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/mm.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/sort.h>
  13 #include <linux/rcupdate.h>
  14 #include <linux/kthread.h>
  15 #include <linux/slab.h>
  16 #include <linux/ratelimit.h>
  17 #include <linux/percpu_counter.h>
  18 #include <linux/lockdep.h>
  19 #include <linux/crc32c.h>
  20 #include "tree-log.h"
  21 #include "disk-io.h"
  22 #include "print-tree.h"
  23 #include "volumes.h"
  24 #include "raid56.h"
  25 #include "locking.h"
  26 #include "free-space-cache.h"
  27 #include "free-space-tree.h"
  28 #include "math.h"
  29 #include "sysfs.h"
  30 #include "qgroup.h"
  31 #include "ref-verify.h"
  32 #include "space-info.h"
  33 #include "block-rsv.h"
  34 #include "delalloc-space.h"
  35
  36 #undef SCRAMBLE_DELAYED_REFS
  37
  38
  39 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  40                                struct btrfs_delayed_ref_node *node, u64 parent,
  41                                u64 root_objectid, u64 owner_objectid,
  42                                u64 owner_offset, int refs_to_drop,
  43                                struct btrfs_delayed_extent_op *extra_op);
  44 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  45                                     struct extent_buffer *leaf,
  46                                     struct btrfs_extent_item *ei);
  47 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  48                                       u64 parent, u64 root_objectid,
  49                                       u64 flags, u64 owner, u64 offset,
  50                                       struct btrfs_key *ins, int ref_mod);
  51 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  52                                      struct btrfs_delayed_ref_node *node,
  53                                      struct btrfs_delayed_extent_op *extent_op);
  54 static int find_next_key(struct btrfs_path *path, int level,
  55                          struct btrfs_key *key);
  56
  57 static noinline int
  58 block_group_cache_done(struct btrfs_block_group_cache *cache)
  59 {
  60         smp_mb();
  61         return cache->cached == BTRFS_CACHE_FINISHED ||
  62                 cache->cached == BTRFS_CACHE_ERROR;
  63 }
  64
  65 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  66 {
  67         return (cache->flags & bits) == bits;
  68 }
  69
  70 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
  71 {
  72         atomic_inc(&cache->count);
  73 }
  74
  75 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
  76 {
  77         if (atomic_dec_and_test(&cache->count)) {
  78                 WARN_ON(cache->pinned > 0);
  79                 WARN_ON(cache->reserved > 0);
  80
  81                 /*
  82                  * If not empty, someone is still holding mutex of
  83                  * full_stripe_lock, which can only be released by caller.
  84                  * And it will definitely cause use-after-free when caller
  85                  * tries to release full stripe lock.
  86                  *
  87                  * No better way to resolve, but only to warn.
  88                  */
  89                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
  90                 kfree(cache->free_space_ctl);
  91                 kfree(cache);
  92         }
  93 }
  94
  95 /*
  96  * this adds the block group to the fs_info rb tree for the block group
  97  * cache
  98  */
  99 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 100                                 struct btrfs_block_group_cache *block_group)
 101 {
 102         struct rb_node **p;
 103         struct rb_node *parent = NULL;
 104         struct btrfs_block_group_cache *cache;
 105
 106         spin_lock(&info->block_group_cache_lock);
 107         p = &info->block_group_cache_tree.rb_node;
 108
 109         while (*p) {
 110                 parent = *p;
 111                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 112                                  cache_node);
 113                 if (block_group->key.objectid < cache->key.objectid) {
 114                         p = &(*p)->rb_left;
 115                 } else if (block_group->key.objectid > cache->key.objectid) {
 116                         p = &(*p)->rb_right;
 117                 } else {
 118                         spin_unlock(&info->block_group_cache_lock);
 119                         return -EEXIST;
 120                 }
 121         }
 122
 123         rb_link_node(&block_group->cache_node, parent, p);
 124         rb_insert_color(&block_group->cache_node,
 125                         &info->block_group_cache_tree);
 126
 127         if (info->first_logical_byte > block_group->key.objectid)
 128                 info->first_logical_byte = block_group->key.objectid;
 129
 130         spin_unlock(&info->block_group_cache_lock);
 131
 132         return 0;
 133 }
 134
 135 /*
 136  * This will return the block group at or after bytenr if contains is 0, else
 137  * it will return the block group that contains the bytenr
 138  */
 139 static struct btrfs_block_group_cache *
 140 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 141                               int contains)
 142 {
 143         struct btrfs_block_group_cache *cache, *ret = NULL;
 144         struct rb_node *n;
 145         u64 end, start;
 146
 147         spin_lock(&info->block_group_cache_lock);
 148         n = info->block_group_cache_tree.rb_node;
 149
 150         while (n) {
 151                 cache = rb_entry(n, struct btrfs_block_group_cache,
 152                                  cache_node);
 153                 end = cache->key.objectid + cache->key.offset - 1;
 154                 start = cache->key.objectid;
 155
 156                 if (bytenr < start) {
 157                         if (!contains && (!ret || start < ret->key.objectid))
 158                                 ret = cache;
 159                         n = n->rb_left;
 160                 } else if (bytenr > start) {
 161                         if (contains && bytenr <= end) {
 162                                 ret = cache;
 163                                 break;
 164                         }
 165                         n = n->rb_right;
 166                 } else {
 167                         ret = cache;
 168                         break;
 169                 }
 170         }
 171         if (ret) {
 172                 btrfs_get_block_group(ret);
 173                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 174                         info->first_logical_byte = ret->key.objectid;
 175         }
 176         spin_unlock(&info->block_group_cache_lock);
 177
 178         return ret;
 179 }
 180
 181 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 182                                u64 start, u64 num_bytes)
 183 {
 184         u64 end = start + num_bytes - 1;
 185         set_extent_bits(&fs_info->freed_extents[0],
 186                         start, end, EXTENT_UPTODATE);
 187         set_extent_bits(&fs_info->freed_extents[1],
 188                         start, end, EXTENT_UPTODATE);
 189         return 0;
 190 }
 191
 192 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 193 {
 194         struct btrfs_fs_info *fs_info = cache->fs_info;
 195         u64 start, end;
 196
 197         start = cache->key.objectid;
 198         end = start + cache->key.offset - 1;
 199
 200         clear_extent_bits(&fs_info->freed_extents[0],
 201                           start, end, EXTENT_UPTODATE);
 202         clear_extent_bits(&fs_info->freed_extents[1],
 203                           start, end, EXTENT_UPTODATE);
 204 }
 205
 206 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 207 {
 208         struct btrfs_fs_info *fs_info = cache->fs_info;
 209         u64 bytenr;
 210         u64 *logical;
 211         int stripe_len;
 212         int i, nr, ret;
 213
 214         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 215                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 216                 cache->bytes_super += stripe_len;
 217                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 218                                           stripe_len);
 219                 if (ret)
 220                         return ret;
 221         }
 222
 223         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 224                 bytenr = btrfs_sb_offset(i);
 225                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 226                                        bytenr, &logical, &nr, &stripe_len);
 227                 if (ret)
 228                         return ret;
 229
 230                 while (nr--) {
 231                         u64 start, len;
 232
 233                         if (logical[nr] > cache->key.objectid +
 234                             cache->key.offset)
 235                                 continue;
 236
 237                         if (logical[nr] + stripe_len <= cache->key.objectid)
 238                                 continue;
 239
 240                         start = logical[nr];
 241                         if (start < cache->key.objectid) {
 242                                 start = cache->key.objectid;
 243                                 len = (logical[nr] + stripe_len) - start;
 244                         } else {
 245                                 len = min_t(u64, stripe_len,
 246                                             cache->key.objectid +
 247                                             cache->key.offset - start);
 248                         }
 249
 250                         cache->bytes_super += len;
 251                         ret = add_excluded_extent(fs_info, start, len);
 252                         if (ret) {
 253                                 kfree(logical);
 254                                 return ret;
 255                         }
 256                 }
 257
 258                 kfree(logical);
 259         }
 260         return 0;
 261 }
 262
 263 static struct btrfs_caching_control *
 264 get_caching_control(struct btrfs_block_group_cache *cache)
 265 {
 266         struct btrfs_caching_control *ctl;
 267
 268         spin_lock(&cache->lock);
 269         if (!cache->caching_ctl) {
 270                 spin_unlock(&cache->lock);
 271                 return NULL;
 272         }
 273
 274         ctl = cache->caching_ctl;
 275         refcount_inc(&ctl->count);
 276         spin_unlock(&cache->lock);
 277         return ctl;
 278 }
 279
 280 static void put_caching_control(struct btrfs_caching_control *ctl)
 281 {
 282         if (refcount_dec_and_test(&ctl->count))
 283                 kfree(ctl);
 284 }
 285
 286 #ifdef CONFIG_BTRFS_DEBUG
 287 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 288 {
 289         struct btrfs_fs_info *fs_info = block_group->fs_info;
 290         u64 start = block_group->key.objectid;
 291         u64 len = block_group->key.offset;
 292         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 293                 fs_info->nodesize : fs_info->sectorsize;
 294         u64 step = chunk << 1;
 295
 296         while (len > chunk) {
 297                 btrfs_remove_free_space(block_group, start, chunk);
 298                 start += step;
 299                 if (len < step)
 300                         len = 0;
 301                 else
 302                         len -= step;
 303         }
 304 }
 305 #endif
 306
 307 /*
 308  * this is only called by cache_block_group, since we could have freed extents
 309  * we need to check the pinned_extents for any extents that can't be used yet
 310  * since their free space will be released as soon as the transaction commits.
 311  */
 312 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 313                        u64 start, u64 end)
 314 {
 315         struct btrfs_fs_info *info = block_group->fs_info;
 316         u64 extent_start, extent_end, size, total_added = 0;
 317         int ret;
 318
 319         while (start < end) {
 320                 ret = find_first_extent_bit(info->pinned_extents, start,
 321                                             &extent_start, &extent_end,
 322                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 323                                             NULL);
 324                 if (ret)
 325                         break;
 326
 327                 if (extent_start <= start) {
 328                         start = extent_end + 1;
 329                 } else if (extent_start > start && extent_start < end) {
 330                         size = extent_start - start;
 331                         total_added += size;
 332                         ret = btrfs_add_free_space(block_group, start,
 333                                                    size);
 334                         BUG_ON(ret); /* -ENOMEM or logic error */
 335                         start = extent_end + 1;
 336                 } else {
 337                         break;
 338                 }
 339         }
 340
 341         if (start < end) {
 342                 size = end - start;
 343                 total_added += size;
 344                 ret = btrfs_add_free_space(block_group, start, size);
 345                 BUG_ON(ret); /* -ENOMEM or logic error */
 346         }
 347
 348         return total_added;
 349 }
 350
 351 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 352 {
 353         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 354         struct btrfs_fs_info *fs_info = block_group->fs_info;
 355         struct btrfs_root *extent_root = fs_info->extent_root;
 356         struct btrfs_path *path;
 357         struct extent_buffer *leaf;
 358         struct btrfs_key key;
 359         u64 total_found = 0;
 360         u64 last = 0;
 361         u32 nritems;
 362         int ret;
 363         bool wakeup = true;
 364
 365         path = btrfs_alloc_path();
 366         if (!path)
 367                 return -ENOMEM;
 368
 369         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 370
 371 #ifdef CONFIG_BTRFS_DEBUG
 372         /*
 373          * If we're fragmenting we don't want to make anybody think we can
 374          * allocate from this block group until we've had a chance to fragment
 375          * the free space.
 376          */
 377         if (btrfs_should_fragment_free_space(block_group))
 378                 wakeup = false;
 379 #endif
 380         /*
 381          * We don't want to deadlock with somebody trying to allocate a new
 382          * extent for the extent root while also trying to search the extent
 383          * root to add free space.  So we skip locking and search the commit
 384          * root, since its read-only
 385          */
 386         path->skip_locking = 1;
 387         path->search_commit_root = 1;
 388         path->reada = READA_FORWARD;
 389
 390         key.objectid = last;
 391         key.offset = 0;
 392         key.type = BTRFS_EXTENT_ITEM_KEY;
 393
 394 next:
 395         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 396         if (ret < 0)
 397                 goto out;
 398
 399         leaf = path->nodes[0];
 400         nritems = btrfs_header_nritems(leaf);
 401
 402         while (1) {
 403                 if (btrfs_fs_closing(fs_info) > 1) {
 404                         last = (u64)-1;
 405                         break;
 406                 }
 407
 408                 if (path->slots[0] < nritems) {
 409                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 410                 } else {
 411                         ret = find_next_key(path, 0, &key);
 412                         if (ret)
 413                                 break;
 414
 415                         if (need_resched() ||
 416                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 417                                 if (wakeup)
 418                                         caching_ctl->progress = last;
 419                                 btrfs_release_path(path);
 420                                 up_read(&fs_info->commit_root_sem);
 421                                 mutex_unlock(&caching_ctl->mutex);
 422                                 cond_resched();
 423                                 mutex_lock(&caching_ctl->mutex);
 424                                 down_read(&fs_info->commit_root_sem);
 425                                 goto next;
 426                         }
 427
 428                         ret = btrfs_next_leaf(extent_root, path);
 429                         if (ret < 0)
 430                                 goto out;
 431                         if (ret)
 432                                 break;
 433                         leaf = path->nodes[0];
 434                         nritems = btrfs_header_nritems(leaf);
 435                         continue;
 436                 }
 437
 438                 if (key.objectid < last) {
 439                         key.objectid = last;
 440                         key.offset = 0;
 441                         key.type = BTRFS_EXTENT_ITEM_KEY;
 442
 443                         if (wakeup)
 444                                 caching_ctl->progress = last;
 445                         btrfs_release_path(path);
 446                         goto next;
 447                 }
 448
 449                 if (key.objectid < block_group->key.objectid) {
 450                         path->slots[0]++;
 451                         continue;
 452                 }
 453
 454                 if (key.objectid >= block_group->key.objectid +
 455                     block_group->key.offset)
 456                         break;
 457
 458                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 459                     key.type == BTRFS_METADATA_ITEM_KEY) {
 460                         total_found += add_new_free_space(block_group, last,
 461                                                           key.objectid);
 462                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 463                                 last = key.objectid +
 464                                         fs_info->nodesize;
 465                         else
 466                                 last = key.objectid + key.offset;
 467
 468                         if (total_found > CACHING_CTL_WAKE_UP) {
 469                                 total_found = 0;
 470                                 if (wakeup)
 471                                         wake_up(&caching_ctl->wait);
 472                         }
 473                 }
 474                 path->slots[0]++;
 475         }
 476         ret = 0;
 477
 478         total_found += add_new_free_space(block_group, last,
 479                                           block_group->key.objectid +
 480                                           block_group->key.offset);
 481         caching_ctl->progress = (u64)-1;
 482
 483 out:
 484         btrfs_free_path(path);
 485         return ret;
 486 }
 487
 488 static noinline void caching_thread(struct btrfs_work *work)
 489 {
 490         struct btrfs_block_group_cache *block_group;
 491         struct btrfs_fs_info *fs_info;
 492         struct btrfs_caching_control *caching_ctl;
 493         int ret;
 494
 495         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 496         block_group = caching_ctl->block_group;
 497         fs_info = block_group->fs_info;
 498
 499         mutex_lock(&caching_ctl->mutex);
 500         down_read(&fs_info->commit_root_sem);
 501
 502         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 503                 ret = load_free_space_tree(caching_ctl);
 504         else
 505                 ret = load_extent_tree_free(caching_ctl);
 506
 507         spin_lock(&block_group->lock);
 508         block_group->caching_ctl = NULL;
 509         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 510         spin_unlock(&block_group->lock);
 511
 512 #ifdef CONFIG_BTRFS_DEBUG
 513         if (btrfs_should_fragment_free_space(block_group)) {
 514                 u64 bytes_used;
 515
 516                 spin_lock(&block_group->space_info->lock);
 517                 spin_lock(&block_group->lock);
 518                 bytes_used = block_group->key.offset -
 519                         btrfs_block_group_used(&block_group->item);
 520                 block_group->space_info->bytes_used += bytes_used >> 1;
 521                 spin_unlock(&block_group->lock);
 522                 spin_unlock(&block_group->space_info->lock);
 523                 fragment_free_space(block_group);
 524         }
 525 #endif
 526
 527         caching_ctl->progress = (u64)-1;
 528
 529         up_read(&fs_info->commit_root_sem);
 530         free_excluded_extents(block_group);
 531         mutex_unlock(&caching_ctl->mutex);
 532
 533         wake_up(&caching_ctl->wait);
 534
 535         put_caching_control(caching_ctl);
 536         btrfs_put_block_group(block_group);
 537 }
 538
 539 static int cache_block_group(struct btrfs_block_group_cache *cache,
 540                              int load_cache_only)
 541 {
 542         DEFINE_WAIT(wait);
 543         struct btrfs_fs_info *fs_info = cache->fs_info;
 544         struct btrfs_caching_control *caching_ctl;
 545         int ret = 0;
 546
 547         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 548         if (!caching_ctl)
 549                 return -ENOMEM;
 550
 551         INIT_LIST_HEAD(&caching_ctl->list);
 552         mutex_init(&caching_ctl->mutex);
 553         init_waitqueue_head(&caching_ctl->wait);
 554         caching_ctl->block_group = cache;
 555         caching_ctl->progress = cache->key.objectid;
 556         refcount_set(&caching_ctl->count, 1);
 557         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 558                         caching_thread, NULL, NULL);
 559
 560         spin_lock(&cache->lock);
 561         /*
 562          * This should be a rare occasion, but this could happen I think in the
 563          * case where one thread starts to load the space cache info, and then
 564          * some other thread starts a transaction commit which tries to do an
 565          * allocation while the other thread is still loading the space cache
 566          * info.  The previous loop should have kept us from choosing this block
 567          * group, but if we've moved to the state where we will wait on caching
 568          * block groups we need to first check if we're doing a fast load here,
 569          * so we can wait for it to finish, otherwise we could end up allocating
 570          * from a block group who's cache gets evicted for one reason or
 571          * another.
 572          */
 573         while (cache->cached == BTRFS_CACHE_FAST) {
 574                 struct btrfs_caching_control *ctl;
 575
 576                 ctl = cache->caching_ctl;
 577                 refcount_inc(&ctl->count);
 578                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 579                 spin_unlock(&cache->lock);
 580
 581                 schedule();
 582
 583                 finish_wait(&ctl->wait, &wait);
 584                 put_caching_control(ctl);
 585                 spin_lock(&cache->lock);
 586         }
 587
 588         if (cache->cached != BTRFS_CACHE_NO) {
 589                 spin_unlock(&cache->lock);
 590                 kfree(caching_ctl);
 591                 return 0;
 592         }
 593         WARN_ON(cache->caching_ctl);
 594         cache->caching_ctl = caching_ctl;
 595         cache->cached = BTRFS_CACHE_FAST;
 596         spin_unlock(&cache->lock);
 597
 598         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 599                 mutex_lock(&caching_ctl->mutex);
 600                 ret = load_free_space_cache(cache);
 601
 602                 spin_lock(&cache->lock);
 603                 if (ret == 1) {
 604                         cache->caching_ctl = NULL;
 605                         cache->cached = BTRFS_CACHE_FINISHED;
 606                         cache->last_byte_to_unpin = (u64)-1;
 607                         caching_ctl->progress = (u64)-1;
 608                 } else {
 609                         if (load_cache_only) {
 610                                 cache->caching_ctl = NULL;
 611                                 cache->cached = BTRFS_CACHE_NO;
 612                         } else {
 613                                 cache->cached = BTRFS_CACHE_STARTED;
 614                                 cache->has_caching_ctl = 1;
 615                         }
 616                 }
 617                 spin_unlock(&cache->lock);
 618 #ifdef CONFIG_BTRFS_DEBUG
 619                 if (ret == 1 &&
 620                     btrfs_should_fragment_free_space(cache)) {
 621                         u64 bytes_used;
 622
 623                         spin_lock(&cache->space_info->lock);
 624                         spin_lock(&cache->lock);
 625                         bytes_used = cache->key.offset -
 626                                 btrfs_block_group_used(&cache->item);
 627                         cache->space_info->bytes_used += bytes_used >> 1;
 628                         spin_unlock(&cache->lock);
 629                         spin_unlock(&cache->space_info->lock);
 630                         fragment_free_space(cache);
 631                 }
 632 #endif
 633                 mutex_unlock(&caching_ctl->mutex);
 634
 635                 wake_up(&caching_ctl->wait);
 636                 if (ret == 1) {
 637                         put_caching_control(caching_ctl);
 638                         free_excluded_extents(cache);
 639                         return 0;
 640                 }
 641         } else {
 642                 /*
 643                  * We're either using the free space tree or no caching at all.
 644                  * Set cached to the appropriate value and wakeup any waiters.
 645                  */
 646                 spin_lock(&cache->lock);
 647                 if (load_cache_only) {
 648                         cache->caching_ctl = NULL;
 649                         cache->cached = BTRFS_CACHE_NO;
 650                 } else {
 651                         cache->cached = BTRFS_CACHE_STARTED;
 652                         cache->has_caching_ctl = 1;
 653                 }
 654                 spin_unlock(&cache->lock);
 655                 wake_up(&caching_ctl->wait);
 656         }
 657
 658         if (load_cache_only) {
 659                 put_caching_control(caching_ctl);
 660                 return 0;
 661         }
 662
 663         down_write(&fs_info->commit_root_sem);
 664         refcount_inc(&caching_ctl->count);
 665         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 666         up_write(&fs_info->commit_root_sem);
 667
 668         btrfs_get_block_group(cache);
 669
 670         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 671
 672         return ret;
 673 }
 674
 675 /*
 676  * return the block group that starts at or after bytenr
 677  */
 678 static struct btrfs_block_group_cache *
 679 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 680 {
 681         return block_group_cache_tree_search(info, bytenr, 0);
 682 }
 683
 684 /*
 685  * return the block group that contains the given bytenr
 686  */
 687 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 688                                                  struct btrfs_fs_info *info,
 689                                                  u64 bytenr)
 690 {
 691         return block_group_cache_tree_search(info, bytenr, 1);
 692 }
 693
 694 static u64 generic_ref_to_space_flags(struct btrfs_ref *ref)
 695 {
 696         if (ref->type == BTRFS_REF_METADATA) {
 697                 if (ref->tree_ref.root == BTRFS_CHUNK_TREE_OBJECTID)
 698                         return BTRFS_BLOCK_GROUP_SYSTEM;
 699                 else
 700                         return BTRFS_BLOCK_GROUP_METADATA;
 701         }
 702         return BTRFS_BLOCK_GROUP_DATA;
 703 }
 704
 705 static void add_pinned_bytes(struct btrfs_fs_info *fs_info,
 706                              struct btrfs_ref *ref)
 707 {
 708         struct btrfs_space_info *space_info;
 709         u64 flags = generic_ref_to_space_flags(ref);
 710
 711         space_info = btrfs_find_space_info(fs_info, flags);
 712         ASSERT(space_info);
 713         percpu_counter_add_batch(&space_info->total_bytes_pinned, ref->len,
 714                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 715 }
 716
 717 static void sub_pinned_bytes(struct btrfs_fs_info *fs_info,
 718                              struct btrfs_ref *ref)
 719 {
 720         struct btrfs_space_info *space_info;
 721         u64 flags = generic_ref_to_space_flags(ref);
 722
 723         space_info = btrfs_find_space_info(fs_info, flags);
 724         ASSERT(space_info);
 725         percpu_counter_add_batch(&space_info->total_bytes_pinned, -ref->len,
 726                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 727 }
 728
 729 /* simple helper to search for an existing data extent at a given offset */
 730 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 731 {
 732         int ret;
 733         struct btrfs_key key;
 734         struct btrfs_path *path;
 735
 736         path = btrfs_alloc_path();
 737         if (!path)
 738                 return -ENOMEM;
 739
 740         key.objectid = start;
 741         key.offset = len;
 742         key.type = BTRFS_EXTENT_ITEM_KEY;
 743         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 744         btrfs_free_path(path);
 745         return ret;
 746 }
 747
 748 /*
 749  * helper function to lookup reference count and flags of a tree block.
 750  *
 751  * the head node for delayed ref is used to store the sum of all the
 752  * reference count modifications queued up in the rbtree. the head
 753  * node may also store the extent flags to set. This way you can check
 754  * to see what the reference count and extent flags would be if all of
 755  * the delayed refs are not processed.
 756  */
 757 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 758                              struct btrfs_fs_info *fs_info, u64 bytenr,
 759                              u64 offset, int metadata, u64 *refs, u64 *flags)
 760 {
 761         struct btrfs_delayed_ref_head *head;
 762         struct btrfs_delayed_ref_root *delayed_refs;
 763         struct btrfs_path *path;
 764         struct btrfs_extent_item *ei;
 765         struct extent_buffer *leaf;
 766         struct btrfs_key key;
 767         u32 item_size;
 768         u64 num_refs;
 769         u64 extent_flags;
 770         int ret;
 771
 772         /*
 773          * If we don't have skinny metadata, don't bother doing anything
 774          * different
 775          */
 776         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 777                 offset = fs_info->nodesize;
 778                 metadata = 0;
 779         }
 780
 781         path = btrfs_alloc_path();
 782         if (!path)
 783                 return -ENOMEM;
 784
 785         if (!trans) {
 786                 path->skip_locking = 1;
 787                 path->search_commit_root = 1;
 788         }
 789
 790 search_again:
 791         key.objectid = bytenr;
 792         key.offset = offset;
 793         if (metadata)
 794                 key.type = BTRFS_METADATA_ITEM_KEY;
 795         else
 796                 key.type = BTRFS_EXTENT_ITEM_KEY;
 797
 798         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 799         if (ret < 0)
 800                 goto out_free;
 801
 802         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 803                 if (path->slots[0]) {
 804                         path->slots[0]--;
 805                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 806                                               path->slots[0]);
 807                         if (key.objectid == bytenr &&
 808                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 809                             key.offset == fs_info->nodesize)
 810                                 ret = 0;
 811                 }
 812         }
 813
 814         if (ret == 0) {
 815                 leaf = path->nodes[0];
 816                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 817                 if (item_size >= sizeof(*ei)) {
 818                         ei = btrfs_item_ptr(leaf, path->slots[0],
 819                                             struct btrfs_extent_item);
 820                         num_refs = btrfs_extent_refs(leaf, ei);
 821                         extent_flags = btrfs_extent_flags(leaf, ei);
 822                 } else {
 823                         ret = -EINVAL;
 824                         btrfs_print_v0_err(fs_info);
 825                         if (trans)
 826                                 btrfs_abort_transaction(trans, ret);
 827                         else
 828                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 829
 830                         goto out_free;
 831                 }
 832
 833                 BUG_ON(num_refs == 0);
 834         } else {
 835                 num_refs = 0;
 836                 extent_flags = 0;
 837                 ret = 0;
 838         }
 839
 840         if (!trans)
 841                 goto out;
 842
 843         delayed_refs = &trans->transaction->delayed_refs;
 844         spin_lock(&delayed_refs->lock);
 845         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 846         if (head) {
 847                 if (!mutex_trylock(&head->mutex)) {
 848                         refcount_inc(&head->refs);
 849                         spin_unlock(&delayed_refs->lock);
 850
 851                         btrfs_release_path(path);
 852
 853                         /*
 854                          * Mutex was contended, block until it's released and try
 855                          * again
 856                          */
 857                         mutex_lock(&head->mutex);
 858                         mutex_unlock(&head->mutex);
 859                         btrfs_put_delayed_ref_head(head);
 860                         goto search_again;
 861                 }
 862                 spin_lock(&head->lock);
 863                 if (head->extent_op && head->extent_op->update_flags)
 864                         extent_flags |= head->extent_op->flags_to_set;
 865                 else
 866                         BUG_ON(num_refs == 0);
 867
 868                 num_refs += head->ref_mod;
 869                 spin_unlock(&head->lock);
 870                 mutex_unlock(&head->mutex);
 871         }
 872         spin_unlock(&delayed_refs->lock);
 873 out:
 874         WARN_ON(num_refs == 0);
 875         if (refs)
 876                 *refs = num_refs;
 877         if (flags)
 878                 *flags = extent_flags;
 879 out_free:
 880         btrfs_free_path(path);
 881         return ret;
 882 }
 883
 884 /*
 885  * Back reference rules.  Back refs have three main goals:
 886  *
 887  * 1) differentiate between all holders of references to an extent so that
 888  *    when a reference is dropped we can make sure it was a valid reference
 889  *    before freeing the extent.
 890  *
 891  * 2) Provide enough information to quickly find the holders of an extent
 892  *    if we notice a given block is corrupted or bad.
 893  *
 894  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 895  *    maintenance.  This is actually the same as #2, but with a slightly
 896  *    different use case.
 897  *
 898  * There are two kinds of back refs. The implicit back refs is optimized
 899  * for pointers in non-shared tree blocks. For a given pointer in a block,
 900  * back refs of this kind provide information about the block's owner tree
 901  * and the pointer's key. These information allow us to find the block by
 902  * b-tree searching. The full back refs is for pointers in tree blocks not
 903  * referenced by their owner trees. The location of tree block is recorded
 904  * in the back refs. Actually the full back refs is generic, and can be
 905  * used in all cases the implicit back refs is used. The major shortcoming
 906  * of the full back refs is its overhead. Every time a tree block gets
 907  * COWed, we have to update back refs entry for all pointers in it.
 908  *
 909  * For a newly allocated tree block, we use implicit back refs for
 910  * pointers in it. This means most tree related operations only involve
 911  * implicit back refs. For a tree block created in old transaction, the
 912  * only way to drop a reference to it is COW it. So we can detect the
 913  * event that tree block loses its owner tree's reference and do the
 914  * back refs conversion.
 915  *
 916  * When a tree block is COWed through a tree, there are four cases:
 917  *
 918  * The reference count of the block is one and the tree is the block's
 919  * owner tree. Nothing to do in this case.
 920  *
 921  * The reference count of the block is one and the tree is not the
 922  * block's owner tree. In this case, full back refs is used for pointers
 923  * in the block. Remove these full back refs, add implicit back refs for
 924  * every pointers in the new block.
 925  *
 926  * The reference count of the block is greater than one and the tree is
 927  * the block's owner tree. In this case, implicit back refs is used for
 928  * pointers in the block. Add full back refs for every pointers in the
 929  * block, increase lower level extents' reference counts. The original
 930  * implicit back refs are entailed to the new block.
 931  *
 932  * The reference count of the block is greater than one and the tree is
 933  * not the block's owner tree. Add implicit back refs for every pointer in
 934  * the new block, increase lower level extents' reference count.
 935  *
 936  * Back Reference Key composing:
 937  *
 938  * The key objectid corresponds to the first byte in the extent,
 939  * The key type is used to differentiate between types of back refs.
 940  * There are different meanings of the key offset for different types
 941  * of back refs.
 942  *
 943  * File extents can be referenced by:
 944  *
 945  * - multiple snapshots, subvolumes, or different generations in one subvol
 946  * - different files inside a single subvolume
 947  * - different offsets inside a file (bookend extents in file.c)
 948  *
 949  * The extent ref structure for the implicit back refs has fields for:
 950  *
 951  * - Objectid of the subvolume root
 952  * - objectid of the file holding the reference
 953  * - original offset in the file
 954  * - how many bookend extents
 955  *
 956  * The key offset for the implicit back refs is hash of the first
 957  * three fields.
 958  *
 959  * The extent ref structure for the full back refs has field for:
 960  *
 961  * - number of pointers in the tree leaf
 962  *
 963  * The key offset for the implicit back refs is the first byte of
 964  * the tree leaf
 965  *
 966  * When a file extent is allocated, The implicit back refs is used.
 967  * the fields are filled in:
 968  *
 969  *     (root_key.objectid, inode objectid, offset in file, 1)
 970  *
 971  * When a file extent is removed file truncation, we find the
 972  * corresponding implicit back refs and check the following fields:
 973  *
 974  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 975  *
 976  * Btree extents can be referenced by:
 977  *
 978  * - Different subvolumes
 979  *
 980  * Both the implicit back refs and the full back refs for tree blocks
 981  * only consist of key. The key offset for the implicit back refs is
 982  * objectid of block's owner tree. The key offset for the full back refs
 983  * is the first byte of parent block.
 984  *
 985  * When implicit back refs is used, information about the lowest key and
 986  * level of the tree block are required. These information are stored in
 987  * tree block info structure.
 988  */
 989
 990 /*
 991  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 992  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 993  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 994  */
 995 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 996                                      struct btrfs_extent_inline_ref *iref,
 997                                      enum btrfs_inline_ref_type is_data)
 998 {
 999         int type = btrfs_extent_inline_ref_type(eb, iref);
1000         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1001
1002         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1003             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1004             type == BTRFS_SHARED_DATA_REF_KEY ||
1005             type == BTRFS_EXTENT_DATA_REF_KEY) {
1006                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1007                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1008                                 return type;
1009                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1010                                 ASSERT(eb->fs_info);
1011                                 /*
1012                                  * Every shared one has parent tree
1013                                  * block, which must be aligned to
1014                                  * nodesize.
1015                                  */
1016                                 if (offset &&
1017                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1018                                         return type;
1019                         }
1020                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1021                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1022                                 return type;
1023                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1024                                 ASSERT(eb->fs_info);
1025                                 /*
1026                                  * Every shared one has parent tree
1027                                  * block, which must be aligned to
1028                                  * nodesize.
1029                                  */
1030                                 if (offset &&
1031                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1032                                         return type;
1033                         }
1034                 } else {
1035                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1036                         return type;
1037                 }
1038         }
1039
1040         btrfs_print_leaf((struct extent_buffer *)eb);
1041         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1042                   eb->start, type);
1043         WARN_ON(1);
1044
1045         return BTRFS_REF_TYPE_INVALID;
1046 }
1047
1048 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1049 {
1050         u32 high_crc = ~(u32)0;
1051         u32 low_crc = ~(u32)0;
1052         __le64 lenum;
1053
1054         lenum = cpu_to_le64(root_objectid);
1055         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1056         lenum = cpu_to_le64(owner);
1057         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1058         lenum = cpu_to_le64(offset);
1059         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1060
1061         return ((u64)high_crc << 31) ^ (u64)low_crc;
1062 }
1063
1064 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1065                                      struct btrfs_extent_data_ref *ref)
1066 {
1067         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1068                                     btrfs_extent_data_ref_objectid(leaf, ref),
1069                                     btrfs_extent_data_ref_offset(leaf, ref));
1070 }
1071
1072 static int match_extent_data_ref(struct extent_buffer *leaf,
1073                                  struct btrfs_extent_data_ref *ref,
1074                                  u64 root_objectid, u64 owner, u64 offset)
1075 {
1076         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1077             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1078             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1079                 return 0;
1080         return 1;
1081 }
1082
1083 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1084                                            struct btrfs_path *path,
1085                                            u64 bytenr, u64 parent,
1086                                            u64 root_objectid,
1087                                            u64 owner, u64 offset)
1088 {
1089         struct btrfs_root *root = trans->fs_info->extent_root;
1090         struct btrfs_key key;
1091         struct btrfs_extent_data_ref *ref;
1092         struct extent_buffer *leaf;
1093         u32 nritems;
1094         int ret;
1095         int recow;
1096         int err = -ENOENT;
1097
1098         key.objectid = bytenr;
1099         if (parent) {
1100                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1101                 key.offset = parent;
1102         } else {
1103                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1104                 key.offset = hash_extent_data_ref(root_objectid,
1105                                                   owner, offset);
1106         }
1107 again:
1108         recow = 0;
1109         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1110         if (ret < 0) {
1111                 err = ret;
1112                 goto fail;
1113         }
1114
1115         if (parent) {
1116                 if (!ret)
1117                         return 0;
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_path *path,
1161                                            u64 bytenr, u64 parent,
1162                                            u64 root_objectid, u64 owner,
1163                                            u64 offset, int refs_to_add)
1164 {
1165         struct btrfs_root *root = trans->fs_info->extent_root;
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_path *path,
1240                                            int refs_to_drop, int *last_ref)
1241 {
1242         struct btrfs_key key;
1243         struct btrfs_extent_data_ref *ref1 = NULL;
1244         struct btrfs_shared_data_ref *ref2 = NULL;
1245         struct extent_buffer *leaf;
1246         u32 num_refs = 0;
1247         int ret = 0;
1248
1249         leaf = path->nodes[0];
1250         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1251
1252         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1253                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1254                                       struct btrfs_extent_data_ref);
1255                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1256         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1257                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_shared_data_ref);
1259                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1260         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1261                 btrfs_print_v0_err(trans->fs_info);
1262                 btrfs_abort_transaction(trans, -EINVAL);
1263                 return -EINVAL;
1264         } else {
1265                 BUG();
1266         }
1267
1268         BUG_ON(num_refs < refs_to_drop);
1269         num_refs -= refs_to_drop;
1270
1271         if (num_refs == 0) {
1272                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1273                 *last_ref = 1;
1274         } else {
1275                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1276                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1277                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1278                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1279                 btrfs_mark_buffer_dirty(leaf);
1280         }
1281         return ret;
1282 }
1283
1284 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1285                                           struct btrfs_extent_inline_ref *iref)
1286 {
1287         struct btrfs_key key;
1288         struct extent_buffer *leaf;
1289         struct btrfs_extent_data_ref *ref1;
1290         struct btrfs_shared_data_ref *ref2;
1291         u32 num_refs = 0;
1292         int type;
1293
1294         leaf = path->nodes[0];
1295         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1296
1297         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1298         if (iref) {
1299                 /*
1300                  * If type is invalid, we should have bailed out earlier than
1301                  * this call.
