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 int btrfs_pin_extent(struct btrfs_root *root,
 109                      u64 bytenr, u64 num_bytes, int reserved);
 110
 111 static noinline int
 112 block_group_cache_done(struct btrfs_block_group_cache *cache)
 113 {
 114         smp_mb();
 115         return cache->cached == BTRFS_CACHE_FINISHED;
 116 }
 117
 118 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 119 {
 120         return (cache->flags & bits) == bits;
 121 }
 122
 123 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         atomic_inc(&cache->count);
 126 }
 127
 128 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         if (atomic_dec_and_test(&cache->count)) {
 131                 WARN_ON(cache->pinned > 0);
 132                 WARN_ON(cache->reserved > 0);
 133                 kfree(cache->free_space_ctl);
 134                 kfree(cache);
 135         }
 136 }
 137
 138 /*
 139  * this adds the block group to the fs_info rb tree for the block group
 140  * cache
 141  */
 142 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 143                                 struct btrfs_block_group_cache *block_group)
 144 {
 145         struct rb_node **p;
 146         struct rb_node *parent = NULL;
 147         struct btrfs_block_group_cache *cache;
 148
 149         spin_lock(&info->block_group_cache_lock);
 150         p = &info->block_group_cache_tree.rb_node;
 151
 152         while (*p) {
 153                 parent = *p;
 154                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 155                                  cache_node);
 156                 if (block_group->key.objectid < cache->key.objectid) {
 157                         p = &(*p)->rb_left;
 158                 } else if (block_group->key.objectid > cache->key.objectid) {
 159                         p = &(*p)->rb_right;
 160                 } else {
 161                         spin_unlock(&info->block_group_cache_lock);
 162                         return -EEXIST;
 163                 }
 164         }
 165
 166         rb_link_node(&block_group->cache_node, parent, p);
 167         rb_insert_color(&block_group->cache_node,
 168                         &info->block_group_cache_tree);
 169
 170         if (info->first_logical_byte > block_group->key.objectid)
 171                 info->first_logical_byte = block_group->key.objectid;
 172
 173         spin_unlock(&info->block_group_cache_lock);
 174
 175         return 0;
 176 }
 177
 178 /*
 179  * This will return the block group at or after bytenr if contains is 0, else
 180  * it will return the block group that contains the bytenr
 181  */
 182 static struct btrfs_block_group_cache *
 183 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 184                               int contains)
 185 {
 186         struct btrfs_block_group_cache *cache, *ret = NULL;
 187         struct rb_node *n;
 188         u64 end, start;
 189
 190         spin_lock(&info->block_group_cache_lock);
 191         n = info->block_group_cache_tree.rb_node;
 192
 193         while (n) {
 194                 cache = rb_entry(n, struct btrfs_block_group_cache,
 195                                  cache_node);
 196                 end = cache->key.objectid + cache->key.offset - 1;
 197                 start = cache->key.objectid;
 198
 199                 if (bytenr < start) {
 200                         if (!contains && (!ret || start < ret->key.objectid))
 201                                 ret = cache;
 202                         n = n->rb_left;
 203                 } else if (bytenr > start) {
 204                         if (contains && bytenr <= end) {
 205                                 ret = cache;
 206                                 break;
 207                         }
 208                         n = n->rb_right;
 209                 } else {
 210                         ret = cache;
 211                         break;
 212                 }
 213         }
 214         if (ret) {
 215                 btrfs_get_block_group(ret);
 216                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 217                         info->first_logical_byte = ret->key.objectid;
 218         }
 219         spin_unlock(&info->block_group_cache_lock);
 220
 221         return ret;
 222 }
 223
 224 static int add_excluded_extent(struct btrfs_root *root,
 225                                u64 start, u64 num_bytes)
 226 {
 227         u64 end = start + num_bytes - 1;
 228         set_extent_bits(&root->fs_info->freed_extents[0],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         set_extent_bits(&root->fs_info->freed_extents[1],
 231                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 232         return 0;
 233 }
 234
 235 static void free_excluded_extents(struct btrfs_root *root,
 236                                   struct btrfs_block_group_cache *cache)
 237 {
 238         u64 start, end;
 239
 240         start = cache->key.objectid;
 241         end = start + cache->key.offset - 1;
 242
 243         clear_extent_bits(&root->fs_info->freed_extents[0],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245         clear_extent_bits(&root->fs_info->freed_extents[1],
 246                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 247 }
 248
 249 static int exclude_super_stripes(struct btrfs_root *root,
 250                                  struct btrfs_block_group_cache *cache)
 251 {
 252         u64 bytenr;
 253         u64 *logical;
 254         int stripe_len;
 255         int i, nr, ret;
 256
 257         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 258                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 259                 cache->bytes_super += stripe_len;
 260                 ret = add_excluded_extent(root, cache->key.objectid,
 261                                           stripe_len);
 262                 if (ret)
 263                         return ret;
 264         }
 265
 266         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 267                 bytenr = btrfs_sb_offset(i);
 268                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 269                                        cache->key.objectid, bytenr,
 270                                        0, &logical, &nr, &stripe_len);
 271                 if (ret)
 272                         return ret;
 273
 274                 while (nr--) {
 275                         u64 start, len;
 276
 277                         if (logical[nr] > cache->key.objectid +
 278                             cache->key.offset)
 279                                 continue;
 280
 281                         if (logical[nr] + stripe_len <= cache->key.objectid)
 282                                 continue;
 283
 284                         start = logical[nr];
 285                         if (start < cache->key.objectid) {
 286                                 start = cache->key.objectid;
 287                                 len = (logical[nr] + stripe_len) - start;
 288                         } else {
 289                                 len = min_t(u64, stripe_len,
 290                                             cache->key.objectid +
 291                                             cache->key.offset - start);
 292                         }
 293
 294                         cache->bytes_super += len;
 295                         ret = add_excluded_extent(root, start, len);
 296                         if (ret) {
 297                                 kfree(logical);
 298                                 return ret;
 299                         }
 300                 }
 301
 302                 kfree(logical);
 303         }
 304         return 0;
 305 }
 306
 307 static struct btrfs_caching_control *
 308 get_caching_control(struct btrfs_block_group_cache *cache)
 309 {
 310         struct btrfs_caching_control *ctl;
 311
 312         spin_lock(&cache->lock);
 313         if (cache->cached != BTRFS_CACHE_STARTED) {
 314                 spin_unlock(&cache->lock);
 315                 return NULL;
 316         }
 317
 318         /* We're loading it the fast way, so we don't have a caching_ctl. */
 319         if (!cache->caching_ctl) {
 320                 spin_unlock(&cache->lock);
 321                 return NULL;
 322         }
 323
 324         ctl = cache->caching_ctl;
 325         atomic_inc(&ctl->count);
 326         spin_unlock(&cache->lock);
 327         return ctl;
 328 }
 329
 330 static void put_caching_control(struct btrfs_caching_control *ctl)
 331 {
 332         if (atomic_dec_and_test(&ctl->count))
 333                 kfree(ctl);
 334 }
 335
 336 /*
 337  * this is only called by cache_block_group, since we could have freed extents
 338  * we need to check the pinned_extents for any extents that can't be used yet
 339  * since their free space will be released as soon as the transaction commits.
 340  */
 341 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 342                               struct btrfs_fs_info *info, u64 start, u64 end)
 343 {
 344         u64 extent_start, extent_end, size, total_added = 0;
 345         int ret;
 346
 347         while (start < end) {
 348                 ret = find_first_extent_bit(info->pinned_extents, start,
 349                                             &extent_start, &extent_end,
 350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 351                                             NULL);
 352                 if (ret)
 353                         break;
 354
 355                 if (extent_start <= start) {
 356                         start = extent_end + 1;
 357                 } else if (extent_start > start && extent_start < end) {
 358                         size = extent_start - start;
 359                         total_added += size;
 360                         ret = btrfs_add_free_space(block_group, start,
 361                                                    size);
 362                         BUG_ON(ret); /* -ENOMEM or logic error */
 363                         start = extent_end + 1;
 364                 } else {
 365                         break;
 366                 }
 367         }
 368
 369         if (start < end) {
 370                 size = end - start;
 371                 total_added += size;
 372                 ret = btrfs_add_free_space(block_group, start, size);
 373                 BUG_ON(ret); /* -ENOMEM or logic error */
 374         }
 375
 376         return total_added;
 377 }
 378
 379 static noinline void caching_thread(struct btrfs_work *work)
 380 {
 381         struct btrfs_block_group_cache *block_group;
 382         struct btrfs_fs_info *fs_info;
 383         struct btrfs_caching_control *caching_ctl;
 384         struct btrfs_root *extent_root;
 385         struct btrfs_path *path;
 386         struct extent_buffer *leaf;
 387         struct btrfs_key key;
 388         u64 total_found = 0;
 389         u64 last = 0;
 390         u32 nritems;
 391         int ret = 0;
 392
 393         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 394         block_group = caching_ctl->block_group;
 395         fs_info = block_group->fs_info;
 396         extent_root = fs_info->extent_root;
 397
 398         path = btrfs_alloc_path();
 399         if (!path)
 400                 goto out;
 401
 402         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 403
 404         /*
 405          * We don't want to deadlock with somebody trying to allocate a new
 406          * extent for the extent root while also trying to search the extent
 407          * root to add free space.  So we skip locking and search the commit
 408          * root, since its read-only
 409          */
 410         path->skip_locking = 1;
 411         path->search_commit_root = 1;
 412         path->reada = 1;
 413
 414         key.objectid = last;
 415         key.offset = 0;
 416         key.type = BTRFS_EXTENT_ITEM_KEY;
 417 again:
 418         mutex_lock(&caching_ctl->mutex);
 419         /* need to make sure the commit_root doesn't disappear */
 420         down_read(&fs_info->extent_commit_sem);
 421
 422         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 423         if (ret < 0)
 424                 goto err;
 425
 426         leaf = path->nodes[0];
 427         nritems = btrfs_header_nritems(leaf);
 428
 429         while (1) {
 430                 if (btrfs_fs_closing(fs_info) > 1) {
 431                         last = (u64)-1;
 432                         break;
 433                 }
 434
 435                 if (path->slots[0] < nritems) {
 436                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 437                 } else {
 438                         ret = find_next_key(path, 0, &key);
 439                         if (ret)
 440                                 break;
 441
 442                         if (need_resched()) {
 443                                 caching_ctl->progress = last;
 444                                 btrfs_release_path(path);
 445                                 up_read(&fs_info->extent_commit_sem);
 446                                 mutex_unlock(&caching_ctl->mutex);
 447                                 cond_resched();
 448                                 goto again;
 449                         }
 450
 451                         ret = btrfs_next_leaf(extent_root, path);
 452                         if (ret < 0)
 453                                 goto err;
 454                         if (ret)
 455                                 break;
 456                         leaf = path->nodes[0];
 457                         nritems = btrfs_header_nritems(leaf);
 458                         continue;
 459                 }
 460
 461                 if (key.objectid < block_group->key.objectid) {
 462                         path->slots[0]++;
 463                         continue;
 464                 }
 465
 466                 if (key.objectid >= block_group->key.objectid +
 467                     block_group->key.offset)
 468                         break;
 469
 470                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 471                     key.type == BTRFS_METADATA_ITEM_KEY) {
 472                         total_found += add_new_free_space(block_group,
 473                                                           fs_info, last,
 474                                                           key.objectid);
 475                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 476                                 last = key.objectid +
 477                                         fs_info->tree_root->leafsize;
 478                         else
 479                                 last = key.objectid + key.offset;
 480
 481                         if (total_found > (1024 * 1024 * 2)) {
 482                                 total_found = 0;
 483                                 wake_up(&caching_ctl->wait);
 484                         }
 485                 }
 486                 path->slots[0]++;
 487         }
 488         ret = 0;
 489
 490         total_found += add_new_free_space(block_group, fs_info, last,
 491                                           block_group->key.objectid +
 492                                           block_group->key.offset);
 493         caching_ctl->progress = (u64)-1;
 494
 495         spin_lock(&block_group->lock);
 496         block_group->caching_ctl = NULL;
 497         block_group->cached = BTRFS_CACHE_FINISHED;
 498         spin_unlock(&block_group->lock);
 499
 500 err:
 501         btrfs_free_path(path);
 502         up_read(&fs_info->extent_commit_sem);
 503
 504         free_excluded_extents(extent_root, block_group);
 505
 506         mutex_unlock(&caching_ctl->mutex);
 507 out:
 508         wake_up(&caching_ctl->wait);
 509
 510         put_caching_control(caching_ctl);
 511         btrfs_put_block_group(block_group);
 512 }
 513
 514 static int cache_block_group(struct btrfs_block_group_cache *cache,
 515                              int load_cache_only)
 516 {
 517         DEFINE_WAIT(wait);
 518         struct btrfs_fs_info *fs_info = cache->fs_info;
 519         struct btrfs_caching_control *caching_ctl;
 520         int ret = 0;
 521
 522         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 523         if (!caching_ctl)
 524                 return -ENOMEM;
 525
 526         INIT_LIST_HEAD(&caching_ctl->list);
 527         mutex_init(&caching_ctl->mutex);
 528         init_waitqueue_head(&caching_ctl->wait);
 529         caching_ctl->block_group = cache;
 530         caching_ctl->progress = cache->key.objectid;
 531         atomic_set(&caching_ctl->count, 1);
 532         caching_ctl->work.func = caching_thread;
 533
 534         spin_lock(&cache->lock);
 535         /*
 536          * This should be a rare occasion, but this could happen I think in the
 537          * case where one thread starts to load the space cache info, and then
 538          * some other thread starts a transaction commit which tries to do an
 539          * allocation while the other thread is still loading the space cache
 540          * info.  The previous loop should have kept us from choosing this block
 541          * group, but if we've moved to the state where we will wait on caching
 542          * block groups we need to first check if we're doing a fast load here,
 543          * so we can wait for it to finish, otherwise we could end up allocating
 544          * from a block group who's cache gets evicted for one reason or
 545          * another.
