         return (global->size << 1);
4325 }
4326
4327 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4328                               struct btrfs_space_info *sinfo, int force)
4329 {
4330         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4331         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4332         u64 thresh;
4333
4334         if (force == CHUNK_ALLOC_FORCE)
4335                 return 1;
4336
4337         /*
4338          * We need to take into account the global rsv because for all intents
4339          * and purposes it's used space.  Don't worry about locking the
4340          * global_rsv, it doesn't change except when the transaction commits.
4341          */
4342         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4343                 bytes_used += calc_global_rsv_need_space(global_rsv);
4344
4345         /*
4346          * in limited mode, we want to have some free space up to
4347          * about 1% of the FS size.
4348          */
4349         if (force == CHUNK_ALLOC_LIMITED) {
4350                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4351                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4352
4353                 if (sinfo->total_bytes - bytes_used < thresh)
4354                         return 1;
4355         }
4356
4357         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4358                 return 0;
4359         return 1;
4360 }
4361
4362 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4363 {
4364         u64 num_dev;
4365
4366         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4367                     BTRFS_BLOCK_GROUP_RAID0 |
4368                     BTRFS_BLOCK_GROUP_RAID5 |
4369                     BTRFS_BLOCK_GROUP_RAID6))
4370                 num_dev = fs_info->fs_devices->rw_devices;
4371         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4372                 num_dev = 2;
4373         else
4374                 num_dev = 1;    /* DUP or single */
4375
4376         return num_dev;
4377 }
4378
4379 /*
4380  * If @is_allocation is true, reserve space in the system space info necessary
4381  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4382  * removing a chunk.
4383  */
4384 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4385 {
4386         struct btrfs_fs_info *fs_info = trans->fs_info;
4387         struct btrfs_space_info *info;
4388         u64 left;
4389         u64 thresh;
4390         int ret = 0;
4391         u64 num_devs;
4392
4393         /*
4394          * Needed because we can end up allocating a system chunk and for an
4395          * atomic and race free space reservation in the chunk block reserve.
4396          */
4397         lockdep_assert_held(&fs_info->chunk_mutex);
4398
4399         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4400         spin_lock(&info->lock);
4401         left = info->total_bytes - btrfs_space_info_used(info, true);
4402         spin_unlock(&info->lock);
4403
4404         num_devs = get_profile_num_devs(fs_info, type);
4405
4406         /* num_devs device items to update and 1 chunk item to add or remove */
4407         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4408                 btrfs_calc_trans_metadata_size(fs_info, 1);
4409
4410         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4411                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4412                            left, thresh, type);
4413                 dump_space_info(fs_info, info, 0, 0);
4414         }
4415
4416         if (left < thresh) {
4417                 u64 flags = btrfs_system_alloc_profile(fs_info);
4418
4419                 /*
4420                  * Ignore failure to create system chunk. We might end up not
4421                  * needing it, as we might not need to COW all nodes/leafs from
4422                  * the paths we visit in the chunk tree (they were already COWed
4423                  * or created in the current transaction for example).
4424                  */
4425                 ret = btrfs_alloc_chunk(trans, flags);
4426         }
4427
4428         if (!ret) {
4429                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4430                                           &fs_info->chunk_block_rsv,
4431                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4432                 if (!ret)
4433                         trans->chunk_bytes_reserved += thresh;
4434         }
4435 }
4436
4437 /*
4438  * If force is CHUNK_ALLOC_FORCE:
4439  *    - return 1 if it successfully allocates a chunk,
4440  *    - return errors including -ENOSPC otherwise.
4441  * If force is NOT CHUNK_ALLOC_FORCE:
4442  *    - return 0 if it doesn't need to allocate a new chunk,
4443  *    - return 1 if it successfully allocates a chunk,
4444  *    - return errors including -ENOSPC otherwise.
4445  */
4446 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4447                           int force)
4448 {
4449         struct btrfs_fs_info *fs_info = trans->fs_info;
4450         struct btrfs_space_info *space_info;
4451         bool wait_for_alloc = false;
4452         bool should_alloc = false;
4453         int ret = 0;
4454
4455         /* Don't re-enter if we're already allocating a chunk */
4456         if (trans->allocating_chunk)
4457                 return -ENOSPC;
4458
4459         space_info = __find_space_info(fs_info, flags);
4460         ASSERT(space_info);
4461
4462         do {
4463                 spin_lock(&space_info->lock);
4464                 if (force < space_info->force_alloc)
4465                         force = space_info->force_alloc;
4466                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4467                 if (space_info->full) {
4468                         /* No more free physical space */
4469                         if (should_alloc)
4470                                 ret = -ENOSPC;
4471                         else
4472                                 ret = 0;
4473                         spin_unlock(&space_info->lock);
4474                         return ret;
4475                 } else if (!should_alloc) {
4476                         spin_unlock(&space_info->lock);
4477                         return 0;
4478                 } else if (space_info->chunk_alloc) {
4479                         /*
4480                          * Someone is already allocating, so we need to block
4481                          * until this someone is finished and then loop to
4482                          * recheck if we should continue with our allocation
4483                          * attempt.
4484                          */
4485                         wait_for_alloc = true;
4486                         spin_unlock(&space_info->lock);
4487                         mutex_lock(&fs_info->chunk_mutex);
4488                         mutex_unlock(&fs_info->chunk_mutex);
4489                 } else {
4490                         /* Proceed with allocation */
4491                         space_info->chunk_alloc = 1;
4492                         wait_for_alloc = false;
4493                         spin_unlock(&space_info->lock);
4494                 }
4495
4496                 cond_resched();
4497         } while (wait_for_alloc);
4498
4499         mutex_lock(&fs_info->chunk_mutex);
4500         trans->allocating_chunk = true;
4501
4502         /*
4503          * If we have mixed data/metadata chunks we want to make sure we keep
4504          * allocating mixed chunks instead of individual chunks.
4505          */
4506         if (btrfs_mixed_space_info(space_info))
4507                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4508
4509         /*
4510          * if we're doing a data chunk, go ahead and make sure that
4511          * we keep a reasonable number of metadata chunks allocated in the
4512          * FS as well.
4513          */
4514         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4515                 fs_info->data_chunk_allocations++;
4516                 if (!(fs_info->data_chunk_allocations %
4517                       fs_info->metadata_ratio))
4518                         force_metadata_allocation(fs_info);
4519         }
4520
4521         /*
4522          * Check if we have enough space in SYSTEM chunk because we may need
4523          * to update devices.
4524          */
4525         check_system_chunk(trans, flags);
4526
4527         ret = btrfs_alloc_chunk(trans, flags);
4528         trans->allocating_chunk = false;
4529
4530         spin_lock(&space_info->lock);
4531         if (ret < 0) {
4532                 if (ret == -ENOSPC)
4533                         space_info->full = 1;
4534                 else
4535                         goto out;
4536         } else {
4537                 ret = 1;
4538                 space_info->max_extent_size = 0;
4539         }
4540
4541         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4542 out:
4543         space_info->chunk_alloc = 0;
4544         spin_unlock(&space_info->lock);
4545         mutex_unlock(&fs_info->chunk_mutex);
4546         /*
4547          * When we allocate a new chunk we reserve space in the chunk block
4548          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4549          * add new nodes/leafs to it if we end up needing to do it when
4550          * inserting the chunk item and updating device items as part of the
4551          * second phase of chunk allocation, performed by
4552          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4553          * large number of new block groups to create in our transaction
4554          * handle's new_bgs list to avoid exhausting the chunk block reserve
4555          * in extreme cases - like having a single transaction create many new
4556          * block groups when starting to write out the free space caches of all
4557          * the block groups that were made dirty during the lifetime of the
4558          * transaction.
4559          */
4560         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4561                 btrfs_create_pending_block_groups(trans);
4562
4563         return ret;
4564 }
4565
4566 static int can_overcommit(struct btrfs_fs_info *fs_info,
4567                           struct btrfs_space_info *space_info, u64 bytes,
4568                           enum btrfs_reserve_flush_enum flush,
4569                           bool system_chunk)
4570 {
4571         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4572         u64 profile;
4573         u64 space_size;
4574         u64 avail;
4575         u64 used;
4576         int factor;
4577
4578         /* Don't overcommit when in mixed mode. */
4579         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4580                 return 0;
4581
4582         if (system_chunk)
4583                 profile = btrfs_system_alloc_profile(fs_info);
4584         else
4585                 profile = btrfs_metadata_alloc_profile(fs_info);
4586
4587         used = btrfs_space_info_used(space_info, false);
4588
4589         /*
4590          * We only want to allow over committing if we have lots of actual space
4591          * free, but if we don't have enough space to handle the global reserve
4592          * space then we could end up having a real enospc problem when trying
4593          * to allocate a chunk or some other such important allocation.
4594          */
4595         spin_lock(&global_rsv->lock);
4596         space_size = calc_global_rsv_need_space(global_rsv);
4597         spin_unlock(&global_rsv->lock);
4598         if (used + space_size >= space_info->total_bytes)
4599                 return 0;
4600
4601         used += space_info->bytes_may_use;
4602
4603         avail = atomic64_read(&fs_info->free_chunk_space);
4604
4605         /*
4606          * If we have dup, raid1 or raid10 then only half of the free
4607          * space is actually useable.  For raid56, the space info used
4608          * doesn't include the parity drive, so we don't have to
4609          * change the math
4610          */
4611         factor = btrfs_bg_type_to_factor(profile);
4612         avail = div_u64(avail, factor);
4613
4614         /*
4615          * If we aren't flushing all things, let us overcommit up to
4616          * 1/2th of the space. If we can flush, don't let us overcommit
4617          * too much, let it overcommit up to 1/8 of the space.
4618          */
4619         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4620                 avail >>= 3;
4621         else
4622                 avail >>= 1;
4623
4624         if (used + bytes < space_info->total_bytes + avail)
4625                 return 1;
4626         return 0;
4627 }
4628
4629 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4630                                          unsigned long nr_pages, int nr_items)
4631 {
4632         struct super_block *sb = fs_info->sb;
4633
4634         if (down_read_trylock(&sb->s_umount)) {
4635                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4636                 up_read(&sb->s_umount);
4637         } else {
4638                 /*
4639                  * We needn't worry the filesystem going from r/w to r/o though
4640                  * we don't acquire ->s_umount mutex, because the filesystem
4641                  * should guarantee the delalloc inodes list be empty after
4642                  * the filesystem is readonly(all dirty pages are written to
4643                  * the disk).
4644                  */
4645                 btrfs_start_delalloc_roots(fs_info, nr_items);
4646                 if (!current->journal_info)
4647                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4648         }
4649 }
4650
4651 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4652                                         u64 to_reclaim)
4653 {
4654         u64 bytes;
4655         u64 nr;
4656
4657         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4658         nr = div64_u64(to_reclaim, bytes);
4659         if (!nr)
4660                 nr = 1;
4661         return nr;
4662 }
4663
4664 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4665
4666 /*
4667  * shrink metadata reservation for delalloc
4668  */
4669 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4670                             u64 orig, bool wait_ordered)
4671 {
4672         struct btrfs_space_info *space_info;
4673         struct btrfs_trans_handle *trans;
4674         u64 delalloc_bytes;
4675         u64 max_reclaim;
4676         u64 items;
4677         long time_left;
4678         unsigned long nr_pages;
4679         int loops;
4680
4681         /* Calc the number of the pages we need flush for space reservation */
4682         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4683         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4684
4685         trans = (struct btrfs_trans_handle *)current->journal_info;
4686         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4687
4688         delalloc_bytes = percpu_counter_sum_positive(
4689                                                 &fs_info->delalloc_bytes);
4690         if (delalloc_bytes == 0) {
4691                 if (trans)
4692                         return;
4693                 if (wait_ordered)
4694                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4695                 return;
4696         }
4697
4698         loops = 0;
4699         while (delalloc_bytes && loops < 3) {
4700                 max_reclaim = min(delalloc_bytes, to_reclaim);
4701                 nr_pages = max_reclaim >> PAGE_SHIFT;
4702                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4703                 /*
4704                  * We need to wait for the async pages to actually start before
4705                  * we do anything.
4706                  */
4707                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4708                 if (!max_reclaim)
4709                         goto skip_async;
4710
4711                 if (max_reclaim <= nr_pages)
4712                         max_reclaim = 0;
4713                 else
4714                         max_reclaim -= nr_pages;
4715
4716                 wait_event(fs_info->async_submit_wait,
4717                            atomic_read(&fs_info->async_delalloc_pages) <=
4718                            (int)max_reclaim);
4719 skip_async:
4720                 spin_lock(&space_info->lock);
4721                 if (list_empty(&space_info->tickets) &&
4722                     list_empty(&space_info->priority_tickets)) {
4723                         spin_unlock(&space_info->lock);
4724                         break;
4725                 }
4726                 spin_unlock(&space_info->lock);
4727
4728                 loops++;
4729                 if (wait_ordered && !trans) {
4730                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4731                 } else {
4732                         time_left = schedule_timeout_killable(1);
4733                         if (time_left)
4734                                 break;
4735                 }
4736                 delalloc_bytes = percpu_counter_sum_positive(
4737                                                 &fs_info->delalloc_bytes);
4738         }
4739 }
4740
4741 struct reserve_ticket {
4742         u64 bytes;
4743         int error;
4744         struct list_head list;
4745         wait_queue_head_t wait;
4746 };
4747
4748 /**
4749  * maybe_commit_transaction - possibly commit the transaction if its ok to
4750  * @root - the root we're allocating for
4751  * @bytes - the number of bytes we want to reserve
4752  * @force - force the commit
4753  *
4754  * This will check to make sure that committing the transaction will actually
4755  * get us somewhere and then commit the transaction if it does.  Otherwise it
4756  * will return -ENOSPC.
4757  */
4758 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4759                                   struct btrfs_space_info *space_info)
4760 {
4761         struct reserve_ticket *ticket = NULL;
4762         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4763         struct btrfs_trans_handle *trans;
4764         u64 bytes;
4765
4766         trans = (struct btrfs_trans_handle *)current->journal_info;
4767         if (trans)
4768                 return -EAGAIN;
4769
4770         spin_lock(&space_info->lock);
4771         if (!list_empty(&space_info->priority_tickets))
4772                 ticket = list_first_entry(&space_info->priority_tickets,
4773                                           struct reserve_ticket, list);
4774         else if (!list_empty(&space_info->tickets))
4775                 ticket = list_first_entry(&space_info->tickets,
4776                                           struct reserve_ticket, list);
4777         bytes = (ticket) ? ticket->bytes : 0;
4778         spin_unlock(&space_info->lock);
4779
4780         if (!bytes)
4781                 return 0;
4782
4783         /* See if there is enough pinned space to make this reservation */
4784         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4785                                    bytes,
4786                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4787                 goto commit;
4788
4789         /*
4790          * See if there is some space in the delayed insertion reservation for
4791          * this reservation.
4792          */
4793         if (space_info != delayed_rsv->space_info)
4794                 return -ENOSPC;
4795
4796         spin_lock(&delayed_rsv->lock);
4797         if (delayed_rsv->size > bytes)
4798                 bytes = 0;
4799         else
4800                 bytes -= delayed_rsv->size;
4801         spin_unlock(&delayed_rsv->lock);
4802
4803         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4804                                    bytes,
4805                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) {
4806                 return -ENOSPC;
4807         }
4808
4809 commit:
4810         trans = btrfs_join_transaction(fs_info->extent_root);
4811         if (IS_ERR(trans))
4812                 return -ENOSPC;
4813
4814         return btrfs_commit_transaction(trans);
4815 }
4816
4817 /*
4818  * Try to flush some data based on policy set by @state. This is only advisory
4819  * and may fail for various reasons. The caller is supposed to examine the
4820  * state of @space_info to detect the outcome.
4821  */
4822 static void flush_space(struct btrfs_fs_info *fs_info,
4823                        struct btrfs_space_info *space_info, u64 num_bytes,
4824                        int state)
4825 {
4826         struct btrfs_root *root = fs_info->extent_root;
4827         struct btrfs_trans_handle *trans;
4828         int nr;
4829         int ret = 0;
4830
4831         switch (state) {
4832         case FLUSH_DELAYED_ITEMS_NR:
4833         case FLUSH_DELAYED_ITEMS:
4834                 if (state == FLUSH_DELAYED_ITEMS_NR)
4835                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4836                 else
4837                         nr = -1;
4838
4839                 trans = btrfs_join_transaction(root);
4840                 if (IS_ERR(trans)) {
4841                         ret = PTR_ERR(trans);
4842                         break;
4843                 }
4844                 ret = btrfs_run_delayed_items_nr(trans, nr);
4845                 btrfs_end_transaction(trans);
4846                 break;
4847         case FLUSH_DELALLOC:
4848         case FLUSH_DELALLOC_WAIT:
4849                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4850                                 state == FLUSH_DELALLOC_WAIT);
4851                 break;
4852         case ALLOC_CHUNK:
4853                 trans = btrfs_join_transaction(root);
4854                 if (IS_ERR(trans)) {
4855                         ret = PTR_ERR(trans);
4856                         break;
4857                 }
4858                 ret = do_chunk_alloc(trans,
4859                                      btrfs_metadata_alloc_profile(fs_info),
4860                                      CHUNK_ALLOC_NO_FORCE);
4861                 btrfs_end_transaction(trans);
4862                 if (ret > 0 || ret == -ENOSPC)
4863                         ret = 0;
4864                 break;
4865         case COMMIT_TRANS:
4866                 ret = may_commit_transaction(fs_info, space_info);
4867                 break;
4868         default:
4869                 ret = -ENOSPC;
4870                 break;
4871         }
4872
4873         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4874                                 ret);
4875         return;
4876 }
4877
4878 static inline u64
4879 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4880                                  struct btrfs_space_info *space_info,
4881                                  bool system_chunk)
4882 {
4883         struct reserve_ticket *ticket;
4884         u64 used;
4885         u64 expected;
4886         u64 to_reclaim = 0;
4887
4888         list_for_each_entry(ticket, &space_info->tickets, list)
4889                 to_reclaim += ticket->bytes;
4890         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4891                 to_reclaim += ticket->bytes;
4892         if (to_reclaim)
4893                 return to_reclaim;
4894
4895         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4896         if (can_overcommit(fs_info, space_info, to_reclaim,
4897                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4898                 return 0;
4899
4900         used = btrfs_space_info_used(space_info, true);
4901
4902         if (can_overcommit(fs_info, space_info, SZ_1M,
4903                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4904                 expected = div_factor_fine(space_info->total_bytes, 95);
4905         else
4906                 expected = div_factor_fine(space_info->total_bytes, 90);
4907
4908         if (used > expected)
4909                 to_reclaim = used - expected;
4910         else
4911                 to_reclaim = 0;
4912         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4913                                      space_info->bytes_reserved);
4914         return to_reclaim;
4915 }
4916
4917 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4918                                         struct btrfs_space_info *space_info,
4919                                         u64 used, bool system_chunk)
4920 {
4921         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4922
4923         /* If we're just plain full then async reclaim just slows us down. */
4924         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4925                 return 0;
4926
4927         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4928                                               system_chunk))
4929                 return 0;
4930
4931         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4932                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4933 }
4934
4935 static void wake_all_tickets(struct list_head *head)
4936 {
4937         struct reserve_ticket *ticket;
4938
4939         while (!list_empty(head)) {
4940                 ticket = list_first_entry(head, struct reserve_ticket, list);
4941                 list_del_init(&ticket->list);
4942                 ticket->error = -ENOSPC;
4943                 wake_up(&ticket->wait);
4944         }
4945 }
4946
4947 /*
4948  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4949  * will loop and continuously try to flush as long as we are making progress.
4950  * We count progress as clearing off tickets each time we have to loop.
4951  */
4952 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4953 {
4954         struct btrfs_fs_info *fs_info;
4955         struct btrfs_space_info *space_info;
4956         u64 to_reclaim;
4957         int flush_state;
4958         int commit_cycles = 0;
4959         u64 last_tickets_id;
4960
4961         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4962         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4963
4964         spin_lock(&space_info->lock);
4965         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4966                                                       false);
4967         if (!to_reclaim) {
4968                 space_info->flush = 0;
4969                 spin_unlock(&space_info->lock);
4970                 return;
4971         }
4972         last_tickets_id = space_info->tickets_id;
4973         spin_unlock(&space_info->lock);
4974
4975         flush_state = FLUSH_DELAYED_ITEMS_NR;
4976         do {
4977                 flush_space(fs_info, space_info, to_reclaim, flush_state);
4978                 spin_lock(&space_info->lock);
4979                 if (list_empty(&space_info->tickets)) {
4980                         space_info->flush = 0;
4981                         spin_unlock(&space_info->lock);
4982                         return;
4983                 }
4984                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
4985                                                               space_info,
4986                                                               false);
4987                 if (last_tickets_id == space_info->tickets_id) {
4988                         flush_state++;
4989                 } else {
4990                         last_tickets_id = space_info->tickets_id;
4991                         flush_state = FLUSH_DELAYED_ITEMS_NR;
4992                         if (commit_cycles)
4993                                 commit_cycles--;
4994                 }
4995
4996                 if (flush_state > COMMIT_TRANS) {
4997                         commit_cycles++;
4998                         if (commit_cycles > 2) {
4999                                 wake_all_tickets(&space_info->tickets);
5000                                 space_info->flush = 0;
5001                         } else {
5002                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5003                         }
5004                 }
5005                 spin_unlock(&space_info->lock);
5006         } while (flush_state <= COMMIT_TRANS);
5007 }
5008
5009 void btrfs_init_async_reclaim_work(struct work_struct *work)
5010 {
5011         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5012 }
5013
5014 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5015                                             struct btrfs_space_info *space_info,
5016                                             struct reserve_ticket *ticket)
5017 {
5018         u64 to_reclaim;
5019         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5020
5021         spin_lock(&space_info->lock);
5022         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5023                                                       false);
5024         if (!to_reclaim) {
5025                 spin_unlock(&space_info->lock);
5026                 return;
5027         }
5028         spin_unlock(&space_info->lock);
5029
5030         do {
5031                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5032                 flush_state++;
5033                 spin_lock(&space_info->lock);
5034                 if (ticket->bytes == 0) {
5035                         spin_unlock(&space_info->lock);
5036                         return;
5037                 }
5038                 spin_unlock(&space_info->lock);
5039
5040                 /*
5041                  * Priority flushers can't wait on delalloc without
5042                  * deadlocking.
5043                  */
5044                 if (flush_state == FLUSH_DELALLOC ||
5045                     flush_state == FLUSH_DELALLOC_WAIT)
5046                         flush_state = ALLOC_CHUNK;
5047         } while (flush_state < COMMIT_TRANS);
5048 }
5049
5050 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5051                                struct btrfs_space_info *space_info,
5052                                struct reserve_ticket *ticket, u64 orig_bytes)
5053
5054 {
5055         DEFINE_WAIT(wait);
5056         int ret = 0;
5057
5058         spin_lock(&space_info->lock);
5059         while (ticket->bytes > 0 && ticket->error == 0) {
5060                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5061                 if (ret) {
5062                         ret = -EINTR;
5063                         break;
5064                 }
5065                 spin_unlock(&space_info->lock);
5066
5067                 schedule();
5068
5069                 finish_wait(&ticket->wait, &wait);
5070                 spin_lock(&space_info->lock);
5071         }
5072         if (!ret)
5073                 ret = ticket->error;
5074         if (!list_empty(&ticket->list))
5075                 list_del_init(&ticket->list);
5076         if (ticket->bytes && ticket->bytes < orig_bytes) {
5077                 u64 num_bytes = orig_bytes - ticket->bytes;
5078                 space_info->bytes_may_use -= num_bytes;
5079                 trace_btrfs_space_reservation(fs_info, "space_info",
5080                                               space_info->flags, num_bytes, 0);
5081         }
5082         spin_unlock(&space_info->lock);
5083
5084         return ret;
5085 }
5086
5087 /**
5088  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5089  * @root - the root we're allocating for
5090  * @space_info - the space info we want to allocate from
5091  * @orig_bytes - the number of bytes we want
5092  * @flush - whether or not we can flush to make our reservation
5093  *
5094  * This will reserve orig_bytes number of bytes from the space info associated
5095  * with the block_rsv.  If there is not enough space it will make an attempt to
5096  * flush out space to make room.  It will do this by flushing delalloc if
5097  * possible or committing the transaction.  If flush is 0 then no attempts to
5098  * regain reservations will be made and this will fail if there is not enough
5099  * space already.
5100  */
5101 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5102                                     struct btrfs_space_info *space_info,
5103                                     u64 orig_bytes,
5104                                     enum btrfs_reserve_flush_enum flush,
5105                                     bool system_chunk)
5106 {
5107         struct reserve_ticket ticket;
5108         u64 used;
5109         int ret = 0;
5110
5111         ASSERT(orig_bytes);
5112         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5113
5114         spin_lock(&space_info->lock);
5115         ret = -ENOSPC;
5116         used = btrfs_space_info_used(space_info, true);
5117
5118         /*
5119          * If we have enough space then hooray, make our reservation and carry
5120          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5121          * If not things get more complicated.
5122          */
5123         if (used + orig_bytes <= space_info->total_bytes) {
5124                 space_info->bytes_may_use += orig_bytes;
5125                 trace_btrfs_space_reservation(fs_info, "space_info",
5126                                               space_info->flags, orig_bytes, 1);
5127                 ret = 0;
5128         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5129                                   system_chunk)) {
5130                 space_info->bytes_may_use += orig_bytes;
5131                 trace_btrfs_space_reservation(fs_info, "space_info",
5132                                               space_info->flags, orig_bytes, 1);
5133                 ret = 0;
5134         }
5135
5136         /*
5137          * If we couldn't make a reservation then setup our reservation ticket
5138          * and kick the async worker if it's not already running.
5139          *
5140          * If we are a priority flusher then we just need to add our ticket to
5141          * the list and we will do our own flushing further down.
5142          */
5143         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5144                 ticket.bytes = orig_bytes;
5145                 ticket.error = 0;
5146                 init_waitqueue_head(&ticket.wait);
5147                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5148                         list_add_tail(&ticket.list, &space_info->tickets);
5149                         if (!space_info->flush) {
5150                                 space_info->flush = 1;
5151                                 trace_btrfs_trigger_flush(fs_info,
5152                                                           space_info->flags,
5153                                                           orig_bytes, flush,
5154                                                           "enospc");
5155                                 queue_work(system_unbound_wq,
5156                                            &fs_info->async_reclaim_work);
5157                         }
5158                 } else {
5159                         list_add_tail(&ticket.list,
5160                                       &space_info->priority_tickets);
5161                 }
5162         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5163                 used += orig_bytes;
5164                 /*
5165                  * We will do the space reservation dance during log replay,
5166                  * which means we won't have fs_info->fs_root set, so don't do
5167                  * the async reclaim as we will panic.