1302                  */
1303                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1304                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1305                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1306                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1307                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1308                 } else {
1309                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1310                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1311                 }
1312         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1313                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1314                                       struct btrfs_extent_data_ref);
1315                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1316         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1317                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_shared_data_ref);
1319                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1320         } else {
1321                 WARN_ON(1);
1322         }
1323         return num_refs;
1324 }
1325
1326 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1327                                           struct btrfs_path *path,
1328                                           u64 bytenr, u64 parent,
1329                                           u64 root_objectid)
1330 {
1331         struct btrfs_root *root = trans->fs_info->extent_root;
1332         struct btrfs_key key;
1333         int ret;
1334
1335         key.objectid = bytenr;
1336         if (parent) {
1337                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1338                 key.offset = parent;
1339         } else {
1340                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1341                 key.offset = root_objectid;
1342         }
1343
1344         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1345         if (ret > 0)
1346                 ret = -ENOENT;
1347         return ret;
1348 }
1349
1350 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1351                                           struct btrfs_path *path,
1352                                           u64 bytenr, u64 parent,
1353                                           u64 root_objectid)
1354 {
1355         struct btrfs_key key;
1356         int ret;
1357
1358         key.objectid = bytenr;
1359         if (parent) {
1360                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1361                 key.offset = parent;
1362         } else {
1363                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1364                 key.offset = root_objectid;
1365         }
1366
1367         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1368                                       path, &key, 0);
1369         btrfs_release_path(path);
1370         return ret;
1371 }
1372
1373 static inline int extent_ref_type(u64 parent, u64 owner)
1374 {
1375         int type;
1376         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1377                 if (parent > 0)
1378                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1379                 else
1380                         type = BTRFS_TREE_BLOCK_REF_KEY;
1381         } else {
1382                 if (parent > 0)
1383                         type = BTRFS_SHARED_DATA_REF_KEY;
1384                 else
1385                         type = BTRFS_EXTENT_DATA_REF_KEY;
1386         }
1387         return type;
1388 }
1389
1390 static int find_next_key(struct btrfs_path *path, int level,
1391                          struct btrfs_key *key)
1392
1393 {
1394         for (; level < BTRFS_MAX_LEVEL; level++) {
1395                 if (!path->nodes[level])
1396                         break;
1397                 if (path->slots[level] + 1 >=
1398                     btrfs_header_nritems(path->nodes[level]))
1399                         continue;
1400                 if (level == 0)
1401                         btrfs_item_key_to_cpu(path->nodes[level], key,
1402                                               path->slots[level] + 1);
1403                 else
1404                         btrfs_node_key_to_cpu(path->nodes[level], key,
1405                                               path->slots[level] + 1);
1406                 return 0;
1407         }
1408         return 1;
1409 }
1410
1411 /*
1412  * look for inline back ref. if back ref is found, *ref_ret is set
1413  * to the address of inline back ref, and 0 is returned.
1414  *
1415  * if back ref isn't found, *ref_ret is set to the address where it
1416  * should be inserted, and -ENOENT is returned.
1417  *
1418  * if insert is true and there are too many inline back refs, the path
1419  * points to the extent item, and -EAGAIN is returned.
1420  *
1421  * NOTE: inline back refs are ordered in the same way that back ref
1422  *       items in the tree are ordered.
1423  */
1424 static noinline_for_stack
1425 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1426                                  struct btrfs_path *path,
1427                                  struct btrfs_extent_inline_ref **ref_ret,
1428                                  u64 bytenr, u64 num_bytes,
1429                                  u64 parent, u64 root_objectid,
1430                                  u64 owner, u64 offset, int insert)
1431 {
1432         struct btrfs_fs_info *fs_info = trans->fs_info;
1433         struct btrfs_root *root = fs_info->extent_root;
1434         struct btrfs_key key;
1435         struct extent_buffer *leaf;
1436         struct btrfs_extent_item *ei;
1437         struct btrfs_extent_inline_ref *iref;
1438         u64 flags;
1439         u64 item_size;
1440         unsigned long ptr;
1441         unsigned long end;
1442         int extra_size;
1443         int type;
1444         int want;
1445         int ret;
1446         int err = 0;
1447         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1448         int needed;
1449
1450         key.objectid = bytenr;
1451         key.type = BTRFS_EXTENT_ITEM_KEY;
1452         key.offset = num_bytes;
1453
1454         want = extent_ref_type(parent, owner);
1455         if (insert) {
1456                 extra_size = btrfs_extent_inline_ref_size(want);
1457                 path->keep_locks = 1;
1458         } else
1459                 extra_size = -1;
1460
1461         /*
1462          * Owner is our level, so we can just add one to get the level for the
1463          * block we are interested in.
1464          */
1465         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1466                 key.type = BTRFS_METADATA_ITEM_KEY;
1467                 key.offset = owner;
1468         }
1469
1470 again:
1471         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1472         if (ret < 0) {
1473                 err = ret;
1474                 goto out;
1475         }
1476
1477         /*
1478          * We may be a newly converted file system which still has the old fat
1479          * extent entries for metadata, so try and see if we have one of those.
1480          */
1481         if (ret > 0 && skinny_metadata) {
1482                 skinny_metadata = false;
1483                 if (path->slots[0]) {
1484                         path->slots[0]--;
1485                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1486                                               path->slots[0]);
1487                         if (key.objectid == bytenr &&
1488                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1489                             key.offset == num_bytes)
1490                                 ret = 0;
1491                 }
1492                 if (ret) {
1493                         key.objectid = bytenr;
1494                         key.type = BTRFS_EXTENT_ITEM_KEY;
1495                         key.offset = num_bytes;
1496                         btrfs_release_path(path);
1497                         goto again;
1498                 }
1499         }
1500
1501         if (ret && !insert) {
1502                 err = -ENOENT;
1503                 goto out;
1504         } else if (WARN_ON(ret)) {
1505                 err = -EIO;
1506                 goto out;
1507         }
1508
1509         leaf = path->nodes[0];
1510         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1511         if (unlikely(item_size < sizeof(*ei))) {
1512                 err = -EINVAL;
1513                 btrfs_print_v0_err(fs_info);
1514                 btrfs_abort_transaction(trans, err);
1515                 goto out;
1516         }
1517
1518         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1519         flags = btrfs_extent_flags(leaf, ei);
1520
1521         ptr = (unsigned long)(ei + 1);
1522         end = (unsigned long)ei + item_size;
1523
1524         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1525                 ptr += sizeof(struct btrfs_tree_block_info);
1526                 BUG_ON(ptr > end);
1527         }
1528
1529         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1530                 needed = BTRFS_REF_TYPE_DATA;
1531         else
1532                 needed = BTRFS_REF_TYPE_BLOCK;
1533
1534         err = -ENOENT;
1535         while (1) {
1536                 if (ptr >= end) {
1537                         WARN_ON(ptr > end);
1538                         break;
1539                 }
1540                 iref = (struct btrfs_extent_inline_ref *)ptr;
1541                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1542                 if (type == BTRFS_REF_TYPE_INVALID) {
1543                         err = -EUCLEAN;
1544                         goto out;
1545                 }
1546
1547                 if (want < type)
1548                         break;
1549                 if (want > type) {
1550                         ptr += btrfs_extent_inline_ref_size(type);
1551                         continue;
1552                 }
1553
1554                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1555                         struct btrfs_extent_data_ref *dref;
1556                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1557                         if (match_extent_data_ref(leaf, dref, root_objectid,
1558                                                   owner, offset)) {
1559                                 err = 0;
1560                                 break;
1561                         }
1562                         if (hash_extent_data_ref_item(leaf, dref) <
1563                             hash_extent_data_ref(root_objectid, owner, offset))
1564                                 break;
1565                 } else {
1566                         u64 ref_offset;
1567                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1568                         if (parent > 0) {
1569                                 if (parent == ref_offset) {
1570                                         err = 0;
1571                                         break;
1572                                 }
1573                                 if (ref_offset < parent)
1574                                         break;
1575                         } else {
1576                                 if (root_objectid == ref_offset) {
1577                                         err = 0;
1578                                         break;
1579                                 }
1580                                 if (ref_offset < root_objectid)
1581                                         break;
1582                         }
1583                 }
1584                 ptr += btrfs_extent_inline_ref_size(type);
1585         }
1586         if (err == -ENOENT && insert) {
1587                 if (item_size + extra_size >=
1588                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1589                         err = -EAGAIN;
1590                         goto out;
1591                 }
1592                 /*
1593                  * To add new inline back ref, we have to make sure
1594                  * there is no corresponding back ref item.
1595                  * For simplicity, we just do not add new inline back
1596                  * ref if there is any kind of item for this block
1597                  */
1598                 if (find_next_key(path, 0, &key) == 0 &&
1599                     key.objectid == bytenr &&
1600                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1601                         err = -EAGAIN;
1602                         goto out;
1603                 }
1604         }
1605         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1606 out:
1607         if (insert) {
1608                 path->keep_locks = 0;
1609                 btrfs_unlock_up_safe(path, 1);
1610         }
1611         return err;
1612 }
1613
1614 /*
1615  * helper to add new inline back ref
1616  */
1617 static noinline_for_stack
1618 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1619                                  struct btrfs_path *path,
1620                                  struct btrfs_extent_inline_ref *iref,
1621                                  u64 parent, u64 root_objectid,
1622                                  u64 owner, u64 offset, int refs_to_add,
1623                                  struct btrfs_delayed_extent_op *extent_op)
1624 {
1625         struct extent_buffer *leaf;
1626         struct btrfs_extent_item *ei;
1627         unsigned long ptr;
1628         unsigned long end;
1629         unsigned long item_offset;
1630         u64 refs;
1631         int size;
1632         int type;
1633
1634         leaf = path->nodes[0];
1635         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1636         item_offset = (unsigned long)iref - (unsigned long)ei;
1637
1638         type = extent_ref_type(parent, owner);
1639         size = btrfs_extent_inline_ref_size(type);
1640
1641         btrfs_extend_item(path, size);
1642
1643         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1644         refs = btrfs_extent_refs(leaf, ei);
1645         refs += refs_to_add;
1646         btrfs_set_extent_refs(leaf, ei, refs);
1647         if (extent_op)
1648                 __run_delayed_extent_op(extent_op, leaf, ei);
1649
1650         ptr = (unsigned long)ei + item_offset;
1651         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1652         if (ptr < end - size)
1653                 memmove_extent_buffer(leaf, ptr + size, ptr,
1654                                       end - size - ptr);
1655
1656         iref = (struct btrfs_extent_inline_ref *)ptr;
1657         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1658         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1659                 struct btrfs_extent_data_ref *dref;
1660                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1661                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1662                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1663                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1664                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1665         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1666                 struct btrfs_shared_data_ref *sref;
1667                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1668                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1669                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1670         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1671                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1672         } else {
1673                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1674         }
1675         btrfs_mark_buffer_dirty(leaf);
1676 }
1677
1678 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1679                                  struct btrfs_path *path,
1680                                  struct btrfs_extent_inline_ref **ref_ret,
1681                                  u64 bytenr, u64 num_bytes, u64 parent,
1682                                  u64 root_objectid, u64 owner, u64 offset)
1683 {
1684         int ret;
1685
1686         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1687                                            num_bytes, parent, root_objectid,
1688                                            owner, offset, 0);
1689         if (ret != -ENOENT)
1690                 return ret;
1691
1692         btrfs_release_path(path);
1693         *ref_ret = NULL;
1694
1695         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1696                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1697                                             root_objectid);
1698         } else {
1699                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1700                                              root_objectid, owner, offset);
1701         }
1702         return ret;
1703 }
1704
1705 /*
1706  * helper to update/remove inline back ref
1707  */
1708 static noinline_for_stack
1709 void update_inline_extent_backref(struct btrfs_path *path,
1710                                   struct btrfs_extent_inline_ref *iref,
1711                                   int refs_to_mod,
1712                                   struct btrfs_delayed_extent_op *extent_op,
1713                                   int *last_ref)
1714 {
1715         struct extent_buffer *leaf = path->nodes[0];
1716         struct btrfs_extent_item *ei;
1717         struct btrfs_extent_data_ref *dref = NULL;
1718         struct btrfs_shared_data_ref *sref = NULL;
1719         unsigned long ptr;
1720         unsigned long end;
1721         u32 item_size;
1722         int size;
1723         int type;
1724         u64 refs;
1725
1726         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1727         refs = btrfs_extent_refs(leaf, ei);
1728         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1729         refs += refs_to_mod;
1730         btrfs_set_extent_refs(leaf, ei, refs);
1731         if (extent_op)
1732                 __run_delayed_extent_op(extent_op, leaf, ei);
1733
1734         /*
1735          * If type is invalid, we should have bailed out after
1736          * lookup_inline_extent_backref().
1737          */
1738         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1739         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1740
1741         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1742                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1743                 refs = btrfs_extent_data_ref_count(leaf, dref);
1744         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1745                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1746                 refs = btrfs_shared_data_ref_count(leaf, sref);
1747         } else {
1748                 refs = 1;
1749                 BUG_ON(refs_to_mod != -1);
1750         }
1751
1752         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1753         refs += refs_to_mod;
1754
1755         if (refs > 0) {
1756                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1757                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1758                 else
1759                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1760         } else {
1761                 *last_ref = 1;
1762                 size =  btrfs_extent_inline_ref_size(type);
1763                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1764                 ptr = (unsigned long)iref;
1765                 end = (unsigned long)ei + item_size;
1766                 if (ptr + size < end)
1767                         memmove_extent_buffer(leaf, ptr, ptr + size,
1768                                               end - ptr - size);
1769                 item_size -= size;
1770                 btrfs_truncate_item(path, item_size, 1);
1771         }
1772         btrfs_mark_buffer_dirty(leaf);
1773 }
1774
1775 static noinline_for_stack
1776 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1777                                  struct btrfs_path *path,
1778                                  u64 bytenr, u64 num_bytes, u64 parent,
1779                                  u64 root_objectid, u64 owner,
1780                                  u64 offset, int refs_to_add,
1781                                  struct btrfs_delayed_extent_op *extent_op)
1782 {
1783         struct btrfs_extent_inline_ref *iref;
1784         int ret;
1785
1786         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1787                                            num_bytes, parent, root_objectid,
1788                                            owner, offset, 1);
1789         if (ret == 0) {
1790                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1791                 update_inline_extent_backref(path, iref, refs_to_add,
1792                                              extent_op, NULL);
1793         } else if (ret == -ENOENT) {
1794                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1795                                             root_objectid, owner, offset,
1796                                             refs_to_add, extent_op);
1797                 ret = 0;
1798         }
1799         return ret;
1800 }
1801
1802 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_path *path,
1804                                  u64 bytenr, u64 parent, u64 root_objectid,
1805                                  u64 owner, u64 offset, int refs_to_add)
1806 {
1807         int ret;
1808         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1809                 BUG_ON(refs_to_add != 1);
1810                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1811                                             root_objectid);
1812         } else {
1813                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1814                                              root_objectid, owner, offset,
1815                                              refs_to_add);
1816         }
1817         return ret;
1818 }
1819
1820 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1821                                  struct btrfs_path *path,
1822                                  struct btrfs_extent_inline_ref *iref,
1823                                  int refs_to_drop, int is_data, int *last_ref)
1824 {
1825         int ret = 0;
1826
1827         BUG_ON(!is_data && refs_to_drop != 1);
1828         if (iref) {
1829                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1830                                              last_ref);
1831         } else if (is_data) {
1832                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1833                                              last_ref);
1834         } else {
1835                 *last_ref = 1;
1836                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1837         }
1838         return ret;
1839 }
1840
1841 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1842                                u64 *discarded_bytes)
1843 {
1844         int j, ret = 0;
1845         u64 bytes_left, end;
1846         u64 aligned_start = ALIGN(start, 1 << 9);
1847
1848         if (WARN_ON(start != aligned_start)) {
1849                 len -= aligned_start - start;
1850                 len = round_down(len, 1 << 9);
1851                 start = aligned_start;
1852         }
1853
1854         *discarded_bytes = 0;
1855
1856         if (!len)
1857                 return 0;
1858
1859         end = start + len;
1860         bytes_left = len;
1861
1862         /* Skip any superblocks on this device. */
1863         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1864                 u64 sb_start = btrfs_sb_offset(j);
1865                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1866                 u64 size = sb_start - start;
1867
1868                 if (!in_range(sb_start, start, bytes_left) &&
1869                     !in_range(sb_end, start, bytes_left) &&
1870                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1871                         continue;
1872
1873                 /*
1874                  * Superblock spans beginning of range.  Adjust start and
1875                  * try again.
1876                  */
1877                 if (sb_start <= start) {
1878                         start += sb_end - start;
1879                         if (start > end) {
1880                                 bytes_left = 0;
1881                                 break;
1882                         }
1883                         bytes_left = end - start;
1884                         continue;
1885                 }
1886
1887                 if (size) {
1888                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1889                                                    GFP_NOFS, 0);
1890                         if (!ret)
1891                                 *discarded_bytes += size;
1892                         else if (ret != -EOPNOTSUPP)
1893                                 return ret;
1894                 }
1895
1896                 start = sb_end;
1897                 if (start > end) {
1898                         bytes_left = 0;
1899                         break;
1900                 }
1901                 bytes_left = end - start;
1902         }
1903
1904         if (bytes_left) {
1905                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1906                                            GFP_NOFS, 0);
1907                 if (!ret)
1908                         *discarded_bytes += bytes_left;
1909         }
1910         return ret;
1911 }
1912
1913 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1914                          u64 num_bytes, u64 *actual_bytes)
1915 {
1916         int ret;
1917         u64 discarded_bytes = 0;
1918         struct btrfs_bio *bbio = NULL;
1919
1920
1921         /*
1922          * Avoid races with device replace and make sure our bbio has devices
1923          * associated to its stripes that don't go away while we are discarding.
1924          */
1925         btrfs_bio_counter_inc_blocked(fs_info);
1926         /* Tell the block device(s) that the sectors can be discarded */
1927         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1928                               &bbio, 0);
1929         /* Error condition is -ENOMEM */
1930         if (!ret) {
1931                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1932                 int i;
1933
1934
1935                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1936                         u64 bytes;
1937                         struct request_queue *req_q;
1938
1939                         if (!stripe->dev->bdev) {
1940                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1941                                 continue;
1942                         }
1943                         req_q = bdev_get_queue(stripe->dev->bdev);
1944                         if (!blk_queue_discard(req_q))
1945                                 continue;
1946
1947                         ret = btrfs_issue_discard(stripe->dev->bdev,
1948                                                   stripe->physical,
1949                                                   stripe->length,
1950                                                   &bytes);
1951                         if (!ret)
1952                                 discarded_bytes += bytes;
1953                         else if (ret != -EOPNOTSUPP)
1954                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1955
1956                         /*
1957                          * Just in case we get back EOPNOTSUPP for some reason,
1958                          * just ignore the return value so we don't screw up
1959                          * people calling discard_extent.
1960                          */
1961                         ret = 0;
1962                 }
1963                 btrfs_put_bbio(bbio);
1964         }
1965         btrfs_bio_counter_dec(fs_info);
1966
1967         if (actual_bytes)
1968                 *actual_bytes = discarded_bytes;
1969
1970
1971         if (ret == -EOPNOTSUPP)
1972                 ret = 0;
1973         return ret;
1974 }
1975
1976 /* Can return -ENOMEM */
1977 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1978                          struct btrfs_ref *generic_ref)
1979 {
1980         struct btrfs_fs_info *fs_info = trans->fs_info;
1981         int old_ref_mod, new_ref_mod;
1982         int ret;
1983
1984         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1985                generic_ref->action);
1986         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1987                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
1988
1989         if (generic_ref->type == BTRFS_REF_METADATA)
1990                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref,
1991                                 NULL, &old_ref_mod, &new_ref_mod);
1992         else
1993                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0,
1994                                                  &old_ref_mod, &new_ref_mod);
1995
1996         btrfs_ref_tree_mod(fs_info, generic_ref);
1997
1998         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
1999                 sub_pinned_bytes(fs_info, generic_ref);
2000
2001         return ret;
2002 }
2003
2004 /*
2005  * __btrfs_inc_extent_ref - insert backreference for a given extent
2006  *
2007  * @trans:          Handle of transaction
2008  *
2009  * @node:           The delayed ref node used to get the bytenr/length for
2010  *                  extent whose references are incremented.
2011  *
2012  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2013  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2014  *                  bytenr of the parent block. Since new extents are always
2015  *                  created with indirect references, this will only be the case
2016  *                  when relocating a shared extent. In that case, root_objectid
2017  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2018  *                  be 0
2019  *
2020  * @root_objectid:  The id of the root where this modification has originated,
2021  *                  this can be either one of the well-known metadata trees or
2022  *                  the subvolume id which references this extent.
2023  *
2024  * @owner:          For data extents it is the inode number of the owning file.
2025  *                  For metadata extents this parameter holds the level in the
2026  *                  tree of the extent.
2027  *
2028  * @offset:         For metadata extents the offset is ignored and is currently
2029  *                  always passed as 0. For data extents it is the fileoffset
2030  *                  this extent belongs to.
2031  *
2032  * @refs_to_add     Number of references to add
2033  *
2034  * @extent_op       Pointer to a structure, holding information necessary when
2035  *                  updating a tree block's flags
2036  *
2037  */
2038 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2039                                   struct btrfs_delayed_ref_node *node,
2040                                   u64 parent, u64 root_objectid,
2041                                   u64 owner, u64 offset, int refs_to_add,
2042                                   struct btrfs_delayed_extent_op *extent_op)
2043 {
2044         struct btrfs_path *path;
2045         struct extent_buffer *leaf;
2046         struct btrfs_extent_item *item;
2047         struct btrfs_key key;
2048         u64 bytenr = node->bytenr;
2049         u64 num_bytes = node->num_bytes;
2050         u64 refs;
2051         int ret;
2052
2053         path = btrfs_alloc_path();
2054         if (!path)
2055                 return -ENOMEM;
2056
2057         path->reada = READA_FORWARD;
2058         path->leave_spinning = 1;
2059         /* this will setup the path even if it fails to insert the back ref */
2060         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2061                                            parent, root_objectid, owner,
2062                                            offset, refs_to_add, extent_op);
2063         if ((ret < 0 && ret != -EAGAIN) || !ret)
2064                 goto out;
2065
2066         /*
2067          * Ok we had -EAGAIN which means we didn't have space to insert and
2068          * inline extent ref, so just update the reference count and add a
2069          * normal backref.
2070          */
2071         leaf = path->nodes[0];
2072         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2073         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2074         refs = btrfs_extent_refs(leaf, item);
2075         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2076         if (extent_op)
2077                 __run_delayed_extent_op(extent_op, leaf, item);
2078
2079         btrfs_mark_buffer_dirty(leaf);
2080         btrfs_release_path(path);
2081
2082         path->reada = READA_FORWARD;
2083         path->leave_spinning = 1;
2084         /* now insert the actual backref */
2085         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2086                                     owner, offset, refs_to_add);
2087         if (ret)
2088                 btrfs_abort_transaction(trans, ret);
2089 out:
2090         btrfs_free_path(path);
2091         return ret;
2092 }
2093
2094 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2095                                 struct btrfs_delayed_ref_node *node,
2096                                 struct btrfs_delayed_extent_op *extent_op,
2097                                 int insert_reserved)
2098 {
2099         int ret = 0;
2100         struct btrfs_delayed_data_ref *ref;
2101         struct btrfs_key ins;
2102         u64 parent = 0;
2103         u64 ref_root = 0;
2104         u64 flags = 0;
2105
2106         ins.objectid = node->bytenr;
2107         ins.offset = node->num_bytes;
2108         ins.type = BTRFS_EXTENT_ITEM_KEY;
2109
2110         ref = btrfs_delayed_node_to_data_ref(node);
2111         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2112
2113         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2114                 parent = ref->parent;
2115         ref_root = ref->root;
2116
2117         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2118                 if (extent_op)
2119                         flags |= extent_op->flags_to_set;
2120                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2121                                                  flags, ref->objectid,
2122                                                  ref->offset, &ins,
2123                                                  node->ref_mod);
2124         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2125                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2126                                              ref->objectid, ref->offset,
2127                                              node->ref_mod, extent_op);
2128         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2129                 ret = __btrfs_free_extent(trans, node, parent,
2130                                           ref_root, ref->objectid,
2131                                           ref->offset, node->ref_mod,
2132                                           extent_op);
2133         } else {
2134                 BUG();
2135         }
2136         return ret;
2137 }
2138
2139 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2140                                     struct extent_buffer *leaf,
2141                                     struct btrfs_extent_item *ei)
2142 {
2143         u64 flags = btrfs_extent_flags(leaf, ei);
2144         if (extent_op->update_flags) {
2145                 flags |= extent_op->flags_to_set;
2146                 btrfs_set_extent_flags(leaf, ei, flags);
2147         }
2148
2149         if (extent_op->update_key) {
2150                 struct btrfs_tree_block_info *bi;
2151                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2152                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2153                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2154         }
2155 }
2156
2157 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2158                                  struct btrfs_delayed_ref_head *head,
2159                                  struct btrfs_delayed_extent_op *extent_op)
2160 {
2161         struct btrfs_fs_info *fs_info = trans->fs_info;
2162         struct btrfs_key key;
2163         struct btrfs_path *path;
2164         struct btrfs_extent_item *ei;
2165         struct extent_buffer *leaf;
2166         u32 item_size;
2167         int ret;
2168         int err = 0;
2169         int metadata = !extent_op->is_data;
2170
2171         if (trans->aborted)
2172                 return 0;
2173
2174         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2175                 metadata = 0;
2176
2177         path = btrfs_alloc_path();
2178         if (!path)
2179                 return -ENOMEM;
2180
2181         key.objectid = head->bytenr;
2182
2183         if (metadata) {
2184                 key.type = BTRFS_METADATA_ITEM_KEY;
2185                 key.offset = extent_op->level;
2186         } else {
2187                 key.type = BTRFS_EXTENT_ITEM_KEY;
2188                 key.offset = head->num_bytes;
2189         }
2190
2191 again:
2192         path->reada = READA_FORWARD;
2193         path->leave_spinning = 1;
2194         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2195         if (ret < 0) {
2196                 err = ret;
2197                 goto out;
2198         }
2199         if (ret > 0) {
2200                 if (metadata) {
2201                         if (path->slots[0] > 0) {
2202                                 path->slots[0]--;
2203                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2204                                                       path->slots[0]);
2205                                 if (key.objectid == head->bytenr &&
2206                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2207                                     key.offset == head->num_bytes)
2208                                         ret = 0;
2209                         }
2210                         if (ret > 0) {
2211                                 btrfs_release_path(path);
2212                                 metadata = 0;
2213
2214                                 key.objectid = head->bytenr;
2215                                 key.offset = head->num_bytes;
2216                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2217                                 goto again;
2218                         }
2219                 } else {
2220                         err = -EIO;
2221                         goto out;
2222                 }
2223         }
2224
2225         leaf = path->nodes[0];
2226         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2227
2228         if (unlikely(item_size < sizeof(*ei))) {
2229                 err = -EINVAL;
2230                 btrfs_print_v0_err(fs_info);
2231                 btrfs_abort_transaction(trans, err);
2232                 goto out;
2233         }
2234
2235         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2236         __run_delayed_extent_op(extent_op, leaf, ei);
2237
2238         btrfs_mark_buffer_dirty(leaf);
2239 out:
2240         btrfs_free_path(path);
2241         return err;
2242 }
2243
2244 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2245                                 struct btrfs_delayed_ref_node *node,
2246                                 struct btrfs_delayed_extent_op *extent_op,
2247                                 int insert_reserved)
2248 {
2249         int ret = 0;
2250         struct btrfs_delayed_tree_ref *ref;
2251         u64 parent = 0;
2252         u64 ref_root = 0;
2253
2254         ref = btrfs_delayed_node_to_tree_ref(node);
2255         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2256
2257         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2258                 parent = ref->parent;
2259         ref_root = ref->root;
2260
2261         if (node->ref_mod != 1) {
2262                 btrfs_err(trans->fs_info,
2263         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2264                           node->bytenr, node->ref_mod, node->action, ref_root,
2265                           parent);
2266                 return -EIO;
2267         }
2268         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2269                 BUG_ON(!extent_op || !extent_op->update_flags);
2270                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2271         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2272                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2273                                              ref->level, 0, 1, extent_op);
2274         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2275                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2276                                           ref->level, 0, 1, extent_op);
2277         } else {
2278                 BUG();
2279         }
2280         return ret;
2281 }
2282
2283 /* helper function to actually process a single delayed ref entry */
2284 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2285                                struct btrfs_delayed_ref_node *node,
2286                                struct btrfs_delayed_extent_op *extent_op,
2287                                int insert_reserved)
2288 {
2289         int ret = 0;
2290
2291         if (trans->aborted) {
2292                 if (insert_reserved)
2293                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2294                                          node->num_bytes, 1);
2295                 return 0;
2296         }
2297
2298         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2299             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2300                 ret = run_delayed_tree_ref(trans, node, extent_op,
2301                                            insert_reserved);
2302         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2303                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2304                 ret = run_delayed_data_ref(trans, node, extent_op,
2305                                            insert_reserved);
2306         else
2307                 BUG();
2308         if (ret && insert_reserved)
2309                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2310                                  node->num_bytes, 1);
2311         return ret;
2312 }
2313
2314 static inline struct btrfs_delayed_ref_node *
2315 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2316 {
2317         struct btrfs_delayed_ref_node *ref;
2318
2319         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2320                 return NULL;
2321
2322         /*
2323          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2324          * This is to prevent a ref count from going down to zero, which deletes
2325          * the extent item from the extent tree, when there still are references
2326          * to add, which would fail because they would not find the extent item.
2327          */
2328         if (!list_empty(&head->ref_add_list))
2329                 return list_first_entry(&head->ref_add_list,
2330                                 struct btrfs_delayed_ref_node, add_list);
2331
2332         ref = rb_entry(rb_first_cached(&head->ref_tree),
2333                        struct btrfs_delayed_ref_node, ref_node);
2334         ASSERT(list_empty(&ref->add_list));
2335         return ref;
2336 }
2337
2338 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2339                                       struct btrfs_delayed_ref_head *head)
2340 {
2341         spin_lock(&delayed_refs->lock);
2342         head->processing = 0;
2343         delayed_refs->num_heads_ready++;
2344         spin_unlock(&delayed_refs->lock);
2345         btrfs_delayed_ref_unlock(head);
2346 }
2347
2348 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2349                                 struct btrfs_delayed_ref_head *head)
2350 {
2351         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2352
2353         if (!extent_op)
2354                 return NULL;
2355
2356         if (head->must_insert_reserved) {
2357                 head->extent_op = NULL;
2358                 btrfs_free_delayed_extent_op(extent_op);
2359                 return NULL;
2360         }
2361         return extent_op;
2362 }
2363
2364 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2365                                      struct btrfs_delayed_ref_head *head)
2366 {
2367         struct btrfs_delayed_extent_op *extent_op;
2368         int ret;
2369
2370         extent_op = cleanup_extent_op(head);
2371         if (!extent_op)
2372                 return 0;
2373         head->extent_op = NULL;
2374         spin_unlock(&head->lock);
2375         ret = run_delayed_extent_op(trans, head, extent_op);
2376         btrfs_free_delayed_extent_op(extent_op);
2377         return ret ? ret : 1;
2378 }
2379
2380 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2381                                   struct btrfs_delayed_ref_root *delayed_refs,
2382                                   struct btrfs_delayed_ref_head *head)
2383 {
2384         int nr_items = 1;       /* Dropping this ref head update. */
2385
2386         if (head->total_ref_mod < 0) {
2387                 struct btrfs_space_info *space_info;
2388                 u64 flags;
2389
2390                 if (head->is_data)
2391                         flags = BTRFS_BLOCK_GROUP_DATA;
2392                 else if (head->is_system)
2393                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2394                 else
2395                         flags = BTRFS_BLOCK_GROUP_METADATA;
2396                 space_info = btrfs_find_space_info(fs_info, flags);
2397                 ASSERT(space_info);
2398                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2399                                    -head->num_bytes,
2400                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2401
2402                 /*
2403                  * We had csum deletions accounted for in our delayed refs rsv,
2404                  * we need to drop the csum leaves for this update from our
2405                  * delayed_refs_rsv.
2406                  */
2407                 if (head->is_data) {
2408                         spin_lock(&delayed_refs->lock);
2409                         delayed_refs->pending_csums -= head->num_bytes;
2410                         spin_unlock(&delayed_refs->lock);
2411                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2412                                 head->num_bytes);
2413                 }
2414         }
2415
2416         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2417 }
2418
2419 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2420                             struct btrfs_delayed_ref_head *head)
2421 {
2422
2423         struct btrfs_fs_info *fs_info = trans->fs_info;
2424         struct btrfs_delayed_ref_root *delayed_refs;
2425         int ret;
2426
2427         delayed_refs = &trans->transaction->delayed_refs;
2428
2429         ret = run_and_cleanup_extent_op(trans, head);
2430         if (ret < 0) {
2431                 unselect_delayed_ref_head(delayed_refs, head);
2432                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2433                 return ret;
2434         } else if (ret) {
2435                 return ret;
2436         }
2437
2438         /*
2439          * Need to drop our head ref lock and re-acquire the delayed ref lock
2440          * and then re-check to make sure nobody got added.
