fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38
  39 #undef SCRAMBLE_DELAYED_REFS
  40
  41 /*
  42  * control flags for do_chunk_alloc's force field
  43  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  44  * if we really need one.
  45  *
  46  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  47  * if we have very few chunks already allocated.  This is
  48  * used as part of the clustering code to help make sure
  49  * we have a good pool of storage to cluster in, without
  50  * filling the FS with empty chunks
  51  *
  52  * CHUNK_ALLOC_FORCE means it must try to allocate one
  53  *
  54  */
  55 enum {
  56         CHUNK_ALLOC_NO_FORCE = 0,
  57         CHUNK_ALLOC_LIMITED = 1,
  58         CHUNK_ALLOC_FORCE = 2,
  59 };
  60
  61 /*
  62  * Control how reservations are dealt with.
  63  *
  64  * RESERVE_FREE - freeing a reservation.
  65  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  66  *   ENOSPC accounting
  67  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  68  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  69  */
  70 enum {
  71         RESERVE_FREE = 0,
  72         RESERVE_ALLOC = 1,
  73         RESERVE_ALLOC_NO_ACCOUNT = 2,
  74 };
  75
  76 static int update_block_group(struct btrfs_root *root,
  77                               u64 bytenr, u64 num_bytes, int alloc);
  78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  79                                 struct btrfs_root *root,
  80                                 u64 bytenr, u64 num_bytes, u64 parent,
  81                                 u64 root_objectid, u64 owner_objectid,
  82                                 u64 owner_offset, int refs_to_drop,
  83                                 struct btrfs_delayed_extent_op *extra_op);
  84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  85                                     struct extent_buffer *leaf,
  86                                     struct btrfs_extent_item *ei);
  87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  88                                       struct btrfs_root *root,
  89                                       u64 parent, u64 root_objectid,
  90                                       u64 flags, u64 owner, u64 offset,
  91                                       struct btrfs_key *ins, int ref_mod);
  92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  93                                      struct btrfs_root *root,
  94                                      u64 parent, u64 root_objectid,
  95                                      u64 flags, struct btrfs_disk_key *key,
  96                                      int level, struct btrfs_key *ins);
  97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  98                           struct btrfs_root *extent_root, u64 flags,
  99                           int force);
 100 static int find_next_key(struct btrfs_path *path, int level,
 101                          struct btrfs_key *key);
 102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 103                             int dump_block_groups);
 104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 105                                        u64 num_bytes, int reserve);
 106 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 107                                u64 num_bytes);
 108
 109 static noinline int
 110 block_group_cache_done(struct btrfs_block_group_cache *cache)
 111 {
 112         smp_mb();
 113         return cache->cached == BTRFS_CACHE_FINISHED;
 114 }
 115
 116 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 117 {
 118         return (cache->flags & bits) == bits;
 119 }
 120
 121 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         atomic_inc(&cache->count);
 124 }
 125
 126 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 127 {
 128         if (atomic_dec_and_test(&cache->count)) {
 129                 WARN_ON(cache->pinned > 0);
 130                 WARN_ON(cache->reserved > 0);
 131                 kfree(cache->free_space_ctl);
 132                 kfree(cache);
 133         }
 134 }
 135
 136 /*
 137  * this adds the block group to the fs_info rb tree for the block group
 138  * cache
 139  */
 140 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 141                                 struct btrfs_block_group_cache *block_group)
 142 {
 143         struct rb_node **p;
 144         struct rb_node *parent = NULL;
 145         struct btrfs_block_group_cache *cache;
 146
 147         spin_lock(&info->block_group_cache_lock);
 148         p = &info->block_group_cache_tree.rb_node;
 149
 150         while (*p) {
 151                 parent = *p;
 152                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 153                                  cache_node);
 154                 if (block_group->key.objectid < cache->key.objectid) {
 155                         p = &(*p)->rb_left;
 156                 } else if (block_group->key.objectid > cache->key.objectid) {
 157                         p = &(*p)->rb_right;
 158                 } else {
 159                         spin_unlock(&info->block_group_cache_lock);
 160                         return -EEXIST;
 161                 }
 162         }
 163
 164         rb_link_node(&block_group->cache_node, parent, p);
 165         rb_insert_color(&block_group->cache_node,
 166                         &info->block_group_cache_tree);
 167
 168         if (info->first_logical_byte > block_group->key.objectid)
 169                 info->first_logical_byte = block_group->key.objectid;
 170
 171         spin_unlock(&info->block_group_cache_lock);
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * This will return the block group at or after bytenr if contains is 0, else
 178  * it will return the block group that contains the bytenr
 179  */
 180 static struct btrfs_block_group_cache *
 181 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 182                               int contains)
 183 {
 184         struct btrfs_block_group_cache *cache, *ret = NULL;
 185         struct rb_node *n;
 186         u64 end, start;
 187
 188         spin_lock(&info->block_group_cache_lock);
 189         n = info->block_group_cache_tree.rb_node;
 190
 191         while (n) {
 192                 cache = rb_entry(n, struct btrfs_block_group_cache,
 193                                  cache_node);
 194                 end = cache->key.objectid + cache->key.offset - 1;
 195                 start = cache->key.objectid;
 196
 197                 if (bytenr < start) {
 198                         if (!contains && (!ret || start < ret->key.objectid))
 199                                 ret = cache;
 200                         n = n->rb_left;
 201                 } else if (bytenr > start) {
 202                         if (contains && bytenr <= end) {
 203                                 ret = cache;
 204                                 break;
 205                         }
 206                         n = n->rb_right;
 207                 } else {
 208                         ret = cache;
 209                         break;
 210                 }
 211         }
 212         if (ret) {
 213                 btrfs_get_block_group(ret);
 214                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 215                         info->first_logical_byte = ret->key.objectid;
 216         }
 217         spin_unlock(&info->block_group_cache_lock);
 218
 219         return ret;
 220 }
 221
 222 static int add_excluded_extent(struct btrfs_root *root,
 223                                u64 start, u64 num_bytes)
 224 {
 225         u64 end = start + num_bytes - 1;
 226         set_extent_bits(&root->fs_info->freed_extents[0],
 227                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 228         set_extent_bits(&root->fs_info->freed_extents[1],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         return 0;
 231 }
 232
 233 static void free_excluded_extents(struct btrfs_root *root,
 234                                   struct btrfs_block_group_cache *cache)
 235 {
 236         u64 start, end;
 237
 238         start = cache->key.objectid;
 239         end = start + cache->key.offset - 1;
 240
 241         clear_extent_bits(&root->fs_info->freed_extents[0],
 242                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 243         clear_extent_bits(&root->fs_info->freed_extents[1],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245 }
 246
 247 static int exclude_super_stripes(struct btrfs_root *root,
 248                                  struct btrfs_block_group_cache *cache)
 249 {
 250         u64 bytenr;
 251         u64 *logical;
 252         int stripe_len;
 253         int i, nr, ret;
 254
 255         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 256                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 257                 cache->bytes_super += stripe_len;
 258                 ret = add_excluded_extent(root, cache->key.objectid,
 259                                           stripe_len);
 260                 BUG_ON(ret); /* -ENOMEM */
 261         }
 262
 263         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 264                 bytenr = btrfs_sb_offset(i);
 265                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 266                                        cache->key.objectid, bytenr,
 267                                        0, &logical, &nr, &stripe_len);
 268                 BUG_ON(ret); /* -ENOMEM */
 269
 270                 while (nr--) {
 271                         cache->bytes_super += stripe_len;
 272                         ret = add_excluded_extent(root, logical[nr],
 273                                                   stripe_len);
 274                         BUG_ON(ret); /* -ENOMEM */
 275                 }
 276
 277                 kfree(logical);
 278         }
 279         return 0;
 280 }
 281
 282 static struct btrfs_caching_control *
 283 get_caching_control(struct btrfs_block_group_cache *cache)
 284 {
 285         struct btrfs_caching_control *ctl;
 286
 287         spin_lock(&cache->lock);
 288         if (cache->cached != BTRFS_CACHE_STARTED) {
 289                 spin_unlock(&cache->lock);
 290                 return NULL;
 291         }
 292
 293         /* We're loading it the fast way, so we don't have a caching_ctl. */
 294         if (!cache->caching_ctl) {
 295                 spin_unlock(&cache->lock);
 296                 return NULL;
 297         }
 298
 299         ctl = cache->caching_ctl;
 300         atomic_inc(&ctl->count);
 301         spin_unlock(&cache->lock);
 302         return ctl;
 303 }
 304
 305 static void put_caching_control(struct btrfs_caching_control *ctl)
 306 {
 307         if (atomic_dec_and_test(&ctl->count))
 308                 kfree(ctl);
 309 }
 310
 311 /*
 312  * this is only called by cache_block_group, since we could have freed extents
 313  * we need to check the pinned_extents for any extents that can't be used yet
 314  * since their free space will be released as soon as the transaction commits.
 315  */
 316 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 317                               struct btrfs_fs_info *info, u64 start, u64 end)
 318 {
 319         u64 extent_start, extent_end, size, total_added = 0;
 320         int ret;
 321
 322         while (start < end) {
 323                 ret = find_first_extent_bit(info->pinned_extents, start,
 324                                             &extent_start, &extent_end,
 325                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 326                                             NULL);
 327                 if (ret)
 328                         break;
 329
 330                 if (extent_start <= start) {
 331                         start = extent_end + 1;
 332                 } else if (extent_start > start && extent_start < end) {
 333                         size = extent_start - start;
 334                         total_added += size;
 335                         ret = btrfs_add_free_space(block_group, start,
 336                                                    size);
 337                         BUG_ON(ret); /* -ENOMEM or logic error */
 338                         start = extent_end + 1;
 339                 } else {
 340                         break;
 341                 }
 342         }
 343
 344         if (start < end) {
 345                 size = end - start;
 346                 total_added += size;
 347                 ret = btrfs_add_free_space(block_group, start, size);
 348                 BUG_ON(ret); /* -ENOMEM or logic error */
 349         }
 350
 351         return total_added;
 352 }
 353
 354 static noinline void caching_thread(struct btrfs_work *work)
 355 {
 356         struct btrfs_block_group_cache *block_group;
 357         struct btrfs_fs_info *fs_info;
 358         struct btrfs_caching_control *caching_ctl;
 359         struct btrfs_root *extent_root;
 360         struct btrfs_path *path;
 361         struct extent_buffer *leaf;
 362         struct btrfs_key key;
 363         u64 total_found = 0;
 364         u64 last = 0;
 365         u32 nritems;
 366         int ret = 0;
 367
 368         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 369         block_group = caching_ctl->block_group;
 370         fs_info = block_group->fs_info;
 371         extent_root = fs_info->extent_root;
 372
 373         path = btrfs_alloc_path();
 374         if (!path)
 375                 goto out;
 376
 377         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 378
 379         /*
 380          * We don't want to deadlock with somebody trying to allocate a new
 381          * extent for the extent root while also trying to search the extent
 382          * root to add free space.  So we skip locking and search the commit
 383          * root, since its read-only
 384          */
 385         path->skip_locking = 1;
 386         path->search_commit_root = 1;
 387         path->reada = 1;
 388
 389         key.objectid = last;
 390         key.offset = 0;
 391         key.type = BTRFS_EXTENT_ITEM_KEY;
 392 again:
 393         mutex_lock(&caching_ctl->mutex);
 394         /* need to make sure the commit_root doesn't disappear */
 395         down_read(&fs_info->extent_commit_sem);
 396
 397         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 398         if (ret < 0)
 399                 goto err;
 400
 401         leaf = path->nodes[0];
 402         nritems = btrfs_header_nritems(leaf);
 403
 404         while (1) {
 405                 if (btrfs_fs_closing(fs_info) > 1) {
 406                         last = (u64)-1;
 407                         break;
 408                 }
 409
 410                 if (path->slots[0] < nritems) {
 411                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 412                 } else {
 413                         ret = find_next_key(path, 0, &key);
 414                         if (ret)
 415                                 break;
 416
 417                         if (need_resched() ||
 418                             btrfs_next_leaf(extent_root, path)) {
 419                                 caching_ctl->progress = last;
 420                                 btrfs_release_path(path);
 421                                 up_read(&fs_info->extent_commit_sem);
 422                                 mutex_unlock(&caching_ctl->mutex);
 423                                 cond_resched();
 424                                 goto again;
 425                         }
 426                         leaf = path->nodes[0];
 427                         nritems = btrfs_header_nritems(leaf);
 428                         continue;
 429                 }
 430
 431                 if (key.objectid < block_group->key.objectid) {
 432                         path->slots[0]++;
 433                         continue;
 434                 }
 435
 436                 if (key.objectid >= block_group->key.objectid +
 437                     block_group->key.offset)
 438                         break;
 439
 440                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 441                         total_found += add_new_free_space(block_group,
 442                                                           fs_info, last,
 443                                                           key.objectid);
 444                         last = key.objectid + key.offset;
 445
 446                         if (total_found > (1024 * 1024 * 2)) {
 447                                 total_found = 0;
 448                                 wake_up(&caching_ctl->wait);
 449                         }
 450                 }
 451                 path->slots[0]++;
 452         }
 453         ret = 0;
 454
 455         total_found += add_new_free_space(block_group, fs_info, last,
 456                                           block_group->key.objectid +
 457                                           block_group->key.offset);
 458         caching_ctl->progress = (u64)-1;
 459
 460         spin_lock(&block_group->lock);
 461         block_group->caching_ctl = NULL;
 462         block_group->cached = BTRFS_CACHE_FINISHED;
 463         spin_unlock(&block_group->lock);
 464
 465 err:
 466         btrfs_free_path(path);
 467         up_read(&fs_info->extent_commit_sem);
 468
 469         free_excluded_extents(extent_root, block_group);
 470
 471         mutex_unlock(&caching_ctl->mutex);
 472 out:
 473         wake_up(&caching_ctl->wait);
 474
 475         put_caching_control(caching_ctl);
 476         btrfs_put_block_group(block_group);
 477 }
 478
 479 static int cache_block_group(struct btrfs_block_group_cache *cache,
 480                              int load_cache_only)
 481 {
 482         DEFINE_WAIT(wait);
 483         struct btrfs_fs_info *fs_info = cache->fs_info;
 484         struct btrfs_caching_control *caching_ctl;
 485         int ret = 0;
 486
 487         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 488         if (!caching_ctl)
 489                 return -ENOMEM;
 490
 491         INIT_LIST_HEAD(&caching_ctl->list);
 492         mutex_init(&caching_ctl->mutex);
 493         init_waitqueue_head(&caching_ctl->wait);
 494         caching_ctl->block_group = cache;
 495         caching_ctl->progress = cache->key.objectid;
 496         atomic_set(&caching_ctl->count, 1);
 497         caching_ctl->work.func = caching_thread;
 498
 499         spin_lock(&cache->lock);
 500         /*
 501          * This should be a rare occasion, but this could happen I think in the
 502          * case where one thread starts to load the space cache info, and then
 503          * some other thread starts a transaction commit which tries to do an
 504          * allocation while the other thread is still loading the space cache
 505          * info.  The previous loop should have kept us from choosing this block
 506          * group, but if we've moved to the state where we will wait on caching
 507          * block groups we need to first check if we're doing a fast load here,
 508          * so we can wait for it to finish, otherwise we could end up allocating
 509          * from a block group who's cache gets evicted for one reason or
 510          * another.
 511          */
 512         while (cache->cached == BTRFS_CACHE_FAST) {
 513                 struct btrfs_caching_control *ctl;
 514
 515                 ctl = cache->caching_ctl;
 516                 atomic_inc(&ctl->count);
 517                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 518                 spin_unlock(&cache->lock);
 519
 520                 schedule();
 521
 522                 finish_wait(&ctl->wait, &wait);
 523                 put_caching_control(ctl);
 524                 spin_lock(&cache->lock);
 525         }
 526
 527         if (cache->cached != BTRFS_CACHE_NO) {
 528                 spin_unlock(&cache->lock);
 529                 kfree(caching_ctl);
 530                 return 0;
 531         }
 532         WARN_ON(cache->caching_ctl);
 533         cache->caching_ctl = caching_ctl;
 534         cache->cached = BTRFS_CACHE_FAST;
 535         spin_unlock(&cache->lock);
 536
 537         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 538                 ret = load_free_space_cache(fs_info, cache);
 539
 540                 spin_lock(&cache->lock);
 541                 if (ret == 1) {
 542                         cache->caching_ctl = NULL;
 543                         cache->cached = BTRFS_CACHE_FINISHED;
 544                         cache->last_byte_to_unpin = (u64)-1;
 545                 } else {
 546                         if (load_cache_only) {
 547                                 cache->caching_ctl = NULL;
 548                                 cache->cached = BTRFS_CACHE_NO;
 549                         } else {
 550                                 cache->cached = BTRFS_CACHE_STARTED;
 551                         }
 552                 }
 553                 spin_unlock(&cache->lock);
 554                 wake_up(&caching_ctl->wait);
 555                 if (ret == 1) {
 556                         put_caching_control(caching_ctl);
 557                         free_excluded_extents(fs_info->extent_root, cache);
 558                         return 0;
 559                 }
 560         } else {
 561                 /*
 562                  * We are not going to do the fast caching, set cached to the
 563                  * appropriate value and wakeup any waiters.
 564                  */
 565                 spin_lock(&cache->lock);
 566                 if (load_cache_only) {
 567                         cache->caching_ctl = NULL;
 568                         cache->cached = BTRFS_CACHE_NO;
 569                 } else {
 570                         cache->cached = BTRFS_CACHE_STARTED;
 571                 }
 572                 spin_unlock(&cache->lock);
 573                 wake_up(&caching_ctl->wait);
 574         }
 575
 576         if (load_cache_only) {
 577                 put_caching_control(caching_ctl);
 578                 return 0;
 579         }
 580
 581         down_write(&fs_info->extent_commit_sem);
 582         atomic_inc(&caching_ctl->count);
 583         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 584         up_write(&fs_info->extent_commit_sem);
 585
 586         btrfs_get_block_group(cache);
 587
 588         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 589
 590         return ret;
 591 }
 592
 593 /*
 594  * return the block group that starts at or after bytenr
 595  */
 596 static struct btrfs_block_group_cache *
 597 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 598 {
 599         struct btrfs_block_group_cache *cache;
 600
 601         cache = block_group_cache_tree_search(info, bytenr, 0);
 602
 603         return cache;
 604 }
 605
 606 /*
 607  * return the block group that contains the given bytenr
 608  */
 609 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 610                                                  struct btrfs_fs_info *info,
 611                                                  u64 bytenr)
 612 {
 613         struct btrfs_block_group_cache *cache;
 614
 615         cache = block_group_cache_tree_search(info, bytenr, 1);
 616
 617         return cache;
 618 }
 619
 620 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 621                                                   u64 flags)
 622 {
 623         struct list_head *head = &info->space_info;
 624         struct btrfs_space_info *found;
 625
 626         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 627
 628         rcu_read_lock();
 629         list_for_each_entry_rcu(found, head, list) {
 630                 if (found->flags & flags) {
 631                         rcu_read_unlock();
 632                         return found;
 633                 }
 634         }
 635         rcu_read_unlock();
 636         return NULL;
 637 }
 638
 639 /*
 640  * after adding space to the filesystem, we need to clear the full flags
 641  * on all the space infos.
 642  */
 643 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 644 {
 645         struct list_head *head = &info->space_info;
 646         struct btrfs_space_info *found;
 647
 648         rcu_read_lock();
 649         list_for_each_entry_rcu(found, head, list)
 650                 found->full = 0;
 651         rcu_read_unlock();
 652 }
 653
 654 u64 btrfs_find_block_group(struct btrfs_root *root,
 655                            u64 search_start, u64 search_hint, int owner)
 656 {
 657         struct btrfs_block_group_cache *cache;
 658         u64 used;
 659         u64 last = max(search_hint, search_start);
 660         u64 group_start = 0;
 661         int full_search = 0;
 662         int factor = 9;
 663         int wrapped = 0;
 664 again:
 665         while (1) {
 666                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 667                 if (!cache)
 668                         break;
 669
 670                 spin_lock(&cache->lock);
 671                 last = cache->key.objectid + cache->key.offset;
 672                 used = btrfs_block_group_used(&cache->item);
 673
 674                 if ((full_search || !cache->ro) &&
 675                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 676                         if (used + cache->pinned + cache->reserved <
 677                             div_factor(cache->key.offset, factor)) {
 678                                 group_start = cache->key.objectid;
 679                                 spin_unlock(&cache->lock);
 680                                 btrfs_put_block_group(cache);
 681                                 goto found;
 682                         }
 683                 }
 684                 spin_unlock(&cache->lock);
 685                 btrfs_put_block_group(cache);
 686                 cond_resched();
 687         }
 688         if (!wrapped) {
 689                 last = search_start;
 690                 wrapped = 1;
 691                 goto again;
 692         }
 693         if (!full_search && factor < 10) {
 694                 last = search_start;
 695                 full_search = 1;
 696                 factor = 10;
 697                 goto again;
 698         }
 699 found:
 700         return group_start;
 701 }
 702
 703 /* simple helper to search for an existing extent at a given offset */
 704 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 705 {
 706         int ret;
 707         struct btrfs_key key;
 708         struct btrfs_path *path;
 709
 710         path = btrfs_alloc_path();
 711         if (!path)
 712                 return -ENOMEM;
 713
 714         key.objectid = start;
 715         key.offset = len;
 716         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 717         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 718                                 0, 0);
 719         btrfs_free_path(path);
 720         return ret;
 721 }
 722
 723 /*
 724  * helper function to lookup reference count and flags of extent.
 725  *
 726  * the head node for delayed ref is used to store the sum of all the
 727  * reference count modifications queued up in the rbtree. the head
 728  * node may also store the extent flags to set. This way you can check
 729  * to see what the reference count and extent flags would be if all of
 730  * the delayed refs are not processed.
 731  */
 732 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 733                              struct btrfs_root *root, u64 bytenr,
 734                              u64 num_bytes, u64 *refs, u64 *flags)
 735 {
 736         struct btrfs_delayed_ref_head *head;
 737         struct btrfs_delayed_ref_root *delayed_refs;
 738         struct btrfs_path *path;
 739         struct btrfs_extent_item *ei;
 740         struct extent_buffer *leaf;
 741         struct btrfs_key key;
 742         u32 item_size;
 743         u64 num_refs;
 744         u64 extent_flags;
 745         int ret;
 746
 747         path = btrfs_alloc_path();
 748         if (!path)
 749                 return -ENOMEM;
 750
 751         key.objectid = bytenr;
 752         key.type = BTRFS_EXTENT_ITEM_KEY;
 753         key.offset = num_bytes;
 754         if (!trans) {
 755                 path->skip_locking = 1;
 756                 path->search_commit_root = 1;
 757         }
 758 again:
 759         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 760                                 &key, path, 0, 0);
 761         if (ret < 0)
 762                 goto out_free;
 763
 764         if (ret == 0) {
 765                 leaf = path->nodes[0];
 766                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 767                 if (item_size >= sizeof(*ei)) {
 768                         ei = btrfs_item_ptr(leaf, path->slots[0],
 769                                             struct btrfs_extent_item);
 770                         num_refs = btrfs_extent_refs(leaf, ei);
 771                         extent_flags = btrfs_extent_flags(leaf, ei);
 772                 } else {
 773 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 774                         struct btrfs_extent_item_v0 *ei0;
 775                         BUG_ON(item_size != sizeof(*ei0));
 776                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 777                                              struct btrfs_extent_item_v0);
 778                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 779                         /* FIXME: this isn't correct for data */
 780                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 781 #else
 782                         BUG();
 783 #endif
 784                 }
 785                 BUG_ON(num_refs == 0);
 786         } else {
 787                 num_refs = 0;
 788                 extent_flags = 0;
 789                 ret = 0;
 790         }
 791
 792         if (!trans)
 793                 goto out;
 794
 795         delayed_refs = &trans->transaction->delayed_refs;
 796         spin_lock(&delayed_refs->lock);
 797         head = btrfs_find_delayed_ref_head(trans, bytenr);
 798         if (head) {
 799                 if (!mutex_trylock(&head->mutex)) {
 800                         atomic_inc(&head->node.refs);
 801                         spin_unlock(&delayed_refs->lock);
 802
 803                         btrfs_release_path(path);
 804
 805                         /*
 806                          * Mutex was contended, block until it's released and try
 807                          * again
 808                          */
 809                         mutex_lock(&head->mutex);
 810                         mutex_unlock(&head->mutex);
 811                         btrfs_put_delayed_ref(&head->node);
 812                         goto again;
 813                 }
 814                 if (head->extent_op && head->extent_op->update_flags)
 815                         extent_flags |= head->extent_op->flags_to_set;
 816                 else
 817                         BUG_ON(num_refs == 0);
 818
 819                 num_refs += head->node.ref_mod;
 820                 mutex_unlock(&head->mutex);
 821         }
 822         spin_unlock(&delayed_refs->lock);
 823 out:
 824         WARN_ON(num_refs == 0);
 825         if (refs)
 826                 *refs = num_refs;
 827         if (flags)
 828                 *flags = extent_flags;
 829 out_free:
 830         btrfs_free_path(path);
 831         return ret;
 832 }
 833
 834 /*
 835  * Back reference rules.  Back refs have three main goals:
 836  *
 837  * 1) differentiate between all holders of references to an extent so that
 838  *    when a reference is dropped we can make sure it was a valid reference
 839  *    before freeing the extent.
 840  *
 841  * 2) Provide enough information to quickly find the holders of an extent
 842  *    if we notice a given block is corrupted or bad.
 843  *
 844  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 845  *    maintenance.  This is actually the same as #2, but with a slightly
 846  *    different use case.
 847  *
 848  * There are two kinds of back refs. The implicit back refs is optimized
 849  * for pointers in non-shared tree blocks. For a given pointer in a block,
 850  * back refs of this kind provide information about the block's owner tree
 851  * and the pointer's key. These information allow us to find the block by
 852  * b-tree searching. The full back refs is for pointers in tree blocks not
 853  * referenced by their owner trees. The location of tree block is recorded
 854  * in the back refs. Actually the full back refs is generic, and can be
 855  * used in all cases the implicit back refs is used. The major shortcoming
 856  * of the full back refs is its overhead. Every time a tree block gets
 857  * COWed, we have to update back refs entry for all pointers in it.
 858  *
 859  * For a newly allocated tree block, we use implicit back refs for
 860  * pointers in it. This means most tree related operations only involve
 861  * implicit back refs. For a tree block created in old transaction, the
 862  * only way to drop a reference to it is COW it. So we can detect the
 863  * event that tree block loses its owner tree's reference and do the
 864  * back refs conversion.
 865  *
 866  * When a tree block is COW'd through a tree, there are four cases:
 867  *
 868  * The reference count of the block is one and the tree is the block's
 869  * owner tree. Nothing to do in this case.
 870  *
 871  * The reference count of the block is one and the tree is not the
 872  * block's owner tree. In this case, full back refs is used for pointers
 873  * in the block. Remove these full back refs, add implicit back refs for
 874  * every pointers in the new block.
 875  *
 876  * The reference count of the block is greater than one and the tree is
 877  * the block's owner tree. In this case, implicit back refs is used for
 878  * pointers in the block. Add full back refs for every pointers in the
 879  * block, increase lower level extents' reference counts. The original
 880  * implicit back refs are entailed to the new block.
 881  *
 882  * The reference count of the block is greater than one and the tree is
 883  * not the block's owner tree. Add implicit back refs for every pointer in
 884  * the new block, increase lower level extents' reference count.
 885  *
 886  * Back Reference Key composing:
 887  *
 888  * The key objectid corresponds to the first byte in the extent,
 889  * The key type is used to differentiate between types of back refs.
 890  * There are different meanings of the key offset for different types
 891  * of back refs.
 892  *
 893  * File extents can be referenced by:
 894  *
 895  * - multiple snapshots, subvolumes, or different generations in one subvol
 896  * - different files inside a single subvolume
 897  * - different offsets inside a file (bookend extents in file.c)
 898  *
 899  * The extent ref structure for the implicit back refs has fields for:
 900  *
 901  * - Objectid of the subvolume root
 902  * - objectid of the file holding the reference
 903  * - original offset in the file
 904  * - how many bookend extents
 905  *
 906  * The key offset for the implicit back refs is hash of the first
 907  * three fields.
 908  *
 909  * The extent ref structure for the full back refs has field for:
 910  *
 911  * - number of pointers in the tree leaf
 912  *
 913  * The key offset for the implicit back refs is the first byte of
 914  * the tree leaf
 915  *
 916  * When a file extent is allocated, The implicit back refs is used.
 917  * the fields are filled in:
 918  *
 919  *     (root_key.objectid, inode objectid, offset in file, 1)
 920  *
 921  * When a file extent is removed file truncation, we find the
 922  * corresponding implicit back refs and check the following fields:
 923  *
 924  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 925  *
 926  * Btree extents can be referenced by:
 927  *
 928  * - Different subvolumes
 929  *
 930  * Both the implicit back refs and the full back refs for tree blocks
 931  * only consist of key. The key offset for the implicit back refs is
 932  * objectid of block's owner tree. The key offset for the full back refs
 933  * is the first byte of parent block.
 934  *
 935  * When implicit back refs is used, information about the lowest key and
 936  * level of the tree block are required. These information are stored in
 937  * tree block info structure.