5168                  */
5169                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5170                     need_do_async_reclaim(fs_info, space_info,
5171                                           used, system_chunk) &&
5172                     !work_busy(&fs_info->async_reclaim_work)) {
5173                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5174                                                   orig_bytes, flush, "preempt");
5175                         queue_work(system_unbound_wq,
5176                                    &fs_info->async_reclaim_work);
5177                 }
5178         }
5179         spin_unlock(&space_info->lock);
5180         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5181                 return ret;
5182
5183         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5184                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5185                                            orig_bytes);
5186
5187         ret = 0;
5188         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5189         spin_lock(&space_info->lock);
5190         if (ticket.bytes) {
5191                 if (ticket.bytes < orig_bytes) {
5192                         u64 num_bytes = orig_bytes - ticket.bytes;
5193                         space_info->bytes_may_use -= num_bytes;
5194                         trace_btrfs_space_reservation(fs_info, "space_info",
5195                                                       space_info->flags,
5196                                                       num_bytes, 0);
5197
5198                 }
5199                 list_del_init(&ticket.list);
5200                 ret = -ENOSPC;
5201         }
5202         spin_unlock(&space_info->lock);
5203         ASSERT(list_empty(&ticket.list));
5204         return ret;
5205 }
5206
5207 /**
5208  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5209  * @root - the root we're allocating for
5210  * @block_rsv - the block_rsv we're allocating for
5211  * @orig_bytes - the number of bytes we want
5212  * @flush - whether or not we can flush to make our reservation
5213  *
5214  * This will reserve orgi_bytes number of bytes from the space info associated
5215  * with the block_rsv.  If there is not enough space it will make an attempt to
5216  * flush out space to make room.  It will do this by flushing delalloc if
5217  * possible or committing the transaction.  If flush is 0 then no attempts to
5218  * regain reservations will be made and this will fail if there is not enough
5219  * space already.
5220  */
5221 static int reserve_metadata_bytes(struct btrfs_root *root,
5222                                   struct btrfs_block_rsv *block_rsv,
5223                                   u64 orig_bytes,
5224                                   enum btrfs_reserve_flush_enum flush)
5225 {
5226         struct btrfs_fs_info *fs_info = root->fs_info;
5227         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5228         int ret;
5229         bool system_chunk = (root == fs_info->chunk_root);
5230
5231         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5232                                        orig_bytes, flush, system_chunk);
5233         if (ret == -ENOSPC &&
5234             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5235                 if (block_rsv != global_rsv &&
5236                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5237                         ret = 0;
5238         }
5239         if (ret == -ENOSPC) {
5240                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5241                                               block_rsv->space_info->flags,
5242                                               orig_bytes, 1);
5243
5244                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5245                         dump_space_info(fs_info, block_rsv->space_info,
5246                                         orig_bytes, 0);
5247         }
5248         return ret;
5249 }
5250
5251 static struct btrfs_block_rsv *get_block_rsv(
5252                                         const struct btrfs_trans_handle *trans,
5253                                         const struct btrfs_root *root)
5254 {
5255         struct btrfs_fs_info *fs_info = root->fs_info;
5256         struct btrfs_block_rsv *block_rsv = NULL;
5257
5258         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5259             (root == fs_info->csum_root && trans->adding_csums) ||
5260             (root == fs_info->uuid_root))
5261                 block_rsv = trans->block_rsv;
5262
5263         if (!block_rsv)
5264                 block_rsv = root->block_rsv;
5265
5266         if (!block_rsv)
5267                 block_rsv = &fs_info->empty_block_rsv;
5268
5269         return block_rsv;
5270 }
5271
5272 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5273                                u64 num_bytes)
5274 {
5275         int ret = -ENOSPC;
5276         spin_lock(&block_rsv->lock);
5277         if (block_rsv->reserved >= num_bytes) {
5278                 block_rsv->reserved -= num_bytes;
5279                 if (block_rsv->reserved < block_rsv->size)
5280                         block_rsv->full = 0;
5281                 ret = 0;
5282         }
5283         spin_unlock(&block_rsv->lock);
5284         return ret;
5285 }
5286
5287 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5288                                 u64 num_bytes, int update_size)
5289 {
5290         spin_lock(&block_rsv->lock);
5291         block_rsv->reserved += num_bytes;
5292         if (update_size)
5293                 block_rsv->size += num_bytes;
5294         else if (block_rsv->reserved >= block_rsv->size)
5295                 block_rsv->full = 1;
5296         spin_unlock(&block_rsv->lock);
5297 }
5298
5299 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5300                              struct btrfs_block_rsv *dest, u64 num_bytes,
5301                              int min_factor)
5302 {
5303         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5304         u64 min_bytes;
5305
5306         if (global_rsv->space_info != dest->space_info)
5307                 return -ENOSPC;
5308
5309         spin_lock(&global_rsv->lock);
5310         min_bytes = div_factor(global_rsv->size, min_factor);
5311         if (global_rsv->reserved < min_bytes + num_bytes) {
5312                 spin_unlock(&global_rsv->lock);
5313                 return -ENOSPC;
5314         }
5315         global_rsv->reserved -= num_bytes;
5316         if (global_rsv->reserved < global_rsv->size)
5317                 global_rsv->full = 0;
5318         spin_unlock(&global_rsv->lock);
5319
5320         block_rsv_add_bytes(dest, num_bytes, 1);
5321         return 0;
5322 }
5323
5324 /*
5325  * This is for space we already have accounted in space_info->bytes_may_use, so
5326  * basically when we're returning space from block_rsv's.
5327  */
5328 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5329                                      struct btrfs_space_info *space_info,
5330                                      u64 num_bytes)
5331 {
5332         struct reserve_ticket *ticket;
5333         struct list_head *head;
5334         u64 used;
5335         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5336         bool check_overcommit = false;
5337
5338         spin_lock(&space_info->lock);
5339         head = &space_info->priority_tickets;
5340
5341         /*
5342          * If we are over our limit then we need to check and see if we can
5343          * overcommit, and if we can't then we just need to free up our space
5344          * and not satisfy any requests.
5345          */
5346         used = btrfs_space_info_used(space_info, true);
5347         if (used - num_bytes >= space_info->total_bytes)
5348                 check_overcommit = true;
5349 again:
5350         while (!list_empty(head) && num_bytes) {
5351                 ticket = list_first_entry(head, struct reserve_ticket,
5352                                           list);
5353                 /*
5354                  * We use 0 bytes because this space is already reserved, so
5355                  * adding the ticket space would be a double count.
5356                  */
5357                 if (check_overcommit &&
5358                     !can_overcommit(fs_info, space_info, 0, flush, false))
5359                         break;
5360                 if (num_bytes >= ticket->bytes) {
5361                         list_del_init(&ticket->list);
5362                         num_bytes -= ticket->bytes;
5363                         ticket->bytes = 0;
5364                         space_info->tickets_id++;
5365                         wake_up(&ticket->wait);
5366                 } else {
5367                         ticket->bytes -= num_bytes;
5368                         num_bytes = 0;
5369                 }
5370         }
5371
5372         if (num_bytes && head == &space_info->priority_tickets) {
5373                 head = &space_info->tickets;
5374                 flush = BTRFS_RESERVE_FLUSH_ALL;
5375                 goto again;
5376         }
5377         space_info->bytes_may_use -= num_bytes;
5378         trace_btrfs_space_reservation(fs_info, "space_info",
5379                                       space_info->flags, num_bytes, 0);
5380         spin_unlock(&space_info->lock);
5381 }
5382
5383 /*
5384  * This is for newly allocated space that isn't accounted in
5385  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5386  * we use this helper.
5387  */
5388 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5389                                      struct btrfs_space_info *space_info,
5390                                      u64 num_bytes)
5391 {
5392         struct reserve_ticket *ticket;
5393         struct list_head *head = &space_info->priority_tickets;
5394
5395 again:
5396         while (!list_empty(head) && num_bytes) {
5397                 ticket = list_first_entry(head, struct reserve_ticket,
5398                                           list);
5399                 if (num_bytes >= ticket->bytes) {
5400                         trace_btrfs_space_reservation(fs_info, "space_info",
5401                                                       space_info->flags,
5402                                                       ticket->bytes, 1);
5403                         list_del_init(&ticket->list);
5404                         num_bytes -= ticket->bytes;
5405                         space_info->bytes_may_use += ticket->bytes;
5406                         ticket->bytes = 0;
5407                         space_info->tickets_id++;
5408                         wake_up(&ticket->wait);
5409                 } else {
5410                         trace_btrfs_space_reservation(fs_info, "space_info",
5411                                                       space_info->flags,
5412                                                       num_bytes, 1);
5413                         space_info->bytes_may_use += num_bytes;
5414                         ticket->bytes -= num_bytes;
5415                         num_bytes = 0;
5416                 }
5417         }
5418
5419         if (num_bytes && head == &space_info->priority_tickets) {
5420                 head = &space_info->tickets;
5421                 goto again;
5422         }
5423 }
5424
5425 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5426                                     struct btrfs_block_rsv *block_rsv,
5427                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5428                                     u64 *qgroup_to_release_ret)
5429 {
5430         struct btrfs_space_info *space_info = block_rsv->space_info;
5431         u64 qgroup_to_release = 0;
5432         u64 ret;
5433
5434         spin_lock(&block_rsv->lock);
5435         if (num_bytes == (u64)-1) {
5436                 num_bytes = block_rsv->size;
5437                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5438         }
5439         block_rsv->size -= num_bytes;
5440         if (block_rsv->reserved >= block_rsv->size) {
5441                 num_bytes = block_rsv->reserved - block_rsv->size;
5442                 block_rsv->reserved = block_rsv->size;
5443                 block_rsv->full = 1;
5444         } else {
5445                 num_bytes = 0;
5446         }
5447         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5448                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5449                                     block_rsv->qgroup_rsv_size;
5450                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5451         } else {
5452                 qgroup_to_release = 0;
5453         }
5454         spin_unlock(&block_rsv->lock);
5455
5456         ret = num_bytes;
5457         if (num_bytes > 0) {
5458                 if (dest) {
5459                         spin_lock(&dest->lock);
5460                         if (!dest->full) {
5461                                 u64 bytes_to_add;
5462
5463                                 bytes_to_add = dest->size - dest->reserved;
5464                                 bytes_to_add = min(num_bytes, bytes_to_add);
5465                                 dest->reserved += bytes_to_add;
5466                                 if (dest->reserved >= dest->size)
5467                                         dest->full = 1;
5468                                 num_bytes -= bytes_to_add;
5469                         }
5470                         spin_unlock(&dest->lock);
5471                 }
5472                 if (num_bytes)
5473                         space_info_add_old_bytes(fs_info, space_info,
5474                                                  num_bytes);
5475         }
5476         if (qgroup_to_release_ret)
5477                 *qgroup_to_release_ret = qgroup_to_release;
5478         return ret;
5479 }
5480
5481 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5482                             struct btrfs_block_rsv *dst, u64 num_bytes,
5483                             int update_size)
5484 {
5485         int ret;
5486
5487         ret = block_rsv_use_bytes(src, num_bytes);
5488         if (ret)
5489                 return ret;
5490
5491         block_rsv_add_bytes(dst, num_bytes, update_size);
5492         return 0;
5493 }
5494
5495 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5496 {
5497         memset(rsv, 0, sizeof(*rsv));
5498         spin_lock_init(&rsv->lock);
5499         rsv->type = type;
5500 }
5501
5502 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5503                                    struct btrfs_block_rsv *rsv,
5504                                    unsigned short type)
5505 {
5506         btrfs_init_block_rsv(rsv, type);
5507         rsv->space_info = __find_space_info(fs_info,
5508                                             BTRFS_BLOCK_GROUP_METADATA);
5509 }
5510
5511 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5512                                               unsigned short type)
5513 {
5514         struct btrfs_block_rsv *block_rsv;
5515
5516         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5517         if (!block_rsv)
5518                 return NULL;
5519
5520         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5521         return block_rsv;
5522 }
5523
5524 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5525                           struct btrfs_block_rsv *rsv)
5526 {
5527         if (!rsv)
5528                 return;
5529         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5530         kfree(rsv);
5531 }
5532
5533 int btrfs_block_rsv_add(struct btrfs_root *root,
5534                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5535                         enum btrfs_reserve_flush_enum flush)
5536 {
5537         int ret;
5538
5539         if (num_bytes == 0)
5540                 return 0;
5541
5542         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5543         if (!ret) {
5544                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5545                 return 0;
5546         }
5547
5548         return ret;
5549 }
5550
5551 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5552 {
5553         u64 num_bytes = 0;
5554         int ret = -ENOSPC;
5555
5556         if (!block_rsv)
5557                 return 0;
5558
5559         spin_lock(&block_rsv->lock);
5560         num_bytes = div_factor(block_rsv->size, min_factor);
5561         if (block_rsv->reserved >= num_bytes)
5562                 ret = 0;
5563         spin_unlock(&block_rsv->lock);
5564
5565         return ret;
5566 }
5567
5568 int btrfs_block_rsv_refill(struct btrfs_root *root,
5569                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5570                            enum btrfs_reserve_flush_enum flush)
5571 {
5572         u64 num_bytes = 0;
5573         int ret = -ENOSPC;
5574
5575         if (!block_rsv)
5576                 return 0;
5577
5578         spin_lock(&block_rsv->lock);
5579         num_bytes = min_reserved;
5580         if (block_rsv->reserved >= num_bytes)
5581                 ret = 0;
5582         else
5583                 num_bytes -= block_rsv->reserved;
5584         spin_unlock(&block_rsv->lock);
5585
5586         if (!ret)
5587                 return 0;
5588
5589         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5590         if (!ret) {
5591                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5592                 return 0;
5593         }
5594
5595         return ret;
5596 }
5597
5598 /**
5599  * btrfs_inode_rsv_refill - refill the inode block rsv.
5600  * @inode - the inode we are refilling.
5601  * @flush - the flusing restriction.
5602  *
5603  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5604  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5605  * or return if we already have enough space.  This will also handle the resreve
5606  * tracepoint for the reserved amount.
5607  */
5608 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5609                                   enum btrfs_reserve_flush_enum flush)
5610 {
5611         struct btrfs_root *root = inode->root;
5612         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5613         u64 num_bytes = 0;
5614         u64 qgroup_num_bytes = 0;
5615         int ret = -ENOSPC;
5616
5617         spin_lock(&block_rsv->lock);
5618         if (block_rsv->reserved < block_rsv->size)
5619                 num_bytes = block_rsv->size - block_rsv->reserved;
5620         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5621                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5622                                    block_rsv->qgroup_rsv_reserved;
5623         spin_unlock(&block_rsv->lock);
5624
5625         if (num_bytes == 0)
5626                 return 0;
5627
5628         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5629         if (ret)
5630                 return ret;
5631         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5632         if (!ret) {
5633                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5634                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5635                                               btrfs_ino(inode), num_bytes, 1);
5636
5637                 /* Don't forget to increase qgroup_rsv_reserved */
5638                 spin_lock(&block_rsv->lock);
5639                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5640                 spin_unlock(&block_rsv->lock);
5641         } else
5642                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5643         return ret;
5644 }
5645
5646 /**
5647  * btrfs_inode_rsv_release - release any excessive reservation.
5648  * @inode - the inode we need to release from.
5649  * @qgroup_free - free or convert qgroup meta.
5650  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5651  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5652  *   @qgroup_free is true for error handling, and false for normal release.
5653  *
5654  * This is the same as btrfs_block_rsv_release, except that it handles the
5655  * tracepoint for the reservation.
5656  */
5657 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5658 {
5659         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5660         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5661         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5662         u64 released = 0;
5663         u64 qgroup_to_release = 0;
5664
5665         /*
5666          * Since we statically set the block_rsv->size we just want to say we
5667          * are releasing 0 bytes, and then we'll just get the reservation over
5668          * the size free'd.
5669          */
5670         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5671                                            &qgroup_to_release);
5672         if (released > 0)
5673                 trace_btrfs_space_reservation(fs_info, "delalloc",
5674                                               btrfs_ino(inode), released, 0);
5675         if (qgroup_free)
5676                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5677         else
5678                 btrfs_qgroup_convert_reserved_meta(inode->root,
5679                                                    qgroup_to_release);
5680 }
5681
5682 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5683                              struct btrfs_block_rsv *block_rsv,
5684                              u64 num_bytes)
5685 {
5686         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5687
5688         if (global_rsv == block_rsv ||
5689             block_rsv->space_info != global_rsv->space_info)
5690                 global_rsv = NULL;
5691         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5692 }
5693
5694 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5695 {
5696         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5697         struct btrfs_space_info *sinfo = block_rsv->space_info;
5698         u64 num_bytes;
5699
5700         /*
5701          * The global block rsv is based on the size of the extent tree, the
5702          * checksum tree and the root tree.  If the fs is empty we want to set
5703          * it to a minimal amount for safety.
5704          */
5705         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5706                 btrfs_root_used(&fs_info->csum_root->root_item) +
5707                 btrfs_root_used(&fs_info->tree_root->root_item);
5708         num_bytes = max_t(u64, num_bytes, SZ_16M);
5709
5710         spin_lock(&sinfo->lock);
5711         spin_lock(&block_rsv->lock);
5712
5713         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5714
5715         if (block_rsv->reserved < block_rsv->size) {
5716                 num_bytes = btrfs_space_info_used(sinfo, true);
5717                 if (sinfo->total_bytes > num_bytes) {
5718                         num_bytes = sinfo->total_bytes - num_bytes;
5719                         num_bytes = min(num_bytes,
5720                                         block_rsv->size - block_rsv->reserved);
5721                         block_rsv->reserved += num_bytes;
5722                         sinfo->bytes_may_use += num_bytes;
5723                         trace_btrfs_space_reservation(fs_info, "space_info",
5724                                                       sinfo->flags, num_bytes,
5725                                                       1);
5726                 }
5727         } else if (block_rsv->reserved > block_rsv->size) {
5728                 num_bytes = block_rsv->reserved - block_rsv->size;
5729                 sinfo->bytes_may_use -= num_bytes;
5730                 trace_btrfs_space_reservation(fs_info, "space_info",
5731                                       sinfo->flags, num_bytes, 0);
5732                 block_rsv->reserved = block_rsv->size;
5733         }
5734
5735         if (block_rsv->reserved == block_rsv->size)
5736                 block_rsv->full = 1;
5737         else
5738                 block_rsv->full = 0;
5739
5740         spin_unlock(&block_rsv->lock);
5741         spin_unlock(&sinfo->lock);
5742 }
5743
5744 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5745 {
5746         struct btrfs_space_info *space_info;
5747
5748         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5749         fs_info->chunk_block_rsv.space_info = space_info;
5750
5751         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5752         fs_info->global_block_rsv.space_info = space_info;
5753         fs_info->trans_block_rsv.space_info = space_info;
5754         fs_info->empty_block_rsv.space_info = space_info;
5755         fs_info->delayed_block_rsv.space_info = space_info;
5756
5757         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5758         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5759         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5760         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5761         if (fs_info->quota_root)
5762                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5763         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5764
5765         update_global_block_rsv(fs_info);
5766 }
5767
5768 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5769 {
5770         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5771                                 (u64)-1, NULL);
5772         WARN_ON(fs_info->trans_block_rsv.size > 0);
5773         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5774         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5775         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5776         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5777         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5778 }
5779
5780
5781 /*
5782  * To be called after all the new block groups attached to the transaction
5783  * handle have been created (btrfs_create_pending_block_groups()).
5784  */
5785 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5786 {
5787         struct btrfs_fs_info *fs_info = trans->fs_info;
5788
5789         if (!trans->chunk_bytes_reserved)
5790                 return;
5791
5792         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5793
5794         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5795                                 trans->chunk_bytes_reserved, NULL);
5796         trans->chunk_bytes_reserved = 0;
5797 }
5798
5799 /*
5800  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5801  * root: the root of the parent directory
5802  * rsv: block reservation
5803  * items: the number of items that we need do reservation
5804  * use_global_rsv: allow fallback to the global block reservation
5805  *
5806  * This function is used to reserve the space for snapshot/subvolume
5807  * creation and deletion. Those operations are different with the
5808  * common file/directory operations, they change two fs/file trees
5809  * and root tree, the number of items that the qgroup reserves is
5810  * different with the free space reservation. So we can not use
5811  * the space reservation mechanism in start_transaction().
5812  */
5813 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5814                                      struct btrfs_block_rsv *rsv, int items,
5815                                      bool use_global_rsv)
5816 {
5817         u64 qgroup_num_bytes = 0;
5818         u64 num_bytes;
5819         int ret;
5820         struct btrfs_fs_info *fs_info = root->fs_info;
5821         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5822
5823         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5824                 /* One for parent inode, two for dir entries */
5825                 qgroup_num_bytes = 3 * fs_info->nodesize;
5826                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5827                                 qgroup_num_bytes, true);
5828                 if (ret)
5829                         return ret;
5830         }
5831
5832         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5833         rsv->space_info = __find_space_info(fs_info,
5834                                             BTRFS_BLOCK_GROUP_METADATA);
5835         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5836                                   BTRFS_RESERVE_FLUSH_ALL);
5837
5838         if (ret == -ENOSPC && use_global_rsv)
5839                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5840
5841         if (ret && qgroup_num_bytes)
5842                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5843
5844         return ret;
5845 }
5846
5847 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5848                                       struct btrfs_block_rsv *rsv)
5849 {
5850         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5851 }
5852
5853 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5854                                                  struct btrfs_inode *inode)
5855 {
5856         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5857         u64 reserve_size = 0;
5858         u64 qgroup_rsv_size = 0;
5859         u64 csum_leaves;
5860         unsigned outstanding_extents;
5861
5862         lockdep_assert_held(&inode->lock);
5863         outstanding_extents = inode->outstanding_extents;
5864         if (outstanding_extents)
5865                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5866                                                 outstanding_extents + 1);
5867         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
5868                                                  inode->csum_bytes);
5869         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
5870                                                        csum_leaves);
5871         /*
5872          * For qgroup rsv, the calculation is very simple:
5873          * account one nodesize for each outstanding extent
5874          *
5875          * This is overestimating in most cases.
5876          */
5877         qgroup_rsv_size = (u64)outstanding_extents * fs_info->nodesize;
5878
5879         spin_lock(&block_rsv->lock);
5880         block_rsv->size = reserve_size;
5881         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
5882         spin_unlock(&block_rsv->lock);
5883 }
5884
5885 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5886 {
5887         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5888         unsigned nr_extents;
5889         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5890         int ret = 0;
5891         bool delalloc_lock = true;
5892
5893         /* If we are a free space inode we need to not flush since we will be in
5894          * the middle of a transaction commit.  We also don't need the delalloc
5895          * mutex since we won't race with anybody.  We need this mostly to make
5896          * lockdep shut its filthy mouth.
5897          *
5898          * If we have a transaction open (can happen if we call truncate_block
5899          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5900          */
5901         if (btrfs_is_free_space_inode(inode)) {
5902                 flush = BTRFS_RESERVE_NO_FLUSH;
5903                 delalloc_lock = false;
5904         } else {
5905                 if (current->journal_info)
5906                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
5907
5908                 if (btrfs_transaction_in_commit(fs_info))
5909                         schedule_timeout(1);
5910         }
5911
5912         if (delalloc_lock)
5913                 mutex_lock(&inode->delalloc_mutex);
5914
5915         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5916
5917         /* Add our new extents and calculate the new rsv size. */
5918         spin_lock(&inode->lock);
5919         nr_extents = count_max_extents(num_bytes);
5920         btrfs_mod_outstanding_extents(inode, nr_extents);
5921         inode->csum_bytes += num_bytes;
5922         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5923         spin_unlock(&inode->lock);
5924
5925         ret = btrfs_inode_rsv_refill(inode, flush);
5926         if (unlikely(ret))
5927                 goto out_fail;
5928
5929         if (delalloc_lock)
5930                 mutex_unlock(&inode->delalloc_mutex);
5931         return 0;
5932
5933 out_fail:
5934         spin_lock(&inode->lock);
5935         nr_extents = count_max_extents(num_bytes);
5936         btrfs_mod_outstanding_extents(inode, -nr_extents);
5937         inode->csum_bytes -= num_bytes;
5938         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5939         spin_unlock(&inode->lock);
5940
5941         btrfs_inode_rsv_release(inode, true);
5942         if (delalloc_lock)
5943                 mutex_unlock(&inode->delalloc_mutex);
5944         return ret;
5945 }
5946
5947 /**
5948  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5949  * @inode: the inode to release the reservation for.
5950  * @num_bytes: the number of bytes we are releasing.
5951  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
5952  *
5953  * This will release the metadata reservation for an inode.  This can be called
5954  * once we complete IO for a given set of bytes to release their metadata
5955  * reservations, or on error for the same reason.
5956  */
5957 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
5958                                      bool qgroup_free)
5959 {
5960         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5961
5962         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5963         spin_lock(&inode->lock);
5964         inode->csum_bytes -= num_bytes;
5965         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5966         spin_unlock(&inode->lock);
5967
5968         if (btrfs_is_testing(fs_info))
5969                 return;
5970
5971         btrfs_inode_rsv_release(inode, qgroup_free);
5972 }
5973
5974 /**
5975  * btrfs_delalloc_release_extents - release our outstanding_extents
5976  * @inode: the inode to balance the reservation for.
5977  * @num_bytes: the number of bytes we originally reserved with
5978  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
5979  *
5980  * When we reserve space we increase outstanding_extents for the extents we may
5981  * add.  Once we've set the range as delalloc or created our ordered extents we
5982  * have outstanding_extents to track the real usage, so we use this to free our
5983  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
5984  * with btrfs_delalloc_reserve_metadata.
5985  */
5986 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes)
5987 {
5988         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5989         unsigned num_extents;
5990
5991         spin_lock(&inode->lock);
5992         num_extents = count_max_extents(num_bytes);
5993         btrfs_mod_outstanding_extents(inode, -num_extents);
5994         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5995         spin_unlock(&inode->lock);
5996
5997         if (btrfs_is_testing(fs_info))
5998                 return;
5999
6000         btrfs_inode_rsv_release(inode, true);
6001 }
6002
6003 /**
6004  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6005  * delalloc
6006  * @inode: inode we're writing to
6007  * @start: start range we are writing to
6008  * @len: how long the range we are writing to
6009  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6010  *            current reservation.
6011  *
6012  * This will do the following things
6013  *
6014  * o reserve space in data space info for num bytes
6015  *   and reserve precious corresponding qgroup space
6016  *   (Done in check_data_free_space)
6017  *
6018  * o reserve space for metadata space, based on the number of outstanding
6019  *   extents and how much csums will be needed
6020  *   also reserve metadata space in a per root over-reserve method.