2441          */
2442         spin_unlock(&head->lock);
2443         spin_lock(&delayed_refs->lock);
2444         spin_lock(&head->lock);
2445         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2446                 spin_unlock(&head->lock);
2447                 spin_unlock(&delayed_refs->lock);
2448                 return 1;
2449         }
2450         btrfs_delete_ref_head(delayed_refs, head);
2451         spin_unlock(&head->lock);
2452         spin_unlock(&delayed_refs->lock);
2453
2454         if (head->must_insert_reserved) {
2455                 btrfs_pin_extent(fs_info, head->bytenr,
2456                                  head->num_bytes, 1);
2457                 if (head->is_data) {
2458                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2459                                               head->num_bytes);
2460                 }
2461         }
2462
2463         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2464
2465         trace_run_delayed_ref_head(fs_info, head, 0);
2466         btrfs_delayed_ref_unlock(head);
2467         btrfs_put_delayed_ref_head(head);
2468         return 0;
2469 }
2470
2471 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2472                                         struct btrfs_trans_handle *trans)
2473 {
2474         struct btrfs_delayed_ref_root *delayed_refs =
2475                 &trans->transaction->delayed_refs;
2476         struct btrfs_delayed_ref_head *head = NULL;
2477         int ret;
2478
2479         spin_lock(&delayed_refs->lock);
2480         head = btrfs_select_ref_head(delayed_refs);
2481         if (!head) {
2482                 spin_unlock(&delayed_refs->lock);
2483                 return head;
2484         }
2485
2486         /*
2487          * Grab the lock that says we are going to process all the refs for
2488          * this head
2489          */
2490         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2491         spin_unlock(&delayed_refs->lock);
2492
2493         /*
2494          * We may have dropped the spin lock to get the head mutex lock, and
2495          * that might have given someone else time to free the head.  If that's
2496          * true, it has been removed from our list and we can move on.
2497          */
2498         if (ret == -EAGAIN)
2499                 head = ERR_PTR(-EAGAIN);
2500
2501         return head;
2502 }
2503
2504 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2505                                     struct btrfs_delayed_ref_head *locked_ref,
2506                                     unsigned long *run_refs)
2507 {
2508         struct btrfs_fs_info *fs_info = trans->fs_info;
2509         struct btrfs_delayed_ref_root *delayed_refs;
2510         struct btrfs_delayed_extent_op *extent_op;
2511         struct btrfs_delayed_ref_node *ref;
2512         int must_insert_reserved = 0;
2513         int ret;
2514
2515         delayed_refs = &trans->transaction->delayed_refs;
2516
2517         lockdep_assert_held(&locked_ref->mutex);
2518         lockdep_assert_held(&locked_ref->lock);
2519
2520         while ((ref = select_delayed_ref(locked_ref))) {
2521                 if (ref->seq &&
2522                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2523                         spin_unlock(&locked_ref->lock);
2524                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2525                         return -EAGAIN;
2526                 }
2527
2528                 (*run_refs)++;
2529                 ref->in_tree = 0;
2530                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2531                 RB_CLEAR_NODE(&ref->ref_node);
2532                 if (!list_empty(&ref->add_list))
2533                         list_del(&ref->add_list);
2534                 /*
2535                  * When we play the delayed ref, also correct the ref_mod on
2536                  * head
2537                  */
2538                 switch (ref->action) {
2539                 case BTRFS_ADD_DELAYED_REF:
2540                 case BTRFS_ADD_DELAYED_EXTENT:
2541                         locked_ref->ref_mod -= ref->ref_mod;
2542                         break;
2543                 case BTRFS_DROP_DELAYED_REF:
2544                         locked_ref->ref_mod += ref->ref_mod;
2545                         break;
2546                 default:
2547                         WARN_ON(1);
2548                 }
2549                 atomic_dec(&delayed_refs->num_entries);
2550
2551                 /*
2552                  * Record the must_insert_reserved flag before we drop the
2553                  * spin lock.
2554                  */
2555                 must_insert_reserved = locked_ref->must_insert_reserved;
2556                 locked_ref->must_insert_reserved = 0;
2557
2558                 extent_op = locked_ref->extent_op;
2559                 locked_ref->extent_op = NULL;
2560                 spin_unlock(&locked_ref->lock);
2561
2562                 ret = run_one_delayed_ref(trans, ref, extent_op,
2563                                           must_insert_reserved);
2564
2565                 btrfs_free_delayed_extent_op(extent_op);
2566                 if (ret) {
2567                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2568                         btrfs_put_delayed_ref(ref);
2569                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2570                                     ret);
2571                         return ret;
2572                 }
2573
2574                 btrfs_put_delayed_ref(ref);
2575                 cond_resched();
2576
2577                 spin_lock(&locked_ref->lock);
2578                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2579         }
2580
2581         return 0;
2582 }
2583
2584 /*
2585  * Returns 0 on success or if called with an already aborted transaction.
2586  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2587  */
2588 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2589                                              unsigned long nr)
2590 {
2591         struct btrfs_fs_info *fs_info = trans->fs_info;
2592         struct btrfs_delayed_ref_root *delayed_refs;
2593         struct btrfs_delayed_ref_head *locked_ref = NULL;
2594         ktime_t start = ktime_get();
2595         int ret;
2596         unsigned long count = 0;
2597         unsigned long actual_count = 0;
2598
2599         delayed_refs = &trans->transaction->delayed_refs;
2600         do {
2601                 if (!locked_ref) {
2602                         locked_ref = btrfs_obtain_ref_head(trans);
2603                         if (IS_ERR_OR_NULL(locked_ref)) {
2604                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2605                                         continue;
2606                                 } else {
2607                                         break;
2608                                 }
2609                         }
2610                         count++;
2611                 }
2612                 /*
2613                  * We need to try and merge add/drops of the same ref since we
2614                  * can run into issues with relocate dropping the implicit ref
2615                  * and then it being added back again before the drop can
2616                  * finish.  If we merged anything we need to re-loop so we can
2617                  * get a good ref.
2618                  * Or we can get node references of the same type that weren't
2619                  * merged when created due to bumps in the tree mod seq, and
2620                  * we need to merge them to prevent adding an inline extent
2621                  * backref before dropping it (triggering a BUG_ON at
2622                  * insert_inline_extent_backref()).
2623                  */
2624                 spin_lock(&locked_ref->lock);
2625                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2626
2627                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2628                                                       &actual_count);
2629                 if (ret < 0 && ret != -EAGAIN) {
2630                         /*
2631                          * Error, btrfs_run_delayed_refs_for_head already
2632                          * unlocked everything so just bail out
2633                          */
2634                         return ret;
2635                 } else if (!ret) {
2636                         /*
2637                          * Success, perform the usual cleanup of a processed
2638                          * head
2639                          */
2640                         ret = cleanup_ref_head(trans, locked_ref);
2641                         if (ret > 0 ) {
2642                                 /* We dropped our lock, we need to loop. */
2643                                 ret = 0;
2644                                 continue;
2645                         } else if (ret) {
2646                                 return ret;
2647                         }
2648                 }
2649
2650                 /*
2651                  * Either success case or btrfs_run_delayed_refs_for_head
2652                  * returned -EAGAIN, meaning we need to select another head
2653                  */
2654
2655                 locked_ref = NULL;
2656                 cond_resched();
2657         } while ((nr != -1 && count < nr) || locked_ref);
2658
2659         /*
2660          * We don't want to include ref heads since we can have empty ref heads
2661          * and those will drastically skew our runtime down since we just do
2662          * accounting, no actual extent tree updates.
2663          */
2664         if (actual_count > 0) {
2665                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2666                 u64 avg;
2667
2668                 /*
2669                  * We weigh the current average higher than our current runtime
2670                  * to avoid large swings in the average.
2671                  */
2672                 spin_lock(&delayed_refs->lock);
2673                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2674                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2675                 spin_unlock(&delayed_refs->lock);
2676         }
2677         return 0;
2678 }
2679
2680 #ifdef SCRAMBLE_DELAYED_REFS
2681 /*
2682  * Normally delayed refs get processed in ascending bytenr order. This
2683  * correlates in most cases to the order added. To expose dependencies on this
2684  * order, we start to process the tree in the middle instead of the beginning
2685  */
2686 static u64 find_middle(struct rb_root *root)
2687 {
2688         struct rb_node *n = root->rb_node;
2689         struct btrfs_delayed_ref_node *entry;
2690         int alt = 1;
2691         u64 middle;
2692         u64 first = 0, last = 0;
2693
2694         n = rb_first(root);
2695         if (n) {
2696                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2697                 first = entry->bytenr;
2698         }
2699         n = rb_last(root);
2700         if (n) {
2701                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2702                 last = entry->bytenr;
2703         }
2704         n = root->rb_node;
2705
2706         while (n) {
2707                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2708                 WARN_ON(!entry->in_tree);
2709
2710                 middle = entry->bytenr;
2711
2712                 if (alt)
2713                         n = n->rb_left;
2714                 else
2715                         n = n->rb_right;
2716
2717                 alt = 1 - alt;
2718         }
2719         return middle;
2720 }
2721 #endif
2722
2723 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2724 {
2725         u64 num_bytes;
2726
2727         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2728                              sizeof(struct btrfs_extent_inline_ref));
2729         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2730                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2731
2732         /*
2733          * We don't ever fill up leaves all the way so multiply by 2 just to be
2734          * closer to what we're really going to want to use.
2735          */
2736         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2737 }
2738
2739 /*
2740  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2741  * would require to store the csums for that many bytes.
2742  */
2743 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2744 {
2745         u64 csum_size;
2746         u64 num_csums_per_leaf;
2747         u64 num_csums;
2748
2749         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2750         num_csums_per_leaf = div64_u64(csum_size,
2751                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2752         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2753         num_csums += num_csums_per_leaf - 1;
2754         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2755         return num_csums;
2756 }
2757
2758 /*
2759  * this starts processing the delayed reference count updates and
2760  * extent insertions we have queued up so far.  count can be
2761  * 0, which means to process everything in the tree at the start
2762  * of the run (but not newly added entries), or it can be some target
2763  * number you'd like to process.
2764  *
2765  * Returns 0 on success or if called with an aborted transaction
2766  * Returns <0 on error and aborts the transaction
2767  */
2768 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2769                            unsigned long count)
2770 {
2771         struct btrfs_fs_info *fs_info = trans->fs_info;
2772         struct rb_node *node;
2773         struct btrfs_delayed_ref_root *delayed_refs;
2774         struct btrfs_delayed_ref_head *head;
2775         int ret;
2776         int run_all = count == (unsigned long)-1;
2777
2778         /* We'll clean this up in btrfs_cleanup_transaction */
2779         if (trans->aborted)
2780                 return 0;
2781
2782         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2783                 return 0;
2784
2785         delayed_refs = &trans->transaction->delayed_refs;
2786         if (count == 0)
2787                 count = atomic_read(&delayed_refs->num_entries) * 2;
2788
2789 again:
2790 #ifdef SCRAMBLE_DELAYED_REFS
2791         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2792 #endif
2793         ret = __btrfs_run_delayed_refs(trans, count);
2794         if (ret < 0) {
2795                 btrfs_abort_transaction(trans, ret);
2796                 return ret;
2797         }
2798
2799         if (run_all) {
2800                 btrfs_create_pending_block_groups(trans);
2801
2802                 spin_lock(&delayed_refs->lock);
2803                 node = rb_first_cached(&delayed_refs->href_root);
2804                 if (!node) {
2805                         spin_unlock(&delayed_refs->lock);
2806                         goto out;
2807                 }
2808                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2809                                 href_node);
2810                 refcount_inc(&head->refs);
2811                 spin_unlock(&delayed_refs->lock);
2812
2813                 /* Mutex was contended, block until it's released and retry. */
2814                 mutex_lock(&head->mutex);
2815                 mutex_unlock(&head->mutex);
2816
2817                 btrfs_put_delayed_ref_head(head);
2818                 cond_resched();
2819                 goto again;
2820         }
2821 out:
2822         return 0;
2823 }
2824
2825 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2826                                 u64 bytenr, u64 num_bytes, u64 flags,
2827                                 int level, int is_data)
2828 {
2829         struct btrfs_delayed_extent_op *extent_op;
2830         int ret;
2831
2832         extent_op = btrfs_alloc_delayed_extent_op();
2833         if (!extent_op)
2834                 return -ENOMEM;
2835
2836         extent_op->flags_to_set = flags;
2837         extent_op->update_flags = true;
2838         extent_op->update_key = false;
2839         extent_op->is_data = is_data ? true : false;
2840         extent_op->level = level;
2841
2842         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2843         if (ret)
2844                 btrfs_free_delayed_extent_op(extent_op);
2845         return ret;
2846 }
2847
2848 static noinline int check_delayed_ref(struct btrfs_root *root,
2849                                       struct btrfs_path *path,
2850                                       u64 objectid, u64 offset, u64 bytenr)
2851 {
2852         struct btrfs_delayed_ref_head *head;
2853         struct btrfs_delayed_ref_node *ref;
2854         struct btrfs_delayed_data_ref *data_ref;
2855         struct btrfs_delayed_ref_root *delayed_refs;
2856         struct btrfs_transaction *cur_trans;
2857         struct rb_node *node;
2858         int ret = 0;
2859
2860         spin_lock(&root->fs_info->trans_lock);
2861         cur_trans = root->fs_info->running_transaction;
2862         if (cur_trans)
2863                 refcount_inc(&cur_trans->use_count);
2864         spin_unlock(&root->fs_info->trans_lock);
2865         if (!cur_trans)
2866                 return 0;
2867
2868         delayed_refs = &cur_trans->delayed_refs;
2869         spin_lock(&delayed_refs->lock);
2870         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2871         if (!head) {
2872                 spin_unlock(&delayed_refs->lock);
2873                 btrfs_put_transaction(cur_trans);
2874                 return 0;
2875         }
2876
2877         if (!mutex_trylock(&head->mutex)) {
2878                 refcount_inc(&head->refs);
2879                 spin_unlock(&delayed_refs->lock);
2880
2881                 btrfs_release_path(path);
2882
2883                 /*
2884                  * Mutex was contended, block until it's released and let
2885                  * caller try again
2886                  */
2887                 mutex_lock(&head->mutex);
2888                 mutex_unlock(&head->mutex);
2889                 btrfs_put_delayed_ref_head(head);
2890                 btrfs_put_transaction(cur_trans);
2891                 return -EAGAIN;
2892         }
2893         spin_unlock(&delayed_refs->lock);
2894
2895         spin_lock(&head->lock);
2896         /*
2897          * XXX: We should replace this with a proper search function in the
2898          * future.
2899          */
2900         for (node = rb_first_cached(&head->ref_tree); node;
2901              node = rb_next(node)) {
2902                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2903                 /* If it's a shared ref we know a cross reference exists */
2904                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2905                         ret = 1;
2906                         break;
2907                 }
2908
2909                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2910
2911                 /*
2912                  * If our ref doesn't match the one we're currently looking at
2913                  * then we have a cross reference.
2914                  */
2915                 if (data_ref->root != root->root_key.objectid ||
2916                     data_ref->objectid != objectid ||
2917                     data_ref->offset != offset) {
2918                         ret = 1;
2919                         break;
2920                 }
2921         }
2922         spin_unlock(&head->lock);
2923         mutex_unlock(&head->mutex);
2924         btrfs_put_transaction(cur_trans);
2925         return ret;
2926 }
2927
2928 static noinline int check_committed_ref(struct btrfs_root *root,
2929                                         struct btrfs_path *path,
2930                                         u64 objectid, u64 offset, u64 bytenr)
2931 {
2932         struct btrfs_fs_info *fs_info = root->fs_info;
2933         struct btrfs_root *extent_root = fs_info->extent_root;
2934         struct extent_buffer *leaf;
2935         struct btrfs_extent_data_ref *ref;
2936         struct btrfs_extent_inline_ref *iref;
2937         struct btrfs_extent_item *ei;
2938         struct btrfs_key key;
2939         u32 item_size;
2940         int type;
2941         int ret;
2942
2943         key.objectid = bytenr;
2944         key.offset = (u64)-1;
2945         key.type = BTRFS_EXTENT_ITEM_KEY;
2946
2947         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2948         if (ret < 0)
2949                 goto out;
2950         BUG_ON(ret == 0); /* Corruption */
2951
2952         ret = -ENOENT;
2953         if (path->slots[0] == 0)
2954                 goto out;
2955
2956         path->slots[0]--;
2957         leaf = path->nodes[0];
2958         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2959
2960         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2961                 goto out;
2962
2963         ret = 1;
2964         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2965         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2966
2967         if (item_size != sizeof(*ei) +
2968             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2969                 goto out;
2970
2971         if (btrfs_extent_generation(leaf, ei) <=
2972             btrfs_root_last_snapshot(&root->root_item))
2973                 goto out;
2974
2975         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2976
2977         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2978         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2979                 goto out;
2980
2981         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2982         if (btrfs_extent_refs(leaf, ei) !=
2983             btrfs_extent_data_ref_count(leaf, ref) ||
2984             btrfs_extent_data_ref_root(leaf, ref) !=
2985             root->root_key.objectid ||
2986             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2987             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2988                 goto out;
2989
2990         ret = 0;
2991 out:
2992         return ret;
2993 }
2994
2995 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2996                           u64 bytenr)
2997 {
2998         struct btrfs_path *path;
2999         int ret;
3000
3001         path = btrfs_alloc_path();
3002         if (!path)
3003                 return -ENOMEM;
3004
3005         do {
3006                 ret = check_committed_ref(root, path, objectid,
3007                                           offset, bytenr);
3008                 if (ret && ret != -ENOENT)
3009                         goto out;
3010
3011                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3012         } while (ret == -EAGAIN);
3013
3014 out:
3015         btrfs_free_path(path);
3016         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3017                 WARN_ON(ret > 0);
3018         return ret;
3019 }
3020
3021 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3022                            struct btrfs_root *root,
3023                            struct extent_buffer *buf,
3024                            int full_backref, int inc)
3025 {
3026         struct btrfs_fs_info *fs_info = root->fs_info;
3027         u64 bytenr;
3028         u64 num_bytes;
3029         u64 parent;
3030         u64 ref_root;
3031         u32 nritems;
3032         struct btrfs_key key;
3033         struct btrfs_file_extent_item *fi;
3034         struct btrfs_ref generic_ref = { 0 };
3035         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
3036         int i;
3037         int action;
3038         int level;
3039         int ret = 0;
3040
3041         if (btrfs_is_testing(fs_info))
3042                 return 0;
3043
3044         ref_root = btrfs_header_owner(buf);
3045         nritems = btrfs_header_nritems(buf);
3046         level = btrfs_header_level(buf);
3047
3048         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3049                 return 0;
3050
3051         if (full_backref)
3052                 parent = buf->start;
3053         else
3054                 parent = 0;
3055         if (inc)
3056                 action = BTRFS_ADD_DELAYED_REF;
3057         else
3058                 action = BTRFS_DROP_DELAYED_REF;
3059
3060         for (i = 0; i < nritems; i++) {
3061                 if (level == 0) {
3062                         btrfs_item_key_to_cpu(buf, &key, i);
3063                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3064                                 continue;
3065                         fi = btrfs_item_ptr(buf, i,
3066                                             struct btrfs_file_extent_item);
3067                         if (btrfs_file_extent_type(buf, fi) ==
3068                             BTRFS_FILE_EXTENT_INLINE)
3069                                 continue;
3070                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3071                         if (bytenr == 0)
3072                                 continue;
3073
3074                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3075                         key.offset -= btrfs_file_extent_offset(buf, fi);
3076                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3077                                                num_bytes, parent);
3078                         generic_ref.real_root = root->root_key.objectid;
3079                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
3080                                             key.offset);
3081                         generic_ref.skip_qgroup = for_reloc;
3082                         if (inc)
3083                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3084                         else
3085                                 ret = btrfs_free_extent(trans, &generic_ref);
3086                         if (ret)
3087                                 goto fail;
3088                 } else {
3089                         bytenr = btrfs_node_blockptr(buf, i);
3090                         num_bytes = fs_info->nodesize;
3091                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3092                                                num_bytes, parent);
3093                         generic_ref.real_root = root->root_key.objectid;
3094                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
3095                         generic_ref.skip_qgroup = for_reloc;
3096                         if (inc)
3097                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3098                         else
3099                                 ret = btrfs_free_extent(trans, &generic_ref);
3100                         if (ret)
3101                                 goto fail;
3102                 }
3103         }
3104         return 0;
3105 fail:
3106         return ret;
3107 }
3108
3109 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3110                   struct extent_buffer *buf, int full_backref)
3111 {
3112         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3113 }
3114
3115 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3116                   struct extent_buffer *buf, int full_backref)
3117 {
3118         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3119 }
3120
3121 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3122                                  struct btrfs_path *path,
3123                                  struct btrfs_block_group_cache *cache)
3124 {
3125         struct btrfs_fs_info *fs_info = trans->fs_info;
3126         int ret;
3127         struct btrfs_root *extent_root = fs_info->extent_root;
3128         unsigned long bi;
3129         struct extent_buffer *leaf;
3130
3131         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3132         if (ret) {
3133                 if (ret > 0)
3134                         ret = -ENOENT;
3135                 goto fail;
3136         }
3137
3138         leaf = path->nodes[0];
3139         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3140         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3141         btrfs_mark_buffer_dirty(leaf);
3142 fail:
3143         btrfs_release_path(path);
3144         return ret;
3145
3146 }
3147
3148 static struct btrfs_block_group_cache *next_block_group(
3149                 struct btrfs_block_group_cache *cache)
3150 {
3151         struct btrfs_fs_info *fs_info = cache->fs_info;
3152         struct rb_node *node;
3153
3154         spin_lock(&fs_info->block_group_cache_lock);
3155
3156         /* If our block group was removed, we need a full search. */
3157         if (RB_EMPTY_NODE(&cache->cache_node)) {
3158                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3159
3160                 spin_unlock(&fs_info->block_group_cache_lock);
3161                 btrfs_put_block_group(cache);
3162                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3163         }
3164         node = rb_next(&cache->cache_node);
3165         btrfs_put_block_group(cache);
3166         if (node) {
3167                 cache = rb_entry(node, struct btrfs_block_group_cache,
3168                                  cache_node);
3169                 btrfs_get_block_group(cache);
3170         } else
3171                 cache = NULL;
3172         spin_unlock(&fs_info->block_group_cache_lock);
3173         return cache;
3174 }
3175
3176 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3177                             struct btrfs_trans_handle *trans,
3178                             struct btrfs_path *path)
3179 {
3180         struct btrfs_fs_info *fs_info = block_group->fs_info;
3181         struct btrfs_root *root = fs_info->tree_root;
3182         struct inode *inode = NULL;
3183         struct extent_changeset *data_reserved = NULL;
3184         u64 alloc_hint = 0;
3185         int dcs = BTRFS_DC_ERROR;
3186         u64 num_pages = 0;
3187         int retries = 0;
3188         int ret = 0;
3189
3190         /*
3191          * If this block group is smaller than 100 megs don't bother caching the
3192          * block group.
3193          */
3194         if (block_group->key.offset < (100 * SZ_1M)) {
3195                 spin_lock(&block_group->lock);
3196                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3197                 spin_unlock(&block_group->lock);
3198                 return 0;
3199         }
3200
3201         if (trans->aborted)
3202                 return 0;
3203 again:
3204         inode = lookup_free_space_inode(block_group, path);
3205         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3206                 ret = PTR_ERR(inode);
3207                 btrfs_release_path(path);
3208                 goto out;
3209         }
3210
3211         if (IS_ERR(inode)) {
3212                 BUG_ON(retries);
3213                 retries++;
3214
3215                 if (block_group->ro)
3216                         goto out_free;
3217
3218                 ret = create_free_space_inode(trans, block_group, path);
3219                 if (ret)
3220                         goto out_free;
3221                 goto again;
3222         }
3223
3224         /*
3225          * We want to set the generation to 0, that way if anything goes wrong
3226          * from here on out we know not to trust this cache when we load up next
3227          * time.
3228          */
3229         BTRFS_I(inode)->generation = 0;
3230         ret = btrfs_update_inode(trans, root, inode);
3231         if (ret) {
3232                 /*
3233                  * So theoretically we could recover from this, simply set the
3234                  * super cache generation to 0 so we know to invalidate the
3235                  * cache, but then we'd have to keep track of the block groups
3236                  * that fail this way so we know we _have_ to reset this cache
3237                  * before the next commit or risk reading stale cache.  So to
3238                  * limit our exposure to horrible edge cases lets just abort the
3239                  * transaction, this only happens in really bad situations
3240                  * anyway.
3241                  */
3242                 btrfs_abort_transaction(trans, ret);
3243                 goto out_put;
3244         }
3245         WARN_ON(ret);
3246
3247         /* We've already setup this transaction, go ahead and exit */
3248         if (block_group->cache_generation == trans->transid &&
3249             i_size_read(inode)) {
3250                 dcs = BTRFS_DC_SETUP;
3251                 goto out_put;
3252         }
3253
3254         if (i_size_read(inode) > 0) {
3255                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3256                                         &fs_info->global_block_rsv);
3257                 if (ret)
3258                         goto out_put;
3259
3260                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3261                 if (ret)
3262                         goto out_put;
3263         }
3264
3265         spin_lock(&block_group->lock);
3266         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3267             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3268                 /*
3269                  * don't bother trying to write stuff out _if_
3270                  * a) we're not cached,
3271                  * b) we're with nospace_cache mount option,
3272                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3273                  */
3274                 dcs = BTRFS_DC_WRITTEN;
3275                 spin_unlock(&block_group->lock);
3276                 goto out_put;
3277         }
3278         spin_unlock(&block_group->lock);
3279
3280         /*
3281          * We hit an ENOSPC when setting up the cache in this transaction, just
3282          * skip doing the setup, we've already cleared the cache so we're safe.
3283          */
3284         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3285                 ret = -ENOSPC;
3286                 goto out_put;
3287         }
3288
3289         /*
3290          * Try to preallocate enough space based on how big the block group is.
3291          * Keep in mind this has to include any pinned space which could end up
3292          * taking up quite a bit since it's not folded into the other space
3293          * cache.
3294          */
3295         num_pages = div_u64(block_group->key.offset, SZ_256M);
3296         if (!num_pages)
3297                 num_pages = 1;
3298
3299         num_pages *= 16;
3300         num_pages *= PAGE_SIZE;
3301
3302         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3303         if (ret)
3304                 goto out_put;
3305
3306         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3307                                               num_pages, num_pages,
3308                                               &alloc_hint);
3309         /*
3310          * Our cache requires contiguous chunks so that we don't modify a bunch
3311          * of metadata or split extents when writing the cache out, which means
3312          * we can enospc if we are heavily fragmented in addition to just normal
3313          * out of space conditions.  So if we hit this just skip setting up any
3314          * other block groups for this transaction, maybe we'll unpin enough
3315          * space the next time around.
3316          */
3317         if (!ret)
3318                 dcs = BTRFS_DC_SETUP;
3319         else if (ret == -ENOSPC)
3320                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3321
3322 out_put:
3323         iput(inode);
3324 out_free:
3325         btrfs_release_path(path);
3326 out:
3327         spin_lock(&block_group->lock);
3328         if (!ret && dcs == BTRFS_DC_SETUP)
3329                 block_group->cache_generation = trans->transid;
3330         block_group->disk_cache_state = dcs;
3331         spin_unlock(&block_group->lock);
3332
3333         extent_changeset_free(data_reserved);
3334         return ret;
3335 }
3336
3337 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3338 {
3339         struct btrfs_fs_info *fs_info = trans->fs_info;
3340         struct btrfs_block_group_cache *cache, *tmp;
3341         struct btrfs_transaction *cur_trans = trans->transaction;
3342         struct btrfs_path *path;
3343
3344         if (list_empty(&cur_trans->dirty_bgs) ||
3345             !btrfs_test_opt(fs_info, SPACE_CACHE))
3346                 return 0;
3347
3348         path = btrfs_alloc_path();
3349         if (!path)
3350                 return -ENOMEM;
3351
3352         /* Could add new block groups, use _safe just in case */
3353         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3354                                  dirty_list) {
3355                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3356                         cache_save_setup(cache, trans, path);
3357         }
3358
3359         btrfs_free_path(path);
3360         return 0;
3361 }
3362
3363 /*
3364  * transaction commit does final block group cache writeback during a
3365  * critical section where nothing is allowed to change the FS.  This is
3366  * required in order for the cache to actually match the block group,
3367  * but can introduce a lot of latency into the commit.
3368  *
3369  * So, btrfs_start_dirty_block_groups is here to kick off block group
3370  * cache IO.  There's a chance we'll have to redo some of it if the
3371  * block group changes again during the commit, but it greatly reduces
3372  * the commit latency by getting rid of the easy block groups while
3373  * we're still allowing others to join the commit.
3374  */
3375 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3376 {
3377         struct btrfs_fs_info *fs_info = trans->fs_info;
3378         struct btrfs_block_group_cache *cache;
3379         struct btrfs_transaction *cur_trans = trans->transaction;
3380         int ret = 0;
3381         int should_put;
3382         struct btrfs_path *path = NULL;
3383         LIST_HEAD(dirty);
3384         struct list_head *io = &cur_trans->io_bgs;
3385         int num_started = 0;
3386         int loops = 0;
3387
3388         spin_lock(&cur_trans->dirty_bgs_lock);
3389         if (list_empty(&cur_trans->dirty_bgs)) {
3390                 spin_unlock(&cur_trans->dirty_bgs_lock);
3391                 return 0;
3392         }
3393         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3394         spin_unlock(&cur_trans->dirty_bgs_lock);
3395
3396 again:
3397         /*
3398          * make sure all the block groups on our dirty list actually
3399          * exist
3400          */
3401         btrfs_create_pending_block_groups(trans);
3402
3403         if (!path) {
3404                 path = btrfs_alloc_path();
3405                 if (!path)
3406                         return -ENOMEM;
3407         }
3408
3409         /*
3410          * cache_write_mutex is here only to save us from balance or automatic
3411          * removal of empty block groups deleting this block group while we are
3412          * writing out the cache
3413          */
3414         mutex_lock(&trans->transaction->cache_write_mutex);
3415         while (!list_empty(&dirty)) {
3416                 bool drop_reserve = true;
3417
3418                 cache = list_first_entry(&dirty,
3419                                          struct btrfs_block_group_cache,
3420                                          dirty_list);
3421                 /*
3422                  * this can happen if something re-dirties a block
3423                  * group that is already under IO.  Just wait for it to
3424                  * finish and then do it all again
3425                  */
3426                 if (!list_empty(&cache->io_list)) {
3427                         list_del_init(&cache->io_list);
3428                         btrfs_wait_cache_io(trans, cache, path);
3429                         btrfs_put_block_group(cache);
3430                 }
3431
3432
3433                 /*
3434                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3435                  * if it should update the cache_state.  Don't delete
3436                  * until after we wait.
3437                  *
3438                  * Since we're not running in the commit critical section
3439                  * we need the dirty_bgs_lock to protect from update_block_group
3440                  */
3441                 spin_lock(&cur_trans->dirty_bgs_lock);
3442                 list_del_init(&cache->dirty_list);
3443                 spin_unlock(&cur_trans->dirty_bgs_lock);
3444
3445                 should_put = 1;
3446
3447                 cache_save_setup(cache, trans, path);
3448
3449                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3450                         cache->io_ctl.inode = NULL;
3451                         ret = btrfs_write_out_cache(trans, cache, path);
3452                         if (ret == 0 && cache->io_ctl.inode) {
3453                                 num_started++;
3454                                 should_put = 0;
3455
3456                                 /*
3457                                  * The cache_write_mutex is protecting the
3458                                  * io_list, also refer to the definition of
3459                                  * btrfs_transaction::io_bgs for more details
3460                                  */
3461                                 list_add_tail(&cache->io_list, io);
3462                         } else {
3463                                 /*
3464                                  * if we failed to write the cache, the
3465                                  * generation will be bad and life goes on
3466                                  */
3467                                 ret = 0;
3468                         }
3469                 }
3470                 if (!ret) {
3471                         ret = write_one_cache_group(trans, path, cache);
3472                         /*
3473                          * Our block group might still be attached to the list
3474                          * of new block groups in the transaction handle of some
3475                          * other task (struct btrfs_trans_handle->new_bgs). This
3476                          * means its block group item isn't yet in the extent
3477                          * tree. If this happens ignore the error, as we will
3478                          * try again later in the critical section of the
3479                          * transaction commit.
3480                          */
3481                         if (ret == -ENOENT) {
3482                                 ret = 0;
3483                                 spin_lock(&cur_trans->dirty_bgs_lock);
3484                                 if (list_empty(&cache->dirty_list)) {
3485                                         list_add_tail(&cache->dirty_list,
3486                                                       &cur_trans->dirty_bgs);
3487                                         btrfs_get_block_group(cache);
3488                                         drop_reserve = false;
3489                                 }
3490                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3491                         } else if (ret) {
3492                                 btrfs_abort_transaction(trans, ret);
3493                         }
3494                 }
3495
3496                 /* if it's not on the io list, we need to put the block group */
3497                 if (should_put)
3498                         btrfs_put_block_group(cache);
3499                 if (drop_reserve)
3500                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3501
3502                 if (ret)
3503                         break;
3504
3505                 /*
3506                  * Avoid blocking other tasks for too long. It might even save
3507                  * us from writing caches for block groups that are going to be
3508                  * removed.
3509                  */
3510                 mutex_unlock(&trans->transaction->cache_write_mutex);
3511                 mutex_lock(&trans->transaction->cache_write_mutex);
3512         }
3513         mutex_unlock(&trans->transaction->cache_write_mutex);
3514
3515         /*
3516          * go through delayed refs for all the stuff we've just kicked off
3517          * and then loop back (just once)
3518          */
3519         ret = btrfs_run_delayed_refs(trans, 0);
3520         if (!ret && loops == 0) {
3521                 loops++;
3522                 spin_lock(&cur_trans->dirty_bgs_lock);
3523                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3524                 /*
3525                  * dirty_bgs_lock protects us from concurrent block group
3526                  * deletes too (not just cache_write_mutex).
3527                  */
3528                 if (!list_empty(&dirty)) {
3529                         spin_unlock(&cur_trans->dirty_bgs_lock);
3530                         goto again;
3531                 }
3532                 spin_unlock(&cur_trans->dirty_bgs_lock);
3533         } else if (ret < 0) {
3534                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3535         }
3536
3537         btrfs_free_path(path);
3538         return ret;
3539 }
3540
3541 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3542 {
3543         struct btrfs_fs_info *fs_info = trans->fs_info;
3544         struct btrfs_block_group_cache *cache;
3545         struct btrfs_transaction *cur_trans = trans->transaction;
3546         int ret = 0;
3547         int should_put;
3548         struct btrfs_path *path;
3549         struct list_head *io = &cur_trans->io_bgs;
3550         int num_started = 0;
3551
3552         path = btrfs_alloc_path();
3553         if (!path)
3554                 return -ENOMEM;
3555
3556         /*
3557          * Even though we are in the critical section of the transaction commit,
3558          * we can still have concurrent tasks adding elements to this
3559          * transaction's list of dirty block groups. These tasks correspond to
3560          * endio free space workers started when writeback finishes for a
3561          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3562          * allocate new block groups as a result of COWing nodes of the root
3563          * tree when updating the free space inode. The writeback for the space
3564          * caches is triggered by an earlier call to
3565          * btrfs_start_dirty_block_groups() and iterations of the following
3566          * loop.
3567          * Also we want to do the cache_save_setup first and then run the
3568          * delayed refs to make sure we have the best chance at doing this all
3569          * in one shot.