 938  */
 939
 940 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 941 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 942                                   struct btrfs_root *root,
 943                                   struct btrfs_path *path,
 944                                   u64 owner, u32 extra_size)
 945 {
 946         struct btrfs_extent_item *item;
 947         struct btrfs_extent_item_v0 *ei0;
 948         struct btrfs_extent_ref_v0 *ref0;
 949         struct btrfs_tree_block_info *bi;
 950         struct extent_buffer *leaf;
 951         struct btrfs_key key;
 952         struct btrfs_key found_key;
 953         u32 new_size = sizeof(*item);
 954         u64 refs;
 955         int ret;
 956
 957         leaf = path->nodes[0];
 958         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 959
 960         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 961         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 962                              struct btrfs_extent_item_v0);
 963         refs = btrfs_extent_refs_v0(leaf, ei0);
 964
 965         if (owner == (u64)-1) {
 966                 while (1) {
 967                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 968                                 ret = btrfs_next_leaf(root, path);
 969                                 if (ret < 0)
 970                                         return ret;
 971                                 BUG_ON(ret > 0); /* Corruption */
 972                                 leaf = path->nodes[0];
 973                         }
 974                         btrfs_item_key_to_cpu(leaf, &found_key,
 975                                               path->slots[0]);
 976                         BUG_ON(key.objectid != found_key.objectid);
 977                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 978                                 path->slots[0]++;
 979                                 continue;
 980                         }
 981                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 982                                               struct btrfs_extent_ref_v0);
 983                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 984                         break;
 985                 }
 986         }
 987         btrfs_release_path(path);
 988
 989         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 990                 new_size += sizeof(*bi);
 991
 992         new_size -= sizeof(*ei0);
 993         ret = btrfs_search_slot(trans, root, &key, path,
 994                                 new_size + extra_size, 1);
 995         if (ret < 0)
 996                 return ret;
 997         BUG_ON(ret); /* Corruption */
 998
 999         btrfs_extend_item(trans, root, path, new_size);
1000
1001         leaf = path->nodes[0];
1002         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1003         btrfs_set_extent_refs(leaf, item, refs);
1004         /* FIXME: get real generation */
1005         btrfs_set_extent_generation(leaf, item, 0);
1006         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1007                 btrfs_set_extent_flags(leaf, item,
1008                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1009                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1010                 bi = (struct btrfs_tree_block_info *)(item + 1);
1011                 /* FIXME: get first key of the block */
1012                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1013                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1014         } else {
1015                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1016         }
1017         btrfs_mark_buffer_dirty(leaf);
1018         return 0;
1019 }
1020 #endif
1021
1022 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1023 {
1024         u32 high_crc = ~(u32)0;
1025         u32 low_crc = ~(u32)0;
1026         __le64 lenum;
1027
1028         lenum = cpu_to_le64(root_objectid);
1029         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1030         lenum = cpu_to_le64(owner);
1031         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1032         lenum = cpu_to_le64(offset);
1033         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1034
1035         return ((u64)high_crc << 31) ^ (u64)low_crc;
1036 }
1037
1038 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1039                                      struct btrfs_extent_data_ref *ref)
1040 {
1041         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1042                                     btrfs_extent_data_ref_objectid(leaf, ref),
1043                                     btrfs_extent_data_ref_offset(leaf, ref));
1044 }
1045
1046 static int match_extent_data_ref(struct extent_buffer *leaf,
1047                                  struct btrfs_extent_data_ref *ref,
1048                                  u64 root_objectid, u64 owner, u64 offset)
1049 {
1050         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1051             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1052             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1053                 return 0;
1054         return 1;
1055 }
1056
1057 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1058                                            struct btrfs_root *root,
1059                                            struct btrfs_path *path,
1060                                            u64 bytenr, u64 parent,
1061                                            u64 root_objectid,
1062                                            u64 owner, u64 offset)
1063 {
1064         struct btrfs_key key;
1065         struct btrfs_extent_data_ref *ref;
1066         struct extent_buffer *leaf;
1067         u32 nritems;
1068         int ret;
1069         int recow;
1070         int err = -ENOENT;
1071
1072         key.objectid = bytenr;
1073         if (parent) {
1074                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1075                 key.offset = parent;
1076         } else {
1077                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1078                 key.offset = hash_extent_data_ref(root_objectid,
1079                                                   owner, offset);
1080         }
1081 again:
1082         recow = 0;
1083         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1084         if (ret < 0) {
1085                 err = ret;
1086                 goto fail;
1087         }
1088
1089         if (parent) {
1090                 if (!ret)
1091                         return 0;
1092 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1093                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1094                 btrfs_release_path(path);
1095                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1096                 if (ret < 0) {
1097                         err = ret;
1098                         goto fail;
1099                 }
1100                 if (!ret)
1101                         return 0;
1102 #endif
1103                 goto fail;
1104         }
1105
1106         leaf = path->nodes[0];
1107         nritems = btrfs_header_nritems(leaf);
1108         while (1) {
1109                 if (path->slots[0] >= nritems) {
1110                         ret = btrfs_next_leaf(root, path);
1111                         if (ret < 0)
1112                                 err = ret;
1113                         if (ret)
1114                                 goto fail;
1115
1116                         leaf = path->nodes[0];
1117                         nritems = btrfs_header_nritems(leaf);
1118                         recow = 1;
1119                 }
1120
1121                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1122                 if (key.objectid != bytenr ||
1123                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1124                         goto fail;
1125
1126                 ref = btrfs_item_ptr(leaf, path->slots[0],
1127                                      struct btrfs_extent_data_ref);
1128
1129                 if (match_extent_data_ref(leaf, ref, root_objectid,
1130                                           owner, offset)) {
1131                         if (recow) {
1132                                 btrfs_release_path(path);
1133                                 goto again;
1134                         }
1135                         err = 0;
1136                         break;
1137                 }
1138                 path->slots[0]++;
1139         }
1140 fail:
1141         return err;
1142 }
1143
1144 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1145                                            struct btrfs_root *root,
1146                                            struct btrfs_path *path,
1147                                            u64 bytenr, u64 parent,
1148                                            u64 root_objectid, u64 owner,
1149                                            u64 offset, int refs_to_add)
1150 {
1151         struct btrfs_key key;
1152         struct extent_buffer *leaf;
1153         u32 size;
1154         u32 num_refs;
1155         int ret;
1156
1157         key.objectid = bytenr;
1158         if (parent) {
1159                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1160                 key.offset = parent;
1161                 size = sizeof(struct btrfs_shared_data_ref);
1162         } else {
1163                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1164                 key.offset = hash_extent_data_ref(root_objectid,
1165                                                   owner, offset);
1166                 size = sizeof(struct btrfs_extent_data_ref);
1167         }
1168
1169         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1170         if (ret && ret != -EEXIST)
1171                 goto fail;
1172
1173         leaf = path->nodes[0];
1174         if (parent) {
1175                 struct btrfs_shared_data_ref *ref;
1176                 ref = btrfs_item_ptr(leaf, path->slots[0],
1177                                      struct btrfs_shared_data_ref);
1178                 if (ret == 0) {
1179                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1180                 } else {
1181                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1182                         num_refs += refs_to_add;
1183                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1184                 }
1185         } else {
1186                 struct btrfs_extent_data_ref *ref;
1187                 while (ret == -EEXIST) {
1188                         ref = btrfs_item_ptr(leaf, path->slots[0],
1189                                              struct btrfs_extent_data_ref);
1190                         if (match_extent_data_ref(leaf, ref, root_objectid,
1191                                                   owner, offset))
1192                                 break;
1193                         btrfs_release_path(path);
1194                         key.offset++;
1195                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1196                                                       size);
1197                         if (ret && ret != -EEXIST)
1198                                 goto fail;
1199
1200                         leaf = path->nodes[0];
1201                 }
1202                 ref = btrfs_item_ptr(leaf, path->slots[0],
1203                                      struct btrfs_extent_data_ref);
1204                 if (ret == 0) {
1205                         btrfs_set_extent_data_ref_root(leaf, ref,
1206                                                        root_objectid);
1207                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1208                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1209                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1210                 } else {
1211                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1212                         num_refs += refs_to_add;
1213                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1214                 }
1215         }
1216         btrfs_mark_buffer_dirty(leaf);
1217         ret = 0;
1218 fail:
1219         btrfs_release_path(path);
1220         return ret;
1221 }
1222
1223 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1224                                            struct btrfs_root *root,
1225                                            struct btrfs_path *path,
1226                                            int refs_to_drop)
1227 {
1228         struct btrfs_key key;
1229         struct btrfs_extent_data_ref *ref1 = NULL;
1230         struct btrfs_shared_data_ref *ref2 = NULL;
1231         struct extent_buffer *leaf;
1232         u32 num_refs = 0;
1233         int ret = 0;
1234
1235         leaf = path->nodes[0];
1236         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1237
1238         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1239                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1240                                       struct btrfs_extent_data_ref);
1241                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1242         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1243                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1244                                       struct btrfs_shared_data_ref);
1245                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1246 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1247         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1248                 struct btrfs_extent_ref_v0 *ref0;
1249                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1250                                       struct btrfs_extent_ref_v0);
1251                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1252 #endif
1253         } else {
1254                 BUG();
1255         }
1256
1257         BUG_ON(num_refs < refs_to_drop);
1258         num_refs -= refs_to_drop;
1259
1260         if (num_refs == 0) {
1261                 ret = btrfs_del_item(trans, root, path);
1262         } else {
1263                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1264                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1265                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1266                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1267 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1268                 else {
1269                         struct btrfs_extent_ref_v0 *ref0;
1270                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1271                                         struct btrfs_extent_ref_v0);
1272                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1273                 }
1274 #endif
1275                 btrfs_mark_buffer_dirty(leaf);
1276         }
1277         return ret;
1278 }
1279
1280 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1281                                           struct btrfs_path *path,
1282                                           struct btrfs_extent_inline_ref *iref)
1283 {
1284         struct btrfs_key key;
1285         struct extent_buffer *leaf;
1286         struct btrfs_extent_data_ref *ref1;
1287         struct btrfs_shared_data_ref *ref2;
1288         u32 num_refs = 0;
1289
1290         leaf = path->nodes[0];
1291         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1292         if (iref) {
1293                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1294                     BTRFS_EXTENT_DATA_REF_KEY) {
1295                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1296                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1297                 } else {
1298                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1299                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1300                 }
1301         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1302                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1303                                       struct btrfs_extent_data_ref);
1304                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1305         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1306                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1307                                       struct btrfs_shared_data_ref);
1308                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1309 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1310         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1311                 struct btrfs_extent_ref_v0 *ref0;
1312                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1313                                       struct btrfs_extent_ref_v0);
1314                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1315 #endif
1316         } else {
1317                 WARN_ON(1);
1318         }
1319         return num_refs;
1320 }
1321
1322 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1323                                           struct btrfs_root *root,
1324                                           struct btrfs_path *path,
1325                                           u64 bytenr, u64 parent,
1326                                           u64 root_objectid)
1327 {
1328         struct btrfs_key key;
1329         int ret;
1330
1331         key.objectid = bytenr;
1332         if (parent) {
1333                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1334                 key.offset = parent;
1335         } else {
1336                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1337                 key.offset = root_objectid;
1338         }
1339
1340         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1341         if (ret > 0)
1342                 ret = -ENOENT;
1343 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1344         if (ret == -ENOENT && parent) {
1345                 btrfs_release_path(path);
1346                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1347                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1348                 if (ret > 0)
1349                         ret = -ENOENT;
1350         }
1351 #endif
1352         return ret;
1353 }
1354
1355 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1356                                           struct btrfs_root *root,
1357                                           struct btrfs_path *path,
1358                                           u64 bytenr, u64 parent,
1359                                           u64 root_objectid)
1360 {
1361         struct btrfs_key key;
1362         int ret;
1363
1364         key.objectid = bytenr;
1365         if (parent) {
1366                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1367                 key.offset = parent;
1368         } else {
1369                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1370                 key.offset = root_objectid;
1371         }
1372
1373         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1374         btrfs_release_path(path);
1375         return ret;
1376 }
1377
1378 static inline int extent_ref_type(u64 parent, u64 owner)
1379 {
1380         int type;
1381         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1382                 if (parent > 0)
1383                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1384                 else
1385                         type = BTRFS_TREE_BLOCK_REF_KEY;
1386         } else {
1387                 if (parent > 0)
1388                         type = BTRFS_SHARED_DATA_REF_KEY;
1389                 else
1390                         type = BTRFS_EXTENT_DATA_REF_KEY;
1391         }
1392         return type;
1393 }
1394
1395 static int find_next_key(struct btrfs_path *path, int level,
1396                          struct btrfs_key *key)
1397
1398 {
1399         for (; level < BTRFS_MAX_LEVEL; level++) {
1400                 if (!path->nodes[level])
1401                         break;
1402                 if (path->slots[level] + 1 >=
1403                     btrfs_header_nritems(path->nodes[level]))
1404                         continue;
1405                 if (level == 0)
1406                         btrfs_item_key_to_cpu(path->nodes[level], key,
1407                                               path->slots[level] + 1);
1408                 else
1409                         btrfs_node_key_to_cpu(path->nodes[level], key,
1410                                               path->slots[level] + 1);
1411                 return 0;
1412         }
1413         return 1;
1414 }
1415
1416 /*
1417  * look for inline back ref. if back ref is found, *ref_ret is set
1418  * to the address of inline back ref, and 0 is returned.
1419  *
1420  * if back ref isn't found, *ref_ret is set to the address where it
1421  * should be inserted, and -ENOENT is returned.
1422  *
1423  * if insert is true and there are too many inline back refs, the path
1424  * points to the extent item, and -EAGAIN is returned.
1425  *
1426  * NOTE: inline back refs are ordered in the same way that back ref
1427  *       items in the tree are ordered.
1428  */
1429 static noinline_for_stack
1430 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1431                                  struct btrfs_root *root,
1432                                  struct btrfs_path *path,
1433                                  struct btrfs_extent_inline_ref **ref_ret,
1434                                  u64 bytenr, u64 num_bytes,
1435                                  u64 parent, u64 root_objectid,
1436                                  u64 owner, u64 offset, int insert)
1437 {
1438         struct btrfs_key key;
1439         struct extent_buffer *leaf;
1440         struct btrfs_extent_item *ei;
1441         struct btrfs_extent_inline_ref *iref;
1442         u64 flags;
1443         u64 item_size;
1444         unsigned long ptr;
1445         unsigned long end;
1446         int extra_size;
1447         int type;
1448         int want;
1449         int ret;
1450         int err = 0;
1451
1452         key.objectid = bytenr;
1453         key.type = BTRFS_EXTENT_ITEM_KEY;
1454         key.offset = num_bytes;
1455
1456         want = extent_ref_type(parent, owner);
1457         if (insert) {
1458                 extra_size = btrfs_extent_inline_ref_size(want);
1459                 path->keep_locks = 1;
1460         } else
1461                 extra_size = -1;
1462         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1463         if (ret < 0) {
1464                 err = ret;
1465                 goto out;
1466         }
1467         if (ret && !insert) {
1468                 err = -ENOENT;
1469                 goto out;
1470         } else if (ret) {
1471                 err = -EIO;
1472                 WARN_ON(1);
1473                 goto out;
1474         }
1475
1476         leaf = path->nodes[0];
1477         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1478 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1479         if (item_size < sizeof(*ei)) {
1480                 if (!insert) {
1481                         err = -ENOENT;
1482                         goto out;
1483                 }
1484                 ret = convert_extent_item_v0(trans, root, path, owner,
1485                                              extra_size);
1486                 if (ret < 0) {
1487                         err = ret;
1488                         goto out;
1489                 }
1490                 leaf = path->nodes[0];
1491                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1492         }
1493 #endif
1494         BUG_ON(item_size < sizeof(*ei));
1495
1496         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1497         flags = btrfs_extent_flags(leaf, ei);
1498
1499         ptr = (unsigned long)(ei + 1);
1500         end = (unsigned long)ei + item_size;
1501
1502         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1503                 ptr += sizeof(struct btrfs_tree_block_info);
1504                 BUG_ON(ptr > end);
1505         } else {
1506                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1507         }
1508
1509         err = -ENOENT;
1510         while (1) {
1511                 if (ptr >= end) {
1512                         WARN_ON(ptr > end);
1513                         break;
1514                 }
1515                 iref = (struct btrfs_extent_inline_ref *)ptr;
1516                 type = btrfs_extent_inline_ref_type(leaf, iref);
1517                 if (want < type)
1518                         break;
1519                 if (want > type) {
1520                         ptr += btrfs_extent_inline_ref_size(type);
1521                         continue;
1522                 }
1523
1524                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1525                         struct btrfs_extent_data_ref *dref;
1526                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1527                         if (match_extent_data_ref(leaf, dref, root_objectid,
1528                                                   owner, offset)) {
1529                                 err = 0;
1530                                 break;
1531                         }
1532                         if (hash_extent_data_ref_item(leaf, dref) <
1533                             hash_extent_data_ref(root_objectid, owner, offset))
1534                                 break;
1535                 } else {
1536                         u64 ref_offset;
1537                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1538                         if (parent > 0) {
1539                                 if (parent == ref_offset) {
1540                                         err = 0;
1541                                         break;
1542                                 }
1543                                 if (ref_offset < parent)
1544                                         break;
1545                         } else {
1546                                 if (root_objectid == ref_offset) {
1547                                         err = 0;
1548                                         break;
1549                                 }
1550                                 if (ref_offset < root_objectid)
1551                                         break;
1552                         }
1553                 }
1554                 ptr += btrfs_extent_inline_ref_size(type);
1555         }
1556         if (err == -ENOENT && insert) {
1557                 if (item_size + extra_size >=
1558                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1559                         err = -EAGAIN;
1560                         goto out;
1561                 }
1562                 /*
1563                  * To add new inline back ref, we have to make sure
1564                  * there is no corresponding back ref item.
1565                  * For simplicity, we just do not add new inline back
1566                  * ref if there is any kind of item for this block
1567                  */
1568                 if (find_next_key(path, 0, &key) == 0 &&
1569                     key.objectid == bytenr &&
1570                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1571                         err = -EAGAIN;
1572                         goto out;
1573                 }
1574         }
1575         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1576 out:
1577         if (insert) {
1578                 path->keep_locks = 0;
1579                 btrfs_unlock_up_safe(path, 1);
1580         }
1581         return err;
1582 }
1583
1584 /*
1585  * helper to add new inline back ref
1586  */
1587 static noinline_for_stack
1588 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1589                                  struct btrfs_root *root,
1590                                  struct btrfs_path *path,
1591                                  struct btrfs_extent_inline_ref *iref,
1592                                  u64 parent, u64 root_objectid,
1593                                  u64 owner, u64 offset, int refs_to_add,
1594                                  struct btrfs_delayed_extent_op *extent_op)
1595 {
1596         struct extent_buffer *leaf;
1597         struct btrfs_extent_item *ei;
1598         unsigned long ptr;
1599         unsigned long end;
1600         unsigned long item_offset;
1601         u64 refs;
1602         int size;
1603         int type;
1604
1605         leaf = path->nodes[0];
1606         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1607         item_offset = (unsigned long)iref - (unsigned long)ei;
1608
1609         type = extent_ref_type(parent, owner);
1610         size = btrfs_extent_inline_ref_size(type);
1611
1612         btrfs_extend_item(trans, root, path, size);
1613
1614         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1615         refs = btrfs_extent_refs(leaf, ei);
1616         refs += refs_to_add;
1617         btrfs_set_extent_refs(leaf, ei, refs);
1618         if (extent_op)
1619                 __run_delayed_extent_op(extent_op, leaf, ei);
1620
1621         ptr = (unsigned long)ei + item_offset;
1622         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1623         if (ptr < end - size)
1624                 memmove_extent_buffer(leaf, ptr + size, ptr,
1625                                       end - size - ptr);
1626
1627         iref = (struct btrfs_extent_inline_ref *)ptr;
1628         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1629         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1630                 struct btrfs_extent_data_ref *dref;
1631                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1632                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1633                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1634                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1635                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1636         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1637                 struct btrfs_shared_data_ref *sref;
1638                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1639                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1640                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1641         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1642                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1643         } else {
1644                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1645         }
1646         btrfs_mark_buffer_dirty(leaf);
1647 }
1648
1649 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1650                                  struct btrfs_root *root,
1651                                  struct btrfs_path *path,
1652                                  struct btrfs_extent_inline_ref **ref_ret,
1653                                  u64 bytenr, u64 num_bytes, u64 parent,
1654                                  u64 root_objectid, u64 owner, u64 offset)
1655 {
1656         int ret;
1657
1658         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1659                                            bytenr, num_bytes, parent,
1660                                            root_objectid, owner, offset, 0);
1661         if (ret != -ENOENT)
1662                 return ret;
1663
1664         btrfs_release_path(path);
1665         *ref_ret = NULL;
1666
1667         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1668                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1669                                             root_objectid);
1670         } else {
1671                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1672                                              root_objectid, owner, offset);
1673         }
1674         return ret;
1675 }
1676
1677 /*
1678  * helper to update/remove inline back ref
1679  */
1680 static noinline_for_stack
1681 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1682                                   struct btrfs_root *root,
1683                                   struct btrfs_path *path,
1684                                   struct btrfs_extent_inline_ref *iref,
1685                                   int refs_to_mod,
1686                                   struct btrfs_delayed_extent_op *extent_op)
1687 {
1688         struct extent_buffer *leaf;
1689         struct btrfs_extent_item *ei;
1690         struct btrfs_extent_data_ref *dref = NULL;
1691         struct btrfs_shared_data_ref *sref = NULL;
1692         unsigned long ptr;
1693         unsigned long end;
1694         u32 item_size;
1695         int size;
1696         int type;
1697         u64 refs;
1698
1699         leaf = path->nodes[0];
1700         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1701         refs = btrfs_extent_refs(leaf, ei);
1702         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1703         refs += refs_to_mod;
1704         btrfs_set_extent_refs(leaf, ei, refs);
1705         if (extent_op)
1706                 __run_delayed_extent_op(extent_op, leaf, ei);
1707
1708         type = btrfs_extent_inline_ref_type(leaf, iref);
1709
1710         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1711                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1712                 refs = btrfs_extent_data_ref_count(leaf, dref);
1713         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1714                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1715                 refs = btrfs_shared_data_ref_count(leaf, sref);
1716         } else {
1717                 refs = 1;
1718                 BUG_ON(refs_to_mod != -1);
1719         }
1720
1721         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1722         refs += refs_to_mod;
1723
1724         if (refs > 0) {
1725                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1726                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1727                 else
1728                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1729         } else {
1730                 size =  btrfs_extent_inline_ref_size(type);
1731                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1732                 ptr = (unsigned long)iref;
1733                 end = (unsigned long)ei + item_size;
1734                 if (ptr + size < end)
1735                         memmove_extent_buffer(leaf, ptr, ptr + size,
1736                                               end - ptr - size);
1737                 item_size -= size;
1738                 btrfs_truncate_item(trans, root, path, item_size, 1);
1739         }
1740         btrfs_mark_buffer_dirty(leaf);
1741 }
1742
1743 static noinline_for_stack
1744 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1745                                  struct btrfs_root *root,
1746                                  struct btrfs_path *path,
1747                                  u64 bytenr, u64 num_bytes, u64 parent,
1748                                  u64 root_objectid, u64 owner,
1749                                  u64 offset, int refs_to_add,
1750                                  struct btrfs_delayed_extent_op *extent_op)
1751 {
1752         struct btrfs_extent_inline_ref *iref;
1753         int ret;
1754
1755         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1756                                            bytenr, num_bytes, parent,
1757                                            root_objectid, owner, offset, 1);
1758         if (ret == 0) {
1759                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1760                 update_inline_extent_backref(trans, root, path, iref,
1761                                              refs_to_add, extent_op);
1762         } else if (ret == -ENOENT) {
1763                 setup_inline_extent_backref(trans, root, path, iref, parent,
1764                                             root_objectid, owner, offset,
1765                                             refs_to_add, extent_op);
1766                 ret = 0;
1767         }
1768         return ret;
1769 }
1770
1771 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1772                                  struct btrfs_root *root,
1773                                  struct btrfs_path *path,
1774                                  u64 bytenr, u64 parent, u64 root_objectid,
1775                                  u64 owner, u64 offset, int refs_to_add)
1776 {
1777         int ret;
1778         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1779                 BUG_ON(refs_to_add != 1);
1780                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1781                                             parent, root_objectid);
1782         } else {
1783                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1784                                              parent, root_objectid,
1785                                              owner, offset, refs_to_add);
1786         }
1787         return ret;
1788 }
1789
1790 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1791                                  struct btrfs_root *root,
1792                                  struct btrfs_path *path,
1793                                  struct btrfs_extent_inline_ref *iref,
1794                                  int refs_to_drop, int is_data)
1795 {
1796         int ret = 0;
1797
1798         BUG_ON(!is_data && refs_to_drop != 1);
1799         if (iref) {
1800                 update_inline_extent_backref(trans, root, path, iref,
1801                                              -refs_to_drop, NULL);
1802         } else if (is_data) {
1803                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1804         } else {
1805                 ret = btrfs_del_item(trans, root, path);
1806         }
1807         return ret;
1808 }
1809
1810 static int btrfs_issue_discard(struct block_device *bdev,
1811                                 u64 start, u64 len)
1812 {
1813         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1814 }
1815
1816 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1817                                 u64 num_bytes, u64 *actual_bytes)
1818 {
1819         int ret;
1820         u64 discarded_bytes = 0;
1821         struct btrfs_bio *bbio = NULL;
1822
1823
1824         /* Tell the block device(s) that the sectors can be discarded */
1825         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1826                               bytenr, &num_bytes, &bbio, 0);
1827         /* Error condition is -ENOMEM */
1828         if (!ret) {
1829                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1830                 int i;
1831
1832
1833                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1834                         if (!stripe->dev->can_discard)
1835                                 continue;
1836
1837                         ret = btrfs_issue_discard(stripe->dev->bdev,
1838                                                   stripe->physical,
1839                                                   stripe->length);
1840                         if (!ret)
1841                                 discarded_bytes += stripe->length;
1842                         else if (ret != -EOPNOTSUPP)
1843                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1844
1845                         /*
1846                          * Just in case we get back EOPNOTSUPP for some reason,
1847                          * just ignore the return value so we don't screw up
1848                          * people calling discard_extent.
1849                          */
1850                         ret = 0;
1851                 }
1852                 kfree(bbio);
1853         }
1854
1855         if (actual_bytes)
1856                 *actual_bytes = discarded_bytes;
1857
1858
1859         if (ret == -EOPNOTSUPP)
1860                 ret = 0;
1861         return ret;
1862 }
1863
1864 /* Can return -ENOMEM */
1865 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1866                          struct btrfs_root *root,
1867                          u64 bytenr, u64 num_bytes, u64 parent,
1868                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1869 {
1870         int ret;
1871         struct btrfs_fs_info *fs_info = root->fs_info;
1872
1873         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1874                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1875
1876         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1877                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1878                                         num_bytes,
1879                                         parent, root_objectid, (int)owner,
1880                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1881         } else {
1882                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1883                                         num_bytes,
1884                                         parent, root_objectid, owner, offset,
1885                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1886         }
1887         return ret;
1888 }
1889
1890 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1891                                   struct btrfs_root *root,
1892                                   u64 bytenr, u64 num_bytes,
1893                                   u64 parent, u64 root_objectid,
1894                                   u64 owner, u64 offset, int refs_to_add,
1895                                   struct btrfs_delayed_extent_op *extent_op)
1896 {
1897         struct btrfs_path *path;
1898         struct extent_buffer *leaf;
1899         struct btrfs_extent_item *item;
1900         u64 refs;
1901         int ret;
1902         int err = 0;
1903
1904         path = btrfs_alloc_path();
1905         if (!path)
1906                 return -ENOMEM;
1907
1908         path->reada = 1;
1909         path->leave_spinning = 1;
1910         /* this will setup the path even if it fails to insert the back ref */
1911         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1912                                            path, bytenr, num_bytes, parent,
1913                                            root_objectid, owner, offset,
1914                                            refs_to_add, extent_op);
1915         if (ret == 0)
1916                 goto out;
1917
1918         if (ret != -EAGAIN) {
1919                 err = ret;
1920                 goto out;
1921         }
1922
1923         leaf = path->nodes[0];
1924         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1925         refs = btrfs_extent_refs(leaf, item);
1926         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1927         if (extent_op)
1928                 __run_delayed_extent_op(extent_op, leaf, item);
1929
1930         btrfs_mark_buffer_dirty(leaf);
1931         btrfs_release_path(path);
1932
1933         path->reada = 1;
1934         path->leave_spinning = 1;
1935
1936         /* now insert the actual backref */
1937         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1938                                     path, bytenr, parent, root_objectid,
1939                                     owner, offset, refs_to_add);
1940         if (ret)
1941                 btrfs_abort_transaction(trans, root, ret);
1942 out:
1943         btrfs_free_path(path);
1944         return err;
1945 }
1946
1947 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1948                                 struct btrfs_root *root,
1949                                 struct btrfs_delayed_ref_node *node,
1950                                 struct btrfs_delayed_extent_op *extent_op,
1951                                 int insert_reserved)
1952 {
1953         int ret = 0;
1954         struct btrfs_delayed_data_ref *ref;
1955         struct btrfs_key ins;
1956         u64 parent = 0;
1957         u64 ref_root = 0;
1958         u64 flags = 0;
1959
1960         ins.objectid = node->bytenr;
1961         ins.offset = node->num_bytes;
1962         ins.type = BTRFS_EXTENT_ITEM_KEY;
1963
1964         ref = btrfs_delayed_node_to_data_ref(node);
1965         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1966                 parent = ref->parent;
1967         else
1968                 ref_root = ref->root;
1969
1970         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1971                 if (extent_op) {
1972                         BUG_ON(extent_op->update_key);
1973                         flags |= extent_op->flags_to_set;
1974                 }
1975                 ret = alloc_reserved_file_extent(trans, root,
1976                                                  parent, ref_root, flags,
1977                                                  ref->objectid, ref->offset,
1978                                                  &ins, node->ref_mod);
1979         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1980                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1981                                              node->num_bytes, parent,
1982                                              ref_root, ref->objectid,
1983                                              ref->offset, node->ref_mod,
1984                                              extent_op);
1985         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1986                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1987                                           node->num_bytes, parent,
1988                                           ref_root, ref->objectid,
1989                                           ref->offset, node->ref_mod,
1990                                           extent_op);
1991         } else {
1992                 BUG();
1993         }
1994         return ret;
1995 }
1996
1997 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1998                                     struct extent_buffer *leaf,
1999                                     struct btrfs_extent_item *ei)
2000 {
2001         u64 flags = btrfs_extent_flags(leaf, ei);
2002         if (extent_op->update_flags) {
2003                 flags |= extent_op->flags_to_set;
2004                 btrfs_set_extent_flags(leaf, ei, flags);
2005         }
2006
2007         if (extent_op->update_key) {
2008                 struct btrfs_tree_block_info *bi;
2009                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2010                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2011                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2012         }
2013 }
2014
2015 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2016                                  struct btrfs_root *root,
2017                                  struct btrfs_delayed_ref_node *node,
2018                                  struct btrfs_delayed_extent_op *extent_op)
2019 {
2020         struct btrfs_key key;
2021         struct btrfs_path *path;
2022         struct btrfs_extent_item *ei;
2023         struct extent_buffer *leaf;
2024         u32 item_size;
2025         int ret;
2026         int err = 0;
2027
2028         if (trans->aborted)
2029                 return 0;
2030
2031         path = btrfs_alloc_path();
2032         if (!path)
2033                 return -ENOMEM;
2034
2035         key.objectid = node->bytenr;
2036         key.type = BTRFS_EXTENT_ITEM_KEY;
2037         key.offset = node->num_bytes;
2038
2039         path->reada = 1;
2040         path->leave_spinning = 1;
2041         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2042                                 path, 0, 1);
2043         if (ret < 0) {
2044                 err = ret;
2045                 goto out;
2046         }
2047         if (ret > 0) {
2048                 err = -EIO;
2049                 goto out;
2050         }
2051
2052         leaf = path->nodes[0];
2053         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2054 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2055         if (item_size < sizeof(*ei)) {
2056                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2057                                              path, (u64)-1, 0);
2058                 if (ret < 0) {
2059                         err = ret;
2060                         goto out;
2061                 }
2062                 leaf = path->nodes[0];
2063                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2064         }
2065 #endif
2066         BUG_ON(item_size < sizeof(*ei));
2067         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2068         __run_delayed_extent_op(extent_op, leaf, ei);
2069
2070         btrfs_mark_buffer_dirty(leaf);
2071 out:
2072         btrfs_free_path(path);
2073         return err;
2074 }
2075
2076 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2077                                 struct btrfs_root *root,
2078                                 struct btrfs_delayed_ref_node *node,
2079                                 struct btrfs_delayed_extent_op *extent_op,
2080                                 int insert_reserved)
2081 {
2082         int ret = 0;
2083         struct btrfs_delayed_tree_ref *ref;
2084         struct btrfs_key ins;
2085         u64 parent = 0;
2086         u64 ref_root = 0;
2087
2088         ins.objectid = node->bytenr;
2089         ins.offset = node->num_bytes;
2090         ins.type = BTRFS_EXTENT_ITEM_KEY;
2091
2092         ref = btrfs_delayed_node_to_tree_ref(node);
2093         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2094                 parent = ref->parent;
2095         else
2096                 ref_root = ref->root;
2097
2098         BUG_ON(node->ref_mod != 1);
2099         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2100                 BUG_ON(!extent_op || !extent_op->update_flags ||
2101                        !extent_op->update_key);
2102                 ret = alloc_reserved_tree_block(trans, root,
2103                                                 parent, ref_root,
2104                                                 extent_op->flags_to_set,
2105                                                 &extent_op->key,
2106                                                 ref->level, &ins);
2107         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2108                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2109                                              node->num_bytes, parent, ref_root,
2110                                              ref->level, 0, 1, extent_op);
2111         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2112                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2113                                           node->num_bytes, parent, ref_root,
2114                                           ref->level, 0, 1, extent_op);
2115         } else {
2116                 BUG();
2117         }
2118         return ret;
2119 }
2120
2121 /* helper function to actually process a single delayed ref entry */
2122 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2123                                struct btrfs_root *root,
2124                                struct btrfs_delayed_ref_node *node,
2125                                struct btrfs_delayed_extent_op *extent_op,
2126                                int insert_reserved)
2127 {
2128         int ret = 0;
2129
2130         if (trans->aborted)
2131                 return 0;
2132
2133         if (btrfs_delayed_ref_is_head(node)) {
2134                 struct btrfs_delayed_ref_head *head;
2135                 /*
2136                  * we've hit the end of the chain and we were supposed
2137                  * to insert this extent into the tree.  But, it got
2138                  * deleted before we ever needed to insert it, so all
2139                  * we have to do is clean up the accounting
2140                  */
2141                 BUG_ON(extent_op);
2142                 head = btrfs_delayed_node_to_head(node);
2143                 if (insert_reserved) {
2144                         btrfs_pin_extent(root, node->bytenr,
2145                                          node->num_bytes, 1);
2146                         if (head->is_data) {
2147                                 ret = btrfs_del_csums(trans, root,
2148                                                       node->bytenr,
2149                                                       node->num_bytes);
2150                         }
2151                 }
2152                 return ret;
2153         }
2154
2155         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2156             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2157                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2158                                            insert_reserved);
2159         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2160                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2161                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2162                                            insert_reserved);
2163         else
2164                 BUG();
2165         return ret;
2166 }
2167
2168 static noinline struct btrfs_delayed_ref_node *
2169 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2170 {
2171         struct rb_node *node;
2172         struct btrfs_delayed_ref_node *ref;
2173         int action = BTRFS_ADD_DELAYED_REF;
2174 again:
2175         /*
2176          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2177          * this prevents ref count from going down to zero when
2178          * there still are pending delayed ref.