6021  * o add to the inodes->delalloc_bytes
6022  * o add it to the fs_info's delalloc inodes list.
6023  *   (Above 3 all done in delalloc_reserve_metadata)
6024  *
6025  * Return 0 for success
6026  * Return <0 for error(-ENOSPC or -EQUOT)
6027  */
6028 int btrfs_delalloc_reserve_space(struct inode *inode,
6029                         struct extent_changeset **reserved, u64 start, u64 len)
6030 {
6031         int ret;
6032
6033         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6034         if (ret < 0)
6035                 return ret;
6036         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6037         if (ret < 0)
6038                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6039         return ret;
6040 }
6041
6042 /**
6043  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6044  * @inode: inode we're releasing space for
6045  * @start: start position of the space already reserved
6046  * @len: the len of the space already reserved
6047  * @release_bytes: the len of the space we consumed or didn't use
6048  *
6049  * This function will release the metadata space that was not used and will
6050  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6051  * list if there are no delalloc bytes left.
6052  * Also it will handle the qgroup reserved space.
6053  */
6054 void btrfs_delalloc_release_space(struct inode *inode,
6055                                   struct extent_changeset *reserved,
6056                                   u64 start, u64 len, bool qgroup_free)
6057 {
6058         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6059         btrfs_free_reserved_data_space(inode, reserved, start, len);
6060 }
6061
6062 static int update_block_group(struct btrfs_trans_handle *trans,
6063                               struct btrfs_fs_info *info, u64 bytenr,
6064                               u64 num_bytes, int alloc)
6065 {
6066         struct btrfs_block_group_cache *cache = NULL;
6067         u64 total = num_bytes;
6068         u64 old_val;
6069         u64 byte_in_group;
6070         int factor;
6071
6072         /* block accounting for super block */
6073         spin_lock(&info->delalloc_root_lock);
6074         old_val = btrfs_super_bytes_used(info->super_copy);
6075         if (alloc)
6076                 old_val += num_bytes;
6077         else
6078                 old_val -= num_bytes;
6079         btrfs_set_super_bytes_used(info->super_copy, old_val);
6080         spin_unlock(&info->delalloc_root_lock);
6081
6082         while (total) {
6083                 cache = btrfs_lookup_block_group(info, bytenr);
6084                 if (!cache)
6085                         return -ENOENT;
6086                 factor = btrfs_bg_type_to_factor(cache->flags);
6087
6088                 /*
6089                  * If this block group has free space cache written out, we
6090                  * need to make sure to load it if we are removing space.  This
6091                  * is because we need the unpinning stage to actually add the
6092                  * space back to the block group, otherwise we will leak space.
6093                  */
6094                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6095                         cache_block_group(cache, 1);
6096
6097                 byte_in_group = bytenr - cache->key.objectid;
6098                 WARN_ON(byte_in_group > cache->key.offset);
6099
6100                 spin_lock(&cache->space_info->lock);
6101                 spin_lock(&cache->lock);
6102
6103                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6104                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6105                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6106
6107                 old_val = btrfs_block_group_used(&cache->item);
6108                 num_bytes = min(total, cache->key.offset - byte_in_group);
6109                 if (alloc) {
6110                         old_val += num_bytes;
6111                         btrfs_set_block_group_used(&cache->item, old_val);
6112                         cache->reserved -= num_bytes;
6113                         cache->space_info->bytes_reserved -= num_bytes;
6114                         cache->space_info->bytes_used += num_bytes;
6115                         cache->space_info->disk_used += num_bytes * factor;
6116                         spin_unlock(&cache->lock);
6117                         spin_unlock(&cache->space_info->lock);
6118                 } else {
6119                         old_val -= num_bytes;
6120                         btrfs_set_block_group_used(&cache->item, old_val);
6121                         cache->pinned += num_bytes;
6122                         cache->space_info->bytes_pinned += num_bytes;
6123                         cache->space_info->bytes_used -= num_bytes;
6124                         cache->space_info->disk_used -= num_bytes * factor;
6125                         spin_unlock(&cache->lock);
6126                         spin_unlock(&cache->space_info->lock);
6127
6128                         trace_btrfs_space_reservation(info, "pinned",
6129                                                       cache->space_info->flags,
6130                                                       num_bytes, 1);
6131                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6132                                            num_bytes,
6133                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6134                         set_extent_dirty(info->pinned_extents,
6135                                          bytenr, bytenr + num_bytes - 1,
6136                                          GFP_NOFS | __GFP_NOFAIL);
6137                 }
6138
6139                 spin_lock(&trans->transaction->dirty_bgs_lock);
6140                 if (list_empty(&cache->dirty_list)) {
6141                         list_add_tail(&cache->dirty_list,
6142                                       &trans->transaction->dirty_bgs);
6143                         trans->transaction->num_dirty_bgs++;
6144                         btrfs_get_block_group(cache);
6145                 }
6146                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6147
6148                 /*
6149                  * No longer have used bytes in this block group, queue it for
6150                  * deletion. We do this after adding the block group to the
6151                  * dirty list to avoid races between cleaner kthread and space
6152                  * cache writeout.
6153                  */
6154                 if (!alloc && old_val == 0)
6155                         btrfs_mark_bg_unused(cache);
6156
6157                 btrfs_put_block_group(cache);
6158                 total -= num_bytes;
6159                 bytenr += num_bytes;
6160         }
6161         return 0;
6162 }
6163
6164 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6165 {
6166         struct btrfs_block_group_cache *cache;
6167         u64 bytenr;
6168
6169         spin_lock(&fs_info->block_group_cache_lock);
6170         bytenr = fs_info->first_logical_byte;
6171         spin_unlock(&fs_info->block_group_cache_lock);
6172
6173         if (bytenr < (u64)-1)
6174                 return bytenr;
6175
6176         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6177         if (!cache)
6178                 return 0;
6179
6180         bytenr = cache->key.objectid;
6181         btrfs_put_block_group(cache);
6182
6183         return bytenr;
6184 }
6185
6186 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6187                            struct btrfs_block_group_cache *cache,
6188                            u64 bytenr, u64 num_bytes, int reserved)
6189 {
6190         spin_lock(&cache->space_info->lock);
6191         spin_lock(&cache->lock);
6192         cache->pinned += num_bytes;
6193         cache->space_info->bytes_pinned += num_bytes;
6194         if (reserved) {
6195                 cache->reserved -= num_bytes;
6196                 cache->space_info->bytes_reserved -= num_bytes;
6197         }
6198         spin_unlock(&cache->lock);
6199         spin_unlock(&cache->space_info->lock);
6200
6201         trace_btrfs_space_reservation(fs_info, "pinned",
6202                                       cache->space_info->flags, num_bytes, 1);
6203         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6204                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6205         set_extent_dirty(fs_info->pinned_extents, bytenr,
6206                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6207         return 0;
6208 }
6209
6210 /*
6211  * this function must be called within transaction
6212  */
6213 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6214                      u64 bytenr, u64 num_bytes, int reserved)
6215 {
6216         struct btrfs_block_group_cache *cache;
6217
6218         cache = btrfs_lookup_block_group(fs_info, bytenr);
6219         BUG_ON(!cache); /* Logic error */
6220
6221         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6222
6223         btrfs_put_block_group(cache);
6224         return 0;
6225 }
6226
6227 /*
6228  * this function must be called within transaction
6229  */
6230 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6231                                     u64 bytenr, u64 num_bytes)
6232 {
6233         struct btrfs_block_group_cache *cache;
6234         int ret;
6235
6236         cache = btrfs_lookup_block_group(fs_info, bytenr);
6237         if (!cache)
6238                 return -EINVAL;
6239
6240         /*
6241          * pull in the free space cache (if any) so that our pin
6242          * removes the free space from the cache.  We have load_only set
6243          * to one because the slow code to read in the free extents does check
6244          * the pinned extents.
6245          */
6246         cache_block_group(cache, 1);
6247
6248         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6249
6250         /* remove us from the free space cache (if we're there at all) */
6251         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6252         btrfs_put_block_group(cache);
6253         return ret;
6254 }
6255
6256 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6257                                    u64 start, u64 num_bytes)
6258 {
6259         int ret;
6260         struct btrfs_block_group_cache *block_group;
6261         struct btrfs_caching_control *caching_ctl;
6262
6263         block_group = btrfs_lookup_block_group(fs_info, start);
6264         if (!block_group)
6265                 return -EINVAL;
6266
6267         cache_block_group(block_group, 0);
6268         caching_ctl = get_caching_control(block_group);
6269
6270         if (!caching_ctl) {
6271                 /* Logic error */
6272                 BUG_ON(!block_group_cache_done(block_group));
6273                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6274         } else {
6275                 mutex_lock(&caching_ctl->mutex);
6276
6277                 if (start >= caching_ctl->progress) {
6278                         ret = add_excluded_extent(fs_info, start, num_bytes);
6279                 } else if (start + num_bytes <= caching_ctl->progress) {
6280                         ret = btrfs_remove_free_space(block_group,
6281                                                       start, num_bytes);
6282                 } else {
6283                         num_bytes = caching_ctl->progress - start;
6284                         ret = btrfs_remove_free_space(block_group,
6285                                                       start, num_bytes);
6286                         if (ret)
6287                                 goto out_lock;
6288
6289                         num_bytes = (start + num_bytes) -
6290                                 caching_ctl->progress;
6291                         start = caching_ctl->progress;
6292                         ret = add_excluded_extent(fs_info, start, num_bytes);
6293                 }
6294 out_lock:
6295                 mutex_unlock(&caching_ctl->mutex);
6296                 put_caching_control(caching_ctl);
6297         }
6298         btrfs_put_block_group(block_group);
6299         return ret;
6300 }
6301
6302 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6303                                  struct extent_buffer *eb)
6304 {
6305         struct btrfs_file_extent_item *item;
6306         struct btrfs_key key;
6307         int found_type;
6308         int i;
6309         int ret = 0;
6310
6311         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6312                 return 0;
6313
6314         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6315                 btrfs_item_key_to_cpu(eb, &key, i);
6316                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6317                         continue;
6318                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6319                 found_type = btrfs_file_extent_type(eb, item);
6320                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6321                         continue;
6322                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6323                         continue;
6324                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6325                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6326                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6327                 if (ret)
6328                         break;
6329         }
6330
6331         return ret;
6332 }
6333
6334 static void
6335 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6336 {
6337         atomic_inc(&bg->reservations);
6338 }
6339
6340 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6341                                         const u64 start)
6342 {
6343         struct btrfs_block_group_cache *bg;
6344
6345         bg = btrfs_lookup_block_group(fs_info, start);
6346         ASSERT(bg);
6347         if (atomic_dec_and_test(&bg->reservations))
6348                 wake_up_var(&bg->reservations);
6349         btrfs_put_block_group(bg);
6350 }
6351
6352 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6353 {
6354         struct btrfs_space_info *space_info = bg->space_info;
6355
6356         ASSERT(bg->ro);
6357
6358         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6359                 return;
6360
6361         /*
6362          * Our block group is read only but before we set it to read only,
6363          * some task might have had allocated an extent from it already, but it
6364          * has not yet created a respective ordered extent (and added it to a
6365          * root's list of ordered extents).
6366          * Therefore wait for any task currently allocating extents, since the
6367          * block group's reservations counter is incremented while a read lock
6368          * on the groups' semaphore is held and decremented after releasing
6369          * the read access on that semaphore and creating the ordered extent.
6370          */
6371         down_write(&space_info->groups_sem);
6372         up_write(&space_info->groups_sem);
6373
6374         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6375 }
6376
6377 /**
6378  * btrfs_add_reserved_bytes - update the block_group and space info counters
6379  * @cache:      The cache we are manipulating
6380  * @ram_bytes:  The number of bytes of file content, and will be same to
6381  *              @num_bytes except for the compress path.
6382  * @num_bytes:  The number of bytes in question
6383  * @delalloc:   The blocks are allocated for the delalloc write
6384  *
6385  * This is called by the allocator when it reserves space. If this is a
6386  * reservation and the block group has become read only we cannot make the
6387  * reservation and return -EAGAIN, otherwise this function always succeeds.
6388  */
6389 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6390                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6391 {
6392         struct btrfs_space_info *space_info = cache->space_info;
6393         int ret = 0;
6394
6395         spin_lock(&space_info->lock);
6396         spin_lock(&cache->lock);
6397         if (cache->ro) {
6398                 ret = -EAGAIN;
6399         } else {
6400                 cache->reserved += num_bytes;
6401                 space_info->bytes_reserved += num_bytes;
6402
6403                 trace_btrfs_space_reservation(cache->fs_info,
6404                                 "space_info", space_info->flags,
6405                                 ram_bytes, 0);
6406                 space_info->bytes_may_use -= ram_bytes;
6407                 if (delalloc)
6408                         cache->delalloc_bytes += num_bytes;
6409         }
6410         spin_unlock(&cache->lock);
6411         spin_unlock(&space_info->lock);
6412         return ret;
6413 }
6414
6415 /**
6416  * btrfs_free_reserved_bytes - update the block_group and space info counters
6417  * @cache:      The cache we are manipulating
6418  * @num_bytes:  The number of bytes in question
6419  * @delalloc:   The blocks are allocated for the delalloc write
6420  *
6421  * This is called by somebody who is freeing space that was never actually used
6422  * on disk.  For example if you reserve some space for a new leaf in transaction
6423  * A and before transaction A commits you free that leaf, you call this with
6424  * reserve set to 0 in order to clear the reservation.
6425  */
6426
6427 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6428                                      u64 num_bytes, int delalloc)
6429 {
6430         struct btrfs_space_info *space_info = cache->space_info;
6431         int ret = 0;
6432
6433         spin_lock(&space_info->lock);
6434         spin_lock(&cache->lock);
6435         if (cache->ro)
6436                 space_info->bytes_readonly += num_bytes;
6437         cache->reserved -= num_bytes;
6438         space_info->bytes_reserved -= num_bytes;
6439         space_info->max_extent_size = 0;
6440
6441         if (delalloc)
6442                 cache->delalloc_bytes -= num_bytes;
6443         spin_unlock(&cache->lock);
6444         spin_unlock(&space_info->lock);
6445         return ret;
6446 }
6447 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6448 {
6449         struct btrfs_caching_control *next;
6450         struct btrfs_caching_control *caching_ctl;
6451         struct btrfs_block_group_cache *cache;
6452
6453         down_write(&fs_info->commit_root_sem);
6454
6455         list_for_each_entry_safe(caching_ctl, next,
6456                                  &fs_info->caching_block_groups, list) {
6457                 cache = caching_ctl->block_group;
6458                 if (block_group_cache_done(cache)) {
6459                         cache->last_byte_to_unpin = (u64)-1;
6460                         list_del_init(&caching_ctl->list);
6461                         put_caching_control(caching_ctl);
6462                 } else {
6463                         cache->last_byte_to_unpin = caching_ctl->progress;
6464                 }
6465         }
6466
6467         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6468                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6469         else
6470                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6471
6472         up_write(&fs_info->commit_root_sem);
6473
6474         update_global_block_rsv(fs_info);
6475 }
6476
6477 /*
6478  * Returns the free cluster for the given space info and sets empty_cluster to
6479  * what it should be based on the mount options.
6480  */
6481 static struct btrfs_free_cluster *
6482 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6483                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6484 {
6485         struct btrfs_free_cluster *ret = NULL;
6486
6487         *empty_cluster = 0;
6488         if (btrfs_mixed_space_info(space_info))
6489                 return ret;
6490
6491         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6492                 ret = &fs_info->meta_alloc_cluster;
6493                 if (btrfs_test_opt(fs_info, SSD))
6494                         *empty_cluster = SZ_2M;
6495                 else
6496                         *empty_cluster = SZ_64K;
6497         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6498                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6499                 *empty_cluster = SZ_2M;
6500                 ret = &fs_info->data_alloc_cluster;
6501         }
6502
6503         return ret;
6504 }
6505
6506 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6507                               u64 start, u64 end,
6508                               const bool return_free_space)
6509 {
6510         struct btrfs_block_group_cache *cache = NULL;
6511         struct btrfs_space_info *space_info;
6512         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6513         struct btrfs_free_cluster *cluster = NULL;
6514         u64 len;
6515         u64 total_unpinned = 0;
6516         u64 empty_cluster = 0;
6517         bool readonly;
6518
6519         while (start <= end) {
6520                 readonly = false;
6521                 if (!cache ||
6522                     start >= cache->key.objectid + cache->key.offset) {
6523                         if (cache)
6524                                 btrfs_put_block_group(cache);
6525                         total_unpinned = 0;
6526                         cache = btrfs_lookup_block_group(fs_info, start);
6527                         BUG_ON(!cache); /* Logic error */
6528
6529                         cluster = fetch_cluster_info(fs_info,
6530                                                      cache->space_info,
6531                                                      &empty_cluster);
6532                         empty_cluster <<= 1;
6533                 }
6534
6535                 len = cache->key.objectid + cache->key.offset - start;
6536                 len = min(len, end + 1 - start);
6537
6538                 if (start < cache->last_byte_to_unpin) {
6539                         len = min(len, cache->last_byte_to_unpin - start);
6540                         if (return_free_space)
6541                                 btrfs_add_free_space(cache, start, len);
6542                 }
6543
6544                 start += len;
6545                 total_unpinned += len;
6546                 space_info = cache->space_info;
6547
6548                 /*
6549                  * If this space cluster has been marked as fragmented and we've
6550                  * unpinned enough in this block group to potentially allow a
6551                  * cluster to be created inside of it go ahead and clear the
6552                  * fragmented check.
6553                  */
6554                 if (cluster && cluster->fragmented &&
6555                     total_unpinned > empty_cluster) {
6556                         spin_lock(&cluster->lock);
6557                         cluster->fragmented = 0;
6558                         spin_unlock(&cluster->lock);
6559                 }
6560
6561                 spin_lock(&space_info->lock);
6562                 spin_lock(&cache->lock);
6563                 cache->pinned -= len;
6564                 space_info->bytes_pinned -= len;
6565
6566                 trace_btrfs_space_reservation(fs_info, "pinned",
6567                                               space_info->flags, len, 0);
6568                 space_info->max_extent_size = 0;
6569                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6570                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6571                 if (cache->ro) {
6572                         space_info->bytes_readonly += len;
6573                         readonly = true;
6574                 }
6575                 spin_unlock(&cache->lock);
6576                 if (!readonly && return_free_space &&
6577                     global_rsv->space_info == space_info) {
6578                         u64 to_add = len;
6579
6580                         spin_lock(&global_rsv->lock);
6581                         if (!global_rsv->full) {
6582                                 to_add = min(len, global_rsv->size -
6583                                              global_rsv->reserved);
6584                                 global_rsv->reserved += to_add;
6585                                 space_info->bytes_may_use += to_add;
6586                                 if (global_rsv->reserved >= global_rsv->size)
6587                                         global_rsv->full = 1;
6588                                 trace_btrfs_space_reservation(fs_info,
6589                                                               "space_info",
6590                                                               space_info->flags,
6591                                                               to_add, 1);
6592                                 len -= to_add;
6593                         }
6594                         spin_unlock(&global_rsv->lock);
6595                         /* Add to any tickets we may have */
6596                         if (len)
6597                                 space_info_add_new_bytes(fs_info, space_info,
6598                                                          len);
6599                 }
6600                 spin_unlock(&space_info->lock);
6601         }
6602
6603         if (cache)
6604                 btrfs_put_block_group(cache);
6605         return 0;
6606 }
6607
6608 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6609 {
6610         struct btrfs_fs_info *fs_info = trans->fs_info;
6611         struct btrfs_block_group_cache *block_group, *tmp;
6612         struct list_head *deleted_bgs;
6613         struct extent_io_tree *unpin;
6614         u64 start;
6615         u64 end;
6616         int ret;
6617
6618         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6619                 unpin = &fs_info->freed_extents[1];
6620         else
6621                 unpin = &fs_info->freed_extents[0];
6622
6623         while (!trans->aborted) {
6624                 struct extent_state *cached_state = NULL;
6625
6626                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6627                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6628                                             EXTENT_DIRTY, &cached_state);
6629                 if (ret) {
6630                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6631                         break;
6632                 }
6633
6634                 if (btrfs_test_opt(fs_info, DISCARD))
6635                         ret = btrfs_discard_extent(fs_info, start,
6636                                                    end + 1 - start, NULL);
6637
6638                 clear_extent_dirty(unpin, start, end, &cached_state);
6639                 unpin_extent_range(fs_info, start, end, true);
6640                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6641                 free_extent_state(cached_state);
6642                 cond_resched();
6643         }
6644
6645         /*
6646          * Transaction is finished.  We don't need the lock anymore.  We
6647          * do need to clean up the block groups in case of a transaction
6648          * abort.
6649          */
6650         deleted_bgs = &trans->transaction->deleted_bgs;
6651         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6652                 u64 trimmed = 0;
6653
6654                 ret = -EROFS;
6655                 if (!trans->aborted)
6656                         ret = btrfs_discard_extent(fs_info,
6657                                                    block_group->key.objectid,
6658                                                    block_group->key.offset,
6659                                                    &trimmed);
6660
6661                 list_del_init(&block_group->bg_list);
6662                 btrfs_put_block_group_trimming(block_group);
6663                 btrfs_put_block_group(block_group);
6664
6665                 if (ret) {
6666                         const char *errstr = btrfs_decode_error(ret);
6667                         btrfs_warn(fs_info,
6668                            "discard failed while removing blockgroup: errno=%d %s",
6669                                    ret, errstr);
6670                 }
6671         }
6672
6673         return 0;
6674 }
6675
6676 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6677                                struct btrfs_delayed_ref_node *node, u64 parent,
6678                                u64 root_objectid, u64 owner_objectid,
6679                                u64 owner_offset, int refs_to_drop,
6680                                struct btrfs_delayed_extent_op *extent_op)
6681 {
6682         struct btrfs_fs_info *info = trans->fs_info;
6683         struct btrfs_key key;
6684         struct btrfs_path *path;
6685         struct btrfs_root *extent_root = info->extent_root;
6686         struct extent_buffer *leaf;
6687         struct btrfs_extent_item *ei;
6688         struct btrfs_extent_inline_ref *iref;
6689         int ret;
6690         int is_data;
6691         int extent_slot = 0;
6692         int found_extent = 0;
6693         int num_to_del = 1;
6694         u32 item_size;
6695         u64 refs;
6696         u64 bytenr = node->bytenr;
6697         u64 num_bytes = node->num_bytes;
6698         int last_ref = 0;
6699         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6700
6701         path = btrfs_alloc_path();
6702         if (!path)
6703                 return -ENOMEM;
6704
6705         path->reada = READA_FORWARD;
6706         path->leave_spinning = 1;
6707
6708         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6709         BUG_ON(!is_data && refs_to_drop != 1);
6710
6711         if (is_data)
6712                 skinny_metadata = false;
6713
6714         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6715                                     parent, root_objectid, owner_objectid,
6716                                     owner_offset);
6717         if (ret == 0) {
6718                 extent_slot = path->slots[0];
6719                 while (extent_slot >= 0) {
6720                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6721                                               extent_slot);
6722                         if (key.objectid != bytenr)
6723                                 break;
6724                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6725                             key.offset == num_bytes) {
6726                                 found_extent = 1;
6727                                 break;
6728                         }
6729                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6730                             key.offset == owner_objectid) {
6731                                 found_extent = 1;
6732                                 break;
6733                         }
6734                         if (path->slots[0] - extent_slot > 5)
6735                                 break;
6736                         extent_slot--;
6737                 }
6738
6739                 if (!found_extent) {
6740                         BUG_ON(iref);
6741                         ret = remove_extent_backref(trans, path, NULL,
6742                                                     refs_to_drop,
6743                                                     is_data, &last_ref);
6744                         if (ret) {
6745                                 btrfs_abort_transaction(trans, ret);
6746                                 goto out;
6747                         }
6748                         btrfs_release_path(path);
6749                         path->leave_spinning = 1;
6750
6751                         key.objectid = bytenr;
6752                         key.type = BTRFS_EXTENT_ITEM_KEY;
6753                         key.offset = num_bytes;
6754
6755                         if (!is_data && skinny_metadata) {
6756                                 key.type = BTRFS_METADATA_ITEM_KEY;
6757                                 key.offset = owner_objectid;
6758                         }
6759
6760                         ret = btrfs_search_slot(trans, extent_root,
6761                                                 &key, path, -1, 1);
6762                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6763                                 /*
6764                                  * Couldn't find our skinny metadata item,
6765                                  * see if we have ye olde extent item.