3570          */
3571         spin_lock(&cur_trans->dirty_bgs_lock);
3572         while (!list_empty(&cur_trans->dirty_bgs)) {
3573                 cache = list_first_entry(&cur_trans->dirty_bgs,
3574                                          struct btrfs_block_group_cache,
3575                                          dirty_list);
3576
3577                 /*
3578                  * this can happen if cache_save_setup re-dirties a block
3579                  * group that is already under IO.  Just wait for it to
3580                  * finish and then do it all again
3581                  */
3582                 if (!list_empty(&cache->io_list)) {
3583                         spin_unlock(&cur_trans->dirty_bgs_lock);
3584                         list_del_init(&cache->io_list);
3585                         btrfs_wait_cache_io(trans, cache, path);
3586                         btrfs_put_block_group(cache);
3587                         spin_lock(&cur_trans->dirty_bgs_lock);
3588                 }
3589
3590                 /*
3591                  * don't remove from the dirty list until after we've waited
3592                  * on any pending IO
3593                  */
3594                 list_del_init(&cache->dirty_list);
3595                 spin_unlock(&cur_trans->dirty_bgs_lock);
3596                 should_put = 1;
3597
3598                 cache_save_setup(cache, trans, path);
3599
3600                 if (!ret)
3601                         ret = btrfs_run_delayed_refs(trans,
3602                                                      (unsigned long) -1);
3603
3604                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3605                         cache->io_ctl.inode = NULL;
3606                         ret = btrfs_write_out_cache(trans, cache, path);
3607                         if (ret == 0 && cache->io_ctl.inode) {
3608                                 num_started++;
3609                                 should_put = 0;
3610                                 list_add_tail(&cache->io_list, io);
3611                         } else {
3612                                 /*
3613                                  * if we failed to write the cache, the
3614                                  * generation will be bad and life goes on
3615                                  */
3616                                 ret = 0;
3617                         }
3618                 }
3619                 if (!ret) {
3620                         ret = write_one_cache_group(trans, path, cache);
3621                         /*
3622                          * One of the free space endio workers might have
3623                          * created a new block group while updating a free space
3624                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3625                          * and hasn't released its transaction handle yet, in
3626                          * which case the new block group is still attached to
3627                          * its transaction handle and its creation has not
3628                          * finished yet (no block group item in the extent tree
3629                          * yet, etc). If this is the case, wait for all free
3630                          * space endio workers to finish and retry. This is a
3631                          * a very rare case so no need for a more efficient and
3632                          * complex approach.
3633                          */
3634                         if (ret == -ENOENT) {
3635                                 wait_event(cur_trans->writer_wait,
3636                                    atomic_read(&cur_trans->num_writers) == 1);
3637                                 ret = write_one_cache_group(trans, path, cache);
3638                         }
3639                         if (ret)
3640                                 btrfs_abort_transaction(trans, ret);
3641                 }
3642
3643                 /* if its not on the io list, we need to put the block group */
3644                 if (should_put)
3645                         btrfs_put_block_group(cache);
3646                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3647                 spin_lock(&cur_trans->dirty_bgs_lock);
3648         }
3649         spin_unlock(&cur_trans->dirty_bgs_lock);
3650
3651         /*
3652          * Refer to the definition of io_bgs member for details why it's safe
3653          * to use it without any locking
3654          */
3655         while (!list_empty(io)) {
3656                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3657                                          io_list);
3658                 list_del_init(&cache->io_list);
3659                 btrfs_wait_cache_io(trans, cache, path);
3660                 btrfs_put_block_group(cache);
3661         }
3662
3663         btrfs_free_path(path);
3664         return ret;
3665 }
3666
3667 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3668 {
3669         struct btrfs_block_group_cache *block_group;
3670         int readonly = 0;
3671
3672         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3673         if (!block_group || block_group->ro)
3674                 readonly = 1;
3675         if (block_group)
3676                 btrfs_put_block_group(block_group);
3677         return readonly;
3678 }
3679
3680 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3681 {
3682         struct btrfs_block_group_cache *bg;
3683         bool ret = true;
3684
3685         bg = btrfs_lookup_block_group(fs_info, bytenr);
3686         if (!bg)
3687                 return false;
3688
3689         spin_lock(&bg->lock);
3690         if (bg->ro)
3691                 ret = false;
3692         else
3693                 atomic_inc(&bg->nocow_writers);
3694         spin_unlock(&bg->lock);
3695
3696         /* no put on block group, done by btrfs_dec_nocow_writers */
3697         if (!ret)
3698                 btrfs_put_block_group(bg);
3699
3700         return ret;
3701
3702 }
3703
3704 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3705 {
3706         struct btrfs_block_group_cache *bg;
3707
3708         bg = btrfs_lookup_block_group(fs_info, bytenr);
3709         ASSERT(bg);
3710         if (atomic_dec_and_test(&bg->nocow_writers))
3711                 wake_up_var(&bg->nocow_writers);
3712         /*
3713          * Once for our lookup and once for the lookup done by a previous call
3714          * to btrfs_inc_nocow_writers()
3715          */
3716         btrfs_put_block_group(bg);
3717         btrfs_put_block_group(bg);
3718 }
3719
3720 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3721 {
3722         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3723 }
3724
3725 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3726 {
3727         u64 extra_flags = chunk_to_extended(flags) &
3728                                 BTRFS_EXTENDED_PROFILE_MASK;
3729
3730         write_seqlock(&fs_info->profiles_lock);
3731         if (flags & BTRFS_BLOCK_GROUP_DATA)
3732                 fs_info->avail_data_alloc_bits |= extra_flags;
3733         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3734                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3735         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3736                 fs_info->avail_system_alloc_bits |= extra_flags;
3737         write_sequnlock(&fs_info->profiles_lock);
3738 }
3739
3740 /*
3741  * returns target flags in extended format or 0 if restripe for this
3742  * chunk_type is not in progress
3743  *
3744  * should be called with balance_lock held
3745  */
3746 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3747 {
3748         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3749         u64 target = 0;
3750
3751         if (!bctl)
3752                 return 0;
3753
3754         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3755             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3756                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3757         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3758                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3759                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3760         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3761                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3762                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3763         }
3764
3765         return target;
3766 }
3767
3768 /*
3769  * @flags: available profiles in extended format (see ctree.h)
3770  *
3771  * Returns reduced profile in chunk format.  If profile changing is in
3772  * progress (either running or paused) picks the target profile (if it's
3773  * already available), otherwise falls back to plain reducing.
3774  */
3775 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
3776 {
3777         u64 num_devices = fs_info->fs_devices->rw_devices;
3778         u64 target;
3779         u64 raid_type;
3780         u64 allowed = 0;
3781
3782         /*
3783          * see if restripe for this chunk_type is in progress, if so
3784          * try to reduce to the target profile
3785          */
3786         spin_lock(&fs_info->balance_lock);
3787         target = get_restripe_target(fs_info, flags);
3788         if (target) {
3789                 /* pick target profile only if it's already available */
3790                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3791                         spin_unlock(&fs_info->balance_lock);
3792                         return extended_to_chunk(target);
3793                 }
3794         }
3795         spin_unlock(&fs_info->balance_lock);
3796
3797         /* First, mask out the RAID levels which aren't possible */
3798         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3799                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3800                         allowed |= btrfs_raid_array[raid_type].bg_flag;
3801         }
3802         allowed &= flags;
3803
3804         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
3805                 allowed = BTRFS_BLOCK_GROUP_RAID6;
3806         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
3807                 allowed = BTRFS_BLOCK_GROUP_RAID5;
3808         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
3809                 allowed = BTRFS_BLOCK_GROUP_RAID10;
3810         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
3811                 allowed = BTRFS_BLOCK_GROUP_RAID1;
3812         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
3813                 allowed = BTRFS_BLOCK_GROUP_RAID0;
3814
3815         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
3816
3817         return extended_to_chunk(flags | allowed);
3818 }
3819
3820 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
3821 {
3822         unsigned seq;
3823         u64 flags;
3824
3825         do {
3826                 flags = orig_flags;
3827                 seq = read_seqbegin(&fs_info->profiles_lock);
3828
3829                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3830                         flags |= fs_info->avail_data_alloc_bits;
3831                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3832                         flags |= fs_info->avail_system_alloc_bits;
3833                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3834                         flags |= fs_info->avail_metadata_alloc_bits;
3835         } while (read_seqretry(&fs_info->profiles_lock, seq));
3836
3837         return btrfs_reduce_alloc_profile(fs_info, flags);
3838 }
3839
3840 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
3841 {
3842         struct btrfs_fs_info *fs_info = root->fs_info;
3843         u64 flags;
3844         u64 ret;
3845
3846         if (data)
3847                 flags = BTRFS_BLOCK_GROUP_DATA;
3848         else if (root == fs_info->chunk_root)
3849                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3850         else
3851                 flags = BTRFS_BLOCK_GROUP_METADATA;
3852
3853         ret = get_alloc_profile(fs_info, flags);
3854         return ret;
3855 }
3856
3857 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
3858 {
3859         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
3860 }
3861
3862 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
3863 {
3864         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3865 }
3866
3867 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
3868 {
3869         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3870 }
3871
3872 static void force_metadata_allocation(struct btrfs_fs_info *info)
3873 {
3874         struct list_head *head = &info->space_info;
3875         struct btrfs_space_info *found;
3876
3877         rcu_read_lock();
3878         list_for_each_entry_rcu(found, head, list) {
3879                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3880                         found->force_alloc = CHUNK_ALLOC_FORCE;
3881         }
3882         rcu_read_unlock();
3883 }
3884
3885 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3886                               struct btrfs_space_info *sinfo, int force)
3887 {
3888         u64 bytes_used = btrfs_space_info_used(sinfo, false);
3889         u64 thresh;
3890
3891         if (force == CHUNK_ALLOC_FORCE)
3892                 return 1;
3893
3894         /*
3895          * in limited mode, we want to have some free space up to
3896          * about 1% of the FS size.
3897          */
3898         if (force == CHUNK_ALLOC_LIMITED) {
3899                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3900                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3901
3902                 if (sinfo->total_bytes - bytes_used < thresh)
3903                         return 1;
3904         }
3905
3906         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3907                 return 0;
3908         return 1;
3909 }
3910
3911 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3912 {
3913         u64 num_dev;
3914
3915         num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3916         if (!num_dev)
3917                 num_dev = fs_info->fs_devices->rw_devices;
3918
3919         return num_dev;
3920 }
3921
3922 /*
3923  * If @is_allocation is true, reserve space in the system space info necessary
3924  * for allocating a chunk, otherwise if it's false, reserve space necessary for
3925  * removing a chunk.
3926  */
3927 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3928 {
3929         struct btrfs_fs_info *fs_info = trans->fs_info;
3930         struct btrfs_space_info *info;
3931         u64 left;
3932         u64 thresh;
3933         int ret = 0;
3934         u64 num_devs;
3935
3936         /*
3937          * Needed because we can end up allocating a system chunk and for an
3938          * atomic and race free space reservation in the chunk block reserve.
3939          */
3940         lockdep_assert_held(&fs_info->chunk_mutex);
3941
3942         info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3943         spin_lock(&info->lock);
3944         left = info->total_bytes - btrfs_space_info_used(info, true);
3945         spin_unlock(&info->lock);
3946
3947         num_devs = get_profile_num_devs(fs_info, type);
3948
3949         /* num_devs device items to update and 1 chunk item to add or remove */
3950         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
3951                 btrfs_calc_trans_metadata_size(fs_info, 1);
3952
3953         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3954                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3955                            left, thresh, type);
3956                 btrfs_dump_space_info(fs_info, info, 0, 0);
3957         }
3958
3959         if (left < thresh) {
3960                 u64 flags = btrfs_system_alloc_profile(fs_info);
3961
3962                 /*
3963                  * Ignore failure to create system chunk. We might end up not
3964                  * needing it, as we might not need to COW all nodes/leafs from
3965                  * the paths we visit in the chunk tree (they were already COWed
3966                  * or created in the current transaction for example).
3967                  */
3968                 ret = btrfs_alloc_chunk(trans, flags);
3969         }
3970
3971         if (!ret) {
3972                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
3973                                           &fs_info->chunk_block_rsv,
3974                                           thresh, BTRFS_RESERVE_NO_FLUSH);
3975                 if (!ret)
3976                         trans->chunk_bytes_reserved += thresh;
3977         }
3978 }
3979
3980 /*
3981  * If force is CHUNK_ALLOC_FORCE:
3982  *    - return 1 if it successfully allocates a chunk,
3983  *    - return errors including -ENOSPC otherwise.
3984  * If force is NOT CHUNK_ALLOC_FORCE:
3985  *    - return 0 if it doesn't need to allocate a new chunk,
3986  *    - return 1 if it successfully allocates a chunk,
3987  *    - return errors including -ENOSPC otherwise.
3988  */
3989 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3990                       enum btrfs_chunk_alloc_enum force)
3991 {
3992         struct btrfs_fs_info *fs_info = trans->fs_info;
3993         struct btrfs_space_info *space_info;
3994         bool wait_for_alloc = false;
3995         bool should_alloc = false;
3996         int ret = 0;
3997
3998         /* Don't re-enter if we're already allocating a chunk */
3999         if (trans->allocating_chunk)
4000                 return -ENOSPC;
4001
4002         space_info = btrfs_find_space_info(fs_info, flags);
4003         ASSERT(space_info);
4004
4005         do {
4006                 spin_lock(&space_info->lock);
4007                 if (force < space_info->force_alloc)
4008                         force = space_info->force_alloc;
4009                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4010                 if (space_info->full) {
4011                         /* No more free physical space */
4012                         if (should_alloc)
4013                                 ret = -ENOSPC;
4014                         else
4015                                 ret = 0;
4016                         spin_unlock(&space_info->lock);
4017                         return ret;
4018                 } else if (!should_alloc) {
4019                         spin_unlock(&space_info->lock);
4020                         return 0;
4021                 } else if (space_info->chunk_alloc) {
4022                         /*
4023                          * Someone is already allocating, so we need to block
4024                          * until this someone is finished and then loop to
4025                          * recheck if we should continue with our allocation
4026                          * attempt.
4027                          */
4028                         wait_for_alloc = true;
4029                         spin_unlock(&space_info->lock);
4030                         mutex_lock(&fs_info->chunk_mutex);
4031                         mutex_unlock(&fs_info->chunk_mutex);
4032                 } else {
4033                         /* Proceed with allocation */
4034                         space_info->chunk_alloc = 1;
4035                         wait_for_alloc = false;
4036                         spin_unlock(&space_info->lock);
4037                 }
4038
4039                 cond_resched();
4040         } while (wait_for_alloc);
4041
4042         mutex_lock(&fs_info->chunk_mutex);
4043         trans->allocating_chunk = true;
4044
4045         /*
4046          * If we have mixed data/metadata chunks we want to make sure we keep
4047          * allocating mixed chunks instead of individual chunks.
4048          */
4049         if (btrfs_mixed_space_info(space_info))
4050                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4051
4052         /*
4053          * if we're doing a data chunk, go ahead and make sure that
4054          * we keep a reasonable number of metadata chunks allocated in the
4055          * FS as well.
4056          */
4057         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4058                 fs_info->data_chunk_allocations++;
4059                 if (!(fs_info->data_chunk_allocations %
4060                       fs_info->metadata_ratio))
4061                         force_metadata_allocation(fs_info);
4062         }
4063
4064         /*
4065          * Check if we have enough space in SYSTEM chunk because we may need
4066          * to update devices.
4067          */
4068         check_system_chunk(trans, flags);
4069
4070         ret = btrfs_alloc_chunk(trans, flags);
4071         trans->allocating_chunk = false;
4072
4073         spin_lock(&space_info->lock);
4074         if (ret < 0) {
4075                 if (ret == -ENOSPC)
4076                         space_info->full = 1;
4077                 else
4078                         goto out;
4079         } else {
4080                 ret = 1;
4081                 space_info->max_extent_size = 0;
4082         }
4083
4084         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4085 out:
4086         space_info->chunk_alloc = 0;
4087         spin_unlock(&space_info->lock);
4088         mutex_unlock(&fs_info->chunk_mutex);
4089         /*
4090          * When we allocate a new chunk we reserve space in the chunk block
4091          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4092          * add new nodes/leafs to it if we end up needing to do it when
4093          * inserting the chunk item and updating device items as part of the
4094          * second phase of chunk allocation, performed by
4095          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4096          * large number of new block groups to create in our transaction
4097          * handle's new_bgs list to avoid exhausting the chunk block reserve
4098          * in extreme cases - like having a single transaction create many new
4099          * block groups when starting to write out the free space caches of all
4100          * the block groups that were made dirty during the lifetime of the
4101          * transaction.
4102          */
4103         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4104                 btrfs_create_pending_block_groups(trans);
4105
4106         return ret;
4107 }
4108
4109 static int update_block_group(struct btrfs_trans_handle *trans,
4110                               u64 bytenr, u64 num_bytes, int alloc)
4111 {
4112         struct btrfs_fs_info *info = trans->fs_info;
4113         struct btrfs_block_group_cache *cache = NULL;
4114         u64 total = num_bytes;
4115         u64 old_val;
4116         u64 byte_in_group;
4117         int factor;
4118         int ret = 0;
4119
4120         /* block accounting for super block */
4121         spin_lock(&info->delalloc_root_lock);
4122         old_val = btrfs_super_bytes_used(info->super_copy);
4123         if (alloc)
4124                 old_val += num_bytes;
4125         else
4126                 old_val -= num_bytes;
4127         btrfs_set_super_bytes_used(info->super_copy, old_val);
4128         spin_unlock(&info->delalloc_root_lock);
4129
4130         while (total) {
4131                 cache = btrfs_lookup_block_group(info, bytenr);
4132                 if (!cache) {
4133                         ret = -ENOENT;
4134                         break;
4135                 }
4136                 factor = btrfs_bg_type_to_factor(cache->flags);
4137
4138                 /*
4139                  * If this block group has free space cache written out, we
4140                  * need to make sure to load it if we are removing space.  This
4141                  * is because we need the unpinning stage to actually add the
4142                  * space back to the block group, otherwise we will leak space.
4143                  */
4144                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4145                         cache_block_group(cache, 1);
4146
4147                 byte_in_group = bytenr - cache->key.objectid;
4148                 WARN_ON(byte_in_group > cache->key.offset);
4149
4150                 spin_lock(&cache->space_info->lock);
4151                 spin_lock(&cache->lock);
4152
4153                 if (btrfs_test_opt(info, SPACE_CACHE) &&
4154                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4155                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4156
4157                 old_val = btrfs_block_group_used(&cache->item);
4158                 num_bytes = min(total, cache->key.offset - byte_in_group);
4159                 if (alloc) {
4160                         old_val += num_bytes;
4161                         btrfs_set_block_group_used(&cache->item, old_val);
4162                         cache->reserved -= num_bytes;
4163                         cache->space_info->bytes_reserved -= num_bytes;
4164                         cache->space_info->bytes_used += num_bytes;
4165                         cache->space_info->disk_used += num_bytes * factor;
4166                         spin_unlock(&cache->lock);
4167                         spin_unlock(&cache->space_info->lock);
4168                 } else {
4169                         old_val -= num_bytes;
4170                         btrfs_set_block_group_used(&cache->item, old_val);
4171                         cache->pinned += num_bytes;
4172                         btrfs_space_info_update_bytes_pinned(info,
4173                                         cache->space_info, num_bytes);
4174                         cache->space_info->bytes_used -= num_bytes;
4175                         cache->space_info->disk_used -= num_bytes * factor;
4176                         spin_unlock(&cache->lock);
4177                         spin_unlock(&cache->space_info->lock);
4178
4179                         trace_btrfs_space_reservation(info, "pinned",
4180                                                       cache->space_info->flags,
4181                                                       num_bytes, 1);
4182                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
4183                                            num_bytes,
4184                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
4185                         set_extent_dirty(info->pinned_extents,
4186                                          bytenr, bytenr + num_bytes - 1,
4187                                          GFP_NOFS | __GFP_NOFAIL);
4188                 }
4189
4190                 spin_lock(&trans->transaction->dirty_bgs_lock);
4191                 if (list_empty(&cache->dirty_list)) {
4192                         list_add_tail(&cache->dirty_list,
4193                                       &trans->transaction->dirty_bgs);
4194                         trans->delayed_ref_updates++;
4195                         btrfs_get_block_group(cache);
4196                 }
4197                 spin_unlock(&trans->transaction->dirty_bgs_lock);
4198
4199                 /*
4200                  * No longer have used bytes in this block group, queue it for
4201                  * deletion. We do this after adding the block group to the
4202                  * dirty list to avoid races between cleaner kthread and space
4203                  * cache writeout.
4204                  */
4205                 if (!alloc && old_val == 0)
4206                         btrfs_mark_bg_unused(cache);
4207
4208                 btrfs_put_block_group(cache);
4209                 total -= num_bytes;
4210                 bytenr += num_bytes;
4211         }
4212
4213         /* Modified block groups are accounted for in the delayed_refs_rsv. */
4214         btrfs_update_delayed_refs_rsv(trans);
4215         return ret;
4216 }
4217
4218 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
4219 {
4220         struct btrfs_block_group_cache *cache;
4221         u64 bytenr;
4222
4223         spin_lock(&fs_info->block_group_cache_lock);
4224         bytenr = fs_info->first_logical_byte;
4225         spin_unlock(&fs_info->block_group_cache_lock);
4226
4227         if (bytenr < (u64)-1)
4228                 return bytenr;
4229
4230         cache = btrfs_lookup_first_block_group(fs_info, search_start);
4231         if (!cache)
4232                 return 0;
4233
4234         bytenr = cache->key.objectid;
4235         btrfs_put_block_group(cache);
4236
4237         return bytenr;
4238 }
4239
4240 static int pin_down_extent(struct btrfs_block_group_cache *cache,
4241                            u64 bytenr, u64 num_bytes, int reserved)
4242 {
4243         struct btrfs_fs_info *fs_info = cache->fs_info;
4244
4245         spin_lock(&cache->space_info->lock);
4246         spin_lock(&cache->lock);
4247         cache->pinned += num_bytes;
4248         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
4249                                              num_bytes);
4250         if (reserved) {
4251                 cache->reserved -= num_bytes;
4252                 cache->space_info->bytes_reserved -= num_bytes;
4253         }
4254         spin_unlock(&cache->lock);
4255         spin_unlock(&cache->space_info->lock);
4256
4257         trace_btrfs_space_reservation(fs_info, "pinned",
4258                                       cache->space_info->flags, num_bytes, 1);
4259         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
4260                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
4261         set_extent_dirty(fs_info->pinned_extents, bytenr,
4262                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4263         return 0;
4264 }
4265
4266 /*
4267  * this function must be called within transaction
4268  */
4269 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
4270                      u64 bytenr, u64 num_bytes, int reserved)
4271 {
4272         struct btrfs_block_group_cache *cache;
4273
4274         cache = btrfs_lookup_block_group(fs_info, bytenr);
4275         BUG_ON(!cache); /* Logic error */
4276
4277         pin_down_extent(cache, bytenr, num_bytes, reserved);
4278
4279         btrfs_put_block_group(cache);
4280         return 0;
4281 }
4282
4283 /*
4284  * this function must be called within transaction
4285  */
4286 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
4287                                     u64 bytenr, u64 num_bytes)
4288 {
4289         struct btrfs_block_group_cache *cache;
4290         int ret;
4291
4292         cache = btrfs_lookup_block_group(fs_info, bytenr);
4293         if (!cache)
4294                 return -EINVAL;
4295
4296         /*
4297          * pull in the free space cache (if any) so that our pin
4298          * removes the free space from the cache.  We have load_only set
4299          * to one because the slow code to read in the free extents does check
4300          * the pinned extents.
4301          */
4302         cache_block_group(cache, 1);
4303
4304         pin_down_extent(cache, bytenr, num_bytes, 0);
4305
4306         /* remove us from the free space cache (if we're there at all) */
4307         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
4308         btrfs_put_block_group(cache);
4309         return ret;
4310 }
4311
4312 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
4313                                    u64 start, u64 num_bytes)
4314 {
4315         int ret;
4316         struct btrfs_block_group_cache *block_group;
4317         struct btrfs_caching_control *caching_ctl;
4318
4319         block_group = btrfs_lookup_block_group(fs_info, start);
4320         if (!block_group)
4321                 return -EINVAL;
4322
4323         cache_block_group(block_group, 0);
4324         caching_ctl = get_caching_control(block_group);
4325
4326         if (!caching_ctl) {
4327                 /* Logic error */
4328                 BUG_ON(!block_group_cache_done(block_group));
4329                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
4330         } else {
4331                 mutex_lock(&caching_ctl->mutex);
4332
4333                 if (start >= caching_ctl->progress) {
4334                         ret = add_excluded_extent(fs_info, start, num_bytes);
4335                 } else if (start + num_bytes <= caching_ctl->progress) {
4336                         ret = btrfs_remove_free_space(block_group,
4337                                                       start, num_bytes);
4338                 } else {
4339                         num_bytes = caching_ctl->progress - start;
4340                         ret = btrfs_remove_free_space(block_group,
4341                                                       start, num_bytes);
4342                         if (ret)
4343                                 goto out_lock;
4344
4345                         num_bytes = (start + num_bytes) -
4346                                 caching_ctl->progress;
4347                         start = caching_ctl->progress;
4348                         ret = add_excluded_extent(fs_info, start, num_bytes);
4349                 }
4350 out_lock:
4351                 mutex_unlock(&caching_ctl->mutex);
4352                 put_caching_control(caching_ctl);
4353         }
4354         btrfs_put_block_group(block_group);
4355         return ret;
4356 }
4357
4358 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
4359 {
4360         struct btrfs_fs_info *fs_info = eb->fs_info;
4361         struct btrfs_file_extent_item *item;
4362         struct btrfs_key key;
4363         int found_type;
4364         int i;
4365         int ret = 0;
4366
4367         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
4368                 return 0;
4369
4370         for (i = 0; i < btrfs_header_nritems(eb); i++) {
4371                 btrfs_item_key_to_cpu(eb, &key, i);
4372                 if (key.type != BTRFS_EXTENT_DATA_KEY)
4373                         continue;
4374                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
4375                 found_type = btrfs_file_extent_type(eb, item);
4376                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
4377                         continue;
4378                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
4379                         continue;
4380                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
4381                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
4382                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
4383                 if (ret)
4384                         break;
4385         }
4386
4387         return ret;
4388 }
4389
4390 static void
4391 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
4392 {
4393         atomic_inc(&bg->reservations);
4394 }
4395
4396 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
4397                                         const u64 start)
4398 {
4399         struct btrfs_block_group_cache *bg;
4400
4401         bg = btrfs_lookup_block_group(fs_info, start);
4402         ASSERT(bg);
4403         if (atomic_dec_and_test(&bg->reservations))
4404                 wake_up_var(&bg->reservations);
4405         btrfs_put_block_group(bg);
4406 }
4407
4408 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
4409 {
4410         struct btrfs_space_info *space_info = bg->space_info;
4411
4412         ASSERT(bg->ro);
4413
4414         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
4415                 return;
4416
4417         /*
4418          * Our block group is read only but before we set it to read only,
4419          * some task might have had allocated an extent from it already, but it
4420          * has not yet created a respective ordered extent (and added it to a
4421          * root's list of ordered extents).
4422          * Therefore wait for any task currently allocating extents, since the
4423          * block group's reservations counter is incremented while a read lock
4424          * on the groups' semaphore is held and decremented after releasing
4425          * the read access on that semaphore and creating the ordered extent.
4426          */
4427         down_write(&space_info->groups_sem);
4428         up_write(&space_info->groups_sem);
4429
4430         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
4431 }
4432
4433 /**
4434  * btrfs_add_reserved_bytes - update the block_group and space info counters
4435  * @cache:      The cache we are manipulating
4436  * @ram_bytes:  The number of bytes of file content, and will be same to
4437  *              @num_bytes except for the compress path.
4438  * @num_bytes:  The number of bytes in question
4439  * @delalloc:   The blocks are allocated for the delalloc write
4440  *
4441  * This is called by the allocator when it reserves space. If this is a
4442  * reservation and the block group has become read only we cannot make the
4443  * reservation and return -EAGAIN, otherwise this function always succeeds.
4444  */
4445 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
4446                                     u64 ram_bytes, u64 num_bytes, int delalloc)
4447 {
4448         struct btrfs_space_info *space_info = cache->space_info;
4449         int ret = 0;
4450
4451         spin_lock(&space_info->lock);
4452         spin_lock(&cache->lock);
4453         if (cache->ro) {
4454                 ret = -EAGAIN;
4455         } else {
4456                 cache->reserved += num_bytes;
4457                 space_info->bytes_reserved += num_bytes;
4458                 btrfs_space_info_update_bytes_may_use(cache->fs_info,
4459                                                       space_info, -ram_bytes);
4460                 if (delalloc)
4461                         cache->delalloc_bytes += num_bytes;
4462         }
4463         spin_unlock(&cache->lock);
4464         spin_unlock(&space_info->lock);
4465         return ret;
4466 }
4467
4468 /**
4469  * btrfs_free_reserved_bytes - update the block_group and space info counters
4470  * @cache:      The cache we are manipulating
4471  * @num_bytes:  The number of bytes in question
4472  * @delalloc:   The blocks are allocated for the delalloc write
4473  *
4474  * This is called by somebody who is freeing space that was never actually used
4475  * on disk.  For example if you reserve some space for a new leaf in transaction
4476  * A and before transaction A commits you free that leaf, you call this with
4477  * reserve set to 0 in order to clear the reservation.
4478  */
4479
4480 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
4481                                       u64 num_bytes, int delalloc)
4482 {
4483         struct btrfs_space_info *space_info = cache->space_info;
4484
4485         spin_lock(&space_info->lock);
4486         spin_lock(&cache->lock);
4487         if (cache->ro)
4488                 space_info->bytes_readonly += num_bytes;
4489         cache->reserved -= num_bytes;
4490         space_info->bytes_reserved -= num_bytes;
4491         space_info->max_extent_size = 0;
4492
4493         if (delalloc)
4494                 cache->delalloc_bytes -= num_bytes;
4495         spin_unlock(&cache->lock);
4496         spin_unlock(&space_info->lock);
4497 }
4498 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
4499 {
4500         struct btrfs_caching_control *next;
4501         struct btrfs_caching_control *caching_ctl;
4502         struct btrfs_block_group_cache *cache;
4503
4504         down_write(&fs_info->commit_root_sem);
4505
4506         list_for_each_entry_safe(caching_ctl, next,
4507                                  &fs_info->caching_block_groups, list) {
4508                 cache = caching_ctl->block_group;
4509                 if (block_group_cache_done(cache)) {
4510                         cache->last_byte_to_unpin = (u64)-1;
4511                         list_del_init(&caching_ctl->list);
4512                         put_caching_control(caching_ctl);
4513                 } else {
4514                         cache->last_byte_to_unpin = caching_ctl->progress;
4515                 }
4516         }
4517
4518         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4519                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4520         else
4521                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4522
4523         up_write(&fs_info->commit_root_sem);
4524
4525         btrfs_update_global_block_rsv(fs_info);
4526 }
4527
4528 /*
4529  * Returns the free cluster for the given space info and sets empty_cluster to
4530  * what it should be based on the mount options.
4531  */
4532 static struct btrfs_free_cluster *
4533 fetch_cluster_info(struct btrfs_fs_info *fs_info,
4534                    struct btrfs_space_info *space_info, u64 *empty_cluster)
4535 {
4536         struct btrfs_free_cluster *ret = NULL;
4537
4538         *empty_cluster = 0;
4539         if (btrfs_mixed_space_info(space_info))
4540                 return ret;
4541
4542         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4543                 ret = &fs_info->meta_alloc_cluster;
4544                 if (btrfs_test_opt(fs_info, SSD))
4545                         *empty_cluster = SZ_2M;
4546                 else
4547                         *empty_cluster = SZ_64K;
4548         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
4549                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
4550                 *empty_cluster = SZ_2M;
4551                 ret = &fs_info->data_alloc_cluster;
4552         }
4553
4554         return ret;
4555 }
4556
4557 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
4558                               u64 start, u64 end,
4559                               const bool return_free_space)
4560 {
4561         struct btrfs_block_group_cache *cache = NULL;
4562         struct btrfs_space_info *space_info;
4563         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4564         struct btrfs_free_cluster *cluster = NULL;
4565         u64 len;
4566         u64 total_unpinned = 0;
4567         u64 empty_cluster = 0;
4568         bool readonly;
4569
4570         while (start <= end) {
4571                 readonly = false;
4572                 if (!cache ||
4573                     start >= cache->key.objectid + cache->key.offset) {
4574                         if (cache)
4575                                 btrfs_put_block_group(cache);
4576                         total_unpinned = 0;
4577                         cache = btrfs_lookup_block_group(fs_info, start);
4578                         BUG_ON(!cache); /* Logic error */
4579
4580                         cluster = fetch_cluster_info(fs_info,
4581                                                      cache->space_info,
4582                                                      &empty_cluster);
4583                         empty_cluster <<= 1;
4584                 }
4585
4586                 len = cache->key.objectid + cache->key.offset - start;
4587                 len = min(len, end + 1 - start);
4588
4589                 if (start < cache->last_byte_to_unpin) {
4590                         len = min(len, cache->last_byte_to_unpin - start);
4591                         if (return_free_space)
4592                                 btrfs_add_free_space(cache, start, len);
4593                 }
4594
4595                 start += len;
4596                 total_unpinned += len;
4597                 space_info = cache->space_info;
4598
4599                 /*
4600                  * If this space cluster has been marked as fragmented and we've
4601                  * unpinned enough in this block group to potentially allow a
4602                  * cluster to be created inside of it go ahead and clear the
4603                  * fragmented check.
4604                  */
4605                 if (cluster && cluster->fragmented &&
4606                     total_unpinned > empty_cluster) {
4607                         spin_lock(&cluster->lock);
4608                         cluster->fragmented = 0;
4609                         spin_unlock(&cluster->lock);
4610                 }
4611
4612                 spin_lock(&space_info->lock);
4613                 spin_lock(&cache->lock);
4614                 cache->pinned -= len;
4615                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
4616
4617                 trace_btrfs_space_reservation(fs_info, "pinned",
4618                                               space_info->flags, len, 0);
4619                 space_info->max_extent_size = 0;
4620                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
4621                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
4622                 if (cache->ro) {
4623                         space_info->bytes_readonly += len;
4624                         readonly = true;
4625                 }
4626                 spin_unlock(&cache->lock);
4627                 if (!readonly && return_free_space &&
4628                     global_rsv->space_info == space_info) {
4629                         u64 to_add = len;
4630
4631                         spin_lock(&global_rsv->lock);
4632                         if (!global_rsv->full) {
4633                                 to_add = min(len, global_rsv->size -
4634                                              global_rsv->reserved);
4635                                 global_rsv->reserved += to_add;
4636                                 btrfs_space_info_update_bytes_may_use(fs_info,
4637                                                 space_info, to_add);
4638                                 if (global_rsv->reserved >= global_rsv->size)
4639                                         global_rsv->full = 1;
4640                                 trace_btrfs_space_reservation(fs_info,
4641                                                               "space_info",
4642                                                               space_info->flags,
4643                                                               to_add, 1);
4644                                 len -= to_add;
4645                         }
4646                         spin_unlock(&global_rsv->lock);
4647                         /* Add to any tickets we may have */
4648                         if (len)
4649                                 btrfs_space_info_add_new_bytes(fs_info,
4650                                                 space_info, len);
4651                 }
4652                 spin_unlock(&space_info->lock);
4653         }
4654
4655         if (cache)
4656                 btrfs_put_block_group(cache);
4657         return 0;
4658 }
4659
4660 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
4661 {
4662         struct btrfs_fs_info *fs_info = trans->fs_info;
4663         struct btrfs_block_group_cache *block_group, *tmp;
4664         struct list_head *deleted_bgs;
4665         struct extent_io_tree *unpin;
4666         u64 start;
4667         u64 end;
4668         int ret;
4669
4670         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4671                 unpin = &fs_info->freed_extents[1];
4672         else
4673                 unpin = &fs_info->freed_extents[0];
4674
4675         while (!trans->aborted) {
4676                 struct extent_state *cached_state = NULL;
4677
4678                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
4679                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4680                                             EXTENT_DIRTY, &cached_state);
4681                 if (ret) {
4682                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
4683                         break;
4684                 }
4685
4686                 if (btrfs_test_opt(fs_info, DISCARD))
4687                         ret = btrfs_discard_extent(fs_info, start,
4688                                                    end + 1 - start, NULL);
4689
4690                 clear_extent_dirty(unpin, start, end, &cached_state);
4691                 unpin_extent_range(fs_info, start, end, true);
4692                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
4693                 free_extent_state(cached_state);
4694                 cond_resched();
4695         }
4696
4697         /*
4698          * Transaction is finished.  We don't need the lock anymore.  We
4699          * do need to clean up the block groups in case of a transaction
4700          * abort.