2179          */
2180         node = rb_prev(&head->node.rb_node);
2181         while (1) {
2182                 if (!node)
2183                         break;
2184                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2185                                 rb_node);
2186                 if (ref->bytenr != head->node.bytenr)
2187                         break;
2188                 if (ref->action == action)
2189                         return ref;
2190                 node = rb_prev(node);
2191         }
2192         if (action == BTRFS_ADD_DELAYED_REF) {
2193                 action = BTRFS_DROP_DELAYED_REF;
2194                 goto again;
2195         }
2196         return NULL;
2197 }
2198
2199 /*
2200  * Returns 0 on success or if called with an already aborted transaction.
2201  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2202  */
2203 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2204                                        struct btrfs_root *root,
2205                                        struct list_head *cluster)
2206 {
2207         struct btrfs_delayed_ref_root *delayed_refs;
2208         struct btrfs_delayed_ref_node *ref;
2209         struct btrfs_delayed_ref_head *locked_ref = NULL;
2210         struct btrfs_delayed_extent_op *extent_op;
2211         struct btrfs_fs_info *fs_info = root->fs_info;
2212         int ret;
2213         int count = 0;
2214         int must_insert_reserved = 0;
2215
2216         delayed_refs = &trans->transaction->delayed_refs;
2217         while (1) {
2218                 if (!locked_ref) {
2219                         /* pick a new head ref from the cluster list */
2220                         if (list_empty(cluster))
2221                                 break;
2222
2223                         locked_ref = list_entry(cluster->next,
2224                                      struct btrfs_delayed_ref_head, cluster);
2225
2226                         /* grab the lock that says we are going to process
2227                          * all the refs for this head */
2228                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2229
2230                         /*
2231                          * we may have dropped the spin lock to get the head
2232                          * mutex lock, and that might have given someone else
2233                          * time to free the head.  If that's true, it has been
2234                          * removed from our list and we can move on.
2235                          */
2236                         if (ret == -EAGAIN) {
2237                                 locked_ref = NULL;
2238                                 count++;
2239                                 continue;
2240                         }
2241                 }
2242
2243                 /*
2244                  * We need to try and merge add/drops of the same ref since we
2245                  * can run into issues with relocate dropping the implicit ref
2246                  * and then it being added back again before the drop can
2247                  * finish.  If we merged anything we need to re-loop so we can
2248                  * get a good ref.
2249                  */
2250                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2251                                          locked_ref);
2252
2253                 /*
2254                  * locked_ref is the head node, so we have to go one
2255                  * node back for any delayed ref updates
2256                  */
2257                 ref = select_delayed_ref(locked_ref);
2258
2259                 if (ref && ref->seq &&
2260                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2261                         /*
2262                          * there are still refs with lower seq numbers in the
2263                          * process of being added. Don't run this ref yet.
2264                          */
2265                         list_del_init(&locked_ref->cluster);
2266                         btrfs_delayed_ref_unlock(locked_ref);
2267                         locked_ref = NULL;
2268                         delayed_refs->num_heads_ready++;
2269                         spin_unlock(&delayed_refs->lock);
2270                         cond_resched();
2271                         spin_lock(&delayed_refs->lock);
2272                         continue;
2273                 }
2274
2275                 /*
2276                  * record the must insert reserved flag before we
2277                  * drop the spin lock.
2278                  */
2279                 must_insert_reserved = locked_ref->must_insert_reserved;
2280                 locked_ref->must_insert_reserved = 0;
2281
2282                 extent_op = locked_ref->extent_op;
2283                 locked_ref->extent_op = NULL;
2284
2285                 if (!ref) {
2286                         /* All delayed refs have been processed, Go ahead
2287                          * and send the head node to run_one_delayed_ref,
2288                          * so that any accounting fixes can happen
2289                          */
2290                         ref = &locked_ref->node;
2291
2292                         if (extent_op && must_insert_reserved) {
2293                                 btrfs_free_delayed_extent_op(extent_op);
2294                                 extent_op = NULL;
2295                         }
2296
2297                         if (extent_op) {
2298                                 spin_unlock(&delayed_refs->lock);
2299
2300                                 ret = run_delayed_extent_op(trans, root,
2301                                                             ref, extent_op);
2302                                 btrfs_free_delayed_extent_op(extent_op);
2303
2304                                 if (ret) {
2305                                         printk(KERN_DEBUG
2306                                                "btrfs: run_delayed_extent_op "
2307                                                "returned %d\n", ret);
2308                                         spin_lock(&delayed_refs->lock);
2309                                         btrfs_delayed_ref_unlock(locked_ref);
2310                                         return ret;
2311                                 }
2312
2313                                 goto next;
2314                         }
2315                 }
2316
2317                 ref->in_tree = 0;
2318                 rb_erase(&ref->rb_node, &delayed_refs->root);
2319                 delayed_refs->num_entries--;
2320                 if (!btrfs_delayed_ref_is_head(ref)) {
2321                         /*
2322                          * when we play the delayed ref, also correct the
2323                          * ref_mod on head
2324                          */
2325                         switch (ref->action) {
2326                         case BTRFS_ADD_DELAYED_REF:
2327                         case BTRFS_ADD_DELAYED_EXTENT:
2328                                 locked_ref->node.ref_mod -= ref->ref_mod;
2329                                 break;
2330                         case BTRFS_DROP_DELAYED_REF:
2331                                 locked_ref->node.ref_mod += ref->ref_mod;
2332                                 break;
2333                         default:
2334                                 WARN_ON(1);
2335                         }
2336                 }
2337                 spin_unlock(&delayed_refs->lock);
2338
2339                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2340                                           must_insert_reserved);
2341
2342                 btrfs_free_delayed_extent_op(extent_op);
2343                 if (ret) {
2344                         btrfs_delayed_ref_unlock(locked_ref);
2345                         btrfs_put_delayed_ref(ref);
2346                         printk(KERN_DEBUG
2347                                "btrfs: run_one_delayed_ref returned %d\n", ret);
2348                         spin_lock(&delayed_refs->lock);
2349                         return ret;
2350                 }
2351
2352                 /*
2353                  * If this node is a head, that means all the refs in this head
2354                  * have been dealt with, and we will pick the next head to deal
2355                  * with, so we must unlock the head and drop it from the cluster
2356                  * list before we release it.
2357                  */
2358                 if (btrfs_delayed_ref_is_head(ref)) {
2359                         list_del_init(&locked_ref->cluster);
2360                         btrfs_delayed_ref_unlock(locked_ref);
2361                         locked_ref = NULL;
2362                 }
2363                 btrfs_put_delayed_ref(ref);
2364                 count++;
2365 next:
2366                 cond_resched();
2367                 spin_lock(&delayed_refs->lock);
2368         }
2369         return count;
2370 }
2371
2372 #ifdef SCRAMBLE_DELAYED_REFS
2373 /*
2374  * Normally delayed refs get processed in ascending bytenr order. This
2375  * correlates in most cases to the order added. To expose dependencies on this
2376  * order, we start to process the tree in the middle instead of the beginning
2377  */
2378 static u64 find_middle(struct rb_root *root)
2379 {
2380         struct rb_node *n = root->rb_node;
2381         struct btrfs_delayed_ref_node *entry;
2382         int alt = 1;
2383         u64 middle;
2384         u64 first = 0, last = 0;
2385
2386         n = rb_first(root);
2387         if (n) {
2388                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2389                 first = entry->bytenr;
2390         }
2391         n = rb_last(root);
2392         if (n) {
2393                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2394                 last = entry->bytenr;
2395         }
2396         n = root->rb_node;
2397
2398         while (n) {
2399                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2400                 WARN_ON(!entry->in_tree);
2401
2402                 middle = entry->bytenr;
2403
2404                 if (alt)
2405                         n = n->rb_left;
2406                 else
2407                         n = n->rb_right;
2408
2409                 alt = 1 - alt;
2410         }
2411         return middle;
2412 }
2413 #endif
2414
2415 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2416                                          struct btrfs_fs_info *fs_info)
2417 {
2418         struct qgroup_update *qgroup_update;
2419         int ret = 0;
2420
2421         if (list_empty(&trans->qgroup_ref_list) !=
2422             !trans->delayed_ref_elem.seq) {
2423                 /* list without seq or seq without list */
2424                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2425                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2426                         trans->delayed_ref_elem.seq);
2427                 BUG();
2428         }
2429
2430         if (!trans->delayed_ref_elem.seq)
2431                 return 0;
2432
2433         while (!list_empty(&trans->qgroup_ref_list)) {
2434                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2435                                                  struct qgroup_update, list);
2436                 list_del(&qgroup_update->list);
2437                 if (!ret)
2438                         ret = btrfs_qgroup_account_ref(
2439                                         trans, fs_info, qgroup_update->node,
2440                                         qgroup_update->extent_op);
2441                 kfree(qgroup_update);
2442         }
2443
2444         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2445
2446         return ret;
2447 }
2448
2449 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2450                       int count)
2451 {
2452         int val = atomic_read(&delayed_refs->ref_seq);
2453
2454         if (val < seq || val >= seq + count)
2455                 return 1;
2456         return 0;
2457 }
2458
2459 /*
2460  * this starts processing the delayed reference count updates and
2461  * extent insertions we have queued up so far.  count can be
2462  * 0, which means to process everything in the tree at the start
2463  * of the run (but not newly added entries), or it can be some target
2464  * number you'd like to process.
2465  *
2466  * Returns 0 on success or if called with an aborted transaction
2467  * Returns <0 on error and aborts the transaction
2468  */
2469 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2470                            struct btrfs_root *root, unsigned long count)
2471 {
2472         struct rb_node *node;
2473         struct btrfs_delayed_ref_root *delayed_refs;
2474         struct btrfs_delayed_ref_node *ref;
2475         struct list_head cluster;
2476         int ret;
2477         u64 delayed_start;
2478         int run_all = count == (unsigned long)-1;
2479         int run_most = 0;
2480         int loops;
2481
2482         /* We'll clean this up in btrfs_cleanup_transaction */
2483         if (trans->aborted)
2484                 return 0;
2485
2486         if (root == root->fs_info->extent_root)
2487                 root = root->fs_info->tree_root;
2488
2489         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2490
2491         delayed_refs = &trans->transaction->delayed_refs;
2492         INIT_LIST_HEAD(&cluster);
2493         if (count == 0) {
2494                 count = delayed_refs->num_entries * 2;
2495                 run_most = 1;
2496         }
2497
2498         if (!run_all && !run_most) {
2499                 int old;
2500                 int seq = atomic_read(&delayed_refs->ref_seq);
2501
2502 progress:
2503                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2504                 if (old) {
2505                         DEFINE_WAIT(__wait);
2506                         if (delayed_refs->num_entries < 16348)
2507                                 return 0;
2508
2509                         prepare_to_wait(&delayed_refs->wait, &__wait,
2510                                         TASK_UNINTERRUPTIBLE);
2511
2512                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2513                         if (old) {
2514                                 schedule();
2515                                 finish_wait(&delayed_refs->wait, &__wait);
2516
2517                                 if (!refs_newer(delayed_refs, seq, 256))
2518                                         goto progress;
2519                                 else
2520                                         return 0;
2521                         } else {
2522                                 finish_wait(&delayed_refs->wait, &__wait);
2523                                 goto again;
2524                         }
2525                 }
2526
2527         } else {
2528                 atomic_inc(&delayed_refs->procs_running_refs);
2529         }
2530
2531 again:
2532         loops = 0;
2533         spin_lock(&delayed_refs->lock);
2534
2535 #ifdef SCRAMBLE_DELAYED_REFS
2536         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2537 #endif
2538
2539         while (1) {
2540                 if (!(run_all || run_most) &&
2541                     delayed_refs->num_heads_ready < 64)
2542                         break;
2543
2544                 /*
2545                  * go find something we can process in the rbtree.  We start at
2546                  * the beginning of the tree, and then build a cluster
2547                  * of refs to process starting at the first one we are able to
2548                  * lock
2549                  */
2550                 delayed_start = delayed_refs->run_delayed_start;
2551                 ret = btrfs_find_ref_cluster(trans, &cluster,
2552                                              delayed_refs->run_delayed_start);
2553                 if (ret)
2554                         break;
2555
2556                 ret = run_clustered_refs(trans, root, &cluster);
2557                 if (ret < 0) {
2558                         btrfs_release_ref_cluster(&cluster);
2559                         spin_unlock(&delayed_refs->lock);
2560                         btrfs_abort_transaction(trans, root, ret);
2561                         atomic_dec(&delayed_refs->procs_running_refs);
2562                         return ret;
2563                 }
2564
2565                 atomic_add(ret, &delayed_refs->ref_seq);
2566
2567                 count -= min_t(unsigned long, ret, count);
2568
2569                 if (count == 0)
2570                         break;
2571
2572                 if (delayed_start >= delayed_refs->run_delayed_start) {
2573                         if (loops == 0) {
2574                                 /*
2575                                  * btrfs_find_ref_cluster looped. let's do one
2576                                  * more cycle. if we don't run any delayed ref
2577                                  * during that cycle (because we can't because
2578                                  * all of them are blocked), bail out.
2579                                  */
2580                                 loops = 1;
2581                         } else {
2582                                 /*
2583                                  * no runnable refs left, stop trying
2584                                  */
2585                                 BUG_ON(run_all);
2586                                 break;
2587                         }
2588                 }
2589                 if (ret) {
2590                         /* refs were run, let's reset staleness detection */
2591                         loops = 0;
2592                 }
2593         }
2594
2595         if (run_all) {
2596                 if (!list_empty(&trans->new_bgs)) {
2597                         spin_unlock(&delayed_refs->lock);
2598                         btrfs_create_pending_block_groups(trans, root);
2599                         spin_lock(&delayed_refs->lock);
2600                 }
2601
2602                 node = rb_first(&delayed_refs->root);
2603                 if (!node)
2604                         goto out;
2605                 count = (unsigned long)-1;
2606
2607                 while (node) {
2608                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2609                                        rb_node);
2610                         if (btrfs_delayed_ref_is_head(ref)) {
2611                                 struct btrfs_delayed_ref_head *head;
2612
2613                                 head = btrfs_delayed_node_to_head(ref);
2614                                 atomic_inc(&ref->refs);
2615
2616                                 spin_unlock(&delayed_refs->lock);
2617                                 /*
2618                                  * Mutex was contended, block until it's
2619                                  * released and try again
2620                                  */
2621                                 mutex_lock(&head->mutex);
2622                                 mutex_unlock(&head->mutex);
2623
2624                                 btrfs_put_delayed_ref(ref);
2625                                 cond_resched();
2626                                 goto again;
2627                         }
2628                         node = rb_next(node);
2629                 }
2630                 spin_unlock(&delayed_refs->lock);
2631                 schedule_timeout(1);
2632                 goto again;
2633         }
2634 out:
2635         atomic_dec(&delayed_refs->procs_running_refs);
2636         smp_mb();
2637         if (waitqueue_active(&delayed_refs->wait))
2638                 wake_up(&delayed_refs->wait);
2639
2640         spin_unlock(&delayed_refs->lock);
2641         assert_qgroups_uptodate(trans);
2642         return 0;
2643 }
2644
2645 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2646                                 struct btrfs_root *root,
2647                                 u64 bytenr, u64 num_bytes, u64 flags,
2648                                 int is_data)
2649 {
2650         struct btrfs_delayed_extent_op *extent_op;
2651         int ret;
2652
2653         extent_op = btrfs_alloc_delayed_extent_op();
2654         if (!extent_op)
2655                 return -ENOMEM;
2656
2657         extent_op->flags_to_set = flags;
2658         extent_op->update_flags = 1;
2659         extent_op->update_key = 0;
2660         extent_op->is_data = is_data ? 1 : 0;
2661
2662         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2663                                           num_bytes, extent_op);
2664         if (ret)
2665                 btrfs_free_delayed_extent_op(extent_op);
2666         return ret;
2667 }
2668
2669 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2670                                       struct btrfs_root *root,
2671                                       struct btrfs_path *path,
2672                                       u64 objectid, u64 offset, u64 bytenr)
2673 {
2674         struct btrfs_delayed_ref_head *head;
2675         struct btrfs_delayed_ref_node *ref;
2676         struct btrfs_delayed_data_ref *data_ref;
2677         struct btrfs_delayed_ref_root *delayed_refs;
2678         struct rb_node *node;
2679         int ret = 0;
2680
2681         ret = -ENOENT;
2682         delayed_refs = &trans->transaction->delayed_refs;
2683         spin_lock(&delayed_refs->lock);
2684         head = btrfs_find_delayed_ref_head(trans, bytenr);
2685         if (!head)
2686                 goto out;
2687
2688         if (!mutex_trylock(&head->mutex)) {
2689                 atomic_inc(&head->node.refs);
2690                 spin_unlock(&delayed_refs->lock);
2691
2692                 btrfs_release_path(path);
2693
2694                 /*
2695                  * Mutex was contended, block until it's released and let
2696                  * caller try again
2697                  */
2698                 mutex_lock(&head->mutex);
2699                 mutex_unlock(&head->mutex);
2700                 btrfs_put_delayed_ref(&head->node);
2701                 return -EAGAIN;
2702         }
2703
2704         node = rb_prev(&head->node.rb_node);
2705         if (!node)
2706                 goto out_unlock;
2707
2708         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2709
2710         if (ref->bytenr != bytenr)
2711                 goto out_unlock;
2712
2713         ret = 1;
2714         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2715                 goto out_unlock;
2716
2717         data_ref = btrfs_delayed_node_to_data_ref(ref);
2718
2719         node = rb_prev(node);
2720         if (node) {
2721                 int seq = ref->seq;
2722
2723                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2724                 if (ref->bytenr == bytenr && ref->seq == seq)
2725                         goto out_unlock;
2726         }
2727
2728         if (data_ref->root != root->root_key.objectid ||
2729             data_ref->objectid != objectid || data_ref->offset != offset)
2730                 goto out_unlock;
2731
2732         ret = 0;
2733 out_unlock:
2734         mutex_unlock(&head->mutex);
2735 out:
2736         spin_unlock(&delayed_refs->lock);
2737         return ret;
2738 }
2739
2740 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2741                                         struct btrfs_root *root,
2742                                         struct btrfs_path *path,
2743                                         u64 objectid, u64 offset, u64 bytenr)
2744 {
2745         struct btrfs_root *extent_root = root->fs_info->extent_root;
2746         struct extent_buffer *leaf;
2747         struct btrfs_extent_data_ref *ref;
2748         struct btrfs_extent_inline_ref *iref;
2749         struct btrfs_extent_item *ei;
2750         struct btrfs_key key;
2751         u32 item_size;
2752         int ret;
2753
2754         key.objectid = bytenr;
2755         key.offset = (u64)-1;
2756         key.type = BTRFS_EXTENT_ITEM_KEY;
2757
2758         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2759         if (ret < 0)
2760                 goto out;
2761         BUG_ON(ret == 0); /* Corruption */
2762
2763         ret = -ENOENT;
2764         if (path->slots[0] == 0)
2765                 goto out;
2766
2767         path->slots[0]--;
2768         leaf = path->nodes[0];
2769         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2770
2771         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2772                 goto out;
2773
2774         ret = 1;
2775         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2776 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2777         if (item_size < sizeof(*ei)) {
2778                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2779                 goto out;
2780         }
2781 #endif
2782         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2783
2784         if (item_size != sizeof(*ei) +
2785             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2786                 goto out;
2787
2788         if (btrfs_extent_generation(leaf, ei) <=
2789             btrfs_root_last_snapshot(&root->root_item))
2790                 goto out;
2791
2792         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2793         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2794             BTRFS_EXTENT_DATA_REF_KEY)
2795                 goto out;
2796
2797         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2798         if (btrfs_extent_refs(leaf, ei) !=
2799             btrfs_extent_data_ref_count(leaf, ref) ||
2800             btrfs_extent_data_ref_root(leaf, ref) !=
2801             root->root_key.objectid ||
2802             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2803             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2804                 goto out;
2805
2806         ret = 0;
2807 out:
2808         return ret;
2809 }
2810
2811 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2812                           struct btrfs_root *root,
2813                           u64 objectid, u64 offset, u64 bytenr)
2814 {
2815         struct btrfs_path *path;
2816         int ret;
2817         int ret2;
2818
2819         path = btrfs_alloc_path();
2820         if (!path)
2821                 return -ENOENT;
2822
2823         do {
2824                 ret = check_committed_ref(trans, root, path, objectid,
2825                                           offset, bytenr);
2826                 if (ret && ret != -ENOENT)
2827                         goto out;
2828
2829                 ret2 = check_delayed_ref(trans, root, path, objectid,
2830                                          offset, bytenr);
2831         } while (ret2 == -EAGAIN);
2832
2833         if (ret2 && ret2 != -ENOENT) {
2834                 ret = ret2;
2835                 goto out;
2836         }
2837
2838         if (ret != -ENOENT || ret2 != -ENOENT)
2839                 ret = 0;
2840 out:
2841         btrfs_free_path(path);
2842         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2843                 WARN_ON(ret > 0);
2844         return ret;
2845 }
2846
2847 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2848                            struct btrfs_root *root,
2849                            struct extent_buffer *buf,
2850                            int full_backref, int inc, int for_cow)
2851 {
2852         u64 bytenr;
2853         u64 num_bytes;
2854         u64 parent;
2855         u64 ref_root;
2856         u32 nritems;
2857         struct btrfs_key key;
2858         struct btrfs_file_extent_item *fi;
2859         int i;
2860         int level;
2861         int ret = 0;
2862         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2863                             u64, u64, u64, u64, u64, u64, int);
2864
2865         ref_root = btrfs_header_owner(buf);
2866         nritems = btrfs_header_nritems(buf);
2867         level = btrfs_header_level(buf);
2868
2869         if (!root->ref_cows && level == 0)
2870                 return 0;
2871
2872         if (inc)
2873                 process_func = btrfs_inc_extent_ref;
2874         else
2875                 process_func = btrfs_free_extent;
2876
2877         if (full_backref)
2878                 parent = buf->start;
2879         else
2880                 parent = 0;
2881
2882         for (i = 0; i < nritems; i++) {
2883                 if (level == 0) {
2884                         btrfs_item_key_to_cpu(buf, &key, i);
2885                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2886                                 continue;
2887                         fi = btrfs_item_ptr(buf, i,
2888                                             struct btrfs_file_extent_item);
2889                         if (btrfs_file_extent_type(buf, fi) ==
2890                             BTRFS_FILE_EXTENT_INLINE)
2891                                 continue;
2892                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2893                         if (bytenr == 0)
2894                                 continue;
2895
2896                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2897                         key.offset -= btrfs_file_extent_offset(buf, fi);
2898                         ret = process_func(trans, root, bytenr, num_bytes,
2899                                            parent, ref_root, key.objectid,
2900                                            key.offset, for_cow);
2901                         if (ret)
2902                                 goto fail;
2903                 } else {
2904                         bytenr = btrfs_node_blockptr(buf, i);
2905                         num_bytes = btrfs_level_size(root, level - 1);
2906                         ret = process_func(trans, root, bytenr, num_bytes,
2907                                            parent, ref_root, level - 1, 0,
2908                                            for_cow);
2909                         if (ret)
2910                                 goto fail;
2911                 }
2912         }
2913         return 0;
2914 fail:
2915         return ret;
2916 }
2917
2918 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2919                   struct extent_buffer *buf, int full_backref, int for_cow)
2920 {
2921         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2922 }
2923
2924 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2925                   struct extent_buffer *buf, int full_backref, int for_cow)
2926 {
2927         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2928 }
2929
2930 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2931                                  struct btrfs_root *root,
2932                                  struct btrfs_path *path,
2933                                  struct btrfs_block_group_cache *cache)
2934 {
2935         int ret;
2936         struct btrfs_root *extent_root = root->fs_info->extent_root;
2937         unsigned long bi;
2938         struct extent_buffer *leaf;
2939
2940         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2941         if (ret < 0)
2942                 goto fail;
2943         BUG_ON(ret); /* Corruption */
2944
2945         leaf = path->nodes[0];
2946         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2947         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2948         btrfs_mark_buffer_dirty(leaf);
2949         btrfs_release_path(path);
2950 fail:
2951         if (ret) {
2952                 btrfs_abort_transaction(trans, root, ret);
2953                 return ret;
2954         }
2955         return 0;
2956
2957 }
2958
2959 static struct btrfs_block_group_cache *
2960 next_block_group(struct btrfs_root *root,
2961                  struct btrfs_block_group_cache *cache)
2962 {
2963         struct rb_node *node;
2964         spin_lock(&root->fs_info->block_group_cache_lock);
2965         node = rb_next(&cache->cache_node);
2966         btrfs_put_block_group(cache);
2967         if (node) {
2968                 cache = rb_entry(node, struct btrfs_block_group_cache,
2969                                  cache_node);
2970                 btrfs_get_block_group(cache);
2971         } else
2972                 cache = NULL;
2973         spin_unlock(&root->fs_info->block_group_cache_lock);
2974         return cache;
2975 }
2976
2977 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2978                             struct btrfs_trans_handle *trans,
2979                             struct btrfs_path *path)
2980 {
2981         struct btrfs_root *root = block_group->fs_info->tree_root;
2982         struct inode *inode = NULL;
2983         u64 alloc_hint = 0;
2984         int dcs = BTRFS_DC_ERROR;
2985         int num_pages = 0;
2986         int retries = 0;
2987         int ret = 0;
2988
2989         /*
2990          * If this block group is smaller than 100 megs don't bother caching the
2991          * block group.
2992          */
2993         if (block_group->key.offset < (100 * 1024 * 1024)) {
2994                 spin_lock(&block_group->lock);
2995                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2996                 spin_unlock(&block_group->lock);
2997                 return 0;
2998         }
2999
3000 again:
3001         inode = lookup_free_space_inode(root, block_group, path);
3002         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3003                 ret = PTR_ERR(inode);
3004                 btrfs_release_path(path);
3005                 goto out;
3006         }
3007
3008         if (IS_ERR(inode)) {
3009                 BUG_ON(retries);
3010                 retries++;
3011
3012                 if (block_group->ro)
3013                         goto out_free;
3014
3015                 ret = create_free_space_inode(root, trans, block_group, path);
3016                 if (ret)
3017                         goto out_free;
3018                 goto again;
3019         }
3020
3021         /* We've already setup this transaction, go ahead and exit */
3022         if (block_group->cache_generation == trans->transid &&
3023             i_size_read(inode)) {
3024                 dcs = BTRFS_DC_SETUP;
3025                 goto out_put;
3026         }
3027
3028         /*
3029          * We want to set the generation to 0, that way if anything goes wrong
3030          * from here on out we know not to trust this cache when we load up next
3031          * time.
3032          */
3033         BTRFS_I(inode)->generation = 0;
3034         ret = btrfs_update_inode(trans, root, inode);
3035         WARN_ON(ret);
3036
3037         if (i_size_read(inode) > 0) {
3038                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3039                                                       inode);
3040                 if (ret)
3041                         goto out_put;
3042         }
3043
3044         spin_lock(&block_group->lock);
3045         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3046             !btrfs_test_opt(root, SPACE_CACHE)) {
3047                 /*
3048                  * don't bother trying to write stuff out _if_
3049                  * a) we're not cached,
3050                  * b) we're with nospace_cache mount option.
3051                  */
3052                 dcs = BTRFS_DC_WRITTEN;
3053                 spin_unlock(&block_group->lock);
3054                 goto out_put;
3055         }
3056         spin_unlock(&block_group->lock);
3057
3058         /*
3059          * Try to preallocate enough space based on how big the block group is.
3060          * Keep in mind this has to include any pinned space which could end up
3061          * taking up quite a bit since it's not folded into the other space
3062          * cache.
3063          */
3064         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3065         if (!num_pages)
3066                 num_pages = 1;
3067
3068         num_pages *= 16;
3069         num_pages *= PAGE_CACHE_SIZE;
3070
3071         ret = btrfs_check_data_free_space(inode, num_pages);
3072         if (ret)
3073                 goto out_put;
3074
3075         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3076                                               num_pages, num_pages,
3077                                               &alloc_hint);
3078         if (!ret)
3079                 dcs = BTRFS_DC_SETUP;
3080         btrfs_free_reserved_data_space(inode, num_pages);
3081
3082 out_put:
3083         iput(inode);
3084 out_free:
3085         btrfs_release_path(path);
3086 out:
3087         spin_lock(&block_group->lock);
3088         if (!ret && dcs == BTRFS_DC_SETUP)
3089                 block_group->cache_generation = trans->transid;
3090         block_group->disk_cache_state = dcs;
3091         spin_unlock(&block_group->lock);
3092
3093         return ret;
3094 }
3095
3096 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3097                                    struct btrfs_root *root)
3098 {
3099         struct btrfs_block_group_cache *cache;
3100         int err = 0;
3101         struct btrfs_path *path;
3102         u64 last = 0;
3103
3104         path = btrfs_alloc_path();
3105         if (!path)
3106                 return -ENOMEM;
3107
3108 again:
3109         while (1) {
3110                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3111                 while (cache) {
3112                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3113                                 break;
3114                         cache = next_block_group(root, cache);
3115                 }
3116                 if (!cache) {
3117                         if (last == 0)
3118                                 break;
3119                         last = 0;
3120                         continue;
3121                 }
3122                 err = cache_save_setup(cache, trans, path);
3123                 last = cache->key.objectid + cache->key.offset;
3124                 btrfs_put_block_group(cache);
3125         }
3126
3127         while (1) {
3128                 if (last == 0) {
3129                         err = btrfs_run_delayed_refs(trans, root,
3130                                                      (unsigned long)-1);
3131                         if (err) /* File system offline */
3132                                 goto out;
3133                 }
3134
3135                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3136                 while (cache) {
3137                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3138                                 btrfs_put_block_group(cache);
3139                                 goto again;
3140                         }
3141
3142                         if (cache->dirty)
3143                                 break;
3144                         cache = next_block_group(root, cache);
3145                 }
3146                 if (!cache) {
3147                         if (last == 0)
3148                                 break;
3149                         last = 0;
3150                         continue;
3151                 }
3152
3153                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3154                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3155                 cache->dirty = 0;
3156                 last = cache->key.objectid + cache->key.offset;
3157
3158                 err = write_one_cache_group(trans, root, path, cache);
3159                 if (err) /* File system offline */
3160                         goto out;
3161
3162                 btrfs_put_block_group(cache);
3163         }
3164
3165         while (1) {
3166                 /*
3167                  * I don't think this is needed since we're just marking our
3168                  * preallocated extent as written, but just in case it can't
3169                  * hurt.
3170                  */
3171                 if (last == 0) {
3172                         err = btrfs_run_delayed_refs(trans, root,
3173                                                      (unsigned long)-1);
3174                         if (err) /* File system offline */
3175                                 goto out;
3176                 }
3177
3178                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3179                 while (cache) {
3180                         /*
3181                          * Really this shouldn't happen, but it could if we
3182                          * couldn't write the entire preallocated extent and
3183                          * splitting the extent resulted in a new block.
3184                          */
3185                         if (cache->dirty) {
3186                                 btrfs_put_block_group(cache);
3187                                 goto again;
3188                         }
3189                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3190                                 break;
3191                         cache = next_block_group(root, cache);
3192                 }
3193                 if (!cache) {
3194                         if (last == 0)
3195                                 break;
3196                         last = 0;
3197                         continue;
3198                 }
3199
3200                 err = btrfs_write_out_cache(root, trans, cache, path);
3201
3202                 /*
3203                  * If we didn't have an error then the cache state is still
3204                  * NEED_WRITE, so we can set it to WRITTEN.