 546          */
 547         while (cache->cached == BTRFS_CACHE_FAST) {
 548                 struct btrfs_caching_control *ctl;
 549
 550                 ctl = cache->caching_ctl;
 551                 atomic_inc(&ctl->count);
 552                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 553                 spin_unlock(&cache->lock);
 554
 555                 schedule();
 556
 557                 finish_wait(&ctl->wait, &wait);
 558                 put_caching_control(ctl);
 559                 spin_lock(&cache->lock);
 560         }
 561
 562         if (cache->cached != BTRFS_CACHE_NO) {
 563                 spin_unlock(&cache->lock);
 564                 kfree(caching_ctl);
 565                 return 0;
 566         }
 567         WARN_ON(cache->caching_ctl);
 568         cache->caching_ctl = caching_ctl;
 569         cache->cached = BTRFS_CACHE_FAST;
 570         spin_unlock(&cache->lock);
 571
 572         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 573                 ret = load_free_space_cache(fs_info, cache);
 574
 575                 spin_lock(&cache->lock);
 576                 if (ret == 1) {
 577                         cache->caching_ctl = NULL;
 578                         cache->cached = BTRFS_CACHE_FINISHED;
 579                         cache->last_byte_to_unpin = (u64)-1;
 580                 } else {
 581                         if (load_cache_only) {
 582                                 cache->caching_ctl = NULL;
 583                                 cache->cached = BTRFS_CACHE_NO;
 584                         } else {
 585                                 cache->cached = BTRFS_CACHE_STARTED;
 586                         }
 587                 }
 588                 spin_unlock(&cache->lock);
 589                 wake_up(&caching_ctl->wait);
 590                 if (ret == 1) {
 591                         put_caching_control(caching_ctl);
 592                         free_excluded_extents(fs_info->extent_root, cache);
 593                         return 0;
 594                 }
 595         } else {
 596                 /*
 597                  * We are not going to do the fast caching, set cached to the
 598                  * appropriate value and wakeup any waiters.
 599                  */
 600                 spin_lock(&cache->lock);
 601                 if (load_cache_only) {
 602                         cache->caching_ctl = NULL;
 603                         cache->cached = BTRFS_CACHE_NO;
 604                 } else {
 605                         cache->cached = BTRFS_CACHE_STARTED;
 606                 }
 607                 spin_unlock(&cache->lock);
 608                 wake_up(&caching_ctl->wait);
 609         }
 610
 611         if (load_cache_only) {
 612                 put_caching_control(caching_ctl);
 613                 return 0;
 614         }
 615
 616         down_write(&fs_info->extent_commit_sem);
 617         atomic_inc(&caching_ctl->count);
 618         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 619         up_write(&fs_info->extent_commit_sem);
 620
 621         btrfs_get_block_group(cache);
 622
 623         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 624
 625         return ret;
 626 }
 627
 628 /*
 629  * return the block group that starts at or after bytenr
 630  */
 631 static struct btrfs_block_group_cache *
 632 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 633 {
 634         struct btrfs_block_group_cache *cache;
 635
 636         cache = block_group_cache_tree_search(info, bytenr, 0);
 637
 638         return cache;
 639 }
 640
 641 /*
 642  * return the block group that contains the given bytenr
 643  */
 644 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 645                                                  struct btrfs_fs_info *info,
 646                                                  u64 bytenr)
 647 {
 648         struct btrfs_block_group_cache *cache;
 649
 650         cache = block_group_cache_tree_search(info, bytenr, 1);
 651
 652         return cache;
 653 }
 654
 655 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 656                                                   u64 flags)
 657 {
 658         struct list_head *head = &info->space_info;
 659         struct btrfs_space_info *found;
 660
 661         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 662
 663         rcu_read_lock();
 664         list_for_each_entry_rcu(found, head, list) {
 665                 if (found->flags & flags) {
 666                         rcu_read_unlock();
 667                         return found;
 668                 }
 669         }
 670         rcu_read_unlock();
 671         return NULL;
 672 }
 673
 674 /*
 675  * after adding space to the filesystem, we need to clear the full flags
 676  * on all the space infos.
 677  */
 678 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 679 {
 680         struct list_head *head = &info->space_info;
 681         struct btrfs_space_info *found;
 682
 683         rcu_read_lock();
 684         list_for_each_entry_rcu(found, head, list)
 685                 found->full = 0;
 686         rcu_read_unlock();
 687 }
 688
 689 /* simple helper to search for an existing extent at a given offset */
 690 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 691 {
 692         int ret;
 693         struct btrfs_key key;
 694         struct btrfs_path *path;
 695
 696         path = btrfs_alloc_path();
 697         if (!path)
 698                 return -ENOMEM;
 699
 700         key.objectid = start;
 701         key.offset = len;
 702         key.type = BTRFS_EXTENT_ITEM_KEY;
 703         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 704                                 0, 0);
 705         if (ret > 0) {
 706                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 707                 if (key.objectid == start &&
 708                     key.type == BTRFS_METADATA_ITEM_KEY)
 709                         ret = 0;
 710         }
 711         btrfs_free_path(path);
 712         return ret;
 713 }
 714
 715 /*
 716  * helper function to lookup reference count and flags of a tree block.
 717  *
 718  * the head node for delayed ref is used to store the sum of all the
 719  * reference count modifications queued up in the rbtree. the head
 720  * node may also store the extent flags to set. This way you can check
 721  * to see what the reference count and extent flags would be if all of
 722  * the delayed refs are not processed.
 723  */
 724 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 725                              struct btrfs_root *root, u64 bytenr,
 726                              u64 offset, int metadata, u64 *refs, u64 *flags)
 727 {
 728         struct btrfs_delayed_ref_head *head;
 729         struct btrfs_delayed_ref_root *delayed_refs;
 730         struct btrfs_path *path;
 731         struct btrfs_extent_item *ei;
 732         struct extent_buffer *leaf;
 733         struct btrfs_key key;
 734         u32 item_size;
 735         u64 num_refs;
 736         u64 extent_flags;
 737         int ret;
 738
 739         /*
 740          * If we don't have skinny metadata, don't bother doing anything
 741          * different
 742          */
 743         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 744                 offset = root->leafsize;
 745                 metadata = 0;
 746         }
 747
 748         path = btrfs_alloc_path();
 749         if (!path)
 750                 return -ENOMEM;
 751
 752         if (metadata) {
 753                 key.objectid = bytenr;
 754                 key.type = BTRFS_METADATA_ITEM_KEY;
 755                 key.offset = offset;
 756         } else {
 757                 key.objectid = bytenr;
 758                 key.type = BTRFS_EXTENT_ITEM_KEY;
 759                 key.offset = offset;
 760         }
 761
 762         if (!trans) {
 763                 path->skip_locking = 1;
 764                 path->search_commit_root = 1;
 765         }
 766 again:
 767         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 768                                 &key, path, 0, 0);
 769         if (ret < 0)
 770                 goto out_free;
 771
 772         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 773                 key.type = BTRFS_EXTENT_ITEM_KEY;
 774                 key.offset = root->leafsize;
 775                 btrfs_release_path(path);
 776                 goto again;
 777         }
 778
 779         if (ret == 0) {
 780                 leaf = path->nodes[0];
 781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 782                 if (item_size >= sizeof(*ei)) {
 783                         ei = btrfs_item_ptr(leaf, path->slots[0],
 784                                             struct btrfs_extent_item);
 785                         num_refs = btrfs_extent_refs(leaf, ei);
 786                         extent_flags = btrfs_extent_flags(leaf, ei);
 787                 } else {
 788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 789                         struct btrfs_extent_item_v0 *ei0;
 790                         BUG_ON(item_size != sizeof(*ei0));
 791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 792                                              struct btrfs_extent_item_v0);
 793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 794                         /* FIXME: this isn't correct for data */
 795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 796 #else
 797                         BUG();
 798 #endif
 799                 }
 800                 BUG_ON(num_refs == 0);
 801         } else {
 802                 num_refs = 0;
 803                 extent_flags = 0;
 804                 ret = 0;
 805         }
 806
 807         if (!trans)
 808                 goto out;
 809
 810         delayed_refs = &trans->transaction->delayed_refs;
 811         spin_lock(&delayed_refs->lock);
 812         head = btrfs_find_delayed_ref_head(trans, bytenr);
 813         if (head) {
 814                 if (!mutex_trylock(&head->mutex)) {
 815                         atomic_inc(&head->node.refs);
 816                         spin_unlock(&delayed_refs->lock);
 817
 818                         btrfs_release_path(path);
 819
 820                         /*
 821                          * Mutex was contended, block until it's released and try
 822                          * again
 823                          */
 824                         mutex_lock(&head->mutex);
 825                         mutex_unlock(&head->mutex);
 826                         btrfs_put_delayed_ref(&head->node);
 827                         goto again;
 828                 }
 829                 if (head->extent_op && head->extent_op->update_flags)
 830                         extent_flags |= head->extent_op->flags_to_set;
 831                 else
 832                         BUG_ON(num_refs == 0);
 833
 834                 num_refs += head->node.ref_mod;
 835                 mutex_unlock(&head->mutex);
 836         }
 837         spin_unlock(&delayed_refs->lock);
 838 out:
 839         WARN_ON(num_refs == 0);
 840         if (refs)
 841                 *refs = num_refs;
 842         if (flags)
 843                 *flags = extent_flags;
 844 out_free:
 845         btrfs_free_path(path);
 846         return ret;
 847 }
 848
 849 /*
 850  * Back reference rules.  Back refs have three main goals:
 851  *
 852  * 1) differentiate between all holders of references to an extent so that
 853  *    when a reference is dropped we can make sure it was a valid reference
 854  *    before freeing the extent.
 855  *
 856  * 2) Provide enough information to quickly find the holders of an extent
 857  *    if we notice a given block is corrupted or bad.
 858  *
 859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 860  *    maintenance.  This is actually the same as #2, but with a slightly
 861  *    different use case.
 862  *
 863  * There are two kinds of back refs. The implicit back refs is optimized
 864  * for pointers in non-shared tree blocks. For a given pointer in a block,
 865  * back refs of this kind provide information about the block's owner tree
 866  * and the pointer's key. These information allow us to find the block by
 867  * b-tree searching. The full back refs is for pointers in tree blocks not
 868  * referenced by their owner trees. The location of tree block is recorded
 869  * in the back refs. Actually the full back refs is generic, and can be
 870  * used in all cases the implicit back refs is used. The major shortcoming
 871  * of the full back refs is its overhead. Every time a tree block gets
 872  * COWed, we have to update back refs entry for all pointers in it.
 873  *
 874  * For a newly allocated tree block, we use implicit back refs for
 875  * pointers in it. This means most tree related operations only involve
 876  * implicit back refs. For a tree block created in old transaction, the
 877  * only way to drop a reference to it is COW it. So we can detect the
 878  * event that tree block loses its owner tree's reference and do the
 879  * back refs conversion.
 880  *
 881  * When a tree block is COW'd through a tree, there are four cases:
 882  *
 883  * The reference count of the block is one and the tree is the block's
 884  * owner tree. Nothing to do in this case.
 885  *
 886  * The reference count of the block is one and the tree is not the
 887  * block's owner tree. In this case, full back refs is used for pointers
 888  * in the block. Remove these full back refs, add implicit back refs for
 889  * every pointers in the new block.
 890  *
 891  * The reference count of the block is greater than one and the tree is
 892  * the block's owner tree. In this case, implicit back refs is used for
 893  * pointers in the block. Add full back refs for every pointers in the
 894  * block, increase lower level extents' reference counts. The original
 895  * implicit back refs are entailed to the new block.
 896  *
 897  * The reference count of the block is greater than one and the tree is
 898  * not the block's owner tree. Add implicit back refs for every pointer in
 899  * the new block, increase lower level extents' reference count.
 900  *
 901  * Back Reference Key composing:
 902  *
 903  * The key objectid corresponds to the first byte in the extent,
 904  * The key type is used to differentiate between types of back refs.
 905  * There are different meanings of the key offset for different types
 906  * of back refs.
 907  *
 908  * File extents can be referenced by:
 909  *
 910  * - multiple snapshots, subvolumes, or different generations in one subvol
 911  * - different files inside a single subvolume
 912  * - different offsets inside a file (bookend extents in file.c)
 913  *
 914  * The extent ref structure for the implicit back refs has fields for:
 915  *
 916  * - Objectid of the subvolume root
 917  * - objectid of the file holding the reference
 918  * - original offset in the file
 919  * - how many bookend extents
 920  *
 921  * The key offset for the implicit back refs is hash of the first
 922  * three fields.
 923  *
 924  * The extent ref structure for the full back refs has field for:
 925  *
 926  * - number of pointers in the tree leaf
 927  *
 928  * The key offset for the implicit back refs is the first byte of
 929  * the tree leaf
 930  *
 931  * When a file extent is allocated, The implicit back refs is used.
 932  * the fields are filled in:
 933  *
 934  *     (root_key.objectid, inode objectid, offset in file, 1)
 935  *
 936  * When a file extent is removed file truncation, we find the
 937  * corresponding implicit back refs and check the following fields:
 938  *
 939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 940  *
 941  * Btree extents can be referenced by:
 942  *
 943  * - Different subvolumes
 944  *
 945  * Both the implicit back refs and the full back refs for tree blocks
 946  * only consist of key. The key offset for the implicit back refs is
 947  * objectid of block's owner tree. The key offset for the full back refs
 948  * is the first byte of parent block.
 949  *
 950  * When implicit back refs is used, information about the lowest key and
 951  * level of the tree block are required. These information are stored in
 952  * tree block info structure.
 953  */
 954
 955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 957                                   struct btrfs_root *root,
 958                                   struct btrfs_path *path,
 959                                   u64 owner, u32 extra_size)
 960 {
 961         struct btrfs_extent_item *item;
 962         struct btrfs_extent_item_v0 *ei0;
 963         struct btrfs_extent_ref_v0 *ref0;
 964         struct btrfs_tree_block_info *bi;
 965         struct extent_buffer *leaf;
 966         struct btrfs_key key;
 967         struct btrfs_key found_key;
 968         u32 new_size = sizeof(*item);
 969         u64 refs;
 970         int ret;
 971
 972         leaf = path->nodes[0];
 973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 974
 975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 977                              struct btrfs_extent_item_v0);
 978         refs = btrfs_extent_refs_v0(leaf, ei0);
 979
 980         if (owner == (u64)-1) {
 981                 while (1) {
 982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 983                                 ret = btrfs_next_leaf(root, path);
 984                                 if (ret < 0)
 985                                         return ret;
 986                                 BUG_ON(ret > 0); /* Corruption */
 987                                 leaf = path->nodes[0];
 988                         }
 989                         btrfs_item_key_to_cpu(leaf, &found_key,
 990                                               path->slots[0]);
 991                         BUG_ON(key.objectid != found_key.objectid);
 992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 993                                 path->slots[0]++;
 994                                 continue;
 995                         }
 996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 997                                               struct btrfs_extent_ref_v0);
 998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1467                                                  SKINNY_METADATA);
1468
1469         key.objectid = bytenr;
1470         key.type = BTRFS_EXTENT_ITEM_KEY;
1471         key.offset = num_bytes;
1472
1473         want = extent_ref_type(parent, owner);
1474         if (insert) {
1475                 extra_size = btrfs_extent_inline_ref_size(want);
1476                 path->keep_locks = 1;
1477         } else
1478                 extra_size = -1;
1479
1480         /*
1481          * Owner is our parent level, so we can just add one to get the level
1482          * for the block we are interested in.