4701          */
4702         deleted_bgs = &trans->transaction->deleted_bgs;
4703         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
4704                 u64 trimmed = 0;
4705
4706                 ret = -EROFS;
4707                 if (!trans->aborted)
4708                         ret = btrfs_discard_extent(fs_info,
4709                                                    block_group->key.objectid,
4710                                                    block_group->key.offset,
4711                                                    &trimmed);
4712
4713                 list_del_init(&block_group->bg_list);
4714                 btrfs_put_block_group_trimming(block_group);
4715                 btrfs_put_block_group(block_group);
4716
4717                 if (ret) {
4718                         const char *errstr = btrfs_decode_error(ret);
4719                         btrfs_warn(fs_info,
4720                            "discard failed while removing blockgroup: errno=%d %s",
4721                                    ret, errstr);
4722                 }
4723         }
4724
4725         return 0;
4726 }
4727
4728 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4729                                struct btrfs_delayed_ref_node *node, u64 parent,
4730                                u64 root_objectid, u64 owner_objectid,
4731                                u64 owner_offset, int refs_to_drop,
4732                                struct btrfs_delayed_extent_op *extent_op)
4733 {
4734         struct btrfs_fs_info *info = trans->fs_info;
4735         struct btrfs_key key;
4736         struct btrfs_path *path;
4737         struct btrfs_root *extent_root = info->extent_root;
4738         struct extent_buffer *leaf;
4739         struct btrfs_extent_item *ei;
4740         struct btrfs_extent_inline_ref *iref;
4741         int ret;
4742         int is_data;
4743         int extent_slot = 0;
4744         int found_extent = 0;
4745         int num_to_del = 1;
4746         u32 item_size;
4747         u64 refs;
4748         u64 bytenr = node->bytenr;
4749         u64 num_bytes = node->num_bytes;
4750         int last_ref = 0;
4751         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
4752
4753         path = btrfs_alloc_path();
4754         if (!path)
4755                 return -ENOMEM;
4756
4757         path->reada = READA_FORWARD;
4758         path->leave_spinning = 1;
4759
4760         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4761         BUG_ON(!is_data && refs_to_drop != 1);
4762
4763         if (is_data)
4764                 skinny_metadata = false;
4765
4766         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
4767                                     parent, root_objectid, owner_objectid,
4768                                     owner_offset);
4769         if (ret == 0) {
4770                 extent_slot = path->slots[0];
4771                 while (extent_slot >= 0) {
4772                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4773                                               extent_slot);
4774                         if (key.objectid != bytenr)
4775                                 break;
4776                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4777                             key.offset == num_bytes) {
4778                                 found_extent = 1;
4779                                 break;
4780                         }
4781                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
4782                             key.offset == owner_objectid) {
4783                                 found_extent = 1;
4784                                 break;
4785                         }
4786                         if (path->slots[0] - extent_slot > 5)
4787                                 break;
4788                         extent_slot--;
4789                 }
4790
4791                 if (!found_extent) {
4792                         BUG_ON(iref);
4793                         ret = remove_extent_backref(trans, path, NULL,
4794                                                     refs_to_drop,
4795                                                     is_data, &last_ref);
4796                         if (ret) {
4797                                 btrfs_abort_transaction(trans, ret);
4798                                 goto out;
4799                         }
4800                         btrfs_release_path(path);
4801                         path->leave_spinning = 1;
4802
4803                         key.objectid = bytenr;
4804                         key.type = BTRFS_EXTENT_ITEM_KEY;
4805                         key.offset = num_bytes;
4806
4807                         if (!is_data && skinny_metadata) {
4808                                 key.type = BTRFS_METADATA_ITEM_KEY;
4809                                 key.offset = owner_objectid;
4810                         }
4811
4812                         ret = btrfs_search_slot(trans, extent_root,
4813                                                 &key, path, -1, 1);
4814                         if (ret > 0 && skinny_metadata && path->slots[0]) {
4815                                 /*
4816                                  * Couldn't find our skinny metadata item,
4817                                  * see if we have ye olde extent item.
4818                                  */
4819                                 path->slots[0]--;
4820                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
4821                                                       path->slots[0]);
4822                                 if (key.objectid == bytenr &&
4823                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
4824                                     key.offset == num_bytes)
4825                                         ret = 0;
4826                         }
4827
4828                         if (ret > 0 && skinny_metadata) {
4829                                 skinny_metadata = false;
4830                                 key.objectid = bytenr;
4831                                 key.type = BTRFS_EXTENT_ITEM_KEY;
4832                                 key.offset = num_bytes;
4833                                 btrfs_release_path(path);
4834                                 ret = btrfs_search_slot(trans, extent_root,
4835                                                         &key, path, -1, 1);
4836                         }
4837
4838                         if (ret) {
4839                                 btrfs_err(info,
4840                                           "umm, got %d back from search, was looking for %llu",
4841                                           ret, bytenr);
4842                                 if (ret > 0)
4843                                         btrfs_print_leaf(path->nodes[0]);
4844                         }
4845                         if (ret < 0) {
4846                                 btrfs_abort_transaction(trans, ret);
4847                                 goto out;
4848                         }
4849                         extent_slot = path->slots[0];
4850                 }
4851         } else if (WARN_ON(ret == -ENOENT)) {
4852                 btrfs_print_leaf(path->nodes[0]);
4853                 btrfs_err(info,
4854                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
4855                         bytenr, parent, root_objectid, owner_objectid,
4856                         owner_offset);
4857                 btrfs_abort_transaction(trans, ret);
4858                 goto out;
4859         } else {
4860                 btrfs_abort_transaction(trans, ret);
4861                 goto out;
4862         }
4863
4864         leaf = path->nodes[0];
4865         item_size = btrfs_item_size_nr(leaf, extent_slot);
4866         if (unlikely(item_size < sizeof(*ei))) {
4867                 ret = -EINVAL;
4868                 btrfs_print_v0_err(info);
4869                 btrfs_abort_transaction(trans, ret);
4870                 goto out;
4871         }
4872         ei = btrfs_item_ptr(leaf, extent_slot,
4873                             struct btrfs_extent_item);
4874         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
4875             key.type == BTRFS_EXTENT_ITEM_KEY) {
4876                 struct btrfs_tree_block_info *bi;
4877                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4878                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4879                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4880         }
4881
4882         refs = btrfs_extent_refs(leaf, ei);
4883         if (refs < refs_to_drop) {
4884                 btrfs_err(info,
4885                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
4886                           refs_to_drop, refs, bytenr);
4887                 ret = -EINVAL;
4888                 btrfs_abort_transaction(trans, ret);
4889                 goto out;
4890         }
4891         refs -= refs_to_drop;
4892
4893         if (refs > 0) {
4894                 if (extent_op)
4895                         __run_delayed_extent_op(extent_op, leaf, ei);
4896                 /*
4897                  * In the case of inline back ref, reference count will
4898                  * be updated by remove_extent_backref
4899                  */
4900                 if (iref) {
4901                         BUG_ON(!found_extent);
4902                 } else {
4903                         btrfs_set_extent_refs(leaf, ei, refs);
4904                         btrfs_mark_buffer_dirty(leaf);
4905                 }
4906                 if (found_extent) {
4907                         ret = remove_extent_backref(trans, path, iref,
4908                                                     refs_to_drop, is_data,
4909                                                     &last_ref);
4910                         if (ret) {
4911                                 btrfs_abort_transaction(trans, ret);
4912                                 goto out;
4913                         }
4914                 }
4915         } else {
4916                 if (found_extent) {
4917                         BUG_ON(is_data && refs_to_drop !=
4918                                extent_data_ref_count(path, iref));
4919                         if (iref) {
4920                                 BUG_ON(path->slots[0] != extent_slot);
4921                         } else {
4922                                 BUG_ON(path->slots[0] != extent_slot + 1);
4923                                 path->slots[0] = extent_slot;
4924                                 num_to_del = 2;
4925                         }
4926                 }
4927
4928                 last_ref = 1;
4929                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4930                                       num_to_del);
4931                 if (ret) {
4932                         btrfs_abort_transaction(trans, ret);
4933                         goto out;
4934                 }
4935                 btrfs_release_path(path);
4936
4937                 if (is_data) {
4938                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
4939                         if (ret) {
4940                                 btrfs_abort_transaction(trans, ret);
4941                                 goto out;
4942                         }
4943                 }
4944
4945                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
4946                 if (ret) {
4947                         btrfs_abort_transaction(trans, ret);
4948                         goto out;
4949                 }
4950
4951                 ret = update_block_group(trans, bytenr, num_bytes, 0);
4952                 if (ret) {
4953                         btrfs_abort_transaction(trans, ret);
4954                         goto out;
4955                 }
4956         }
4957         btrfs_release_path(path);
4958
4959 out:
4960         btrfs_free_path(path);
4961         return ret;
4962 }
4963
4964 /*
4965  * when we free an block, it is possible (and likely) that we free the last
4966  * delayed ref for that extent as well.  This searches the delayed ref tree for
4967  * a given extent, and if there are no other delayed refs to be processed, it
4968  * removes it from the tree.
4969  */
4970 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4971                                       u64 bytenr)
4972 {
4973         struct btrfs_delayed_ref_head *head;
4974         struct btrfs_delayed_ref_root *delayed_refs;
4975         int ret = 0;
4976
4977         delayed_refs = &trans->transaction->delayed_refs;
4978         spin_lock(&delayed_refs->lock);
4979         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
4980         if (!head)
4981                 goto out_delayed_unlock;
4982
4983         spin_lock(&head->lock);
4984         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
4985                 goto out;
4986
4987         if (cleanup_extent_op(head) != NULL)
4988                 goto out;
4989
4990         /*
4991          * waiting for the lock here would deadlock.  If someone else has it
4992          * locked they are already in the process of dropping it anyway
4993          */
4994         if (!mutex_trylock(&head->mutex))
4995                 goto out;
4996
4997         btrfs_delete_ref_head(delayed_refs, head);
4998         head->processing = 0;
4999
5000         spin_unlock(&head->lock);
5001         spin_unlock(&delayed_refs->lock);
5002
5003         BUG_ON(head->extent_op);
5004         if (head->must_insert_reserved)
5005                 ret = 1;
5006
5007         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
5008         mutex_unlock(&head->mutex);
5009         btrfs_put_delayed_ref_head(head);
5010         return ret;
5011 out:
5012         spin_unlock(&head->lock);
5013
5014 out_delayed_unlock:
5015         spin_unlock(&delayed_refs->lock);
5016         return 0;
5017 }
5018
5019 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5020                            struct btrfs_root *root,
5021                            struct extent_buffer *buf,
5022                            u64 parent, int last_ref)
5023 {
5024         struct btrfs_fs_info *fs_info = root->fs_info;
5025         struct btrfs_ref generic_ref = { 0 };
5026         int pin = 1;
5027         int ret;
5028
5029         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
5030                                buf->start, buf->len, parent);
5031         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
5032                             root->root_key.objectid);
5033
5034         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5035                 int old_ref_mod, new_ref_mod;
5036
5037                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5038                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL,
5039                                                  &old_ref_mod, &new_ref_mod);
5040                 BUG_ON(ret); /* -ENOMEM */
5041                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
5042         }
5043
5044         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
5045                 struct btrfs_block_group_cache *cache;
5046
5047                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5048                         ret = check_ref_cleanup(trans, buf->start);
5049                         if (!ret)
5050                                 goto out;
5051                 }
5052
5053                 pin = 0;
5054                 cache = btrfs_lookup_block_group(fs_info, buf->start);
5055
5056                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5057                         pin_down_extent(cache, buf->start, buf->len, 1);
5058                         btrfs_put_block_group(cache);
5059                         goto out;
5060                 }
5061
5062                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5063
5064                 btrfs_add_free_space(cache, buf->start, buf->len);
5065                 btrfs_free_reserved_bytes(cache, buf->len, 0);
5066                 btrfs_put_block_group(cache);
5067                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
5068         }
5069 out:
5070         if (pin)
5071                 add_pinned_bytes(fs_info, &generic_ref);
5072
5073         if (last_ref) {
5074                 /*
5075                  * Deleting the buffer, clear the corrupt flag since it doesn't
5076                  * matter anymore.
5077                  */
5078                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5079         }
5080 }
5081
5082 /* Can return -ENOMEM */
5083 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
5084 {
5085         struct btrfs_fs_info *fs_info = trans->fs_info;
5086         int old_ref_mod, new_ref_mod;
5087         int ret;
5088
5089         if (btrfs_is_testing(fs_info))
5090                 return 0;
5091
5092         /*
5093          * tree log blocks never actually go into the extent allocation
5094          * tree, just update pinning info and exit early.
5095          */
5096         if ((ref->type == BTRFS_REF_METADATA &&
5097              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
5098             (ref->type == BTRFS_REF_DATA &&
5099              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
5100                 /* unlocks the pinned mutex */
5101                 btrfs_pin_extent(fs_info, ref->bytenr, ref->len, 1);
5102                 old_ref_mod = new_ref_mod = 0;
5103                 ret = 0;
5104         } else if (ref->type == BTRFS_REF_METADATA) {
5105                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL,
5106                                                  &old_ref_mod, &new_ref_mod);
5107         } else {
5108                 ret = btrfs_add_delayed_data_ref(trans, ref, 0,
5109                                                  &old_ref_mod, &new_ref_mod);
5110         }
5111
5112         if (!((ref->type == BTRFS_REF_METADATA &&
5113                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
5114               (ref->type == BTRFS_REF_DATA &&
5115                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
5116                 btrfs_ref_tree_mod(fs_info, ref);
5117
5118         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
5119                 add_pinned_bytes(fs_info, ref);
5120
5121         return ret;
5122 }
5123
5124 /*
5125  * when we wait for progress in the block group caching, its because
5126  * our allocation attempt failed at least once.  So, we must sleep
5127  * and let some progress happen before we try again.
5128  *
5129  * This function will sleep at least once waiting for new free space to
5130  * show up, and then it will check the block group free space numbers
5131  * for our min num_bytes.  Another option is to have it go ahead
5132  * and look in the rbtree for a free extent of a given size, but this
5133  * is a good start.
5134  *
5135  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
5136  * any of the information in this block group.
5137  */
5138 static noinline void
5139 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5140                                 u64 num_bytes)
5141 {
5142         struct btrfs_caching_control *caching_ctl;
5143
5144         caching_ctl = get_caching_control(cache);
5145         if (!caching_ctl)
5146                 return;
5147
5148         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5149                    (cache->free_space_ctl->free_space >= num_bytes));
5150
5151         put_caching_control(caching_ctl);
5152 }
5153
5154 static noinline int
5155 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5156 {
5157         struct btrfs_caching_control *caching_ctl;
5158         int ret = 0;
5159
5160         caching_ctl = get_caching_control(cache);
5161         if (!caching_ctl)
5162                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
5163
5164         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5165         if (cache->cached == BTRFS_CACHE_ERROR)
5166                 ret = -EIO;
5167         put_caching_control(caching_ctl);
5168         return ret;
5169 }
5170
5171 enum btrfs_loop_type {
5172         LOOP_CACHING_NOWAIT,
5173         LOOP_CACHING_WAIT,
5174         LOOP_ALLOC_CHUNK,
5175         LOOP_NO_EMPTY_SIZE,
5176 };
5177
5178 static inline void
5179 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
5180                        int delalloc)
5181 {
5182         if (delalloc)
5183                 down_read(&cache->data_rwsem);
5184 }
5185
5186 static inline void
5187 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
5188                        int delalloc)
5189 {
5190         btrfs_get_block_group(cache);
5191         if (delalloc)
5192                 down_read(&cache->data_rwsem);
5193 }
5194
5195 static struct btrfs_block_group_cache *
5196 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
5197                    struct btrfs_free_cluster *cluster,
5198                    int delalloc)
5199 {
5200         struct btrfs_block_group_cache *used_bg = NULL;
5201
5202         spin_lock(&cluster->refill_lock);
5203         while (1) {
5204                 used_bg = cluster->block_group;
5205                 if (!used_bg)
5206                         return NULL;
5207
5208                 if (used_bg == block_group)
5209                         return used_bg;
5210
5211                 btrfs_get_block_group(used_bg);
5212
5213                 if (!delalloc)
5214                         return used_bg;
5215
5216                 if (down_read_trylock(&used_bg->data_rwsem))
5217                         return used_bg;
5218
5219                 spin_unlock(&cluster->refill_lock);
5220
5221                 /* We should only have one-level nested. */
5222                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
5223
5224                 spin_lock(&cluster->refill_lock);
5225                 if (used_bg == cluster->block_group)
5226                         return used_bg;
5227
5228                 up_read(&used_bg->data_rwsem);
5229                 btrfs_put_block_group(used_bg);
5230         }
5231 }
5232
5233 static inline void
5234 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
5235                          int delalloc)
5236 {
5237         if (delalloc)
5238                 up_read(&cache->data_rwsem);
5239         btrfs_put_block_group(cache);
5240 }
5241
5242 /*
5243  * Structure used internally for find_free_extent() function.  Wraps needed
5244  * parameters.
5245  */
5246 struct find_free_extent_ctl {
5247         /* Basic allocation info */
5248         u64 ram_bytes;
5249         u64 num_bytes;
5250         u64 empty_size;
5251         u64 flags;
5252         int delalloc;
5253
5254         /* Where to start the search inside the bg */
5255         u64 search_start;
5256
5257         /* For clustered allocation */
5258         u64 empty_cluster;
5259
5260         bool have_caching_bg;
5261         bool orig_have_caching_bg;
5262
5263         /* RAID index, converted from flags */
5264         int index;
5265
5266         /*
5267          * Current loop number, check find_free_extent_update_loop() for details
5268          */
5269         int loop;
5270
5271         /*
5272          * Whether we're refilling a cluster, if true we need to re-search
5273          * current block group but don't try to refill the cluster again.
5274          */
5275         bool retry_clustered;
5276
5277         /*
5278          * Whether we're updating free space cache, if true we need to re-search
5279          * current block group but don't try updating free space cache again.
5280          */
5281         bool retry_unclustered;
5282
5283         /* If current block group is cached */
5284         int cached;
5285
5286         /* Max contiguous hole found */
5287         u64 max_extent_size;
5288
5289         /* Total free space from free space cache, not always contiguous */
5290         u64 total_free_space;
5291
5292         /* Found result */
5293         u64 found_offset;
5294 };
5295
5296
5297 /*
5298  * Helper function for find_free_extent().
5299  *
5300  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
5301  * Return -EAGAIN to inform caller that we need to re-search this block group
5302  * Return >0 to inform caller that we find nothing
5303  * Return 0 means we have found a location and set ffe_ctl->found_offset.
5304  */
5305 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
5306                 struct btrfs_free_cluster *last_ptr,
5307                 struct find_free_extent_ctl *ffe_ctl,
5308                 struct btrfs_block_group_cache **cluster_bg_ret)
5309 {
5310         struct btrfs_block_group_cache *cluster_bg;
5311         u64 aligned_cluster;
5312         u64 offset;
5313         int ret;
5314
5315         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
5316         if (!cluster_bg)
5317                 goto refill_cluster;
5318         if (cluster_bg != bg && (cluster_bg->ro ||
5319             !block_group_bits(cluster_bg, ffe_ctl->flags)))
5320                 goto release_cluster;
5321
5322         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
5323                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
5324                         &ffe_ctl->max_extent_size);
5325         if (offset) {
5326                 /* We have a block, we're done */
5327                 spin_unlock(&last_ptr->refill_lock);
5328                 trace_btrfs_reserve_extent_cluster(cluster_bg,
5329                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
5330                 *cluster_bg_ret = cluster_bg;
5331                 ffe_ctl->found_offset = offset;
5332                 return 0;
5333         }
5334         WARN_ON(last_ptr->block_group != cluster_bg);
5335
5336 release_cluster:
5337         /*
5338          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
5339          * lets just skip it and let the allocator find whatever block it can
5340          * find. If we reach this point, we will have tried the cluster
5341          * allocator plenty of times and not have found anything, so we are
5342          * likely way too fragmented for the clustering stuff to find anything.
5343          *
5344          * However, if the cluster is taken from the current block group,
5345          * release the cluster first, so that we stand a better chance of
5346          * succeeding in the unclustered allocation.
5347          */
5348         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
5349                 spin_unlock(&last_ptr->refill_lock);
5350                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
5351                 return -ENOENT;
5352         }
5353
5354         /* This cluster didn't work out, free it and start over */
5355         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5356
5357         if (cluster_bg != bg)
5358                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
5359
5360 refill_cluster:
5361         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
5362                 spin_unlock(&last_ptr->refill_lock);
5363                 return -ENOENT;
5364         }
5365
5366         aligned_cluster = max_t(u64,
5367                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
5368                         bg->full_stripe_len);
5369         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
5370                         ffe_ctl->num_bytes, aligned_cluster);
5371         if (ret == 0) {
5372                 /* Now pull our allocation out of this cluster */
5373                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
5374                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
5375                                 &ffe_ctl->max_extent_size);
5376                 if (offset) {
5377                         /* We found one, proceed */
5378                         spin_unlock(&last_ptr->refill_lock);
5379                         trace_btrfs_reserve_extent_cluster(bg,
5380                                         ffe_ctl->search_start,
5381                                         ffe_ctl->num_bytes);
5382                         ffe_ctl->found_offset = offset;
5383                         return 0;
5384                 }
5385         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
5386                    !ffe_ctl->retry_clustered) {
5387                 spin_unlock(&last_ptr->refill_lock);
5388
5389                 ffe_ctl->retry_clustered = true;
5390                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
5391                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
5392                 return -EAGAIN;
5393         }
5394         /*
5395          * At this point we either didn't find a cluster or we weren't able to
5396          * allocate a block from our cluster.  Free the cluster we've been
5397          * trying to use, and go to the next block group.
5398          */
5399         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5400         spin_unlock(&last_ptr->refill_lock);
5401         return 1;
5402 }
5403
5404 /*
5405  * Return >0 to inform caller that we find nothing
5406  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
5407  * Return -EAGAIN to inform caller that we need to re-search this block group
5408  */
5409 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
5410                 struct btrfs_free_cluster *last_ptr,
5411                 struct find_free_extent_ctl *ffe_ctl)
5412 {
5413         u64 offset;
5414
5415         /*
5416          * We are doing an unclustered allocation, set the fragmented flag so
5417          * we don't bother trying to setup a cluster again until we get more
5418          * space.
5419          */
5420         if (unlikely(last_ptr)) {
5421                 spin_lock(&last_ptr->lock);
5422                 last_ptr->fragmented = 1;
5423                 spin_unlock(&last_ptr->lock);
5424         }
5425         if (ffe_ctl->cached) {
5426                 struct btrfs_free_space_ctl *free_space_ctl;
5427
5428                 free_space_ctl = bg->free_space_ctl;
5429                 spin_lock(&free_space_ctl->tree_lock);
5430                 if (free_space_ctl->free_space <
5431                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
5432                     ffe_ctl->empty_size) {
5433                         ffe_ctl->total_free_space = max_t(u64,
5434                                         ffe_ctl->total_free_space,
5435                                         free_space_ctl->free_space);
5436                         spin_unlock(&free_space_ctl->tree_lock);
5437                         return 1;
5438                 }
5439                 spin_unlock(&free_space_ctl->tree_lock);
5440         }
5441
5442         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
5443                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
5444                         &ffe_ctl->max_extent_size);
5445
5446         /*
5447          * If we didn't find a chunk, and we haven't failed on this block group
5448          * before, and this block group is in the middle of caching and we are
5449          * ok with waiting, then go ahead and wait for progress to be made, and
5450          * set @retry_unclustered to true.
5451          *
5452          * If @retry_unclustered is true then we've already waited on this
5453          * block group once and should move on to the next block group.
5454          */
5455         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
5456             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
5457                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
5458                                                 ffe_ctl->empty_size);
5459                 ffe_ctl->retry_unclustered = true;
5460                 return -EAGAIN;
5461         } else if (!offset) {
5462                 return 1;
5463         }
5464         ffe_ctl->found_offset = offset;
5465         return 0;
5466 }
5467
5468 /*
5469  * Return >0 means caller needs to re-search for free extent
5470  * Return 0 means we have the needed free extent.
5471  * Return <0 means we failed to locate any free extent.
5472  */
5473 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
5474                                         struct btrfs_free_cluster *last_ptr,
5475                                         struct btrfs_key *ins,
5476                                         struct find_free_extent_ctl *ffe_ctl,
5477                                         int full_search, bool use_cluster)
5478 {
5479         struct btrfs_root *root = fs_info->extent_root;
5480         int ret;
5481
5482         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
5483             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
5484                 ffe_ctl->orig_have_caching_bg = true;
5485
5486         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
5487             ffe_ctl->have_caching_bg)
5488                 return 1;
5489
5490         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
5491                 return 1;
5492
5493         if (ins->objectid) {
5494                 if (!use_cluster && last_ptr) {
5495                         spin_lock(&last_ptr->lock);
5496                         last_ptr->window_start = ins->objectid;
5497                         spin_unlock(&last_ptr->lock);
5498                 }
5499                 return 0;
5500         }
5501
5502         /*
5503          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5504          *                      caching kthreads as we move along
5505          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5506          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5507          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5508          *                     again
5509          */
5510         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
5511                 ffe_ctl->index = 0;
5512                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
5513                         /*
5514                          * We want to skip the LOOP_CACHING_WAIT step if we
5515                          * don't have any uncached bgs and we've already done a
5516                          * full search through.
5517                          */
5518                         if (ffe_ctl->orig_have_caching_bg || !full_search)
5519                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
5520                         else
5521                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
5522                 } else {
5523                         ffe_ctl->loop++;
5524                 }
5525
5526                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
5527                         struct btrfs_trans_handle *trans;
5528                         int exist = 0;
5529
5530                         trans = current->journal_info;
5531                         if (trans)
5532                                 exist = 1;
5533                         else
5534                                 trans = btrfs_join_transaction(root);
5535
5536                         if (IS_ERR(trans)) {
5537                                 ret = PTR_ERR(trans);
5538                                 return ret;
5539                         }
5540
5541                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
5542                                                 CHUNK_ALLOC_FORCE);
5543
5544                         /*
5545                          * If we can't allocate a new chunk we've already looped
5546                          * through at least once, move on to the NO_EMPTY_SIZE
5547                          * case.
5548                          */
5549                         if (ret == -ENOSPC)
5550                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
5551
5552                         /* Do not bail out on ENOSPC since we can do more. */
5553                         if (ret < 0 && ret != -ENOSPC)
5554                                 btrfs_abort_transaction(trans, ret);
5555                         else
5556                                 ret = 0;
5557                         if (!exist)
5558                                 btrfs_end_transaction(trans);
5559                         if (ret)
5560                                 return ret;
5561                 }
5562
5563                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
5564                         /*
5565                          * Don't loop again if we already have no empty_size and
5566                          * no empty_cluster.
5567                          */
5568                         if (ffe_ctl->empty_size == 0 &&
5569                             ffe_ctl->empty_cluster == 0)
5570                                 return -ENOSPC;
5571                         ffe_ctl->empty_size = 0;
5572                         ffe_ctl->empty_cluster = 0;
5573                 }
5574                 return 1;
5575         }
5576         return -ENOSPC;
5577 }
5578
5579 /*
5580  * walks the btree of allocated extents and find a hole of a given size.
5581  * The key ins is changed to record the hole:
5582  * ins->objectid == start position
5583  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5584  * ins->offset == the size of the hole.
5585  * Any available blocks before search_start are skipped.
5586  *
5587  * If there is no suitable free space, we will record the max size of
5588  * the free space extent currently.
5589  *
5590  * The overall logic and call chain:
5591  *
5592  * find_free_extent()
5593  * |- Iterate through all block groups
5594  * |  |- Get a valid block group
5595  * |  |- Try to do clustered allocation in that block group
5596  * |  |- Try to do unclustered allocation in that block group
5597  * |  |- Check if the result is valid
5598  * |  |  |- If valid, then exit
5599  * |  |- Jump to next block group
5600  * |
5601  * |- Push harder to find free extents
5602  *    |- If not found, re-iterate all block groups
5603  */
5604 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
5605                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
5606                                 u64 hint_byte, struct btrfs_key *ins,
5607                                 u64 flags, int delalloc)
5608 {
5609         int ret = 0;
5610         struct btrfs_free_cluster *last_ptr = NULL;
5611         struct btrfs_block_group_cache *block_group = NULL;
5612         struct find_free_extent_ctl ffe_ctl = {0};
5613         struct btrfs_space_info *space_info;
5614         bool use_cluster = true;
5615         bool full_search = false;
5616
5617         WARN_ON(num_bytes < fs_info->sectorsize);
5618
5619         ffe_ctl.ram_bytes = ram_bytes;
5620         ffe_ctl.num_bytes = num_bytes;
5621         ffe_ctl.empty_size = empty_size;
5622         ffe_ctl.flags = flags;
5623         ffe_ctl.search_start = 0;
5624         ffe_ctl.retry_clustered = false;
5625         ffe_ctl.retry_unclustered = false;
5626         ffe_ctl.delalloc = delalloc;
5627         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
5628         ffe_ctl.have_caching_bg = false;
5629         ffe_ctl.orig_have_caching_bg = false;
5630         ffe_ctl.found_offset = 0;
5631
5632         ins->type = BTRFS_EXTENT_ITEM_KEY;
5633         ins->objectid = 0;
5634         ins->offset = 0;
5635
5636         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
5637
5638         space_info = btrfs_find_space_info(fs_info, flags);
5639         if (!space_info) {
5640                 btrfs_err(fs_info, "No space info for %llu", flags);
5641                 return -ENOSPC;
5642         }
5643
5644         /*
5645          * If our free space is heavily fragmented we may not be able to make
5646          * big contiguous allocations, so instead of doing the expensive search
5647          * for free space, simply return ENOSPC with our max_extent_size so we
5648          * can go ahead and search for a more manageable chunk.
5649          *
5650          * If our max_extent_size is large enough for our allocation simply
5651          * disable clustering since we will likely not be able to find enough
5652          * space to create a cluster and induce latency trying.
5653          */
5654         if (unlikely(space_info->max_extent_size)) {
5655                 spin_lock(&space_info->lock);
5656                 if (space_info->max_extent_size &&
5657                     num_bytes > space_info->max_extent_size) {
5658                         ins->offset = space_info->max_extent_size;
5659                         spin_unlock(&space_info->lock);
5660                         return -ENOSPC;
5661                 } else if (space_info->max_extent_size) {
5662                         use_cluster = false;
5663                 }
5664                 spin_unlock(&space_info->lock);
5665         }
5666
5667         last_ptr = fetch_cluster_info(fs_info, space_info,
5668                                       &ffe_ctl.empty_cluster);
5669         if (last_ptr) {
5670                 spin_lock(&last_ptr->lock);
5671                 if (last_ptr->block_group)
5672                         hint_byte = last_ptr->window_start;
5673                 if (last_ptr->fragmented) {
5674                         /*
5675                          * We still set window_start so we can keep track of the
5676                          * last place we found an allocation to try and save
5677                          * some time.
5678                          */
5679                         hint_byte = last_ptr->window_start;
5680                         use_cluster = false;
5681                 }
5682                 spin_unlock(&last_ptr->lock);
5683         }
5684
5685         ffe_ctl.search_start = max(ffe_ctl.search_start,
5686                                    first_logical_byte(fs_info, 0));
5687         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
5688         if (ffe_ctl.search_start == hint_byte) {
5689                 block_group = btrfs_lookup_block_group(fs_info,
5690                                                        ffe_ctl.search_start);
5691                 /*
5692                  * we don't want to use the block group if it doesn't match our
5693                  * allocation bits, or if its not cached.
5694                  *
5695                  * However if we are re-searching with an ideal block group
5696                  * picked out then we don't care that the block group is cached.
5697                  */
5698                 if (block_group && block_group_bits(block_group, flags) &&
5699                     block_group->cached != BTRFS_CACHE_NO) {
5700                         down_read(&space_info->groups_sem);
5701                         if (list_empty(&block_group->list) ||
5702                             block_group->ro) {
5703                                 /*
5704                                  * someone is removing this block group,
5705                                  * we can't jump into the have_block_group
5706                                  * target because our list pointers are not
5707                                  * valid
5708                                  */
5709                                 btrfs_put_block_group(block_group);
5710                                 up_read(&space_info->groups_sem);
5711                         } else {
5712                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
5713                                                 block_group->flags);
5714                                 btrfs_lock_block_group(block_group, delalloc);
5715                                 goto have_block_group;
5716                         }
5717                 } else if (block_group) {
5718                         btrfs_put_block_group(block_group);
5719                 }
5720         }
5721 search:
5722         ffe_ctl.have_caching_bg = false;
5723         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
5724             ffe_ctl.index == 0)
5725                 full_search = true;
5726         down_read(&space_info->groups_sem);
5727         list_for_each_entry(block_group,
5728                             &space_info->block_groups[ffe_ctl.index], list) {
5729                 /* If the block group is read-only, we can skip it entirely. */
5730                 if (unlikely(block_group->ro))
5731                         continue;
5732
5733                 btrfs_grab_block_group(block_group, delalloc);
5734                 ffe_ctl.search_start = block_group->key.objectid;
5735
5736                 /*
5737                  * this can happen if we end up cycling through all the
5738                  * raid types, but we want to make sure we only allocate
5739                  * for the proper type.
5740                  */
5741                 if (!block_group_bits(block_group, flags)) {
5742                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
5743                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
5744                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
5745                                 BTRFS_BLOCK_GROUP_RAID10;
5746
5747                         /*
5748                          * if they asked for extra copies and this block group
5749                          * doesn't provide them, bail.  This does allow us to
5750                          * fill raid0 from raid1.
5751                          */
5752                         if ((flags & extra) && !(block_group->flags & extra))
5753                                 goto loop;
5754
5755                         /*
5756                          * This block group has different flags than we want.
5757                          * It's possible that we have MIXED_GROUP flag but no
5758                          * block group is mixed.  Just skip such block group.
5759                          */
5760                         btrfs_release_block_group(block_group, delalloc);
5761                         continue;
5762                 }
5763
5764 have_block_group:
5765                 ffe_ctl.cached = block_group_cache_done(block_group);
5766                 if (unlikely(!ffe_ctl.cached)) {
5767                         ffe_ctl.have_caching_bg = true;
5768                         ret = cache_block_group(block_group, 0);
5769                         BUG_ON(ret < 0);
5770                         ret = 0;
5771                 }
5772
5773                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
5774                         goto loop;
5775
5776                 /*
5777                  * Ok we want to try and use the cluster allocator, so
5778                  * lets look there
5779                  */
5780                 if (last_ptr && use_cluster) {
5781                         struct btrfs_block_group_cache *cluster_bg = NULL;
5782
5783                         ret = find_free_extent_clustered(block_group, last_ptr,
5784                                                          &ffe_ctl, &cluster_bg);
5785
5786                         if (ret == 0) {
5787                                 if (cluster_bg && cluster_bg != block_group) {
5788                                         btrfs_release_block_group(block_group,
5789                                                                   delalloc);
5790                                         block_group = cluster_bg;
5791                                 }
5792                                 goto checks;
5793                         } else if (ret == -EAGAIN) {
5794                                 goto have_block_group;
5795                         } else if (ret > 0) {
5796                                 goto loop;
5797                         }
5798                         /* ret == -ENOENT case falls through */
5799                 }
5800
5801                 ret = find_free_extent_unclustered(block_group, last_ptr,
5802                                                    &ffe_ctl);
5803                 if (ret == -EAGAIN)
5804                         goto have_block_group;
5805                 else if (ret > 0)
5806                         goto loop;
5807                 /* ret == 0 case falls through */
5808 checks:
5809                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
5810                                              fs_info->stripesize);
5811
5812                 /* move on to the next group */
5813                 if (ffe_ctl.search_start + num_bytes >
5814                     block_group->key.objectid + block_group->key.offset) {
5815                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
5816                                              num_bytes);
5817                         goto loop;
5818                 }
5819
5820                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
5821                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
5822                                 ffe_ctl.search_start - ffe_ctl.found_offset);
5823
5824                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
5825                                 num_bytes, delalloc);
5826                 if (ret == -EAGAIN) {
5827                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
5828                                              num_bytes);
5829                         goto loop;
5830                 }
5831                 btrfs_inc_block_group_reservations(block_group);
5832
5833                 /* we are all good, lets return */
5834                 ins->objectid = ffe_ctl.search_start;
5835                 ins->offset = num_bytes;
5836
5837                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
5838                                            num_bytes);
5839                 btrfs_release_block_group(block_group, delalloc);
5840                 break;
5841 loop:
5842                 ffe_ctl.retry_clustered = false;
5843                 ffe_ctl.retry_unclustered = false;
5844                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
5845                        ffe_ctl.index);
5846                 btrfs_release_block_group(block_group, delalloc);
5847                 cond_resched();
5848         }
5849         up_read(&space_info->groups_sem);
5850
5851         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
5852                                            full_search, use_cluster);
5853         if (ret > 0)
5854                 goto search;
5855
5856         if (ret == -ENOSPC) {
5857                 /*
5858                  * Use ffe_ctl->total_free_space as fallback if we can't find
5859                  * any contiguous hole.