3205                  */
3206                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3207                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3208                 last = cache->key.objectid + cache->key.offset;
3209                 btrfs_put_block_group(cache);
3210         }
3211 out:
3212
3213         btrfs_free_path(path);
3214         return err;
3215 }
3216
3217 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3218 {
3219         struct btrfs_block_group_cache *block_group;
3220         int readonly = 0;
3221
3222         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3223         if (!block_group || block_group->ro)
3224                 readonly = 1;
3225         if (block_group)
3226                 btrfs_put_block_group(block_group);
3227         return readonly;
3228 }
3229
3230 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3231                              u64 total_bytes, u64 bytes_used,
3232                              struct btrfs_space_info **space_info)
3233 {
3234         struct btrfs_space_info *found;
3235         int i;
3236         int factor;
3237
3238         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3239                      BTRFS_BLOCK_GROUP_RAID10))
3240                 factor = 2;
3241         else
3242                 factor = 1;
3243
3244         found = __find_space_info(info, flags);
3245         if (found) {
3246                 spin_lock(&found->lock);
3247                 found->total_bytes += total_bytes;
3248                 found->disk_total += total_bytes * factor;
3249                 found->bytes_used += bytes_used;
3250                 found->disk_used += bytes_used * factor;
3251                 found->full = 0;
3252                 spin_unlock(&found->lock);
3253                 *space_info = found;
3254                 return 0;
3255         }
3256         found = kzalloc(sizeof(*found), GFP_NOFS);
3257         if (!found)
3258                 return -ENOMEM;
3259
3260         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3261                 INIT_LIST_HEAD(&found->block_groups[i]);
3262         init_rwsem(&found->groups_sem);
3263         spin_lock_init(&found->lock);
3264         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3265         found->total_bytes = total_bytes;
3266         found->disk_total = total_bytes * factor;
3267         found->bytes_used = bytes_used;
3268         found->disk_used = bytes_used * factor;
3269         found->bytes_pinned = 0;
3270         found->bytes_reserved = 0;
3271         found->bytes_readonly = 0;
3272         found->bytes_may_use = 0;
3273         found->full = 0;
3274         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3275         found->chunk_alloc = 0;
3276         found->flush = 0;
3277         init_waitqueue_head(&found->wait);
3278         *space_info = found;
3279         list_add_rcu(&found->list, &info->space_info);
3280         if (flags & BTRFS_BLOCK_GROUP_DATA)
3281                 info->data_sinfo = found;
3282         return 0;
3283 }
3284
3285 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3286 {
3287         u64 extra_flags = chunk_to_extended(flags) &
3288                                 BTRFS_EXTENDED_PROFILE_MASK;
3289
3290         write_seqlock(&fs_info->profiles_lock);
3291         if (flags & BTRFS_BLOCK_GROUP_DATA)
3292                 fs_info->avail_data_alloc_bits |= extra_flags;
3293         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3294                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3295         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3296                 fs_info->avail_system_alloc_bits |= extra_flags;
3297         write_sequnlock(&fs_info->profiles_lock);
3298 }
3299
3300 /*
3301  * returns target flags in extended format or 0 if restripe for this
3302  * chunk_type is not in progress
3303  *
3304  * should be called with either volume_mutex or balance_lock held
3305  */
3306 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3307 {
3308         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3309         u64 target = 0;
3310
3311         if (!bctl)
3312                 return 0;
3313
3314         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3315             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3316                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3317         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3318                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3319                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3320         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3321                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3322                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3323         }
3324
3325         return target;
3326 }
3327
3328 /*
3329  * @flags: available profiles in extended format (see ctree.h)
3330  *
3331  * Returns reduced profile in chunk format.  If profile changing is in
3332  * progress (either running or paused) picks the target profile (if it's
3333  * already available), otherwise falls back to plain reducing.
3334  */
3335 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3336 {
3337         /*
3338          * we add in the count of missing devices because we want
3339          * to make sure that any RAID levels on a degraded FS
3340          * continue to be honored.
3341          */
3342         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3343                 root->fs_info->fs_devices->missing_devices;
3344         u64 target;
3345         u64 tmp;
3346
3347         /*
3348          * see if restripe for this chunk_type is in progress, if so
3349          * try to reduce to the target profile
3350          */
3351         spin_lock(&root->fs_info->balance_lock);
3352         target = get_restripe_target(root->fs_info, flags);
3353         if (target) {
3354                 /* pick target profile only if it's already available */
3355                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3356                         spin_unlock(&root->fs_info->balance_lock);
3357                         return extended_to_chunk(target);
3358                 }
3359         }
3360         spin_unlock(&root->fs_info->balance_lock);
3361
3362         /* First, mask out the RAID levels which aren't possible */
3363         if (num_devices == 1)
3364                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3365                            BTRFS_BLOCK_GROUP_RAID5);
3366         if (num_devices < 3)
3367                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3368         if (num_devices < 4)
3369                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3370
3371         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3372                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3373                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3374         flags &= ~tmp;
3375
3376         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3377                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3378         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3379                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3380         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3381                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3382         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3383                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3384         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3385                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3386
3387         return extended_to_chunk(flags | tmp);
3388 }
3389
3390 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3391 {
3392         unsigned seq;
3393
3394         do {
3395                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3396
3397                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3398                         flags |= root->fs_info->avail_data_alloc_bits;
3399                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3400                         flags |= root->fs_info->avail_system_alloc_bits;
3401                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3402                         flags |= root->fs_info->avail_metadata_alloc_bits;
3403         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3404
3405         return btrfs_reduce_alloc_profile(root, flags);
3406 }
3407
3408 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3409 {
3410         u64 flags;
3411         u64 ret;
3412
3413         if (data)
3414                 flags = BTRFS_BLOCK_GROUP_DATA;
3415         else if (root == root->fs_info->chunk_root)
3416                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3417         else
3418                 flags = BTRFS_BLOCK_GROUP_METADATA;
3419
3420         ret = get_alloc_profile(root, flags);
3421         return ret;
3422 }
3423
3424 /*
3425  * This will check the space that the inode allocates from to make sure we have
3426  * enough space for bytes.
3427  */
3428 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3429 {
3430         struct btrfs_space_info *data_sinfo;
3431         struct btrfs_root *root = BTRFS_I(inode)->root;
3432         struct btrfs_fs_info *fs_info = root->fs_info;
3433         u64 used;
3434         int ret = 0, committed = 0, alloc_chunk = 1;
3435
3436         /* make sure bytes are sectorsize aligned */
3437         bytes = ALIGN(bytes, root->sectorsize);
3438
3439         if (root == root->fs_info->tree_root ||
3440             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3441                 alloc_chunk = 0;
3442                 committed = 1;
3443         }
3444
3445         data_sinfo = fs_info->data_sinfo;
3446         if (!data_sinfo)
3447                 goto alloc;
3448
3449 again:
3450         /* make sure we have enough space to handle the data first */
3451         spin_lock(&data_sinfo->lock);
3452         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3453                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3454                 data_sinfo->bytes_may_use;
3455
3456         if (used + bytes > data_sinfo->total_bytes) {
3457                 struct btrfs_trans_handle *trans;
3458
3459                 /*
3460                  * if we don't have enough free bytes in this space then we need
3461                  * to alloc a new chunk.
3462                  */
3463                 if (!data_sinfo->full && alloc_chunk) {
3464                         u64 alloc_target;
3465
3466                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3467                         spin_unlock(&data_sinfo->lock);
3468 alloc:
3469                         alloc_target = btrfs_get_alloc_profile(root, 1);
3470                         trans = btrfs_join_transaction(root);
3471                         if (IS_ERR(trans))
3472                                 return PTR_ERR(trans);
3473
3474                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3475                                              alloc_target,
3476                                              CHUNK_ALLOC_NO_FORCE);
3477                         btrfs_end_transaction(trans, root);
3478                         if (ret < 0) {
3479                                 if (ret != -ENOSPC)
3480                                         return ret;
3481                                 else
3482                                         goto commit_trans;
3483                         }
3484
3485                         if (!data_sinfo)
3486                                 data_sinfo = fs_info->data_sinfo;
3487
3488                         goto again;
3489                 }
3490
3491                 /*
3492                  * If we have less pinned bytes than we want to allocate then
3493                  * don't bother committing the transaction, it won't help us.
3494                  */
3495                 if (data_sinfo->bytes_pinned < bytes)
3496                         committed = 1;
3497                 spin_unlock(&data_sinfo->lock);
3498
3499                 /* commit the current transaction and try again */
3500 commit_trans:
3501                 if (!committed &&
3502                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3503                         committed = 1;
3504                         trans = btrfs_join_transaction(root);
3505                         if (IS_ERR(trans))
3506                                 return PTR_ERR(trans);
3507                         ret = btrfs_commit_transaction(trans, root);
3508                         if (ret)
3509                                 return ret;
3510                         goto again;
3511                 }
3512
3513                 return -ENOSPC;
3514         }
3515         data_sinfo->bytes_may_use += bytes;
3516         trace_btrfs_space_reservation(root->fs_info, "space_info",
3517                                       data_sinfo->flags, bytes, 1);
3518         spin_unlock(&data_sinfo->lock);
3519
3520         return 0;
3521 }
3522
3523 /*
3524  * Called if we need to clear a data reservation for this inode.
3525  */
3526 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3527 {
3528         struct btrfs_root *root = BTRFS_I(inode)->root;
3529         struct btrfs_space_info *data_sinfo;
3530
3531         /* make sure bytes are sectorsize aligned */
3532         bytes = ALIGN(bytes, root->sectorsize);
3533
3534         data_sinfo = root->fs_info->data_sinfo;
3535         spin_lock(&data_sinfo->lock);
3536         data_sinfo->bytes_may_use -= bytes;
3537         trace_btrfs_space_reservation(root->fs_info, "space_info",
3538                                       data_sinfo->flags, bytes, 0);
3539         spin_unlock(&data_sinfo->lock);
3540 }
3541
3542 static void force_metadata_allocation(struct btrfs_fs_info *info)
3543 {
3544         struct list_head *head = &info->space_info;
3545         struct btrfs_space_info *found;
3546
3547         rcu_read_lock();
3548         list_for_each_entry_rcu(found, head, list) {
3549                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3550                         found->force_alloc = CHUNK_ALLOC_FORCE;
3551         }
3552         rcu_read_unlock();
3553 }
3554
3555 static int should_alloc_chunk(struct btrfs_root *root,
3556                               struct btrfs_space_info *sinfo, int force)
3557 {
3558         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3559         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3560         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3561         u64 thresh;
3562
3563         if (force == CHUNK_ALLOC_FORCE)
3564                 return 1;
3565
3566         /*
3567          * We need to take into account the global rsv because for all intents
3568          * and purposes it's used space.  Don't worry about locking the
3569          * global_rsv, it doesn't change except when the transaction commits.
3570          */
3571         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3572                 num_allocated += global_rsv->size;
3573
3574         /*
3575          * in limited mode, we want to have some free space up to
3576          * about 1% of the FS size.
3577          */
3578         if (force == CHUNK_ALLOC_LIMITED) {
3579                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3580                 thresh = max_t(u64, 64 * 1024 * 1024,
3581                                div_factor_fine(thresh, 1));
3582
3583                 if (num_bytes - num_allocated < thresh)
3584                         return 1;
3585         }
3586
3587         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3588                 return 0;
3589         return 1;
3590 }
3591
3592 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3593 {
3594         u64 num_dev;
3595
3596         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3597                     BTRFS_BLOCK_GROUP_RAID0 |
3598                     BTRFS_BLOCK_GROUP_RAID5 |
3599                     BTRFS_BLOCK_GROUP_RAID6))
3600                 num_dev = root->fs_info->fs_devices->rw_devices;
3601         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3602                 num_dev = 2;
3603         else
3604                 num_dev = 1;    /* DUP or single */
3605
3606         /* metadata for updaing devices and chunk tree */
3607         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3608 }
3609
3610 static void check_system_chunk(struct btrfs_trans_handle *trans,
3611                                struct btrfs_root *root, u64 type)
3612 {
3613         struct btrfs_space_info *info;
3614         u64 left;
3615         u64 thresh;
3616
3617         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3618         spin_lock(&info->lock);
3619         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3620                 info->bytes_reserved - info->bytes_readonly;
3621         spin_unlock(&info->lock);
3622
3623         thresh = get_system_chunk_thresh(root, type);
3624         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3625                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3626                        left, thresh, type);
3627                 dump_space_info(info, 0, 0);
3628         }
3629
3630         if (left < thresh) {
3631                 u64 flags;
3632
3633                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3634                 btrfs_alloc_chunk(trans, root, flags);
3635         }
3636 }
3637
3638 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3639                           struct btrfs_root *extent_root, u64 flags, int force)
3640 {
3641         struct btrfs_space_info *space_info;
3642         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3643         int wait_for_alloc = 0;
3644         int ret = 0;
3645
3646         /* Don't re-enter if we're already allocating a chunk */
3647         if (trans->allocating_chunk)
3648                 return -ENOSPC;
3649
3650         space_info = __find_space_info(extent_root->fs_info, flags);
3651         if (!space_info) {
3652                 ret = update_space_info(extent_root->fs_info, flags,
3653                                         0, 0, &space_info);
3654                 BUG_ON(ret); /* -ENOMEM */
3655         }
3656         BUG_ON(!space_info); /* Logic error */
3657
3658 again:
3659         spin_lock(&space_info->lock);
3660         if (force < space_info->force_alloc)
3661                 force = space_info->force_alloc;
3662         if (space_info->full) {
3663                 spin_unlock(&space_info->lock);
3664                 return 0;
3665         }
3666
3667         if (!should_alloc_chunk(extent_root, space_info, force)) {
3668                 spin_unlock(&space_info->lock);
3669                 return 0;
3670         } else if (space_info->chunk_alloc) {
3671                 wait_for_alloc = 1;
3672         } else {
3673                 space_info->chunk_alloc = 1;
3674         }
3675
3676         spin_unlock(&space_info->lock);
3677
3678         mutex_lock(&fs_info->chunk_mutex);
3679
3680         /*
3681          * The chunk_mutex is held throughout the entirety of a chunk
3682          * allocation, so once we've acquired the chunk_mutex we know that the
3683          * other guy is done and we need to recheck and see if we should
3684          * allocate.
3685          */
3686         if (wait_for_alloc) {
3687                 mutex_unlock(&fs_info->chunk_mutex);
3688                 wait_for_alloc = 0;
3689                 goto again;
3690         }
3691
3692         trans->allocating_chunk = true;
3693
3694         /*
3695          * If we have mixed data/metadata chunks we want to make sure we keep
3696          * allocating mixed chunks instead of individual chunks.
3697          */
3698         if (btrfs_mixed_space_info(space_info))
3699                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3700
3701         /*
3702          * if we're doing a data chunk, go ahead and make sure that
3703          * we keep a reasonable number of metadata chunks allocated in the
3704          * FS as well.
3705          */
3706         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3707                 fs_info->data_chunk_allocations++;
3708                 if (!(fs_info->data_chunk_allocations %
3709                       fs_info->metadata_ratio))
3710                         force_metadata_allocation(fs_info);
3711         }
3712
3713         /*
3714          * Check if we have enough space in SYSTEM chunk because we may need
3715          * to update devices.
3716          */
3717         check_system_chunk(trans, extent_root, flags);
3718
3719         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3720         trans->allocating_chunk = false;
3721
3722         spin_lock(&space_info->lock);
3723         if (ret < 0 && ret != -ENOSPC)
3724                 goto out;
3725         if (ret)
3726                 space_info->full = 1;
3727         else
3728                 ret = 1;
3729
3730         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3731 out:
3732         space_info->chunk_alloc = 0;
3733         spin_unlock(&space_info->lock);
3734         mutex_unlock(&fs_info->chunk_mutex);
3735         return ret;
3736 }
3737
3738 static int can_overcommit(struct btrfs_root *root,
3739                           struct btrfs_space_info *space_info, u64 bytes,
3740                           enum btrfs_reserve_flush_enum flush)
3741 {
3742         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3743         u64 profile = btrfs_get_alloc_profile(root, 0);
3744         u64 rsv_size = 0;
3745         u64 avail;
3746         u64 used;
3747         u64 to_add;
3748
3749         used = space_info->bytes_used + space_info->bytes_reserved +
3750                 space_info->bytes_pinned + space_info->bytes_readonly;
3751
3752         spin_lock(&global_rsv->lock);
3753         rsv_size = global_rsv->size;
3754         spin_unlock(&global_rsv->lock);
3755
3756         /*
3757          * We only want to allow over committing if we have lots of actual space
3758          * free, but if we don't have enough space to handle the global reserve
3759          * space then we could end up having a real enospc problem when trying
3760          * to allocate a chunk or some other such important allocation.
3761          */
3762         rsv_size <<= 1;
3763         if (used + rsv_size >= space_info->total_bytes)
3764                 return 0;
3765
3766         used += space_info->bytes_may_use;
3767
3768         spin_lock(&root->fs_info->free_chunk_lock);
3769         avail = root->fs_info->free_chunk_space;
3770         spin_unlock(&root->fs_info->free_chunk_lock);
3771
3772         /*
3773          * If we have dup, raid1 or raid10 then only half of the free
3774          * space is actually useable.  For raid56, the space info used
3775          * doesn't include the parity drive, so we don't have to
3776          * change the math
3777          */
3778         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3779                        BTRFS_BLOCK_GROUP_RAID1 |
3780                        BTRFS_BLOCK_GROUP_RAID10))
3781                 avail >>= 1;
3782
3783         to_add = space_info->total_bytes;
3784
3785         /*
3786          * If we aren't flushing all things, let us overcommit up to
3787          * 1/2th of the space. If we can flush, don't let us overcommit
3788          * too much, let it overcommit up to 1/8 of the space.
3789          */
3790         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3791                 to_add >>= 3;
3792         else
3793                 to_add >>= 1;
3794
3795         /*
3796          * Limit the overcommit to the amount of free space we could possibly
3797          * allocate for chunks.
3798          */
3799         to_add = min(avail, to_add);
3800
3801         if (used + bytes < space_info->total_bytes + to_add)
3802                 return 1;
3803         return 0;
3804 }
3805
3806 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3807                                   unsigned long nr_pages)
3808 {
3809         struct super_block *sb = root->fs_info->sb;
3810         int started;
3811
3812         /* If we can not start writeback, just sync all the delalloc file. */
3813         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3814                                                       WB_REASON_FS_FREE_SPACE);
3815         if (!started) {
3816                 /*
3817                  * We needn't worry the filesystem going from r/w to r/o though
3818                  * we don't acquire ->s_umount mutex, because the filesystem
3819                  * should guarantee the delalloc inodes list be empty after
3820                  * the filesystem is readonly(all dirty pages are written to
3821                  * the disk).
3822                  */
3823                 btrfs_start_delalloc_inodes(root, 0);
3824                 btrfs_wait_ordered_extents(root, 0);
3825         }
3826 }
3827
3828 /*
3829  * shrink metadata reservation for delalloc
3830  */
3831 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3832                             bool wait_ordered)
3833 {
3834         struct btrfs_block_rsv *block_rsv;
3835         struct btrfs_space_info *space_info;
3836         struct btrfs_trans_handle *trans;
3837         u64 delalloc_bytes;
3838         u64 max_reclaim;
3839         long time_left;
3840         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3841         int loops = 0;
3842         enum btrfs_reserve_flush_enum flush;
3843
3844         trans = (struct btrfs_trans_handle *)current->journal_info;
3845         block_rsv = &root->fs_info->delalloc_block_rsv;
3846         space_info = block_rsv->space_info;
3847
3848         smp_mb();
3849         delalloc_bytes = percpu_counter_sum_positive(
3850                                                 &root->fs_info->delalloc_bytes);
3851         if (delalloc_bytes == 0) {
3852                 if (trans)
3853                         return;
3854                 btrfs_wait_ordered_extents(root, 0);
3855                 return;
3856         }
3857
3858         while (delalloc_bytes && loops < 3) {
3859                 max_reclaim = min(delalloc_bytes, to_reclaim);
3860                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3861                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3862                 /*
3863                  * We need to wait for the async pages to actually start before
3864                  * we do anything.
3865                  */
3866                 wait_event(root->fs_info->async_submit_wait,
3867                            !atomic_read(&root->fs_info->async_delalloc_pages));
3868
3869                 if (!trans)
3870                         flush = BTRFS_RESERVE_FLUSH_ALL;
3871                 else
3872                         flush = BTRFS_RESERVE_NO_FLUSH;
3873                 spin_lock(&space_info->lock);
3874                 if (can_overcommit(root, space_info, orig, flush)) {
3875                         spin_unlock(&space_info->lock);
3876                         break;
3877                 }
3878                 spin_unlock(&space_info->lock);
3879
3880                 loops++;
3881                 if (wait_ordered && !trans) {
3882                         btrfs_wait_ordered_extents(root, 0);
3883                 } else {
3884                         time_left = schedule_timeout_killable(1);
3885                         if (time_left)
3886                                 break;
3887                 }
3888                 smp_mb();
3889                 delalloc_bytes = percpu_counter_sum_positive(
3890                                                 &root->fs_info->delalloc_bytes);
3891         }
3892 }
3893
3894 /**
3895  * maybe_commit_transaction - possibly commit the transaction if its ok to
3896  * @root - the root we're allocating for
3897  * @bytes - the number of bytes we want to reserve
3898  * @force - force the commit
3899  *
3900  * This will check to make sure that committing the transaction will actually
3901  * get us somewhere and then commit the transaction if it does.  Otherwise it
3902  * will return -ENOSPC.
3903  */
3904 static int may_commit_transaction(struct btrfs_root *root,
3905                                   struct btrfs_space_info *space_info,
3906                                   u64 bytes, int force)
3907 {
3908         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3909         struct btrfs_trans_handle *trans;
3910
3911         trans = (struct btrfs_trans_handle *)current->journal_info;
3912         if (trans)
3913                 return -EAGAIN;
3914
3915         if (force)
3916                 goto commit;
3917
3918         /* See if there is enough pinned space to make this reservation */
3919         spin_lock(&space_info->lock);
3920         if (space_info->bytes_pinned >= bytes) {
3921                 spin_unlock(&space_info->lock);
3922                 goto commit;
3923         }
3924         spin_unlock(&space_info->lock);
3925
3926         /*
3927          * See if there is some space in the delayed insertion reservation for
3928          * this reservation.
3929          */
3930         if (space_info != delayed_rsv->space_info)
3931                 return -ENOSPC;
3932
3933         spin_lock(&space_info->lock);
3934         spin_lock(&delayed_rsv->lock);
3935         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3936                 spin_unlock(&delayed_rsv->lock);
3937                 spin_unlock(&space_info->lock);
3938                 return -ENOSPC;
3939         }
3940         spin_unlock(&delayed_rsv->lock);
3941         spin_unlock(&space_info->lock);
3942
3943 commit:
3944         trans = btrfs_join_transaction(root);
3945         if (IS_ERR(trans))
3946                 return -ENOSPC;
3947
3948         return btrfs_commit_transaction(trans, root);
3949 }
3950
3951 enum flush_state {
3952         FLUSH_DELAYED_ITEMS_NR  =       1,
3953         FLUSH_DELAYED_ITEMS     =       2,
3954         FLUSH_DELALLOC          =       3,
3955         FLUSH_DELALLOC_WAIT     =       4,
3956         ALLOC_CHUNK             =       5,
3957         COMMIT_TRANS            =       6,
3958 };
3959
3960 static int flush_space(struct btrfs_root *root,
3961                        struct btrfs_space_info *space_info, u64 num_bytes,
3962                        u64 orig_bytes, int state)
3963 {
3964         struct btrfs_trans_handle *trans;
3965         int nr;
3966         int ret = 0;
3967
3968         switch (state) {
3969         case FLUSH_DELAYED_ITEMS_NR:
3970         case FLUSH_DELAYED_ITEMS:
3971                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3972                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3973
3974                         nr = (int)div64_u64(num_bytes, bytes);
3975                         if (!nr)
3976                                 nr = 1;
3977                         nr *= 2;
3978                 } else {
3979                         nr = -1;
3980                 }
3981                 trans = btrfs_join_transaction(root);
3982                 if (IS_ERR(trans)) {
3983                         ret = PTR_ERR(trans);
3984                         break;
3985                 }
3986                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3987                 btrfs_end_transaction(trans, root);
3988                 break;
3989         case FLUSH_DELALLOC:
3990         case FLUSH_DELALLOC_WAIT:
3991                 shrink_delalloc(root, num_bytes, orig_bytes,
3992                                 state == FLUSH_DELALLOC_WAIT);
3993                 break;
3994         case ALLOC_CHUNK:
3995                 trans = btrfs_join_transaction(root);
3996                 if (IS_ERR(trans)) {
3997                         ret = PTR_ERR(trans);
3998                         break;
3999                 }
4000                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4001                                      btrfs_get_alloc_profile(root, 0),
4002                                      CHUNK_ALLOC_NO_FORCE);
4003                 btrfs_end_transaction(trans, root);
4004                 if (ret == -ENOSPC)
4005                         ret = 0;
4006                 break;
4007         case COMMIT_TRANS:
4008                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4009                 break;
4010         default:
4011                 ret = -ENOSPC;
4012                 break;
4013         }
4014
4015         return ret;
4016 }
4017 /**
4018  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4019  * @root - the root we're allocating for
4020  * @block_rsv - the block_rsv we're allocating for
4021  * @orig_bytes - the number of bytes we want
4022  * @flush - whether or not we can flush to make our reservation
4023  *
4024  * This will reserve orgi_bytes number of bytes from the space info associated
4025  * with the block_rsv.  If there is not enough space it will make an attempt to
4026  * flush out space to make room.  It will do this by flushing delalloc if
4027  * possible or committing the transaction.  If flush is 0 then no attempts to
4028  * regain reservations will be made and this will fail if there is not enough
4029  * space already.
4030  */
4031 static int reserve_metadata_bytes(struct btrfs_root *root,
4032                                   struct btrfs_block_rsv *block_rsv,
4033                                   u64 orig_bytes,
4034                                   enum btrfs_reserve_flush_enum flush)
4035 {
4036         struct btrfs_space_info *space_info = block_rsv->space_info;
4037         u64 used;
4038         u64 num_bytes = orig_bytes;
4039         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4040         int ret = 0;
4041         bool flushing = false;
4042
4043 again:
4044         ret = 0;
4045         spin_lock(&space_info->lock);
4046         /*
4047          * We only want to wait if somebody other than us is flushing and we
4048          * are actually allowed to flush all things.
4049          */
4050         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4051                space_info->flush) {
4052                 spin_unlock(&space_info->lock);
4053                 /*
4054                  * If we have a trans handle we can't wait because the flusher
4055                  * may have to commit the transaction, which would mean we would
4056                  * deadlock since we are waiting for the flusher to finish, but
4057                  * hold the current transaction open.
4058                  */
4059                 if (current->journal_info)
4060                         return -EAGAIN;
4061                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4062                 /* Must have been killed, return */
4063                 if (ret)
4064                         return -EINTR;
4065
4066                 spin_lock(&space_info->lock);
4067         }
4068
4069         ret = -ENOSPC;
4070         used = space_info->bytes_used + space_info->bytes_reserved +
4071                 space_info->bytes_pinned + space_info->bytes_readonly +
4072                 space_info->bytes_may_use;
4073
4074         /*
4075          * The idea here is that we've not already over-reserved the block group
4076          * then we can go ahead and save our reservation first and then start
4077          * flushing if we need to.  Otherwise if we've already overcommitted
4078          * lets start flushing stuff first and then come back and try to make
4079          * our reservation.
4080          */
4081         if (used <= space_info->total_bytes) {
4082                 if (used + orig_bytes <= space_info->total_bytes) {
4083                         space_info->bytes_may_use += orig_bytes;
4084                         trace_btrfs_space_reservation(root->fs_info,
4085                                 "space_info", space_info->flags, orig_bytes, 1);
4086                         ret = 0;
4087                 } else {
4088                         /*
4089                          * Ok set num_bytes to orig_bytes since we aren't
4090                          * overocmmitted, this way we only try and reclaim what
4091                          * we need.
4092                          */
4093                         num_bytes = orig_bytes;
4094                 }
4095         } else {
4096                 /*
4097                  * Ok we're over committed, set num_bytes to the overcommitted
4098                  * amount plus the amount of bytes that we need for this
4099                  * reservation.
4100                  */
4101                 num_bytes = used - space_info->total_bytes +
4102                         (orig_bytes * 2);
4103         }
4104
4105         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4106                 space_info->bytes_may_use += orig_bytes;
4107                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4108                                               space_info->flags, orig_bytes,
4109                                               1);
4110                 ret = 0;
4111         }
4112
4113         /*
4114          * Couldn't make our reservation, save our place so while we're trying
4115          * to reclaim space we can actually use it instead of somebody else
4116          * stealing it from us.
4117          *
4118          * We make the other tasks wait for the flush only when we can flush
4119          * all things.
4120          */
4121         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4122                 flushing = true;
4123                 space_info->flush = 1;
4124         }
4125
4126         spin_unlock(&space_info->lock);
4127
4128         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4129                 goto out;
4130
4131         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4132                           flush_state);
4133         flush_state++;
4134
4135         /*
4136          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4137          * would happen. So skip delalloc flush.