1483          */
1484         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1485                 key.type = BTRFS_METADATA_ITEM_KEY;
1486                 key.offset = owner;
1487         }
1488
1489 again:
1490         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1491         if (ret < 0) {
1492                 err = ret;
1493                 goto out;
1494         }
1495
1496         /*
1497          * We may be a newly converted file system which still has the old fat
1498          * extent entries for metadata, so try and see if we have one of those.
1499          */
1500         if (ret > 0 && skinny_metadata) {
1501                 skinny_metadata = false;
1502                 if (path->slots[0]) {
1503                         path->slots[0]--;
1504                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1505                                               path->slots[0]);
1506                         if (key.objectid == bytenr &&
1507                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1508                             key.offset == num_bytes)
1509                                 ret = 0;
1510                 }
1511                 if (ret) {
1512                         key.type = BTRFS_EXTENT_ITEM_KEY;
1513                         key.offset = num_bytes;
1514                         btrfs_release_path(path);
1515                         goto again;
1516                 }
1517         }
1518
1519         if (ret && !insert) {
1520                 err = -ENOENT;
1521                 goto out;
1522         } else if (ret) {
1523                 err = -EIO;
1524                 WARN_ON(1);
1525                 goto out;
1526         }
1527
1528         leaf = path->nodes[0];
1529         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1530 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1531         if (item_size < sizeof(*ei)) {
1532                 if (!insert) {
1533                         err = -ENOENT;
1534                         goto out;
1535                 }
1536                 ret = convert_extent_item_v0(trans, root, path, owner,
1537                                              extra_size);
1538                 if (ret < 0) {
1539                         err = ret;
1540                         goto out;
1541                 }
1542                 leaf = path->nodes[0];
1543                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1544         }
1545 #endif
1546         BUG_ON(item_size < sizeof(*ei));
1547
1548         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1549         flags = btrfs_extent_flags(leaf, ei);
1550
1551         ptr = (unsigned long)(ei + 1);
1552         end = (unsigned long)ei + item_size;
1553
1554         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1555                 ptr += sizeof(struct btrfs_tree_block_info);
1556                 BUG_ON(ptr > end);
1557         }
1558
1559         err = -ENOENT;
1560         while (1) {
1561                 if (ptr >= end) {
1562                         WARN_ON(ptr > end);
1563                         break;
1564                 }
1565                 iref = (struct btrfs_extent_inline_ref *)ptr;
1566                 type = btrfs_extent_inline_ref_type(leaf, iref);
1567                 if (want < type)
1568                         break;
1569                 if (want > type) {
1570                         ptr += btrfs_extent_inline_ref_size(type);
1571                         continue;
1572                 }
1573
1574                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1575                         struct btrfs_extent_data_ref *dref;
1576                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1577                         if (match_extent_data_ref(leaf, dref, root_objectid,
1578                                                   owner, offset)) {
1579                                 err = 0;
1580                                 break;
1581                         }
1582                         if (hash_extent_data_ref_item(leaf, dref) <
1583                             hash_extent_data_ref(root_objectid, owner, offset))
1584                                 break;
1585                 } else {
1586                         u64 ref_offset;
1587                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1588                         if (parent > 0) {
1589                                 if (parent == ref_offset) {
1590                                         err = 0;
1591                                         break;
1592                                 }
1593                                 if (ref_offset < parent)
1594                                         break;
1595                         } else {
1596                                 if (root_objectid == ref_offset) {
1597                                         err = 0;
1598                                         break;
1599                                 }
1600                                 if (ref_offset < root_objectid)
1601                                         break;
1602                         }
1603                 }
1604                 ptr += btrfs_extent_inline_ref_size(type);
1605         }
1606         if (err == -ENOENT && insert) {
1607                 if (item_size + extra_size >=
1608                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1609                         err = -EAGAIN;
1610                         goto out;
1611                 }
1612                 /*
1613                  * To add new inline back ref, we have to make sure
1614                  * there is no corresponding back ref item.
1615                  * For simplicity, we just do not add new inline back
1616                  * ref if there is any kind of item for this block
1617                  */
1618                 if (find_next_key(path, 0, &key) == 0 &&
1619                     key.objectid == bytenr &&
1620                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1621                         err = -EAGAIN;
1622                         goto out;
1623                 }
1624         }
1625         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1626 out:
1627         if (insert) {
1628                 path->keep_locks = 0;
1629                 btrfs_unlock_up_safe(path, 1);
1630         }
1631         return err;
1632 }
1633
1634 /*
1635  * helper to add new inline back ref
1636  */
1637 static noinline_for_stack
1638 void setup_inline_extent_backref(struct btrfs_root *root,
1639                                  struct btrfs_path *path,
1640                                  struct btrfs_extent_inline_ref *iref,
1641                                  u64 parent, u64 root_objectid,
1642                                  u64 owner, u64 offset, int refs_to_add,
1643                                  struct btrfs_delayed_extent_op *extent_op)
1644 {
1645         struct extent_buffer *leaf;
1646         struct btrfs_extent_item *ei;
1647         unsigned long ptr;
1648         unsigned long end;
1649         unsigned long item_offset;
1650         u64 refs;
1651         int size;
1652         int type;
1653
1654         leaf = path->nodes[0];
1655         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1656         item_offset = (unsigned long)iref - (unsigned long)ei;
1657
1658         type = extent_ref_type(parent, owner);
1659         size = btrfs_extent_inline_ref_size(type);
1660
1661         btrfs_extend_item(root, path, size);
1662
1663         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1664         refs = btrfs_extent_refs(leaf, ei);
1665         refs += refs_to_add;
1666         btrfs_set_extent_refs(leaf, ei, refs);
1667         if (extent_op)
1668                 __run_delayed_extent_op(extent_op, leaf, ei);
1669
1670         ptr = (unsigned long)ei + item_offset;
1671         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1672         if (ptr < end - size)
1673                 memmove_extent_buffer(leaf, ptr + size, ptr,
1674                                       end - size - ptr);
1675
1676         iref = (struct btrfs_extent_inline_ref *)ptr;
1677         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1678         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1679                 struct btrfs_extent_data_ref *dref;
1680                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1681                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1682                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1683                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1684                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1685         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1686                 struct btrfs_shared_data_ref *sref;
1687                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1688                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1689                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1690         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1691                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1692         } else {
1693                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1694         }
1695         btrfs_mark_buffer_dirty(leaf);
1696 }
1697
1698 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1699                                  struct btrfs_root *root,
1700                                  struct btrfs_path *path,
1701                                  struct btrfs_extent_inline_ref **ref_ret,
1702                                  u64 bytenr, u64 num_bytes, u64 parent,
1703                                  u64 root_objectid, u64 owner, u64 offset)
1704 {
1705         int ret;
1706
1707         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1708                                            bytenr, num_bytes, parent,
1709                                            root_objectid, owner, offset, 0);
1710         if (ret != -ENOENT)
1711                 return ret;
1712
1713         btrfs_release_path(path);
1714         *ref_ret = NULL;
1715
1716         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1717                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1718                                             root_objectid);
1719         } else {
1720                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1721                                              root_objectid, owner, offset);
1722         }
1723         return ret;
1724 }
1725
1726 /*
1727  * helper to update/remove inline back ref
1728  */
1729 static noinline_for_stack
1730 void update_inline_extent_backref(struct btrfs_root *root,
1731                                   struct btrfs_path *path,
1732                                   struct btrfs_extent_inline_ref *iref,
1733                                   int refs_to_mod,
1734                                   struct btrfs_delayed_extent_op *extent_op)
1735 {
1736         struct extent_buffer *leaf;
1737         struct btrfs_extent_item *ei;
1738         struct btrfs_extent_data_ref *dref = NULL;
1739         struct btrfs_shared_data_ref *sref = NULL;
1740         unsigned long ptr;
1741         unsigned long end;
1742         u32 item_size;
1743         int size;
1744         int type;
1745         u64 refs;
1746
1747         leaf = path->nodes[0];
1748         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1749         refs = btrfs_extent_refs(leaf, ei);
1750         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1751         refs += refs_to_mod;
1752         btrfs_set_extent_refs(leaf, ei, refs);
1753         if (extent_op)
1754                 __run_delayed_extent_op(extent_op, leaf, ei);
1755
1756         type = btrfs_extent_inline_ref_type(leaf, iref);
1757
1758         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1759                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1760                 refs = btrfs_extent_data_ref_count(leaf, dref);
1761         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1762                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1763                 refs = btrfs_shared_data_ref_count(leaf, sref);
1764         } else {
1765                 refs = 1;
1766                 BUG_ON(refs_to_mod != -1);
1767         }
1768
1769         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1770         refs += refs_to_mod;
1771
1772         if (refs > 0) {
1773                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1774                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1775                 else
1776                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1777         } else {
1778                 size =  btrfs_extent_inline_ref_size(type);
1779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1780                 ptr = (unsigned long)iref;
1781                 end = (unsigned long)ei + item_size;
1782                 if (ptr + size < end)
1783                         memmove_extent_buffer(leaf, ptr, ptr + size,
1784                                               end - ptr - size);
1785                 item_size -= size;
1786                 btrfs_truncate_item(root, path, item_size, 1);
1787         }
1788         btrfs_mark_buffer_dirty(leaf);
1789 }
1790
1791 static noinline_for_stack
1792 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1793                                  struct btrfs_root *root,
1794                                  struct btrfs_path *path,
1795                                  u64 bytenr, u64 num_bytes, u64 parent,
1796                                  u64 root_objectid, u64 owner,
1797                                  u64 offset, int refs_to_add,
1798                                  struct btrfs_delayed_extent_op *extent_op)
1799 {
1800         struct btrfs_extent_inline_ref *iref;
1801         int ret;
1802
1803         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1804                                            bytenr, num_bytes, parent,
1805                                            root_objectid, owner, offset, 1);
1806         if (ret == 0) {
1807                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1808                 update_inline_extent_backref(root, path, iref,
1809                                              refs_to_add, extent_op);
1810         } else if (ret == -ENOENT) {
1811                 setup_inline_extent_backref(root, path, iref, parent,
1812                                             root_objectid, owner, offset,
1813                                             refs_to_add, extent_op);
1814                 ret = 0;
1815         }
1816         return ret;
1817 }
1818
1819 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1820                                  struct btrfs_root *root,
1821                                  struct btrfs_path *path,
1822                                  u64 bytenr, u64 parent, u64 root_objectid,
1823                                  u64 owner, u64 offset, int refs_to_add)
1824 {
1825         int ret;
1826         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1827                 BUG_ON(refs_to_add != 1);
1828                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1829                                             parent, root_objectid);
1830         } else {
1831                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1832                                              parent, root_objectid,
1833                                              owner, offset, refs_to_add);
1834         }
1835         return ret;
1836 }
1837
1838 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1839                                  struct btrfs_root *root,
1840                                  struct btrfs_path *path,
1841                                  struct btrfs_extent_inline_ref *iref,
1842                                  int refs_to_drop, int is_data)
1843 {
1844         int ret = 0;
1845
1846         BUG_ON(!is_data && refs_to_drop != 1);
1847         if (iref) {
1848                 update_inline_extent_backref(root, path, iref,
1849                                              -refs_to_drop, NULL);
1850         } else if (is_data) {
1851                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1852         } else {
1853                 ret = btrfs_del_item(trans, root, path);
1854         }
1855         return ret;
1856 }
1857
1858 static int btrfs_issue_discard(struct block_device *bdev,
1859                                 u64 start, u64 len)
1860 {
1861         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1862 }
1863
1864 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1865                                 u64 num_bytes, u64 *actual_bytes)
1866 {
1867         int ret;
1868         u64 discarded_bytes = 0;
1869         struct btrfs_bio *bbio = NULL;
1870
1871
1872         /* Tell the block device(s) that the sectors can be discarded */
1873         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1874                               bytenr, &num_bytes, &bbio, 0);
1875         /* Error condition is -ENOMEM */
1876         if (!ret) {
1877                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1878                 int i;
1879
1880
1881                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1882                         if (!stripe->dev->can_discard)
1883                                 continue;
1884
1885                         ret = btrfs_issue_discard(stripe->dev->bdev,
1886                                                   stripe->physical,
1887                                                   stripe->length);
1888                         if (!ret)
1889                                 discarded_bytes += stripe->length;
1890                         else if (ret != -EOPNOTSUPP)
1891                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1892
1893                         /*
1894                          * Just in case we get back EOPNOTSUPP for some reason,
1895                          * just ignore the return value so we don't screw up
1896                          * people calling discard_extent.