6766                                  */
6767                                 path->slots[0]--;
6768                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6769                                                       path->slots[0]);
6770                                 if (key.objectid == bytenr &&
6771                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6772                                     key.offset == num_bytes)
6773                                         ret = 0;
6774                         }
6775
6776                         if (ret > 0 && skinny_metadata) {
6777                                 skinny_metadata = false;
6778                                 key.objectid = bytenr;
6779                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6780                                 key.offset = num_bytes;
6781                                 btrfs_release_path(path);
6782                                 ret = btrfs_search_slot(trans, extent_root,
6783                                                         &key, path, -1, 1);
6784                         }
6785
6786                         if (ret) {
6787                                 btrfs_err(info,
6788                                           "umm, got %d back from search, was looking for %llu",
6789                                           ret, bytenr);
6790                                 if (ret > 0)
6791                                         btrfs_print_leaf(path->nodes[0]);
6792                         }
6793                         if (ret < 0) {
6794                                 btrfs_abort_transaction(trans, ret);
6795                                 goto out;
6796                         }
6797                         extent_slot = path->slots[0];
6798                 }
6799         } else if (WARN_ON(ret == -ENOENT)) {
6800                 btrfs_print_leaf(path->nodes[0]);
6801                 btrfs_err(info,
6802                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6803                         bytenr, parent, root_objectid, owner_objectid,
6804                         owner_offset);
6805                 btrfs_abort_transaction(trans, ret);
6806                 goto out;
6807         } else {
6808                 btrfs_abort_transaction(trans, ret);
6809                 goto out;
6810         }
6811
6812         leaf = path->nodes[0];
6813         item_size = btrfs_item_size_nr(leaf, extent_slot);
6814         if (unlikely(item_size < sizeof(*ei))) {
6815                 ret = -EINVAL;
6816                 btrfs_print_v0_err(info);
6817                 btrfs_abort_transaction(trans, ret);
6818                 goto out;
6819         }
6820         ei = btrfs_item_ptr(leaf, extent_slot,
6821                             struct btrfs_extent_item);
6822         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6823             key.type == BTRFS_EXTENT_ITEM_KEY) {
6824                 struct btrfs_tree_block_info *bi;
6825                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6826                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6827                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6828         }
6829
6830         refs = btrfs_extent_refs(leaf, ei);
6831         if (refs < refs_to_drop) {
6832                 btrfs_err(info,
6833                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
6834                           refs_to_drop, refs, bytenr);
6835                 ret = -EINVAL;
6836                 btrfs_abort_transaction(trans, ret);
6837                 goto out;
6838         }
6839         refs -= refs_to_drop;
6840
6841         if (refs > 0) {
6842                 if (extent_op)
6843                         __run_delayed_extent_op(extent_op, leaf, ei);
6844                 /*
6845                  * In the case of inline back ref, reference count will
6846                  * be updated by remove_extent_backref
6847                  */
6848                 if (iref) {
6849                         BUG_ON(!found_extent);
6850                 } else {
6851                         btrfs_set_extent_refs(leaf, ei, refs);
6852                         btrfs_mark_buffer_dirty(leaf);
6853                 }
6854                 if (found_extent) {
6855                         ret = remove_extent_backref(trans, path, iref,
6856                                                     refs_to_drop, is_data,
6857                                                     &last_ref);
6858                         if (ret) {
6859                                 btrfs_abort_transaction(trans, ret);
6860                                 goto out;
6861                         }
6862                 }
6863         } else {
6864                 if (found_extent) {
6865                         BUG_ON(is_data && refs_to_drop !=
6866                                extent_data_ref_count(path, iref));
6867                         if (iref) {
6868                                 BUG_ON(path->slots[0] != extent_slot);
6869                         } else {
6870                                 BUG_ON(path->slots[0] != extent_slot + 1);
6871                                 path->slots[0] = extent_slot;
6872                                 num_to_del = 2;
6873                         }
6874                 }
6875
6876                 last_ref = 1;
6877                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6878                                       num_to_del);
6879                 if (ret) {
6880                         btrfs_abort_transaction(trans, ret);
6881                         goto out;
6882                 }
6883                 btrfs_release_path(path);
6884
6885                 if (is_data) {
6886                         ret = btrfs_del_csums(trans, info->csum_root, bytenr,
6887                                               num_bytes);
6888                         if (ret) {
6889                                 btrfs_abort_transaction(trans, ret);
6890                                 goto out;
6891                         }
6892                 }
6893
6894                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
6895                 if (ret) {
6896                         btrfs_abort_transaction(trans, ret);
6897                         goto out;
6898                 }
6899
6900                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
6901                 if (ret) {
6902                         btrfs_abort_transaction(trans, ret);
6903                         goto out;
6904                 }
6905         }
6906         btrfs_release_path(path);
6907
6908 out:
6909         btrfs_free_path(path);
6910         return ret;
6911 }
6912
6913 /*
6914  * when we free an block, it is possible (and likely) that we free the last
6915  * delayed ref for that extent as well.  This searches the delayed ref tree for
6916  * a given extent, and if there are no other delayed refs to be processed, it
6917  * removes it from the tree.
6918  */
6919 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6920                                       u64 bytenr)
6921 {
6922         struct btrfs_delayed_ref_head *head;
6923         struct btrfs_delayed_ref_root *delayed_refs;
6924         int ret = 0;
6925
6926         delayed_refs = &trans->transaction->delayed_refs;
6927         spin_lock(&delayed_refs->lock);
6928         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
6929         if (!head)
6930                 goto out_delayed_unlock;
6931
6932         spin_lock(&head->lock);
6933         if (!RB_EMPTY_ROOT(&head->ref_tree))
6934                 goto out;
6935
6936         if (head->extent_op) {
6937                 if (!head->must_insert_reserved)
6938                         goto out;
6939                 btrfs_free_delayed_extent_op(head->extent_op);
6940                 head->extent_op = NULL;
6941         }
6942
6943         /*
6944          * waiting for the lock here would deadlock.  If someone else has it
6945          * locked they are already in the process of dropping it anyway
6946          */
6947         if (!mutex_trylock(&head->mutex))
6948                 goto out;
6949
6950         /*
6951          * at this point we have a head with no other entries.  Go
6952          * ahead and process it.
6953          */
6954         rb_erase(&head->href_node, &delayed_refs->href_root);
6955         RB_CLEAR_NODE(&head->href_node);
6956         atomic_dec(&delayed_refs->num_entries);
6957
6958         /*
6959          * we don't take a ref on the node because we're removing it from the
6960          * tree, so we just steal the ref the tree was holding.
6961          */
6962         delayed_refs->num_heads--;
6963         if (head->processing == 0)
6964                 delayed_refs->num_heads_ready--;
6965         head->processing = 0;
6966         spin_unlock(&head->lock);
6967         spin_unlock(&delayed_refs->lock);
6968
6969         BUG_ON(head->extent_op);
6970         if (head->must_insert_reserved)
6971                 ret = 1;
6972
6973         mutex_unlock(&head->mutex);
6974         btrfs_put_delayed_ref_head(head);
6975         return ret;
6976 out:
6977         spin_unlock(&head->lock);
6978
6979 out_delayed_unlock:
6980         spin_unlock(&delayed_refs->lock);
6981         return 0;
6982 }
6983
6984 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6985                            struct btrfs_root *root,
6986                            struct extent_buffer *buf,
6987                            u64 parent, int last_ref)
6988 {
6989         struct btrfs_fs_info *fs_info = root->fs_info;
6990         int pin = 1;
6991         int ret;
6992
6993         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6994                 int old_ref_mod, new_ref_mod;
6995
6996                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
6997                                    root->root_key.objectid,
6998                                    btrfs_header_level(buf), 0,
6999                                    BTRFS_DROP_DELAYED_REF);
7000                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7001                                                  buf->len, parent,
7002                                                  root->root_key.objectid,
7003                                                  btrfs_header_level(buf),
7004                                                  BTRFS_DROP_DELAYED_REF, NULL,
7005                                                  &old_ref_mod, &new_ref_mod);
7006                 BUG_ON(ret); /* -ENOMEM */
7007                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7008         }
7009
7010         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7011                 struct btrfs_block_group_cache *cache;
7012
7013                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7014                         ret = check_ref_cleanup(trans, buf->start);
7015                         if (!ret)
7016                                 goto out;
7017                 }
7018
7019                 pin = 0;
7020                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7021
7022                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7023                         pin_down_extent(fs_info, cache, buf->start,
7024                                         buf->len, 1);
7025                         btrfs_put_block_group(cache);
7026                         goto out;
7027                 }
7028
7029                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7030
7031                 btrfs_add_free_space(cache, buf->start, buf->len);
7032                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7033                 btrfs_put_block_group(cache);
7034                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7035         }
7036 out:
7037         if (pin)
7038                 add_pinned_bytes(fs_info, buf->len, true,
7039                                  root->root_key.objectid);
7040
7041         if (last_ref) {
7042                 /*
7043                  * Deleting the buffer, clear the corrupt flag since it doesn't
7044                  * matter anymore.
7045                  */
7046                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7047         }
7048 }
7049
7050 /* Can return -ENOMEM */
7051 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7052                       struct btrfs_root *root,
7053                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7054                       u64 owner, u64 offset)
7055 {
7056         struct btrfs_fs_info *fs_info = root->fs_info;
7057         int old_ref_mod, new_ref_mod;
7058         int ret;
7059
7060         if (btrfs_is_testing(fs_info))
7061                 return 0;
7062
7063         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7064                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7065                                    root_objectid, owner, offset,
7066                                    BTRFS_DROP_DELAYED_REF);
7067
7068         /*
7069          * tree log blocks never actually go into the extent allocation
7070          * tree, just update pinning info and exit early.
7071          */
7072         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7073                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7074                 /* unlocks the pinned mutex */
7075                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7076                 old_ref_mod = new_ref_mod = 0;
7077                 ret = 0;
7078         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7079                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7080                                                  num_bytes, parent,
7081                                                  root_objectid, (int)owner,
7082                                                  BTRFS_DROP_DELAYED_REF, NULL,
7083                                                  &old_ref_mod, &new_ref_mod);
7084         } else {
7085                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7086                                                  num_bytes, parent,
7087                                                  root_objectid, owner, offset,
7088                                                  0, BTRFS_DROP_DELAYED_REF,
7089                                                  &old_ref_mod, &new_ref_mod);
7090         }
7091
7092         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7093                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7094
7095                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7096         }
7097
7098         return ret;
7099 }
7100
7101 /*
7102  * when we wait for progress in the block group caching, its because
7103  * our allocation attempt failed at least once.  So, we must sleep
7104  * and let some progress happen before we try again.
7105  *
7106  * This function will sleep at least once waiting for new free space to
7107  * show up, and then it will check the block group free space numbers
7108  * for our min num_bytes.  Another option is to have it go ahead
7109  * and look in the rbtree for a free extent of a given size, but this
7110  * is a good start.
7111  *
7112  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7113  * any of the information in this block group.
7114  */
7115 static noinline void
7116 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7117                                 u64 num_bytes)
7118 {
7119         struct btrfs_caching_control *caching_ctl;
7120
7121         caching_ctl = get_caching_control(cache);
7122         if (!caching_ctl)
7123                 return;
7124
7125         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7126                    (cache->free_space_ctl->free_space >= num_bytes));
7127
7128         put_caching_control(caching_ctl);
7129 }
7130
7131 static noinline int
7132 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7133 {
7134         struct btrfs_caching_control *caching_ctl;
7135         int ret = 0;
7136
7137         caching_ctl = get_caching_control(cache);
7138         if (!caching_ctl)
7139                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7140
7141         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7142         if (cache->cached == BTRFS_CACHE_ERROR)
7143                 ret = -EIO;
7144         put_caching_control(caching_ctl);
7145         return ret;
7146 }
7147
7148 enum btrfs_loop_type {
7149         LOOP_CACHING_NOWAIT = 0,
7150         LOOP_CACHING_WAIT = 1,
7151         LOOP_ALLOC_CHUNK = 2,
7152         LOOP_NO_EMPTY_SIZE = 3,
7153 };
7154
7155 static inline void
7156 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7157                        int delalloc)
7158 {
7159         if (delalloc)
7160                 down_read(&cache->data_rwsem);
7161 }
7162
7163 static inline void
7164 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7165                        int delalloc)
7166 {
7167         btrfs_get_block_group(cache);
7168         if (delalloc)
7169                 down_read(&cache->data_rwsem);
7170 }
7171
7172 static struct btrfs_block_group_cache *
7173 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7174                    struct btrfs_free_cluster *cluster,
7175                    int delalloc)
7176 {
7177         struct btrfs_block_group_cache *used_bg = NULL;
7178
7179         spin_lock(&cluster->refill_lock);
7180         while (1) {
7181                 used_bg = cluster->block_group;
7182                 if (!used_bg)
7183                         return NULL;
7184
7185                 if (used_bg == block_group)
7186                         return used_bg;
7187
7188                 btrfs_get_block_group(used_bg);
7189
7190                 if (!delalloc)
7191                         return used_bg;
7192
7193                 if (down_read_trylock(&used_bg->data_rwsem))
7194                         return used_bg;
7195
7196                 spin_unlock(&cluster->refill_lock);
7197
7198                 /* We should only have one-level nested. */
7199                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7200
7201                 spin_lock(&cluster->refill_lock);
7202                 if (used_bg == cluster->block_group)
7203                         return used_bg;
7204
7205                 up_read(&used_bg->data_rwsem);
7206                 btrfs_put_block_group(used_bg);
7207         }
7208 }
7209
7210 static inline void
7211 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7212                          int delalloc)
7213 {
7214         if (delalloc)
7215                 up_read(&cache->data_rwsem);
7216         btrfs_put_block_group(cache);
7217 }
7218
7219 /*
7220  * walks the btree of allocated extents and find a hole of a given size.
7221  * The key ins is changed to record the hole:
7222  * ins->objectid == start position
7223  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7224  * ins->offset == the size of the hole.
7225  * Any available blocks before search_start are skipped.
7226  *
7227  * If there is no suitable free space, we will record the max size of
7228  * the free space extent currently.
7229  */
7230 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7231                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7232                                 u64 hint_byte, struct btrfs_key *ins,
7233                                 u64 flags, int delalloc)
7234 {
7235         int ret = 0;
7236         struct btrfs_root *root = fs_info->extent_root;
7237         struct btrfs_free_cluster *last_ptr = NULL;
7238         struct btrfs_block_group_cache *block_group = NULL;
7239         u64 search_start = 0;
7240         u64 max_extent_size = 0;
7241         u64 max_free_space = 0;
7242         u64 empty_cluster = 0;
7243         struct btrfs_space_info *space_info;
7244         int loop = 0;
7245         int index = btrfs_bg_flags_to_raid_index(flags);
7246         bool failed_cluster_refill = false;
7247         bool failed_alloc = false;
7248         bool use_cluster = true;
7249         bool have_caching_bg = false;
7250         bool orig_have_caching_bg = false;
7251         bool full_search = false;
7252
7253         WARN_ON(num_bytes < fs_info->sectorsize);
7254         ins->type = BTRFS_EXTENT_ITEM_KEY;
7255         ins->objectid = 0;
7256         ins->offset = 0;
7257
7258         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7259
7260         space_info = __find_space_info(fs_info, flags);
7261         if (!space_info) {
7262                 btrfs_err(fs_info, "No space info for %llu", flags);
7263                 return -ENOSPC;
7264         }
7265
7266         /*
7267          * If our free space is heavily fragmented we may not be able to make
7268          * big contiguous allocations, so instead of doing the expensive search
7269          * for free space, simply return ENOSPC with our max_extent_size so we
7270          * can go ahead and search for a more manageable chunk.
7271          *
7272          * If our max_extent_size is large enough for our allocation simply
7273          * disable clustering since we will likely not be able to find enough
7274          * space to create a cluster and induce latency trying.
7275          */
7276         if (unlikely(space_info->max_extent_size)) {
7277                 spin_lock(&space_info->lock);
7278                 if (space_info->max_extent_size &&
7279                     num_bytes > space_info->max_extent_size) {
7280                         ins->offset = space_info->max_extent_size;
7281                         spin_unlock(&space_info->lock);
7282                         return -ENOSPC;
7283                 } else if (space_info->max_extent_size) {
7284                         use_cluster = false;
7285                 }
7286                 spin_unlock(&space_info->lock);
7287         }
7288
7289         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7290         if (last_ptr) {
7291                 spin_lock(&last_ptr->lock);
7292                 if (last_ptr->block_group)
7293                         hint_byte = last_ptr->window_start;
7294                 if (last_ptr->fragmented) {
7295                         /*
7296                          * We still set window_start so we can keep track of the
7297                          * last place we found an allocation to try and save
7298                          * some time.
7299                          */
7300                         hint_byte = last_ptr->window_start;
7301                         use_cluster = false;
7302                 }
7303                 spin_unlock(&last_ptr->lock);
7304         }
7305
7306         search_start = max(search_start, first_logical_byte(fs_info, 0));
7307         search_start = max(search_start, hint_byte);
7308         if (search_start == hint_byte) {
7309                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7310                 /*
7311                  * we don't want to use the block group if it doesn't match our
7312                  * allocation bits, or if its not cached.
7313                  *
7314                  * However if we are re-searching with an ideal block group
7315                  * picked out then we don't care that the block group is cached.
7316                  */
7317                 if (block_group && block_group_bits(block_group, flags) &&
7318                     block_group->cached != BTRFS_CACHE_NO) {
7319                         down_read(&space_info->groups_sem);
7320                         if (list_empty(&block_group->list) ||
7321                             block_group->ro) {
7322                                 /*
7323                                  * someone is removing this block group,
7324                                  * we can't jump into the have_block_group
7325                                  * target because our list pointers are not
7326                                  * valid
7327                                  */
7328                                 btrfs_put_block_group(block_group);
7329                                 up_read(&space_info->groups_sem);
7330                         } else {
7331                                 index = btrfs_bg_flags_to_raid_index(
7332                                                 block_group->flags);
7333                                 btrfs_lock_block_group(block_group, delalloc);
7334                                 goto have_block_group;
7335                         }
7336                 } else if (block_group) {
7337                         btrfs_put_block_group(block_group);
7338                 }
7339         }
7340 search:
7341         have_caching_bg = false;
7342         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7343                 full_search = true;
7344         down_read(&space_info->groups_sem);
7345         list_for_each_entry(block_group, &space_info->block_groups[index],
7346                             list) {
7347                 u64 offset;
7348                 int cached;
7349
7350                 /* If the block group is read-only, we can skip it entirely. */
7351                 if (unlikely(block_group->ro))
7352                         continue;
7353
7354                 btrfs_grab_block_group(block_group, delalloc);
7355                 search_start = block_group->key.objectid;
7356
7357                 /*
7358                  * this can happen if we end up cycling through all the
7359                  * raid types, but we want to make sure we only allocate
7360                  * for the proper type.
7361                  */
7362                 if (!block_group_bits(block_group, flags)) {
7363                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7364                                 BTRFS_BLOCK_GROUP_RAID1 |
7365                                 BTRFS_BLOCK_GROUP_RAID5 |
7366                                 BTRFS_BLOCK_GROUP_RAID6 |
7367                                 BTRFS_BLOCK_GROUP_RAID10;
7368
7369                         /*
7370                          * if they asked for extra copies and this block group
7371                          * doesn't provide them, bail.  This does allow us to
7372                          * fill raid0 from raid1.
7373                          */
7374                         if ((flags & extra) && !(block_group->flags & extra))
7375                                 goto loop;
7376
7377                         /*
7378                          * This block group has different flags than we want.
7379                          * It's possible that we have MIXED_GROUP flag but no
7380                          * block group is mixed.  Just skip such block group.
7381                          */
7382                         btrfs_release_block_group(block_group, delalloc);
7383                         continue;
7384                 }
7385
7386 have_block_group:
7387                 cached = block_group_cache_done(block_group);
7388                 if (unlikely(!cached)) {
7389                         have_caching_bg = true;
7390                         ret = cache_block_group(block_group, 0);
7391                         BUG_ON(ret < 0);
7392                         ret = 0;
7393                 }
7394
7395                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7396                         goto loop;
7397
7398                 /*
7399                  * Ok we want to try and use the cluster allocator, so
7400                  * lets look there
7401                  */
7402                 if (last_ptr && use_cluster) {
7403                         struct btrfs_block_group_cache *used_block_group;
7404                         unsigned long aligned_cluster;
7405                         /*
7406                          * the refill lock keeps out other
7407                          * people trying to start a new cluster
7408                          */
7409                         used_block_group = btrfs_lock_cluster(block_group,
7410                                                               last_ptr,
7411                                                               delalloc);
7412                         if (!used_block_group)
7413                                 goto refill_cluster;
7414
7415                         if (used_block_group != block_group &&
7416                             (used_block_group->ro ||
7417                              !block_group_bits(used_block_group, flags)))
7418                                 goto release_cluster;
7419
7420                         offset = btrfs_alloc_from_cluster(used_block_group,
7421                                                 last_ptr,
7422                                                 num_bytes,
7423                                                 used_block_group->key.objectid,
7424                                                 &max_extent_size);
7425                         if (offset) {
7426                                 /* we have a block, we're done */
7427                                 spin_unlock(&last_ptr->refill_lock);
7428                                 trace_btrfs_reserve_extent_cluster(
7429                                                 used_block_group,
7430                                                 search_start, num_bytes);
7431                                 if (used_block_group != block_group) {
7432                                         btrfs_release_block_group(block_group,
7433                                                                   delalloc);
7434                                         block_group = used_block_group;
7435                                 }
7436                                 goto checks;
7437                         }
7438
7439                         WARN_ON(last_ptr->block_group != used_block_group);
7440 release_cluster:
7441                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7442                          * set up a new clusters, so lets just skip it
7443                          * and let the allocator find whatever block
7444                          * it can find.  If we reach this point, we
7445                          * will have tried the cluster allocator
7446                          * plenty of times and not have found
7447                          * anything, so we are likely way too
7448                          * fragmented for the clustering stuff to find
7449                          * anything.
7450                          *
7451                          * However, if the cluster is taken from the
7452                          * current block group, release the cluster
7453                          * first, so that we stand a better chance of
7454                          * succeeding in the unclustered
7455                          * allocation.  */
7456                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7457                             used_block_group != block_group) {
7458                                 spin_unlock(&last_ptr->refill_lock);
7459                                 btrfs_release_block_group(used_block_group,
7460                                                           delalloc);
7461                                 goto unclustered_alloc;
7462                         }
7463
7464                         /*
7465                          * this cluster didn't work out, free it and
7466                          * start over
7467                          */
7468                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7469
7470                         if (used_block_group != block_group)
7471                                 btrfs_release_block_group(used_block_group,
7472                                                           delalloc);
7473 refill_cluster:
7474                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7475                                 spin_unlock(&last_ptr->refill_lock);
7476                                 goto unclustered_alloc;
7477                         }
7478
7479                         aligned_cluster = max_t(unsigned long,
7480                                                 empty_cluster + empty_size,
7481                                               block_group->full_stripe_len);
7482
7483                         /* allocate a cluster in this block group */
7484                         ret = btrfs_find_space_cluster(fs_info, block_group,
7485                                                        last_ptr, search_start,
7486                                                        num_bytes,
7487                                                        aligned_cluster);
7488                         if (ret == 0) {
7489                                 /*
7490                                  * now pull our allocation out of this
7491                                  * cluster
7492                                  */
7493                                 offset = btrfs_alloc_from_cluster(block_group,
7494                                                         last_ptr,
7495                                                         num_bytes,
7496                                                         search_start,
7497                                                         &max_extent_size);
7498                                 if (offset) {
7499                                         /* we found one, proceed */
7500                                         spin_unlock(&last_ptr->refill_lock);
7501                                         trace_btrfs_reserve_extent_cluster(
7502                                                 block_group, search_start,
7503                                                 num_bytes);
7504                                         goto checks;
7505                                 }
7506                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7507                                    && !failed_cluster_refill) {
7508                                 spin_unlock(&last_ptr->refill_lock);
7509
7510                                 failed_cluster_refill = true;
7511                                 wait_block_group_cache_progress(block_group,
7512                                        num_bytes + empty_cluster + empty_size);
7513                                 goto have_block_group;
7514                         }
7515
7516                         /*
7517                          * at this point we either didn't find a cluster
7518                          * or we weren't able to allocate a block from our
7519                          * cluster.  Free the cluster we've been trying
7520                          * to use, and go to the next block group
7521                          */
7522                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7523                         spin_unlock(&last_ptr->refill_lock);
7524                         goto loop;
7525                 }
7526
7527 unclustered_alloc:
7528                 /*
7529                  * We are doing an unclustered alloc, set the fragmented flag so
7530                  * we don't bother trying to setup a cluster again until we get
7531                  * more space.
7532                  */
7533                 if (unlikely(last_ptr)) {
7534                         spin_lock(&last_ptr->lock);
7535                         last_ptr->fragmented = 1;
7536                         spin_unlock(&last_ptr->lock);
7537                 }
7538                 if (cached) {
7539                         struct btrfs_free_space_ctl *ctl =
7540                                 block_group->free_space_ctl;
7541
7542                         spin_lock(&ctl->tree_lock);
7543                         if (ctl->free_space <
7544                             num_bytes + empty_cluster + empty_size) {
7545                                 max_free_space = max(max_free_space,
7546                                                      ctl->free_space);
7547                                 spin_unlock(&ctl->tree_lock);
7548                                 goto loop;
7549                         }
7550                         spin_unlock(&ctl->tree_lock);
7551                 }
7552
7553                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7554                                                     num_bytes, empty_size,
7555                                                     &max_extent_size);
7556                 /*
7557                  * If we didn't find a chunk, and we haven't failed on this
7558                  * block group before, and this block group is in the middle of
7559                  * caching and we are ok with waiting, then go ahead and wait
7560                  * for progress to be made, and set failed_alloc to true.
7561                  *
7562                  * If failed_alloc is true then we've already waited on this
7563                  * block group once and should move on to the next block group.
7564                  */
7565                 if (!offset && !failed_alloc && !cached &&
7566                     loop > LOOP_CACHING_NOWAIT) {
7567                         wait_block_group_cache_progress(block_group,
7568                                                 num_bytes + empty_size);
7569                         failed_alloc = true;
7570                         goto have_block_group;
7571                 } else if (!offset) {
7572                         goto loop;
7573                 }
7574 checks:
7575                 search_start = round_up(offset, fs_info->stripesize);
7576
7577                 /* move on to the next group */
7578                 if (search_start + num_bytes >
7579                     block_group->key.objectid + block_group->key.offset) {
7580                         btrfs_add_free_space(block_group, offset, num_bytes);
7581                         goto loop;
7582                 }
7583
7584                 if (offset < search_start)
7585                         btrfs_add_free_space(block_group, offset,
7586                                              search_start - offset);
7587
7588                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7589                                 num_bytes, delalloc);
7590                 if (ret == -EAGAIN) {
7591                         btrfs_add_free_space(block_group, offset, num_bytes);
7592                         goto loop;
7593                 }
7594                 btrfs_inc_block_group_reservations(block_group);
7595
7596                 /* we are all good, lets return */
7597                 ins->objectid = search_start;
7598                 ins->offset = num_bytes;
7599
7600                 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7601                 btrfs_release_block_group(block_group, delalloc);
7602                 break;
7603 loop:
7604                 failed_cluster_refill = false;
7605                 failed_alloc = false;
7606                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7607                        index);
7608                 btrfs_release_block_group(block_group, delalloc);
7609                 cond_resched();
7610         }
7611         up_read(&space_info->groups_sem);
7612
7613         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7614                 && !orig_have_caching_bg)
7615                 orig_have_caching_bg = true;
7616
7617         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7618                 goto search;
7619
7620         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7621                 goto search;
7622
7623         /*
7624          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7625          *                      caching kthreads as we move along
7626          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7627          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7628          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7629          *                      again
7630          */
7631         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7632                 index = 0;
7633                 if (loop == LOOP_CACHING_NOWAIT) {
7634                         /*
7635                          * We want to skip the LOOP_CACHING_WAIT step if we
7636                          * don't have any uncached bgs and we've already done a
7637                          * full search through.
7638                          */
7639                         if (orig_have_caching_bg || !full_search)
7640                                 loop = LOOP_CACHING_WAIT;
7641                         else
7642                                 loop = LOOP_ALLOC_CHUNK;
7643                 } else {
7644                         loop++;
7645                 }
7646
7647                 if (loop == LOOP_ALLOC_CHUNK) {
7648                         struct btrfs_trans_handle *trans;
7649                         int exist = 0;
7650
7651                         trans = current->journal_info;
7652                         if (trans)
7653                                 exist = 1;
7654                         else
7655                                 trans = btrfs_join_transaction(root);
7656
7657                         if (IS_ERR(trans)) {
7658                                 ret = PTR_ERR(trans);
7659                                 goto out;
7660                         }
7661
7662                         ret = do_chunk_alloc(trans, flags, CHUNK_ALLOC_FORCE);
7663
7664                         /*
7665                          * If we can't allocate a new chunk we've already looped
7666                          * through at least once, move on to the NO_EMPTY_SIZE
7667                          * case.