5860                  */
5861                 if (!ffe_ctl.max_extent_size)
5862                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
5863                 spin_lock(&space_info->lock);
5864                 space_info->max_extent_size = ffe_ctl.max_extent_size;
5865                 spin_unlock(&space_info->lock);
5866                 ins->offset = ffe_ctl.max_extent_size;
5867         }
5868         return ret;
5869 }
5870
5871 /*
5872  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
5873  *                        hole that is at least as big as @num_bytes.
5874  *
5875  * @root           -    The root that will contain this extent
5876  *
5877  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
5878  *                      is used for accounting purposes. This value differs
5879  *                      from @num_bytes only in the case of compressed extents.
5880  *
5881  * @num_bytes      -    Number of bytes to allocate on-disk.
5882  *
5883  * @min_alloc_size -    Indicates the minimum amount of space that the
5884  *                      allocator should try to satisfy. In some cases
5885  *                      @num_bytes may be larger than what is required and if
5886  *                      the filesystem is fragmented then allocation fails.
5887  *                      However, the presence of @min_alloc_size gives a
5888  *                      chance to try and satisfy the smaller allocation.
5889  *
5890  * @empty_size     -    A hint that you plan on doing more COW. This is the
5891  *                      size in bytes the allocator should try to find free
5892  *                      next to the block it returns.  This is just a hint and
5893  *                      may be ignored by the allocator.
5894  *
5895  * @hint_byte      -    Hint to the allocator to start searching above the byte
5896  *                      address passed. It might be ignored.
5897  *
5898  * @ins            -    This key is modified to record the found hole. It will
5899  *                      have the following values:
5900  *                      ins->objectid == start position
5901  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
5902  *                      ins->offset == the size of the hole.
5903  *
5904  * @is_data        -    Boolean flag indicating whether an extent is
5905  *                      allocated for data (true) or metadata (false)
5906  *
5907  * @delalloc       -    Boolean flag indicating whether this allocation is for
5908  *                      delalloc or not. If 'true' data_rwsem of block groups
5909  *                      is going to be acquired.
5910  *
5911  *
5912  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
5913  * case -ENOSPC is returned then @ins->offset will contain the size of the
5914  * largest available hole the allocator managed to find.
5915  */
5916 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
5917                          u64 num_bytes, u64 min_alloc_size,
5918                          u64 empty_size, u64 hint_byte,
5919                          struct btrfs_key *ins, int is_data, int delalloc)
5920 {
5921         struct btrfs_fs_info *fs_info = root->fs_info;
5922         bool final_tried = num_bytes == min_alloc_size;
5923         u64 flags;
5924         int ret;
5925
5926         flags = get_alloc_profile_by_root(root, is_data);
5927 again:
5928         WARN_ON(num_bytes < fs_info->sectorsize);
5929         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
5930                                hint_byte, ins, flags, delalloc);
5931         if (!ret && !is_data) {
5932                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
5933         } else if (ret == -ENOSPC) {
5934                 if (!final_tried && ins->offset) {
5935                         num_bytes = min(num_bytes >> 1, ins->offset);
5936                         num_bytes = round_down(num_bytes,
5937                                                fs_info->sectorsize);
5938                         num_bytes = max(num_bytes, min_alloc_size);
5939                         ram_bytes = num_bytes;
5940                         if (num_bytes == min_alloc_size)
5941                                 final_tried = true;
5942                         goto again;
5943                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
5944                         struct btrfs_space_info *sinfo;
5945
5946                         sinfo = btrfs_find_space_info(fs_info, flags);
5947                         btrfs_err(fs_info,
5948                                   "allocation failed flags %llu, wanted %llu",
5949                                   flags, num_bytes);
5950                         if (sinfo)
5951                                 btrfs_dump_space_info(fs_info, sinfo,
5952                                                       num_bytes, 1);
5953                 }
5954         }
5955
5956         return ret;
5957 }
5958
5959 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
5960                                         u64 start, u64 len,
5961                                         int pin, int delalloc)
5962 {
5963         struct btrfs_block_group_cache *cache;
5964         int ret = 0;
5965
5966         cache = btrfs_lookup_block_group(fs_info, start);
5967         if (!cache) {
5968                 btrfs_err(fs_info, "Unable to find block group for %llu",
5969                           start);
5970                 return -ENOSPC;
5971         }
5972
5973         if (pin)
5974                 pin_down_extent(cache, start, len, 1);
5975         else {
5976                 if (btrfs_test_opt(fs_info, DISCARD))
5977                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
5978                 btrfs_add_free_space(cache, start, len);
5979                 btrfs_free_reserved_bytes(cache, len, delalloc);
5980                 trace_btrfs_reserved_extent_free(fs_info, start, len);
5981         }
5982
5983         btrfs_put_block_group(cache);
5984         return ret;
5985 }
5986
5987 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
5988                                u64 start, u64 len, int delalloc)
5989 {
5990         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
5991 }
5992
5993 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
5994                                        u64 start, u64 len)
5995 {
5996         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
5997 }
5998
5999 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6000                                       u64 parent, u64 root_objectid,
6001                                       u64 flags, u64 owner, u64 offset,
6002                                       struct btrfs_key *ins, int ref_mod)
6003 {
6004         struct btrfs_fs_info *fs_info = trans->fs_info;
6005         int ret;
6006         struct btrfs_extent_item *extent_item;
6007         struct btrfs_extent_inline_ref *iref;
6008         struct btrfs_path *path;
6009         struct extent_buffer *leaf;
6010         int type;
6011         u32 size;
6012
6013         if (parent > 0)
6014                 type = BTRFS_SHARED_DATA_REF_KEY;
6015         else
6016                 type = BTRFS_EXTENT_DATA_REF_KEY;
6017
6018         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6019
6020         path = btrfs_alloc_path();
6021         if (!path)
6022                 return -ENOMEM;
6023
6024         path->leave_spinning = 1;
6025         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6026                                       ins, size);
6027         if (ret) {
6028                 btrfs_free_path(path);
6029                 return ret;
6030         }
6031
6032         leaf = path->nodes[0];
6033         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6034                                      struct btrfs_extent_item);
6035         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6036         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6037         btrfs_set_extent_flags(leaf, extent_item,
6038                                flags | BTRFS_EXTENT_FLAG_DATA);
6039
6040         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6041         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6042         if (parent > 0) {
6043                 struct btrfs_shared_data_ref *ref;
6044                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6045                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6046                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6047         } else {
6048                 struct btrfs_extent_data_ref *ref;
6049                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6050                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6051                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6052                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6053                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6054         }
6055
6056         btrfs_mark_buffer_dirty(path->nodes[0]);
6057         btrfs_free_path(path);
6058
6059         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
6060         if (ret)
6061                 return ret;
6062
6063         ret = update_block_group(trans, ins->objectid, ins->offset, 1);
6064         if (ret) { /* -ENOENT, logic error */
6065                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6066                         ins->objectid, ins->offset);
6067                 BUG();
6068         }
6069         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
6070         return ret;
6071 }
6072
6073 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6074                                      struct btrfs_delayed_ref_node *node,
6075                                      struct btrfs_delayed_extent_op *extent_op)
6076 {
6077         struct btrfs_fs_info *fs_info = trans->fs_info;
6078         int ret;
6079         struct btrfs_extent_item *extent_item;
6080         struct btrfs_key extent_key;
6081         struct btrfs_tree_block_info *block_info;
6082         struct btrfs_extent_inline_ref *iref;
6083         struct btrfs_path *path;
6084         struct extent_buffer *leaf;
6085         struct btrfs_delayed_tree_ref *ref;
6086         u32 size = sizeof(*extent_item) + sizeof(*iref);
6087         u64 num_bytes;
6088         u64 flags = extent_op->flags_to_set;
6089         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
6090
6091         ref = btrfs_delayed_node_to_tree_ref(node);
6092
6093         extent_key.objectid = node->bytenr;
6094         if (skinny_metadata) {
6095                 extent_key.offset = ref->level;
6096                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
6097                 num_bytes = fs_info->nodesize;
6098         } else {
6099                 extent_key.offset = node->num_bytes;
6100                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
6101                 size += sizeof(*block_info);
6102                 num_bytes = node->num_bytes;
6103         }
6104
6105         path = btrfs_alloc_path();
6106         if (!path)
6107                 return -ENOMEM;
6108
6109         path->leave_spinning = 1;
6110         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6111                                       &extent_key, size);
6112         if (ret) {
6113                 btrfs_free_path(path);
6114                 return ret;
6115         }
6116
6117         leaf = path->nodes[0];
6118         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6119                                      struct btrfs_extent_item);
6120         btrfs_set_extent_refs(leaf, extent_item, 1);
6121         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6122         btrfs_set_extent_flags(leaf, extent_item,
6123                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6124
6125         if (skinny_metadata) {
6126                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6127         } else {
6128                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6129                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
6130                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
6131                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6132         }
6133
6134         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
6135                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6136                 btrfs_set_extent_inline_ref_type(leaf, iref,
6137                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6138                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
6139         } else {
6140                 btrfs_set_extent_inline_ref_type(leaf, iref,
6141                                                  BTRFS_TREE_BLOCK_REF_KEY);
6142                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
6143         }
6144
6145         btrfs_mark_buffer_dirty(leaf);
6146         btrfs_free_path(path);
6147
6148         ret = remove_from_free_space_tree(trans, extent_key.objectid,
6149                                           num_bytes);
6150         if (ret)
6151                 return ret;
6152
6153         ret = update_block_group(trans, extent_key.objectid,
6154                                  fs_info->nodesize, 1);
6155         if (ret) { /* -ENOENT, logic error */
6156                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6157                         extent_key.objectid, extent_key.offset);
6158                 BUG();
6159         }
6160
6161         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
6162                                           fs_info->nodesize);
6163         return ret;
6164 }
6165
6166 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6167                                      struct btrfs_root *root, u64 owner,
6168                                      u64 offset, u64 ram_bytes,
6169                                      struct btrfs_key *ins)
6170 {
6171         struct btrfs_ref generic_ref = { 0 };
6172         int ret;
6173
6174         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
6175
6176         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
6177                                ins->objectid, ins->offset, 0);
6178         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
6179         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
6180         ret = btrfs_add_delayed_data_ref(trans, &generic_ref,
6181                                          ram_bytes, NULL, NULL);
6182         return ret;
6183 }
6184
6185 /*
6186  * this is used by the tree logging recovery code.  It records that
6187  * an extent has been allocated and makes sure to clear the free
6188  * space cache bits as well
6189  */
6190 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6191                                    u64 root_objectid, u64 owner, u64 offset,
6192                                    struct btrfs_key *ins)
6193 {
6194         struct btrfs_fs_info *fs_info = trans->fs_info;
6195         int ret;
6196         struct btrfs_block_group_cache *block_group;
6197         struct btrfs_space_info *space_info;
6198
6199         /*
6200          * Mixed block groups will exclude before processing the log so we only
6201          * need to do the exclude dance if this fs isn't mixed.
6202          */
6203         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
6204                 ret = __exclude_logged_extent(fs_info, ins->objectid,
6205                                               ins->offset);
6206                 if (ret)
6207                         return ret;
6208         }
6209
6210         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
6211         if (!block_group)
6212                 return -EINVAL;
6213
6214         space_info = block_group->space_info;
6215         spin_lock(&space_info->lock);
6216         spin_lock(&block_group->lock);
6217         space_info->bytes_reserved += ins->offset;
6218         block_group->reserved += ins->offset;
6219         spin_unlock(&block_group->lock);
6220         spin_unlock(&space_info->lock);
6221
6222         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
6223                                          offset, ins, 1);
6224         btrfs_put_block_group(block_group);
6225         return ret;
6226 }
6227
6228 static struct extent_buffer *
6229 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6230                       u64 bytenr, int level, u64 owner)
6231 {
6232         struct btrfs_fs_info *fs_info = root->fs_info;
6233         struct extent_buffer *buf;
6234
6235         buf = btrfs_find_create_tree_block(fs_info, bytenr);
6236         if (IS_ERR(buf))
6237                 return buf;
6238
6239         /*
6240          * Extra safety check in case the extent tree is corrupted and extent
6241          * allocator chooses to use a tree block which is already used and
6242          * locked.
6243          */
6244         if (buf->lock_owner == current->pid) {
6245                 btrfs_err_rl(fs_info,
6246 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
6247                         buf->start, btrfs_header_owner(buf), current->pid);
6248                 free_extent_buffer(buf);
6249                 return ERR_PTR(-EUCLEAN);
6250         }
6251
6252         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6253         btrfs_tree_lock(buf);
6254         btrfs_clean_tree_block(buf);
6255         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6256
6257         btrfs_set_lock_blocking_write(buf);
6258         set_extent_buffer_uptodate(buf);
6259
6260         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
6261         btrfs_set_header_level(buf, level);
6262         btrfs_set_header_bytenr(buf, buf->start);
6263         btrfs_set_header_generation(buf, trans->transid);
6264         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
6265         btrfs_set_header_owner(buf, owner);
6266         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
6267         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
6268         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6269                 buf->log_index = root->log_transid % 2;
6270                 /*
6271                  * we allow two log transactions at a time, use different
6272                  * EXTENT bit to differentiate dirty pages.
6273                  */
6274                 if (buf->log_index == 0)
6275                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6276                                         buf->start + buf->len - 1, GFP_NOFS);
6277                 else
6278                         set_extent_new(&root->dirty_log_pages, buf->start,
6279                                         buf->start + buf->len - 1);
6280         } else {
6281                 buf->log_index = -1;
6282                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6283                          buf->start + buf->len - 1, GFP_NOFS);
6284         }
6285         trans->dirty = true;
6286         /* this returns a buffer locked for blocking */
6287         return buf;
6288 }
6289
6290 /*
6291  * finds a free extent and does all the dirty work required for allocation
6292  * returns the tree buffer or an ERR_PTR on error.
6293  */
6294 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
6295                                              struct btrfs_root *root,
6296                                              u64 parent, u64 root_objectid,
6297                                              const struct btrfs_disk_key *key,
6298                                              int level, u64 hint,
6299                                              u64 empty_size)
6300 {
6301         struct btrfs_fs_info *fs_info = root->fs_info;
6302         struct btrfs_key ins;
6303         struct btrfs_block_rsv *block_rsv;
6304         struct extent_buffer *buf;
6305         struct btrfs_delayed_extent_op *extent_op;
6306         struct btrfs_ref generic_ref = { 0 };
6307         u64 flags = 0;
6308         int ret;
6309         u32 blocksize = fs_info->nodesize;
6310         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
6311
6312 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
6313         if (btrfs_is_testing(fs_info)) {
6314                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
6315                                             level, root_objectid);
6316                 if (!IS_ERR(buf))
6317                         root->alloc_bytenr += blocksize;
6318                 return buf;
6319         }
6320 #endif
6321
6322         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
6323         if (IS_ERR(block_rsv))
6324                 return ERR_CAST(block_rsv);
6325
6326         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
6327                                    empty_size, hint, &ins, 0, 0);
6328         if (ret)
6329                 goto out_unuse;
6330
6331         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
6332                                     root_objectid);
6333         if (IS_ERR(buf)) {
6334                 ret = PTR_ERR(buf);
6335                 goto out_free_reserved;
6336         }
6337
6338         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6339                 if (parent == 0)
6340                         parent = ins.objectid;
6341                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6342         } else
6343                 BUG_ON(parent > 0);
6344
6345         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6346                 extent_op = btrfs_alloc_delayed_extent_op();
6347                 if (!extent_op) {
6348                         ret = -ENOMEM;
6349                         goto out_free_buf;
6350                 }
6351                 if (key)
6352                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6353                 else
6354                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6355                 extent_op->flags_to_set = flags;
6356                 extent_op->update_key = skinny_metadata ? false : true;
6357                 extent_op->update_flags = true;
6358                 extent_op->is_data = false;
6359                 extent_op->level = level;
6360
6361                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
6362                                        ins.objectid, ins.offset, parent);
6363                 generic_ref.real_root = root->root_key.objectid;
6364                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
6365                 btrfs_ref_tree_mod(fs_info, &generic_ref);
6366                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref,
6367                                                  extent_op, NULL, NULL);
6368                 if (ret)
6369                         goto out_free_delayed;
6370         }
6371         return buf;
6372
6373 out_free_delayed:
6374         btrfs_free_delayed_extent_op(extent_op);
6375 out_free_buf:
6376         free_extent_buffer(buf);
6377 out_free_reserved:
6378         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
6379 out_unuse:
6380         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
6381         return ERR_PTR(ret);
6382 }
6383
6384 struct walk_control {
6385         u64 refs[BTRFS_MAX_LEVEL];
6386         u64 flags[BTRFS_MAX_LEVEL];
6387         struct btrfs_key update_progress;
6388         struct btrfs_key drop_progress;
6389         int drop_level;
6390         int stage;
6391         int level;
6392         int shared_level;
6393         int update_ref;
6394         int keep_locks;
6395         int reada_slot;
6396         int reada_count;
6397         int restarted;
6398 };
6399
6400 #define DROP_REFERENCE  1
6401 #define UPDATE_BACKREF  2
6402
6403 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6404                                      struct btrfs_root *root,
6405                                      struct walk_control *wc,
6406                                      struct btrfs_path *path)
6407 {
6408         struct btrfs_fs_info *fs_info = root->fs_info;
6409         u64 bytenr;
6410         u64 generation;
6411         u64 refs;
6412         u64 flags;
6413         u32 nritems;
6414         struct btrfs_key key;
6415         struct extent_buffer *eb;
6416         int ret;
6417         int slot;
6418         int nread = 0;
6419
6420         if (path->slots[wc->level] < wc->reada_slot) {
6421                 wc->reada_count = wc->reada_count * 2 / 3;
6422                 wc->reada_count = max(wc->reada_count, 2);
6423         } else {
6424                 wc->reada_count = wc->reada_count * 3 / 2;
6425                 wc->reada_count = min_t(int, wc->reada_count,
6426                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
6427         }
6428
6429         eb = path->nodes[wc->level];
6430         nritems = btrfs_header_nritems(eb);
6431
6432         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6433                 if (nread >= wc->reada_count)
6434                         break;
6435
6436                 cond_resched();
6437                 bytenr = btrfs_node_blockptr(eb, slot);
6438                 generation = btrfs_node_ptr_generation(eb, slot);
6439
6440                 if (slot == path->slots[wc->level])
6441                         goto reada;
6442
6443                 if (wc->stage == UPDATE_BACKREF &&
6444                     generation <= root->root_key.offset)
6445                         continue;
6446
6447                 /* We don't lock the tree block, it's OK to be racy here */
6448                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
6449                                                wc->level - 1, 1, &refs,
6450                                                &flags);
6451                 /* We don't care about errors in readahead. */
6452                 if (ret < 0)
6453                         continue;
6454                 BUG_ON(refs == 0);
6455
6456                 if (wc->stage == DROP_REFERENCE) {
6457                         if (refs == 1)
6458                                 goto reada;
6459
6460                         if (wc->level == 1 &&
6461                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6462                                 continue;
6463                         if (!wc->update_ref ||
6464                             generation <= root->root_key.offset)
6465                                 continue;
6466                         btrfs_node_key_to_cpu(eb, &key, slot);
6467                         ret = btrfs_comp_cpu_keys(&key,
6468                                                   &wc->update_progress);
6469                         if (ret < 0)
6470                                 continue;
6471                 } else {
6472                         if (wc->level == 1 &&
6473                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6474                                 continue;
6475                 }
6476 reada:
6477                 readahead_tree_block(fs_info, bytenr);
6478                 nread++;
6479         }
6480         wc->reada_slot = slot;
6481 }
6482
6483 /*
6484  * helper to process tree block while walking down the tree.
6485  *
6486  * when wc->stage == UPDATE_BACKREF, this function updates
6487  * back refs for pointers in the block.
6488  *
6489  * NOTE: return value 1 means we should stop walking down.
6490  */
6491 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6492                                    struct btrfs_root *root,
6493                                    struct btrfs_path *path,
6494                                    struct walk_control *wc, int lookup_info)
6495 {
6496         struct btrfs_fs_info *fs_info = root->fs_info;
6497         int level = wc->level;
6498         struct extent_buffer *eb = path->nodes[level];
6499         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6500         int ret;
6501
6502         if (wc->stage == UPDATE_BACKREF &&
6503             btrfs_header_owner(eb) != root->root_key.objectid)
6504                 return 1;
6505
6506         /*
6507          * when reference count of tree block is 1, it won't increase
6508          * again. once full backref flag is set, we never clear it.
6509          */
6510         if (lookup_info &&
6511             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6512              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6513                 BUG_ON(!path->locks[level]);
6514                 ret = btrfs_lookup_extent_info(trans, fs_info,
6515                                                eb->start, level, 1,
6516                                                &wc->refs[level],
6517                                                &wc->flags[level]);
6518                 BUG_ON(ret == -ENOMEM);
6519                 if (ret)
6520                         return ret;
6521                 BUG_ON(wc->refs[level] == 0);
6522         }
6523
6524         if (wc->stage == DROP_REFERENCE) {
6525                 if (wc->refs[level] > 1)
6526                         return 1;
6527
6528                 if (path->locks[level] && !wc->keep_locks) {
6529                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6530                         path->locks[level] = 0;
6531                 }
6532                 return 0;
6533         }
6534
6535         /* wc->stage == UPDATE_BACKREF */
6536         if (!(wc->flags[level] & flag)) {
6537                 BUG_ON(!path->locks[level]);
6538                 ret = btrfs_inc_ref(trans, root, eb, 1);
6539                 BUG_ON(ret); /* -ENOMEM */
6540                 ret = btrfs_dec_ref(trans, root, eb, 0);
6541                 BUG_ON(ret); /* -ENOMEM */
6542                 ret = btrfs_set_disk_extent_flags(trans, eb->start,
6543                                                   eb->len, flag,
6544                                                   btrfs_header_level(eb), 0);
6545                 BUG_ON(ret); /* -ENOMEM */
6546                 wc->flags[level] |= flag;
6547         }
6548
6549         /*
6550          * the block is shared by multiple trees, so it's not good to
6551          * keep the tree lock
6552          */
6553         if (path->locks[level] && level > 0) {
6554                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6555                 path->locks[level] = 0;
6556         }
6557         return 0;
6558 }
6559
6560 /*
6561  * This is used to verify a ref exists for this root to deal with a bug where we
6562  * would have a drop_progress key that hadn't been updated properly.
6563  */
6564 static int check_ref_exists(struct btrfs_trans_handle *trans,
6565                             struct btrfs_root *root, u64 bytenr, u64 parent,
6566                             int level)
6567 {
6568         struct btrfs_path *path;
6569         struct btrfs_extent_inline_ref *iref;
6570         int ret;
6571
6572         path = btrfs_alloc_path();
6573         if (!path)
6574                 return -ENOMEM;
6575
6576         ret = lookup_extent_backref(trans, path, &iref, bytenr,
6577                                     root->fs_info->nodesize, parent,
6578                                     root->root_key.objectid, level, 0);
6579         btrfs_free_path(path);
6580         if (ret == -ENOENT)
6581                 return 0;
6582         if (ret < 0)
6583                 return ret;
6584         return 1;
6585 }
6586
6587 /*
6588  * helper to process tree block pointer.
6589  *
6590  * when wc->stage == DROP_REFERENCE, this function checks
6591  * reference count of the block pointed to. if the block
6592  * is shared and we need update back refs for the subtree
6593  * rooted at the block, this function changes wc->stage to
6594  * UPDATE_BACKREF. if the block is shared and there is no
6595  * need to update back, this function drops the reference
6596  * to the block.
6597  *
6598  * NOTE: return value 1 means we should stop walking down.
6599  */
6600 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6601                                  struct btrfs_root *root,
6602                                  struct btrfs_path *path,
6603                                  struct walk_control *wc, int *lookup_info)
6604 {
6605         struct btrfs_fs_info *fs_info = root->fs_info;
6606         u64 bytenr;
6607         u64 generation;
6608         u64 parent;
6609         struct btrfs_key key;
6610         struct btrfs_key first_key;
6611         struct btrfs_ref ref = { 0 };
6612         struct extent_buffer *next;
6613         int level = wc->level;
6614         int reada = 0;
6615         int ret = 0;
6616         bool need_account = false;
6617
6618         generation = btrfs_node_ptr_generation(path->nodes[level],
6619                                                path->slots[level]);
6620         /*
6621          * if the lower level block was created before the snapshot
6622          * was created, we know there is no need to update back refs
6623          * for the subtree
6624          */
6625         if (wc->stage == UPDATE_BACKREF &&
6626             generation <= root->root_key.offset) {
6627                 *lookup_info = 1;
6628                 return 1;
6629         }
6630
6631         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6632         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
6633                               path->slots[level]);
6634
6635         next = find_extent_buffer(fs_info, bytenr);
6636         if (!next) {
6637                 next = btrfs_find_create_tree_block(fs_info, bytenr);
6638                 if (IS_ERR(next))
6639                         return PTR_ERR(next);
6640
6641                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
6642                                                level - 1);
6643                 reada = 1;
6644         }
6645         btrfs_tree_lock(next);
6646         btrfs_set_lock_blocking_write(next);
6647
6648         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
6649                                        &wc->refs[level - 1],
6650                                        &wc->flags[level - 1]);
6651         if (ret < 0)
6652                 goto out_unlock;
6653
6654         if (unlikely(wc->refs[level - 1] == 0)) {
6655                 btrfs_err(fs_info, "Missing references.");
6656                 ret = -EIO;
6657                 goto out_unlock;
6658         }
6659         *lookup_info = 0;
6660
6661         if (wc->stage == DROP_REFERENCE) {
6662                 if (wc->refs[level - 1] > 1) {
6663                         need_account = true;
6664                         if (level == 1 &&
6665                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6666                                 goto skip;
6667
6668                         if (!wc->update_ref ||
6669                             generation <= root->root_key.offset)
6670                                 goto skip;
6671
6672                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6673                                               path->slots[level]);
6674                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6675                         if (ret < 0)
6676                                 goto skip;
6677
6678                         wc->stage = UPDATE_BACKREF;
6679                         wc->shared_level = level - 1;
6680                 }
6681         } else {
6682                 if (level == 1 &&
6683                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6684                         goto skip;
6685         }
6686
6687         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6688                 btrfs_tree_unlock(next);
6689                 free_extent_buffer(next);
6690                 next = NULL;
6691                 *lookup_info = 1;
6692         }
6693
6694         if (!next) {
6695                 if (reada && level == 1)
6696                         reada_walk_down(trans, root, wc, path);
6697                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
6698                                        &first_key);
6699                 if (IS_ERR(next)) {
6700                         return PTR_ERR(next);
6701                 } else if (!extent_buffer_uptodate(next)) {
6702                         free_extent_buffer(next);
6703                         return -EIO;
6704                 }
6705                 btrfs_tree_lock(next);
6706                 btrfs_set_lock_blocking_write(next);
6707         }
6708
6709         level--;
6710         ASSERT(level == btrfs_header_level(next));
6711         if (level != btrfs_header_level(next)) {
6712                 btrfs_err(root->fs_info, "mismatched level");
6713                 ret = -EIO;
6714                 goto out_unlock;
6715         }
6716         path->nodes[level] = next;
6717         path->slots[level] = 0;
6718         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6719         wc->level = level;
6720         if (wc->level == 1)
6721                 wc->reada_slot = 0;
6722         return 0;
6723 skip:
6724         wc->refs[level - 1] = 0;
6725         wc->flags[level - 1] = 0;
6726         if (wc->stage == DROP_REFERENCE) {
6727                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6728                         parent = path->nodes[level]->start;
6729                 } else {
6730                         ASSERT(root->root_key.objectid ==
6731                                btrfs_header_owner(path->nodes[level]));
6732                         if (root->root_key.objectid !=
6733                             btrfs_header_owner(path->nodes[level])) {
6734                                 btrfs_err(root->fs_info,
6735                                                 "mismatched block owner");
6736                                 ret = -EIO;
6737                                 goto out_unlock;
6738                         }
6739                         parent = 0;
6740                 }
6741
6742                 /*
6743                  * If we had a drop_progress we need to verify the refs are set
6744                  * as expected.  If we find our ref then we know that from here
6745                  * on out everything should be correct, and we can clear the
6746                  * ->restarted flag.
6747                  */
6748                 if (wc->restarted) {
6749                         ret = check_ref_exists(trans, root, bytenr, parent,
6750                                                level - 1);
6751                         if (ret < 0)
6752                                 goto out_unlock;
6753                         if (ret == 0)
6754                                 goto no_delete;
6755                         ret = 0;
6756                         wc->restarted = 0;
6757                 }
6758
6759                 /*
6760                  * Reloc tree doesn't contribute to qgroup numbers, and we have
6761                  * already accounted them at merge time (replace_path),
6762                  * thus we could skip expensive subtree trace here.
6763                  */
6764                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
6765                     need_account) {
6766                         ret = btrfs_qgroup_trace_subtree(trans, next,
6767                                                          generation, level - 1);
6768                         if (ret) {
6769                                 btrfs_err_rl(fs_info,
6770                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
6771                                              ret);
6772                         }
6773                 }
6774
6775                 /*
6776                  * We need to update the next key in our walk control so we can
6777                  * update the drop_progress key accordingly.  We don't care if
6778                  * find_next_key doesn't find a key because that means we're at
6779                  * the end and are going to clean up now.
6780                  */
6781                 wc->drop_level = level;
6782                 find_next_key(path, level, &wc->drop_progress);
6783
6784                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
6785                                        fs_info->nodesize, parent);
6786                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
6787                 ret = btrfs_free_extent(trans, &ref);
6788                 if (ret)
6789                         goto out_unlock;
6790         }
6791 no_delete:
6792         *lookup_info = 1;
6793         ret = 1;
6794
6795 out_unlock:
6796         btrfs_tree_unlock(next);
6797         free_extent_buffer(next);
6798
6799         return ret;
6800 }
6801
6802 /*
6803  * helper to process tree block while walking up the tree.
6804  *
6805  * when wc->stage == DROP_REFERENCE, this function drops
6806  * reference count on the block.
6807  *
6808  * when wc->stage == UPDATE_BACKREF, this function changes
6809  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6810  * to UPDATE_BACKREF previously while processing the block.
6811  *
6812  * NOTE: return value 1 means we should stop walking up.
6813  */
6814 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6815                                  struct btrfs_root *root,
6816                                  struct btrfs_path *path,
6817                                  struct walk_control *wc)
6818 {
6819         struct btrfs_fs_info *fs_info = root->fs_info;
6820         int ret;
6821         int level = wc->level;
6822         struct extent_buffer *eb = path->nodes[level];
6823         u64 parent = 0;
6824
6825         if (wc->stage == UPDATE_BACKREF) {
6826                 BUG_ON(wc->shared_level < level);
6827                 if (level < wc->shared_level)
6828                         goto out;
6829
6830                 ret = find_next_key(path, level + 1, &wc->update_progress);
6831                 if (ret > 0)
6832                         wc->update_ref = 0;
6833
6834                 wc->stage = DROP_REFERENCE;
6835                 wc->shared_level = -1;
6836                 path->slots[level] = 0;
6837
6838                 /*
6839                  * check reference count again if the block isn't locked.
6840                  * we should start walking down the tree again if reference
6841                  * count is one.
6842                  */
6843                 if (!path->locks[level]) {
6844                         BUG_ON(level == 0);
6845                         btrfs_tree_lock(eb);
6846                         btrfs_set_lock_blocking_write(eb);
6847                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6848
6849                         ret = btrfs_lookup_extent_info(trans, fs_info,
6850                                                        eb->start, level, 1,
6851                                                        &wc->refs[level],
6852                                                        &wc->flags[level]);
6853                         if (ret < 0) {
6854                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6855                                 path->locks[level] = 0;
6856                                 return ret;
6857                         }
6858                         BUG_ON(wc->refs[level] == 0);
6859                         if (wc->refs[level] == 1) {
6860                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6861                                 path->locks[level] = 0;
6862                                 return 1;
6863                         }
6864                 }
6865         }
6866
6867         /* wc->stage == DROP_REFERENCE */
6868         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6869
6870         if (wc->refs[level] == 1) {
6871                 if (level == 0) {
6872                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6873                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6874                         else
6875                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6876                         BUG_ON(ret); /* -ENOMEM */
6877                         if (is_fstree(root->root_key.objectid)) {
6878                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
6879                                 if (ret) {
6880                                         btrfs_err_rl(fs_info,
6881         "error %d accounting leaf items, quota is out of sync, rescan required",
6882                                              ret);
6883                                 }
6884                         }
6885                 }
6886                 /* make block locked assertion in btrfs_clean_tree_block happy */
6887                 if (!path->locks[level] &&
6888                     btrfs_header_generation(eb) == trans->transid) {
6889                         btrfs_tree_lock(eb);
6890                         btrfs_set_lock_blocking_write(eb);
6891                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6892                 }
6893                 btrfs_clean_tree_block(eb);
6894         }
6895
6896         if (eb == root->node) {
6897                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6898                         parent = eb->start;
6899                 else if (root->root_key.objectid != btrfs_header_owner(eb))
6900                         goto owner_mismatch;
6901         } else {
6902                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6903                         parent = path->nodes[level + 1]->start;
6904                 else if (root->root_key.objectid !=
6905                          btrfs_header_owner(path->nodes[level + 1]))
6906                         goto owner_mismatch;
6907         }
6908
6909         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6910 out:
6911         wc->refs[level] = 0;
6912         wc->flags[level] = 0;
6913         return 0;
6914
6915 owner_mismatch:
6916         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
6917                      btrfs_header_owner(eb), root->root_key.objectid);
6918         return -EUCLEAN;
6919 }
6920
6921 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6922                                    struct btrfs_root *root,
6923                                    struct btrfs_path *path,
6924                                    struct walk_control *wc)
6925 {
6926         int level = wc->level;
6927         int lookup_info = 1;
6928         int ret;
6929
6930         while (level >= 0) {
6931                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6932                 if (ret > 0)
6933                         break;
6934
6935                 if (level == 0)
6936                         break;
6937
6938                 if (path->slots[level] >=
6939                     btrfs_header_nritems(path->nodes[level]))
6940                         break;
6941
6942                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6943                 if (ret > 0) {
6944                         path->slots[level]++;
6945                         continue;
6946                 } else if (ret < 0)
6947                         return ret;
6948                 level = wc->level;
6949         }
6950         return 0;
6951 }
6952
6953 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6954                                  struct btrfs_root *root,
6955                                  struct btrfs_path *path,
6956                                  struct walk_control *wc, int max_level)
6957 {
6958         int level = wc->level;
6959         int ret;
6960
6961         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6962         while (level < max_level && path->nodes[level]) {
6963                 wc->level = level;
6964                 if (path->slots[level] + 1 <
6965                     btrfs_header_nritems(path->nodes[level])) {
6966                         path->slots[level]++;
6967                         return 0;
6968                 } else {
6969                         ret = walk_up_proc(trans, root, path, wc);
6970                         if (ret > 0)
6971                                 return 0;
6972                         if (ret < 0)
6973                                 return ret;
6974
6975                         if (path->locks[level]) {
6976                                 btrfs_tree_unlock_rw(path->nodes[level],
6977                                                      path->locks[level]);
6978                                 path->locks[level] = 0;
6979                         }
6980                         free_extent_buffer(path->nodes[level]);
6981                         path->nodes[level] = NULL;
6982                         level++;
6983                 }
6984         }
6985         return 1;
6986 }
6987
6988 /*
6989  * drop a subvolume tree.