1897                          */
1898                         ret = 0;
1899                 }
1900                 kfree(bbio);
1901         }
1902
1903         if (actual_bytes)
1904                 *actual_bytes = discarded_bytes;
1905
1906
1907         if (ret == -EOPNOTSUPP)
1908                 ret = 0;
1909         return ret;
1910 }
1911
1912 /* Can return -ENOMEM */
1913 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1914                          struct btrfs_root *root,
1915                          u64 bytenr, u64 num_bytes, u64 parent,
1916                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1917 {
1918         int ret;
1919         struct btrfs_fs_info *fs_info = root->fs_info;
1920
1921         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1922                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1923
1924         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1925                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1926                                         num_bytes,
1927                                         parent, root_objectid, (int)owner,
1928                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1929         } else {
1930                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1931                                         num_bytes,
1932                                         parent, root_objectid, owner, offset,
1933                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1934         }
1935         return ret;
1936 }
1937
1938 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1939                                   struct btrfs_root *root,
1940                                   u64 bytenr, u64 num_bytes,
1941                                   u64 parent, u64 root_objectid,
1942                                   u64 owner, u64 offset, int refs_to_add,
1943                                   struct btrfs_delayed_extent_op *extent_op)
1944 {
1945         struct btrfs_path *path;
1946         struct extent_buffer *leaf;
1947         struct btrfs_extent_item *item;
1948         u64 refs;
1949         int ret;
1950         int err = 0;
1951
1952         path = btrfs_alloc_path();
1953         if (!path)
1954                 return -ENOMEM;
1955
1956         path->reada = 1;
1957         path->leave_spinning = 1;
1958         /* this will setup the path even if it fails to insert the back ref */
1959         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1960                                            path, bytenr, num_bytes, parent,
1961                                            root_objectid, owner, offset,
1962                                            refs_to_add, extent_op);
1963         if (ret == 0)
1964                 goto out;
1965
1966         if (ret != -EAGAIN) {
1967                 err = ret;
1968                 goto out;
1969         }
1970
1971         leaf = path->nodes[0];
1972         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1973         refs = btrfs_extent_refs(leaf, item);
1974         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1975         if (extent_op)
1976                 __run_delayed_extent_op(extent_op, leaf, item);
1977
1978         btrfs_mark_buffer_dirty(leaf);
1979         btrfs_release_path(path);
1980
1981         path->reada = 1;
1982         path->leave_spinning = 1;
1983
1984         /* now insert the actual backref */
1985         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1986                                     path, bytenr, parent, root_objectid,
1987                                     owner, offset, refs_to_add);
1988         if (ret)
1989                 btrfs_abort_transaction(trans, root, ret);
1990 out:
1991         btrfs_free_path(path);
1992         return err;
1993 }
1994
1995 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1996                                 struct btrfs_root *root,
1997                                 struct btrfs_delayed_ref_node *node,
1998                                 struct btrfs_delayed_extent_op *extent_op,
1999                                 int insert_reserved)
2000 {
2001         int ret = 0;
2002         struct btrfs_delayed_data_ref *ref;
2003         struct btrfs_key ins;
2004         u64 parent = 0;
2005         u64 ref_root = 0;
2006         u64 flags = 0;
2007
2008         ins.objectid = node->bytenr;
2009         ins.offset = node->num_bytes;
2010         ins.type = BTRFS_EXTENT_ITEM_KEY;
2011
2012         ref = btrfs_delayed_node_to_data_ref(node);
2013         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2014                 parent = ref->parent;
2015         else
2016                 ref_root = ref->root;
2017
2018         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2019                 if (extent_op)
2020                         flags |= extent_op->flags_to_set;
2021                 ret = alloc_reserved_file_extent(trans, root,
2022                                                  parent, ref_root, flags,
2023                                                  ref->objectid, ref->offset,
2024                                                  &ins, node->ref_mod);
2025         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2026                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2027                                              node->num_bytes, parent,
2028                                              ref_root, ref->objectid,
2029                                              ref->offset, node->ref_mod,
2030                                              extent_op);
2031         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2032                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2033                                           node->num_bytes, parent,
2034                                           ref_root, ref->objectid,
2035                                           ref->offset, node->ref_mod,
2036                                           extent_op);
2037         } else {
2038                 BUG();
2039         }
2040         return ret;
2041 }
2042
2043 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2044                                     struct extent_buffer *leaf,
2045                                     struct btrfs_extent_item *ei)
2046 {
2047         u64 flags = btrfs_extent_flags(leaf, ei);
2048         if (extent_op->update_flags) {
2049                 flags |= extent_op->flags_to_set;
2050                 btrfs_set_extent_flags(leaf, ei, flags);
2051         }
2052
2053         if (extent_op->update_key) {
2054                 struct btrfs_tree_block_info *bi;
2055                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2056                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2057                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2058         }
2059 }
2060
2061 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2062                                  struct btrfs_root *root,
2063                                  struct btrfs_delayed_ref_node *node,
2064                                  struct btrfs_delayed_extent_op *extent_op)
2065 {
2066         struct btrfs_key key;
2067         struct btrfs_path *path;
2068         struct btrfs_extent_item *ei;
2069         struct extent_buffer *leaf;
2070         u32 item_size;
2071         int ret;
2072         int err = 0;
2073         int metadata = (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2074                         node->type == BTRFS_SHARED_BLOCK_REF_KEY);
2075
2076         if (trans->aborted)
2077                 return 0;
2078
2079         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2080                 metadata = 0;
2081
2082         path = btrfs_alloc_path();
2083         if (!path)
2084                 return -ENOMEM;
2085
2086         key.objectid = node->bytenr;
2087
2088         if (metadata) {
2089                 struct btrfs_delayed_tree_ref *tree_ref;
2090
2091                 tree_ref = btrfs_delayed_node_to_tree_ref(node);
2092                 key.type = BTRFS_METADATA_ITEM_KEY;
2093                 key.offset = tree_ref->level;
2094         } else {
2095                 key.type = BTRFS_EXTENT_ITEM_KEY;
2096                 key.offset = node->num_bytes;
2097         }
2098
2099 again:
2100         path->reada = 1;
2101         path->leave_spinning = 1;
2102         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2103                                 path, 0, 1);
2104         if (ret < 0) {
2105                 err = ret;
2106                 goto out;
2107         }
2108         if (ret > 0) {
2109                 if (metadata) {
2110                         btrfs_release_path(path);
2111                         metadata = 0;
2112
2113                         key.offset = node->num_bytes;
2114                         key.type = BTRFS_EXTENT_ITEM_KEY;
2115                         goto again;
2116                 }
2117                 err = -EIO;
2118                 goto out;
2119         }
2120
2121         leaf = path->nodes[0];
2122         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2123 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2124         if (item_size < sizeof(*ei)) {
2125                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2126                                              path, (u64)-1, 0);
2127                 if (ret < 0) {
2128                         err = ret;
2129                         goto out;
2130                 }
2131                 leaf = path->nodes[0];
2132                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2133         }
2134 #endif
2135         BUG_ON(item_size < sizeof(*ei));
2136         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2137         __run_delayed_extent_op(extent_op, leaf, ei);
2138
2139         btrfs_mark_buffer_dirty(leaf);
2140 out:
2141         btrfs_free_path(path);
2142         return err;
2143 }
2144
2145 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2146                                 struct btrfs_root *root,
2147                                 struct btrfs_delayed_ref_node *node,
2148                                 struct btrfs_delayed_extent_op *extent_op,
2149                                 int insert_reserved)
2150 {
2151         int ret = 0;
2152         struct btrfs_delayed_tree_ref *ref;
2153         struct btrfs_key ins;
2154         u64 parent = 0;
2155         u64 ref_root = 0;
2156         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2157                                                  SKINNY_METADATA);
2158
2159         ref = btrfs_delayed_node_to_tree_ref(node);
2160         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2161                 parent = ref->parent;
2162         else
2163                 ref_root = ref->root;
2164
2165         ins.objectid = node->bytenr;
2166         if (skinny_metadata) {
2167                 ins.offset = ref->level;
2168                 ins.type = BTRFS_METADATA_ITEM_KEY;
2169         } else {
2170                 ins.offset = node->num_bytes;
2171                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2172         }
2173
2174         BUG_ON(node->ref_mod != 1);
2175         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2176                 BUG_ON(!extent_op || !extent_op->update_flags);
2177                 ret = alloc_reserved_tree_block(trans, root,
2178                                                 parent, ref_root,
2179                                                 extent_op->flags_to_set,
2180                                                 &extent_op->key,
2181                                                 ref->level, &ins);
2182         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2183                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2184                                              node->num_bytes, parent, ref_root,
2185                                              ref->level, 0, 1, extent_op);
2186         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2187                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2188                                           node->num_bytes, parent, ref_root,
2189                                           ref->level, 0, 1, extent_op);
2190         } else {
2191                 BUG();
2192         }
2193         return ret;
2194 }
2195
2196 /* helper function to actually process a single delayed ref entry */
2197 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2198                                struct btrfs_root *root,
2199                                struct btrfs_delayed_ref_node *node,
2200                                struct btrfs_delayed_extent_op *extent_op,
2201                                int insert_reserved)
2202 {
2203         int ret = 0;
2204
2205         if (trans->aborted)
2206                 return 0;
2207
2208         if (btrfs_delayed_ref_is_head(node)) {
2209                 struct btrfs_delayed_ref_head *head;
2210                 /*
2211                  * we've hit the end of the chain and we were supposed
2212                  * to insert this extent into the tree.  But, it got
2213                  * deleted before we ever needed to insert it, so all
2214                  * we have to do is clean up the accounting
2215                  */
2216                 BUG_ON(extent_op);
2217                 head = btrfs_delayed_node_to_head(node);
2218                 if (insert_reserved) {
2219                         btrfs_pin_extent(root, node->bytenr,
2220                                          node->num_bytes, 1);
2221                         if (head->is_data) {
2222                                 ret = btrfs_del_csums(trans, root,
2223                                                       node->bytenr,
2224                                                       node->num_bytes);
2225                         }
2226                 }
2227                 return ret;
2228         }
2229
2230         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2231             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2232                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2233                                            insert_reserved);
2234         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2235                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2236                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2237                                            insert_reserved);
2238         else
2239                 BUG();
2240         return ret;
2241 }
2242
2243 static noinline struct btrfs_delayed_ref_node *
2244 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2245 {
2246         struct rb_node *node;
2247         struct btrfs_delayed_ref_node *ref;
2248         int action = BTRFS_ADD_DELAYED_REF;
2249 again:
2250         /*
2251          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2252          * this prevents ref count from going down to zero when
2253          * there still are pending delayed ref.
2254          */
2255         node = rb_prev(&head->node.rb_node);
2256         while (1) {
2257                 if (!node)
2258                         break;
2259                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2260                                 rb_node);
2261                 if (ref->bytenr != head->node.bytenr)
2262                         break;
2263                 if (ref->action == action)
2264                         return ref;
2265                 node = rb_prev(node);
2266         }
2267         if (action == BTRFS_ADD_DELAYED_REF) {
2268                 action = BTRFS_DROP_DELAYED_REF;
2269                 goto again;
2270         }
2271         return NULL;
2272 }
2273
2274 /*
2275  * Returns 0 on success or if called with an already aborted transaction.
2276  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2277  */
2278 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2279                                        struct btrfs_root *root,
2280                                        struct list_head *cluster)
2281 {
2282         struct btrfs_delayed_ref_root *delayed_refs;
2283         struct btrfs_delayed_ref_node *ref;
2284         struct btrfs_delayed_ref_head *locked_ref = NULL;
2285         struct btrfs_delayed_extent_op *extent_op;
2286         struct btrfs_fs_info *fs_info = root->fs_info;
2287         int ret;
2288         int count = 0;
2289         int must_insert_reserved = 0;
2290
2291         delayed_refs = &trans->transaction->delayed_refs;
2292         while (1) {
2293                 if (!locked_ref) {
2294                         /* pick a new head ref from the cluster list */
2295                         if (list_empty(cluster))
2296                                 break;
2297
2298                         locked_ref = list_entry(cluster->next,
2299                                      struct btrfs_delayed_ref_head, cluster);
2300
2301                         /* grab the lock that says we are going to process
2302                          * all the refs for this head */
2303                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2304
2305                         /*
2306                          * we may have dropped the spin lock to get the head
2307                          * mutex lock, and that might have given someone else
2308                          * time to free the head.  If that's true, it has been
2309                          * removed from our list and we can move on.
2310                          */
2311                         if (ret == -EAGAIN) {
2312                                 locked_ref = NULL;
2313                                 count++;
2314                                 continue;
2315                         }
2316                 }
2317
2318                 /*
2319                  * We need to try and merge add/drops of the same ref since we
2320                  * can run into issues with relocate dropping the implicit ref
2321                  * and then it being added back again before the drop can
2322                  * finish.  If we merged anything we need to re-loop so we can
2323                  * get a good ref.
2324                  */
2325                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2326                                          locked_ref);
2327
2328                 /*
2329                  * locked_ref is the head node, so we have to go one
2330                  * node back for any delayed ref updates
2331                  */
2332                 ref = select_delayed_ref(locked_ref);
2333
2334                 if (ref && ref->seq &&
2335                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2336                         /*
2337                          * there are still refs with lower seq numbers in the
2338                          * process of being added. Don't run this ref yet.
2339                          */
2340                         list_del_init(&locked_ref->cluster);
2341                         btrfs_delayed_ref_unlock(locked_ref);
2342                         locked_ref = NULL;
2343                         delayed_refs->num_heads_ready++;
2344                         spin_unlock(&delayed_refs->lock);
2345                         cond_resched();
2346                         spin_lock(&delayed_refs->lock);
2347                         continue;
2348                 }
2349
2350                 /*
2351                  * record the must insert reserved flag before we
2352                  * drop the spin lock.
2353                  */
2354                 must_insert_reserved = locked_ref->must_insert_reserved;
2355                 locked_ref->must_insert_reserved = 0;
2356
2357                 extent_op = locked_ref->extent_op;
2358                 locked_ref->extent_op = NULL;
2359
2360                 if (!ref) {
2361                         /* All delayed refs have been processed, Go ahead
2362                          * and send the head node to run_one_delayed_ref,
2363                          * so that any accounting fixes can happen
2364                          */
2365                         ref = &locked_ref->node;
2366
2367                         if (extent_op && must_insert_reserved) {
2368                                 btrfs_free_delayed_extent_op(extent_op);
2369                                 extent_op = NULL;
2370                         }
2371
2372                         if (extent_op) {
2373                                 spin_unlock(&delayed_refs->lock);
2374
2375                                 ret = run_delayed_extent_op(trans, root,
2376                                                             ref, extent_op);
2377                                 btrfs_free_delayed_extent_op(extent_op);
2378
2379                                 if (ret) {
2380                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2381                                         spin_lock(&delayed_refs->lock);
2382                                         btrfs_delayed_ref_unlock(locked_ref);
2383                                         return ret;
2384                                 }
2385
2386                                 goto next;
2387                         }
2388                 }
2389
2390                 ref->in_tree = 0;
2391                 rb_erase(&ref->rb_node, &delayed_refs->root);
2392                 delayed_refs->num_entries--;
2393                 if (!btrfs_delayed_ref_is_head(ref)) {
2394                         /*
2395                          * when we play the delayed ref, also correct the
2396                          * ref_mod on head
2397                          */
2398                         switch (ref->action) {
2399                         case BTRFS_ADD_DELAYED_REF:
2400                         case BTRFS_ADD_DELAYED_EXTENT:
2401                                 locked_ref->node.ref_mod -= ref->ref_mod;
2402                                 break;
2403                         case BTRFS_DROP_DELAYED_REF:
2404                                 locked_ref->node.ref_mod += ref->ref_mod;
2405                                 break;
2406                         default:
2407                                 WARN_ON(1);
2408                         }
2409                 }
2410                 spin_unlock(&delayed_refs->lock);
2411
2412                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2413                                           must_insert_reserved);
2414
2415                 btrfs_free_delayed_extent_op(extent_op);
2416                 if (ret) {
2417                         btrfs_delayed_ref_unlock(locked_ref);
2418                         btrfs_put_delayed_ref(ref);
2419                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2420                         spin_lock(&delayed_refs->lock);
2421                         return ret;
2422                 }
2423
2424                 /*
2425                  * If this node is a head, that means all the refs in this head
2426                  * have been dealt with, and we will pick the next head to deal
2427                  * with, so we must unlock the head and drop it from the cluster
2428                  * list before we release it.