7668                          */
7669                         if (ret == -ENOSPC)
7670                                 loop = LOOP_NO_EMPTY_SIZE;
7671
7672                         /*
7673                          * Do not bail out on ENOSPC since we
7674                          * can do more things.
7675                          */
7676                         if (ret < 0 && ret != -ENOSPC)
7677                                 btrfs_abort_transaction(trans, ret);
7678                         else
7679                                 ret = 0;
7680                         if (!exist)
7681                                 btrfs_end_transaction(trans);
7682                         if (ret)
7683                                 goto out;
7684                 }
7685
7686                 if (loop == LOOP_NO_EMPTY_SIZE) {
7687                         /*
7688                          * Don't loop again if we already have no empty_size and
7689                          * no empty_cluster.
7690                          */
7691                         if (empty_size == 0 &&
7692                             empty_cluster == 0) {
7693                                 ret = -ENOSPC;
7694                                 goto out;
7695                         }
7696                         empty_size = 0;
7697                         empty_cluster = 0;
7698                 }
7699
7700                 goto search;
7701         } else if (!ins->objectid) {
7702                 ret = -ENOSPC;
7703         } else if (ins->objectid) {
7704                 if (!use_cluster && last_ptr) {
7705                         spin_lock(&last_ptr->lock);
7706                         last_ptr->window_start = ins->objectid;
7707                         spin_unlock(&last_ptr->lock);
7708                 }
7709                 ret = 0;
7710         }
7711 out:
7712         if (ret == -ENOSPC) {
7713                 if (!max_extent_size)
7714                         max_extent_size = max_free_space;
7715                 spin_lock(&space_info->lock);
7716                 space_info->max_extent_size = max_extent_size;
7717                 spin_unlock(&space_info->lock);
7718                 ins->offset = max_extent_size;
7719         }
7720         return ret;
7721 }
7722
7723 static void dump_space_info(struct btrfs_fs_info *fs_info,
7724                             struct btrfs_space_info *info, u64 bytes,
7725                             int dump_block_groups)
7726 {
7727         struct btrfs_block_group_cache *cache;
7728         int index = 0;
7729
7730         spin_lock(&info->lock);
7731         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7732                    info->flags,
7733                    info->total_bytes - btrfs_space_info_used(info, true),
7734                    info->full ? "" : "not ");
7735         btrfs_info(fs_info,
7736                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7737                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7738                 info->bytes_reserved, info->bytes_may_use,
7739                 info->bytes_readonly);
7740         spin_unlock(&info->lock);
7741
7742         if (!dump_block_groups)
7743                 return;
7744
7745         down_read(&info->groups_sem);
7746 again:
7747         list_for_each_entry(cache, &info->block_groups[index], list) {
7748                 spin_lock(&cache->lock);
7749                 btrfs_info(fs_info,
7750                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7751                         cache->key.objectid, cache->key.offset,
7752                         btrfs_block_group_used(&cache->item), cache->pinned,
7753                         cache->reserved, cache->ro ? "[readonly]" : "");
7754                 btrfs_dump_free_space(cache, bytes);
7755                 spin_unlock(&cache->lock);
7756         }
7757         if (++index < BTRFS_NR_RAID_TYPES)
7758                 goto again;
7759         up_read(&info->groups_sem);
7760 }
7761
7762 /*
7763  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7764  *                        hole that is at least as big as @num_bytes.
7765  *
7766  * @root           -    The root that will contain this extent
7767  *
7768  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7769  *                      is used for accounting purposes. This value differs
7770  *                      from @num_bytes only in the case of compressed extents.
7771  *
7772  * @num_bytes      -    Number of bytes to allocate on-disk.
7773  *
7774  * @min_alloc_size -    Indicates the minimum amount of space that the
7775  *                      allocator should try to satisfy. In some cases
7776  *                      @num_bytes may be larger than what is required and if
7777  *                      the filesystem is fragmented then allocation fails.
7778  *                      However, the presence of @min_alloc_size gives a
7779  *                      chance to try and satisfy the smaller allocation.
7780  *
7781  * @empty_size     -    A hint that you plan on doing more COW. This is the
7782  *                      size in bytes the allocator should try to find free
7783  *                      next to the block it returns.  This is just a hint and
7784  *                      may be ignored by the allocator.
7785  *
7786  * @hint_byte      -    Hint to the allocator to start searching above the byte
7787  *                      address passed. It might be ignored.
7788  *
7789  * @ins            -    This key is modified to record the found hole. It will
7790  *                      have the following values:
7791  *                      ins->objectid == start position
7792  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7793  *                      ins->offset == the size of the hole.
7794  *
7795  * @is_data        -    Boolean flag indicating whether an extent is
7796  *                      allocated for data (true) or metadata (false)
7797  *
7798  * @delalloc       -    Boolean flag indicating whether this allocation is for
7799  *                      delalloc or not. If 'true' data_rwsem of block groups
7800  *                      is going to be acquired.
7801  *
7802  *
7803  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7804  * case -ENOSPC is returned then @ins->offset will contain the size of the
7805  * largest available hole the allocator managed to find.
7806  */
7807 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7808                          u64 num_bytes, u64 min_alloc_size,
7809                          u64 empty_size, u64 hint_byte,
7810                          struct btrfs_key *ins, int is_data, int delalloc)
7811 {
7812         struct btrfs_fs_info *fs_info = root->fs_info;
7813         bool final_tried = num_bytes == min_alloc_size;
7814         u64 flags;
7815         int ret;
7816
7817         flags = get_alloc_profile_by_root(root, is_data);
7818 again:
7819         WARN_ON(num_bytes < fs_info->sectorsize);
7820         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7821                                hint_byte, ins, flags, delalloc);
7822         if (!ret && !is_data) {
7823                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7824         } else if (ret == -ENOSPC) {
7825                 if (!final_tried && ins->offset) {
7826                         num_bytes = min(num_bytes >> 1, ins->offset);
7827                         num_bytes = round_down(num_bytes,
7828                                                fs_info->sectorsize);
7829                         num_bytes = max(num_bytes, min_alloc_size);
7830                         ram_bytes = num_bytes;
7831                         if (num_bytes == min_alloc_size)
7832                                 final_tried = true;
7833                         goto again;
7834                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7835                         struct btrfs_space_info *sinfo;
7836
7837                         sinfo = __find_space_info(fs_info, flags);
7838                         btrfs_err(fs_info,
7839                                   "allocation failed flags %llu, wanted %llu",
7840                                   flags, num_bytes);
7841                         if (sinfo)
7842                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7843                 }
7844         }
7845
7846         return ret;
7847 }
7848
7849 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7850                                         u64 start, u64 len,
7851                                         int pin, int delalloc)
7852 {
7853         struct btrfs_block_group_cache *cache;
7854         int ret = 0;
7855
7856         cache = btrfs_lookup_block_group(fs_info, start);
7857         if (!cache) {
7858                 btrfs_err(fs_info, "Unable to find block group for %llu",
7859                           start);
7860                 return -ENOSPC;
7861         }
7862
7863         if (pin)
7864                 pin_down_extent(fs_info, cache, start, len, 1);
7865         else {
7866                 if (btrfs_test_opt(fs_info, DISCARD))
7867                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
7868                 btrfs_add_free_space(cache, start, len);
7869                 btrfs_free_reserved_bytes(cache, len, delalloc);
7870                 trace_btrfs_reserved_extent_free(fs_info, start, len);
7871         }
7872
7873         btrfs_put_block_group(cache);
7874         return ret;
7875 }
7876
7877 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7878                                u64 start, u64 len, int delalloc)
7879 {
7880         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
7881 }
7882
7883 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
7884                                        u64 start, u64 len)
7885 {
7886         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
7887 }
7888
7889 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7890                                       u64 parent, u64 root_objectid,
7891                                       u64 flags, u64 owner, u64 offset,
7892                                       struct btrfs_key *ins, int ref_mod)
7893 {
7894         struct btrfs_fs_info *fs_info = trans->fs_info;
7895         int ret;
7896         struct btrfs_extent_item *extent_item;
7897         struct btrfs_extent_inline_ref *iref;
7898         struct btrfs_path *path;
7899         struct extent_buffer *leaf;
7900         int type;
7901         u32 size;
7902
7903         if (parent > 0)
7904                 type = BTRFS_SHARED_DATA_REF_KEY;
7905         else
7906                 type = BTRFS_EXTENT_DATA_REF_KEY;
7907
7908         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7909
7910         path = btrfs_alloc_path();
7911         if (!path)
7912                 return -ENOMEM;
7913
7914         path->leave_spinning = 1;
7915         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7916                                       ins, size);
7917         if (ret) {
7918                 btrfs_free_path(path);
7919                 return ret;
7920         }
7921
7922         leaf = path->nodes[0];
7923         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7924                                      struct btrfs_extent_item);
7925         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7926         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7927         btrfs_set_extent_flags(leaf, extent_item,
7928                                flags | BTRFS_EXTENT_FLAG_DATA);
7929
7930         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7931         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7932         if (parent > 0) {
7933                 struct btrfs_shared_data_ref *ref;
7934                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7935                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7936                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7937         } else {
7938                 struct btrfs_extent_data_ref *ref;
7939                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7940                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7941                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7942                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7943                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7944         }
7945
7946         btrfs_mark_buffer_dirty(path->nodes[0]);
7947         btrfs_free_path(path);
7948
7949         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
7950         if (ret)
7951                 return ret;
7952
7953         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
7954         if (ret) { /* -ENOENT, logic error */
7955                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7956                         ins->objectid, ins->offset);
7957                 BUG();
7958         }
7959         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
7960         return ret;
7961 }
7962
7963 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7964                                      struct btrfs_delayed_ref_node *node,
7965                                      struct btrfs_delayed_extent_op *extent_op)
7966 {
7967         struct btrfs_fs_info *fs_info = trans->fs_info;
7968         int ret;
7969         struct btrfs_extent_item *extent_item;
7970         struct btrfs_key extent_key;
7971         struct btrfs_tree_block_info *block_info;
7972         struct btrfs_extent_inline_ref *iref;
7973         struct btrfs_path *path;
7974         struct extent_buffer *leaf;
7975         struct btrfs_delayed_tree_ref *ref;
7976         u32 size = sizeof(*extent_item) + sizeof(*iref);
7977         u64 num_bytes;
7978         u64 flags = extent_op->flags_to_set;
7979         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
7980
7981         ref = btrfs_delayed_node_to_tree_ref(node);
7982
7983         extent_key.objectid = node->bytenr;
7984         if (skinny_metadata) {
7985                 extent_key.offset = ref->level;
7986                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
7987                 num_bytes = fs_info->nodesize;
7988         } else {
7989                 extent_key.offset = node->num_bytes;
7990                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
7991                 size += sizeof(*block_info);
7992                 num_bytes = node->num_bytes;
7993         }
7994
7995         path = btrfs_alloc_path();
7996         if (!path)
7997                 return -ENOMEM;
7998
7999         path->leave_spinning = 1;
8000         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8001                                       &extent_key, size);
8002         if (ret) {
8003                 btrfs_free_path(path);
8004                 return ret;
8005         }
8006
8007         leaf = path->nodes[0];
8008         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8009                                      struct btrfs_extent_item);
8010         btrfs_set_extent_refs(leaf, extent_item, 1);
8011         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8012         btrfs_set_extent_flags(leaf, extent_item,
8013                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8014
8015         if (skinny_metadata) {
8016                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8017         } else {
8018                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8019                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8020                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8021                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8022         }
8023
8024         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8025                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8026                 btrfs_set_extent_inline_ref_type(leaf, iref,
8027                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8028                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8029         } else {
8030                 btrfs_set_extent_inline_ref_type(leaf, iref,
8031                                                  BTRFS_TREE_BLOCK_REF_KEY);
8032                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8033         }
8034
8035         btrfs_mark_buffer_dirty(leaf);
8036         btrfs_free_path(path);
8037
8038         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8039                                           num_bytes);
8040         if (ret)
8041                 return ret;
8042
8043         ret = update_block_group(trans, fs_info, extent_key.objectid,
8044                                  fs_info->nodesize, 1);
8045         if (ret) { /* -ENOENT, logic error */
8046                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8047                         extent_key.objectid, extent_key.offset);
8048                 BUG();
8049         }
8050
8051         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8052                                           fs_info->nodesize);
8053         return ret;
8054 }
8055
8056 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8057                                      struct btrfs_root *root, u64 owner,
8058                                      u64 offset, u64 ram_bytes,
8059                                      struct btrfs_key *ins)
8060 {
8061         int ret;
8062
8063         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8064
8065         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8066                            root->root_key.objectid, owner, offset,
8067                            BTRFS_ADD_DELAYED_EXTENT);
8068
8069         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8070                                          ins->offset, 0,
8071                                          root->root_key.objectid, owner,
8072                                          offset, ram_bytes,
8073                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8074         return ret;
8075 }
8076
8077 /*
8078  * this is used by the tree logging recovery code.  It records that
8079  * an extent has been allocated and makes sure to clear the free
8080  * space cache bits as well
8081  */
8082 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8083                                    u64 root_objectid, u64 owner, u64 offset,
8084                                    struct btrfs_key *ins)
8085 {
8086         struct btrfs_fs_info *fs_info = trans->fs_info;
8087         int ret;
8088         struct btrfs_block_group_cache *block_group;
8089         struct btrfs_space_info *space_info;
8090
8091         /*
8092          * Mixed block groups will exclude before processing the log so we only
8093          * need to do the exclude dance if this fs isn't mixed.
8094          */
8095         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8096                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8097                                               ins->offset);
8098                 if (ret)
8099                         return ret;
8100         }
8101
8102         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8103         if (!block_group)
8104                 return -EINVAL;
8105
8106         space_info = block_group->space_info;
8107         spin_lock(&space_info->lock);
8108         spin_lock(&block_group->lock);
8109         space_info->bytes_reserved += ins->offset;
8110         block_group->reserved += ins->offset;
8111         spin_unlock(&block_group->lock);
8112         spin_unlock(&space_info->lock);
8113
8114         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8115                                          offset, ins, 1);
8116         btrfs_put_block_group(block_group);
8117         return ret;
8118 }
8119
8120 static struct extent_buffer *
8121 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8122                       u64 bytenr, int level, u64 owner)
8123 {
8124         struct btrfs_fs_info *fs_info = root->fs_info;
8125         struct extent_buffer *buf;
8126
8127         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8128         if (IS_ERR(buf))
8129                 return buf;
8130
8131         /*
8132          * Extra safety check in case the extent tree is corrupted and extent
8133          * allocator chooses to use a tree block which is already used and
8134          * locked.
8135          */
8136         if (buf->lock_owner == current->pid) {
8137                 btrfs_err_rl(fs_info,
8138 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8139                         buf->start, btrfs_header_owner(buf), current->pid);
8140                 free_extent_buffer(buf);
8141                 return ERR_PTR(-EUCLEAN);
8142         }
8143
8144         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8145         btrfs_tree_lock(buf);
8146         clean_tree_block(fs_info, buf);
8147         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8148
8149         btrfs_set_lock_blocking(buf);
8150         set_extent_buffer_uptodate(buf);
8151
8152         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8153         btrfs_set_header_level(buf, level);
8154         btrfs_set_header_bytenr(buf, buf->start);
8155         btrfs_set_header_generation(buf, trans->transid);
8156         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8157         btrfs_set_header_owner(buf, owner);
8158         write_extent_buffer_fsid(buf, fs_info->fsid);
8159         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8160         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8161                 buf->log_index = root->log_transid % 2;
8162                 /*
8163                  * we allow two log transactions at a time, use different
8164                  * EXENT bit to differentiate dirty pages.
8165                  */
8166                 if (buf->log_index == 0)
8167                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8168                                         buf->start + buf->len - 1, GFP_NOFS);
8169                 else
8170                         set_extent_new(&root->dirty_log_pages, buf->start,
8171                                         buf->start + buf->len - 1);
8172         } else {
8173                 buf->log_index = -1;
8174                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8175                          buf->start + buf->len - 1, GFP_NOFS);
8176         }
8177         trans->dirty = true;
8178         /* this returns a buffer locked for blocking */
8179         return buf;
8180 }
8181
8182 static struct btrfs_block_rsv *
8183 use_block_rsv(struct btrfs_trans_handle *trans,
8184               struct btrfs_root *root, u32 blocksize)
8185 {
8186         struct btrfs_fs_info *fs_info = root->fs_info;
8187         struct btrfs_block_rsv *block_rsv;
8188         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8189         int ret;
8190         bool global_updated = false;
8191
8192         block_rsv = get_block_rsv(trans, root);
8193
8194         if (unlikely(block_rsv->size == 0))
8195                 goto try_reserve;
8196 again:
8197         ret = block_rsv_use_bytes(block_rsv, blocksize);
8198         if (!ret)
8199                 return block_rsv;
8200
8201         if (block_rsv->failfast)
8202                 return ERR_PTR(ret);
8203
8204         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8205                 global_updated = true;
8206                 update_global_block_rsv(fs_info);
8207                 goto again;
8208         }
8209
8210         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8211                 static DEFINE_RATELIMIT_STATE(_rs,
8212                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8213                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8214                 if (__ratelimit(&_rs))
8215                         WARN(1, KERN_DEBUG
8216                                 "BTRFS: block rsv returned %d\n", ret);
8217         }
8218 try_reserve:
8219         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8220                                      BTRFS_RESERVE_NO_FLUSH);
8221         if (!ret)
8222                 return block_rsv;
8223         /*
8224          * If we couldn't reserve metadata bytes try and use some from
8225          * the global reserve if its space type is the same as the global
8226          * reservation.
8227          */
8228         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8229             block_rsv->space_info == global_rsv->space_info) {
8230                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8231                 if (!ret)
8232                         return global_rsv;
8233         }
8234         return ERR_PTR(ret);
8235 }
8236
8237 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8238                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8239 {
8240         block_rsv_add_bytes(block_rsv, blocksize, 0);
8241         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8242 }
8243
8244 /*
8245  * finds a free extent and does all the dirty work required for allocation
8246  * returns the tree buffer or an ERR_PTR on error.
8247  */
8248 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8249                                              struct btrfs_root *root,
8250                                              u64 parent, u64 root_objectid,
8251                                              const struct btrfs_disk_key *key,
8252                                              int level, u64 hint,
8253                                              u64 empty_size)
8254 {
8255         struct btrfs_fs_info *fs_info = root->fs_info;
8256         struct btrfs_key ins;
8257         struct btrfs_block_rsv *block_rsv;
8258         struct extent_buffer *buf;
8259         struct btrfs_delayed_extent_op *extent_op;
8260         u64 flags = 0;
8261         int ret;
8262         u32 blocksize = fs_info->nodesize;
8263         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8264
8265 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8266         if (btrfs_is_testing(fs_info)) {
8267                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8268                                             level, root_objectid);
8269                 if (!IS_ERR(buf))
8270                         root->alloc_bytenr += blocksize;
8271                 return buf;
8272         }
8273 #endif
8274
8275         block_rsv = use_block_rsv(trans, root, blocksize);
8276         if (IS_ERR(block_rsv))
8277                 return ERR_CAST(block_rsv);
8278
8279         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8280                                    empty_size, hint, &ins, 0, 0);
8281         if (ret)
8282                 goto out_unuse;
8283
8284         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8285                                     root_objectid);
8286         if (IS_ERR(buf)) {
8287                 ret = PTR_ERR(buf);
8288                 goto out_free_reserved;
8289         }
8290
8291         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8292                 if (parent == 0)
8293                         parent = ins.objectid;
8294                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8295         } else
8296                 BUG_ON(parent > 0);
8297
8298         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8299                 extent_op = btrfs_alloc_delayed_extent_op();
8300                 if (!extent_op) {
8301                         ret = -ENOMEM;
8302                         goto out_free_buf;
8303                 }
8304                 if (key)
8305                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8306                 else
8307                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8308                 extent_op->flags_to_set = flags;
8309                 extent_op->update_key = skinny_metadata ? false : true;
8310                 extent_op->update_flags = true;
8311                 extent_op->is_data = false;
8312                 extent_op->level = level;
8313
8314                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8315                                    root_objectid, level, 0,
8316                                    BTRFS_ADD_DELAYED_EXTENT);
8317                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8318                                                  ins.offset, parent,
8319                                                  root_objectid, level,
8320                                                  BTRFS_ADD_DELAYED_EXTENT,
8321                                                  extent_op, NULL, NULL);
8322                 if (ret)
8323                         goto out_free_delayed;
8324         }
8325         return buf;
8326
8327 out_free_delayed:
8328         btrfs_free_delayed_extent_op(extent_op);
8329 out_free_buf:
8330         btrfs_tree_unlock(buf);
8331         free_extent_buffer(buf);
8332 out_free_reserved:
8333         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8334 out_unuse:
8335         unuse_block_rsv(fs_info, block_rsv, blocksize);
8336         return ERR_PTR(ret);
8337 }
8338
8339 struct walk_control {
8340         u64 refs[BTRFS_MAX_LEVEL];
8341         u64 flags[BTRFS_MAX_LEVEL];
8342         struct btrfs_key update_progress;
8343         int stage;
8344         int level;
8345         int shared_level;
8346         int update_ref;
8347         int keep_locks;
8348         int reada_slot;
8349         int reada_count;
8350 };
8351
8352 #define DROP_REFERENCE  1
8353 #define UPDATE_BACKREF  2
8354
8355 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8356                                      struct btrfs_root *root,
8357                                      struct walk_control *wc,
8358                                      struct btrfs_path *path)
8359 {
8360         struct btrfs_fs_info *fs_info = root->fs_info;
8361         u64 bytenr;
8362         u64 generation;
8363         u64 refs;
8364         u64 flags;
8365         u32 nritems;
8366         struct btrfs_key key;
8367         struct extent_buffer *eb;
8368         int ret;
8369         int slot;
8370         int nread = 0;
8371
8372         if (path->slots[wc->level] < wc->reada_slot) {
8373                 wc->reada_count = wc->reada_count * 2 / 3;
8374                 wc->reada_count = max(wc->reada_count, 2);
8375         } else {
8376                 wc->reada_count = wc->reada_count * 3 / 2;
8377                 wc->reada_count = min_t(int, wc->reada_count,
8378                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8379         }
8380
8381         eb = path->nodes[wc->level];
8382         nritems = btrfs_header_nritems(eb);
8383
8384         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8385                 if (nread >= wc->reada_count)
8386                         break;
8387
8388                 cond_resched();
8389                 bytenr = btrfs_node_blockptr(eb, slot);
8390                 generation = btrfs_node_ptr_generation(eb, slot);
8391
8392                 if (slot == path->slots[wc->level])
8393                         goto reada;
8394
8395                 if (wc->stage == UPDATE_BACKREF &&
8396                     generation <= root->root_key.offset)
8397                         continue;
8398
8399                 /* We don't lock the tree block, it's OK to be racy here */
8400                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8401                                                wc->level - 1, 1, &refs,
8402                                                &flags);
8403                 /* We don't care about errors in readahead. */
8404                 if (ret < 0)
8405                         continue;
8406                 BUG_ON(refs == 0);
8407
8408                 if (wc->stage == DROP_REFERENCE) {
8409                         if (refs == 1)
8410                                 goto reada;
8411
8412                         if (wc->level == 1 &&
8413                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8414                                 continue;
8415                         if (!wc->update_ref ||
8416                             generation <= root->root_key.offset)
8417                                 continue;
8418                         btrfs_node_key_to_cpu(eb, &key, slot);
8419                         ret = btrfs_comp_cpu_keys(&key,
8420                                                   &wc->update_progress);
8421                         if (ret < 0)
8422                                 continue;
8423                 } else {
8424                         if (wc->level == 1 &&
8425                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8426                                 continue;
8427                 }
8428 reada:
8429                 readahead_tree_block(fs_info, bytenr);
8430                 nread++;
8431         }
8432         wc->reada_slot = slot;
8433 }
8434
8435 /*
8436  * helper to process tree block while walking down the tree.
8437  *
8438  * when wc->stage == UPDATE_BACKREF, this function updates
8439  * back refs for pointers in the block.
8440  *
8441  * NOTE: return value 1 means we should stop walking down.
8442  */
8443 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8444                                    struct btrfs_root *root,
8445                                    struct btrfs_path *path,
8446                                    struct walk_control *wc, int lookup_info)
8447 {
8448         struct btrfs_fs_info *fs_info = root->fs_info;
8449         int level = wc->level;
8450         struct extent_buffer *eb = path->nodes[level];
8451         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8452         int ret;
8453
8454         if (wc->stage == UPDATE_BACKREF &&
8455             btrfs_header_owner(eb) != root->root_key.objectid)
8456                 return 1;
8457
8458         /*
8459          * when reference count of tree block is 1, it won't increase
8460          * again. once full backref flag is set, we never clear it.
8461          */
8462         if (lookup_info &&
8463             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8464              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8465                 BUG_ON(!path->locks[level]);
8466                 ret = btrfs_lookup_extent_info(trans, fs_info,
8467                                                eb->start, level, 1,
8468                                                &wc->refs[level],
8469                                                &wc->flags[level]);
8470                 BUG_ON(ret == -ENOMEM);
8471                 if (ret)
8472                         return ret;
8473                 BUG_ON(wc->refs[level] == 0);
8474         }
8475
8476         if (wc->stage == DROP_REFERENCE) {
8477                 if (wc->refs[level] > 1)
8478                         return 1;
8479
8480                 if (path->locks[level] && !wc->keep_locks) {
8481                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8482                         path->locks[level] = 0;
8483                 }
8484                 return 0;
8485         }
8486
8487         /* wc->stage == UPDATE_BACKREF */
8488         if (!(wc->flags[level] & flag)) {
8489                 BUG_ON(!path->locks[level]);
8490                 ret = btrfs_inc_ref(trans, root, eb, 1);
8491                 BUG_ON(ret); /* -ENOMEM */
8492                 ret = btrfs_dec_ref(trans, root, eb, 0);
8493                 BUG_ON(ret); /* -ENOMEM */
8494                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8495                                                   eb->len, flag,
8496                                                   btrfs_header_level(eb), 0);
8497                 BUG_ON(ret); /* -ENOMEM */
8498                 wc->flags[level] |= flag;
8499         }
8500
8501         /*
8502          * the block is shared by multiple trees, so it's not good to
8503          * keep the tree lock
8504          */
8505         if (path->locks[level] && level > 0) {
8506                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8507                 path->locks[level] = 0;
8508         }
8509         return 0;
8510 }
8511
8512 /*
8513  * helper to process tree block pointer.