6990  *
6991  * this function traverses the tree freeing any blocks that only
6992  * referenced by the tree.
6993  *
6994  * when a shared tree block is found. this function decreases its
6995  * reference count by one. if update_ref is true, this function
6996  * also make sure backrefs for the shared block and all lower level
6997  * blocks are properly updated.
6998  *
6999  * If called with for_reloc == 0, may exit early with -EAGAIN
7000  */
7001 int btrfs_drop_snapshot(struct btrfs_root *root,
7002                          struct btrfs_block_rsv *block_rsv, int update_ref,
7003                          int for_reloc)
7004 {
7005         struct btrfs_fs_info *fs_info = root->fs_info;
7006         struct btrfs_path *path;
7007         struct btrfs_trans_handle *trans;
7008         struct btrfs_root *tree_root = fs_info->tree_root;
7009         struct btrfs_root_item *root_item = &root->root_item;
7010         struct walk_control *wc;
7011         struct btrfs_key key;
7012         int err = 0;
7013         int ret;
7014         int level;
7015         bool root_dropped = false;
7016
7017         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
7018
7019         path = btrfs_alloc_path();
7020         if (!path) {
7021                 err = -ENOMEM;
7022                 goto out;
7023         }
7024
7025         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7026         if (!wc) {
7027                 btrfs_free_path(path);
7028                 err = -ENOMEM;
7029                 goto out;
7030         }
7031
7032         trans = btrfs_start_transaction(tree_root, 0);
7033         if (IS_ERR(trans)) {
7034                 err = PTR_ERR(trans);
7035                 goto out_free;
7036         }
7037
7038         err = btrfs_run_delayed_items(trans);
7039         if (err)
7040                 goto out_end_trans;
7041
7042         if (block_rsv)
7043                 trans->block_rsv = block_rsv;
7044
7045         /*
7046          * This will help us catch people modifying the fs tree while we're
7047          * dropping it.  It is unsafe to mess with the fs tree while it's being
7048          * dropped as we unlock the root node and parent nodes as we walk down
7049          * the tree, assuming nothing will change.  If something does change
7050          * then we'll have stale information and drop references to blocks we've
7051          * already dropped.
7052          */
7053         set_bit(BTRFS_ROOT_DELETING, &root->state);
7054         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7055                 level = btrfs_header_level(root->node);
7056                 path->nodes[level] = btrfs_lock_root_node(root);
7057                 btrfs_set_lock_blocking_write(path->nodes[level]);
7058                 path->slots[level] = 0;
7059                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7060                 memset(&wc->update_progress, 0,
7061                        sizeof(wc->update_progress));
7062         } else {
7063                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7064                 memcpy(&wc->update_progress, &key,
7065                        sizeof(wc->update_progress));
7066
7067                 level = root_item->drop_level;
7068                 BUG_ON(level == 0);
7069                 path->lowest_level = level;
7070                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7071                 path->lowest_level = 0;
7072                 if (ret < 0) {
7073                         err = ret;
7074                         goto out_end_trans;
7075                 }
7076                 WARN_ON(ret > 0);
7077
7078                 /*
7079                  * unlock our path, this is safe because only this
7080                  * function is allowed to delete this snapshot
7081                  */
7082                 btrfs_unlock_up_safe(path, 0);
7083
7084                 level = btrfs_header_level(root->node);
7085                 while (1) {
7086                         btrfs_tree_lock(path->nodes[level]);
7087                         btrfs_set_lock_blocking_write(path->nodes[level]);
7088                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7089
7090                         ret = btrfs_lookup_extent_info(trans, fs_info,
7091                                                 path->nodes[level]->start,
7092                                                 level, 1, &wc->refs[level],
7093                                                 &wc->flags[level]);
7094                         if (ret < 0) {
7095                                 err = ret;
7096                                 goto out_end_trans;
7097                         }
7098                         BUG_ON(wc->refs[level] == 0);
7099
7100                         if (level == root_item->drop_level)
7101                                 break;
7102
7103                         btrfs_tree_unlock(path->nodes[level]);
7104                         path->locks[level] = 0;
7105                         WARN_ON(wc->refs[level] != 1);
7106                         level--;
7107                 }
7108         }
7109
7110         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
7111         wc->level = level;
7112         wc->shared_level = -1;
7113         wc->stage = DROP_REFERENCE;
7114         wc->update_ref = update_ref;
7115         wc->keep_locks = 0;
7116         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
7117
7118         while (1) {
7119
7120                 ret = walk_down_tree(trans, root, path, wc);
7121                 if (ret < 0) {
7122                         err = ret;
7123                         break;
7124                 }
7125
7126                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7127                 if (ret < 0) {
7128                         err = ret;
7129                         break;
7130                 }
7131
7132                 if (ret > 0) {
7133                         BUG_ON(wc->stage != DROP_REFERENCE);
7134                         break;
7135                 }
7136
7137                 if (wc->stage == DROP_REFERENCE) {
7138                         wc->drop_level = wc->level;
7139                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
7140                                               &wc->drop_progress,
7141                                               path->slots[wc->drop_level]);
7142                 }
7143                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
7144                                       &wc->drop_progress);
7145                 root_item->drop_level = wc->drop_level;
7146
7147                 BUG_ON(wc->level == 0);
7148                 if (btrfs_should_end_transaction(trans) ||
7149                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
7150                         ret = btrfs_update_root(trans, tree_root,
7151                                                 &root->root_key,
7152                                                 root_item);
7153                         if (ret) {
7154                                 btrfs_abort_transaction(trans, ret);
7155                                 err = ret;
7156                                 goto out_end_trans;
7157                         }
7158
7159                         btrfs_end_transaction_throttle(trans);
7160                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
7161                                 btrfs_debug(fs_info,
7162                                             "drop snapshot early exit");
7163                                 err = -EAGAIN;
7164                                 goto out_free;
7165                         }
7166
7167                         trans = btrfs_start_transaction(tree_root, 0);
7168                         if (IS_ERR(trans)) {
7169                                 err = PTR_ERR(trans);
7170                                 goto out_free;
7171                         }
7172                         if (block_rsv)
7173                                 trans->block_rsv = block_rsv;
7174                 }
7175         }
7176         btrfs_release_path(path);
7177         if (err)
7178                 goto out_end_trans;
7179
7180         ret = btrfs_del_root(trans, &root->root_key);
7181         if (ret) {
7182                 btrfs_abort_transaction(trans, ret);
7183                 err = ret;
7184                 goto out_end_trans;
7185         }
7186
7187         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7188                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7189                                       NULL, NULL);
7190                 if (ret < 0) {
7191                         btrfs_abort_transaction(trans, ret);
7192                         err = ret;
7193                         goto out_end_trans;
7194                 } else if (ret > 0) {
7195                         /* if we fail to delete the orphan item this time
7196                          * around, it'll get picked up the next time.
7197                          *
7198                          * The most common failure here is just -ENOENT.
7199                          */
7200                         btrfs_del_orphan_item(trans, tree_root,
7201                                               root->root_key.objectid);
7202                 }
7203         }
7204
7205         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
7206                 btrfs_add_dropped_root(trans, root);
7207         } else {
7208                 free_extent_buffer(root->node);
7209                 free_extent_buffer(root->commit_root);
7210                 btrfs_put_fs_root(root);
7211         }
7212         root_dropped = true;
7213 out_end_trans:
7214         btrfs_end_transaction_throttle(trans);
7215 out_free:
7216         kfree(wc);
7217         btrfs_free_path(path);
7218 out:
7219         /*
7220          * So if we need to stop dropping the snapshot for whatever reason we
7221          * need to make sure to add it back to the dead root list so that we
7222          * keep trying to do the work later.  This also cleans up roots if we
7223          * don't have it in the radix (like when we recover after a power fail
7224          * or unmount) so we don't leak memory.
7225          */
7226         if (!for_reloc && !root_dropped)
7227                 btrfs_add_dead_root(root);
7228         if (err && err != -EAGAIN)
7229                 btrfs_handle_fs_error(fs_info, err, NULL);
7230         return err;
7231 }
7232
7233 /*
7234  * drop subtree rooted at tree block 'node'.
7235  *
7236  * NOTE: this function will unlock and release tree block 'node'
7237  * only used by relocation code
7238  */
7239 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7240                         struct btrfs_root *root,
7241                         struct extent_buffer *node,
7242                         struct extent_buffer *parent)
7243 {
7244         struct btrfs_fs_info *fs_info = root->fs_info;
7245         struct btrfs_path *path;
7246         struct walk_control *wc;
7247         int level;
7248         int parent_level;
7249         int ret = 0;
7250         int wret;
7251
7252         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7253
7254         path = btrfs_alloc_path();
7255         if (!path)
7256                 return -ENOMEM;
7257
7258         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7259         if (!wc) {
7260                 btrfs_free_path(path);
7261                 return -ENOMEM;
7262         }
7263
7264         btrfs_assert_tree_locked(parent);
7265         parent_level = btrfs_header_level(parent);
7266         extent_buffer_get(parent);
7267         path->nodes[parent_level] = parent;
7268         path->slots[parent_level] = btrfs_header_nritems(parent);
7269
7270         btrfs_assert_tree_locked(node);
7271         level = btrfs_header_level(node);
7272         path->nodes[level] = node;
7273         path->slots[level] = 0;
7274         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7275
7276         wc->refs[parent_level] = 1;
7277         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7278         wc->level = level;
7279         wc->shared_level = -1;
7280         wc->stage = DROP_REFERENCE;
7281         wc->update_ref = 0;
7282         wc->keep_locks = 1;
7283         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
7284
7285         while (1) {
7286                 wret = walk_down_tree(trans, root, path, wc);
7287                 if (wret < 0) {
7288                         ret = wret;
7289                         break;
7290                 }
7291
7292                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7293                 if (wret < 0)
7294                         ret = wret;
7295                 if (wret != 0)
7296                         break;
7297         }
7298
7299         kfree(wc);
7300         btrfs_free_path(path);
7301         return ret;
7302 }
7303
7304 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
7305 {
7306         u64 num_devices;
7307         u64 stripped;
7308
7309         /*
7310          * if restripe for this chunk_type is on pick target profile and
7311          * return, otherwise do the usual balance
7312          */
7313         stripped = get_restripe_target(fs_info, flags);
7314         if (stripped)
7315                 return extended_to_chunk(stripped);
7316
7317         num_devices = fs_info->fs_devices->rw_devices;
7318
7319         stripped = BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID56_MASK |
7320                 BTRFS_BLOCK_GROUP_RAID1_MASK | BTRFS_BLOCK_GROUP_RAID10;
7321
7322         if (num_devices == 1) {
7323                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7324                 stripped = flags & ~stripped;
7325
7326                 /* turn raid0 into single device chunks */
7327                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7328                         return stripped;
7329
7330                 /* turn mirroring into duplication */
7331                 if (flags & (BTRFS_BLOCK_GROUP_RAID1_MASK |
7332                              BTRFS_BLOCK_GROUP_RAID10))
7333                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7334         } else {
7335                 /* they already had raid on here, just return */
7336                 if (flags & stripped)
7337                         return flags;
7338
7339                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7340                 stripped = flags & ~stripped;
7341
7342                 /* switch duplicated blocks with raid1 */
7343                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7344                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7345
7346                 /* this is drive concat, leave it alone */
7347         }
7348
7349         return flags;
7350 }
7351
7352 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7353 {
7354         struct btrfs_space_info *sinfo = cache->space_info;
7355         u64 num_bytes;
7356         u64 sinfo_used;
7357         u64 min_allocable_bytes;
7358         int ret = -ENOSPC;
7359
7360         /*
7361          * We need some metadata space and system metadata space for
7362          * allocating chunks in some corner cases until we force to set
7363          * it to be readonly.
7364          */
7365         if ((sinfo->flags &
7366              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7367             !force)
7368                 min_allocable_bytes = SZ_1M;
7369         else
7370                 min_allocable_bytes = 0;
7371
7372         spin_lock(&sinfo->lock);
7373         spin_lock(&cache->lock);
7374
7375         if (cache->ro) {
7376                 cache->ro++;
7377                 ret = 0;
7378                 goto out;
7379         }
7380
7381         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7382                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7383         sinfo_used = btrfs_space_info_used(sinfo, true);
7384
7385         if (sinfo_used + num_bytes + min_allocable_bytes <=
7386             sinfo->total_bytes) {
7387                 sinfo->bytes_readonly += num_bytes;
7388                 cache->ro++;
7389                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
7390                 ret = 0;
7391         }
7392 out:
7393         spin_unlock(&cache->lock);
7394         spin_unlock(&sinfo->lock);
7395         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
7396                 btrfs_info(cache->fs_info,
7397                         "unable to make block group %llu ro",
7398                         cache->key.objectid);
7399                 btrfs_info(cache->fs_info,
7400                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
7401                         sinfo_used, num_bytes, min_allocable_bytes);
7402                 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
7403         }
7404         return ret;
7405 }
7406
7407 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
7408
7409 {
7410         struct btrfs_fs_info *fs_info = cache->fs_info;
7411         struct btrfs_trans_handle *trans;
7412         u64 alloc_flags;
7413         int ret;
7414
7415 again:
7416         trans = btrfs_join_transaction(fs_info->extent_root);
7417         if (IS_ERR(trans))
7418                 return PTR_ERR(trans);
7419
7420         /*
7421          * we're not allowed to set block groups readonly after the dirty
7422          * block groups cache has started writing.  If it already started,
7423          * back off and let this transaction commit
7424          */
7425         mutex_lock(&fs_info->ro_block_group_mutex);
7426         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
7427                 u64 transid = trans->transid;
7428
7429                 mutex_unlock(&fs_info->ro_block_group_mutex);
7430                 btrfs_end_transaction(trans);
7431
7432                 ret = btrfs_wait_for_commit(fs_info, transid);
7433                 if (ret)
7434                         return ret;
7435                 goto again;
7436         }
7437
7438         /*
7439          * if we are changing raid levels, try to allocate a corresponding
7440          * block group with the new raid level.
7441          */
7442         alloc_flags = update_block_group_flags(fs_info, cache->flags);
7443         if (alloc_flags != cache->flags) {
7444                 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
7445                 /*
7446                  * ENOSPC is allowed here, we may have enough space
7447                  * already allocated at the new raid level to
7448                  * carry on
7449                  */
7450                 if (ret == -ENOSPC)
7451                         ret = 0;
7452                 if (ret < 0)
7453                         goto out;
7454         }
7455
7456         ret = inc_block_group_ro(cache, 0);
7457         if (!ret)
7458                 goto out;
7459         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
7460         ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
7461         if (ret < 0)
7462                 goto out;
7463         ret = inc_block_group_ro(cache, 0);
7464 out:
7465         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
7466                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
7467                 mutex_lock(&fs_info->chunk_mutex);
7468                 check_system_chunk(trans, alloc_flags);
7469                 mutex_unlock(&fs_info->chunk_mutex);
7470         }
7471         mutex_unlock(&fs_info->ro_block_group_mutex);
7472
7473         btrfs_end_transaction(trans);
7474         return ret;
7475 }
7476
7477 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
7478 {
7479         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
7480
7481         return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
7482 }
7483
7484 /*
7485  * helper to account the unused space of all the readonly block group in the
7486  * space_info. takes mirrors into account.
7487  */
7488 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7489 {
7490         struct btrfs_block_group_cache *block_group;
7491         u64 free_bytes = 0;
7492         int factor;
7493
7494         /* It's df, we don't care if it's racy */
7495         if (list_empty(&sinfo->ro_bgs))
7496                 return 0;
7497
7498         spin_lock(&sinfo->lock);
7499         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
7500                 spin_lock(&block_group->lock);
7501
7502                 if (!block_group->ro) {
7503                         spin_unlock(&block_group->lock);
7504                         continue;
7505                 }
7506
7507                 factor = btrfs_bg_type_to_factor(block_group->flags);
7508                 free_bytes += (block_group->key.offset -
7509                                btrfs_block_group_used(&block_group->item)) *
7510                                factor;
7511
7512                 spin_unlock(&block_group->lock);
7513         }
7514         spin_unlock(&sinfo->lock);
7515
7516         return free_bytes;
7517 }
7518
7519 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
7520 {
7521         struct btrfs_space_info *sinfo = cache->space_info;
7522         u64 num_bytes;
7523
7524         BUG_ON(!cache->ro);
7525
7526         spin_lock(&sinfo->lock);
7527         spin_lock(&cache->lock);
7528         if (!--cache->ro) {
7529                 num_bytes = cache->key.offset - cache->reserved -
7530                             cache->pinned - cache->bytes_super -
7531                             btrfs_block_group_used(&cache->item);
7532                 sinfo->bytes_readonly -= num_bytes;
7533                 list_del_init(&cache->ro_list);
7534         }
7535         spin_unlock(&cache->lock);
7536         spin_unlock(&sinfo->lock);
7537 }
7538
7539 /*
7540  * Checks to see if it's even possible to relocate this block group.
7541  *
7542  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7543  * ok to go ahead and try.
7544  */
7545 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
7546 {
7547         struct btrfs_block_group_cache *block_group;
7548         struct btrfs_space_info *space_info;
7549         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
7550         struct btrfs_device *device;
7551         u64 min_free;
7552         u64 dev_min = 1;
7553         u64 dev_nr = 0;
7554         u64 target;
7555         int debug;
7556         int index;
7557         int full = 0;
7558         int ret = 0;
7559
7560         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
7561
7562         block_group = btrfs_lookup_block_group(fs_info, bytenr);
7563
7564         /* odd, couldn't find the block group, leave it alone */
7565         if (!block_group) {
7566                 if (debug)
7567                         btrfs_warn(fs_info,
7568                                    "can't find block group for bytenr %llu",
7569                                    bytenr);
7570                 return -1;
7571         }
7572
7573         min_free = btrfs_block_group_used(&block_group->item);
7574
7575         /* no bytes used, we're good */
7576         if (!min_free)
7577                 goto out;
7578
7579         space_info = block_group->space_info;
7580         spin_lock(&space_info->lock);
7581
7582         full = space_info->full;
7583
7584         /*
7585          * if this is the last block group we have in this space, we can't
7586          * relocate it unless we're able to allocate a new chunk below.
7587          *
7588          * Otherwise, we need to make sure we have room in the space to handle
7589          * all of the extents from this block group.  If we can, we're good
7590          */
7591         if ((space_info->total_bytes != block_group->key.offset) &&
7592             (btrfs_space_info_used(space_info, false) + min_free <
7593              space_info->total_bytes)) {
7594                 spin_unlock(&space_info->lock);
7595                 goto out;
7596         }
7597         spin_unlock(&space_info->lock);
7598
7599         /*
7600          * ok we don't have enough space, but maybe we have free space on our
7601          * devices to allocate new chunks for relocation, so loop through our
7602          * alloc devices and guess if we have enough space.  if this block
7603          * group is going to be restriped, run checks against the target
7604          * profile instead of the current one.
7605          */
7606         ret = -1;
7607
7608         /*
7609          * index:
7610          *      0: raid10
7611          *      1: raid1
7612          *      2: dup
7613          *      3: raid0
7614          *      4: single
7615          */
7616         target = get_restripe_target(fs_info, block_group->flags);
7617         if (target) {
7618                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
7619         } else {
7620                 /*
7621                  * this is just a balance, so if we were marked as full
7622                  * we know there is no space for a new chunk
7623                  */
7624                 if (full) {
7625                         if (debug)
7626                                 btrfs_warn(fs_info,
7627                                            "no space to alloc new chunk for block group %llu",
7628                                            block_group->key.objectid);
7629                         goto out;
7630                 }
7631
7632                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
7633         }
7634
7635         if (index == BTRFS_RAID_RAID10) {
7636                 dev_min = 4;
7637                 /* Divide by 2 */
7638                 min_free >>= 1;
7639         } else if (index == BTRFS_RAID_RAID1) {
7640                 dev_min = 2;
7641         } else if (index == BTRFS_RAID_DUP) {
7642                 /* Multiply by 2 */
7643                 min_free <<= 1;
7644         } else if (index == BTRFS_RAID_RAID0) {
7645                 dev_min = fs_devices->rw_devices;
7646                 min_free = div64_u64(min_free, dev_min);
7647         }
7648
7649         mutex_lock(&fs_info->chunk_mutex);
7650         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7651                 u64 dev_offset;
7652
7653                 /*
7654                  * check to make sure we can actually find a chunk with enough
7655                  * space to fit our block group in.
7656                  */
7657                 if (device->total_bytes > device->bytes_used + min_free &&
7658                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
7659                         ret = find_free_dev_extent(device, min_free,
7660                                                    &dev_offset, NULL);
7661                         if (!ret)
7662                                 dev_nr++;
7663
7664                         if (dev_nr >= dev_min)
7665                                 break;
7666
7667                         ret = -1;
7668                 }
7669         }
7670         if (debug && ret == -1)
7671                 btrfs_warn(fs_info,
7672                            "no space to allocate a new chunk for block group %llu",
7673                            block_group->key.objectid);
7674         mutex_unlock(&fs_info->chunk_mutex);
7675 out:
7676         btrfs_put_block_group(block_group);
7677         return ret;
7678 }
7679
7680 static int find_first_block_group(struct btrfs_fs_info *fs_info,
7681                                   struct btrfs_path *path,
7682                                   struct btrfs_key *key)
7683 {
7684         struct btrfs_root *root = fs_info->extent_root;
7685         int ret = 0;
7686         struct btrfs_key found_key;
7687         struct extent_buffer *leaf;
7688         struct btrfs_block_group_item bg;
7689         u64 flags;
7690         int slot;
7691
7692         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7693         if (ret < 0)
7694                 goto out;
7695
7696         while (1) {
7697                 slot = path->slots[0];
7698                 leaf = path->nodes[0];
7699                 if (slot >= btrfs_header_nritems(leaf)) {
7700                         ret = btrfs_next_leaf(root, path);
7701                         if (ret == 0)
7702                                 continue;
7703                         if (ret < 0)
7704                                 goto out;
7705                         break;
7706                 }
7707                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7708
7709                 if (found_key.objectid >= key->objectid &&
7710                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7711                         struct extent_map_tree *em_tree;
7712                         struct extent_map *em;
7713
7714                         em_tree = &root->fs_info->mapping_tree;
7715                         read_lock(&em_tree->lock);
7716                         em = lookup_extent_mapping(em_tree, found_key.objectid,
7717                                                    found_key.offset);
7718                         read_unlock(&em_tree->lock);
7719                         if (!em) {
7720                                 btrfs_err(fs_info,
7721                         "logical %llu len %llu found bg but no related chunk",
7722                                           found_key.objectid, found_key.offset);
7723                                 ret = -ENOENT;
7724                         } else if (em->start != found_key.objectid ||
7725                                    em->len != found_key.offset) {
7726                                 btrfs_err(fs_info,
7727                 "block group %llu len %llu mismatch with chunk %llu len %llu",
7728                                           found_key.objectid, found_key.offset,
7729                                           em->start, em->len);
7730                                 ret = -EUCLEAN;
7731                         } else {
7732                                 read_extent_buffer(leaf, &bg,
7733                                         btrfs_item_ptr_offset(leaf, slot),
7734                                         sizeof(bg));
7735                                 flags = btrfs_block_group_flags(&bg) &
7736                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
7737
7738                                 if (flags != (em->map_lookup->type &
7739                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
7740                                         btrfs_err(fs_info,
7741 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
7742                                                 found_key.objectid,
7743                                                 found_key.offset, flags,
7744                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
7745                                                  em->map_lookup->type));
7746                                         ret = -EUCLEAN;
7747                                 } else {
7748                                         ret = 0;
7749                                 }
7750                         }
7751                         free_extent_map(em);
7752                         goto out;
7753                 }
7754                 path->slots[0]++;
7755         }
7756 out:
7757         return ret;
7758 }
7759
7760 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7761 {
7762         struct btrfs_block_group_cache *block_group;
7763         u64 last = 0;
7764
7765         while (1) {
7766                 struct inode *inode;
7767
7768                 block_group = btrfs_lookup_first_block_group(info, last);
7769                 while (block_group) {
7770                         wait_block_group_cache_done(block_group);
7771                         spin_lock(&block_group->lock);
7772                         if (block_group->iref)
7773                                 break;
7774                         spin_unlock(&block_group->lock);
7775                         block_group = next_block_group(block_group);
7776                 }
7777                 if (!block_group) {
7778                         if (last == 0)
7779                                 break;
7780                         last = 0;
7781                         continue;
7782                 }
7783
7784                 inode = block_group->inode;
7785                 block_group->iref = 0;
7786                 block_group->inode = NULL;
7787                 spin_unlock(&block_group->lock);
7788                 ASSERT(block_group->io_ctl.inode == NULL);
7789                 iput(inode);
7790                 last = block_group->key.objectid + block_group->key.offset;
7791                 btrfs_put_block_group(block_group);
7792         }
7793 }
7794
7795 /*
7796  * Must be called only after stopping all workers, since we could have block
7797  * group caching kthreads running, and therefore they could race with us if we
7798  * freed the block groups before stopping them.
7799  */
7800 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7801 {
7802         struct btrfs_block_group_cache *block_group;
7803         struct btrfs_space_info *space_info;
7804         struct btrfs_caching_control *caching_ctl;
7805         struct rb_node *n;
7806
7807         down_write(&info->commit_root_sem);
7808         while (!list_empty(&info->caching_block_groups)) {
7809                 caching_ctl = list_entry(info->caching_block_groups.next,
7810                                          struct btrfs_caching_control, list);
7811                 list_del(&caching_ctl->list);
7812                 put_caching_control(caching_ctl);
7813         }
7814         up_write(&info->commit_root_sem);
7815
7816         spin_lock(&info->unused_bgs_lock);
7817         while (!list_empty(&info->unused_bgs)) {
7818                 block_group = list_first_entry(&info->unused_bgs,
7819                                                struct btrfs_block_group_cache,
7820                                                bg_list);
7821                 list_del_init(&block_group->bg_list);
7822                 btrfs_put_block_group(block_group);
7823         }
7824         spin_unlock(&info->unused_bgs_lock);
7825
7826         spin_lock(&info->block_group_cache_lock);
7827         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7828                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7829                                        cache_node);
7830                 rb_erase(&block_group->cache_node,
7831                          &info->block_group_cache_tree);
7832                 RB_CLEAR_NODE(&block_group->cache_node);
7833                 spin_unlock(&info->block_group_cache_lock);
7834
7835                 down_write(&block_group->space_info->groups_sem);
7836                 list_del(&block_group->list);
7837                 up_write(&block_group->space_info->groups_sem);
7838
7839                 /*
7840                  * We haven't cached this block group, which means we could
7841                  * possibly have excluded extents on this block group.
7842                  */
7843                 if (block_group->cached == BTRFS_CACHE_NO ||
7844                     block_group->cached == BTRFS_CACHE_ERROR)
7845                         free_excluded_extents(block_group);
7846
7847                 btrfs_remove_free_space_cache(block_group);
7848                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
7849                 ASSERT(list_empty(&block_group->dirty_list));
7850                 ASSERT(list_empty(&block_group->io_list));
7851                 ASSERT(list_empty(&block_group->bg_list));
7852                 ASSERT(atomic_read(&block_group->count) == 1);
7853                 btrfs_put_block_group(block_group);
7854
7855                 spin_lock(&info->block_group_cache_lock);
7856         }
7857         spin_unlock(&info->block_group_cache_lock);
7858
7859         /* now that all the block groups are freed, go through and
7860          * free all the space_info structs.  This is only called during
7861          * the final stages of unmount, and so we know nobody is
7862          * using them.  We call synchronize_rcu() once before we start,
7863          * just to be on the safe side.
7864          */
7865         synchronize_rcu();
7866
7867         btrfs_release_global_block_rsv(info);
7868
7869         while (!list_empty(&info->space_info)) {
7870                 int i;
7871
7872                 space_info = list_entry(info->space_info.next,
7873                                         struct btrfs_space_info,
7874                                         list);
7875
7876                 /*
7877                  * Do not hide this behind enospc_debug, this is actually
7878                  * important and indicates a real bug if this happens.
7879                  */
7880                 if (WARN_ON(space_info->bytes_pinned > 0 ||
7881                             space_info->bytes_reserved > 0 ||
7882                             space_info->bytes_may_use > 0))
7883                         btrfs_dump_space_info(info, space_info, 0, 0);
7884                 list_del(&space_info->list);
7885                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
7886                         struct kobject *kobj;
7887                         kobj = space_info->block_group_kobjs[i];
7888                         space_info->block_group_kobjs[i] = NULL;
7889                         if (kobj) {
7890                                 kobject_del(kobj);
7891                                 kobject_put(kobj);
7892                         }
7893                 }
7894                 kobject_del(&space_info->kobj);
7895                 kobject_put(&space_info->kobj);
7896         }
7897         return 0;
7898 }
7899
7900 static void link_block_group(struct btrfs_block_group_cache *cache)
7901 {
7902         struct btrfs_space_info *space_info = cache->space_info;
7903         struct btrfs_fs_info *fs_info = cache->fs_info;
7904         int index = btrfs_bg_flags_to_raid_index(cache->flags);
7905         bool first = false;
7906
7907         down_write(&space_info->groups_sem);
7908         if (list_empty(&space_info->block_groups[index]))
7909                 first = true;
7910         list_add_tail(&cache->list, &space_info->block_groups[index]);
7911         up_write(&space_info->groups_sem);
7912
7913         if (first) {
7914                 struct raid_kobject *rkobj;
7915                 unsigned int nofs_flag;
7916                 int ret;
7917
7918                 /*
7919                  * Setup a NOFS context because kobject_add(), deep in its call
7920                  * chain, does GFP_KERNEL allocations, and we are often called
7921                  * in a context where if reclaim is triggered we can deadlock
7922                  * (we are either holding a transaction handle or some lock
7923                  * required for a transaction commit).
7924                  */
7925                 nofs_flag = memalloc_nofs_save();
7926                 rkobj = kzalloc(sizeof(*rkobj), GFP_KERNEL);
7927                 if (!rkobj) {
7928                         memalloc_nofs_restore(nofs_flag);
7929                         btrfs_warn(cache->fs_info,
7930                                 "couldn't alloc memory for raid level kobject");
7931                         return;
7932                 }
7933                 rkobj->flags = cache->flags;
7934                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
7935                 ret = kobject_add(&rkobj->kobj, &space_info->kobj, "%s",
7936                                   btrfs_bg_type_to_raid_name(rkobj->flags));
7937                 memalloc_nofs_restore(nofs_flag);
7938                 if (ret) {
7939                         kobject_put(&rkobj->kobj);
7940                         btrfs_warn(fs_info,
7941                            "failed to add kobject for block cache, ignoring");
7942                         return;
7943                 }
7944                 space_info->block_group_kobjs[index] = &rkobj->kobj;
7945         }
7946 }
7947
7948 static struct btrfs_block_group_cache *
7949 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
7950                                u64 start, u64 size)
7951 {
7952         struct btrfs_block_group_cache *cache;
7953
7954         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7955         if (!cache)
7956                 return NULL;
7957
7958         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7959                                         GFP_NOFS);
7960         if (!cache->free_space_ctl) {
7961                 kfree(cache);
7962                 return NULL;
7963         }
7964
7965         cache->key.objectid = start;
7966         cache->key.offset = size;
7967         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7968
7969         cache->fs_info = fs_info;
7970         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
7971         set_free_space_tree_thresholds(cache);
7972
7973         atomic_set(&cache->count, 1);
7974         spin_lock_init(&cache->lock);
7975         init_rwsem(&cache->data_rwsem);
7976         INIT_LIST_HEAD(&cache->list);
7977         INIT_LIST_HEAD(&cache->cluster_list);
7978         INIT_LIST_HEAD(&cache->bg_list);
7979         INIT_LIST_HEAD(&cache->ro_list);
7980         INIT_LIST_HEAD(&cache->dirty_list);
7981         INIT_LIST_HEAD(&cache->io_list);
7982         btrfs_init_free_space_ctl(cache);
7983         atomic_set(&cache->trimming, 0);
7984         mutex_init(&cache->free_space_lock);
7985         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
7986
7987         return cache;
7988 }
7989
7990
7991 /*
7992  * Iterate all chunks and verify that each of them has the corresponding block
7993  * group
7994  */
7995 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
7996 {
7997         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
7998         struct extent_map *em;
7999         struct btrfs_block_group_cache *bg;
8000         u64 start = 0;
8001         int ret = 0;
8002
8003         while (1) {
8004                 read_lock(&map_tree->lock);
8005                 /*
8006                  * lookup_extent_mapping will return the first extent map
8007                  * intersecting the range, so setting @len to 1 is enough to
8008                  * get the first chunk.
8009                  */
8010                 em = lookup_extent_mapping(map_tree, start, 1);
8011                 read_unlock(&map_tree->lock);
8012                 if (!em)
8013                         break;
8014
8015                 bg = btrfs_lookup_block_group(fs_info, em->start);
8016                 if (!bg) {
8017                         btrfs_err(fs_info,
8018         "chunk start=%llu len=%llu doesn't have corresponding block group",
8019                                      em->start, em->len);
8020                         ret = -EUCLEAN;
8021                         free_extent_map(em);
8022                         break;
8023                 }
8024                 if (bg->key.objectid != em->start ||
8025                     bg->key.offset != em->len ||
8026                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
8027                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
8028                         btrfs_err(fs_info,
8029 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
8030                                 em->start, em->len,
8031                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
8032                                 bg->key.objectid, bg->key.offset,
8033                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
8034                         ret = -EUCLEAN;
8035                         free_extent_map(em);
8036                         btrfs_put_block_group(bg);
8037                         break;
8038                 }
8039                 start = em->start + em->len;
8040                 free_extent_map(em);
8041                 btrfs_put_block_group(bg);
8042         }
8043         return ret;
8044 }
8045
8046 int btrfs_read_block_groups(struct btrfs_fs_info *info)
8047 {
8048         struct btrfs_path *path;
8049         int ret;
8050         struct btrfs_block_group_cache *cache;
8051         struct btrfs_space_info *space_info;
8052         struct btrfs_key key;
8053         struct btrfs_key found_key;
8054         struct extent_buffer *leaf;
8055         int need_clear = 0;
8056         u64 cache_gen;
8057         u64 feature;
8058         int mixed;
8059
8060         feature = btrfs_super_incompat_flags(info->super_copy);
8061         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
8062
8063         key.objectid = 0;
8064         key.offset = 0;
8065         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8066         path = btrfs_alloc_path();
8067         if (!path)
8068                 return -ENOMEM;
8069         path->reada = READA_FORWARD;
8070
8071         cache_gen = btrfs_super_cache_generation(info->super_copy);
8072         if (btrfs_test_opt(info, SPACE_CACHE) &&
8073             btrfs_super_generation(info->super_copy) != cache_gen)
8074                 need_clear = 1;
8075         if (btrfs_test_opt(info, CLEAR_CACHE))
8076                 need_clear = 1;
8077
8078         while (1) {
8079                 ret = find_first_block_group(info, path, &key);
8080                 if (ret > 0)
8081                         break;
8082                 if (ret != 0)
8083                         goto error;
8084
8085                 leaf = path->nodes[0];
8086                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8087
8088                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
8089                                                        found_key.offset);
8090                 if (!cache) {
8091                         ret = -ENOMEM;
8092                         goto error;
8093                 }
8094
8095                 if (need_clear) {
8096                         /*
8097                          * When we mount with old space cache, we need to
8098                          * set BTRFS_DC_CLEAR and set dirty flag.