2429                  */
2430                 if (btrfs_delayed_ref_is_head(ref)) {
2431                         list_del_init(&locked_ref->cluster);
2432                         btrfs_delayed_ref_unlock(locked_ref);
2433                         locked_ref = NULL;
2434                 }
2435                 btrfs_put_delayed_ref(ref);
2436                 count++;
2437 next:
2438                 cond_resched();
2439                 spin_lock(&delayed_refs->lock);
2440         }
2441         return count;
2442 }
2443
2444 #ifdef SCRAMBLE_DELAYED_REFS
2445 /*
2446  * Normally delayed refs get processed in ascending bytenr order. This
2447  * correlates in most cases to the order added. To expose dependencies on this
2448  * order, we start to process the tree in the middle instead of the beginning
2449  */
2450 static u64 find_middle(struct rb_root *root)
2451 {
2452         struct rb_node *n = root->rb_node;
2453         struct btrfs_delayed_ref_node *entry;
2454         int alt = 1;
2455         u64 middle;
2456         u64 first = 0, last = 0;
2457
2458         n = rb_first(root);
2459         if (n) {
2460                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2461                 first = entry->bytenr;
2462         }
2463         n = rb_last(root);
2464         if (n) {
2465                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2466                 last = entry->bytenr;
2467         }
2468         n = root->rb_node;
2469
2470         while (n) {
2471                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2472                 WARN_ON(!entry->in_tree);
2473
2474                 middle = entry->bytenr;
2475
2476                 if (alt)
2477                         n = n->rb_left;
2478                 else
2479                         n = n->rb_right;
2480
2481                 alt = 1 - alt;
2482         }
2483         return middle;
2484 }
2485 #endif
2486
2487 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2488                                          struct btrfs_fs_info *fs_info)
2489 {
2490         struct qgroup_update *qgroup_update;
2491         int ret = 0;
2492
2493         if (list_empty(&trans->qgroup_ref_list) !=
2494             !trans->delayed_ref_elem.seq) {
2495                 /* list without seq or seq without list */
2496                 btrfs_err(fs_info,
2497                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2498                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2499                         (u32)(trans->delayed_ref_elem.seq >> 32),
2500                         (u32)trans->delayed_ref_elem.seq);
2501                 BUG();
2502         }
2503
2504         if (!trans->delayed_ref_elem.seq)
2505                 return 0;
2506
2507         while (!list_empty(&trans->qgroup_ref_list)) {
2508                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2509                                                  struct qgroup_update, list);
2510                 list_del(&qgroup_update->list);
2511                 if (!ret)
2512                         ret = btrfs_qgroup_account_ref(
2513                                         trans, fs_info, qgroup_update->node,
2514                                         qgroup_update->extent_op);
2515                 kfree(qgroup_update);
2516         }
2517
2518         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2519
2520         return ret;
2521 }
2522
2523 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2524                       int count)
2525 {
2526         int val = atomic_read(&delayed_refs->ref_seq);
2527
2528         if (val < seq || val >= seq + count)
2529                 return 1;
2530         return 0;
2531 }
2532
2533 /*
2534  * this starts processing the delayed reference count updates and
2535  * extent insertions we have queued up so far.  count can be
2536  * 0, which means to process everything in the tree at the start
2537  * of the run (but not newly added entries), or it can be some target
2538  * number you'd like to process.
2539  *
2540  * Returns 0 on success or if called with an aborted transaction
2541  * Returns <0 on error and aborts the transaction
2542  */
2543 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2544                            struct btrfs_root *root, unsigned long count)
2545 {
2546         struct rb_node *node;
2547         struct btrfs_delayed_ref_root *delayed_refs;
2548         struct btrfs_delayed_ref_node *ref;
2549         struct list_head cluster;
2550         int ret;
2551         u64 delayed_start;
2552         int run_all = count == (unsigned long)-1;
2553         int run_most = 0;
2554         int loops;
2555
2556         /* We'll clean this up in btrfs_cleanup_transaction */
2557         if (trans->aborted)
2558                 return 0;
2559
2560         if (root == root->fs_info->extent_root)
2561                 root = root->fs_info->tree_root;
2562
2563         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2564
2565         delayed_refs = &trans->transaction->delayed_refs;
2566         INIT_LIST_HEAD(&cluster);
2567         if (count == 0) {
2568                 count = delayed_refs->num_entries * 2;
2569                 run_most = 1;
2570         }
2571
2572         if (!run_all && !run_most) {
2573                 int old;
2574                 int seq = atomic_read(&delayed_refs->ref_seq);
2575
2576 progress:
2577                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2578                 if (old) {
2579                         DEFINE_WAIT(__wait);
2580                         if (delayed_refs->num_entries < 16348)
2581                                 return 0;
2582
2583                         prepare_to_wait(&delayed_refs->wait, &__wait,
2584                                         TASK_UNINTERRUPTIBLE);
2585
2586                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2587                         if (old) {
2588                                 schedule();
2589                                 finish_wait(&delayed_refs->wait, &__wait);
2590
2591                                 if (!refs_newer(delayed_refs, seq, 256))
2592                                         goto progress;
2593                                 else
2594                                         return 0;
2595                         } else {
2596                                 finish_wait(&delayed_refs->wait, &__wait);
2597                                 goto again;
2598                         }
2599                 }
2600
2601         } else {
2602                 atomic_inc(&delayed_refs->procs_running_refs);
2603         }
2604
2605 again:
2606         loops = 0;
2607         spin_lock(&delayed_refs->lock);
2608
2609 #ifdef SCRAMBLE_DELAYED_REFS
2610         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2611 #endif
2612
2613         while (1) {
2614                 if (!(run_all || run_most) &&
2615                     delayed_refs->num_heads_ready < 64)
2616                         break;
2617
2618                 /*
2619                  * go find something we can process in the rbtree.  We start at
2620                  * the beginning of the tree, and then build a cluster
2621                  * of refs to process starting at the first one we are able to
2622                  * lock
2623                  */
2624                 delayed_start = delayed_refs->run_delayed_start;
2625                 ret = btrfs_find_ref_cluster(trans, &cluster,
2626                                              delayed_refs->run_delayed_start);
2627                 if (ret)
2628                         break;
2629
2630                 ret = run_clustered_refs(trans, root, &cluster);
2631                 if (ret < 0) {
2632                         btrfs_release_ref_cluster(&cluster);
2633                         spin_unlock(&delayed_refs->lock);
2634                         btrfs_abort_transaction(trans, root, ret);
2635                         atomic_dec(&delayed_refs->procs_running_refs);
2636                         return ret;
2637                 }
2638
2639                 atomic_add(ret, &delayed_refs->ref_seq);
2640
2641                 count -= min_t(unsigned long, ret, count);
2642
2643                 if (count == 0)
2644                         break;
2645
2646                 if (delayed_start >= delayed_refs->run_delayed_start) {
2647                         if (loops == 0) {
2648                                 /*
2649                                  * btrfs_find_ref_cluster looped. let's do one
2650                                  * more cycle. if we don't run any delayed ref
2651                                  * during that cycle (because we can't because
2652                                  * all of them are blocked), bail out.
2653                                  */
2654                                 loops = 1;
2655                         } else {
2656                                 /*
2657                                  * no runnable refs left, stop trying
2658                                  */
2659                                 BUG_ON(run_all);
2660                                 break;
2661                         }
2662                 }
2663                 if (ret) {
2664                         /* refs were run, let's reset staleness detection */
2665                         loops = 0;
2666                 }
2667         }
2668
2669         if (run_all) {
2670                 if (!list_empty(&trans->new_bgs)) {
2671                         spin_unlock(&delayed_refs->lock);
2672                         btrfs_create_pending_block_groups(trans, root);
2673                         spin_lock(&delayed_refs->lock);
2674                 }
2675
2676                 node = rb_first(&delayed_refs->root);
2677                 if (!node)
2678                         goto out;
2679                 count = (unsigned long)-1;
2680
2681                 while (node) {
2682                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2683                                        rb_node);
2684                         if (btrfs_delayed_ref_is_head(ref)) {
2685                                 struct btrfs_delayed_ref_head *head;
2686
2687                                 head = btrfs_delayed_node_to_head(ref);
2688                                 atomic_inc(&ref->refs);
2689
2690                                 spin_unlock(&delayed_refs->lock);
2691                                 /*
2692                                  * Mutex was contended, block until it's
2693                                  * released and try again
2694                                  */
2695                                 mutex_lock(&head->mutex);
2696                                 mutex_unlock(&head->mutex);
2697
2698                                 btrfs_put_delayed_ref(ref);
2699                                 cond_resched();
2700                                 goto again;
2701                         }
2702                         node = rb_next(node);
2703                 }
2704                 spin_unlock(&delayed_refs->lock);
2705                 schedule_timeout(1);
2706                 goto again;
2707         }
2708 out:
2709         atomic_dec(&delayed_refs->procs_running_refs);
2710         smp_mb();
2711         if (waitqueue_active(&delayed_refs->wait))
2712                 wake_up(&delayed_refs->wait);
2713
2714         spin_unlock(&delayed_refs->lock);
2715         assert_qgroups_uptodate(trans);
2716         return 0;
2717 }
2718
2719 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2720                                 struct btrfs_root *root,
2721                                 u64 bytenr, u64 num_bytes, u64 flags,
2722                                 int is_data)
2723 {
2724         struct btrfs_delayed_extent_op *extent_op;
2725         int ret;
2726
2727         extent_op = btrfs_alloc_delayed_extent_op();
2728         if (!extent_op)
2729                 return -ENOMEM;
2730
2731         extent_op->flags_to_set = flags;
2732         extent_op->update_flags = 1;
2733         extent_op->update_key = 0;
2734         extent_op->is_data = is_data ? 1 : 0;
2735
2736         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2737                                           num_bytes, extent_op);
2738         if (ret)
2739                 btrfs_free_delayed_extent_op(extent_op);
2740         return ret;
2741 }
2742
2743 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2744                                       struct btrfs_root *root,
2745                                       struct btrfs_path *path,
2746                                       u64 objectid, u64 offset, u64 bytenr)
2747 {
2748         struct btrfs_delayed_ref_head *head;
2749         struct btrfs_delayed_ref_node *ref;
2750         struct btrfs_delayed_data_ref *data_ref;
2751         struct btrfs_delayed_ref_root *delayed_refs;
2752         struct rb_node *node;
2753         int ret = 0;
2754
2755         ret = -ENOENT;
2756         delayed_refs = &trans->transaction->delayed_refs;
2757         spin_lock(&delayed_refs->lock);
2758         head = btrfs_find_delayed_ref_head(trans, bytenr);
2759         if (!head)
2760                 goto out;
2761
2762         if (!mutex_trylock(&head->mutex)) {
2763                 atomic_inc(&head->node.refs);
2764                 spin_unlock(&delayed_refs->lock);
2765
2766                 btrfs_release_path(path);
2767
2768                 /*
2769                  * Mutex was contended, block until it's released and let
2770                  * caller try again
2771                  */
2772                 mutex_lock(&head->mutex);
2773                 mutex_unlock(&head->mutex);
2774                 btrfs_put_delayed_ref(&head->node);
2775                 return -EAGAIN;
2776         }
2777
2778         node = rb_prev(&head->node.rb_node);
2779         if (!node)
2780                 goto out_unlock;
2781
2782         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2783
2784         if (ref->bytenr != bytenr)
2785                 goto out_unlock;
2786
2787         ret = 1;
2788         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2789                 goto out_unlock;
2790
2791         data_ref = btrfs_delayed_node_to_data_ref(ref);
2792
2793         node = rb_prev(node);
2794         if (node) {
2795                 int seq = ref->seq;
2796
2797                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2798                 if (ref->bytenr == bytenr && ref->seq == seq)
2799                         goto out_unlock;
2800         }
2801
2802         if (data_ref->root != root->root_key.objectid ||
2803             data_ref->objectid != objectid || data_ref->offset != offset)
2804                 goto out_unlock;
2805
2806         ret = 0;
2807 out_unlock:
2808         mutex_unlock(&head->mutex);
2809 out:
2810         spin_unlock(&delayed_refs->lock);
2811         return ret;
2812 }
2813
2814 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2815                                         struct btrfs_root *root,
2816                                         struct btrfs_path *path,
2817                                         u64 objectid, u64 offset, u64 bytenr)
2818 {
2819         struct btrfs_root *extent_root = root->fs_info->extent_root;
2820         struct extent_buffer *leaf;
2821         struct btrfs_extent_data_ref *ref;
2822         struct btrfs_extent_inline_ref *iref;
2823         struct btrfs_extent_item *ei;
2824         struct btrfs_key key;
2825         u32 item_size;
2826         int ret;
2827
2828         key.objectid = bytenr;
2829         key.offset = (u64)-1;
2830         key.type = BTRFS_EXTENT_ITEM_KEY;
2831
2832         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2833         if (ret < 0)
2834                 goto out;
2835         BUG_ON(ret == 0); /* Corruption */
2836
2837         ret = -ENOENT;
2838         if (path->slots[0] == 0)
2839                 goto out;
2840
2841         path->slots[0]--;
2842         leaf = path->nodes[0];
2843         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2844
2845         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2846                 goto out;
2847
2848         ret = 1;
2849         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2850 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2851         if (item_size < sizeof(*ei)) {
2852                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2853                 goto out;
2854         }
2855 #endif
2856         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2857