8514  *
8515  * when wc->stage == DROP_REFERENCE, this function checks
8516  * reference count of the block pointed to. if the block
8517  * is shared and we need update back refs for the subtree
8518  * rooted at the block, this function changes wc->stage to
8519  * UPDATE_BACKREF. if the block is shared and there is no
8520  * need to update back, this function drops the reference
8521  * to the block.
8522  *
8523  * NOTE: return value 1 means we should stop walking down.
8524  */
8525 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8526                                  struct btrfs_root *root,
8527                                  struct btrfs_path *path,
8528                                  struct walk_control *wc, int *lookup_info)
8529 {
8530         struct btrfs_fs_info *fs_info = root->fs_info;
8531         u64 bytenr;
8532         u64 generation;
8533         u64 parent;
8534         u32 blocksize;
8535         struct btrfs_key key;
8536         struct btrfs_key first_key;
8537         struct extent_buffer *next;
8538         int level = wc->level;
8539         int reada = 0;
8540         int ret = 0;
8541         bool need_account = false;
8542
8543         generation = btrfs_node_ptr_generation(path->nodes[level],
8544                                                path->slots[level]);
8545         /*
8546          * if the lower level block was created before the snapshot
8547          * was created, we know there is no need to update back refs
8548          * for the subtree
8549          */
8550         if (wc->stage == UPDATE_BACKREF &&
8551             generation <= root->root_key.offset) {
8552                 *lookup_info = 1;
8553                 return 1;
8554         }
8555
8556         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8557         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8558                               path->slots[level]);
8559         blocksize = fs_info->nodesize;
8560
8561         next = find_extent_buffer(fs_info, bytenr);
8562         if (!next) {
8563                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8564                 if (IS_ERR(next))
8565                         return PTR_ERR(next);
8566
8567                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8568                                                level - 1);
8569                 reada = 1;
8570         }
8571         btrfs_tree_lock(next);
8572         btrfs_set_lock_blocking(next);
8573
8574         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8575                                        &wc->refs[level - 1],
8576                                        &wc->flags[level - 1]);
8577         if (ret < 0)
8578                 goto out_unlock;
8579
8580         if (unlikely(wc->refs[level - 1] == 0)) {
8581                 btrfs_err(fs_info, "Missing references.");
8582                 ret = -EIO;
8583                 goto out_unlock;
8584         }
8585         *lookup_info = 0;
8586
8587         if (wc->stage == DROP_REFERENCE) {
8588                 if (wc->refs[level - 1] > 1) {
8589                         need_account = true;
8590                         if (level == 1 &&
8591                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8592                                 goto skip;
8593
8594                         if (!wc->update_ref ||
8595                             generation <= root->root_key.offset)
8596                                 goto skip;
8597
8598                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8599                                               path->slots[level]);
8600                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8601                         if (ret < 0)
8602                                 goto skip;
8603
8604                         wc->stage = UPDATE_BACKREF;
8605                         wc->shared_level = level - 1;
8606                 }
8607         } else {
8608                 if (level == 1 &&
8609                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8610                         goto skip;
8611         }
8612
8613         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8614                 btrfs_tree_unlock(next);
8615                 free_extent_buffer(next);
8616                 next = NULL;
8617                 *lookup_info = 1;
8618         }
8619
8620         if (!next) {
8621                 if (reada && level == 1)
8622                         reada_walk_down(trans, root, wc, path);
8623                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8624                                        &first_key);
8625                 if (IS_ERR(next)) {
8626                         return PTR_ERR(next);
8627                 } else if (!extent_buffer_uptodate(next)) {
8628                         free_extent_buffer(next);
8629                         return -EIO;
8630                 }
8631                 btrfs_tree_lock(next);
8632                 btrfs_set_lock_blocking(next);
8633         }
8634
8635         level--;
8636         ASSERT(level == btrfs_header_level(next));
8637         if (level != btrfs_header_level(next)) {
8638                 btrfs_err(root->fs_info, "mismatched level");
8639                 ret = -EIO;
8640                 goto out_unlock;
8641         }
8642         path->nodes[level] = next;
8643         path->slots[level] = 0;
8644         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8645         wc->level = level;
8646         if (wc->level == 1)
8647                 wc->reada_slot = 0;
8648         return 0;
8649 skip:
8650         wc->refs[level - 1] = 0;
8651         wc->flags[level - 1] = 0;
8652         if (wc->stage == DROP_REFERENCE) {
8653                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8654                         parent = path->nodes[level]->start;
8655                 } else {
8656                         ASSERT(root->root_key.objectid ==
8657                                btrfs_header_owner(path->nodes[level]));
8658                         if (root->root_key.objectid !=
8659                             btrfs_header_owner(path->nodes[level])) {
8660                                 btrfs_err(root->fs_info,
8661                                                 "mismatched block owner");
8662                                 ret = -EIO;
8663                                 goto out_unlock;
8664                         }
8665                         parent = 0;
8666                 }
8667
8668                 if (need_account) {
8669                         ret = btrfs_qgroup_trace_subtree(trans, next,
8670                                                          generation, level - 1);
8671                         if (ret) {
8672                                 btrfs_err_rl(fs_info,
8673                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8674                                              ret);
8675                         }
8676                 }
8677                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8678                                         parent, root->root_key.objectid,
8679                                         level - 1, 0);
8680                 if (ret)
8681                         goto out_unlock;
8682         }
8683
8684         *lookup_info = 1;
8685         ret = 1;
8686
8687 out_unlock:
8688         btrfs_tree_unlock(next);
8689         free_extent_buffer(next);
8690
8691         return ret;
8692 }
8693
8694 /*
8695  * helper to process tree block while walking up the tree.
8696  *
8697  * when wc->stage == DROP_REFERENCE, this function drops
8698  * reference count on the block.
8699  *
8700  * when wc->stage == UPDATE_BACKREF, this function changes
8701  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8702  * to UPDATE_BACKREF previously while processing the block.
8703  *
8704  * NOTE: return value 1 means we should stop walking up.
8705  */
8706 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8707                                  struct btrfs_root *root,
8708                                  struct btrfs_path *path,
8709                                  struct walk_control *wc)
8710 {
8711         struct btrfs_fs_info *fs_info = root->fs_info;
8712         int ret;
8713         int level = wc->level;
8714         struct extent_buffer *eb = path->nodes[level];
8715         u64 parent = 0;
8716
8717         if (wc->stage == UPDATE_BACKREF) {
8718                 BUG_ON(wc->shared_level < level);
8719                 if (level < wc->shared_level)
8720                         goto out;
8721
8722                 ret = find_next_key(path, level + 1, &wc->update_progress);
8723                 if (ret > 0)
8724                         wc->update_ref = 0;
8725
8726                 wc->stage = DROP_REFERENCE;
8727                 wc->shared_level = -1;
8728                 path->slots[level] = 0;
8729
8730                 /*
8731                  * check reference count again if the block isn't locked.
8732                  * we should start walking down the tree again if reference
8733                  * count is one.
8734                  */
8735                 if (!path->locks[level]) {
8736                         BUG_ON(level == 0);
8737                         btrfs_tree_lock(eb);
8738                         btrfs_set_lock_blocking(eb);
8739                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8740
8741                         ret = btrfs_lookup_extent_info(trans, fs_info,
8742                                                        eb->start, level, 1,
8743                                                        &wc->refs[level],
8744                                                        &wc->flags[level]);
8745                         if (ret < 0) {
8746                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8747                                 path->locks[level] = 0;
8748                                 return ret;
8749                         }
8750                         BUG_ON(wc->refs[level] == 0);
8751                         if (wc->refs[level] == 1) {
8752                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8753                                 path->locks[level] = 0;
8754                                 return 1;
8755                         }
8756                 }
8757         }
8758
8759         /* wc->stage == DROP_REFERENCE */
8760         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8761
8762         if (wc->refs[level] == 1) {
8763                 if (level == 0) {
8764                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8765                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8766                         else
8767                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8768                         BUG_ON(ret); /* -ENOMEM */
8769                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
8770                         if (ret) {
8771                                 btrfs_err_rl(fs_info,
8772                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8773                                              ret);
8774                         }
8775                 }
8776                 /* make block locked assertion in clean_tree_block happy */
8777                 if (!path->locks[level] &&
8778                     btrfs_header_generation(eb) == trans->transid) {
8779                         btrfs_tree_lock(eb);
8780                         btrfs_set_lock_blocking(eb);
8781                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8782                 }
8783                 clean_tree_block(fs_info, eb);
8784         }
8785
8786         if (eb == root->node) {
8787                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8788                         parent = eb->start;
8789                 else if (root->root_key.objectid != btrfs_header_owner(eb))
8790                         goto owner_mismatch;
8791         } else {
8792                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8793                         parent = path->nodes[level + 1]->start;
8794                 else if (root->root_key.objectid !=
8795                          btrfs_header_owner(path->nodes[level + 1]))
8796                         goto owner_mismatch;
8797         }
8798
8799         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8800 out:
8801         wc->refs[level] = 0;
8802         wc->flags[level] = 0;
8803         return 0;
8804
8805 owner_mismatch:
8806         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
8807                      btrfs_header_owner(eb), root->root_key.objectid);
8808         return -EUCLEAN;
8809 }
8810
8811 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8812                                    struct btrfs_root *root,
8813                                    struct btrfs_path *path,
8814                                    struct walk_control *wc)
8815 {
8816         int level = wc->level;
8817         int lookup_info = 1;
8818         int ret;
8819
8820         while (level >= 0) {
8821                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8822                 if (ret > 0)
8823                         break;
8824
8825                 if (level == 0)
8826                         break;
8827
8828                 if (path->slots[level] >=
8829                     btrfs_header_nritems(path->nodes[level]))
8830                         break;
8831
8832                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8833                 if (ret > 0) {
8834                         path->slots[level]++;
8835                         continue;
8836                 } else if (ret < 0)
8837                         return ret;
8838                 level = wc->level;
8839         }
8840         return 0;
8841 }
8842
8843 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8844                                  struct btrfs_root *root,
8845                                  struct btrfs_path *path,
8846                                  struct walk_control *wc, int max_level)
8847 {
8848         int level = wc->level;
8849         int ret;
8850
8851         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8852         while (level < max_level && path->nodes[level]) {
8853                 wc->level = level;
8854                 if (path->slots[level] + 1 <
8855                     btrfs_header_nritems(path->nodes[level])) {
8856                         path->slots[level]++;
8857                         return 0;
8858                 } else {
8859                         ret = walk_up_proc(trans, root, path, wc);
8860                         if (ret > 0)
8861                                 return 0;
8862                         if (ret < 0)
8863                                 return ret;
8864
8865                         if (path->locks[level]) {
8866                                 btrfs_tree_unlock_rw(path->nodes[level],
8867                                                      path->locks[level]);
8868                                 path->locks[level] = 0;
8869                         }
8870                         free_extent_buffer(path->nodes[level]);
8871                         path->nodes[level] = NULL;
8872                         level++;
8873                 }
8874         }
8875         return 1;
8876 }
8877
8878 /*
8879  * drop a subvolume tree.
8880  *
8881  * this function traverses the tree freeing any blocks that only
8882  * referenced by the tree.
8883  *
8884  * when a shared tree block is found. this function decreases its
8885  * reference count by one. if update_ref is true, this function
8886  * also make sure backrefs for the shared block and all lower level
8887  * blocks are properly updated.
8888  *
8889  * If called with for_reloc == 0, may exit early with -EAGAIN
8890  */
8891 int btrfs_drop_snapshot(struct btrfs_root *root,
8892                          struct btrfs_block_rsv *block_rsv, int update_ref,
8893                          int for_reloc)
8894 {
8895         struct btrfs_fs_info *fs_info = root->fs_info;
8896         struct btrfs_path *path;
8897         struct btrfs_trans_handle *trans;
8898         struct btrfs_root *tree_root = fs_info->tree_root;
8899         struct btrfs_root_item *root_item = &root->root_item;
8900         struct walk_control *wc;
8901         struct btrfs_key key;
8902         int err = 0;
8903         int ret;
8904         int level;
8905         bool root_dropped = false;
8906
8907         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
8908
8909         path = btrfs_alloc_path();
8910         if (!path) {
8911                 err = -ENOMEM;
8912                 goto out;
8913         }
8914
8915         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8916         if (!wc) {
8917                 btrfs_free_path(path);
8918                 err = -ENOMEM;
8919                 goto out;
8920         }
8921
8922         trans = btrfs_start_transaction(tree_root, 0);
8923         if (IS_ERR(trans)) {
8924                 err = PTR_ERR(trans);
8925                 goto out_free;
8926         }
8927
8928         err = btrfs_run_delayed_items(trans);
8929         if (err)
8930                 goto out_end_trans;
8931
8932         if (block_rsv)
8933                 trans->block_rsv = block_rsv;
8934
8935         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8936                 level = btrfs_header_level(root->node);
8937                 path->nodes[level] = btrfs_lock_root_node(root);
8938                 btrfs_set_lock_blocking(path->nodes[level]);
8939                 path->slots[level] = 0;
8940                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8941                 memset(&wc->update_progress, 0,
8942                        sizeof(wc->update_progress));
8943         } else {
8944                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8945                 memcpy(&wc->update_progress, &key,
8946                        sizeof(wc->update_progress));
8947
8948                 level = root_item->drop_level;
8949                 BUG_ON(level == 0);
8950                 path->lowest_level = level;
8951                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8952                 path->lowest_level = 0;
8953                 if (ret < 0) {
8954                         err = ret;
8955                         goto out_end_trans;
8956                 }
8957                 WARN_ON(ret > 0);
8958
8959                 /*
8960                  * unlock our path, this is safe because only this
8961                  * function is allowed to delete this snapshot
8962                  */
8963                 btrfs_unlock_up_safe(path, 0);
8964
8965                 level = btrfs_header_level(root->node);
8966                 while (1) {
8967                         btrfs_tree_lock(path->nodes[level]);
8968                         btrfs_set_lock_blocking(path->nodes[level]);
8969                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8970
8971                         ret = btrfs_lookup_extent_info(trans, fs_info,
8972                                                 path->nodes[level]->start,
8973                                                 level, 1, &wc->refs[level],
8974                                                 &wc->flags[level]);
8975                         if (ret < 0) {
8976                                 err = ret;
8977                                 goto out_end_trans;
8978                         }
8979                         BUG_ON(wc->refs[level] == 0);
8980
8981                         if (level == root_item->drop_level)
8982                                 break;
8983
8984                         btrfs_tree_unlock(path->nodes[level]);
8985                         path->locks[level] = 0;
8986                         WARN_ON(wc->refs[level] != 1);
8987                         level--;
8988                 }
8989         }
8990
8991         wc->level = level;
8992         wc->shared_level = -1;
8993         wc->stage = DROP_REFERENCE;
8994         wc->update_ref = update_ref;
8995         wc->keep_locks = 0;
8996         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
8997
8998         while (1) {
8999
9000                 ret = walk_down_tree(trans, root, path, wc);
9001                 if (ret < 0) {
9002                         err = ret;
9003                         break;
9004                 }
9005
9006                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9007                 if (ret < 0) {
9008                         err = ret;
9009                         break;
9010                 }
9011
9012                 if (ret > 0) {
9013                         BUG_ON(wc->stage != DROP_REFERENCE);
9014                         break;
9015                 }
9016
9017                 if (wc->stage == DROP_REFERENCE) {
9018                         level = wc->level;
9019                         btrfs_node_key(path->nodes[level],
9020                                        &root_item->drop_progress,
9021                                        path->slots[level]);
9022                         root_item->drop_level = level;
9023                 }
9024
9025                 BUG_ON(wc->level == 0);
9026                 if (btrfs_should_end_transaction(trans) ||
9027                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9028                         ret = btrfs_update_root(trans, tree_root,
9029                                                 &root->root_key,
9030                                                 root_item);
9031                         if (ret) {
9032                                 btrfs_abort_transaction(trans, ret);
9033                                 err = ret;
9034                                 goto out_end_trans;
9035                         }
9036
9037                         btrfs_end_transaction_throttle(trans);
9038                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9039                                 btrfs_debug(fs_info,
9040                                             "drop snapshot early exit");
9041                                 err = -EAGAIN;
9042                                 goto out_free;
9043                         }
9044
9045                         trans = btrfs_start_transaction(tree_root, 0);
9046                         if (IS_ERR(trans)) {
9047                                 err = PTR_ERR(trans);
9048                                 goto out_free;
9049                         }
9050                         if (block_rsv)
9051                                 trans->block_rsv = block_rsv;
9052                 }
9053         }
9054         btrfs_release_path(path);
9055         if (err)
9056                 goto out_end_trans;
9057
9058         ret = btrfs_del_root(trans, &root->root_key);
9059         if (ret) {
9060                 btrfs_abort_transaction(trans, ret);
9061                 err = ret;
9062                 goto out_end_trans;
9063         }
9064
9065         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9066                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9067                                       NULL, NULL);
9068                 if (ret < 0) {
9069                         btrfs_abort_transaction(trans, ret);
9070                         err = ret;
9071                         goto out_end_trans;
9072                 } else if (ret > 0) {
9073                         /* if we fail to delete the orphan item this time
9074                          * around, it'll get picked up the next time.
9075                          *
9076                          * The most common failure here is just -ENOENT.
9077                          */
9078                         btrfs_del_orphan_item(trans, tree_root,
9079                                               root->root_key.objectid);
9080                 }
9081         }
9082
9083         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9084                 btrfs_add_dropped_root(trans, root);
9085         } else {
9086                 free_extent_buffer(root->node);
9087                 free_extent_buffer(root->commit_root);
9088                 btrfs_put_fs_root(root);
9089         }
9090         root_dropped = true;
9091 out_end_trans:
9092         btrfs_end_transaction_throttle(trans);
9093 out_free:
9094         kfree(wc);
9095         btrfs_free_path(path);
9096 out:
9097         /*
9098          * So if we need to stop dropping the snapshot for whatever reason we
9099          * need to make sure to add it back to the dead root list so that we
9100          * keep trying to do the work later.  This also cleans up roots if we
9101          * don't have it in the radix (like when we recover after a power fail
9102          * or unmount) so we don't leak memory.
9103          */
9104         if (!for_reloc && !root_dropped)
9105                 btrfs_add_dead_root(root);
9106         return err;
9107 }
9108
9109 /*
9110  * drop subtree rooted at tree block 'node'.
9111  *
9112  * NOTE: this function will unlock and release tree block 'node'
9113  * only used by relocation code
9114  */
9115 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9116                         struct btrfs_root *root,
9117                         struct extent_buffer *node,
9118                         struct extent_buffer *parent)
9119 {
9120         struct btrfs_fs_info *fs_info = root->fs_info;
9121         struct btrfs_path *path;
9122         struct walk_control *wc;
9123         int level;
9124         int parent_level;
9125         int ret = 0;
9126         int wret;
9127
9128         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9129
9130         path = btrfs_alloc_path();
9131         if (!path)
9132                 return -ENOMEM;
9133
9134         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9135         if (!wc) {
9136                 btrfs_free_path(path);
9137                 return -ENOMEM;
9138         }
9139
9140         btrfs_assert_tree_locked(parent);
9141         parent_level = btrfs_header_level(parent);
9142         extent_buffer_get(parent);
9143         path->nodes[parent_level] = parent;
9144         path->slots[parent_level] = btrfs_header_nritems(parent);
9145
9146         btrfs_assert_tree_locked(node);
9147         level = btrfs_header_level(node);
9148         path->nodes[level] = node;
9149         path->slots[level] = 0;
9150         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9151
9152         wc->refs[parent_level] = 1;
9153         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9154         wc->level = level;
9155         wc->shared_level = -1;
9156         wc->stage = DROP_REFERENCE;
9157         wc->update_ref = 0;
9158         wc->keep_locks = 1;
9159         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9160
9161         while (1) {
9162                 wret = walk_down_tree(trans, root, path, wc);
9163                 if (wret < 0) {
9164                         ret = wret;
9165                         break;
9166                 }
9167
9168                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9169                 if (wret < 0)
9170                         ret = wret;
9171                 if (wret != 0)
9172                         break;
9173         }
9174
9175         kfree(wc);
9176         btrfs_free_path(path);
9177         return ret;
9178 }
9179
9180 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9181 {
9182         u64 num_devices;
9183         u64 stripped;
9184
9185         /*
9186          * if restripe for this chunk_type is on pick target profile and
9187          * return, otherwise do the usual balance
9188          */
9189         stripped = get_restripe_target(fs_info, flags);
9190         if (stripped)
9191                 return extended_to_chunk(stripped);
9192
9193         num_devices = fs_info->fs_devices->rw_devices;
9194
9195         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9196                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9197                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9198
9199         if (num_devices == 1) {
9200                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9201                 stripped = flags & ~stripped;
9202
9203                 /* turn raid0 into single device chunks */
9204                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9205                         return stripped;
9206
9207                 /* turn mirroring into duplication */
9208                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9209                              BTRFS_BLOCK_GROUP_RAID10))
9210                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9211         } else {
9212                 /* they already had raid on here, just return */
9213                 if (flags & stripped)
9214                         return flags;
9215
9216                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9217                 stripped = flags & ~stripped;
9218
9219                 /* switch duplicated blocks with raid1 */
9220                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9221                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9222
9223                 /* this is drive concat, leave it alone */
9224         }
9225
9226         return flags;
9227 }
9228
9229 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9230 {
9231         struct btrfs_space_info *sinfo = cache->space_info;
9232         u64 num_bytes;
9233         u64 min_allocable_bytes;
9234         int ret = -ENOSPC;
9235
9236         /*
9237          * We need some metadata space and system metadata space for
9238          * allocating chunks in some corner cases until we force to set
9239          * it to be readonly.
9240          */
9241         if ((sinfo->flags &
9242              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9243             !force)
9244                 min_allocable_bytes = SZ_1M;
9245         else
9246                 min_allocable_bytes = 0;
9247
9248         spin_lock(&sinfo->lock);
9249         spin_lock(&cache->lock);
9250
9251         if (cache->ro) {
9252                 cache->ro++;
9253                 ret = 0;
9254                 goto out;
9255         }
9256
9257         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9258                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9259
9260         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9261             min_allocable_bytes <= sinfo->total_bytes) {
9262                 sinfo->bytes_readonly += num_bytes;
9263                 cache->ro++;
9264                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9265                 ret = 0;
9266         }
9267 out:
9268         spin_unlock(&cache->lock);
9269         spin_unlock(&sinfo->lock);
9270         return ret;
9271 }
9272
9273 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9274
9275 {
9276         struct btrfs_fs_info *fs_info = cache->fs_info;
9277         struct btrfs_trans_handle *trans;
9278         u64 alloc_flags;
9279         int ret;
9280
9281 again:
9282         trans = btrfs_join_transaction(fs_info->extent_root);
9283         if (IS_ERR(trans))
9284                 return PTR_ERR(trans);
9285
9286         /*
9287          * we're not allowed to set block groups readonly after the dirty
9288          * block groups cache has started writing.  If it already started,
9289          * back off and let this transaction commit
9290          */
9291         mutex_lock(&fs_info->ro_block_group_mutex);
9292         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9293                 u64 transid = trans->transid;
9294
9295                 mutex_unlock(&fs_info->ro_block_group_mutex);
9296                 btrfs_end_transaction(trans);
9297
9298                 ret = btrfs_wait_for_commit(fs_info, transid);
9299                 if (ret)
9300                         return ret;
9301                 goto again;
9302         }
9303
9304         /*
9305          * if we are changing raid levels, try to allocate a corresponding
9306          * block group with the new raid level.
9307          */
9308         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9309         if (alloc_flags != cache->flags) {
9310                 ret = do_chunk_alloc(trans, alloc_flags,
9311                                      CHUNK_ALLOC_FORCE);
9312                 /*
9313                  * ENOSPC is allowed here, we may have enough space
9314                  * already allocated at the new raid level to
9315                  * carry on
9316                  */
9317                 if (ret == -ENOSPC)
9318                         ret = 0;
9319                 if (ret < 0)
9320                         goto out;
9321         }
9322
9323         ret = inc_block_group_ro(cache, 0);
9324         if (!ret)
9325                 goto out;
9326         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9327         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9328         if (ret < 0)
9329                 goto out;
9330         ret = inc_block_group_ro(cache, 0);
9331 out:
9332         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9333                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9334                 mutex_lock(&fs_info->chunk_mutex);
9335                 check_system_chunk(trans, alloc_flags);
9336                 mutex_unlock(&fs_info->chunk_mutex);
9337         }
9338         mutex_unlock(&fs_info->ro_block_group_mutex);
9339
9340         btrfs_end_transaction(trans);
9341         return ret;
9342 }
9343
9344 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9345 {
9346         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9347
9348         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9349 }
9350
9351 /*
9352  * helper to account the unused space of all the readonly block group in the
9353  * space_info. takes mirrors into account.
9354  */
9355 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9356 {
9357         struct btrfs_block_group_cache *block_group;
9358         u64 free_bytes = 0;
9359         int factor;
9360
9361         /* It's df, we don't care if it's racy */
9362         if (list_empty(&sinfo->ro_bgs))
9363                 return 0;
9364
9365         spin_lock(&sinfo->lock);
9366         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9367                 spin_lock(&block_group->lock);
9368
9369                 if (!block_group->ro) {
9370                         spin_unlock(&block_group->lock);
9371                         continue;
9372                 }
9373
9374                 factor = btrfs_bg_type_to_factor(block_group->flags);
9375                 free_bytes += (block_group->key.offset -
9376                                btrfs_block_group_used(&block_group->item)) *
9377                                factor;
9378
9379                 spin_unlock(&block_group->lock);
9380         }
9381         spin_unlock(&sinfo->lock);
9382
9383         return free_bytes;
9384 }
9385
9386 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9387 {
9388         struct btrfs_space_info *sinfo = cache->space_info;
9389         u64 num_bytes;
9390
9391         BUG_ON(!cache->ro);
9392
9393         spin_lock(&sinfo->lock);
9394         spin_lock(&cache->lock);
9395         if (!--cache->ro) {
9396                 num_bytes = cache->key.offset - cache->reserved -
9397                             cache->pinned - cache->bytes_super -
9398                             btrfs_block_group_used(&cache->item);
9399                 sinfo->bytes_readonly -= num_bytes;
9400                 list_del_init(&cache->ro_list);
9401         }
9402         spin_unlock(&cache->lock);
9403         spin_unlock(&sinfo->lock);
9404 }
9405
9406 /*
9407  * checks to see if its even possible to relocate this block group.
9408  *
9409  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9410  * ok to go ahead and try.