4138          */
4139         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4140             (flush_state == FLUSH_DELALLOC ||
4141              flush_state == FLUSH_DELALLOC_WAIT))
4142                 flush_state = ALLOC_CHUNK;
4143
4144         if (!ret)
4145                 goto again;
4146         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4147                  flush_state < COMMIT_TRANS)
4148                 goto again;
4149         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4150                  flush_state <= COMMIT_TRANS)
4151                 goto again;
4152
4153 out:
4154         if (ret == -ENOSPC &&
4155             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4156                 struct btrfs_block_rsv *global_rsv =
4157                         &root->fs_info->global_block_rsv;
4158
4159                 if (block_rsv != global_rsv &&
4160                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4161                         ret = 0;
4162         }
4163         if (flushing) {
4164                 spin_lock(&space_info->lock);
4165                 space_info->flush = 0;
4166                 wake_up_all(&space_info->wait);
4167                 spin_unlock(&space_info->lock);
4168         }
4169         return ret;
4170 }
4171
4172 static struct btrfs_block_rsv *get_block_rsv(
4173                                         const struct btrfs_trans_handle *trans,
4174                                         const struct btrfs_root *root)
4175 {
4176         struct btrfs_block_rsv *block_rsv = NULL;
4177
4178         if (root->ref_cows)
4179                 block_rsv = trans->block_rsv;
4180
4181         if (root == root->fs_info->csum_root && trans->adding_csums)
4182                 block_rsv = trans->block_rsv;
4183
4184         if (!block_rsv)
4185                 block_rsv = root->block_rsv;
4186
4187         if (!block_rsv)
4188                 block_rsv = &root->fs_info->empty_block_rsv;
4189
4190         return block_rsv;
4191 }
4192
4193 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4194                                u64 num_bytes)
4195 {
4196         int ret = -ENOSPC;
4197         spin_lock(&block_rsv->lock);
4198         if (block_rsv->reserved >= num_bytes) {
4199                 block_rsv->reserved -= num_bytes;
4200                 if (block_rsv->reserved < block_rsv->size)
4201                         block_rsv->full = 0;
4202                 ret = 0;
4203         }
4204         spin_unlock(&block_rsv->lock);
4205         return ret;
4206 }
4207
4208 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4209                                 u64 num_bytes, int update_size)
4210 {
4211         spin_lock(&block_rsv->lock);
4212         block_rsv->reserved += num_bytes;
4213         if (update_size)
4214                 block_rsv->size += num_bytes;
4215         else if (block_rsv->reserved >= block_rsv->size)
4216                 block_rsv->full = 1;
4217         spin_unlock(&block_rsv->lock);
4218 }
4219
4220 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4221                                     struct btrfs_block_rsv *block_rsv,
4222                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4223 {
4224         struct btrfs_space_info *space_info = block_rsv->space_info;
4225
4226         spin_lock(&block_rsv->lock);
4227         if (num_bytes == (u64)-1)
4228                 num_bytes = block_rsv->size;
4229         block_rsv->size -= num_bytes;
4230         if (block_rsv->reserved >= block_rsv->size) {
4231                 num_bytes = block_rsv->reserved - block_rsv->size;
4232                 block_rsv->reserved = block_rsv->size;
4233                 block_rsv->full = 1;
4234         } else {
4235                 num_bytes = 0;
4236         }
4237         spin_unlock(&block_rsv->lock);
4238
4239         if (num_bytes > 0) {
4240                 if (dest) {
4241                         spin_lock(&dest->lock);
4242                         if (!dest->full) {
4243                                 u64 bytes_to_add;
4244
4245                                 bytes_to_add = dest->size - dest->reserved;
4246                                 bytes_to_add = min(num_bytes, bytes_to_add);
4247                                 dest->reserved += bytes_to_add;
4248                                 if (dest->reserved >= dest->size)
4249                                         dest->full = 1;
4250                                 num_bytes -= bytes_to_add;
4251                         }
4252                         spin_unlock(&dest->lock);
4253                 }
4254                 if (num_bytes) {
4255                         spin_lock(&space_info->lock);
4256                         space_info->bytes_may_use -= num_bytes;
4257                         trace_btrfs_space_reservation(fs_info, "space_info",
4258                                         space_info->flags, num_bytes, 0);
4259                         space_info->reservation_progress++;
4260                         spin_unlock(&space_info->lock);
4261                 }
4262         }
4263 }
4264
4265 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4266                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4267 {
4268         int ret;
4269
4270         ret = block_rsv_use_bytes(src, num_bytes);
4271         if (ret)
4272                 return ret;
4273
4274         block_rsv_add_bytes(dst, num_bytes, 1);
4275         return 0;
4276 }
4277
4278 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4279 {
4280         memset(rsv, 0, sizeof(*rsv));
4281         spin_lock_init(&rsv->lock);
4282         rsv->type = type;
4283 }
4284
4285 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4286                                               unsigned short type)
4287 {
4288         struct btrfs_block_rsv *block_rsv;
4289         struct btrfs_fs_info *fs_info = root->fs_info;
4290
4291         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4292         if (!block_rsv)
4293                 return NULL;
4294
4295         btrfs_init_block_rsv(block_rsv, type);
4296         block_rsv->space_info = __find_space_info(fs_info,
4297                                                   BTRFS_BLOCK_GROUP_METADATA);
4298         return block_rsv;
4299 }
4300
4301 void btrfs_free_block_rsv(struct btrfs_root *root,
4302                           struct btrfs_block_rsv *rsv)
4303 {
4304         if (!rsv)
4305                 return;
4306         btrfs_block_rsv_release(root, rsv, (u64)-1);
4307         kfree(rsv);
4308 }
4309
4310 int btrfs_block_rsv_add(struct btrfs_root *root,
4311                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4312                         enum btrfs_reserve_flush_enum flush)
4313 {
4314         int ret;
4315
4316         if (num_bytes == 0)
4317                 return 0;
4318
4319         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4320         if (!ret) {
4321                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4322                 return 0;
4323         }
4324
4325         return ret;
4326 }
4327
4328 int btrfs_block_rsv_check(struct btrfs_root *root,
4329                           struct btrfs_block_rsv *block_rsv, int min_factor)
4330 {
4331         u64 num_bytes = 0;
4332         int ret = -ENOSPC;
4333
4334         if (!block_rsv)
4335                 return 0;
4336
4337         spin_lock(&block_rsv->lock);
4338         num_bytes = div_factor(block_rsv->size, min_factor);
4339         if (block_rsv->reserved >= num_bytes)
4340                 ret = 0;
4341         spin_unlock(&block_rsv->lock);
4342
4343         return ret;
4344 }
4345
4346 int btrfs_block_rsv_refill(struct btrfs_root *root,
4347                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4348                            enum btrfs_reserve_flush_enum flush)
4349 {
4350         u64 num_bytes = 0;
4351         int ret = -ENOSPC;
4352
4353         if (!block_rsv)
4354                 return 0;
4355
4356         spin_lock(&block_rsv->lock);
4357         num_bytes = min_reserved;
4358         if (block_rsv->reserved >= num_bytes)
4359                 ret = 0;
4360         else
4361                 num_bytes -= block_rsv->reserved;
4362         spin_unlock(&block_rsv->lock);
4363
4364         if (!ret)
4365                 return 0;
4366
4367         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4368         if (!ret) {
4369                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4370                 return 0;
4371         }
4372
4373         return ret;
4374 }
4375
4376 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4377                             struct btrfs_block_rsv *dst_rsv,
4378                             u64 num_bytes)
4379 {
4380         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4381 }
4382
4383 void btrfs_block_rsv_release(struct btrfs_root *root,
4384                              struct btrfs_block_rsv *block_rsv,
4385                              u64 num_bytes)
4386 {
4387         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4388         if (global_rsv->full || global_rsv == block_rsv ||
4389             block_rsv->space_info != global_rsv->space_info)
4390                 global_rsv = NULL;
4391         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4392                                 num_bytes);
4393 }
4394
4395 /*
4396  * helper to calculate size of global block reservation.
4397  * the desired value is sum of space used by extent tree,
4398  * checksum tree and root tree
4399  */
4400 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4401 {
4402         struct btrfs_space_info *sinfo;
4403         u64 num_bytes;
4404         u64 meta_used;
4405         u64 data_used;
4406         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4407
4408         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4409         spin_lock(&sinfo->lock);
4410         data_used = sinfo->bytes_used;
4411         spin_unlock(&sinfo->lock);
4412
4413         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4414         spin_lock(&sinfo->lock);
4415         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4416                 data_used = 0;
4417         meta_used = sinfo->bytes_used;
4418         spin_unlock(&sinfo->lock);
4419
4420         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4421                     csum_size * 2;
4422         num_bytes += div64_u64(data_used + meta_used, 50);
4423
4424         if (num_bytes * 3 > meta_used)
4425                 num_bytes = div64_u64(meta_used, 3);
4426
4427         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4428 }
4429
4430 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4431 {
4432         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4433         struct btrfs_space_info *sinfo = block_rsv->space_info;
4434         u64 num_bytes;
4435
4436         num_bytes = calc_global_metadata_size(fs_info);
4437
4438         spin_lock(&sinfo->lock);
4439         spin_lock(&block_rsv->lock);
4440
4441         block_rsv->size = num_bytes;
4442
4443         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4444                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4445                     sinfo->bytes_may_use;
4446
4447         if (sinfo->total_bytes > num_bytes) {
4448                 num_bytes = sinfo->total_bytes - num_bytes;
4449                 block_rsv->reserved += num_bytes;
4450                 sinfo->bytes_may_use += num_bytes;
4451                 trace_btrfs_space_reservation(fs_info, "space_info",
4452                                       sinfo->flags, num_bytes, 1);
4453         }
4454
4455         if (block_rsv->reserved >= block_rsv->size) {
4456                 num_bytes = block_rsv->reserved - block_rsv->size;
4457                 sinfo->bytes_may_use -= num_bytes;
4458                 trace_btrfs_space_reservation(fs_info, "space_info",
4459                                       sinfo->flags, num_bytes, 0);
4460                 sinfo->reservation_progress++;
4461                 block_rsv->reserved = block_rsv->size;
4462                 block_rsv->full = 1;
4463         }
4464
4465         spin_unlock(&block_rsv->lock);
4466         spin_unlock(&sinfo->lock);
4467 }
4468
4469 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4470 {
4471         struct btrfs_space_info *space_info;
4472
4473         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4474         fs_info->chunk_block_rsv.space_info = space_info;
4475
4476         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4477         fs_info->global_block_rsv.space_info = space_info;
4478         fs_info->delalloc_block_rsv.space_info = space_info;
4479         fs_info->trans_block_rsv.space_info = space_info;
4480         fs_info->empty_block_rsv.space_info = space_info;
4481         fs_info->delayed_block_rsv.space_info = space_info;
4482
4483         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4484         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4485         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4486         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4487         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4488
4489         update_global_block_rsv(fs_info);
4490 }
4491
4492 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4493 {
4494         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4495                                 (u64)-1);
4496         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4497         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4498         WARN_ON(fs_info->trans_block_rsv.size > 0);
4499         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4500         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4501         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4502         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4503         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4504 }
4505
4506 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4507                                   struct btrfs_root *root)
4508 {
4509         if (!trans->block_rsv)
4510                 return;
4511
4512         if (!trans->bytes_reserved)
4513                 return;
4514
4515         trace_btrfs_space_reservation(root->fs_info, "transaction",
4516                                       trans->transid, trans->bytes_reserved, 0);
4517         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4518         trans->bytes_reserved = 0;
4519 }
4520
4521 /* Can only return 0 or -ENOSPC */
4522 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4523                                   struct inode *inode)
4524 {
4525         struct btrfs_root *root = BTRFS_I(inode)->root;
4526         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4527         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4528
4529         /*
4530          * We need to hold space in order to delete our orphan item once we've
4531          * added it, so this takes the reservation so we can release it later
4532          * when we are truly done with the orphan item.
4533          */
4534         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4535         trace_btrfs_space_reservation(root->fs_info, "orphan",
4536                                       btrfs_ino(inode), num_bytes, 1);
4537         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4538 }
4539
4540 void btrfs_orphan_release_metadata(struct inode *inode)
4541 {
4542         struct btrfs_root *root = BTRFS_I(inode)->root;
4543         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4544         trace_btrfs_space_reservation(root->fs_info, "orphan",
4545                                       btrfs_ino(inode), num_bytes, 0);
4546         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4547 }
4548
4549 /*
4550  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4551  * root: the root of the parent directory
4552  * rsv: block reservation
4553  * items: the number of items that we need do reservation
4554  * qgroup_reserved: used to return the reserved size in qgroup
4555  *
4556  * This function is used to reserve the space for snapshot/subvolume
4557  * creation and deletion. Those operations are different with the
4558  * common file/directory operations, they change two fs/file trees
4559  * and root tree, the number of items that the qgroup reserves is
4560  * different with the free space reservation. So we can not use
4561  * the space reseravtion mechanism in start_transaction().
4562  */
4563 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4564                                      struct btrfs_block_rsv *rsv,
4565                                      int items,
4566                                      u64 *qgroup_reserved)
4567 {
4568         u64 num_bytes;
4569         int ret;
4570
4571         if (root->fs_info->quota_enabled) {
4572                 /* One for parent inode, two for dir entries */
4573                 num_bytes = 3 * root->leafsize;
4574                 ret = btrfs_qgroup_reserve(root, num_bytes);
4575                 if (ret)
4576                         return ret;
4577         } else {
4578                 num_bytes = 0;
4579         }
4580
4581         *qgroup_reserved = num_bytes;
4582
4583         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4584         rsv->space_info = __find_space_info(root->fs_info,
4585                                             BTRFS_BLOCK_GROUP_METADATA);
4586         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4587                                   BTRFS_RESERVE_FLUSH_ALL);
4588         if (ret) {
4589                 if (*qgroup_reserved)
4590                         btrfs_qgroup_free(root, *qgroup_reserved);
4591         }
4592
4593         return ret;
4594 }
4595
4596 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4597                                       struct btrfs_block_rsv *rsv,
4598                                       u64 qgroup_reserved)
4599 {
4600         btrfs_block_rsv_release(root, rsv, (u64)-1);
4601         if (qgroup_reserved)
4602                 btrfs_qgroup_free(root, qgroup_reserved);
4603 }
4604
4605 /**
4606  * drop_outstanding_extent - drop an outstanding extent
4607  * @inode: the inode we're dropping the extent for
4608  *
4609  * This is called when we are freeing up an outstanding extent, either called
4610  * after an error or after an extent is written.  This will return the number of
4611  * reserved extents that need to be freed.  This must be called with
4612  * BTRFS_I(inode)->lock held.
4613  */
4614 static unsigned drop_outstanding_extent(struct inode *inode)
4615 {
4616         unsigned drop_inode_space = 0;
4617         unsigned dropped_extents = 0;
4618
4619         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4620         BTRFS_I(inode)->outstanding_extents--;
4621
4622         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4623             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4624                                &BTRFS_I(inode)->runtime_flags))
4625                 drop_inode_space = 1;
4626
4627         /*
4628          * If we have more or the same amount of outsanding extents than we have
4629          * reserved then we need to leave the reserved extents count alone.
4630          */
4631         if (BTRFS_I(inode)->outstanding_extents >=
4632             BTRFS_I(inode)->reserved_extents)
4633                 return drop_inode_space;
4634
4635         dropped_extents = BTRFS_I(inode)->reserved_extents -
4636                 BTRFS_I(inode)->outstanding_extents;
4637         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4638         return dropped_extents + drop_inode_space;
4639 }
4640
4641 /**
4642  * calc_csum_metadata_size - return the amount of metada space that must be
4643  *      reserved/free'd for the given bytes.
4644  * @inode: the inode we're manipulating
4645  * @num_bytes: the number of bytes in question
4646  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4647  *
4648  * This adjusts the number of csum_bytes in the inode and then returns the
4649  * correct amount of metadata that must either be reserved or freed.  We
4650  * calculate how many checksums we can fit into one leaf and then divide the
4651  * number of bytes that will need to be checksumed by this value to figure out
4652  * how many checksums will be required.  If we are adding bytes then the number
4653  * may go up and we will return the number of additional bytes that must be
4654  * reserved.  If it is going down we will return the number of bytes that must
4655  * be freed.
4656  *
4657  * This must be called with BTRFS_I(inode)->lock held.
4658  */
4659 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4660                                    int reserve)
4661 {
4662         struct btrfs_root *root = BTRFS_I(inode)->root;
4663         u64 csum_size;
4664         int num_csums_per_leaf;
4665         int num_csums;
4666         int old_csums;
4667
4668         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4669             BTRFS_I(inode)->csum_bytes == 0)
4670                 return 0;
4671
4672         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4673         if (reserve)
4674                 BTRFS_I(inode)->csum_bytes += num_bytes;
4675         else
4676                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4677         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4678         num_csums_per_leaf = (int)div64_u64(csum_size,
4679                                             sizeof(struct btrfs_csum_item) +
4680                                             sizeof(struct btrfs_disk_key));
4681         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4682         num_csums = num_csums + num_csums_per_leaf - 1;
4683         num_csums = num_csums / num_csums_per_leaf;
4684
4685         old_csums = old_csums + num_csums_per_leaf - 1;
4686         old_csums = old_csums / num_csums_per_leaf;
4687
4688         /* No change, no need to reserve more */
4689         if (old_csums == num_csums)
4690                 return 0;
4691
4692         if (reserve)
4693                 return btrfs_calc_trans_metadata_size(root,
4694                                                       num_csums - old_csums);
4695
4696         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4697 }
4698
4699 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4700 {
4701         struct btrfs_root *root = BTRFS_I(inode)->root;
4702         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4703         u64 to_reserve = 0;
4704         u64 csum_bytes;
4705         unsigned nr_extents = 0;
4706         int extra_reserve = 0;
4707         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4708         int ret = 0;
4709         bool delalloc_lock = true;
4710         u64 to_free = 0;
4711         unsigned dropped;
4712
4713         /* If we are a free space inode we need to not flush since we will be in
4714          * the middle of a transaction commit.  We also don't need the delalloc
4715          * mutex since we won't race with anybody.  We need this mostly to make
4716          * lockdep shut its filthy mouth.
4717          */
4718         if (btrfs_is_free_space_inode(inode)) {
4719                 flush = BTRFS_RESERVE_NO_FLUSH;
4720                 delalloc_lock = false;
4721         }
4722
4723         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4724             btrfs_transaction_in_commit(root->fs_info))
4725                 schedule_timeout(1);
4726
4727         if (delalloc_lock)
4728                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4729
4730         num_bytes = ALIGN(num_bytes, root->sectorsize);
4731
4732         spin_lock(&BTRFS_I(inode)->lock);
4733         BTRFS_I(inode)->outstanding_extents++;
4734
4735         if (BTRFS_I(inode)->outstanding_extents >
4736             BTRFS_I(inode)->reserved_extents)
4737                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4738                         BTRFS_I(inode)->reserved_extents;
4739
4740         /*
4741          * Add an item to reserve for updating the inode when we complete the
4742          * delalloc io.
4743          */
4744         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4745                       &BTRFS_I(inode)->runtime_flags)) {
4746                 nr_extents++;
4747                 extra_reserve = 1;
4748         }
4749
4750         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4751         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4752         csum_bytes = BTRFS_I(inode)->csum_bytes;
4753         spin_unlock(&BTRFS_I(inode)->lock);
4754
4755         if (root->fs_info->quota_enabled) {
4756                 ret = btrfs_qgroup_reserve(root, num_bytes +
4757                                            nr_extents * root->leafsize);
4758                 if (ret)
4759                         goto out_fail;
4760         }
4761
4762         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4763         if (unlikely(ret)) {
4764                 if (root->fs_info->quota_enabled)
4765                         btrfs_qgroup_free(root, num_bytes +
4766                                                 nr_extents * root->leafsize);
4767                 goto out_fail;
4768         }
4769
4770         spin_lock(&BTRFS_I(inode)->lock);
4771         if (extra_reserve) {
4772                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4773                         &BTRFS_I(inode)->runtime_flags);
4774                 nr_extents--;
4775         }
4776         BTRFS_I(inode)->reserved_extents += nr_extents;
4777         spin_unlock(&BTRFS_I(inode)->lock);
4778
4779         if (delalloc_lock)
4780                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4781
4782         if (to_reserve)
4783                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4784                                               btrfs_ino(inode), to_reserve, 1);
4785         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4786
4787         return 0;
4788
4789 out_fail:
4790         spin_lock(&BTRFS_I(inode)->lock);
4791         dropped = drop_outstanding_extent(inode);
4792         /*
4793          * If the inodes csum_bytes is the same as the original
4794          * csum_bytes then we know we haven't raced with any free()ers
4795          * so we can just reduce our inodes csum bytes and carry on.
4796          * Otherwise we have to do the normal free thing to account for
4797          * the case that the free side didn't free up its reserve
4798          * because of this outstanding reservation.
4799          */
4800         if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4801                 calc_csum_metadata_size(inode, num_bytes, 0);
4802         else
4803                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4804         spin_unlock(&BTRFS_I(inode)->lock);
4805         if (dropped)
4806                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4807
4808         if (to_free) {
4809                 btrfs_block_rsv_release(root, block_rsv, to_free);
4810                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4811                                               btrfs_ino(inode), to_free, 0);
4812         }
4813         if (delalloc_lock)
4814                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4815         return ret;
4816 }
4817
4818 /**
4819  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4820  * @inode: the inode to release the reservation for
4821  * @num_bytes: the number of bytes we're releasing
4822  *
4823  * This will release the metadata reservation for an inode.  This can be called
4824  * once we complete IO for a given set of bytes to release their metadata
4825  * reservations.
4826  */
4827 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4828 {
4829         struct btrfs_root *root = BTRFS_I(inode)->root;
4830         u64 to_free = 0;
4831         unsigned dropped;
4832
4833         num_bytes = ALIGN(num_bytes, root->sectorsize);
4834         spin_lock(&BTRFS_I(inode)->lock);
4835         dropped = drop_outstanding_extent(inode);
4836
4837         if (num_bytes)
4838                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4839         spin_unlock(&BTRFS_I(inode)->lock);
4840         if (dropped > 0)
4841                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4842
4843         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4844                                       btrfs_ino(inode), to_free, 0);
4845         if (root->fs_info->quota_enabled) {
4846                 btrfs_qgroup_free(root, num_bytes +
4847                                         dropped * root->leafsize);
4848         }
4849
4850         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4851                                 to_free);
4852 }
4853
4854 /**
4855  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4856  * @inode: inode we're writing to
4857  * @num_bytes: the number of bytes we want to allocate
4858  *
4859  * This will do the following things
4860  *
4861  * o reserve space in the data space info for num_bytes
4862  * o reserve space in the metadata space info based on number of outstanding
4863  *   extents and how much csums will be needed
4864  * o add to the inodes ->delalloc_bytes
4865  * o add it to the fs_info's delalloc inodes list.
4866  *
4867  * This will return 0 for success and -ENOSPC if there is no space left.
4868  */
4869 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4870 {
4871         int ret;
4872
4873         ret = btrfs_check_data_free_space(inode, num_bytes);
4874         if (ret)
4875                 return ret;
4876
4877         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4878         if (ret) {
4879                 btrfs_free_reserved_data_space(inode, num_bytes);
4880                 return ret;
4881         }
4882
4883         return 0;
4884 }
4885
4886 /**
4887  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4888  * @inode: inode we're releasing space for
4889  * @num_bytes: the number of bytes we want to free up
4890  *
4891  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4892  * called in the case that we don't need the metadata AND data reservations
4893  * anymore.  So if there is an error or we insert an inline extent.
4894  *
4895  * This function will release the metadata space that was not used and will
4896  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4897  * list if there are no delalloc bytes left.
4898  */
4899 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4900 {
4901         btrfs_delalloc_release_metadata(inode, num_bytes);
4902         btrfs_free_reserved_data_space(inode, num_bytes);
4903 }
4904
4905 static int update_block_group(struct btrfs_root *root,
4906                               u64 bytenr, u64 num_bytes, int alloc)
4907 {
4908         struct btrfs_block_group_cache *cache = NULL;
4909         struct btrfs_fs_info *info = root->fs_info;
4910         u64 total = num_bytes;
4911         u64 old_val;
4912         u64 byte_in_group;
4913         int factor;
4914
4915         /* block accounting for super block */
4916         spin_lock(&info->delalloc_lock);
4917         old_val = btrfs_super_bytes_used(info->super_copy);
4918         if (alloc)
4919                 old_val += num_bytes;
4920         else
4921                 old_val -= num_bytes;
4922         btrfs_set_super_bytes_used(info->super_copy, old_val);
4923         spin_unlock(&info->delalloc_lock);
4924
4925         while (total) {
4926                 cache = btrfs_lookup_block_group(info, bytenr);
4927                 if (!cache)
4928                         return -ENOENT;
4929                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4930                                     BTRFS_BLOCK_GROUP_RAID1 |
4931                                     BTRFS_BLOCK_GROUP_RAID10))
4932                         factor = 2;
4933                 else
4934                         factor = 1;
4935                 /*
4936                  * If this block group has free space cache written out, we
4937                  * need to make sure to load it if we are removing space.  This
4938                  * is because we need the unpinning stage to actually add the
4939                  * space back to the block group, otherwise we will leak space.
4940                  */
4941                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4942                         cache_block_group(cache, 1);
4943
4944                 byte_in_group = bytenr - cache->key.objectid;
4945                 WARN_ON(byte_in_group > cache->key.offset);
4946
4947                 spin_lock(&cache->space_info->lock);
4948                 spin_lock(&cache->lock);
4949
4950                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4951                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4952                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4953
4954                 cache->dirty = 1;
4955                 old_val = btrfs_block_group_used(&cache->item);
4956                 num_bytes = min(total, cache->key.offset - byte_in_group);
4957                 if (alloc) {
4958                         old_val += num_bytes;
4959                         btrfs_set_block_group_used(&cache->item, old_val);
4960                         cache->reserved -= num_bytes;
4961                         cache->space_info->bytes_reserved -= num_bytes;
4962                         cache->space_info->bytes_used += num_bytes;
4963                         cache->space_info->disk_used += num_bytes * factor;
4964                         spin_unlock(&cache->lock);
4965                         spin_unlock(&cache->space_info->lock);
4966                 } else {
4967                         old_val -= num_bytes;
4968                         btrfs_set_block_group_used(&cache->item, old_val);
4969                         cache->pinned += num_bytes;
4970                         cache->space_info->bytes_pinned += num_bytes;
4971                         cache->space_info->bytes_used -= num_bytes;
4972                         cache->space_info->disk_used -= num_bytes * factor;
4973                         spin_unlock(&cache->lock);
4974                         spin_unlock(&cache->space_info->lock);
4975
4976                         set_extent_dirty(info->pinned_extents,
4977                                          bytenr, bytenr + num_bytes - 1,
4978                                          GFP_NOFS | __GFP_NOFAIL);
4979                 }
4980                 btrfs_put_block_group(cache);
4981                 total -= num_bytes;
4982                 bytenr += num_bytes;
4983         }
4984         return 0;
4985 }
4986
4987 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4988 {
4989         struct btrfs_block_group_cache *cache;
4990         u64 bytenr;
4991
4992         spin_lock(&root->fs_info->block_group_cache_lock);
4993         bytenr = root->fs_info->first_logical_byte;
4994         spin_unlock(&root->fs_info->block_group_cache_lock);
4995
4996         if (bytenr < (u64)-1)
4997                 return bytenr;
4998
4999         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5000         if (!cache)
5001                 return 0;
5002
5003         bytenr = cache->key.objectid;
5004         btrfs_put_block_group(cache);
5005
5006         return bytenr;
5007 }
5008
5009 static int pin_down_extent(struct btrfs_root *root,
5010                            struct btrfs_block_group_cache *cache,
5011                            u64 bytenr, u64 num_bytes, int reserved)
5012 {
5013         spin_lock(&cache->space_info->lock);
5014         spin_lock(&cache->lock);
5015         cache->pinned += num_bytes;
5016         cache->space_info->bytes_pinned += num_bytes;
5017         if (reserved) {
5018                 cache->reserved -= num_bytes;
5019                 cache->space_info->bytes_reserved -= num_bytes;
5020         }
5021         spin_unlock(&cache->lock);
5022         spin_unlock(&cache->space_info->lock);
5023
5024         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5025                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5026         return 0;
5027 }
5028
5029 /*
5030  * this function must be called within transaction
5031  */
5032 int btrfs_pin_extent(struct btrfs_root *root,
5033                      u64 bytenr, u64 num_bytes, int reserved)
5034 {
5035         struct btrfs_block_group_cache *cache;
5036
5037         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5038         BUG_ON(!cache); /* Logic error */
5039
5040         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5041
5042         btrfs_put_block_group(cache);
5043         return 0;
5044 }
5045
5046 /*
5047  * this function must be called within transaction
5048  */
5049 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5050                                     u64 bytenr, u64 num_bytes)
5051 {
5052         struct btrfs_block_group_cache *cache;
5053
5054         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5055         BUG_ON(!cache); /* Logic error */
5056
5057         /*
5058          * pull in the free space cache (if any) so that our pin
5059          * removes the free space from the cache.  We have load_only set
5060          * to one because the slow code to read in the free extents does check
5061          * the pinned extents.
5062          */
5063         cache_block_group(cache, 1);
5064
5065         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5066
5067         /* remove us from the free space cache (if we're there at all) */
5068         btrfs_remove_free_space(cache, bytenr, num_bytes);
5069         btrfs_put_block_group(cache);
5070         return 0;
5071 }
5072
5073 /**
5074  * btrfs_update_reserved_bytes - update the block_group and space info counters
5075  * @cache:      The cache we are manipulating
5076  * @num_bytes:  The number of bytes in question
5077  * @reserve:    One of the reservation enums
5078  *
5079  * This is called by the allocator when it reserves space, or by somebody who is
5080  * freeing space that was never actually used on disk.  For example if you
5081  * reserve some space for a new leaf in transaction A and before transaction A
5082  * commits you free that leaf, you call this with reserve set to 0 in order to
5083  * clear the reservation.
5084  *
5085  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5086  * ENOSPC accounting.  For data we handle the reservation through clearing the
5087  * delalloc bits in the io_tree.  We have to do this since we could end up
5088  * allocating less disk space for the amount of data we have reserved in the
5089  * case of compression.
5090  *
5091  * If this is a reservation and the block group has become read only we cannot
5092  * make the reservation and return -EAGAIN, otherwise this function always
5093  * succeeds.