8099                          *
8100                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8101                          *    truncate the old free space cache inode and
8102                          *    setup a new one.
8103                          * b) Setting 'dirty flag' makes sure that we flush
8104                          *    the new space cache info onto disk.
8105                          */
8106                         if (btrfs_test_opt(info, SPACE_CACHE))
8107                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
8108                 }
8109
8110                 read_extent_buffer(leaf, &cache->item,
8111                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8112                                    sizeof(cache->item));
8113                 cache->flags = btrfs_block_group_flags(&cache->item);
8114                 if (!mixed &&
8115                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
8116                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
8117                         btrfs_err(info,
8118 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
8119                                   cache->key.objectid);
8120                         btrfs_put_block_group(cache);
8121                         ret = -EINVAL;
8122                         goto error;
8123                 }
8124
8125                 key.objectid = found_key.objectid + found_key.offset;
8126                 btrfs_release_path(path);
8127
8128                 /*
8129                  * We need to exclude the super stripes now so that the space
8130                  * info has super bytes accounted for, otherwise we'll think
8131                  * we have more space than we actually do.
8132                  */
8133                 ret = exclude_super_stripes(cache);
8134                 if (ret) {
8135                         /*
8136                          * We may have excluded something, so call this just in
8137                          * case.
8138                          */
8139                         free_excluded_extents(cache);
8140                         btrfs_put_block_group(cache);
8141                         goto error;
8142                 }
8143
8144                 /*
8145                  * check for two cases, either we are full, and therefore
8146                  * don't need to bother with the caching work since we won't
8147                  * find any space, or we are empty, and we can just add all
8148                  * the space in and be done with it.  This saves us _a_lot_ of
8149                  * time, particularly in the full case.
8150                  */
8151                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8152                         cache->last_byte_to_unpin = (u64)-1;
8153                         cache->cached = BTRFS_CACHE_FINISHED;
8154                         free_excluded_extents(cache);
8155                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8156                         cache->last_byte_to_unpin = (u64)-1;
8157                         cache->cached = BTRFS_CACHE_FINISHED;
8158                         add_new_free_space(cache, found_key.objectid,
8159                                            found_key.objectid +
8160                                            found_key.offset);
8161                         free_excluded_extents(cache);
8162                 }
8163
8164                 ret = btrfs_add_block_group_cache(info, cache);
8165                 if (ret) {
8166                         btrfs_remove_free_space_cache(cache);
8167                         btrfs_put_block_group(cache);
8168                         goto error;
8169                 }
8170
8171                 trace_btrfs_add_block_group(info, cache, 0);
8172                 btrfs_update_space_info(info, cache->flags, found_key.offset,
8173                                         btrfs_block_group_used(&cache->item),
8174                                         cache->bytes_super, &space_info);
8175
8176                 cache->space_info = space_info;
8177
8178                 link_block_group(cache);
8179
8180                 set_avail_alloc_bits(info, cache->flags);
8181                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
8182                         inc_block_group_ro(cache, 1);
8183                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8184                         ASSERT(list_empty(&cache->bg_list));
8185                         btrfs_mark_bg_unused(cache);
8186                 }
8187         }
8188
8189         list_for_each_entry_rcu(space_info, &info->space_info, list) {
8190                 if (!(get_alloc_profile(info, space_info->flags) &
8191                       (BTRFS_BLOCK_GROUP_RAID10 |
8192                        BTRFS_BLOCK_GROUP_RAID1_MASK |
8193                        BTRFS_BLOCK_GROUP_RAID56_MASK |
8194                        BTRFS_BLOCK_GROUP_DUP)))
8195                         continue;
8196                 /*
8197                  * avoid allocating from un-mirrored block group if there are
8198                  * mirrored block groups.
8199                  */
8200                 list_for_each_entry(cache,
8201                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8202                                 list)
8203                         inc_block_group_ro(cache, 1);
8204                 list_for_each_entry(cache,
8205                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8206                                 list)
8207                         inc_block_group_ro(cache, 1);
8208         }
8209
8210         btrfs_init_global_block_rsv(info);
8211         ret = check_chunk_block_group_mappings(info);
8212 error:
8213         btrfs_free_path(path);
8214         return ret;
8215 }
8216
8217 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
8218 {
8219         struct btrfs_fs_info *fs_info = trans->fs_info;
8220         struct btrfs_block_group_cache *block_group;
8221         struct btrfs_root *extent_root = fs_info->extent_root;
8222         struct btrfs_block_group_item item;
8223         struct btrfs_key key;
8224         int ret = 0;
8225
8226         if (!trans->can_flush_pending_bgs)
8227                 return;
8228
8229         while (!list_empty(&trans->new_bgs)) {
8230                 block_group = list_first_entry(&trans->new_bgs,
8231                                                struct btrfs_block_group_cache,
8232                                                bg_list);
8233                 if (ret)
8234                         goto next;
8235
8236                 spin_lock(&block_group->lock);
8237                 memcpy(&item, &block_group->item, sizeof(item));
8238                 memcpy(&key, &block_group->key, sizeof(key));
8239                 spin_unlock(&block_group->lock);
8240
8241                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8242                                         sizeof(item));
8243                 if (ret)
8244                         btrfs_abort_transaction(trans, ret);
8245                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
8246                 if (ret)
8247                         btrfs_abort_transaction(trans, ret);
8248                 add_block_group_free_space(trans, block_group);
8249                 /* already aborted the transaction if it failed. */
8250 next:
8251                 btrfs_delayed_refs_rsv_release(fs_info, 1);
8252                 list_del_init(&block_group->bg_list);
8253         }
8254         btrfs_trans_release_chunk_metadata(trans);
8255 }
8256
8257 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
8258                            u64 type, u64 chunk_offset, u64 size)
8259 {
8260         struct btrfs_fs_info *fs_info = trans->fs_info;
8261         struct btrfs_block_group_cache *cache;
8262         int ret;
8263
8264         btrfs_set_log_full_commit(trans);
8265
8266         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
8267         if (!cache)
8268                 return -ENOMEM;
8269
8270         btrfs_set_block_group_used(&cache->item, bytes_used);
8271         btrfs_set_block_group_chunk_objectid(&cache->item,
8272                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
8273         btrfs_set_block_group_flags(&cache->item, type);
8274
8275         cache->flags = type;
8276         cache->last_byte_to_unpin = (u64)-1;
8277         cache->cached = BTRFS_CACHE_FINISHED;
8278         cache->needs_free_space = 1;
8279         ret = exclude_super_stripes(cache);
8280         if (ret) {
8281                 /*
8282                  * We may have excluded something, so call this just in
8283                  * case.
8284                  */
8285                 free_excluded_extents(cache);
8286                 btrfs_put_block_group(cache);
8287                 return ret;
8288         }
8289
8290         add_new_free_space(cache, chunk_offset, chunk_offset + size);
8291
8292         free_excluded_extents(cache);
8293
8294 #ifdef CONFIG_BTRFS_DEBUG
8295         if (btrfs_should_fragment_free_space(cache)) {
8296                 u64 new_bytes_used = size - bytes_used;
8297
8298                 bytes_used += new_bytes_used >> 1;
8299                 fragment_free_space(cache);
8300         }
8301 #endif
8302         /*
8303          * Ensure the corresponding space_info object is created and
8304          * assigned to our block group. We want our bg to be added to the rbtree
8305          * with its ->space_info set.
8306          */
8307         cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
8308         ASSERT(cache->space_info);
8309
8310         ret = btrfs_add_block_group_cache(fs_info, cache);
8311         if (ret) {
8312                 btrfs_remove_free_space_cache(cache);
8313                 btrfs_put_block_group(cache);
8314                 return ret;
8315         }
8316
8317         /*
8318          * Now that our block group has its ->space_info set and is inserted in
8319          * the rbtree, update the space info's counters.
8320          */
8321         trace_btrfs_add_block_group(fs_info, cache, 1);
8322         btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
8323                                 cache->bytes_super, &cache->space_info);
8324         btrfs_update_global_block_rsv(fs_info);
8325
8326         link_block_group(cache);
8327
8328         list_add_tail(&cache->bg_list, &trans->new_bgs);
8329         trans->delayed_ref_updates++;
8330         btrfs_update_delayed_refs_rsv(trans);
8331
8332         set_avail_alloc_bits(fs_info, type);
8333         return 0;
8334 }
8335
8336 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8337 {
8338         u64 extra_flags = chunk_to_extended(flags) &
8339                                 BTRFS_EXTENDED_PROFILE_MASK;
8340
8341         write_seqlock(&fs_info->profiles_lock);
8342         if (flags & BTRFS_BLOCK_GROUP_DATA)
8343                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8344         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8345                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8346         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8347                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8348         write_sequnlock(&fs_info->profiles_lock);
8349 }
8350
8351 /*
8352  * Clear incompat bits for the following feature(s):
8353  *
8354  * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
8355  *            in the whole filesystem
8356  */
8357 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
8358 {
8359         if (flags & BTRFS_BLOCK_GROUP_RAID56_MASK) {
8360                 struct list_head *head = &fs_info->space_info;
8361                 struct btrfs_space_info *sinfo;
8362
8363                 list_for_each_entry_rcu(sinfo, head, list) {
8364                         bool found = false;
8365
8366                         down_read(&sinfo->groups_sem);
8367                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
8368                                 found = true;
8369                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
8370                                 found = true;
8371                         up_read(&sinfo->groups_sem);
8372
8373                         if (found)
8374                                 return;
8375                 }
8376                 btrfs_clear_fs_incompat(fs_info, RAID56);
8377         }
8378 }
8379
8380 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8381                              u64 group_start, struct extent_map *em)
8382 {
8383         struct btrfs_fs_info *fs_info = trans->fs_info;
8384         struct btrfs_root *root = fs_info->extent_root;
8385         struct btrfs_path *path;
8386         struct btrfs_block_group_cache *block_group;
8387         struct btrfs_free_cluster *cluster;
8388         struct btrfs_root *tree_root = fs_info->tree_root;
8389         struct btrfs_key key;
8390         struct inode *inode;
8391         struct kobject *kobj = NULL;
8392         int ret;
8393         int index;
8394         int factor;
8395         struct btrfs_caching_control *caching_ctl = NULL;
8396         bool remove_em;
8397         bool remove_rsv = false;
8398
8399         block_group = btrfs_lookup_block_group(fs_info, group_start);
8400         BUG_ON(!block_group);
8401         BUG_ON(!block_group->ro);
8402
8403         trace_btrfs_remove_block_group(block_group);
8404         /*
8405          * Free the reserved super bytes from this block group before
8406          * remove it.
8407          */
8408         free_excluded_extents(block_group);
8409         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
8410                                   block_group->key.offset);
8411
8412         memcpy(&key, &block_group->key, sizeof(key));
8413         index = btrfs_bg_flags_to_raid_index(block_group->flags);
8414         factor = btrfs_bg_type_to_factor(block_group->flags);
8415
8416         /* make sure this block group isn't part of an allocation cluster */
8417         cluster = &fs_info->data_alloc_cluster;
8418         spin_lock(&cluster->refill_lock);
8419         btrfs_return_cluster_to_free_space(block_group, cluster);
8420         spin_unlock(&cluster->refill_lock);
8421
8422         /*
8423          * make sure this block group isn't part of a metadata
8424          * allocation cluster
8425          */
8426         cluster = &fs_info->meta_alloc_cluster;
8427         spin_lock(&cluster->refill_lock);
8428         btrfs_return_cluster_to_free_space(block_group, cluster);
8429         spin_unlock(&cluster->refill_lock);
8430
8431         path = btrfs_alloc_path();
8432         if (!path) {
8433                 ret = -ENOMEM;
8434                 goto out;
8435         }
8436
8437         /*
8438          * get the inode first so any iput calls done for the io_list
8439          * aren't the final iput (no unlinks allowed now)
8440          */
8441         inode = lookup_free_space_inode(block_group, path);
8442
8443         mutex_lock(&trans->transaction->cache_write_mutex);
8444         /*
8445          * Make sure our free space cache IO is done before removing the
8446          * free space inode
8447          */
8448         spin_lock(&trans->transaction->dirty_bgs_lock);
8449         if (!list_empty(&block_group->io_list)) {
8450                 list_del_init(&block_group->io_list);
8451
8452                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
8453
8454                 spin_unlock(&trans->transaction->dirty_bgs_lock);
8455                 btrfs_wait_cache_io(trans, block_group, path);
8456                 btrfs_put_block_group(block_group);
8457                 spin_lock(&trans->transaction->dirty_bgs_lock);
8458         }
8459
8460         if (!list_empty(&block_group->dirty_list)) {
8461                 list_del_init(&block_group->dirty_list);
8462                 remove_rsv = true;
8463                 btrfs_put_block_group(block_group);
8464         }
8465         spin_unlock(&trans->transaction->dirty_bgs_lock);
8466         mutex_unlock(&trans->transaction->cache_write_mutex);
8467
8468         if (!IS_ERR(inode)) {
8469                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
8470                 if (ret) {
8471                         btrfs_add_delayed_iput(inode);
8472                         goto out;
8473                 }
8474                 clear_nlink(inode);
8475                 /* One for the block groups ref */
8476                 spin_lock(&block_group->lock);
8477                 if (block_group->iref) {
8478                         block_group->iref = 0;
8479                         block_group->inode = NULL;
8480                         spin_unlock(&block_group->lock);
8481                         iput(inode);
8482                 } else {
8483                         spin_unlock(&block_group->lock);
8484                 }
8485                 /* One for our lookup ref */
8486                 btrfs_add_delayed_iput(inode);
8487         }
8488
8489         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8490         key.offset = block_group->key.objectid;
8491         key.type = 0;
8492
8493         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8494         if (ret < 0)
8495                 goto out;
8496         if (ret > 0)
8497                 btrfs_release_path(path);
8498         if (ret == 0) {
8499                 ret = btrfs_del_item(trans, tree_root, path);
8500                 if (ret)
8501                         goto out;
8502                 btrfs_release_path(path);
8503         }
8504
8505         spin_lock(&fs_info->block_group_cache_lock);
8506         rb_erase(&block_group->cache_node,
8507                  &fs_info->block_group_cache_tree);
8508         RB_CLEAR_NODE(&block_group->cache_node);
8509
8510         if (fs_info->first_logical_byte == block_group->key.objectid)
8511                 fs_info->first_logical_byte = (u64)-1;
8512         spin_unlock(&fs_info->block_group_cache_lock);
8513
8514         down_write(&block_group->space_info->groups_sem);
8515         /*
8516          * we must use list_del_init so people can check to see if they
8517          * are still on the list after taking the semaphore
8518          */
8519         list_del_init(&block_group->list);
8520         if (list_empty(&block_group->space_info->block_groups[index])) {
8521                 kobj = block_group->space_info->block_group_kobjs[index];
8522                 block_group->space_info->block_group_kobjs[index] = NULL;
8523                 clear_avail_alloc_bits(fs_info, block_group->flags);
8524         }
8525         up_write(&block_group->space_info->groups_sem);
8526         clear_incompat_bg_bits(fs_info, block_group->flags);
8527         if (kobj) {
8528                 kobject_del(kobj);
8529                 kobject_put(kobj);
8530         }
8531
8532         if (block_group->has_caching_ctl)
8533                 caching_ctl = get_caching_control(block_group);
8534         if (block_group->cached == BTRFS_CACHE_STARTED)
8535                 wait_block_group_cache_done(block_group);
8536         if (block_group->has_caching_ctl) {
8537                 down_write(&fs_info->commit_root_sem);
8538                 if (!caching_ctl) {
8539                         struct btrfs_caching_control *ctl;
8540
8541                         list_for_each_entry(ctl,
8542                                     &fs_info->caching_block_groups, list)
8543                                 if (ctl->block_group == block_group) {
8544                                         caching_ctl = ctl;
8545                                         refcount_inc(&caching_ctl->count);
8546                                         break;
8547                                 }
8548                 }
8549                 if (caching_ctl)
8550                         list_del_init(&caching_ctl->list);
8551                 up_write(&fs_info->commit_root_sem);
8552                 if (caching_ctl) {
8553                         /* Once for the caching bgs list and once for us. */
8554                         put_caching_control(caching_ctl);
8555                         put_caching_control(caching_ctl);
8556                 }
8557         }
8558
8559         spin_lock(&trans->transaction->dirty_bgs_lock);
8560         WARN_ON(!list_empty(&block_group->dirty_list));
8561         WARN_ON(!list_empty(&block_group->io_list));
8562         spin_unlock(&trans->transaction->dirty_bgs_lock);
8563
8564         btrfs_remove_free_space_cache(block_group);
8565
8566         spin_lock(&block_group->space_info->lock);
8567         list_del_init(&block_group->ro_list);
8568
8569         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8570                 WARN_ON(block_group->space_info->total_bytes
8571                         < block_group->key.offset);
8572                 WARN_ON(block_group->space_info->bytes_readonly
8573                         < block_group->key.offset);
8574                 WARN_ON(block_group->space_info->disk_total
8575                         < block_group->key.offset * factor);
8576         }
8577         block_group->space_info->total_bytes -= block_group->key.offset;
8578         block_group->space_info->bytes_readonly -= block_group->key.offset;
8579         block_group->space_info->disk_total -= block_group->key.offset * factor;
8580
8581         spin_unlock(&block_group->space_info->lock);
8582
8583         memcpy(&key, &block_group->key, sizeof(key));
8584
8585         mutex_lock(&fs_info->chunk_mutex);
8586         spin_lock(&block_group->lock);
8587         block_group->removed = 1;
8588         /*
8589          * At this point trimming can't start on this block group, because we
8590          * removed the block group from the tree fs_info->block_group_cache_tree
8591          * so no one can't find it anymore and even if someone already got this
8592          * block group before we removed it from the rbtree, they have already
8593          * incremented block_group->trimming - if they didn't, they won't find
8594          * any free space entries because we already removed them all when we
8595          * called btrfs_remove_free_space_cache().
8596          *
8597          * And we must not remove the extent map from the fs_info->mapping_tree
8598          * to prevent the same logical address range and physical device space
8599          * ranges from being reused for a new block group. This is because our
8600          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
8601          * completely transactionless, so while it is trimming a range the
8602          * currently running transaction might finish and a new one start,
8603          * allowing for new block groups to be created that can reuse the same
8604          * physical device locations unless we take this special care.
8605          *
8606          * There may also be an implicit trim operation if the file system
8607          * is mounted with -odiscard. The same protections must remain
8608          * in place until the extents have been discarded completely when
8609          * the transaction commit has completed.
8610          */
8611         remove_em = (atomic_read(&block_group->trimming) == 0);
8612         spin_unlock(&block_group->lock);
8613
8614         mutex_unlock(&fs_info->chunk_mutex);
8615
8616         ret = remove_block_group_free_space(trans, block_group);
8617         if (ret)
8618                 goto out;
8619
8620         btrfs_put_block_group(block_group);
8621         btrfs_put_block_group(block_group);
8622
8623         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8624         if (ret > 0)
8625                 ret = -EIO;
8626         if (ret < 0)
8627                 goto out;
8628
8629         ret = btrfs_del_item(trans, root, path);
8630         if (ret)
8631                 goto out;
8632
8633         if (remove_em) {
8634                 struct extent_map_tree *em_tree;
8635
8636                 em_tree = &fs_info->mapping_tree;
8637                 write_lock(&em_tree->lock);
8638                 remove_extent_mapping(em_tree, em);
8639                 write_unlock(&em_tree->lock);
8640                 /* once for the tree */
8641                 free_extent_map(em);
8642         }
8643 out:
8644         if (remove_rsv)
8645                 btrfs_delayed_refs_rsv_release(fs_info, 1);
8646         btrfs_free_path(path);
8647         return ret;
8648 }
8649
8650 struct btrfs_trans_handle *
8651 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
8652                                      const u64 chunk_offset)
8653 {
8654         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
8655         struct extent_map *em;
8656         struct map_lookup *map;
8657         unsigned int num_items;
8658
8659         read_lock(&em_tree->lock);
8660         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
8661         read_unlock(&em_tree->lock);
8662         ASSERT(em && em->start == chunk_offset);
8663
8664         /*
8665          * We need to reserve 3 + N units from the metadata space info in order
8666          * to remove a block group (done at btrfs_remove_chunk() and at
8667          * btrfs_remove_block_group()), which are used for:
8668          *
8669          * 1 unit for adding the free space inode's orphan (located in the tree
8670          * of tree roots).
8671          * 1 unit for deleting the block group item (located in the extent
8672          * tree).
8673          * 1 unit for deleting the free space item (located in tree of tree
8674          * roots).
8675          * N units for deleting N device extent items corresponding to each
8676          * stripe (located in the device tree).
8677          *
8678          * In order to remove a block group we also need to reserve units in the
8679          * system space info in order to update the chunk tree (update one or
8680          * more device items and remove one chunk item), but this is done at
8681          * btrfs_remove_chunk() through a call to check_system_chunk().
8682          */
8683         map = em->map_lookup;
8684         num_items = 3 + map->num_stripes;
8685         free_extent_map(em);
8686
8687         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
8688                                                            num_items, 1);
8689 }
8690
8691 /*
8692  * Process the unused_bgs list and remove any that don't have any allocated
8693  * space inside of them.
8694  */
8695 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
8696 {
8697         struct btrfs_block_group_cache *block_group;
8698         struct btrfs_space_info *space_info;
8699         struct btrfs_trans_handle *trans;
8700         int ret = 0;
8701
8702         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
8703                 return;
8704
8705         spin_lock(&fs_info->unused_bgs_lock);
8706         while (!list_empty(&fs_info->unused_bgs)) {
8707                 u64 start, end;
8708                 int trimming;
8709
8710                 block_group = list_first_entry(&fs_info->unused_bgs,
8711                                                struct btrfs_block_group_cache,
8712                                                bg_list);
8713                 list_del_init(&block_group->bg_list);
8714
8715                 space_info = block_group->space_info;
8716
8717                 if (ret || btrfs_mixed_space_info(space_info)) {
8718                         btrfs_put_block_group(block_group);
8719                         continue;
8720                 }
8721                 spin_unlock(&fs_info->unused_bgs_lock);
8722
8723                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
8724
8725                 /* Don't want to race with allocators so take the groups_sem */
8726                 down_write(&space_info->groups_sem);
8727                 spin_lock(&block_group->lock);
8728                 if (block_group->reserved || block_group->pinned ||
8729                     btrfs_block_group_used(&block_group->item) ||
8730                     block_group->ro ||
8731                     list_is_singular(&block_group->list)) {
8732                         /*
8733                          * We want to bail if we made new allocations or have
8734                          * outstanding allocations in this block group.  We do
8735                          * the ro check in case balance is currently acting on
8736                          * this block group.
8737                          */
8738                         trace_btrfs_skip_unused_block_group(block_group);
8739                         spin_unlock(&block_group->lock);
8740                         up_write(&space_info->groups_sem);
8741                         goto next;
8742                 }
8743                 spin_unlock(&block_group->lock);
8744
8745                 /* We don't want to force the issue, only flip if it's ok. */
8746                 ret = inc_block_group_ro(block_group, 0);
8747                 up_write(&space_info->groups_sem);
8748                 if (ret < 0) {
8749                         ret = 0;
8750                         goto next;
8751                 }
8752
8753                 /*
8754                  * Want to do this before we do anything else so we can recover
8755                  * properly if we fail to join the transaction.
8756                  */
8757                 trans = btrfs_start_trans_remove_block_group(fs_info,
8758                                                      block_group->key.objectid);
8759                 if (IS_ERR(trans)) {
8760                         btrfs_dec_block_group_ro(block_group);
8761                         ret = PTR_ERR(trans);
8762                         goto next;
8763                 }
8764
8765                 /*
8766                  * We could have pending pinned extents for this block group,
8767                  * just delete them, we don't care about them anymore.
8768                  */
8769                 start = block_group->key.objectid;
8770                 end = start + block_group->key.offset - 1;
8771                 /*
8772                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
8773                  * btrfs_finish_extent_commit(). If we are at transaction N,
8774                  * another task might be running finish_extent_commit() for the
8775                  * previous transaction N - 1, and have seen a range belonging
8776                  * to the block group in freed_extents[] before we were able to
8777                  * clear the whole block group range from freed_extents[]. This
8778                  * means that task can lookup for the block group after we
8779                  * unpinned it from freed_extents[] and removed it, leading to
8780                  * a BUG_ON() at btrfs_unpin_extent_range().
8781                  */
8782                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
8783                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
8784                                   EXTENT_DIRTY);
8785                 if (ret) {
8786                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
8787                         btrfs_dec_block_group_ro(block_group);
8788                         goto end_trans;
8789                 }
8790                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
8791                                   EXTENT_DIRTY);
8792                 if (ret) {
8793                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
8794                         btrfs_dec_block_group_ro(block_group);
8795                         goto end_trans;
8796                 }
8797                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
8798
8799                 /* Reset pinned so btrfs_put_block_group doesn't complain */
8800                 spin_lock(&space_info->lock);
8801                 spin_lock(&block_group->lock);
8802
8803                 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
8804                                                      -block_group->pinned);
8805                 space_info->bytes_readonly += block_group->pinned;
8806                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
8807                                    -block_group->pinned,
8808                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
8809                 block_group->pinned = 0;
8810
8811                 spin_unlock(&block_group->lock);
8812                 spin_unlock(&space_info->lock);
8813
8814                 /* DISCARD can flip during remount */
8815                 trimming = btrfs_test_opt(fs_info, DISCARD);
8816
8817                 /* Implicit trim during transaction commit. */
8818                 if (trimming)
8819                         btrfs_get_block_group_trimming(block_group);
8820
8821                 /*
8822                  * Btrfs_remove_chunk will abort the transaction if things go
8823                  * horribly wrong.
8824                  */
8825                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
8826
8827                 if (ret) {
8828                         if (trimming)
8829                                 btrfs_put_block_group_trimming(block_group);
8830                         goto end_trans;
8831                 }
8832
8833                 /*
8834                  * If we're not mounted with -odiscard, we can just forget
8835                  * about this block group. Otherwise we'll need to wait
8836                  * until transaction commit to do the actual discard.
8837                  */
8838                 if (trimming) {
8839                         spin_lock(&fs_info->unused_bgs_lock);
8840                         /*
8841                          * A concurrent scrub might have added us to the list
8842                          * fs_info->unused_bgs, so use a list_move operation
8843                          * to add the block group to the deleted_bgs list.
8844                          */
8845                         list_move(&block_group->bg_list,
8846                                   &trans->transaction->deleted_bgs);
8847                         spin_unlock(&fs_info->unused_bgs_lock);
8848                         btrfs_get_block_group(block_group);
8849                 }
8850 end_trans:
8851                 btrfs_end_transaction(trans);
8852 next:
8853                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
8854                 btrfs_put_block_group(block_group);
8855                 spin_lock(&fs_info->unused_bgs_lock);
8856         }
8857         spin_unlock(&fs_info->unused_bgs_lock);
8858 }
8859
8860 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
8861                                    u64 start, u64 end)
8862 {
8863         return unpin_extent_range(fs_info, start, end, false);
8864 }
8865
8866 /*
8867  * It used to be that old block groups would be left around forever.
8868  * Iterating over them would be enough to trim unused space.  Since we
8869  * now automatically remove them, we also need to iterate over unallocated
8870  * space.
8871  *
8872  * We don't want a transaction for this since the discard may take a
8873  * substantial amount of time.  We don't require that a transaction be
8874  * running, but we do need to take a running transaction into account
8875  * to ensure that we're not discarding chunks that were released or
8876  * allocated in the current transaction.
8877  *
8878  * Holding the chunks lock will prevent other threads from allocating
8879  * or releasing chunks, but it won't prevent a running transaction
8880  * from committing and releasing the memory that the pending chunks
8881  * list head uses.  For that, we need to take a reference to the
8882  * transaction and hold the commit root sem.  We only need to hold
8883  * it while performing the free space search since we have already
8884  * held back allocations.
8885  */
8886 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
8887 {
8888         u64 start = SZ_1M, len = 0, end = 0;
8889         int ret;
8890
8891         *trimmed = 0;
8892
8893         /* Discard not supported = nothing to do. */
8894         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
8895                 return 0;
8896
8897         /* Not writable = nothing to do. */
8898         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
8899                 return 0;
8900
8901         /* No free space = nothing to do. */
8902         if (device->total_bytes <= device->bytes_used)
8903                 return 0;
8904
8905         ret = 0;
8906
8907         while (1) {
8908                 struct btrfs_fs_info *fs_info = device->fs_info;
8909                 u64 bytes;
8910
8911                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
8912                 if (ret)
8913                         break;
8914
8915                 find_first_clear_extent_bit(&device->alloc_state, start,
8916                                             &start, &end,
8917                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
8918
8919                 /* Ensure we skip the reserved area in the first 1M */
8920                 start = max_t(u64, start, SZ_1M);
8921
8922                 /*
8923                  * If find_first_clear_extent_bit find a range that spans the
8924                  * end of the device it will set end to -1, in this case it's up
8925                  * to the caller to trim the value to the size of the device.
8926                  */
8927                 end = min(end, device->total_bytes - 1);
8928
8929                 len = end - start + 1;
8930
8931                 /* We didn't find any extents */
8932                 if (!len) {
8933                         mutex_unlock(&fs_info->chunk_mutex);
8934                         ret = 0;
8935                         break;
8936                 }
8937
8938                 ret = btrfs_issue_discard(device->bdev, start, len,
8939                                           &bytes);
8940                 if (!ret)
8941                         set_extent_bits(&device->alloc_state, start,
8942                                         start + bytes - 1,
8943                                         CHUNK_TRIMMED);
8944                 mutex_unlock(&fs_info->chunk_mutex);
8945
8946                 if (ret)
8947                         break;
8948
8949                 start += len;
8950                 *trimmed += bytes;
8951
8952                 if (fatal_signal_pending(current)) {
8953                         ret = -ERESTARTSYS;
8954                         break;
8955                 }
8956
8957                 cond_resched();
8958         }
8959
8960         return ret;
8961 }
8962
8963 /*
8964  * Trim the whole filesystem by:
8965  * 1) trimming the free space in each block group
8966  * 2) trimming the unallocated space on each device
8967  *
8968  * This will also continue trimming even if a block group or device encounters
8969  * an error.  The return value will be the last error, or 0 if nothing bad
8970  * happens.
8971  */
8972 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
8973 {
8974         struct btrfs_block_group_cache *cache = NULL;
8975         struct btrfs_device *device;
8976         struct list_head *devices;
8977         u64 group_trimmed;
8978         u64 range_end = U64_MAX;
8979         u64 start;
8980         u64 end;
8981         u64 trimmed = 0;
8982         u64 bg_failed = 0;
8983         u64 dev_failed = 0;
8984         int bg_ret = 0;
8985         int dev_ret = 0;
8986         int ret = 0;
8987
8988         /*
8989          * Check range overflow if range->len is set.
8990          * The default range->len is U64_MAX.
8991          */
8992         if (range->len != U64_MAX &&
8993             check_add_overflow(range->start, range->len, &range_end))
8994                 return -EINVAL;
8995
8996         cache = btrfs_lookup_first_block_group(fs_info, range->start);
8997         for (; cache; cache = next_block_group(cache)) {
8998                 if (cache->key.objectid >= range_end) {
8999                         btrfs_put_block_group(cache);
9000                         break;
9001                 }
9002
9003                 start = max(range->start, cache->key.objectid);
9004                 end = min(range_end, cache->key.objectid + cache->key.offset);
9005
9006                 if (end - start >= range->minlen) {
9007                         if (!block_group_cache_done(cache)) {
9008                                 ret = cache_block_group(cache, 0);
9009                                 if (ret) {
9010                                         bg_failed++;
9011                                         bg_ret = ret;
9012                                         continue;
9013                                 }
9014                                 ret = wait_block_group_cache_done(cache);
9015                                 if (ret) {
9016                                         bg_failed++;
9017                                         bg_ret = ret;
9018                                         continue;
9019                                 }
9020                         }
9021                         ret = btrfs_trim_block_group(cache,
9022                                                      &group_trimmed,
9023                                                      start,
9024                                                      end,
9025                                                      range->minlen);
9026
9027                         trimmed += group_trimmed;
9028                         if (ret) {
9029                                 bg_failed++;
9030                                 bg_ret = ret;
9031                                 continue;
9032                         }
9033                 }
9034         }
9035
9036         if (bg_failed)
9037                 btrfs_warn(fs_info,
9038                         "failed to trim %llu block group(s), last error %d",
9039                         bg_failed, bg_ret);
9040         mutex_lock(&fs_info->fs_devices->device_list_mutex);
9041         devices = &fs_info->fs_devices->devices;
9042         list_for_each_entry(device, devices, dev_list) {
9043                 ret = btrfs_trim_free_extents(device, &group_trimmed);
9044                 if (ret) {
9045                         dev_failed++;
9046                         dev_ret = ret;
9047                         break;
9048                 }
9049
9050                 trimmed += group_trimmed;
9051         }
9052         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
9053
9054         if (dev_failed)
9055                 btrfs_warn(fs_info,
9056                         "failed to trim %llu device(s), last error %d",
9057                         dev_failed, dev_ret);
9058         range->len = trimmed;
9059         if (bg_ret)
9060                 return bg_ret;
9061         return dev_ret;
9062 }
9063
9064 /*
9065  * btrfs_{start,end}_write_no_snapshotting() are similar to
9066  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
9067  * data into the page cache through nocow before the subvolume is snapshoted,
9068  * but flush the data into disk after the snapshot creation, or to prevent
9069  * operations while snapshotting is ongoing and that cause the snapshot to be
9070  * inconsistent (writes followed by expanding truncates for example).
9071  */
9072 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
9073 {
9074         percpu_counter_dec(&root->subv_writers->counter);
9075         cond_wake_up(&root->subv_writers->wait);
9076 }
9077
9078 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
9079 {
9080         if (atomic_read(&root->will_be_snapshotted))
9081                 return 0;
9082
9083         percpu_counter_inc(&root->subv_writers->counter);
9084         /*
9085          * Make sure counter is updated before we check for snapshot creation.
9086          */
9087         smp_mb();
9088         if (atomic_read(&root->will_be_snapshotted)) {
9089                 btrfs_end_write_no_snapshotting(root);
9090                 return 0;
9091         }
9092         return 1;
9093 }
9094
9095 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
9096 {
9097         while (true) {
9098                 int ret;
9099
9100                 ret = btrfs_start_write_no_snapshotting(root);
9101                 if (ret)
9102                         break;
9103                 wait_var_event(&root->will_be_snapshotted,
9104                                !atomic_read(&root->will_be_snapshotted));
9105         }
9106 }
9107
9108 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
9109 {
9110         struct btrfs_fs_info *fs_info = bg->fs_info;
9111
9112         spin_lock(&fs_info->unused_bgs_lock);
9113         if (list_empty(&bg->bg_list)) {
9114                 btrfs_get_block_group(bg);
9115                 trace_btrfs_add_unused_block_group(bg);
9116                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
9117         }
9118         spin_unlock(&fs_info->unused_bgs_lock);
9119 }