2858         if (item_size != sizeof(*ei) +
2859             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2860                 goto out;
2861
2862         if (btrfs_extent_generation(leaf, ei) <=
2863             btrfs_root_last_snapshot(&root->root_item))
2864                 goto out;
2865
2866         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2867         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2868             BTRFS_EXTENT_DATA_REF_KEY)
2869                 goto out;
2870
2871         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2872         if (btrfs_extent_refs(leaf, ei) !=
2873             btrfs_extent_data_ref_count(leaf, ref) ||
2874             btrfs_extent_data_ref_root(leaf, ref) !=
2875             root->root_key.objectid ||
2876             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2877             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2878                 goto out;
2879
2880         ret = 0;
2881 out:
2882         return ret;
2883 }
2884
2885 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2886                           struct btrfs_root *root,
2887                           u64 objectid, u64 offset, u64 bytenr)
2888 {
2889         struct btrfs_path *path;
2890         int ret;
2891         int ret2;
2892
2893         path = btrfs_alloc_path();
2894         if (!path)
2895                 return -ENOENT;
2896
2897         do {
2898                 ret = check_committed_ref(trans, root, path, objectid,
2899                                           offset, bytenr);
2900                 if (ret && ret != -ENOENT)
2901                         goto out;
2902
2903                 ret2 = check_delayed_ref(trans, root, path, objectid,
2904                                          offset, bytenr);
2905         } while (ret2 == -EAGAIN);
2906
2907         if (ret2 && ret2 != -ENOENT) {
2908                 ret = ret2;
2909                 goto out;
2910         }
2911
2912         if (ret != -ENOENT || ret2 != -ENOENT)
2913                 ret = 0;
2914 out:
2915         btrfs_free_path(path);
2916         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2917                 WARN_ON(ret > 0);
2918         return ret;
2919 }
2920
2921 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2922                            struct btrfs_root *root,
2923                            struct extent_buffer *buf,
2924                            int full_backref, int inc, int for_cow)
2925 {
2926         u64 bytenr;
2927         u64 num_bytes;
2928         u64 parent;
2929         u64 ref_root;
2930         u32 nritems;
2931         struct btrfs_key key;
2932         struct btrfs_file_extent_item *fi;
2933         int i;
2934         int level;
2935         int ret = 0;
2936         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2937                             u64, u64, u64, u64, u64, u64, int);
2938
2939         ref_root = btrfs_header_owner(buf);
2940         nritems = btrfs_header_nritems(buf);
2941         level = btrfs_header_level(buf);
2942
2943         if (!root->ref_cows && level == 0)
2944                 return 0;
2945
2946         if (inc)
2947                 process_func = btrfs_inc_extent_ref;
2948         else
2949                 process_func = btrfs_free_extent;
2950
2951         if (full_backref)
2952                 parent = buf->start;
2953         else
2954                 parent = 0;
2955
2956         for (i = 0; i < nritems; i++) {
2957                 if (level == 0) {
2958                         btrfs_item_key_to_cpu(buf, &key, i);
2959                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2960                                 continue;
2961                         fi = btrfs_item_ptr(buf, i,
2962                                             struct btrfs_file_extent_item);
2963                         if (btrfs_file_extent_type(buf, fi) ==
2964                             BTRFS_FILE_EXTENT_INLINE)
2965                                 continue;
2966                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2967                         if (bytenr == 0)
2968                                 continue;
2969
2970                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2971                         key.offset -= btrfs_file_extent_offset(buf, fi);
2972                         ret = process_func(trans, root, bytenr, num_bytes,
2973                                            parent, ref_root, key.objectid,
2974                                            key.offset, for_cow);
2975                         if (ret)
2976                                 goto fail;
2977                 } else {
2978                         bytenr = btrfs_node_blockptr(buf, i);
2979                         num_bytes = btrfs_level_size(root, level - 1);
2980                         ret = process_func(trans, root, bytenr, num_bytes,
2981                                            parent, ref_root, level - 1, 0,
2982                                            for_cow);
2983                         if (ret)
2984                                 goto fail;
2985                 }
2986         }
2987         return 0;
2988 fail:
2989         return ret;
2990 }
2991
2992 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2993                   struct extent_buffer *buf, int full_backref, int for_cow)
2994 {
2995         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2996 }
2997
2998 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2999                   struct extent_buffer *buf, int full_backref, int for_cow)
3000 {
3001         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3002 }
3003
3004 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3005                                  struct btrfs_root *root,
3006                                  struct btrfs_path *path,
3007                                  struct btrfs_block_group_cache *cache)
3008 {
3009         int ret;
3010         struct btrfs_root *extent_root = root->fs_info->extent_root;
3011         unsigned long bi;
3012         struct extent_buffer *leaf;
3013
3014         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3015         if (ret < 0)
3016                 goto fail;
3017         BUG_ON(ret); /* Corruption */
3018
3019         leaf = path->nodes[0];
3020         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3021         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3022         btrfs_mark_buffer_dirty(leaf);
3023         btrfs_release_path(path);
3024 fail:
3025         if (ret) {
3026                 btrfs_abort_transaction(trans, root, ret);
3027                 return ret;
3028         }
3029         return 0;
3030
3031 }
3032
3033 static struct btrfs_block_group_cache *
3034 next_block_group(struct btrfs_root *root,
3035                  struct btrfs_block_group_cache *cache)
3036 {
3037         struct rb_node *node;
3038         spin_lock(&root->fs_info->block_group_cache_lock);
3039         node = rb_next(&cache->cache_node);
3040         btrfs_put_block_group(cache);
3041         if (node) {
3042                 cache = rb_entry(node, struct btrfs_block_group_cache,
3043                                  cache_node);
3044                 btrfs_get_block_group(cache);
3045         } else
3046                 cache = NULL;
3047         spin_unlock(&root->fs_info->block_group_cache_lock);
3048         return cache;
3049 }
3050
3051 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3052                             struct btrfs_trans_handle *trans,
3053                             struct btrfs_path *path)
3054 {
3055         struct btrfs_root *root = block_group->fs_info->tree_root;
3056         struct inode *inode = NULL;
3057         u64 alloc_hint = 0;
3058         int dcs = BTRFS_DC_ERROR;
3059         int num_pages = 0;
3060         int retries = 0;
3061         int ret = 0;
3062
3063         /*
3064          * If this block group is smaller than 100 megs don't bother caching the
3065          * block group.
3066          */
3067         if (block_group->key.offset < (100 * 1024 * 1024)) {
3068                 spin_lock(&block_group->lock);
3069                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3070                 spin_unlock(&block_group->lock);
3071                 return 0;
3072         }
3073
3074 again:
3075         inode = lookup_free_space_inode(root, block_group, path);
3076         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3077                 ret = PTR_ERR(inode);
3078                 btrfs_release_path(path);
3079                 goto out;
3080         }
3081
3082         if (IS_ERR(inode)) {
3083                 BUG_ON(retries);
3084                 retries++;
3085
3086                 if (block_group->ro)
3087                         goto out_free;
3088
3089                 ret = create_free_space_inode(root, trans, block_group, path);
3090                 if (ret)
3091                         goto out_free;
3092                 goto again;
3093         }
3094
3095         /* We've already setup this transaction, go ahead and exit */
3096         if (block_group->cache_generation == trans->transid &&
3097             i_size_read(inode)) {
3098                 dcs = BTRFS_DC_SETUP;
3099                 goto out_put;
3100         }
3101
3102         /*
3103          * We want to set the generation to 0, that way if anything goes wrong
3104          * from here on out we know not to trust this cache when we load up next
3105          * time.
3106          */
3107         BTRFS_I(inode)->generation = 0;
3108         ret = btrfs_update_inode(trans, root, inode);
3109         WARN_ON(ret);
3110
3111         if (i_size_read(inode) > 0) {
3112                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3113                                                       inode);
3114                 if (ret)
3115                         goto out_put;
3116         }
3117
3118         spin_lock(&block_group->lock);
3119         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3120             !btrfs_test_opt(root, SPACE_CACHE)) {
3121                 /*
3122                  * don't bother trying to write stuff out _if_
3123                  * a) we're not cached,
3124                  * b) we're with nospace_cache mount option.
3125                  */
3126                 dcs = BTRFS_DC_WRITTEN;
3127                 spin_unlock(&block_group->lock);
3128                 goto out_put;
3129         }
3130         spin_unlock(&block_group->lock);
3131
3132         /*
3133          * Try to preallocate enough space based on how big the block group is.
3134          * Keep in mind this has to include any pinned space which could end up
3135          * taking up quite a bit since it's not folded into the other space
3136          * cache.
3137          */
3138         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3139         if (!num_pages)
3140                 num_pages = 1;
3141
3142         num_pages *= 16;
3143         num_pages *= PAGE_CACHE_SIZE;
3144
3145         ret = btrfs_check_data_free_space(inode, num_pages);
3146         if (ret)
3147                 goto out_put;
3148
3149         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3150                                               num_pages, num_pages,
3151                                               &alloc_hint);
3152         if (!ret)
3153                 dcs = BTRFS_DC_SETUP;
3154         btrfs_free_reserved_data_space(inode, num_pages);
3155
3156 out_put:
3157         iput(inode);
3158 out_free:
3159         btrfs_release_path(path);
3160 out:
3161         spin_lock(&block_group->lock);
3162         if (!ret && dcs == BTRFS_DC_SETUP)
3163                 block_group->cache_generation = trans->transid;
3164         block_group->disk_cache_state = dcs;
3165         spin_unlock(&block_group->lock);
3166
3167         return ret;
3168 }
3169
3170 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3171                                    struct btrfs_root *root)
3172 {
3173         struct btrfs_block_group_cache *cache;
3174         int err = 0;
3175         struct btrfs_path *path;
3176         u64 last = 0;
3177
3178         path = btrfs_alloc_path();
3179         if (!path)
3180                 return -ENOMEM;
3181
3182 again:
3183         while (1) {
3184                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3185                 while (cache) {
3186                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3187                                 break;
3188                         cache = next_block_group(root, cache);
3189                 }
3190                 if (!cache) {
3191                         if (last == 0)
3192                                 break;
3193                         last = 0;
3194                         continue;
3195                 }
3196                 err = cache_save_setup(cache, trans, path);
3197                 last = cache->key.objectid + cache->key.offset;
3198                 btrfs_put_block_group(cache);
3199         }
3200
3201         while (1) {
3202                 if (last == 0) {
3203                         err = btrfs_run_delayed_refs(trans, root,
3204                                                      (unsigned long)-1);
3205                         if (err) /* File system offline */
3206                                 goto out;
3207                 }
3208
3209                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3210                 while (cache) {
3211                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3212                                 btrfs_put_block_group(cache);
3213                                 goto again;
3214                         }
3215
3216                         if (cache->dirty)
3217                                 break;
3218                         cache = next_block_group(root, cache);
3219                 }
3220                 if (!cache) {
3221                         if (last == 0)
3222                                 break;
3223                         last = 0;
3224                         continue;
3225                 }
3226
3227                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3228                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3229                 cache->dirty = 0;
3230                 last = cache->key.objectid + cache->key.offset;
3231
3232                 err = write_one_cache_group(trans, root, path, cache);
3233                 if (err) /* File system offline */
3234                         goto out;
3235
3236                 btrfs_put_block_group(cache);
3237         }
3238
3239         while (1) {
3240                 /*
3241                  * I don't think this is needed since we're just marking our
3242                  * preallocated extent as written, but just in case it can't
3243                  * hurt.
3244                  */
3245                 if (last == 0) {
3246                         err = btrfs_run_delayed_refs(trans, root,
3247                                                      (unsigned long)-1);
3248                         if (err) /* File system offline */
3249                                 goto out;
3250                 }
3251
3252                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3253                 while (cache) {
3254                         /*
3255                          * Really this shouldn't happen, but it could if we
3256                          * couldn't write the entire preallocated extent and
3257                          * splitting the extent resulted in a new block.
3258                          */
3259                         if (cache->dirty) {
3260                                 btrfs_put_block_group(cache);
3261                                 goto again;
3262                         }
3263                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3264                                 break;
3265                         cache = next_block_group(root, cache);
3266                 }
3267                 if (!cache) {
3268                         if (last == 0)
3269                                 break;
3270                         last = 0;
3271                         continue;
3272                 }
3273
3274                 err = btrfs_write_out_cache(root, trans, cache, path);
3275
3276                 /*
3277                  * If we didn't have an error then the cache state is still
3278                  * NEED_WRITE, so we can set it to WRITTEN.