5094  */
5095 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5096                                        u64 num_bytes, int reserve)
5097 {
5098         struct btrfs_space_info *space_info = cache->space_info;
5099         int ret = 0;
5100
5101         spin_lock(&space_info->lock);
5102         spin_lock(&cache->lock);
5103         if (reserve != RESERVE_FREE) {
5104                 if (cache->ro) {
5105                         ret = -EAGAIN;
5106                 } else {
5107                         cache->reserved += num_bytes;
5108                         space_info->bytes_reserved += num_bytes;
5109                         if (reserve == RESERVE_ALLOC) {
5110                                 trace_btrfs_space_reservation(cache->fs_info,
5111                                                 "space_info", space_info->flags,
5112                                                 num_bytes, 0);
5113                                 space_info->bytes_may_use -= num_bytes;
5114                         }
5115                 }
5116         } else {
5117                 if (cache->ro)
5118                         space_info->bytes_readonly += num_bytes;
5119                 cache->reserved -= num_bytes;
5120                 space_info->bytes_reserved -= num_bytes;
5121                 space_info->reservation_progress++;
5122         }
5123         spin_unlock(&cache->lock);
5124         spin_unlock(&space_info->lock);
5125         return ret;
5126 }
5127
5128 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5129                                 struct btrfs_root *root)
5130 {
5131         struct btrfs_fs_info *fs_info = root->fs_info;
5132         struct btrfs_caching_control *next;
5133         struct btrfs_caching_control *caching_ctl;
5134         struct btrfs_block_group_cache *cache;
5135
5136         down_write(&fs_info->extent_commit_sem);
5137
5138         list_for_each_entry_safe(caching_ctl, next,
5139                                  &fs_info->caching_block_groups, list) {
5140                 cache = caching_ctl->block_group;
5141                 if (block_group_cache_done(cache)) {
5142                         cache->last_byte_to_unpin = (u64)-1;
5143                         list_del_init(&caching_ctl->list);
5144                         put_caching_control(caching_ctl);
5145                 } else {
5146                         cache->last_byte_to_unpin = caching_ctl->progress;
5147                 }
5148         }
5149
5150         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5151                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5152         else
5153                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5154
5155         up_write(&fs_info->extent_commit_sem);
5156
5157         update_global_block_rsv(fs_info);
5158 }
5159
5160 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5161 {
5162         struct btrfs_fs_info *fs_info = root->fs_info;
5163         struct btrfs_block_group_cache *cache = NULL;
5164         struct btrfs_space_info *space_info;
5165         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5166         u64 len;
5167         bool readonly;
5168
5169         while (start <= end) {
5170                 readonly = false;
5171                 if (!cache ||
5172                     start >= cache->key.objectid + cache->key.offset) {
5173                         if (cache)
5174                                 btrfs_put_block_group(cache);
5175                         cache = btrfs_lookup_block_group(fs_info, start);
5176                         BUG_ON(!cache); /* Logic error */
5177                 }
5178
5179                 len = cache->key.objectid + cache->key.offset - start;
5180                 len = min(len, end + 1 - start);
5181
5182                 if (start < cache->last_byte_to_unpin) {
5183                         len = min(len, cache->last_byte_to_unpin - start);
5184                         btrfs_add_free_space(cache, start, len);
5185                 }
5186
5187                 start += len;
5188                 space_info = cache->space_info;
5189
5190                 spin_lock(&space_info->lock);
5191                 spin_lock(&cache->lock);
5192                 cache->pinned -= len;
5193                 space_info->bytes_pinned -= len;
5194                 if (cache->ro) {
5195                         space_info->bytes_readonly += len;
5196                         readonly = true;
5197                 }
5198                 spin_unlock(&cache->lock);
5199                 if (!readonly && global_rsv->space_info == space_info) {
5200                         spin_lock(&global_rsv->lock);
5201                         if (!global_rsv->full) {
5202                                 len = min(len, global_rsv->size -
5203                                           global_rsv->reserved);
5204                                 global_rsv->reserved += len;
5205                                 space_info->bytes_may_use += len;
5206                                 if (global_rsv->reserved >= global_rsv->size)
5207                                         global_rsv->full = 1;
5208                         }
5209                         spin_unlock(&global_rsv->lock);
5210                 }
5211                 spin_unlock(&space_info->lock);
5212         }
5213
5214         if (cache)
5215                 btrfs_put_block_group(cache);
5216         return 0;
5217 }
5218
5219 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5220                                struct btrfs_root *root)
5221 {
5222         struct btrfs_fs_info *fs_info = root->fs_info;
5223         struct extent_io_tree *unpin;
5224         u64 start;
5225         u64 end;
5226         int ret;
5227
5228         if (trans->aborted)
5229                 return 0;
5230
5231         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5232                 unpin = &fs_info->freed_extents[1];
5233         else
5234                 unpin = &fs_info->freed_extents[0];
5235
5236         while (1) {
5237                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5238                                             EXTENT_DIRTY, NULL);
5239                 if (ret)
5240                         break;
5241
5242                 if (btrfs_test_opt(root, DISCARD))
5243                         ret = btrfs_discard_extent(root, start,
5244                                                    end + 1 - start, NULL);
5245
5246                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5247                 unpin_extent_range(root, start, end);
5248                 cond_resched();
5249         }
5250
5251         return 0;
5252 }
5253
5254 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5255                                 struct btrfs_root *root,
5256                                 u64 bytenr, u64 num_bytes, u64 parent,
5257                                 u64 root_objectid, u64 owner_objectid,
5258                                 u64 owner_offset, int refs_to_drop,
5259                                 struct btrfs_delayed_extent_op *extent_op)
5260 {
5261         struct btrfs_key key;
5262         struct btrfs_path *path;
5263         struct btrfs_fs_info *info = root->fs_info;
5264         struct btrfs_root *extent_root = info->extent_root;
5265         struct extent_buffer *leaf;
5266         struct btrfs_extent_item *ei;
5267         struct btrfs_extent_inline_ref *iref;
5268         int ret;
5269         int is_data;
5270         int extent_slot = 0;
5271         int found_extent = 0;
5272         int num_to_del = 1;
5273         u32 item_size;
5274         u64 refs;
5275
5276         path = btrfs_alloc_path();
5277         if (!path)
5278                 return -ENOMEM;
5279
5280         path->reada = 1;
5281         path->leave_spinning = 1;
5282
5283         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5284         BUG_ON(!is_data && refs_to_drop != 1);
5285
5286         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5287                                     bytenr, num_bytes, parent,
5288                                     root_objectid, owner_objectid,
5289                                     owner_offset);
5290         if (ret == 0) {
5291                 extent_slot = path->slots[0];
5292                 while (extent_slot >= 0) {
5293                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5294                                               extent_slot);
5295                         if (key.objectid != bytenr)
5296                                 break;
5297                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5298                             key.offset == num_bytes) {
5299                                 found_extent = 1;
5300                                 break;
5301                         }
5302                         if (path->slots[0] - extent_slot > 5)
5303                                 break;
5304                         extent_slot--;
5305                 }
5306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5307                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5308                 if (found_extent && item_size < sizeof(*ei))
5309                         found_extent = 0;
5310 #endif
5311                 if (!found_extent) {
5312                         BUG_ON(iref);
5313                         ret = remove_extent_backref(trans, extent_root, path,
5314                                                     NULL, refs_to_drop,
5315                                                     is_data);
5316                         if (ret) {
5317                                 btrfs_abort_transaction(trans, extent_root, ret);
5318                                 goto out;
5319                         }
5320                         btrfs_release_path(path);
5321                         path->leave_spinning = 1;
5322
5323                         key.objectid = bytenr;
5324                         key.type = BTRFS_EXTENT_ITEM_KEY;
5325                         key.offset = num_bytes;
5326
5327                         ret = btrfs_search_slot(trans, extent_root,
5328                                                 &key, path, -1, 1);
5329                         if (ret) {
5330                                 printk(KERN_ERR "umm, got %d back from search"
5331                                        ", was looking for %llu\n", ret,
5332                                        (unsigned long long)bytenr);
5333                                 if (ret > 0)
5334                                         btrfs_print_leaf(extent_root,
5335                                                          path->nodes[0]);
5336                         }
5337                         if (ret < 0) {
5338                                 btrfs_abort_transaction(trans, extent_root, ret);
5339                                 goto out;
5340                         }
5341                         extent_slot = path->slots[0];
5342                 }
5343         } else if (ret == -ENOENT) {
5344                 btrfs_print_leaf(extent_root, path->nodes[0]);
5345                 WARN_ON(1);
5346                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5347                        "parent %llu root %llu  owner %llu offset %llu\n",
5348                        (unsigned long long)bytenr,
5349                        (unsigned long long)parent,
5350                        (unsigned long long)root_objectid,
5351                        (unsigned long long)owner_objectid,
5352                        (unsigned long long)owner_offset);
5353         } else {
5354                 btrfs_abort_transaction(trans, extent_root, ret);
5355                 goto out;
5356         }
5357
5358         leaf = path->nodes[0];
5359         item_size = btrfs_item_size_nr(leaf, extent_slot);
5360 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5361         if (item_size < sizeof(*ei)) {
5362                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5363                 ret = convert_extent_item_v0(trans, extent_root, path,
5364                                              owner_objectid, 0);
5365                 if (ret < 0) {
5366                         btrfs_abort_transaction(trans, extent_root, ret);
5367                         goto out;
5368                 }
5369
5370                 btrfs_release_path(path);
5371                 path->leave_spinning = 1;
5372
5373                 key.objectid = bytenr;
5374                 key.type = BTRFS_EXTENT_ITEM_KEY;
5375                 key.offset = num_bytes;
5376
5377                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5378                                         -1, 1);
5379                 if (ret) {
5380                         printk(KERN_ERR "umm, got %d back from search"
5381                                ", was looking for %llu\n", ret,
5382                                (unsigned long long)bytenr);
5383                         btrfs_print_leaf(extent_root, path->nodes[0]);
5384                 }
5385                 if (ret < 0) {
5386                         btrfs_abort_transaction(trans, extent_root, ret);
5387                         goto out;
5388                 }
5389
5390                 extent_slot = path->slots[0];
5391                 leaf = path->nodes[0];
5392                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5393         }
5394 #endif
5395         BUG_ON(item_size < sizeof(*ei));
5396         ei = btrfs_item_ptr(leaf, extent_slot,
5397                             struct btrfs_extent_item);
5398         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5399                 struct btrfs_tree_block_info *bi;
5400                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5401                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5402                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5403         }
5404
5405         refs = btrfs_extent_refs(leaf, ei);
5406         BUG_ON(refs < refs_to_drop);
5407         refs -= refs_to_drop;
5408
5409         if (refs > 0) {
5410                 if (extent_op)
5411                         __run_delayed_extent_op(extent_op, leaf, ei);
5412                 /*
5413                  * In the case of inline back ref, reference count will
5414                  * be updated by remove_extent_backref
5415                  */
5416                 if (iref) {
5417                         BUG_ON(!found_extent);
5418                 } else {
5419                         btrfs_set_extent_refs(leaf, ei, refs);
5420                         btrfs_mark_buffer_dirty(leaf);
5421                 }
5422                 if (found_extent) {
5423                         ret = remove_extent_backref(trans, extent_root, path,
5424                                                     iref, refs_to_drop,
5425                                                     is_data);
5426                         if (ret) {
5427                                 btrfs_abort_transaction(trans, extent_root, ret);
5428                                 goto out;
5429                         }
5430                 }
5431         } else {
5432                 if (found_extent) {
5433                         BUG_ON(is_data && refs_to_drop !=
5434                                extent_data_ref_count(root, path, iref));
5435                         if (iref) {
5436                                 BUG_ON(path->slots[0] != extent_slot);
5437                         } else {
5438                                 BUG_ON(path->slots[0] != extent_slot + 1);
5439                                 path->slots[0] = extent_slot;
5440                                 num_to_del = 2;
5441                         }
5442                 }
5443
5444                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5445                                       num_to_del);
5446                 if (ret) {
5447                         btrfs_abort_transaction(trans, extent_root, ret);
5448                         goto out;
5449                 }
5450                 btrfs_release_path(path);
5451
5452                 if (is_data) {
5453                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5454                         if (ret) {
5455                                 btrfs_abort_transaction(trans, extent_root, ret);
5456                                 goto out;
5457                         }
5458                 }
5459
5460                 ret = update_block_group(root, bytenr, num_bytes, 0);
5461                 if (ret) {
5462                         btrfs_abort_transaction(trans, extent_root, ret);
5463                         goto out;
5464                 }
5465         }
5466 out:
5467         btrfs_free_path(path);
5468         return ret;
5469 }
5470
5471 /*
5472  * when we free an block, it is possible (and likely) that we free the last
5473  * delayed ref for that extent as well.  This searches the delayed ref tree for
5474  * a given extent, and if there are no other delayed refs to be processed, it
5475  * removes it from the tree.
5476  */
5477 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5478                                       struct btrfs_root *root, u64 bytenr)
5479 {
5480         struct btrfs_delayed_ref_head *head;
5481         struct btrfs_delayed_ref_root *delayed_refs;
5482         struct btrfs_delayed_ref_node *ref;
5483         struct rb_node *node;
5484         int ret = 0;
5485
5486         delayed_refs = &trans->transaction->delayed_refs;
5487         spin_lock(&delayed_refs->lock);
5488         head = btrfs_find_delayed_ref_head(trans, bytenr);
5489         if (!head)
5490                 goto out;
5491
5492         node = rb_prev(&head->node.rb_node);
5493         if (!node)
5494                 goto out;
5495
5496         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5497
5498         /* there are still entries for this ref, we can't drop it */
5499         if (ref->bytenr == bytenr)
5500                 goto out;
5501
5502         if (head->extent_op) {
5503                 if (!head->must_insert_reserved)
5504                         goto out;
5505                 btrfs_free_delayed_extent_op(head->extent_op);
5506                 head->extent_op = NULL;
5507         }
5508
5509         /*
5510          * waiting for the lock here would deadlock.  If someone else has it
5511          * locked they are already in the process of dropping it anyway
5512          */
5513         if (!mutex_trylock(&head->mutex))
5514                 goto out;
5515
5516         /*
5517          * at this point we have a head with no other entries.  Go
5518          * ahead and process it.
5519          */
5520         head->node.in_tree = 0;
5521         rb_erase(&head->node.rb_node, &delayed_refs->root);
5522
5523         delayed_refs->num_entries--;
5524
5525         /*
5526          * we don't take a ref on the node because we're removing it from the
5527          * tree, so we just steal the ref the tree was holding.
5528          */
5529         delayed_refs->num_heads--;
5530         if (list_empty(&head->cluster))
5531                 delayed_refs->num_heads_ready--;
5532
5533         list_del_init(&head->cluster);
5534         spin_unlock(&delayed_refs->lock);
5535
5536         BUG_ON(head->extent_op);
5537         if (head->must_insert_reserved)
5538                 ret = 1;
5539
5540         mutex_unlock(&head->mutex);
5541         btrfs_put_delayed_ref(&head->node);
5542         return ret;
5543 out:
5544         spin_unlock(&delayed_refs->lock);
5545         return 0;
5546 }
5547
5548 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5549                            struct btrfs_root *root,
5550                            struct extent_buffer *buf,
5551                            u64 parent, int last_ref)
5552 {
5553         struct btrfs_block_group_cache *cache = NULL;
5554         int ret;
5555
5556         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5557                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5558                                         buf->start, buf->len,
5559                                         parent, root->root_key.objectid,
5560                                         btrfs_header_level(buf),
5561                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5562                 BUG_ON(ret); /* -ENOMEM */
5563         }
5564
5565         if (!last_ref)
5566                 return;
5567
5568         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5569
5570         if (btrfs_header_generation(buf) == trans->transid) {
5571                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5572                         ret = check_ref_cleanup(trans, root, buf->start);
5573                         if (!ret)
5574                                 goto out;
5575                 }
5576
5577                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5578                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5579                         goto out;
5580                 }
5581
5582                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5583
5584                 btrfs_add_free_space(cache, buf->start, buf->len);
5585                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5586         }
5587 out:
5588         /*
5589          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5590          * anymore.
5591          */
5592         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5593         btrfs_put_block_group(cache);
5594 }
5595
5596 /* Can return -ENOMEM */
5597 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5598                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5599                       u64 owner, u64 offset, int for_cow)
5600 {
5601         int ret;
5602         struct btrfs_fs_info *fs_info = root->fs_info;
5603
5604         /*
5605          * tree log blocks never actually go into the extent allocation
5606          * tree, just update pinning info and exit early.
5607          */
5608         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5609                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5610                 /* unlocks the pinned mutex */
5611                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5612                 ret = 0;
5613         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5614                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5615                                         num_bytes,
5616                                         parent, root_objectid, (int)owner,
5617                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5618         } else {
5619                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5620                                                 num_bytes,
5621                                                 parent, root_objectid, owner,
5622                                                 offset, BTRFS_DROP_DELAYED_REF,
5623                                                 NULL, for_cow);
5624         }
5625         return ret;
5626 }
5627
5628 static u64 stripe_align(struct btrfs_root *root,
5629                         struct btrfs_block_group_cache *cache,
5630                         u64 val, u64 num_bytes)
5631 {
5632         u64 ret = ALIGN(val, root->stripesize);
5633         return ret;
5634 }
5635
5636 /*
5637  * when we wait for progress in the block group caching, its because
5638  * our allocation attempt failed at least once.  So, we must sleep
5639  * and let some progress happen before we try again.
5640  *
5641  * This function will sleep at least once waiting for new free space to
5642  * show up, and then it will check the block group free space numbers
5643  * for our min num_bytes.  Another option is to have it go ahead
5644  * and look in the rbtree for a free extent of a given size, but this
5645  * is a good start.
5646  */
5647 static noinline int
5648 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5649                                 u64 num_bytes)
5650 {
5651         struct btrfs_caching_control *caching_ctl;
5652
5653         caching_ctl = get_caching_control(cache);
5654         if (!caching_ctl)
5655                 return 0;
5656
5657         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5658                    (cache->free_space_ctl->free_space >= num_bytes));
5659
5660         put_caching_control(caching_ctl);
5661         return 0;
5662 }
5663
5664 static noinline int
5665 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5666 {
5667         struct btrfs_caching_control *caching_ctl;
5668
5669         caching_ctl = get_caching_control(cache);
5670         if (!caching_ctl)
5671                 return 0;
5672
5673         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5674
5675         put_caching_control(caching_ctl);
5676         return 0;
5677 }
5678
5679 int __get_raid_index(u64 flags)
5680 {
5681         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5682                 return BTRFS_RAID_RAID10;
5683         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5684                 return BTRFS_RAID_RAID1;
5685         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5686                 return BTRFS_RAID_DUP;
5687         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5688                 return BTRFS_RAID_RAID0;
5689         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5690                 return BTRFS_RAID_RAID5;
5691         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5692                 return BTRFS_RAID_RAID6;
5693
5694         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5695 }
5696
5697 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5698 {
5699         return __get_raid_index(cache->flags);
5700 }
5701
5702 enum btrfs_loop_type {
5703         LOOP_CACHING_NOWAIT = 0,
5704         LOOP_CACHING_WAIT = 1,
5705         LOOP_ALLOC_CHUNK = 2,
5706         LOOP_NO_EMPTY_SIZE = 3,
5707 };
5708
5709 /*
5710  * walks the btree of allocated extents and find a hole of a given size.
5711  * The key ins is changed to record the hole:
5712  * ins->objectid == block start
5713  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5714  * ins->offset == number of blocks
5715  * Any available blocks before search_start are skipped.
5716  */
5717 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5718                                      struct btrfs_root *orig_root,
5719                                      u64 num_bytes, u64 empty_size,
5720                                      u64 hint_byte, struct btrfs_key *ins,
5721                                      u64 data)
5722 {
5723         int ret = 0;
5724         struct btrfs_root *root = orig_root->fs_info->extent_root;
5725         struct btrfs_free_cluster *last_ptr = NULL;
5726         struct btrfs_block_group_cache *block_group = NULL;
5727         struct btrfs_block_group_cache *used_block_group;
5728         u64 search_start = 0;
5729         int empty_cluster = 2 * 1024 * 1024;
5730         struct btrfs_space_info *space_info;
5731         int loop = 0;
5732         int index = __get_raid_index(data);
5733         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5734                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5735         bool found_uncached_bg = false;
5736         bool failed_cluster_refill = false;
5737         bool failed_alloc = false;
5738         bool use_cluster = true;
5739         bool have_caching_bg = false;
5740
5741         WARN_ON(num_bytes < root->sectorsize);
5742         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5743         ins->objectid = 0;
5744         ins->offset = 0;
5745
5746         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5747
5748         space_info = __find_space_info(root->fs_info, data);
5749         if (!space_info) {
5750                 printk(KERN_ERR "No space info for %llu\n", data);
5751                 return -ENOSPC;
5752         }
5753
5754         /*
5755          * If the space info is for both data and metadata it means we have a
5756          * small filesystem and we can't use the clustering stuff.
5757          */
5758         if (btrfs_mixed_space_info(space_info))
5759                 use_cluster = false;
5760
5761         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5762                 last_ptr = &root->fs_info->meta_alloc_cluster;
5763                 if (!btrfs_test_opt(root, SSD))
5764                         empty_cluster = 64 * 1024;
5765         }
5766
5767         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5768             btrfs_test_opt(root, SSD)) {
5769                 last_ptr = &root->fs_info->data_alloc_cluster;
5770         }
5771
5772         if (last_ptr) {
5773                 spin_lock(&last_ptr->lock);
5774                 if (last_ptr->block_group)
5775                         hint_byte = last_ptr->window_start;
5776                 spin_unlock(&last_ptr->lock);
5777         }
5778
5779         search_start = max(search_start, first_logical_byte(root, 0));
5780         search_start = max(search_start, hint_byte);
5781
5782         if (!last_ptr)
5783                 empty_cluster = 0;
5784
5785         if (search_start == hint_byte) {
5786                 block_group = btrfs_lookup_block_group(root->fs_info,
5787                                                        search_start);
5788                 used_block_group = block_group;
5789                 /*
5790                  * we don't want to use the block group if it doesn't match our
5791                  * allocation bits, or if its not cached.
5792                  *
5793                  * However if we are re-searching with an ideal block group
5794                  * picked out then we don't care that the block group is cached.
5795                  */
5796                 if (block_group && block_group_bits(block_group, data) &&
5797                     block_group->cached != BTRFS_CACHE_NO) {
5798                         down_read(&space_info->groups_sem);
5799                         if (list_empty(&block_group->list) ||
5800                             block_group->ro) {
5801                                 /*
5802                                  * someone is removing this block group,
5803                                  * we can't jump into the have_block_group
5804                                  * target because our list pointers are not
5805                                  * valid
5806                                  */
5807                                 btrfs_put_block_group(block_group);
5808                                 up_read(&space_info->groups_sem);
5809                         } else {
5810                                 index = get_block_group_index(block_group);
5811                                 goto have_block_group;
5812                         }
5813                 } else if (block_group) {
5814                         btrfs_put_block_group(block_group);
5815                 }
5816         }
5817 search:
5818         have_caching_bg = false;
5819         down_read(&space_info->groups_sem);
5820         list_for_each_entry(block_group, &space_info->block_groups[index],
5821                             list) {
5822                 u64 offset;
5823                 int cached;
5824
5825                 used_block_group = block_group;
5826                 btrfs_get_block_group(block_group);
5827                 search_start = block_group->key.objectid;
5828
5829                 /*
5830                  * this can happen if we end up cycling through all the
5831                  * raid types, but we want to make sure we only allocate
5832                  * for the proper type.
5833                  */
5834                 if (!block_group_bits(block_group, data)) {
5835                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5836                                 BTRFS_BLOCK_GROUP_RAID1 |
5837                                 BTRFS_BLOCK_GROUP_RAID5 |
5838                                 BTRFS_BLOCK_GROUP_RAID6 |
5839                                 BTRFS_BLOCK_GROUP_RAID10;
5840
5841                         /*
5842                          * if they asked for extra copies and this block group
5843                          * doesn't provide them, bail.  This does allow us to
5844                          * fill raid0 from raid1.
5845                          */
5846                         if ((data & extra) && !(block_group->flags & extra))
5847                                 goto loop;
5848                 }
5849
5850 have_block_group:
5851                 cached = block_group_cache_done(block_group);
5852                 if (unlikely(!cached)) {
5853                         found_uncached_bg = true;
5854                         ret = cache_block_group(block_group, 0);
5855                         BUG_ON(ret < 0);
5856                         ret = 0;
5857                 }
5858
5859                 if (unlikely(block_group->ro))
5860                         goto loop;
5861
5862                 /*
5863                  * Ok we want to try and use the cluster allocator, so
5864                  * lets look there
5865                  */
5866                 if (last_ptr) {
5867                         unsigned long aligned_cluster;
5868                         /*
5869                          * the refill lock keeps out other
5870                          * people trying to start a new cluster
5871                          */
5872                         spin_lock(&last_ptr->refill_lock);
5873                         used_block_group = last_ptr->block_group;
5874                         if (used_block_group != block_group &&
5875                             (!used_block_group ||
5876                              used_block_group->ro ||
5877                              !block_group_bits(used_block_group, data))) {
5878                                 used_block_group = block_group;
5879                                 goto refill_cluster;
5880                         }
5881
5882                         if (used_block_group != block_group)
5883                                 btrfs_get_block_group(used_block_group);
5884
5885                         offset = btrfs_alloc_from_cluster(used_block_group,
5886                           last_ptr, num_bytes, used_block_group->key.objectid);
5887                         if (offset) {
5888                                 /* we have a block, we're done */
5889                                 spin_unlock(&last_ptr->refill_lock);
5890                                 trace_btrfs_reserve_extent_cluster(root,
5891                                         block_group, search_start, num_bytes);
5892                                 goto checks;
5893                         }
5894
5895                         WARN_ON(last_ptr->block_group != used_block_group);
5896                         if (used_block_group != block_group) {
5897                                 btrfs_put_block_group(used_block_group);
5898                                 used_block_group = block_group;
5899                         }
5900 refill_cluster:
5901                         BUG_ON(used_block_group != block_group);
5902                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5903                          * set up a new clusters, so lets just skip it
5904                          * and let the allocator find whatever block
5905                          * it can find.  If we reach this point, we
5906                          * will have tried the cluster allocator
5907                          * plenty of times and not have found
5908                          * anything, so we are likely way too
5909                          * fragmented for the clustering stuff to find
5910                          * anything.
5911                          *
5912                          * However, if the cluster is taken from the
5913                          * current block group, release the cluster
5914                          * first, so that we stand a better chance of
5915                          * succeeding in the unclustered
5916                          * allocation.  */
5917                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5918                             last_ptr->block_group != block_group) {
5919                                 spin_unlock(&last_ptr->refill_lock);
5920                                 goto unclustered_alloc;
5921                         }
5922
5923                         /*
5924                          * this cluster didn't work out, free it and
5925                          * start over
5926                          */
5927                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5928
5929                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5930                                 spin_unlock(&last_ptr->refill_lock);
5931                                 goto unclustered_alloc;
5932                         }
5933
5934                         aligned_cluster = max_t(unsigned long,
5935                                                 empty_cluster + empty_size,
5936                                               block_group->full_stripe_len);
5937
5938                         /* allocate a cluster in this block group */
5939                         ret = btrfs_find_space_cluster(trans, root,
5940                                                block_group, last_ptr,
5941                                                search_start, num_bytes,
5942                                                aligned_cluster);
5943                         if (ret == 0) {
5944                                 /*
5945                                  * now pull our allocation out of this
5946                                  * cluster
5947                                  */
5948                                 offset = btrfs_alloc_from_cluster(block_group,
5949                                                   last_ptr, num_bytes,
5950                                                   search_start);
5951                                 if (offset) {
5952                                         /* we found one, proceed */
5953                                         spin_unlock(&last_ptr->refill_lock);
5954                                         trace_btrfs_reserve_extent_cluster(root,
5955                                                 block_group, search_start,
5956                                                 num_bytes);
5957                                         goto checks;
5958                                 }
5959                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5960                                    && !failed_cluster_refill) {
5961                                 spin_unlock(&last_ptr->refill_lock);
5962
5963                                 failed_cluster_refill = true;
5964                                 wait_block_group_cache_progress(block_group,
5965                                        num_bytes + empty_cluster + empty_size);
5966                                 goto have_block_group;
5967                         }
5968
5969                         /*
5970                          * at this point we either didn't find a cluster
5971                          * or we weren't able to allocate a block from our
5972                          * cluster.  Free the cluster we've been trying
5973                          * to use, and go to the next block group
5974                          */
5975                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5976                         spin_unlock(&last_ptr->refill_lock);
5977                         goto loop;
5978                 }
5979
5980 unclustered_alloc:
5981                 spin_lock(&block_group->free_space_ctl->tree_lock);
5982                 if (cached &&
5983                     block_group->free_space_ctl->free_space <
5984                     num_bytes + empty_cluster + empty_size) {
5985                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5986                         goto loop;
5987                 }
5988                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5989
5990                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5991                                                     num_bytes, empty_size);
5992                 /*
5993                  * If we didn't find a chunk, and we haven't failed on this
5994                  * block group before, and this block group is in the middle of
5995                  * caching and we are ok with waiting, then go ahead and wait
5996                  * for progress to be made, and set failed_alloc to true.
5997                  *
5998                  * If failed_alloc is true then we've already waited on this
5999                  * block group once and should move on to the next block group.
6000                  */
6001                 if (!offset && !failed_alloc && !cached &&
6002                     loop > LOOP_CACHING_NOWAIT) {
6003                         wait_block_group_cache_progress(block_group,
6004                                                 num_bytes + empty_size);
6005                         failed_alloc = true;
6006                         goto have_block_group;
6007                 } else if (!offset) {
6008                         if (!cached)
6009                                 have_caching_bg = true;
6010                         goto loop;
6011                 }
6012 checks:
6013                 search_start = stripe_align(root, used_block_group,
6014                                             offset, num_bytes);
6015
6016                 /* move on to the next group */
6017                 if (search_start + num_bytes >
6018                     used_block_group->key.objectid + used_block_group->key.offset) {
6019                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6020                         goto loop;
6021                 }
6022
6023                 if (offset < search_start)
6024                         btrfs_add_free_space(used_block_group, offset,
6025                                              search_start - offset);
6026                 BUG_ON(offset > search_start);
6027
6028                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6029                                                   alloc_type);
6030                 if (ret == -EAGAIN) {
6031                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6032                         goto loop;
6033                 }
6034
6035                 /* we are all good, lets return */
6036                 ins->objectid = search_start;
6037                 ins->offset = num_bytes;
6038
6039                 trace_btrfs_reserve_extent(orig_root, block_group,
6040                                            search_start, num_bytes);
6041                 if (used_block_group != block_group)
6042                         btrfs_put_block_group(used_block_group);
6043                 btrfs_put_block_group(block_group);
6044                 break;
6045 loop:
6046                 failed_cluster_refill = false;
6047                 failed_alloc = false;
6048                 BUG_ON(index != get_block_group_index(block_group));
6049                 if (used_block_group != block_group)
6050                         btrfs_put_block_group(used_block_group);
6051                 btrfs_put_block_group(block_group);
6052         }
6053         up_read(&space_info->groups_sem);
6054
6055         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6056                 goto search;
6057
6058         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6059                 goto search;
6060
6061         /*
6062          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6063          *                      caching kthreads as we move along
6064          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6065          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6066          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6067          *                      again
6068          */
6069         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6070                 index = 0;
6071                 loop++;
6072                 if (loop == LOOP_ALLOC_CHUNK) {
6073                         ret = do_chunk_alloc(trans, root, data,
6074                                              CHUNK_ALLOC_FORCE);
6075                         /*
6076                          * Do not bail out on ENOSPC since we
6077                          * can do more things.
6078                          */
6079                         if (ret < 0 && ret != -ENOSPC) {
6080                                 btrfs_abort_transaction(trans,
6081                                                         root, ret);
6082                                 goto out;
6083                         }
6084                 }
6085
6086                 if (loop == LOOP_NO_EMPTY_SIZE) {
6087                         empty_size = 0;
6088                         empty_cluster = 0;
6089                 }
6090
6091                 goto search;
6092         } else if (!ins->objectid) {
6093                 ret = -ENOSPC;
6094         } else if (ins->objectid) {
6095                 ret = 0;
6096         }
6097 out:
6098
6099         return ret;
6100 }
6101
6102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6103                             int dump_block_groups)
6104 {
6105         struct btrfs_block_group_cache *cache;
6106         int index = 0;
6107
6108         spin_lock(&info->lock);
6109         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6110                (unsigned long long)info->flags,
6111                (unsigned long long)(info->total_bytes - info->bytes_used -
6112                                     info->bytes_pinned - info->bytes_reserved -
6113                                     info->bytes_readonly),
6114                (info->full) ? "" : "not ");
6115         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6116                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6117                (unsigned long long)info->total_bytes,
6118                (unsigned long long)info->bytes_used,
6119                (unsigned long long)info->bytes_pinned,
6120                (unsigned long long)info->bytes_reserved,
6121                (unsigned long long)info->bytes_may_use,
6122                (unsigned long long)info->bytes_readonly);
6123         spin_unlock(&info->lock);
6124
6125         if (!dump_block_groups)
6126                 return;
6127
6128         down_read(&info->groups_sem);
6129 again:
6130         list_for_each_entry(cache, &info->block_groups[index], list) {
6131                 spin_lock(&cache->lock);
6132                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6133                        (unsigned long long)cache->key.objectid,
6134                        (unsigned long long)cache->key.offset,
6135                        (unsigned long long)btrfs_block_group_used(&cache->item),
6136                        (unsigned long long)cache->pinned,
6137                        (unsigned long long)cache->reserved,
6138                        cache->ro ? "[readonly]" : "");
6139                 btrfs_dump_free_space(cache, bytes);
6140                 spin_unlock(&cache->lock);
6141         }
6142         if (++index < BTRFS_NR_RAID_TYPES)
6143                 goto again;
6144         up_read(&info->groups_sem);
6145 }
6146
6147 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6148                          struct btrfs_root *root,
6149                          u64 num_bytes, u64 min_alloc_size,
6150                          u64 empty_size, u64 hint_byte,
6151                          struct btrfs_key *ins, u64 data)
6152 {
6153         bool final_tried = false;
6154         int ret;
6155
6156         data = btrfs_get_alloc_profile(root, data);
6157 again:
6158         WARN_ON(num_bytes < root->sectorsize);
6159         ret = find_free_extent(trans, root, num_bytes, empty_size,
6160                                hint_byte, ins, data);
6161
6162         if (ret == -ENOSPC) {
6163                 if (!final_tried) {
6164                         num_bytes = num_bytes >> 1;
6165                         num_bytes = round_down(num_bytes, root->sectorsize);
6166                         num_bytes = max(num_bytes, min_alloc_size);
6167                         if (num_bytes == min_alloc_size)
6168                                 final_tried = true;
6169                         goto again;
6170                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6171                         struct btrfs_space_info *sinfo;
6172
6173                         sinfo = __find_space_info(root->fs_info, data);
6174                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
6175                                "wanted %llu\n", (unsigned long long)data,
6176                                (unsigned long long)num_bytes);
6177                         if (sinfo)
6178                                 dump_space_info(sinfo, num_bytes, 1);
6179                 }
6180         }
6181
6182         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6183
6184         return ret;
6185 }
6186
6187 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6188                                         u64 start, u64 len, int pin)
6189 {
6190         struct btrfs_block_group_cache *cache;
6191         int ret = 0;
6192
6193         cache = btrfs_lookup_block_group(root->fs_info, start);
6194         if (!cache) {
6195                 printk(KERN_ERR "Unable to find block group for %llu\n",
6196                        (unsigned long long)start);
6197                 return -ENOSPC;
6198         }
6199
6200         if (btrfs_test_opt(root, DISCARD))
6201                 ret = btrfs_discard_extent(root, start, len, NULL);
6202
6203         if (pin)
6204                 pin_down_extent(root, cache, start, len, 1);
6205         else {
6206                 btrfs_add_free_space(cache, start, len);
6207                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6208         }
6209         btrfs_put_block_group(cache);
6210
6211         trace_btrfs_reserved_extent_free(root, start, len);
6212
6213         return ret;
6214 }
6215
6216 int btrfs_free_reserved_extent(struct btrfs_root *root,
6217                                         u64 start, u64 len)
6218 {
6219         return __btrfs_free_reserved_extent(root, start, len, 0);
6220 }
6221
6222 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6223                                        u64 start, u64 len)
6224 {
6225         return __btrfs_free_reserved_extent(root, start, len, 1);
6226 }
6227
6228 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6229                                       struct btrfs_root *root,
6230                                       u64 parent, u64 root_objectid,
6231                                       u64 flags, u64 owner, u64 offset,
6232                                       struct btrfs_key *ins, int ref_mod)
6233 {
6234         int ret;
6235         struct btrfs_fs_info *fs_info = root->fs_info;
6236         struct btrfs_extent_item *extent_item;
6237         struct btrfs_extent_inline_ref *iref;
6238         struct btrfs_path *path;
6239         struct extent_buffer *leaf;
6240         int type;
6241         u32 size;
6242
6243         if (parent > 0)
6244                 type = BTRFS_SHARED_DATA_REF_KEY;
6245         else
6246                 type = BTRFS_EXTENT_DATA_REF_KEY;
6247
6248         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6249
6250         path = btrfs_alloc_path();
6251         if (!path)
6252                 return -ENOMEM;
6253
6254         path->leave_spinning = 1;
6255         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6256                                       ins, size);
6257         if (ret) {
6258                 btrfs_free_path(path);
6259                 return ret;
6260         }
6261
6262         leaf = path->nodes[0];
6263         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6264                                      struct btrfs_extent_item);
6265         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6266         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6267         btrfs_set_extent_flags(leaf, extent_item,
6268                                flags | BTRFS_EXTENT_FLAG_DATA);
6269
6270         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6271         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6272         if (parent > 0) {
6273                 struct btrfs_shared_data_ref *ref;
6274                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6275                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6276                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6277         } else {
6278                 struct btrfs_extent_data_ref *ref;
6279                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6280                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6281                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6282                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6283                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6284         }
6285
6286         btrfs_mark_buffer_dirty(path->nodes[0]);
6287         btrfs_free_path(path);
6288
6289         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6290         if (ret) { /* -ENOENT, logic error */
6291                 printk(KERN_ERR "btrfs update block group failed for %llu "
6292                        "%llu\n", (unsigned long long)ins->objectid,
6293                        (unsigned long long)ins->offset);
6294                 BUG();
6295         }
6296         return ret;
6297 }
6298
6299 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6300                                      struct btrfs_root *root,
6301                                      u64 parent, u64 root_objectid,
6302                                      u64 flags, struct btrfs_disk_key *key,
6303                                      int level, struct btrfs_key *ins)
6304 {
6305         int ret;
6306         struct btrfs_fs_info *fs_info = root->fs_info;
6307         struct btrfs_extent_item *extent_item;
6308         struct btrfs_tree_block_info *block_info;
6309         struct btrfs_extent_inline_ref *iref;
6310         struct btrfs_path *path;
6311         struct extent_buffer *leaf;
6312         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6313
6314         path = btrfs_alloc_path();
6315         if (!path)
6316                 return -ENOMEM;
6317
6318         path->leave_spinning = 1;
6319         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6320                                       ins, size);
6321         if (ret) {
6322                 btrfs_free_path(path);
6323                 return ret;
6324         }
6325
6326         leaf = path->nodes[0];
6327         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6328                                      struct btrfs_extent_item);
6329         btrfs_set_extent_refs(leaf, extent_item, 1);
6330         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6331         btrfs_set_extent_flags(leaf, extent_item,
6332                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6333         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6334
6335         btrfs_set_tree_block_key(leaf, block_info, key);
6336         btrfs_set_tree_block_level(leaf, block_info, level);
6337
6338         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6339         if (parent > 0) {
6340                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6341                 btrfs_set_extent_inline_ref_type(leaf, iref,
6342                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6343                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6344         } else {
6345                 btrfs_set_extent_inline_ref_type(leaf, iref,
6346                                                  BTRFS_TREE_BLOCK_REF_KEY);
6347                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6348         }
6349
6350         btrfs_mark_buffer_dirty(leaf);
6351         btrfs_free_path(path);
6352
6353         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6354         if (ret) { /* -ENOENT, logic error */
6355                 printk(KERN_ERR "btrfs update block group failed for %llu "
6356                        "%llu\n", (unsigned long long)ins->objectid,
6357                        (unsigned long long)ins->offset);
6358                 BUG();
6359         }
6360         return ret;
6361 }
6362
6363 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6364                                      struct btrfs_root *root,
6365                                      u64 root_objectid, u64 owner,
6366                                      u64 offset, struct btrfs_key *ins)
6367 {
6368         int ret;
6369
6370         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6371
6372         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6373                                          ins->offset, 0,
6374                                          root_objectid, owner, offset,
6375                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6376         return ret;
6377 }
6378
6379 /*
6380  * this is used by the tree logging recovery code.  It records that
6381  * an extent has been allocated and makes sure to clear the free
6382  * space cache bits as well
6383  */
6384 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6385                                    struct btrfs_root *root,
6386                                    u64 root_objectid, u64 owner, u64 offset,
6387                                    struct btrfs_key *ins)
6388 {
6389         int ret;
6390         struct btrfs_block_group_cache *block_group;
6391         struct btrfs_caching_control *caching_ctl;
6392         u64 start = ins->objectid;
6393         u64 num_bytes = ins->offset;
6394
6395         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6396         cache_block_group(block_group, 0);
6397         caching_ctl = get_caching_control(block_group);
6398
6399         if (!caching_ctl) {
6400                 BUG_ON(!block_group_cache_done(block_group));
6401                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6402                 BUG_ON(ret); /* -ENOMEM */
6403         } else {
6404                 mutex_lock(&caching_ctl->mutex);
6405
6406                 if (start >= caching_ctl->progress) {
6407                         ret = add_excluded_extent(root, start, num_bytes);
6408                         BUG_ON(ret); /* -ENOMEM */
6409                 } else if (start + num_bytes <= caching_ctl->progress) {
6410                         ret = btrfs_remove_free_space(block_group,
6411                                                       start, num_bytes);
6412                         BUG_ON(ret); /* -ENOMEM */
6413                 } else {
6414                         num_bytes = caching_ctl->progress - start;
6415                         ret = btrfs_remove_free_space(block_group,
6416                                                       start, num_bytes);
6417                         BUG_ON(ret); /* -ENOMEM */
6418
6419                         start = caching_ctl->progress;
6420                         num_bytes = ins->objectid + ins->offset -
6421                                     caching_ctl->progress;
6422                         ret = add_excluded_extent(root, start, num_bytes);
6423                         BUG_ON(ret); /* -ENOMEM */
6424                 }
6425
6426                 mutex_unlock(&caching_ctl->mutex);
6427                 put_caching_control(caching_ctl);
6428         }
6429
6430         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6431                                           RESERVE_ALLOC_NO_ACCOUNT);
6432         BUG_ON(ret); /* logic error */
6433         btrfs_put_block_group(block_group);
6434         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6435                                          0, owner, offset, ins, 1);
6436         return ret;
6437 }
6438
6439 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6440                                             struct btrfs_root *root,
6441                                             u64 bytenr, u32 blocksize,
6442                                             int level)
6443 {
6444         struct extent_buffer *buf;
6445
6446         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6447         if (!buf)
6448                 return ERR_PTR(-ENOMEM);
6449         btrfs_set_header_generation(buf, trans->transid);
6450         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6451         btrfs_tree_lock(buf);
6452         clean_tree_block(trans, root, buf);
6453         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6454
6455         btrfs_set_lock_blocking(buf);
6456         btrfs_set_buffer_uptodate(buf);
6457
6458         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6459                 /*
6460                  * we allow two log transactions at a time, use different
6461                  * EXENT bit to differentiate dirty pages.