9411  */
9412 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9413 {
9414         struct btrfs_root *root = fs_info->extent_root;
9415         struct btrfs_block_group_cache *block_group;
9416         struct btrfs_space_info *space_info;
9417         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9418         struct btrfs_device *device;
9419         struct btrfs_trans_handle *trans;
9420         u64 min_free;
9421         u64 dev_min = 1;
9422         u64 dev_nr = 0;
9423         u64 target;
9424         int debug;
9425         int index;
9426         int full = 0;
9427         int ret = 0;
9428
9429         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9430
9431         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9432
9433         /* odd, couldn't find the block group, leave it alone */
9434         if (!block_group) {
9435                 if (debug)
9436                         btrfs_warn(fs_info,
9437                                    "can't find block group for bytenr %llu",
9438                                    bytenr);
9439                 return -1;
9440         }
9441
9442         min_free = btrfs_block_group_used(&block_group->item);
9443
9444         /* no bytes used, we're good */
9445         if (!min_free)
9446                 goto out;
9447
9448         space_info = block_group->space_info;
9449         spin_lock(&space_info->lock);
9450
9451         full = space_info->full;
9452
9453         /*
9454          * if this is the last block group we have in this space, we can't
9455          * relocate it unless we're able to allocate a new chunk below.
9456          *
9457          * Otherwise, we need to make sure we have room in the space to handle
9458          * all of the extents from this block group.  If we can, we're good
9459          */
9460         if ((space_info->total_bytes != block_group->key.offset) &&
9461             (btrfs_space_info_used(space_info, false) + min_free <
9462              space_info->total_bytes)) {
9463                 spin_unlock(&space_info->lock);
9464                 goto out;
9465         }
9466         spin_unlock(&space_info->lock);
9467
9468         /*
9469          * ok we don't have enough space, but maybe we have free space on our
9470          * devices to allocate new chunks for relocation, so loop through our
9471          * alloc devices and guess if we have enough space.  if this block
9472          * group is going to be restriped, run checks against the target
9473          * profile instead of the current one.
9474          */
9475         ret = -1;
9476
9477         /*
9478          * index:
9479          *      0: raid10
9480          *      1: raid1
9481          *      2: dup
9482          *      3: raid0
9483          *      4: single
9484          */
9485         target = get_restripe_target(fs_info, block_group->flags);
9486         if (target) {
9487                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9488         } else {
9489                 /*
9490                  * this is just a balance, so if we were marked as full
9491                  * we know there is no space for a new chunk
9492                  */
9493                 if (full) {
9494                         if (debug)
9495                                 btrfs_warn(fs_info,
9496                                            "no space to alloc new chunk for block group %llu",
9497                                            block_group->key.objectid);
9498                         goto out;
9499                 }
9500
9501                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9502         }
9503
9504         if (index == BTRFS_RAID_RAID10) {
9505                 dev_min = 4;
9506                 /* Divide by 2 */
9507                 min_free >>= 1;
9508         } else if (index == BTRFS_RAID_RAID1) {
9509                 dev_min = 2;
9510         } else if (index == BTRFS_RAID_DUP) {
9511                 /* Multiply by 2 */
9512                 min_free <<= 1;
9513         } else if (index == BTRFS_RAID_RAID0) {
9514                 dev_min = fs_devices->rw_devices;
9515                 min_free = div64_u64(min_free, dev_min);
9516         }
9517
9518         /* We need to do this so that we can look at pending chunks */
9519         trans = btrfs_join_transaction(root);
9520         if (IS_ERR(trans)) {
9521                 ret = PTR_ERR(trans);
9522                 goto out;
9523         }
9524
9525         mutex_lock(&fs_info->chunk_mutex);
9526         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9527                 u64 dev_offset;
9528
9529                 /*
9530                  * check to make sure we can actually find a chunk with enough
9531                  * space to fit our block group in.
9532                  */
9533                 if (device->total_bytes > device->bytes_used + min_free &&
9534                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9535                         ret = find_free_dev_extent(trans, device, min_free,
9536                                                    &dev_offset, NULL);
9537                         if (!ret)
9538                                 dev_nr++;
9539
9540                         if (dev_nr >= dev_min)
9541                                 break;
9542
9543                         ret = -1;
9544                 }
9545         }
9546         if (debug && ret == -1)
9547                 btrfs_warn(fs_info,
9548                            "no space to allocate a new chunk for block group %llu",
9549                            block_group->key.objectid);
9550         mutex_unlock(&fs_info->chunk_mutex);
9551         btrfs_end_transaction(trans);
9552 out:
9553         btrfs_put_block_group(block_group);
9554         return ret;
9555 }
9556
9557 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9558                                   struct btrfs_path *path,
9559                                   struct btrfs_key *key)
9560 {
9561         struct btrfs_root *root = fs_info->extent_root;
9562         int ret = 0;
9563         struct btrfs_key found_key;
9564         struct extent_buffer *leaf;
9565         struct btrfs_block_group_item bg;
9566         u64 flags;
9567         int slot;
9568
9569         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9570         if (ret < 0)
9571                 goto out;
9572
9573         while (1) {
9574                 slot = path->slots[0];
9575                 leaf = path->nodes[0];
9576                 if (slot >= btrfs_header_nritems(leaf)) {
9577                         ret = btrfs_next_leaf(root, path);
9578                         if (ret == 0)
9579                                 continue;
9580                         if (ret < 0)
9581                                 goto out;
9582                         break;
9583                 }
9584                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9585
9586                 if (found_key.objectid >= key->objectid &&
9587                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9588                         struct extent_map_tree *em_tree;
9589                         struct extent_map *em;
9590
9591                         em_tree = &root->fs_info->mapping_tree.map_tree;
9592                         read_lock(&em_tree->lock);
9593                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9594                                                    found_key.offset);
9595                         read_unlock(&em_tree->lock);
9596                         if (!em) {
9597                                 btrfs_err(fs_info,
9598                         "logical %llu len %llu found bg but no related chunk",
9599                                           found_key.objectid, found_key.offset);
9600                                 ret = -ENOENT;
9601                         } else if (em->start != found_key.objectid ||
9602                                    em->len != found_key.offset) {
9603                                 btrfs_err(fs_info,
9604                 "block group %llu len %llu mismatch with chunk %llu len %llu",
9605                                           found_key.objectid, found_key.offset,
9606                                           em->start, em->len);
9607                                 ret = -EUCLEAN;
9608                         } else {
9609                                 read_extent_buffer(leaf, &bg,
9610                                         btrfs_item_ptr_offset(leaf, slot),
9611                                         sizeof(bg));
9612                                 flags = btrfs_block_group_flags(&bg) &
9613                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
9614
9615                                 if (flags != (em->map_lookup->type &
9616                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9617                                         btrfs_err(fs_info,
9618 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9619                                                 found_key.objectid,
9620                                                 found_key.offset, flags,
9621                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9622                                                  em->map_lookup->type));
9623                                         ret = -EUCLEAN;
9624                                 } else {
9625                                         ret = 0;
9626                                 }
9627                         }
9628                         free_extent_map(em);
9629                         goto out;
9630                 }
9631                 path->slots[0]++;
9632         }
9633 out:
9634         return ret;
9635 }
9636
9637 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9638 {
9639         struct btrfs_block_group_cache *block_group;
9640         u64 last = 0;
9641
9642         while (1) {
9643                 struct inode *inode;
9644
9645                 block_group = btrfs_lookup_first_block_group(info, last);
9646                 while (block_group) {
9647                         wait_block_group_cache_done(block_group);
9648                         spin_lock(&block_group->lock);
9649                         if (block_group->iref)
9650                                 break;
9651                         spin_unlock(&block_group->lock);
9652                         block_group = next_block_group(info, block_group);
9653                 }
9654                 if (!block_group) {
9655                         if (last == 0)
9656                                 break;
9657                         last = 0;
9658                         continue;
9659                 }
9660
9661                 inode = block_group->inode;
9662                 block_group->iref = 0;
9663                 block_group->inode = NULL;
9664                 spin_unlock(&block_group->lock);
9665                 ASSERT(block_group->io_ctl.inode == NULL);
9666                 iput(inode);
9667                 last = block_group->key.objectid + block_group->key.offset;
9668                 btrfs_put_block_group(block_group);
9669         }
9670 }
9671
9672 /*
9673  * Must be called only after stopping all workers, since we could have block
9674  * group caching kthreads running, and therefore they could race with us if we
9675  * freed the block groups before stopping them.
9676  */
9677 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9678 {
9679         struct btrfs_block_group_cache *block_group;
9680         struct btrfs_space_info *space_info;
9681         struct btrfs_caching_control *caching_ctl;
9682         struct rb_node *n;
9683
9684         down_write(&info->commit_root_sem);
9685         while (!list_empty(&info->caching_block_groups)) {
9686                 caching_ctl = list_entry(info->caching_block_groups.next,
9687                                          struct btrfs_caching_control, list);
9688                 list_del(&caching_ctl->list);
9689                 put_caching_control(caching_ctl);
9690         }
9691         up_write(&info->commit_root_sem);
9692
9693         spin_lock(&info->unused_bgs_lock);
9694         while (!list_empty(&info->unused_bgs)) {
9695                 block_group = list_first_entry(&info->unused_bgs,
9696                                                struct btrfs_block_group_cache,
9697                                                bg_list);
9698                 list_del_init(&block_group->bg_list);
9699                 btrfs_put_block_group(block_group);
9700         }
9701         spin_unlock(&info->unused_bgs_lock);
9702
9703         spin_lock(&info->block_group_cache_lock);
9704         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9705                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9706                                        cache_node);
9707                 rb_erase(&block_group->cache_node,
9708                          &info->block_group_cache_tree);
9709                 RB_CLEAR_NODE(&block_group->cache_node);
9710                 spin_unlock(&info->block_group_cache_lock);
9711
9712                 down_write(&block_group->space_info->groups_sem);
9713                 list_del(&block_group->list);
9714                 up_write(&block_group->space_info->groups_sem);
9715
9716                 /*
9717                  * We haven't cached this block group, which means we could
9718                  * possibly have excluded extents on this block group.
9719                  */
9720                 if (block_group->cached == BTRFS_CACHE_NO ||
9721                     block_group->cached == BTRFS_CACHE_ERROR)
9722                         free_excluded_extents(block_group);
9723
9724                 btrfs_remove_free_space_cache(block_group);
9725                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9726                 ASSERT(list_empty(&block_group->dirty_list));
9727                 ASSERT(list_empty(&block_group->io_list));
9728                 ASSERT(list_empty(&block_group->bg_list));
9729                 ASSERT(atomic_read(&block_group->count) == 1);
9730                 btrfs_put_block_group(block_group);
9731
9732                 spin_lock(&info->block_group_cache_lock);
9733         }
9734         spin_unlock(&info->block_group_cache_lock);
9735
9736         /* now that all the block groups are freed, go through and
9737          * free all the space_info structs.  This is only called during
9738          * the final stages of unmount, and so we know nobody is
9739          * using them.  We call synchronize_rcu() once before we start,
9740          * just to be on the safe side.
9741          */
9742         synchronize_rcu();
9743
9744         release_global_block_rsv(info);
9745
9746         while (!list_empty(&info->space_info)) {
9747                 int i;
9748
9749                 space_info = list_entry(info->space_info.next,
9750                                         struct btrfs_space_info,
9751                                         list);
9752
9753                 /*
9754                  * Do not hide this behind enospc_debug, this is actually
9755                  * important and indicates a real bug if this happens.
9756                  */
9757                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9758                             space_info->bytes_reserved > 0 ||
9759                             space_info->bytes_may_use > 0))
9760                         dump_space_info(info, space_info, 0, 0);
9761                 list_del(&space_info->list);
9762                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9763                         struct kobject *kobj;
9764                         kobj = space_info->block_group_kobjs[i];
9765                         space_info->block_group_kobjs[i] = NULL;
9766                         if (kobj) {
9767                                 kobject_del(kobj);
9768                                 kobject_put(kobj);
9769                         }
9770                 }
9771                 kobject_del(&space_info->kobj);
9772                 kobject_put(&space_info->kobj);
9773         }
9774         return 0;
9775 }
9776
9777 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9778 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9779 {
9780         struct btrfs_space_info *space_info;
9781         struct raid_kobject *rkobj;
9782         LIST_HEAD(list);
9783         int index;
9784         int ret = 0;
9785
9786         spin_lock(&fs_info->pending_raid_kobjs_lock);
9787         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9788         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9789
9790         list_for_each_entry(rkobj, &list, list) {
9791                 space_info = __find_space_info(fs_info, rkobj->flags);
9792                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9793
9794                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9795                                   "%s", get_raid_name(index));
9796                 if (ret) {
9797                         kobject_put(&rkobj->kobj);
9798                         break;
9799                 }
9800         }
9801         if (ret)
9802                 btrfs_warn(fs_info,
9803                            "failed to add kobject for block cache, ignoring");
9804 }
9805
9806 static void link_block_group(struct btrfs_block_group_cache *cache)
9807 {
9808         struct btrfs_space_info *space_info = cache->space_info;
9809         struct btrfs_fs_info *fs_info = cache->fs_info;
9810         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9811         bool first = false;
9812
9813         down_write(&space_info->groups_sem);
9814         if (list_empty(&space_info->block_groups[index]))
9815                 first = true;
9816         list_add_tail(&cache->list, &space_info->block_groups[index]);
9817         up_write(&space_info->groups_sem);
9818
9819         if (first) {
9820                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9821                 if (!rkobj) {
9822                         btrfs_warn(cache->fs_info,
9823                                 "couldn't alloc memory for raid level kobject");
9824                         return;
9825                 }
9826                 rkobj->flags = cache->flags;
9827                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9828
9829                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9830                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9831                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9832                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9833         }
9834 }
9835
9836 static struct btrfs_block_group_cache *
9837 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9838                                u64 start, u64 size)
9839 {
9840         struct btrfs_block_group_cache *cache;
9841
9842         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9843         if (!cache)
9844                 return NULL;
9845
9846         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9847                                         GFP_NOFS);
9848         if (!cache->free_space_ctl) {
9849                 kfree(cache);
9850                 return NULL;
9851         }
9852
9853         cache->key.objectid = start;
9854         cache->key.offset = size;
9855         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9856
9857         cache->fs_info = fs_info;
9858         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9859         set_free_space_tree_thresholds(cache);
9860
9861         atomic_set(&cache->count, 1);
9862         spin_lock_init(&cache->lock);
9863         init_rwsem(&cache->data_rwsem);
9864         INIT_LIST_HEAD(&cache->list);
9865         INIT_LIST_HEAD(&cache->cluster_list);
9866         INIT_LIST_HEAD(&cache->bg_list);
9867         INIT_LIST_HEAD(&cache->ro_list);
9868         INIT_LIST_HEAD(&cache->dirty_list);
9869         INIT_LIST_HEAD(&cache->io_list);
9870         btrfs_init_free_space_ctl(cache);
9871         atomic_set(&cache->trimming, 0);
9872         mutex_init(&cache->free_space_lock);
9873         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9874
9875         return cache;
9876 }
9877
9878
9879 /*
9880  * Iterate all chunks and verify that each of them has the corresponding block
9881  * group
9882  */
9883 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
9884 {
9885         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
9886         struct extent_map *em;
9887         struct btrfs_block_group_cache *bg;
9888         u64 start = 0;
9889         int ret = 0;
9890
9891         while (1) {
9892                 read_lock(&map_tree->map_tree.lock);
9893                 /*
9894                  * lookup_extent_mapping will return the first extent map
9895                  * intersecting the range, so setting @len to 1 is enough to
9896                  * get the first chunk.
9897                  */
9898                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
9899                 read_unlock(&map_tree->map_tree.lock);
9900                 if (!em)
9901                         break;
9902
9903                 bg = btrfs_lookup_block_group(fs_info, em->start);
9904                 if (!bg) {
9905                         btrfs_err(fs_info,
9906         "chunk start=%llu len=%llu doesn't have corresponding block group",
9907                                      em->start, em->len);
9908                         ret = -EUCLEAN;
9909                         free_extent_map(em);
9910                         break;
9911                 }
9912                 if (bg->key.objectid != em->start ||
9913                     bg->key.offset != em->len ||
9914                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
9915                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9916                         btrfs_err(fs_info,
9917 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
9918                                 em->start, em->len,
9919                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
9920                                 bg->key.objectid, bg->key.offset,
9921                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
9922                         ret = -EUCLEAN;
9923                         free_extent_map(em);
9924                         btrfs_put_block_group(bg);
9925                         break;
9926                 }
9927                 start = em->start + em->len;
9928                 free_extent_map(em);
9929                 btrfs_put_block_group(bg);
9930         }
9931         return ret;
9932 }
9933
9934 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9935 {
9936         struct btrfs_path *path;
9937         int ret;
9938         struct btrfs_block_group_cache *cache;
9939         struct btrfs_space_info *space_info;
9940         struct btrfs_key key;
9941         struct btrfs_key found_key;
9942         struct extent_buffer *leaf;
9943         int need_clear = 0;
9944         u64 cache_gen;
9945         u64 feature;
9946         int mixed;
9947
9948         feature = btrfs_super_incompat_flags(info->super_copy);
9949         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9950
9951         key.objectid = 0;
9952         key.offset = 0;
9953         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9954         path = btrfs_alloc_path();
9955         if (!path)
9956                 return -ENOMEM;
9957         path->reada = READA_FORWARD;
9958
9959         cache_gen = btrfs_super_cache_generation(info->super_copy);
9960         if (btrfs_test_opt(info, SPACE_CACHE) &&
9961             btrfs_super_generation(info->super_copy) != cache_gen)
9962                 need_clear = 1;
9963         if (btrfs_test_opt(info, CLEAR_CACHE))
9964                 need_clear = 1;
9965
9966         while (1) {
9967                 ret = find_first_block_group(info, path, &key);
9968                 if (ret > 0)
9969                         break;
9970                 if (ret != 0)
9971                         goto error;
9972
9973                 leaf = path->nodes[0];
9974                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9975
9976                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9977                                                        found_key.offset);
9978                 if (!cache) {
9979                         ret = -ENOMEM;
9980                         goto error;
9981                 }
9982
9983                 if (need_clear) {
9984                         /*
9985                          * When we mount with old space cache, we need to
9986                          * set BTRFS_DC_CLEAR and set dirty flag.
9987                          *
9988                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9989                          *    truncate the old free space cache inode and
9990                          *    setup a new one.
9991                          * b) Setting 'dirty flag' makes sure that we flush
9992                          *    the new space cache info onto disk.
9993                          */
9994                         if (btrfs_test_opt(info, SPACE_CACHE))
9995                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9996                 }
9997
9998                 read_extent_buffer(leaf, &cache->item,
9999                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10000                                    sizeof(cache->item));
10001                 cache->flags = btrfs_block_group_flags(&cache->item);
10002                 if (!mixed &&
10003                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10004                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10005                         btrfs_err(info,
10006 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10007                                   cache->key.objectid);
10008                         btrfs_put_block_group(cache);
10009                         ret = -EINVAL;
10010                         goto error;
10011                 }
10012
10013                 key.objectid = found_key.objectid + found_key.offset;
10014                 btrfs_release_path(path);
10015
10016                 /*
10017                  * We need to exclude the super stripes now so that the space
10018                  * info has super bytes accounted for, otherwise we'll think
10019                  * we have more space than we actually do.
10020                  */
10021                 ret = exclude_super_stripes(cache);
10022                 if (ret) {
10023                         /*
10024                          * We may have excluded something, so call this just in
10025                          * case.
10026                          */
10027                         free_excluded_extents(cache);
10028                         btrfs_put_block_group(cache);
10029                         goto error;
10030                 }
10031
10032                 /*
10033                  * check for two cases, either we are full, and therefore
10034                  * don't need to bother with the caching work since we won't
10035                  * find any space, or we are empty, and we can just add all
10036                  * the space in and be done with it.  This saves us _alot_ of
10037                  * time, particularly in the full case.
10038                  */
10039                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10040                         cache->last_byte_to_unpin = (u64)-1;
10041                         cache->cached = BTRFS_CACHE_FINISHED;
10042                         free_excluded_extents(cache);
10043                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10044                         cache->last_byte_to_unpin = (u64)-1;
10045                         cache->cached = BTRFS_CACHE_FINISHED;
10046                         add_new_free_space(cache, found_key.objectid,
10047                                            found_key.objectid +
10048                                            found_key.offset);
10049                         free_excluded_extents(cache);
10050                 }
10051
10052                 ret = btrfs_add_block_group_cache(info, cache);
10053                 if (ret) {
10054                         btrfs_remove_free_space_cache(cache);
10055                         btrfs_put_block_group(cache);
10056                         goto error;
10057                 }
10058
10059                 trace_btrfs_add_block_group(info, cache, 0);
10060                 update_space_info(info, cache->flags, found_key.offset,
10061                                   btrfs_block_group_used(&cache->item),
10062                                   cache->bytes_super, &space_info);
10063
10064                 cache->space_info = space_info;
10065
10066                 link_block_group(cache);
10067
10068                 set_avail_alloc_bits(info, cache->flags);
10069                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10070                         inc_block_group_ro(cache, 1);
10071                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10072                         ASSERT(list_empty(&cache->bg_list));
10073                         btrfs_mark_bg_unused(cache);
10074                 }
10075         }
10076
10077         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10078                 if (!(get_alloc_profile(info, space_info->flags) &
10079                       (BTRFS_BLOCK_GROUP_RAID10 |
10080                        BTRFS_BLOCK_GROUP_RAID1 |
10081                        BTRFS_BLOCK_GROUP_RAID5 |
10082                        BTRFS_BLOCK_GROUP_RAID6 |
10083                        BTRFS_BLOCK_GROUP_DUP)))
10084                         continue;
10085                 /*
10086                  * avoid allocating from un-mirrored block group if there are
10087                  * mirrored block groups.
10088                  */
10089                 list_for_each_entry(cache,
10090                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10091                                 list)
10092                         inc_block_group_ro(cache, 1);
10093                 list_for_each_entry(cache,
10094                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10095                                 list)
10096                         inc_block_group_ro(cache, 1);
10097         }
10098
10099         btrfs_add_raid_kobjects(info);
10100         init_global_block_rsv(info);
10101         ret = check_chunk_block_group_mappings(info);
10102 error:
10103         btrfs_free_path(path);
10104         return ret;
10105 }
10106
10107 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10108 {
10109         struct btrfs_fs_info *fs_info = trans->fs_info;
10110         struct btrfs_block_group_cache *block_group;
10111         struct btrfs_root *extent_root = fs_info->extent_root;
10112         struct btrfs_block_group_item item;
10113         struct btrfs_key key;
10114         int ret = 0;
10115
10116         if (!trans->can_flush_pending_bgs)
10117                 return;
10118
10119         while (!list_empty(&trans->new_bgs)) {
10120                 block_group = list_first_entry(&trans->new_bgs,
10121                                                struct btrfs_block_group_cache,
10122                                                bg_list);
10123                 if (ret)
10124                         goto next;
10125
10126                 spin_lock(&block_group->lock);
10127                 memcpy(&item, &block_group->item, sizeof(item));
10128                 memcpy(&key, &block_group->key, sizeof(key));
10129                 spin_unlock(&block_group->lock);
10130
10131                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10132                                         sizeof(item));
10133                 if (ret)
10134                         btrfs_abort_transaction(trans, ret);
10135                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10136                 if (ret)
10137                         btrfs_abort_transaction(trans, ret);
10138                 add_block_group_free_space(trans, block_group);
10139                 /* already aborted the transaction if it failed. */
10140 next:
10141                 list_del_init(&block_group->bg_list);
10142         }
10143         btrfs_trans_release_chunk_metadata(trans);
10144 }
10145
10146 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10147                            u64 type, u64 chunk_offset, u64 size)
10148 {
10149         struct btrfs_fs_info *fs_info = trans->fs_info;
10150         struct btrfs_block_group_cache *cache;
10151         int ret;
10152
10153         btrfs_set_log_full_commit(fs_info, trans);
10154
10155         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10156         if (!cache)
10157                 return -ENOMEM;
10158
10159         btrfs_set_block_group_used(&cache->item, bytes_used);
10160         btrfs_set_block_group_chunk_objectid(&cache->item,
10161                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10162         btrfs_set_block_group_flags(&cache->item, type);
10163
10164         cache->flags = type;
10165         cache->last_byte_to_unpin = (u64)-1;
10166         cache->cached = BTRFS_CACHE_FINISHED;
10167         cache->needs_free_space = 1;
10168         ret = exclude_super_stripes(cache);
10169         if (ret) {
10170                 /*
10171                  * We may have excluded something, so call this just in
10172                  * case.
10173                  */
10174                 free_excluded_extents(cache);
10175                 btrfs_put_block_group(cache);
10176                 return ret;
10177         }
10178
10179         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10180
10181         free_excluded_extents(cache);
10182
10183 #ifdef CONFIG_BTRFS_DEBUG
10184         if (btrfs_should_fragment_free_space(cache)) {
10185                 u64 new_bytes_used = size - bytes_used;
10186
10187                 bytes_used += new_bytes_used >> 1;
10188                 fragment_free_space(cache);
10189         }
10190 #endif
10191         /*
10192          * Ensure the corresponding space_info object is created and
10193          * assigned to our block group. We want our bg to be added to the rbtree
10194          * with its ->space_info set.
10195          */
10196         cache->space_info = __find_space_info(fs_info, cache->flags);
10197         ASSERT(cache->space_info);
10198
10199         ret = btrfs_add_block_group_cache(fs_info, cache);
10200         if (ret) {
10201                 btrfs_remove_free_space_cache(cache);
10202                 btrfs_put_block_group(cache);
10203                 return ret;
10204         }
10205
10206         /*
10207          * Now that our block group has its ->space_info set and is inserted in
10208          * the rbtree, update the space info's counters.
10209          */
10210         trace_btrfs_add_block_group(fs_info, cache, 1);
10211         update_space_info(fs_info, cache->flags, size, bytes_used,
10212                                 cache->bytes_super, &cache->space_info);
10213         update_global_block_rsv(fs_info);
10214
10215         link_block_group(cache);
10216
10217         list_add_tail(&cache->bg_list, &trans->new_bgs);
10218
10219         set_avail_alloc_bits(fs_info, type);
10220         return 0;
10221 }
10222
10223 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10224 {
10225         u64 extra_flags = chunk_to_extended(flags) &
10226                                 BTRFS_EXTENDED_PROFILE_MASK;
10227
10228         write_seqlock(&fs_info->profiles_lock);
10229         if (flags & BTRFS_BLOCK_GROUP_DATA)
10230                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10231         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10232                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10233         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10234                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10235         write_sequnlock(&fs_info->profiles_lock);
10236 }
10237
10238 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10239                              u64 group_start, struct extent_map *em)
10240 {
10241         struct btrfs_fs_info *fs_info = trans->fs_info;
10242         struct btrfs_root *root = fs_info->extent_root;
10243         struct btrfs_path *path;
10244         struct btrfs_block_group_cache *block_group;
10245         struct btrfs_free_cluster *cluster;
10246         struct btrfs_root *tree_root = fs_info->tree_root;
10247         struct btrfs_key key;
10248         struct inode *inode;
10249         struct kobject *kobj = NULL;
10250         int ret;
10251         int index;
10252         int factor;
10253         struct btrfs_caching_control *caching_ctl = NULL;
10254         bool remove_em;
10255
10256         block_group = btrfs_lookup_block_group(fs_info, group_start);
10257         BUG_ON(!block_group);
10258         BUG_ON(!block_group->ro);
10259
10260         trace_btrfs_remove_block_group(block_group);
10261         /*
10262          * Free the reserved super bytes from this block group before
10263          * remove it.