3279                  */
3280                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3281                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3282                 last = cache->key.objectid + cache->key.offset;
3283                 btrfs_put_block_group(cache);
3284         }
3285 out:
3286
3287         btrfs_free_path(path);
3288         return err;
3289 }
3290
3291 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3292 {
3293         struct btrfs_block_group_cache *block_group;
3294         int readonly = 0;
3295
3296         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3297         if (!block_group || block_group->ro)
3298                 readonly = 1;
3299         if (block_group)
3300                 btrfs_put_block_group(block_group);
3301         return readonly;
3302 }
3303
3304 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3305                              u64 total_bytes, u64 bytes_used,
3306                              struct btrfs_space_info **space_info)
3307 {
3308         struct btrfs_space_info *found;
3309         int i;
3310         int factor;
3311
3312         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3313                      BTRFS_BLOCK_GROUP_RAID10))
3314                 factor = 2;
3315         else
3316                 factor = 1;
3317
3318         found = __find_space_info(info, flags);
3319         if (found) {
3320                 spin_lock(&found->lock);
3321                 found->total_bytes += total_bytes;
3322                 found->disk_total += total_bytes * factor;
3323                 found->bytes_used += bytes_used;
3324                 found->disk_used += bytes_used * factor;
3325                 found->full = 0;
3326                 spin_unlock(&found->lock);
3327                 *space_info = found;
3328                 return 0;
3329         }
3330         found = kzalloc(sizeof(*found), GFP_NOFS);
3331         if (!found)
3332                 return -ENOMEM;
3333
3334         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3335                 INIT_LIST_HEAD(&found->block_groups[i]);
3336         init_rwsem(&found->groups_sem);
3337         spin_lock_init(&found->lock);
3338         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3339         found->total_bytes = total_bytes;
3340         found->disk_total = total_bytes * factor;
3341         found->bytes_used = bytes_used;
3342         found->disk_used = bytes_used * factor;
3343         found->bytes_pinned = 0;
3344         found->bytes_reserved = 0;
3345         found->bytes_readonly = 0;
3346         found->bytes_may_use = 0;
3347         found->full = 0;
3348         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3349         found->chunk_alloc = 0;
3350         found->flush = 0;
3351         init_waitqueue_head(&found->wait);
3352         *space_info = found;
3353         list_add_rcu(&found->list, &info->space_info);
3354         if (flags & BTRFS_BLOCK_GROUP_DATA)
3355                 info->data_sinfo = found;
3356         return 0;
3357 }
3358
3359 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3360 {
3361         u64 extra_flags = chunk_to_extended(flags) &
3362                                 BTRFS_EXTENDED_PROFILE_MASK;
3363
3364         write_seqlock(&fs_info->profiles_lock);
3365         if (flags & BTRFS_BLOCK_GROUP_DATA)
3366                 fs_info->avail_data_alloc_bits |= extra_flags;
3367         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3368                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3369         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3370                 fs_info->avail_system_alloc_bits |= extra_flags;
3371         write_sequnlock(&fs_info->profiles_lock);
3372 }
3373
3374 /*
3375  * returns target flags in extended format or 0 if restripe for this
3376  * chunk_type is not in progress
3377  *
3378  * should be called with either volume_mutex or balance_lock held
3379  */
3380 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3381 {
3382         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3383         u64 target = 0;
3384
3385         if (!bctl)
3386                 return 0;
3387
3388         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3389             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3390                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3391         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3392                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3393                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3394         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3395                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3396                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3397         }
3398
3399         return target;
3400 }
3401
3402 /*
3403  * @flags: available profiles in extended format (see ctree.h)
3404  *
3405  * Returns reduced profile in chunk format.  If profile changing is in
3406  * progress (either running or paused) picks the target profile (if it's
3407  * already available), otherwise falls back to plain reducing.
3408  */
3409 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3410 {
3411         /*
3412          * we add in the count of missing devices because we want
3413          * to make sure that any RAID levels on a degraded FS
3414          * continue to be honored.
3415          */
3416         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3417                 root->fs_info->fs_devices->missing_devices;
3418         u64 target;
3419         u64 tmp;
3420
3421         /*
3422          * see if restripe for this chunk_type is in progress, if so
3423          * try to reduce to the target profile
3424          */
3425         spin_lock(&root->fs_info->balance_lock);
3426         target = get_restripe_target(root->fs_info, flags);
3427         if (target) {
3428                 /* pick target profile only if it's already available */
3429                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3430                         spin_unlock(&root->fs_info->balance_lock);
3431                         return extended_to_chunk(target);
3432                 }
3433         }
3434         spin_unlock(&root->fs_info->balance_lock);
3435
3436         /* First, mask out the RAID levels which aren't possible */
3437         if (num_devices == 1)
3438                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3439                            BTRFS_BLOCK_GROUP_RAID5);
3440         if (num_devices < 3)
3441                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3442         if (num_devices < 4)
3443                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3444
3445         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3446                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3447                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3448         flags &= ~tmp;
3449
3450         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3451                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3452         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3453                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3454         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3455                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3456         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3457                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3458         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3459                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3460
3461         return extended_to_chunk(flags | tmp);
3462 }
3463
3464 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3465 {
3466         unsigned seq;
3467
3468         do {
3469                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3470
3471                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3472                         flags |= root->fs_info->avail_data_alloc_bits;
3473                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3474                         flags |= root->fs_info->avail_system_alloc_bits;
3475                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3476                         flags |= root->fs_info->avail_metadata_alloc_bits;
3477         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3478
3479         return btrfs_reduce_alloc_profile(root, flags);
3480 }
3481
3482 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3483 {
3484         u64 flags;
3485         u64 ret;
3486
3487         if (data)
3488                 flags = BTRFS_BLOCK_GROUP_DATA;
3489         else if (root == root->fs_info->chunk_root)
3490                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3491         else
3492                 flags = BTRFS_BLOCK_GROUP_METADATA;
3493
3494         ret = get_alloc_profile(root, flags);
3495         return ret;
3496 }
3497
3498 /*
3499  * This will check the space that the inode allocates from to make sure we have
3500  * enough space for bytes.
3501  */
3502 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3503 {
3504         struct btrfs_space_info *data_sinfo;
3505         struct btrfs_root *root = BTRFS_I(inode)->root;
3506         struct btrfs_fs_info *fs_info = root->fs_info;
3507         u64 used;
3508         int ret = 0, committed = 0, alloc_chunk = 1;
3509
3510         /* make sure bytes are sectorsize aligned */
3511         bytes = ALIGN(bytes, root->sectorsize);
3512
3513         if (root == root->fs_info->tree_root ||
3514             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3515                 alloc_chunk = 0;
3516                 committed = 1;
3517         }
3518
3519         data_sinfo = fs_info->data_sinfo;
3520         if (!data_sinfo)
3521                 goto alloc;
3522
3523 again:
3524         /* make sure we have enough space to handle the data first */
3525         spin_lock(&data_sinfo->lock);
3526         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3527                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3528                 data_sinfo->bytes_may_use;
3529
3530         if (used + bytes > data_sinfo->total_bytes) {
3531                 struct btrfs_trans_handle *trans;
3532
3533                 /*
3534                  * if we don't have enough free bytes in this space then we need
3535                  * to alloc a new chunk.
3536                  */
3537                 if (!data_sinfo->full && alloc_chunk) {
3538                         u64 alloc_target;
3539
3540                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3541                         spin_unlock(&data_sinfo->lock);
3542 alloc:
3543                         alloc_target = btrfs_get_alloc_profile(root, 1);
3544                         trans = btrfs_join_transaction(root);
3545                         if (IS_ERR(trans))
3546                                 return PTR_ERR(trans);
3547
3548                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3549                                              alloc_target,
3550                                              CHUNK_ALLOC_NO_FORCE);
3551                         btrfs_end_transaction(trans, root);
3552                         if (ret < 0) {
3553                                 if (ret != -ENOSPC)
3554                                         return ret;
3555                                 else
3556                                         goto commit_trans;
3557                         }
3558
3559                         if (!data_sinfo)
3560                                 data_sinfo = fs_info->data_sinfo;
3561
3562                         goto again;
3563                 }
3564
3565                 /*
3566                  * If we have less pinned bytes than we want to allocate then
3567                  * don't bother committing the transaction, it won't help us.
3568                  */
3569                 if (data_sinfo->bytes_pinned < bytes)
3570                         committed = 1;
3571                 spin_unlock(&data_sinfo->lock);
3572
3573                 /* commit the current transaction and try again */
3574 commit_trans:
3575                 if (!committed &&
3576                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3577                         committed = 1;
3578                         trans = btrfs_join_transaction(root);
3579                         if (IS_ERR(trans))
3580                                 return PTR_ERR(trans);
3581                         ret = btrfs_commit_transaction(trans, root);
3582                         if (ret)
3583                                 return ret;
3584                         goto again;
3585                 }
3586
3587                 return -ENOSPC;
3588         }
3589         data_sinfo->bytes_may_use += bytes;
3590         trace_btrfs_space_reservation(root->fs_info, "space_info",
3591                                       data_sinfo->flags, bytes, 1);
3592         spin_unlock(&data_sinfo->lock);
3593
3594         return 0;
3595 }
3596
3597 /*
3598  * Called if we need to clear a data reservation for this inode.
3599  */
3600 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3601 {
3602         struct btrfs_root *root = BTRFS_I(inode)->root;
3603         struct btrfs_space_info *data_sinfo;
3604
3605         /* make sure bytes are sectorsize aligned */
3606         bytes = ALIGN(bytes, root->sectorsize);
3607
3608         data_sinfo = root->fs_info->data_sinfo;
3609         spin_lock(&data_sinfo->lock);
3610         data_sinfo->bytes_may_use -= bytes;
3611         trace_btrfs_space_reservation(root->fs_info, "space_info",
3612                                       data_sinfo->flags, bytes, 0);
3613         spin_unlock(&data_sinfo->lock);
3614 }
3615
3616 static void force_metadata_allocation(struct btrfs_fs_info *info)
3617 {
3618         struct list_head *head = &info->space_info;
3619         struct btrfs_space_info *found;
3620
3621         rcu_read_lock();
3622         list_for_each_entry_rcu(found, head, list) {
3623                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3624                         found->force_alloc = CHUNK_ALLOC_FORCE;
3625         }
3626         rcu_read_unlock();
3627 }
3628
3629 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3630 {
3631         return (global->size << 1);
3632 }
3633
3634 static int should_alloc_chunk(struct btrfs_root *root,
3635                               struct btrfs_space_info *sinfo, int force)
3636 {
3637         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3638         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3639         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3640         u64 thresh;
3641
3642         if (force == CHUNK_ALLOC_FORCE)
3643                 return 1;
3644
3645         /*
3646          * We need to take into account the global rsv because for all intents
3647          * and purposes it's used space.  Don't worry about locking the
3648          * global_rsv, it doesn't change except when the transaction commits.
3649          */
3650         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3651                 num_allocated += calc_global_rsv_need_space(global_rsv);
3652
3653         /*
3654          * in limited mode, we want to have some free space up to
3655          * about 1% of the FS size.
3656          */
3657         if (force == CHUNK_ALLOC_LIMITED) {
3658                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3659                 thresh = max_t(u64, 64 * 1024 * 1024,
3660                                div_factor_fine(thresh, 1));
3661
3662                 if (num_bytes - num_allocated < thresh)
3663                         return 1;
3664         }
3665
3666         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3667                 return 0;
3668         return 1;
3669 }
3670
3671 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3672 {
3673         u64 num_dev;
3674
3675         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3676                     BTRFS_BLOCK_GROUP_RAID0 |
3677                     BTRFS_BLOCK_GROUP_RAID5 |
3678                     BTRFS_BLOCK_GROUP_RAID6))
3679                 num_dev = root->fs_info->fs_devices->rw_devices;
3680         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3681                 num_dev = 2;
3682         else
3683                 num_dev = 1;    /* DUP or single */
3684
3685         /* metadata for updaing devices and chunk tree */
3686         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3687 }
3688
3689 static void check_system_chunk(struct btrfs_trans_handle *trans,
3690                                struct btrfs_root *root, u64 type)
3691 {
3692         struct btrfs_space_info *info;
3693         u64 left;
3694         u64 thresh;
3695
3696         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3697         spin_lock(&info->lock);
3698         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3699                 info->bytes_reserved - info->bytes_readonly;
3700         spin_unlock(&info->lock);
3701
3702         thresh = get_system_chunk_thresh(root, type);
3703         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3704                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3705                         left, thresh, type);
3706                 dump_space_info(info, 0, 0);
3707         }
3708
3709         if (left < thresh) {
3710                 u64 flags;
3711
3712                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3713                 btrfs_alloc_chunk(trans, root, flags);
3714         }
3715 }
3716
3717 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3718                           struct btrfs_root *extent_root, u64 flags, int force)
3719 {
3720         struct btrfs_space_info *space_info;
3721         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3722         int wait_for_alloc = 0;
3723         int ret = 0;
3724
3725         /* Don't re-enter if we're already allocating a chunk */
3726         if (trans->allocating_chunk)
3727                 return -ENOSPC;
3728
3729         space_info = __find_space_info(extent_root->fs_info, flags);
3730         if (!space_info) {
3731                 ret = update_space_info(extent_root->fs_info, flags,
3732                                         0, 0, &space_info);
3733                 BUG_ON(ret); /* -ENOMEM */
3734         }
3735         BUG_ON(!space_info); /* Logic error */
3736
3737 again:
3738         spin_lock(&space_info->lock);
3739         if (force < space_info->force_alloc)
3740                 force = space_info->force_alloc;
3741         if (space_info->full) {
3742                 spin_unlock(&space_info->lock);
3743                 return 0;
3744         }
3745
3746         if (!should_alloc_chunk(extent_root, space_info, force)) {
3747                 spin_unlock(&space_info->lock);
3748                 return 0;
3749         } else if (space_info->chunk_alloc) {
3750                 wait_for_alloc = 1;
3751         } else {
3752                 space_info->chunk_alloc = 1;
3753         }
3754
3755         spin_unlock(&space_info->lock);
3756
3757         mutex_lock(&fs_info->chunk_mutex);
3758
3759         /*
3760          * The chunk_mutex is held throughout the entirety of a chunk
3761          * allocation, so once we've acquired the chunk_mutex we know that the
3762          * other guy is done and we need to recheck and see if we should
3763          * allocate.
3764          */
3765         if (wait_for_alloc) {
3766                 mutex_unlock(&fs_info->chunk_mutex);
3767                 wait_for_alloc = 0;
3768                 goto again;
3769         }
3770
3771         trans->allocating_chunk = true;
3772
3773         /*
3774          * If we have mixed data/metadata chunks we want to make sure we keep
3775          * allocating mixed chunks instead of individual chunks.
3776          */
3777         if (btrfs_mixed_space_info(space_info))
3778                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3779
3780         /*
3781          * if we're doing a data chunk, go ahead and make sure that
3782          * we keep a reasonable number of metadata chunks allocated in the
3783          * FS as well.
3784          */
3785         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3786                 fs_info->data_chunk_allocations++;
3787                 if (!(fs_info->data_chunk_allocations %
3788                       fs_info->metadata_ratio))
3789                         force_metadata_allocation(fs_info);
3790         }
3791
3792         /*
3793          * Check if we have enough space in SYSTEM chunk because we may need
3794          * to update devices.