6462                  */
6463                 if (root->log_transid % 2 == 0)
6464                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6465                                         buf->start + buf->len - 1, GFP_NOFS);
6466                 else
6467                         set_extent_new(&root->dirty_log_pages, buf->start,
6468                                         buf->start + buf->len - 1, GFP_NOFS);
6469         } else {
6470                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6471                          buf->start + buf->len - 1, GFP_NOFS);
6472         }
6473         trans->blocks_used++;
6474         /* this returns a buffer locked for blocking */
6475         return buf;
6476 }
6477
6478 static struct btrfs_block_rsv *
6479 use_block_rsv(struct btrfs_trans_handle *trans,
6480               struct btrfs_root *root, u32 blocksize)
6481 {
6482         struct btrfs_block_rsv *block_rsv;
6483         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6484         int ret;
6485
6486         block_rsv = get_block_rsv(trans, root);
6487
6488         if (block_rsv->size == 0) {
6489                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6490                                              BTRFS_RESERVE_NO_FLUSH);
6491                 /*
6492                  * If we couldn't reserve metadata bytes try and use some from
6493                  * the global reserve.
6494                  */
6495                 if (ret && block_rsv != global_rsv) {
6496                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6497                         if (!ret)
6498                                 return global_rsv;
6499                         return ERR_PTR(ret);
6500                 } else if (ret) {
6501                         return ERR_PTR(ret);
6502                 }
6503                 return block_rsv;
6504         }
6505
6506         ret = block_rsv_use_bytes(block_rsv, blocksize);
6507         if (!ret)
6508                 return block_rsv;
6509         if (ret && !block_rsv->failfast) {
6510                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6511                         static DEFINE_RATELIMIT_STATE(_rs,
6512                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6513                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6514                         if (__ratelimit(&_rs))
6515                                 WARN(1, KERN_DEBUG
6516                                         "btrfs: block rsv returned %d\n", ret);
6517                 }
6518                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6519                                              BTRFS_RESERVE_NO_FLUSH);
6520                 if (!ret) {
6521                         return block_rsv;
6522                 } else if (ret && block_rsv != global_rsv) {
6523                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6524                         if (!ret)
6525                                 return global_rsv;
6526                 }
6527         }
6528
6529         return ERR_PTR(-ENOSPC);
6530 }
6531
6532 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6533                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6534 {
6535         block_rsv_add_bytes(block_rsv, blocksize, 0);
6536         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6537 }
6538
6539 /*
6540  * finds a free extent and does all the dirty work required for allocation
6541  * returns the key for the extent through ins, and a tree buffer for
6542  * the first block of the extent through buf.
6543  *
6544  * returns the tree buffer or NULL.
6545  */
6546 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6547                                         struct btrfs_root *root, u32 blocksize,
6548                                         u64 parent, u64 root_objectid,
6549                                         struct btrfs_disk_key *key, int level,
6550                                         u64 hint, u64 empty_size)
6551 {
6552         struct btrfs_key ins;
6553         struct btrfs_block_rsv *block_rsv;
6554         struct extent_buffer *buf;
6555         u64 flags = 0;
6556         int ret;
6557
6558
6559         block_rsv = use_block_rsv(trans, root, blocksize);
6560         if (IS_ERR(block_rsv))
6561                 return ERR_CAST(block_rsv);
6562
6563         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6564                                    empty_size, hint, &ins, 0);
6565         if (ret) {
6566                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6567                 return ERR_PTR(ret);
6568         }
6569
6570         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6571                                     blocksize, level);
6572         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6573
6574         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6575                 if (parent == 0)
6576                         parent = ins.objectid;
6577                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6578         } else
6579                 BUG_ON(parent > 0);
6580
6581         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6582                 struct btrfs_delayed_extent_op *extent_op;
6583                 extent_op = btrfs_alloc_delayed_extent_op();
6584                 BUG_ON(!extent_op); /* -ENOMEM */
6585                 if (key)
6586                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6587                 else
6588                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6589                 extent_op->flags_to_set = flags;
6590                 extent_op->update_key = 1;
6591                 extent_op->update_flags = 1;
6592                 extent_op->is_data = 0;
6593
6594                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6595                                         ins.objectid,
6596                                         ins.offset, parent, root_objectid,
6597                                         level, BTRFS_ADD_DELAYED_EXTENT,
6598                                         extent_op, 0);
6599                 BUG_ON(ret); /* -ENOMEM */
6600         }
6601         return buf;
6602 }
6603
6604 struct walk_control {
6605         u64 refs[BTRFS_MAX_LEVEL];
6606         u64 flags[BTRFS_MAX_LEVEL];
6607         struct btrfs_key update_progress;
6608         int stage;
6609         int level;
6610         int shared_level;
6611         int update_ref;
6612         int keep_locks;
6613         int reada_slot;
6614         int reada_count;
6615         int for_reloc;
6616 };
6617
6618 #define DROP_REFERENCE  1
6619 #define UPDATE_BACKREF  2
6620
6621 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6622                                      struct btrfs_root *root,
6623                                      struct walk_control *wc,
6624                                      struct btrfs_path *path)
6625 {
6626         u64 bytenr;
6627         u64 generation;
6628         u64 refs;
6629         u64 flags;
6630         u32 nritems;
6631         u32 blocksize;
6632         struct btrfs_key key;
6633         struct extent_buffer *eb;
6634         int ret;
6635         int slot;
6636         int nread = 0;
6637
6638         if (path->slots[wc->level] < wc->reada_slot) {
6639                 wc->reada_count = wc->reada_count * 2 / 3;
6640                 wc->reada_count = max(wc->reada_count, 2);
6641         } else {
6642                 wc->reada_count = wc->reada_count * 3 / 2;
6643                 wc->reada_count = min_t(int, wc->reada_count,
6644                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6645         }
6646
6647         eb = path->nodes[wc->level];
6648         nritems = btrfs_header_nritems(eb);
6649         blocksize = btrfs_level_size(root, wc->level - 1);
6650
6651         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6652                 if (nread >= wc->reada_count)
6653                         break;
6654
6655                 cond_resched();
6656                 bytenr = btrfs_node_blockptr(eb, slot);
6657                 generation = btrfs_node_ptr_generation(eb, slot);
6658
6659                 if (slot == path->slots[wc->level])
6660                         goto reada;
6661
6662                 if (wc->stage == UPDATE_BACKREF &&
6663                     generation <= root->root_key.offset)
6664                         continue;
6665
6666                 /* We don't lock the tree block, it's OK to be racy here */
6667                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6668                                                &refs, &flags);
6669                 /* We don't care about errors in readahead. */
6670                 if (ret < 0)
6671                         continue;
6672                 BUG_ON(refs == 0);
6673
6674                 if (wc->stage == DROP_REFERENCE) {
6675                         if (refs == 1)
6676                                 goto reada;
6677
6678                         if (wc->level == 1 &&
6679                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6680                                 continue;
6681                         if (!wc->update_ref ||
6682                             generation <= root->root_key.offset)
6683                                 continue;
6684                         btrfs_node_key_to_cpu(eb, &key, slot);
6685                         ret = btrfs_comp_cpu_keys(&key,
6686                                                   &wc->update_progress);
6687                         if (ret < 0)
6688                                 continue;
6689                 } else {
6690                         if (wc->level == 1 &&
6691                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6692                                 continue;
6693                 }
6694 reada:
6695                 ret = readahead_tree_block(root, bytenr, blocksize,
6696                                            generation);
6697                 if (ret)
6698                         break;
6699                 nread++;
6700         }
6701         wc->reada_slot = slot;
6702 }
6703
6704 /*
6705  * helper to process tree block while walking down the tree.
6706  *
6707  * when wc->stage == UPDATE_BACKREF, this function updates
6708  * back refs for pointers in the block.
6709  *
6710  * NOTE: return value 1 means we should stop walking down.
6711  */
6712 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6713                                    struct btrfs_root *root,
6714                                    struct btrfs_path *path,
6715                                    struct walk_control *wc, int lookup_info)
6716 {
6717         int level = wc->level;
6718         struct extent_buffer *eb = path->nodes[level];
6719         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6720         int ret;
6721
6722         if (wc->stage == UPDATE_BACKREF &&
6723             btrfs_header_owner(eb) != root->root_key.objectid)
6724                 return 1;
6725
6726         /*
6727          * when reference count of tree block is 1, it won't increase
6728          * again. once full backref flag is set, we never clear it.
6729          */
6730         if (lookup_info &&
6731             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6732              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6733                 BUG_ON(!path->locks[level]);
6734                 ret = btrfs_lookup_extent_info(trans, root,
6735                                                eb->start, eb->len,
6736                                                &wc->refs[level],
6737                                                &wc->flags[level]);
6738                 BUG_ON(ret == -ENOMEM);
6739                 if (ret)
6740                         return ret;
6741                 BUG_ON(wc->refs[level] == 0);
6742         }
6743
6744         if (wc->stage == DROP_REFERENCE) {
6745                 if (wc->refs[level] > 1)
6746                         return 1;
6747
6748                 if (path->locks[level] && !wc->keep_locks) {
6749                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6750                         path->locks[level] = 0;
6751                 }
6752                 return 0;
6753         }
6754
6755         /* wc->stage == UPDATE_BACKREF */
6756         if (!(wc->flags[level] & flag)) {
6757                 BUG_ON(!path->locks[level]);
6758                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6759                 BUG_ON(ret); /* -ENOMEM */
6760                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6761                 BUG_ON(ret); /* -ENOMEM */
6762                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6763                                                   eb->len, flag, 0);
6764                 BUG_ON(ret); /* -ENOMEM */
6765                 wc->flags[level] |= flag;
6766         }
6767
6768         /*
6769          * the block is shared by multiple trees, so it's not good to
6770          * keep the tree lock
6771          */
6772         if (path->locks[level] && level > 0) {
6773                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6774                 path->locks[level] = 0;
6775         }
6776         return 0;
6777 }
6778
6779 /*
6780  * helper to process tree block pointer.
6781  *
6782  * when wc->stage == DROP_REFERENCE, this function checks
6783  * reference count of the block pointed to. if the block
6784  * is shared and we need update back refs for the subtree
6785  * rooted at the block, this function changes wc->stage to
6786  * UPDATE_BACKREF. if the block is shared and there is no
6787  * need to update back, this function drops the reference
6788  * to the block.
6789  *
6790  * NOTE: return value 1 means we should stop walking down.
6791  */
6792 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6793                                  struct btrfs_root *root,
6794                                  struct btrfs_path *path,
6795                                  struct walk_control *wc, int *lookup_info)
6796 {
6797         u64 bytenr;
6798         u64 generation;
6799         u64 parent;
6800         u32 blocksize;
6801         struct btrfs_key key;
6802         struct extent_buffer *next;
6803         int level = wc->level;
6804         int reada = 0;
6805         int ret = 0;
6806
6807         generation = btrfs_node_ptr_generation(path->nodes[level],
6808                                                path->slots[level]);
6809         /*
6810          * if the lower level block was created before the snapshot
6811          * was created, we know there is no need to update back refs
6812          * for the subtree
6813          */
6814         if (wc->stage == UPDATE_BACKREF &&
6815             generation <= root->root_key.offset) {
6816                 *lookup_info = 1;
6817                 return 1;
6818         }
6819
6820         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6821         blocksize = btrfs_level_size(root, level - 1);
6822
6823         next = btrfs_find_tree_block(root, bytenr, blocksize);
6824         if (!next) {
6825                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6826                 if (!next)
6827                         return -ENOMEM;
6828                 reada = 1;
6829         }
6830         btrfs_tree_lock(next);
6831         btrfs_set_lock_blocking(next);
6832
6833         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6834                                        &wc->refs[level - 1],
6835                                        &wc->flags[level - 1]);
6836         if (ret < 0) {
6837                 btrfs_tree_unlock(next);
6838                 return ret;
6839         }
6840
6841         BUG_ON(wc->refs[level - 1] == 0);
6842         *lookup_info = 0;
6843
6844         if (wc->stage == DROP_REFERENCE) {
6845                 if (wc->refs[level - 1] > 1) {
6846                         if (level == 1 &&
6847                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6848                                 goto skip;
6849
6850                         if (!wc->update_ref ||
6851                             generation <= root->root_key.offset)
6852                                 goto skip;
6853
6854                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6855                                               path->slots[level]);
6856                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6857                         if (ret < 0)
6858                                 goto skip;
6859
6860                         wc->stage = UPDATE_BACKREF;
6861                         wc->shared_level = level - 1;
6862                 }
6863         } else {
6864                 if (level == 1 &&
6865                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6866                         goto skip;
6867         }
6868
6869         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6870                 btrfs_tree_unlock(next);
6871                 free_extent_buffer(next);
6872                 next = NULL;
6873                 *lookup_info = 1;
6874         }
6875
6876         if (!next) {
6877                 if (reada && level == 1)
6878                         reada_walk_down(trans, root, wc, path);
6879                 next = read_tree_block(root, bytenr, blocksize, generation);
6880                 if (!next)
6881                         return -EIO;
6882                 btrfs_tree_lock(next);
6883                 btrfs_set_lock_blocking(next);
6884         }
6885
6886         level--;
6887         BUG_ON(level != btrfs_header_level(next));
6888         path->nodes[level] = next;
6889         path->slots[level] = 0;
6890         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6891         wc->level = level;
6892         if (wc->level == 1)
6893                 wc->reada_slot = 0;
6894         return 0;
6895 skip:
6896         wc->refs[level - 1] = 0;
6897         wc->flags[level - 1] = 0;
6898         if (wc->stage == DROP_REFERENCE) {
6899                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6900                         parent = path->nodes[level]->start;
6901                 } else {
6902                         BUG_ON(root->root_key.objectid !=
6903                                btrfs_header_owner(path->nodes[level]));
6904                         parent = 0;
6905                 }
6906
6907                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6908                                 root->root_key.objectid, level - 1, 0, 0);
6909                 BUG_ON(ret); /* -ENOMEM */
6910         }
6911         btrfs_tree_unlock(next);
6912         free_extent_buffer(next);
6913         *lookup_info = 1;
6914         return 1;
6915 }
6916
6917 /*
6918  * helper to process tree block while walking up the tree.
6919  *
6920  * when wc->stage == DROP_REFERENCE, this function drops
6921  * reference count on the block.
6922  *
6923  * when wc->stage == UPDATE_BACKREF, this function changes
6924  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6925  * to UPDATE_BACKREF previously while processing the block.
6926  *
6927  * NOTE: return value 1 means we should stop walking up.
6928  */
6929 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6930                                  struct btrfs_root *root,
6931                                  struct btrfs_path *path,
6932                                  struct walk_control *wc)
6933 {
6934         int ret;
6935         int level = wc->level;
6936         struct extent_buffer *eb = path->nodes[level];
6937         u64 parent = 0;
6938
6939         if (wc->stage == UPDATE_BACKREF) {
6940                 BUG_ON(wc->shared_level < level);
6941                 if (level < wc->shared_level)
6942                         goto out;
6943
6944                 ret = find_next_key(path, level + 1, &wc->update_progress);
6945                 if (ret > 0)
6946                         wc->update_ref = 0;
6947
6948                 wc->stage = DROP_REFERENCE;
6949                 wc->shared_level = -1;
6950                 path->slots[level] = 0;
6951
6952                 /*
6953                  * check reference count again if the block isn't locked.
6954                  * we should start walking down the tree again if reference
6955                  * count is one.
6956                  */
6957                 if (!path->locks[level]) {
6958                         BUG_ON(level == 0);
6959                         btrfs_tree_lock(eb);
6960                         btrfs_set_lock_blocking(eb);
6961                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6962
6963                         ret = btrfs_lookup_extent_info(trans, root,
6964                                                        eb->start, eb->len,
6965                                                        &wc->refs[level],
6966                                                        &wc->flags[level]);
6967                         if (ret < 0) {
6968                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6969                                 path->locks[level] = 0;
6970                                 return ret;
6971                         }
6972                         BUG_ON(wc->refs[level] == 0);
6973                         if (wc->refs[level] == 1) {
6974                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6975                                 path->locks[level] = 0;
6976                                 return 1;
6977                         }
6978                 }
6979         }
6980
6981         /* wc->stage == DROP_REFERENCE */
6982         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6983
6984         if (wc->refs[level] == 1) {
6985                 if (level == 0) {
6986                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6987                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6988                                                     wc->for_reloc);
6989                         else
6990                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6991                                                     wc->for_reloc);
6992                         BUG_ON(ret); /* -ENOMEM */
6993                 }
6994                 /* make block locked assertion in clean_tree_block happy */
6995                 if (!path->locks[level] &&
6996                     btrfs_header_generation(eb) == trans->transid) {
6997                         btrfs_tree_lock(eb);
6998                         btrfs_set_lock_blocking(eb);
6999                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7000                 }
7001                 clean_tree_block(trans, root, eb);
7002         }
7003
7004         if (eb == root->node) {
7005                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7006                         parent = eb->start;
7007                 else
7008                         BUG_ON(root->root_key.objectid !=
7009                                btrfs_header_owner(eb));
7010         } else {
7011                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7012                         parent = path->nodes[level + 1]->start;
7013                 else
7014                         BUG_ON(root->root_key.objectid !=
7015                                btrfs_header_owner(path->nodes[level + 1]));
7016         }
7017
7018         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7019 out:
7020         wc->refs[level] = 0;
7021         wc->flags[level] = 0;
7022         return 0;
7023 }
7024
7025 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7026                                    struct btrfs_root *root,
7027                                    struct btrfs_path *path,
7028                                    struct walk_control *wc)
7029 {
7030         int level = wc->level;
7031         int lookup_info = 1;
7032         int ret;
7033
7034         while (level >= 0) {
7035                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7036                 if (ret > 0)
7037                         break;
7038
7039                 if (level == 0)
7040                         break;
7041
7042                 if (path->slots[level] >=
7043                     btrfs_header_nritems(path->nodes[level]))
7044                         break;
7045
7046                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7047                 if (ret > 0) {
7048                         path->slots[level]++;
7049                         continue;
7050                 } else if (ret < 0)
7051                         return ret;
7052                 level = wc->level;
7053         }
7054         return 0;
7055 }
7056
7057 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7058                                  struct btrfs_root *root,
7059                                  struct btrfs_path *path,
7060                                  struct walk_control *wc, int max_level)
7061 {
7062         int level = wc->level;
7063         int ret;
7064
7065         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7066         while (level < max_level && path->nodes[level]) {
7067                 wc->level = level;
7068                 if (path->slots[level] + 1 <
7069                     btrfs_header_nritems(path->nodes[level])) {
7070                         path->slots[level]++;
7071                         return 0;
7072                 } else {
7073                         ret = walk_up_proc(trans, root, path, wc);
7074                         if (ret > 0)
7075                                 return 0;
7076
7077                         if (path->locks[level]) {
7078                                 btrfs_tree_unlock_rw(path->nodes[level],
7079                                                      path->locks[level]);
7080                                 path->locks[level] = 0;
7081                         }
7082                         free_extent_buffer(path->nodes[level]);
7083                         path->nodes[level] = NULL;
7084                         level++;
7085                 }
7086         }
7087         return 1;
7088 }
7089
7090 /*
7091  * drop a subvolume tree.
7092  *
7093  * this function traverses the tree freeing any blocks that only
7094  * referenced by the tree.
7095  *
7096  * when a shared tree block is found. this function decreases its
7097  * reference count by one. if update_ref is true, this function
7098  * also make sure backrefs for the shared block and all lower level
7099  * blocks are properly updated.
7100  */
7101 int btrfs_drop_snapshot(struct btrfs_root *root,
7102                          struct btrfs_block_rsv *block_rsv, int update_ref,
7103                          int for_reloc)
7104 {
7105         struct btrfs_path *path;
7106         struct btrfs_trans_handle *trans;
7107         struct btrfs_root *tree_root = root->fs_info->tree_root;
7108         struct btrfs_root_item *root_item = &root->root_item;
7109         struct walk_control *wc;
7110         struct btrfs_key key;
7111         int err = 0;
7112         int ret;
7113         int level;
7114
7115         path = btrfs_alloc_path();
7116         if (!path) {
7117                 err = -ENOMEM;
7118                 goto out;
7119         }
7120
7121         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7122         if (!wc) {
7123                 btrfs_free_path(path);
7124                 err = -ENOMEM;
7125                 goto out;
7126         }
7127
7128         trans = btrfs_start_transaction(tree_root, 0);
7129         if (IS_ERR(trans)) {
7130                 err = PTR_ERR(trans);
7131                 goto out_free;
7132         }
7133
7134         if (block_rsv)
7135                 trans->block_rsv = block_rsv;
7136
7137         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7138                 level = btrfs_header_level(root->node);
7139                 path->nodes[level] = btrfs_lock_root_node(root);
7140                 btrfs_set_lock_blocking(path->nodes[level]);
7141                 path->slots[level] = 0;
7142                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7143                 memset(&wc->update_progress, 0,
7144                        sizeof(wc->update_progress));
7145         } else {
7146                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7147                 memcpy(&wc->update_progress, &key,
7148                        sizeof(wc->update_progress));
7149
7150                 level = root_item->drop_level;
7151                 BUG_ON(level == 0);
7152                 path->lowest_level = level;
7153                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7154                 path->lowest_level = 0;
7155                 if (ret < 0) {
7156                         err = ret;
7157                         goto out_end_trans;
7158                 }
7159                 WARN_ON(ret > 0);
7160
7161                 /*
7162                  * unlock our path, this is safe because only this
7163                  * function is allowed to delete this snapshot
7164                  */
7165                 btrfs_unlock_up_safe(path, 0);
7166
7167                 level = btrfs_header_level(root->node);
7168                 while (1) {
7169                         btrfs_tree_lock(path->nodes[level]);
7170                         btrfs_set_lock_blocking(path->nodes[level]);
7171
7172                         ret = btrfs_lookup_extent_info(trans, root,
7173                                                 path->nodes[level]->start,
7174                                                 path->nodes[level]->len,
7175                                                 &wc->refs[level],
7176                                                 &wc->flags[level]);
7177                         if (ret < 0) {
7178                                 err = ret;
7179                                 goto out_end_trans;
7180                         }
7181                         BUG_ON(wc->refs[level] == 0);
7182
7183                         if (level == root_item->drop_level)
7184                                 break;
7185
7186                         btrfs_tree_unlock(path->nodes[level]);
7187                         WARN_ON(wc->refs[level] != 1);
7188                         level--;
7189                 }
7190         }
7191
7192         wc->level = level;
7193         wc->shared_level = -1;
7194         wc->stage = DROP_REFERENCE;
7195         wc->update_ref = update_ref;
7196         wc->keep_locks = 0;
7197         wc->for_reloc = for_reloc;
7198         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7199
7200         while (1) {
7201                 ret = walk_down_tree(trans, root, path, wc);
7202                 if (ret < 0) {
7203                         err = ret;
7204                         break;
7205                 }
7206
7207                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7208                 if (ret < 0) {
7209                         err = ret;
7210                         break;
7211                 }
7212
7213                 if (ret > 0) {
7214                         BUG_ON(wc->stage != DROP_REFERENCE);
7215                         break;
7216                 }
7217
7218                 if (wc->stage == DROP_REFERENCE) {
7219                         level = wc->level;
7220                         btrfs_node_key(path->nodes[level],
7221                                        &root_item->drop_progress,
7222                                        path->slots[level]);
7223                         root_item->drop_level = level;
7224                 }
7225
7226                 BUG_ON(wc->level == 0);
7227                 if (btrfs_should_end_transaction(trans, tree_root)) {
7228                         ret = btrfs_update_root(trans, tree_root,
7229                                                 &root->root_key,
7230                                                 root_item);
7231                         if (ret) {
7232                                 btrfs_abort_transaction(trans, tree_root, ret);
7233                                 err = ret;
7234                                 goto out_end_trans;
7235                         }
7236
7237                         btrfs_end_transaction_throttle(trans, tree_root);
7238                         trans = btrfs_start_transaction(tree_root, 0);
7239                         if (IS_ERR(trans)) {
7240                                 err = PTR_ERR(trans);
7241                                 goto out_free;
7242                         }
7243                         if (block_rsv)
7244                                 trans->block_rsv = block_rsv;
7245                 }
7246         }
7247         btrfs_release_path(path);
7248         if (err)
7249                 goto out_end_trans;
7250
7251         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7252         if (ret) {
7253                 btrfs_abort_transaction(trans, tree_root, ret);
7254                 goto out_end_trans;
7255         }
7256
7257         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7258                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7259                                            NULL, NULL);
7260                 if (ret < 0) {
7261                         btrfs_abort_transaction(trans, tree_root, ret);
7262                         err = ret;
7263                         goto out_end_trans;
7264                 } else if (ret > 0) {
7265                         /* if we fail to delete the orphan item this time
7266                          * around, it'll get picked up the next time.
7267                          *
7268                          * The most common failure here is just -ENOENT.
7269                          */
7270                         btrfs_del_orphan_item(trans, tree_root,
7271                                               root->root_key.objectid);
7272                 }
7273         }
7274
7275         if (root->in_radix) {
7276                 btrfs_free_fs_root(tree_root->fs_info, root);
7277         } else {
7278                 free_extent_buffer(root->node);
7279                 free_extent_buffer(root->commit_root);
7280                 kfree(root);
7281         }
7282 out_end_trans:
7283         btrfs_end_transaction_throttle(trans, tree_root);
7284 out_free:
7285         kfree(wc);
7286         btrfs_free_path(path);
7287 out:
7288         if (err)
7289                 btrfs_std_error(root->fs_info, err);
7290         return err;
7291 }
7292
7293 /*
7294  * drop subtree rooted at tree block 'node'.