10264          */
10265         free_excluded_extents(block_group);
10266         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10267                                   block_group->key.offset);
10268
10269         memcpy(&key, &block_group->key, sizeof(key));
10270         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10271         factor = btrfs_bg_type_to_factor(block_group->flags);
10272
10273         /* make sure this block group isn't part of an allocation cluster */
10274         cluster = &fs_info->data_alloc_cluster;
10275         spin_lock(&cluster->refill_lock);
10276         btrfs_return_cluster_to_free_space(block_group, cluster);
10277         spin_unlock(&cluster->refill_lock);
10278
10279         /*
10280          * make sure this block group isn't part of a metadata
10281          * allocation cluster
10282          */
10283         cluster = &fs_info->meta_alloc_cluster;
10284         spin_lock(&cluster->refill_lock);
10285         btrfs_return_cluster_to_free_space(block_group, cluster);
10286         spin_unlock(&cluster->refill_lock);
10287
10288         path = btrfs_alloc_path();
10289         if (!path) {
10290                 ret = -ENOMEM;
10291                 goto out;
10292         }
10293
10294         /*
10295          * get the inode first so any iput calls done for the io_list
10296          * aren't the final iput (no unlinks allowed now)
10297          */
10298         inode = lookup_free_space_inode(fs_info, block_group, path);
10299
10300         mutex_lock(&trans->transaction->cache_write_mutex);
10301         /*
10302          * make sure our free spache cache IO is done before remove the
10303          * free space inode
10304          */
10305         spin_lock(&trans->transaction->dirty_bgs_lock);
10306         if (!list_empty(&block_group->io_list)) {
10307                 list_del_init(&block_group->io_list);
10308
10309                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10310
10311                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10312                 btrfs_wait_cache_io(trans, block_group, path);
10313                 btrfs_put_block_group(block_group);
10314                 spin_lock(&trans->transaction->dirty_bgs_lock);
10315         }
10316
10317         if (!list_empty(&block_group->dirty_list)) {
10318                 list_del_init(&block_group->dirty_list);
10319                 btrfs_put_block_group(block_group);
10320         }
10321         spin_unlock(&trans->transaction->dirty_bgs_lock);
10322         mutex_unlock(&trans->transaction->cache_write_mutex);
10323
10324         if (!IS_ERR(inode)) {
10325                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10326                 if (ret) {
10327                         btrfs_add_delayed_iput(inode);
10328                         goto out;
10329                 }
10330                 clear_nlink(inode);
10331                 /* One for the block groups ref */
10332                 spin_lock(&block_group->lock);
10333                 if (block_group->iref) {
10334                         block_group->iref = 0;
10335                         block_group->inode = NULL;
10336                         spin_unlock(&block_group->lock);
10337                         iput(inode);
10338                 } else {
10339                         spin_unlock(&block_group->lock);
10340                 }
10341                 /* One for our lookup ref */
10342                 btrfs_add_delayed_iput(inode);
10343         }
10344
10345         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10346         key.offset = block_group->key.objectid;
10347         key.type = 0;
10348
10349         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10350         if (ret < 0)
10351                 goto out;
10352         if (ret > 0)
10353                 btrfs_release_path(path);
10354         if (ret == 0) {
10355                 ret = btrfs_del_item(trans, tree_root, path);
10356                 if (ret)
10357                         goto out;
10358                 btrfs_release_path(path);
10359         }
10360
10361         spin_lock(&fs_info->block_group_cache_lock);
10362         rb_erase(&block_group->cache_node,
10363                  &fs_info->block_group_cache_tree);
10364         RB_CLEAR_NODE(&block_group->cache_node);
10365
10366         /* Once for the block groups rbtree */
10367         btrfs_put_block_group(block_group);
10368
10369         if (fs_info->first_logical_byte == block_group->key.objectid)
10370                 fs_info->first_logical_byte = (u64)-1;
10371         spin_unlock(&fs_info->block_group_cache_lock);
10372
10373         down_write(&block_group->space_info->groups_sem);
10374         /*
10375          * we must use list_del_init so people can check to see if they
10376          * are still on the list after taking the semaphore
10377          */
10378         list_del_init(&block_group->list);
10379         if (list_empty(&block_group->space_info->block_groups[index])) {
10380                 kobj = block_group->space_info->block_group_kobjs[index];
10381                 block_group->space_info->block_group_kobjs[index] = NULL;
10382                 clear_avail_alloc_bits(fs_info, block_group->flags);
10383         }
10384         up_write(&block_group->space_info->groups_sem);
10385         if (kobj) {
10386                 kobject_del(kobj);
10387                 kobject_put(kobj);
10388         }
10389
10390         if (block_group->has_caching_ctl)
10391                 caching_ctl = get_caching_control(block_group);
10392         if (block_group->cached == BTRFS_CACHE_STARTED)
10393                 wait_block_group_cache_done(block_group);
10394         if (block_group->has_caching_ctl) {
10395                 down_write(&fs_info->commit_root_sem);
10396                 if (!caching_ctl) {
10397                         struct btrfs_caching_control *ctl;
10398
10399                         list_for_each_entry(ctl,
10400                                     &fs_info->caching_block_groups, list)
10401                                 if (ctl->block_group == block_group) {
10402                                         caching_ctl = ctl;
10403                                         refcount_inc(&caching_ctl->count);
10404                                         break;
10405                                 }
10406                 }
10407                 if (caching_ctl)
10408                         list_del_init(&caching_ctl->list);
10409                 up_write(&fs_info->commit_root_sem);
10410                 if (caching_ctl) {
10411                         /* Once for the caching bgs list and once for us. */
10412                         put_caching_control(caching_ctl);
10413                         put_caching_control(caching_ctl);
10414                 }
10415         }
10416
10417         spin_lock(&trans->transaction->dirty_bgs_lock);
10418         if (!list_empty(&block_group->dirty_list)) {
10419                 WARN_ON(1);
10420         }
10421         if (!list_empty(&block_group->io_list)) {
10422                 WARN_ON(1);
10423         }
10424         spin_unlock(&trans->transaction->dirty_bgs_lock);
10425         btrfs_remove_free_space_cache(block_group);
10426
10427         spin_lock(&block_group->space_info->lock);
10428         list_del_init(&block_group->ro_list);
10429
10430         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10431                 WARN_ON(block_group->space_info->total_bytes
10432                         < block_group->key.offset);
10433                 WARN_ON(block_group->space_info->bytes_readonly
10434                         < block_group->key.offset);
10435                 WARN_ON(block_group->space_info->disk_total
10436                         < block_group->key.offset * factor);
10437         }
10438         block_group->space_info->total_bytes -= block_group->key.offset;
10439         block_group->space_info->bytes_readonly -= block_group->key.offset;
10440         block_group->space_info->disk_total -= block_group->key.offset * factor;
10441
10442         spin_unlock(&block_group->space_info->lock);
10443
10444         memcpy(&key, &block_group->key, sizeof(key));
10445
10446         mutex_lock(&fs_info->chunk_mutex);
10447         if (!list_empty(&em->list)) {
10448                 /* We're in the transaction->pending_chunks list. */
10449                 free_extent_map(em);
10450         }
10451         spin_lock(&block_group->lock);
10452         block_group->removed = 1;
10453         /*
10454          * At this point trimming can't start on this block group, because we
10455          * removed the block group from the tree fs_info->block_group_cache_tree
10456          * so no one can't find it anymore and even if someone already got this
10457          * block group before we removed it from the rbtree, they have already
10458          * incremented block_group->trimming - if they didn't, they won't find
10459          * any free space entries because we already removed them all when we
10460          * called btrfs_remove_free_space_cache().
10461          *
10462          * And we must not remove the extent map from the fs_info->mapping_tree
10463          * to prevent the same logical address range and physical device space
10464          * ranges from being reused for a new block group. This is because our
10465          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10466          * completely transactionless, so while it is trimming a range the
10467          * currently running transaction might finish and a new one start,
10468          * allowing for new block groups to be created that can reuse the same
10469          * physical device locations unless we take this special care.
10470          *
10471          * There may also be an implicit trim operation if the file system
10472          * is mounted with -odiscard. The same protections must remain
10473          * in place until the extents have been discarded completely when
10474          * the transaction commit has completed.
10475          */
10476         remove_em = (atomic_read(&block_group->trimming) == 0);
10477         /*
10478          * Make sure a trimmer task always sees the em in the pinned_chunks list
10479          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10480          * before checking block_group->removed).
10481          */
10482         if (!remove_em) {
10483                 /*
10484                  * Our em might be in trans->transaction->pending_chunks which
10485                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10486                  * and so is the fs_info->pinned_chunks list.
10487                  *
10488                  * So at this point we must be holding the chunk_mutex to avoid
10489                  * any races with chunk allocation (more specifically at
10490                  * volumes.c:contains_pending_extent()), to ensure it always
10491                  * sees the em, either in the pending_chunks list or in the
10492                  * pinned_chunks list.
10493                  */
10494                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10495         }
10496         spin_unlock(&block_group->lock);
10497
10498         mutex_unlock(&fs_info->chunk_mutex);
10499
10500         ret = remove_block_group_free_space(trans, block_group);
10501         if (ret)
10502                 goto out;
10503
10504         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10505         if (ret > 0)
10506                 ret = -EIO;
10507         if (ret < 0)
10508                 goto out;
10509
10510         ret = btrfs_del_item(trans, root, path);
10511         if (ret)
10512                 goto out;
10513
10514         if (remove_em) {
10515                 struct extent_map_tree *em_tree;
10516
10517                 em_tree = &fs_info->mapping_tree.map_tree;
10518                 write_lock(&em_tree->lock);
10519                 /*
10520                  * The em might be in the pending_chunks list, so make sure the
10521                  * chunk mutex is locked, since remove_extent_mapping() will
10522                  * delete us from that list.
10523                  */
10524                 remove_extent_mapping(em_tree, em);
10525                 write_unlock(&em_tree->lock);
10526                 /* once for the tree */
10527                 free_extent_map(em);
10528         }
10529
10530 out:
10531         /* Once for the lookup reference */
10532         btrfs_put_block_group(block_group);
10533         btrfs_free_path(path);
10534         return ret;
10535 }
10536
10537 struct btrfs_trans_handle *
10538 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10539                                      const u64 chunk_offset)
10540 {
10541         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10542         struct extent_map *em;
10543         struct map_lookup *map;
10544         unsigned int num_items;
10545
10546         read_lock(&em_tree->lock);
10547         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10548         read_unlock(&em_tree->lock);
10549         ASSERT(em && em->start == chunk_offset);
10550
10551         /*
10552          * We need to reserve 3 + N units from the metadata space info in order
10553          * to remove a block group (done at btrfs_remove_chunk() and at
10554          * btrfs_remove_block_group()), which are used for:
10555          *
10556          * 1 unit for adding the free space inode's orphan (located in the tree
10557          * of tree roots).
10558          * 1 unit for deleting the block group item (located in the extent
10559          * tree).
10560          * 1 unit for deleting the free space item (located in tree of tree
10561          * roots).
10562          * N units for deleting N device extent items corresponding to each
10563          * stripe (located in the device tree).
10564          *
10565          * In order to remove a block group we also need to reserve units in the
10566          * system space info in order to update the chunk tree (update one or
10567          * more device items and remove one chunk item), but this is done at
10568          * btrfs_remove_chunk() through a call to check_system_chunk().
10569          */
10570         map = em->map_lookup;
10571         num_items = 3 + map->num_stripes;
10572         free_extent_map(em);
10573
10574         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10575                                                            num_items, 1);
10576 }
10577
10578 /*
10579  * Process the unused_bgs list and remove any that don't have any allocated
10580  * space inside of them.
10581  */
10582 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10583 {
10584         struct btrfs_block_group_cache *block_group;
10585         struct btrfs_space_info *space_info;
10586         struct btrfs_trans_handle *trans;
10587         int ret = 0;
10588
10589         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10590                 return;
10591
10592         spin_lock(&fs_info->unused_bgs_lock);
10593         while (!list_empty(&fs_info->unused_bgs)) {
10594                 u64 start, end;
10595                 int trimming;
10596
10597                 block_group = list_first_entry(&fs_info->unused_bgs,
10598                                                struct btrfs_block_group_cache,
10599                                                bg_list);
10600                 list_del_init(&block_group->bg_list);
10601
10602                 space_info = block_group->space_info;
10603
10604                 if (ret || btrfs_mixed_space_info(space_info)) {
10605                         btrfs_put_block_group(block_group);
10606                         continue;
10607                 }
10608                 spin_unlock(&fs_info->unused_bgs_lock);
10609
10610                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10611
10612                 /* Don't want to race with allocators so take the groups_sem */
10613                 down_write(&space_info->groups_sem);
10614                 spin_lock(&block_group->lock);
10615                 if (block_group->reserved || block_group->pinned ||
10616                     btrfs_block_group_used(&block_group->item) ||
10617                     block_group->ro ||
10618                     list_is_singular(&block_group->list)) {
10619                         /*
10620                          * We want to bail if we made new allocations or have
10621                          * outstanding allocations in this block group.  We do
10622                          * the ro check in case balance is currently acting on
10623                          * this block group.
10624                          */
10625                         trace_btrfs_skip_unused_block_group(block_group);
10626                         spin_unlock(&block_group->lock);
10627                         up_write(&space_info->groups_sem);
10628                         goto next;
10629                 }
10630                 spin_unlock(&block_group->lock);
10631
10632                 /* We don't want to force the issue, only flip if it's ok. */
10633                 ret = inc_block_group_ro(block_group, 0);
10634                 up_write(&space_info->groups_sem);
10635                 if (ret < 0) {
10636                         ret = 0;
10637                         goto next;
10638                 }
10639
10640                 /*
10641                  * Want to do this before we do anything else so we can recover
10642                  * properly if we fail to join the transaction.
10643                  */
10644                 trans = btrfs_start_trans_remove_block_group(fs_info,
10645                                                      block_group->key.objectid);
10646                 if (IS_ERR(trans)) {
10647                         btrfs_dec_block_group_ro(block_group);
10648                         ret = PTR_ERR(trans);
10649                         goto next;
10650                 }
10651
10652                 /*
10653                  * We could have pending pinned extents for this block group,
10654                  * just delete them, we don't care about them anymore.
10655                  */
10656                 start = block_group->key.objectid;
10657                 end = start + block_group->key.offset - 1;
10658                 /*
10659                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10660                  * btrfs_finish_extent_commit(). If we are at transaction N,
10661                  * another task might be running finish_extent_commit() for the
10662                  * previous transaction N - 1, and have seen a range belonging
10663                  * to the block group in freed_extents[] before we were able to
10664                  * clear the whole block group range from freed_extents[]. This
10665                  * means that task can lookup for the block group after we
10666                  * unpinned it from freed_extents[] and removed it, leading to
10667                  * a BUG_ON() at btrfs_unpin_extent_range().
10668                  */
10669                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10670                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10671                                   EXTENT_DIRTY);
10672                 if (ret) {
10673                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10674                         btrfs_dec_block_group_ro(block_group);
10675                         goto end_trans;
10676                 }
10677                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10678                                   EXTENT_DIRTY);
10679                 if (ret) {
10680                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10681                         btrfs_dec_block_group_ro(block_group);
10682                         goto end_trans;
10683                 }
10684                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10685
10686                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10687                 spin_lock(&space_info->lock);
10688                 spin_lock(&block_group->lock);
10689
10690                 space_info->bytes_pinned -= block_group->pinned;
10691                 space_info->bytes_readonly += block_group->pinned;
10692                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
10693                                    -block_group->pinned,
10694                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
10695                 block_group->pinned = 0;
10696
10697                 spin_unlock(&block_group->lock);
10698                 spin_unlock(&space_info->lock);
10699
10700                 /* DISCARD can flip during remount */
10701                 trimming = btrfs_test_opt(fs_info, DISCARD);
10702
10703                 /* Implicit trim during transaction commit. */
10704                 if (trimming)
10705                         btrfs_get_block_group_trimming(block_group);
10706
10707                 /*
10708                  * Btrfs_remove_chunk will abort the transaction if things go
10709                  * horribly wrong.
10710                  */
10711                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
10712
10713                 if (ret) {
10714                         if (trimming)
10715                                 btrfs_put_block_group_trimming(block_group);
10716                         goto end_trans;
10717                 }
10718
10719                 /*
10720                  * If we're not mounted with -odiscard, we can just forget
10721                  * about this block group. Otherwise we'll need to wait
10722                  * until transaction commit to do the actual discard.
10723                  */
10724                 if (trimming) {
10725                         spin_lock(&fs_info->unused_bgs_lock);
10726                         /*
10727                          * A concurrent scrub might have added us to the list
10728                          * fs_info->unused_bgs, so use a list_move operation
10729                          * to add the block group to the deleted_bgs list.
10730                          */
10731                         list_move(&block_group->bg_list,
10732                                   &trans->transaction->deleted_bgs);
10733                         spin_unlock(&fs_info->unused_bgs_lock);
10734                         btrfs_get_block_group(block_group);
10735                 }
10736 end_trans:
10737                 btrfs_end_transaction(trans);
10738 next:
10739                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10740                 btrfs_put_block_group(block_group);
10741                 spin_lock(&fs_info->unused_bgs_lock);
10742         }
10743         spin_unlock(&fs_info->unused_bgs_lock);
10744 }
10745
10746 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10747 {
10748         struct btrfs_super_block *disk_super;
10749         u64 features;
10750         u64 flags;
10751         int mixed = 0;
10752         int ret;
10753
10754         disk_super = fs_info->super_copy;
10755         if (!btrfs_super_root(disk_super))
10756                 return -EINVAL;
10757
10758         features = btrfs_super_incompat_flags(disk_super);
10759         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10760                 mixed = 1;
10761
10762         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10763         ret = create_space_info(fs_info, flags);
10764         if (ret)
10765                 goto out;
10766
10767         if (mixed) {
10768                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10769                 ret = create_space_info(fs_info, flags);
10770         } else {
10771                 flags = BTRFS_BLOCK_GROUP_METADATA;
10772                 ret = create_space_info(fs_info, flags);
10773                 if (ret)
10774                         goto out;
10775
10776                 flags = BTRFS_BLOCK_GROUP_DATA;
10777                 ret = create_space_info(fs_info, flags);
10778         }
10779 out:
10780         return ret;
10781 }
10782
10783 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10784                                    u64 start, u64 end)
10785 {
10786         return unpin_extent_range(fs_info, start, end, false);
10787 }
10788
10789 /*
10790  * It used to be that old block groups would be left around forever.
10791  * Iterating over them would be enough to trim unused space.  Since we
10792  * now automatically remove them, we also need to iterate over unallocated
10793  * space.
10794  *
10795  * We don't want a transaction for this since the discard may take a
10796  * substantial amount of time.  We don't require that a transaction be
10797  * running, but we do need to take a running transaction into account
10798  * to ensure that we're not discarding chunks that were released or
10799  * allocated in the current transaction.
10800  *
10801  * Holding the chunks lock will prevent other threads from allocating
10802  * or releasing chunks, but it won't prevent a running transaction
10803  * from committing and releasing the memory that the pending chunks
10804  * list head uses.  For that, we need to take a reference to the
10805  * transaction and hold the commit root sem.  We only need to hold
10806  * it while performing the free space search since we have already
10807  * held back allocations.
10808  */
10809 static int btrfs_trim_free_extents(struct btrfs_device *device,
10810                                    u64 minlen, u64 *trimmed)
10811 {
10812         u64 start = 0, len = 0;
10813         int ret;
10814
10815         *trimmed = 0;
10816
10817         /* Discard not supported = nothing to do. */
10818         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
10819                 return 0;
10820
10821         /* Not writeable = nothing to do. */
10822         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10823                 return 0;
10824
10825         /* No free space = nothing to do. */
10826         if (device->total_bytes <= device->bytes_used)
10827                 return 0;
10828
10829         ret = 0;
10830
10831         while (1) {
10832                 struct btrfs_fs_info *fs_info = device->fs_info;
10833                 struct btrfs_transaction *trans;
10834                 u64 bytes;
10835
10836                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10837                 if (ret)
10838                         break;
10839
10840                 ret = down_read_killable(&fs_info->commit_root_sem);
10841                 if (ret) {
10842                         mutex_unlock(&fs_info->chunk_mutex);
10843                         break;
10844                 }
10845
10846                 spin_lock(&fs_info->trans_lock);
10847                 trans = fs_info->running_transaction;
10848                 if (trans)
10849                         refcount_inc(&trans->use_count);
10850                 spin_unlock(&fs_info->trans_lock);
10851
10852                 if (!trans)
10853                         up_read(&fs_info->commit_root_sem);
10854
10855                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10856                                                  &start, &len);
10857                 if (trans) {
10858                         up_read(&fs_info->commit_root_sem);
10859                         btrfs_put_transaction(trans);
10860                 }
10861
10862                 if (ret) {
10863                         mutex_unlock(&fs_info->chunk_mutex);
10864                         if (ret == -ENOSPC)
10865                                 ret = 0;
10866                         break;
10867                 }
10868
10869                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10870                 mutex_unlock(&fs_info->chunk_mutex);
10871
10872                 if (ret)
10873                         break;
10874
10875                 start += len;
10876                 *trimmed += bytes;
10877
10878                 if (fatal_signal_pending(current)) {
10879                         ret = -ERESTARTSYS;
10880                         break;
10881                 }
10882
10883                 cond_resched();
10884         }
10885
10886         return ret;
10887 }
10888
10889 /*
10890  * Trim the whole filesystem by:
10891  * 1) trimming the free space in each block group
10892  * 2) trimming the unallocated space on each device
10893  *
10894  * This will also continue trimming even if a block group or device encounters
10895  * an error.  The return value will be the last error, or 0 if nothing bad
10896  * happens.
10897  */
10898 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10899 {
10900         struct btrfs_block_group_cache *cache = NULL;
10901         struct btrfs_device *device;
10902         struct list_head *devices;
10903         u64 group_trimmed;
10904         u64 start;
10905         u64 end;
10906         u64 trimmed = 0;
10907         u64 bg_failed = 0;
10908         u64 dev_failed = 0;
10909         int bg_ret = 0;
10910         int dev_ret = 0;
10911         int ret = 0;
10912
10913         cache = btrfs_lookup_first_block_group(fs_info, range->start);
10914         for (; cache; cache = next_block_group(fs_info, cache)) {
10915                 if (cache->key.objectid >= (range->start + range->len)) {
10916                         btrfs_put_block_group(cache);
10917                         break;
10918                 }
10919
10920                 start = max(range->start, cache->key.objectid);
10921                 end = min(range->start + range->len,
10922                                 cache->key.objectid + cache->key.offset);
10923
10924                 if (end - start >= range->minlen) {
10925                         if (!block_group_cache_done(cache)) {
10926                                 ret = cache_block_group(cache, 0);
10927                                 if (ret) {
10928                                         bg_failed++;
10929                                         bg_ret = ret;
10930                                         continue;
10931                                 }
10932                                 ret = wait_block_group_cache_done(cache);
10933                                 if (ret) {
10934                                         bg_failed++;
10935                                         bg_ret = ret;
10936                                         continue;
10937                                 }
10938                         }
10939                         ret = btrfs_trim_block_group(cache,
10940                                                      &group_trimmed,
10941                                                      start,
10942                                                      end,
10943                                                      range->minlen);
10944
10945                         trimmed += group_trimmed;
10946                         if (ret) {
10947                                 bg_failed++;
10948                                 bg_ret = ret;
10949                                 continue;
10950                         }
10951                 }
10952         }
10953
10954         if (bg_failed)
10955                 btrfs_warn(fs_info,
10956                         "failed to trim %llu block group(s), last error %d",
10957                         bg_failed, bg_ret);
10958         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10959         devices = &fs_info->fs_devices->devices;
10960         list_for_each_entry(device, devices, dev_list) {
10961                 ret = btrfs_trim_free_extents(device, range->minlen,
10962                                               &group_trimmed);
10963                 if (ret) {
10964                         dev_failed++;
10965                         dev_ret = ret;
10966                         break;
10967                 }
10968
10969                 trimmed += group_trimmed;
10970         }
10971         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10972
10973         if (dev_failed)
10974                 btrfs_warn(fs_info,
10975                         "failed to trim %llu device(s), last error %d",
10976                         dev_failed, dev_ret);
10977         range->len = trimmed;
10978         if (bg_ret)
10979                 return bg_ret;
10980         return dev_ret;
10981 }
10982
10983 /*
10984  * btrfs_{start,end}_write_no_snapshotting() are similar to
10985  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10986  * data into the page cache through nocow before the subvolume is snapshoted,
10987  * but flush the data into disk after the snapshot creation, or to prevent
10988  * operations while snapshotting is ongoing and that cause the snapshot to be
10989  * inconsistent (writes followed by expanding truncates for example).
10990  */
10991 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
10992 {
10993         percpu_counter_dec(&root->subv_writers->counter);
10994         cond_wake_up(&root->subv_writers->wait);
10995 }
10996
10997 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
10998 {
10999         if (atomic_read(&root->will_be_snapshotted))
11000                 return 0;
11001
11002         percpu_counter_inc(&root->subv_writers->counter);
11003         /*
11004          * Make sure counter is updated before we check for snapshot creation.
11005          */
11006         smp_mb();
11007         if (atomic_read(&root->will_be_snapshotted)) {
11008                 btrfs_end_write_no_snapshotting(root);
11009                 return 0;
11010         }
11011         return 1;
11012 }
11013
11014 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11015 {
11016         while (true) {
11017                 int ret;
11018
11019                 ret = btrfs_start_write_no_snapshotting(root);
11020                 if (ret)
11021                         break;
11022                 wait_var_event(&root->will_be_snapshotted,
11023                                !atomic_read(&root->will_be_snapshotted));
11024         }
11025 }
11026
11027 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11028 {
11029         struct btrfs_fs_info *fs_info = bg->fs_info;
11030
11031         spin_lock(&fs_info->unused_bgs_lock);
11032         if (list_empty(&bg->bg_list)) {
11033                 btrfs_get_block_group(bg);
11034                 trace_btrfs_add_unused_block_group(bg);
11035                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11036         }
11037         spin_unlock(&fs_info->unused_bgs_lock);
11038 }