7295  *
7296  * NOTE: this function will unlock and release tree block 'node'
7297  * only used by relocation code
7298  */
7299 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7300                         struct btrfs_root *root,
7301                         struct extent_buffer *node,
7302                         struct extent_buffer *parent)
7303 {
7304         struct btrfs_path *path;
7305         struct walk_control *wc;
7306         int level;
7307         int parent_level;
7308         int ret = 0;
7309         int wret;
7310
7311         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7312
7313         path = btrfs_alloc_path();
7314         if (!path)
7315                 return -ENOMEM;
7316
7317         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7318         if (!wc) {
7319                 btrfs_free_path(path);
7320                 return -ENOMEM;
7321         }
7322
7323         btrfs_assert_tree_locked(parent);
7324         parent_level = btrfs_header_level(parent);
7325         extent_buffer_get(parent);
7326         path->nodes[parent_level] = parent;
7327         path->slots[parent_level] = btrfs_header_nritems(parent);
7328
7329         btrfs_assert_tree_locked(node);
7330         level = btrfs_header_level(node);
7331         path->nodes[level] = node;
7332         path->slots[level] = 0;
7333         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7334
7335         wc->refs[parent_level] = 1;
7336         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7337         wc->level = level;
7338         wc->shared_level = -1;
7339         wc->stage = DROP_REFERENCE;
7340         wc->update_ref = 0;
7341         wc->keep_locks = 1;
7342         wc->for_reloc = 1;
7343         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7344
7345         while (1) {
7346                 wret = walk_down_tree(trans, root, path, wc);
7347                 if (wret < 0) {
7348                         ret = wret;
7349                         break;
7350                 }
7351
7352                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7353                 if (wret < 0)
7354                         ret = wret;
7355                 if (wret != 0)
7356                         break;
7357         }
7358
7359         kfree(wc);
7360         btrfs_free_path(path);
7361         return ret;
7362 }
7363
7364 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7365 {
7366         u64 num_devices;
7367         u64 stripped;
7368
7369         /*
7370          * if restripe for this chunk_type is on pick target profile and
7371          * return, otherwise do the usual balance
7372          */
7373         stripped = get_restripe_target(root->fs_info, flags);
7374         if (stripped)
7375                 return extended_to_chunk(stripped);
7376
7377         /*
7378          * we add in the count of missing devices because we want
7379          * to make sure that any RAID levels on a degraded FS
7380          * continue to be honored.
7381          */
7382         num_devices = root->fs_info->fs_devices->rw_devices +
7383                 root->fs_info->fs_devices->missing_devices;
7384
7385         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7386                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7387                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7388
7389         if (num_devices == 1) {
7390                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7391                 stripped = flags & ~stripped;
7392
7393                 /* turn raid0 into single device chunks */
7394                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7395                         return stripped;
7396
7397                 /* turn mirroring into duplication */
7398                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7399                              BTRFS_BLOCK_GROUP_RAID10))
7400                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7401         } else {
7402                 /* they already had raid on here, just return */
7403                 if (flags & stripped)
7404                         return flags;
7405
7406                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7407                 stripped = flags & ~stripped;
7408
7409                 /* switch duplicated blocks with raid1 */
7410                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7411                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7412
7413                 /* this is drive concat, leave it alone */
7414         }
7415
7416         return flags;
7417 }
7418
7419 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7420 {
7421         struct btrfs_space_info *sinfo = cache->space_info;
7422         u64 num_bytes;
7423         u64 min_allocable_bytes;
7424         int ret = -ENOSPC;
7425
7426
7427         /*
7428          * We need some metadata space and system metadata space for
7429          * allocating chunks in some corner cases until we force to set
7430          * it to be readonly.
7431          */
7432         if ((sinfo->flags &
7433              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7434             !force)
7435                 min_allocable_bytes = 1 * 1024 * 1024;
7436         else
7437                 min_allocable_bytes = 0;
7438
7439         spin_lock(&sinfo->lock);
7440         spin_lock(&cache->lock);
7441
7442         if (cache->ro) {
7443                 ret = 0;
7444                 goto out;
7445         }
7446
7447         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7448                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7449
7450         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7451             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7452             min_allocable_bytes <= sinfo->total_bytes) {
7453                 sinfo->bytes_readonly += num_bytes;
7454                 cache->ro = 1;
7455                 ret = 0;
7456         }
7457 out:
7458         spin_unlock(&cache->lock);
7459         spin_unlock(&sinfo->lock);
7460         return ret;
7461 }
7462
7463 int btrfs_set_block_group_ro(struct btrfs_root *root,
7464                              struct btrfs_block_group_cache *cache)
7465
7466 {
7467         struct btrfs_trans_handle *trans;
7468         u64 alloc_flags;
7469         int ret;
7470
7471         BUG_ON(cache->ro);
7472
7473         trans = btrfs_join_transaction(root);
7474         if (IS_ERR(trans))
7475                 return PTR_ERR(trans);
7476
7477         alloc_flags = update_block_group_flags(root, cache->flags);
7478         if (alloc_flags != cache->flags) {
7479                 ret = do_chunk_alloc(trans, root, alloc_flags,
7480                                      CHUNK_ALLOC_FORCE);
7481                 if (ret < 0)
7482                         goto out;
7483         }
7484
7485         ret = set_block_group_ro(cache, 0);
7486         if (!ret)
7487                 goto out;
7488         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7489         ret = do_chunk_alloc(trans, root, alloc_flags,
7490                              CHUNK_ALLOC_FORCE);
7491         if (ret < 0)
7492                 goto out;
7493         ret = set_block_group_ro(cache, 0);
7494 out:
7495         btrfs_end_transaction(trans, root);
7496         return ret;
7497 }
7498
7499 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7500                             struct btrfs_root *root, u64 type)
7501 {
7502         u64 alloc_flags = get_alloc_profile(root, type);
7503         return do_chunk_alloc(trans, root, alloc_flags,
7504                               CHUNK_ALLOC_FORCE);
7505 }
7506
7507 /*
7508  * helper to account the unused space of all the readonly block group in the
7509  * list. takes mirrors into account.
7510  */
7511 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7512 {
7513         struct btrfs_block_group_cache *block_group;
7514         u64 free_bytes = 0;
7515         int factor;
7516
7517         list_for_each_entry(block_group, groups_list, list) {
7518                 spin_lock(&block_group->lock);
7519
7520                 if (!block_group->ro) {
7521                         spin_unlock(&block_group->lock);
7522                         continue;
7523                 }
7524
7525                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7526                                           BTRFS_BLOCK_GROUP_RAID10 |
7527                                           BTRFS_BLOCK_GROUP_DUP))
7528                         factor = 2;
7529                 else
7530                         factor = 1;
7531
7532                 free_bytes += (block_group->key.offset -
7533                                btrfs_block_group_used(&block_group->item)) *
7534                                factor;
7535
7536                 spin_unlock(&block_group->lock);
7537         }
7538
7539         return free_bytes;
7540 }
7541
7542 /*
7543  * helper to account the unused space of all the readonly block group in the
7544  * space_info. takes mirrors into account.
7545  */
7546 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7547 {
7548         int i;
7549         u64 free_bytes = 0;
7550
7551         spin_lock(&sinfo->lock);
7552
7553         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7554                 if (!list_empty(&sinfo->block_groups[i]))
7555                         free_bytes += __btrfs_get_ro_block_group_free_space(
7556                                                 &sinfo->block_groups[i]);
7557
7558         spin_unlock(&sinfo->lock);
7559
7560         return free_bytes;
7561 }
7562
7563 void btrfs_set_block_group_rw(struct btrfs_root *root,
7564                               struct btrfs_block_group_cache *cache)
7565 {
7566         struct btrfs_space_info *sinfo = cache->space_info;
7567         u64 num_bytes;
7568
7569         BUG_ON(!cache->ro);
7570
7571         spin_lock(&sinfo->lock);
7572         spin_lock(&cache->lock);
7573         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7574                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7575         sinfo->bytes_readonly -= num_bytes;
7576         cache->ro = 0;
7577         spin_unlock(&cache->lock);
7578         spin_unlock(&sinfo->lock);
7579 }
7580
7581 /*
7582  * checks to see if its even possible to relocate this block group.
7583  *
7584  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7585  * ok to go ahead and try.
7586  */
7587 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7588 {
7589         struct btrfs_block_group_cache *block_group;
7590         struct btrfs_space_info *space_info;
7591         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7592         struct btrfs_device *device;
7593         u64 min_free;
7594         u64 dev_min = 1;
7595         u64 dev_nr = 0;
7596         u64 target;
7597         int index;
7598         int full = 0;
7599         int ret = 0;
7600
7601         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7602
7603         /* odd, couldn't find the block group, leave it alone */
7604         if (!block_group)
7605                 return -1;
7606
7607         min_free = btrfs_block_group_used(&block_group->item);
7608
7609         /* no bytes used, we're good */
7610         if (!min_free)
7611                 goto out;
7612
7613         space_info = block_group->space_info;
7614         spin_lock(&space_info->lock);
7615
7616         full = space_info->full;
7617
7618         /*
7619          * if this is the last block group we have in this space, we can't
7620          * relocate it unless we're able to allocate a new chunk below.
7621          *
7622          * Otherwise, we need to make sure we have room in the space to handle
7623          * all of the extents from this block group.  If we can, we're good
7624          */
7625         if ((space_info->total_bytes != block_group->key.offset) &&
7626             (space_info->bytes_used + space_info->bytes_reserved +
7627              space_info->bytes_pinned + space_info->bytes_readonly +
7628              min_free < space_info->total_bytes)) {
7629                 spin_unlock(&space_info->lock);
7630                 goto out;
7631         }
7632         spin_unlock(&space_info->lock);
7633
7634         /*
7635          * ok we don't have enough space, but maybe we have free space on our
7636          * devices to allocate new chunks for relocation, so loop through our
7637          * alloc devices and guess if we have enough space.  if this block
7638          * group is going to be restriped, run checks against the target
7639          * profile instead of the current one.
7640          */
7641         ret = -1;
7642
7643         /*
7644          * index:
7645          *      0: raid10
7646          *      1: raid1
7647          *      2: dup
7648          *      3: raid0
7649          *      4: single
7650          */
7651         target = get_restripe_target(root->fs_info, block_group->flags);
7652         if (target) {
7653                 index = __get_raid_index(extended_to_chunk(target));
7654         } else {
7655                 /*
7656                  * this is just a balance, so if we were marked as full
7657                  * we know there is no space for a new chunk
7658                  */
7659                 if (full)
7660                         goto out;
7661
7662                 index = get_block_group_index(block_group);
7663         }
7664
7665         if (index == BTRFS_RAID_RAID10) {
7666                 dev_min = 4;
7667                 /* Divide by 2 */
7668                 min_free >>= 1;
7669         } else if (index == BTRFS_RAID_RAID1) {
7670                 dev_min = 2;
7671         } else if (index == BTRFS_RAID_DUP) {
7672                 /* Multiply by 2 */
7673                 min_free <<= 1;
7674         } else if (index == BTRFS_RAID_RAID0) {
7675                 dev_min = fs_devices->rw_devices;
7676                 do_div(min_free, dev_min);
7677         }
7678
7679         mutex_lock(&root->fs_info->chunk_mutex);
7680         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7681                 u64 dev_offset;
7682
7683                 /*
7684                  * check to make sure we can actually find a chunk with enough
7685                  * space to fit our block group in.
7686                  */
7687                 if (device->total_bytes > device->bytes_used + min_free &&
7688                     !device->is_tgtdev_for_dev_replace) {
7689                         ret = find_free_dev_extent(device, min_free,
7690                                                    &dev_offset, NULL);
7691                         if (!ret)
7692                                 dev_nr++;
7693
7694                         if (dev_nr >= dev_min)
7695                                 break;
7696
7697                         ret = -1;
7698                 }
7699         }
7700         mutex_unlock(&root->fs_info->chunk_mutex);
7701 out:
7702         btrfs_put_block_group(block_group);
7703         return ret;
7704 }
7705
7706 static int find_first_block_group(struct btrfs_root *root,
7707                 struct btrfs_path *path, struct btrfs_key *key)
7708 {
7709         int ret = 0;
7710         struct btrfs_key found_key;
7711         struct extent_buffer *leaf;
7712         int slot;
7713
7714         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7715         if (ret < 0)
7716                 goto out;
7717
7718         while (1) {
7719                 slot = path->slots[0];
7720                 leaf = path->nodes[0];
7721                 if (slot >= btrfs_header_nritems(leaf)) {
7722                         ret = btrfs_next_leaf(root, path);
7723                         if (ret == 0)
7724                                 continue;
7725                         if (ret < 0)
7726                                 goto out;
7727                         break;
7728                 }
7729                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7730
7731                 if (found_key.objectid >= key->objectid &&
7732                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7733                         ret = 0;
7734                         goto out;
7735                 }
7736                 path->slots[0]++;
7737         }
7738 out:
7739         return ret;
7740 }
7741
7742 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7743 {
7744         struct btrfs_block_group_cache *block_group;
7745         u64 last = 0;
7746
7747         while (1) {
7748                 struct inode *inode;
7749
7750                 block_group = btrfs_lookup_first_block_group(info, last);
7751                 while (block_group) {
7752                         spin_lock(&block_group->lock);
7753                         if (block_group->iref)
7754                                 break;
7755                         spin_unlock(&block_group->lock);
7756                         block_group = next_block_group(info->tree_root,
7757                                                        block_group);
7758                 }
7759                 if (!block_group) {
7760                         if (last == 0)
7761                                 break;
7762                         last = 0;
7763                         continue;
7764                 }
7765
7766                 inode = block_group->inode;
7767                 block_group->iref = 0;
7768                 block_group->inode = NULL;
7769                 spin_unlock(&block_group->lock);
7770                 iput(inode);
7771                 last = block_group->key.objectid + block_group->key.offset;
7772                 btrfs_put_block_group(block_group);
7773         }
7774 }
7775
7776 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7777 {
7778         struct btrfs_block_group_cache *block_group;
7779         struct btrfs_space_info *space_info;
7780         struct btrfs_caching_control *caching_ctl;
7781         struct rb_node *n;
7782
7783         down_write(&info->extent_commit_sem);
7784         while (!list_empty(&info->caching_block_groups)) {
7785                 caching_ctl = list_entry(info->caching_block_groups.next,
7786                                          struct btrfs_caching_control, list);
7787                 list_del(&caching_ctl->list);
7788                 put_caching_control(caching_ctl);
7789         }
7790         up_write(&info->extent_commit_sem);
7791
7792         spin_lock(&info->block_group_cache_lock);
7793         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7794                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7795                                        cache_node);
7796                 rb_erase(&block_group->cache_node,
7797                          &info->block_group_cache_tree);
7798                 spin_unlock(&info->block_group_cache_lock);
7799
7800                 down_write(&block_group->space_info->groups_sem);
7801                 list_del(&block_group->list);
7802                 up_write(&block_group->space_info->groups_sem);
7803
7804                 if (block_group->cached == BTRFS_CACHE_STARTED)
7805                         wait_block_group_cache_done(block_group);
7806
7807                 /*
7808                  * We haven't cached this block group, which means we could
7809                  * possibly have excluded extents on this block group.
7810                  */
7811                 if (block_group->cached == BTRFS_CACHE_NO)
7812                         free_excluded_extents(info->extent_root, block_group);
7813
7814                 btrfs_remove_free_space_cache(block_group);
7815                 btrfs_put_block_group(block_group);
7816
7817                 spin_lock(&info->block_group_cache_lock);
7818         }
7819         spin_unlock(&info->block_group_cache_lock);
7820
7821         /* now that all the block groups are freed, go through and
7822          * free all the space_info structs.  This is only called during
7823          * the final stages of unmount, and so we know nobody is
7824          * using them.  We call synchronize_rcu() once before we start,
7825          * just to be on the safe side.
7826          */
7827         synchronize_rcu();
7828
7829         release_global_block_rsv(info);
7830
7831         while(!list_empty(&info->space_info)) {
7832                 space_info = list_entry(info->space_info.next,
7833                                         struct btrfs_space_info,
7834                                         list);
7835                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
7836                         if (space_info->bytes_pinned > 0 ||
7837                             space_info->bytes_reserved > 0 ||
7838                             space_info->bytes_may_use > 0) {
7839                                 WARN_ON(1);
7840                                 dump_space_info(space_info, 0, 0);
7841                         }
7842                 }
7843                 list_del(&space_info->list);
7844                 kfree(space_info);
7845         }
7846         return 0;
7847 }
7848
7849 static void __link_block_group(struct btrfs_space_info *space_info,
7850                                struct btrfs_block_group_cache *cache)
7851 {
7852         int index = get_block_group_index(cache);
7853
7854         down_write(&space_info->groups_sem);
7855         list_add_tail(&cache->list, &space_info->block_groups[index]);
7856         up_write(&space_info->groups_sem);
7857 }
7858
7859 int btrfs_read_block_groups(struct btrfs_root *root)
7860 {
7861         struct btrfs_path *path;
7862         int ret;
7863         struct btrfs_block_group_cache *cache;
7864         struct btrfs_fs_info *info = root->fs_info;
7865         struct btrfs_space_info *space_info;
7866         struct btrfs_key key;
7867         struct btrfs_key found_key;
7868         struct extent_buffer *leaf;
7869         int need_clear = 0;
7870         u64 cache_gen;
7871
7872         root = info->extent_root;
7873         key.objectid = 0;
7874         key.offset = 0;
7875         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7876         path = btrfs_alloc_path();
7877         if (!path)
7878                 return -ENOMEM;
7879         path->reada = 1;
7880
7881         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7882         if (btrfs_test_opt(root, SPACE_CACHE) &&
7883             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7884                 need_clear = 1;
7885         if (btrfs_test_opt(root, CLEAR_CACHE))
7886                 need_clear = 1;
7887
7888         while (1) {
7889                 ret = find_first_block_group(root, path, &key);
7890                 if (ret > 0)
7891                         break;
7892                 if (ret != 0)
7893                         goto error;
7894                 leaf = path->nodes[0];
7895                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7896                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7897                 if (!cache) {
7898                         ret = -ENOMEM;
7899                         goto error;
7900                 }
7901                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7902                                                 GFP_NOFS);
7903                 if (!cache->free_space_ctl) {
7904                         kfree(cache);
7905                         ret = -ENOMEM;
7906                         goto error;
7907                 }
7908
7909                 atomic_set(&cache->count, 1);
7910                 spin_lock_init(&cache->lock);
7911                 cache->fs_info = info;
7912                 INIT_LIST_HEAD(&cache->list);
7913                 INIT_LIST_HEAD(&cache->cluster_list);
7914
7915                 if (need_clear) {
7916                         /*
7917                          * When we mount with old space cache, we need to
7918                          * set BTRFS_DC_CLEAR and set dirty flag.
7919                          *
7920                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7921                          *    truncate the old free space cache inode and
7922                          *    setup a new one.
7923                          * b) Setting 'dirty flag' makes sure that we flush
7924                          *    the new space cache info onto disk.
7925                          */
7926                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7927                         if (btrfs_test_opt(root, SPACE_CACHE))
7928                                 cache->dirty = 1;
7929                 }
7930
7931                 read_extent_buffer(leaf, &cache->item,
7932                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7933                                    sizeof(cache->item));
7934                 memcpy(&cache->key, &found_key, sizeof(found_key));
7935
7936                 key.objectid = found_key.objectid + found_key.offset;
7937                 btrfs_release_path(path);
7938                 cache->flags = btrfs_block_group_flags(&cache->item);
7939                 cache->sectorsize = root->sectorsize;
7940                 cache->full_stripe_len = btrfs_full_stripe_len(root,
7941                                                &root->fs_info->mapping_tree,
7942                                                found_key.objectid);
7943                 btrfs_init_free_space_ctl(cache);
7944
7945                 /*
7946                  * We need to exclude the super stripes now so that the space
7947                  * info has super bytes accounted for, otherwise we'll think
7948                  * we have more space than we actually do.
7949                  */
7950                 exclude_super_stripes(root, cache);
7951
7952                 /*
7953                  * check for two cases, either we are full, and therefore
7954                  * don't need to bother with the caching work since we won't
7955                  * find any space, or we are empty, and we can just add all
7956                  * the space in and be done with it.  This saves us _alot_ of
7957                  * time, particularly in the full case.
7958                  */
7959                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7960                         cache->last_byte_to_unpin = (u64)-1;
7961                         cache->cached = BTRFS_CACHE_FINISHED;
7962                         free_excluded_extents(root, cache);
7963                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7964                         cache->last_byte_to_unpin = (u64)-1;
7965                         cache->cached = BTRFS_CACHE_FINISHED;
7966                         add_new_free_space(cache, root->fs_info,
7967                                            found_key.objectid,
7968                                            found_key.objectid +
7969                                            found_key.offset);
7970                         free_excluded_extents(root, cache);
7971                 }
7972
7973                 ret = update_space_info(info, cache->flags, found_key.offset,
7974                                         btrfs_block_group_used(&cache->item),
7975                                         &space_info);
7976                 BUG_ON(ret); /* -ENOMEM */
7977                 cache->space_info = space_info;
7978                 spin_lock(&cache->space_info->lock);
7979                 cache->space_info->bytes_readonly += cache->bytes_super;
7980                 spin_unlock(&cache->space_info->lock);
7981
7982                 __link_block_group(space_info, cache);
7983
7984                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7985                 BUG_ON(ret); /* Logic error */
7986
7987                 set_avail_alloc_bits(root->fs_info, cache->flags);
7988                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7989                         set_block_group_ro(cache, 1);
7990         }
7991
7992         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7993                 if (!(get_alloc_profile(root, space_info->flags) &
7994                       (BTRFS_BLOCK_GROUP_RAID10 |
7995                        BTRFS_BLOCK_GROUP_RAID1 |
7996                        BTRFS_BLOCK_GROUP_RAID5 |
7997                        BTRFS_BLOCK_GROUP_RAID6 |
7998                        BTRFS_BLOCK_GROUP_DUP)))
7999                         continue;
8000                 /*
8001                  * avoid allocating from un-mirrored block group if there are
8002                  * mirrored block groups.
8003                  */
8004                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8005                         set_block_group_ro(cache, 1);
8006                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8007                         set_block_group_ro(cache, 1);
8008         }
8009
8010         init_global_block_rsv(info);
8011         ret = 0;
8012 error:
8013         btrfs_free_path(path);
8014         return ret;
8015 }
8016
8017 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8018                                        struct btrfs_root *root)
8019 {
8020         struct btrfs_block_group_cache *block_group, *tmp;
8021         struct btrfs_root *extent_root = root->fs_info->extent_root;
8022         struct btrfs_block_group_item item;
8023         struct btrfs_key key;
8024         int ret = 0;
8025
8026         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8027                                  new_bg_list) {
8028                 list_del_init(&block_group->new_bg_list);
8029
8030                 if (ret)
8031                         continue;
8032
8033                 spin_lock(&block_group->lock);
8034                 memcpy(&item, &block_group->item, sizeof(item));
8035                 memcpy(&key, &block_group->key, sizeof(key));
8036                 spin_unlock(&block_group->lock);
8037
8038                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8039                                         sizeof(item));
8040                 if (ret)
8041                         btrfs_abort_transaction(trans, extent_root, ret);
8042         }
8043 }
8044
8045 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8046                            struct btrfs_root *root, u64 bytes_used,
8047                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8048                            u64 size)
8049 {
8050         int ret;
8051         struct btrfs_root *extent_root;
8052         struct btrfs_block_group_cache *cache;
8053
8054         extent_root = root->fs_info->extent_root;
8055
8056         root->fs_info->last_trans_log_full_commit = trans->transid;
8057
8058         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8059         if (!cache)
8060                 return -ENOMEM;
8061         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8062                                         GFP_NOFS);
8063         if (!cache->free_space_ctl) {
8064                 kfree(cache);
8065                 return -ENOMEM;
8066         }
8067
8068         cache->key.objectid = chunk_offset;
8069         cache->key.offset = size;
8070         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8071         cache->sectorsize = root->sectorsize;
8072         cache->fs_info = root->fs_info;
8073         cache->full_stripe_len = btrfs_full_stripe_len(root,
8074                                                &root->fs_info->mapping_tree,
8075                                                chunk_offset);
8076
8077         atomic_set(&cache->count, 1);
8078         spin_lock_init(&cache->lock);
8079         INIT_LIST_HEAD(&cache->list);
8080         INIT_LIST_HEAD(&cache->cluster_list);
8081         INIT_LIST_HEAD(&cache->new_bg_list);
8082
8083         btrfs_init_free_space_ctl(cache);
8084
8085         btrfs_set_block_group_used(&cache->item, bytes_used);
8086         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8087         cache->flags = type;
8088         btrfs_set_block_group_flags(&cache->item, type);
8089
8090         cache->last_byte_to_unpin = (u64)-1;
8091         cache->cached = BTRFS_CACHE_FINISHED;
8092         exclude_super_stripes(root, cache);
8093
8094         add_new_free_space(cache, root->fs_info, chunk_offset,
8095                            chunk_offset + size);
8096
8097         free_excluded_extents(root, cache);
8098
8099         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8100                                 &cache->space_info);
8101         BUG_ON(ret); /* -ENOMEM */
8102         update_global_block_rsv(root->fs_info);
8103
8104         spin_lock(&cache->space_info->lock);
8105         cache->space_info->bytes_readonly += cache->bytes_super;
8106         spin_unlock(&cache->space_info->lock);
8107
8108         __link_block_group(cache->space_info, cache);
8109
8110         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8111         BUG_ON(ret); /* Logic error */
8112
8113         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8114
8115         set_avail_alloc_bits(extent_root->fs_info, type);
8116
8117         return 0;
8118 }
8119
8120 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8121 {
8122         u64 extra_flags = chunk_to_extended(flags) &
8123                                 BTRFS_EXTENDED_PROFILE_MASK;
8124
8125         write_seqlock(&fs_info->profiles_lock);
8126         if (flags & BTRFS_BLOCK_GROUP_DATA)
8127                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8128         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8129                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8130         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8131                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8132         write_sequnlock(&fs_info->profiles_lock);
8133 }
8134
8135 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8136                              struct btrfs_root *root, u64 group_start)
8137 {
8138         struct btrfs_path *path;
8139         struct btrfs_block_group_cache *block_group;
8140         struct btrfs_free_cluster *cluster;
8141         struct btrfs_root *tree_root = root->fs_info->tree_root;
8142         struct btrfs_key key;
8143         struct inode *inode;
8144         int ret;
8145         int index;
8146         int factor;
8147
8148         root = root->fs_info->extent_root;
8149
8150         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8151         BUG_ON(!block_group);
8152         BUG_ON(!block_group->ro);
8153
8154         /*
8155          * Free the reserved super bytes from this block group before
8156          * remove it.
8157          */
8158         free_excluded_extents(root, block_group);
8159
8160         memcpy(&key, &block_group->key, sizeof(key));
8161         index = get_block_group_index(block_group);
8162         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8163                                   BTRFS_BLOCK_GROUP_RAID1 |
8164                                   BTRFS_BLOCK_GROUP_RAID10))
8165                 factor = 2;
8166         else
8167                 factor = 1;
8168
8169         /* make sure this block group isn't part of an allocation cluster */
8170         cluster = &root->fs_info->data_alloc_cluster;
8171         spin_lock(&cluster->refill_lock);
8172         btrfs_return_cluster_to_free_space(block_group, cluster);
8173         spin_unlock(&cluster->refill_lock);
8174
8175         /*
8176          * make sure this block group isn't part of a metadata
8177          * allocation cluster
8178          */
8179         cluster = &root->fs_info->meta_alloc_cluster;
8180         spin_lock(&cluster->refill_lock);
8181         btrfs_return_cluster_to_free_space(block_group, cluster);
8182         spin_unlock(&cluster->refill_lock);
8183
8184         path = btrfs_alloc_path();
8185         if (!path) {
8186                 ret = -ENOMEM;
8187                 goto out;
8188         }
8189
8190         inode = lookup_free_space_inode(tree_root, block_group, path);
8191         if (!IS_ERR(inode)) {
8192                 ret = btrfs_orphan_add(trans, inode);
8193                 if (ret) {
8194                         btrfs_add_delayed_iput(inode);
8195                         goto out;
8196                 }
8197                 clear_nlink(inode);
8198                 /* One for the block groups ref */
8199                 spin_lock(&block_group->lock);
8200                 if (block_group->iref) {
8201                         block_group->iref = 0;
8202                         block_group->inode = NULL;
8203                         spin_unlock(&block_group->lock);
8204                         iput(inode);
8205                 } else {
8206                         spin_unlock(&block_group->lock);
8207                 }
8208                 /* One for our lookup ref */
8209                 btrfs_add_delayed_iput(inode);
8210         }
8211
8212         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8213         key.offset = block_group->key.objectid;
8214         key.type = 0;
8215
8216         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8217         if (ret < 0)
8218                 goto out;
8219         if (ret > 0)
8220                 btrfs_release_path(path);
8221         if (ret == 0) {
8222                 ret = btrfs_del_item(trans, tree_root, path);
8223                 if (ret)
8224                         goto out;
8225                 btrfs_release_path(path);
8226         }
8227
8228         spin_lock(&root->fs_info->block_group_cache_lock);
8229         rb_erase(&block_group->cache_node,
8230                  &root->fs_info->block_group_cache_tree);
8231
8232         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8233                 root->fs_info->first_logical_byte = (u64)-1;
8234         spin_unlock(&root->fs_info->block_group_cache_lock);
8235
8236         down_write(&block_group->space_info->groups_sem);
8237         /*
8238          * we must use list_del_init so people can check to see if they
8239          * are still on the list after taking the semaphore
8240          */
8241         list_del_init(&block_group->list);
8242         if (list_empty(&block_group->space_info->block_groups[index]))
8243                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8244         up_write(&block_group->space_info->groups_sem);
8245
8246         if (block_group->cached == BTRFS_CACHE_STARTED)
8247                 wait_block_group_cache_done(block_group);
8248
8249         btrfs_remove_free_space_cache(block_group);
8250
8251         spin_lock(&block_group->space_info->lock);
8252         block_group->space_info->total_bytes -= block_group->key.offset;
8253         block_group->space_info->bytes_readonly -= block_group->key.offset;
8254         block_group->space_info->disk_total -= block_group->key.offset * factor;
8255         spin_unlock(&block_group->space_info->lock);
8256
8257         memcpy(&key, &block_group->key, sizeof(key));
8258
8259         btrfs_clear_space_info_full(root->fs_info);
8260
8261         btrfs_put_block_group(block_group);
8262         btrfs_put_block_group(block_group);
8263
8264         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8265         if (ret > 0)
8266                 ret = -EIO;
8267         if (ret < 0)
8268                 goto out;
8269
8270         ret = btrfs_del_item(trans, root, path);
8271 out:
8272         btrfs_free_path(path);
8273         return ret;
8274 }
8275
8276 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8277 {
8278         struct btrfs_space_info *space_info;
8279         struct btrfs_super_block *disk_super;
8280         u64 features;
8281         u64 flags;
8282         int mixed = 0;
8283         int ret;
8284
8285         disk_super = fs_info->super_copy;
8286         if (!btrfs_super_root(disk_super))
8287                 return 1;
8288
8289         features = btrfs_super_incompat_flags(disk_super);
8290         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8291                 mixed = 1;
8292
8293         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8294         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8295         if (ret)
8296                 goto out;
8297
8298         if (mixed) {
8299                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8300                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8301         } else {
8302                 flags = BTRFS_BLOCK_GROUP_METADATA;
8303                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8304                 if (ret)
8305                         goto out;
8306
8307                 flags = BTRFS_BLOCK_GROUP_DATA;
8308                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8309         }
8310 out:
8311         return ret;
8312 }
8313
8314 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8315 {
8316         return unpin_extent_range(root, start, end);
8317 }
8318
8319 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8320                                u64 num_bytes, u64 *actual_bytes)
8321 {
8322         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8323 }
8324
8325 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8326 {
8327         struct btrfs_fs_info *fs_info = root->fs_info;
8328         struct btrfs_block_group_cache *cache = NULL;
8329         u64 group_trimmed;
8330         u64 start;
8331         u64 end;
8332         u64 trimmed = 0;
8333         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8334         int ret = 0;
8335
8336         /*
8337          * try to trim all FS space, our block group may start from non-zero.
8338          */
8339         if (range->len == total_bytes)
8340                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8341         else
8342                 cache = btrfs_lookup_block_group(fs_info, range->start);
8343
8344         while (cache) {
8345                 if (cache->key.objectid >= (range->start + range->len)) {
8346                         btrfs_put_block_group(cache);
8347                         break;
8348                 }
8349
8350                 start = max(range->start, cache->key.objectid);
8351                 end = min(range->start + range->len,
8352                                 cache->key.objectid + cache->key.offset);
8353
8354                 if (end - start >= range->minlen) {
8355                         if (!block_group_cache_done(cache)) {
8356                                 ret = cache_block_group(cache, 0);
8357                                 if (!ret)
8358                                         wait_block_group_cache_done(cache);
8359                         }
8360                         ret = btrfs_trim_block_group(cache,
8361                                                      &group_trimmed,
8362                                                      start,
8363                                                      end,
8364                                                      range->minlen);
8365
8366                         trimmed += group_trimmed;
8367                         if (ret) {
8368                                 btrfs_put_block_group(cache);
8369                                 break;
8370                         }
8371                 }
8372
8373                 cache = next_block_group(fs_info->tree_root, cache);
8374         }
8375
8376         range->len = trimmed;
8377         return ret;
8378 }