3795          */
3796         check_system_chunk(trans, extent_root, flags);
3797
3798         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3799         trans->allocating_chunk = false;
3800
3801         spin_lock(&space_info->lock);
3802         if (ret < 0 && ret != -ENOSPC)
3803                 goto out;
3804         if (ret)
3805                 space_info->full = 1;
3806         else
3807                 ret = 1;
3808
3809         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3810 out:
3811         space_info->chunk_alloc = 0;
3812         spin_unlock(&space_info->lock);
3813         mutex_unlock(&fs_info->chunk_mutex);
3814         return ret;
3815 }
3816
3817 static int can_overcommit(struct btrfs_root *root,
3818                           struct btrfs_space_info *space_info, u64 bytes,
3819                           enum btrfs_reserve_flush_enum flush)
3820 {
3821         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3822         u64 profile = btrfs_get_alloc_profile(root, 0);
3823         u64 space_size;
3824         u64 avail;
3825         u64 used;
3826         u64 to_add;
3827
3828         used = space_info->bytes_used + space_info->bytes_reserved +
3829                 space_info->bytes_pinned + space_info->bytes_readonly;
3830
3831         /*
3832          * We only want to allow over committing if we have lots of actual space
3833          * free, but if we don't have enough space to handle the global reserve
3834          * space then we could end up having a real enospc problem when trying
3835          * to allocate a chunk or some other such important allocation.
3836          */
3837         spin_lock(&global_rsv->lock);
3838         space_size = calc_global_rsv_need_space(global_rsv);
3839         spin_unlock(&global_rsv->lock);
3840         if (used + space_size >= space_info->total_bytes)
3841                 return 0;
3842
3843         used += space_info->bytes_may_use;
3844
3845         spin_lock(&root->fs_info->free_chunk_lock);
3846         avail = root->fs_info->free_chunk_space;
3847         spin_unlock(&root->fs_info->free_chunk_lock);
3848
3849         /*
3850          * If we have dup, raid1 or raid10 then only half of the free
3851          * space is actually useable.  For raid56, the space info used
3852          * doesn't include the parity drive, so we don't have to
3853          * change the math
3854          */
3855         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3856                        BTRFS_BLOCK_GROUP_RAID1 |
3857                        BTRFS_BLOCK_GROUP_RAID10))
3858                 avail >>= 1;
3859
3860         to_add = space_info->total_bytes;
3861
3862         /*
3863          * If we aren't flushing all things, let us overcommit up to
3864          * 1/2th of the space. If we can flush, don't let us overcommit
3865          * too much, let it overcommit up to 1/8 of the space.
3866          */
3867         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3868                 to_add >>= 3;
3869         else
3870                 to_add >>= 1;
3871
3872         /*
3873          * Limit the overcommit to the amount of free space we could possibly
3874          * allocate for chunks.
3875          */
3876         to_add = min(avail, to_add);
3877
3878         if (used + bytes < space_info->total_bytes + to_add)
3879                 return 1;
3880         return 0;
3881 }
3882
3883 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3884                                          unsigned long nr_pages)
3885 {
3886         struct super_block *sb = root->fs_info->sb;
3887         int started;
3888
3889         /* If we can not start writeback, just sync all the delalloc file. */
3890         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3891                                                       WB_REASON_FS_FREE_SPACE);
3892         if (!started) {
3893                 /*
3894                  * We needn't worry the filesystem going from r/w to r/o though
3895                  * we don't acquire ->s_umount mutex, because the filesystem
3896                  * should guarantee the delalloc inodes list be empty after
3897                  * the filesystem is readonly(all dirty pages are written to
3898                  * the disk).
3899                  */
3900                 btrfs_start_delalloc_inodes(root, 0);
3901                 if (!current->journal_info)
3902                         btrfs_wait_ordered_extents(root, 0);
3903         }
3904 }
3905
3906 /*
3907  * shrink metadata reservation for delalloc
3908  */
3909 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3910                             bool wait_ordered)
3911 {
3912         struct btrfs_block_rsv *block_rsv;
3913         struct btrfs_space_info *space_info;
3914         struct btrfs_trans_handle *trans;
3915         u64 delalloc_bytes;
3916         u64 max_reclaim;
3917         long time_left;
3918         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3919         int loops = 0;
3920         enum btrfs_reserve_flush_enum flush;
3921
3922         trans = (struct btrfs_trans_handle *)current->journal_info;
3923         block_rsv = &root->fs_info->delalloc_block_rsv;
3924         space_info = block_rsv->space_info;
3925
3926         smp_mb();
3927         delalloc_bytes = percpu_counter_sum_positive(
3928                                                 &root->fs_info->delalloc_bytes);
3929         if (delalloc_bytes == 0) {
3930                 if (trans)
3931                         return;
3932                 btrfs_wait_ordered_extents(root, 0);
3933                 return;
3934         }
3935
3936         while (delalloc_bytes && loops < 3) {
3937                 max_reclaim = min(delalloc_bytes, to_reclaim);
3938                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3939                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3940                 /*
3941                  * We need to wait for the async pages to actually start before
3942                  * we do anything.
3943                  */
3944                 wait_event(root->fs_info->async_submit_wait,
3945                            !atomic_read(&root->fs_info->async_delalloc_pages));
3946
3947                 if (!trans)
3948                         flush = BTRFS_RESERVE_FLUSH_ALL;
3949                 else
3950                         flush = BTRFS_RESERVE_NO_FLUSH;
3951                 spin_lock(&space_info->lock);
3952                 if (can_overcommit(root, space_info, orig, flush)) {
3953                         spin_unlock(&space_info->lock);
3954                         break;
3955                 }
3956                 spin_unlock(&space_info->lock);
3957
3958                 loops++;
3959                 if (wait_ordered && !trans) {
3960                         btrfs_wait_ordered_extents(root, 0);
3961                 } else {
3962                         time_left = schedule_timeout_killable(1);
3963                         if (time_left)
3964                                 break;
3965                 }
3966                 smp_mb();
3967                 delalloc_bytes = percpu_counter_sum_positive(
3968                                                 &root->fs_info->delalloc_bytes);
3969         }
3970 }
3971
3972 /**
3973  * maybe_commit_transaction - possibly commit the transaction if its ok to
3974  * @root - the root we're allocating for
3975  * @bytes - the number of bytes we want to reserve
3976  * @force - force the commit
3977  *
3978  * This will check to make sure that committing the transaction will actually
3979  * get us somewhere and then commit the transaction if it does.  Otherwise it
3980  * will return -ENOSPC.
3981  */
3982 static int may_commit_transaction(struct btrfs_root *root,
3983                                   struct btrfs_space_info *space_info,
3984                                   u64 bytes, int force)
3985 {
3986         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3987         struct btrfs_trans_handle *trans;
3988
3989         trans = (struct btrfs_trans_handle *)current->journal_info;
3990         if (trans)
3991                 return -EAGAIN;
3992
3993         if (force)
3994                 goto commit;
3995
3996         /* See if there is enough pinned space to make this reservation */
3997         spin_lock(&space_info->lock);
3998         if (space_info->bytes_pinned >= bytes) {
3999                 spin_unlock(&space_info->lock);
4000                 goto commit;
4001         }
4002         spin_unlock(&space_info->lock);
4003
4004         /*
4005          * See if there is some space in the delayed insertion reservation for
4006          * this reservation.
4007          */
4008         if (space_info != delayed_rsv->space_info)
4009                 return -ENOSPC;
4010
4011         spin_lock(&space_info->lock);
4012         spin_lock(&delayed_rsv->lock);
4013         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
4014                 spin_unlock(&delayed_rsv->lock);
4015                 spin_unlock(&space_info->lock);
4016                 return -ENOSPC;
4017         }
4018         spin_unlock(&delayed_rsv->lock);
4019         spin_unlock(&space_info->lock);
4020
4021 commit:
4022         trans = btrfs_join_transaction(root);
4023         if (IS_ERR(trans))
4024                 return -ENOSPC;
4025
4026         return btrfs_commit_transaction(trans, root);
4027 }
4028
4029 enum flush_state {
4030         FLUSH_DELAYED_ITEMS_NR  =       1,
4031         FLUSH_DELAYED_ITEMS     =       2,
4032         FLUSH_DELALLOC          =       3,
4033         FLUSH_DELALLOC_WAIT     =       4,
4034         ALLOC_CHUNK             =       5,
4035         COMMIT_TRANS            =       6,
4036 };
4037
4038 static int flush_space(struct btrfs_root *root,
4039                        struct btrfs_space_info *space_info, u64 num_bytes,
4040                        u64 orig_bytes, int state)
4041 {
4042         struct btrfs_trans_handle *trans;
4043         int nr;
4044         int ret = 0;
4045
4046         switch (state) {
4047         case FLUSH_DELAYED_ITEMS_NR:
4048         case FLUSH_DELAYED_ITEMS:
4049                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4050                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4051
4052                         nr = (int)div64_u64(num_bytes, bytes);
4053                         if (!nr)
4054                                 nr = 1;
4055                         nr *= 2;
4056                 } else {
4057                         nr = -1;
4058                 }
4059                 trans = btrfs_join_transaction(root);
4060                 if (IS_ERR(trans)) {
4061                         ret = PTR_ERR(trans);
4062                         break;
4063                 }
4064                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4065                 btrfs_end_transaction(trans, root);
4066                 break;
4067         case FLUSH_DELALLOC:
4068         case FLUSH_DELALLOC_WAIT:
4069                 shrink_delalloc(root, num_bytes, orig_bytes,
4070                                 state == FLUSH_DELALLOC_WAIT);
4071                 break;
4072         case ALLOC_CHUNK:
4073                 trans = btrfs_join_transaction(root);
4074                 if (IS_ERR(trans)) {
4075                         ret = PTR_ERR(trans);
4076                         break;
4077                 }
4078                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4079                                      btrfs_get_alloc_profile(root, 0),
4080                                      CHUNK_ALLOC_NO_FORCE);
4081                 btrfs_end_transaction(trans, root);
4082                 if (ret == -ENOSPC)
4083                         ret = 0;
4084                 break;
4085         case COMMIT_TRANS:
4086                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4087                 break;
4088         default:
4089                 ret = -ENOSPC;
4090                 break;
4091         }
4092
4093         return ret;
4094 }
4095 /**
4096  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4097  * @root - the root we're allocating for
4098  * @block_rsv - the block_rsv we're allocating for
4099  * @orig_bytes - the number of bytes we want
4100  * @flush - whether or not we can flush to make our reservation
4101  *
4102  * This will reserve orgi_bytes number of bytes from the space info associated
4103  * with the block_rsv.  If there is not enough space it will make an attempt to
4104  * flush out space to make room.  It will do this by flushing delalloc if
4105  * possible or committing the transaction.  If flush is 0 then no attempts to
4106  * regain reservations will be made and this will fail if there is not enough
4107  * space already.
4108  */
4109 static int reserve_metadata_bytes(struct btrfs_root *root,
4110                                   struct btrfs_block_rsv *block_rsv,
4111                                   u64 orig_bytes,
4112                                   enum btrfs_reserve_flush_enum flush)
4113 {
4114         struct btrfs_space_info *space_info = block_rsv->space_info;
4115         u64 used;
4116         u64 num_bytes = orig_bytes;
4117         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4118         int ret = 0;
4119         bool flushing = false;
4120
4121 again:
4122         ret = 0;
4123         spin_lock(&space_info->lock);
4124         /*
4125          * We only want to wait if somebody other than us is flushing and we
4126          * are actually allowed to flush all things.
4127          */
4128         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4129                space_info->flush) {
4130                 spin_unlock(&space_info->lock);
4131                 /*
4132                  * If we have a trans handle we can't wait because the flusher
4133                  * may have to commit the transaction, which would mean we would
4134                  * deadlock since we are waiting for the flusher to finish, but
4135                  * hold the current transaction open.
4136                  */
4137                 if (current->journal_info)
4138                         return -EAGAIN;
4139                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4140                 /* Must have been killed, return */
4141                 if (ret)
4142                         return -EINTR;
4143
4144                 spin_lock(&space_info->lock);
4145         }
4146
4147         ret = -ENOSPC;
4148         used = space_info->bytes_used + space_info->bytes_reserved +
4149                 space_info->bytes_pinned + space_info->bytes_readonly +
4150                 space_info->bytes_may_use;
4151
4152         /*
4153          * The idea here is that we've not already over-reserved the block group
4154          * then we can go ahead and save our reservation first and then start
4155          * flushing if we need to.  Otherwise if we've already overcommitted
4156          * lets start flushing stuff first and then come back and try to make
4157          * our reservation.
4158          */
4159         if (used <= space_info->total_bytes) {
4160                 if (used + orig_bytes <= space_info->total_bytes) {
4161                         space_info->bytes_may_use += orig_bytes;
4162                         trace_btrfs_space_reservation(root->fs_info,
4163                                 "space_info", space_info->flags, orig_bytes, 1);
4164                         ret = 0;
4165                 } else {
4166                         /*
4167                          * Ok set num_bytes to orig_bytes since we aren't
4168                          * overocmmitted, this way we only try and reclaim what
4169                          * we need.
4170                          */
4171                         num_bytes = orig_bytes;
4172                 }
4173         } else {
4174                 /*
4175                  * Ok we're over committed, set num_bytes to the overcommitted
4176                  * amount plus the amount of bytes that we need for this
4177                  * reservation.
4178                  */
4179                 num_bytes = used - space_info->total_bytes +
4180                         (orig_bytes * 2);
4181         }
4182
4183         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4184                 space_info->bytes_may_use += orig_bytes;
4185                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4186                                               space_info->flags, orig_bytes,
4187                                               1);
4188                 ret = 0;
4189         }
4190
4191         /*
4192          * Couldn't make our reservation, save our place so while we're trying
4193          * to reclaim space we can actually use it instead of somebody else
4194          * stealing it from us.
4195          *
4196          * We make the other tasks wait for the flush only when we can flush
4197          * all things.
4198          */
4199         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4200                 flushing = true;
4201                 space_info->flush = 1;
4202         }
4203
4204         spin_unlock(&space_info->lock);
4205
4206         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4207                 goto out;
4208
4209         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4210                           flush_state);
4211         flush_state++;
4212
4213         /*
4214          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4215          * would happen. So skip delalloc flush.
4216          */
4217         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4218             (flush_state == FLUSH_DELALLOC ||
4219              flush_state == FLUSH_DELALLOC_WAIT))
4220                 flush_state = ALLOC_CHUNK;
4221
4222         if (!ret)
4223                 goto again;
4224         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4225                  flush_state < COMMIT_TRANS)
4226                 goto again;
4227         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4228                  flush_state <= COMMIT_TRANS)
4229                 goto again;
4230
4231 out:
4232         if (ret == -ENOSPC &&
4233             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4234                 struct btrfs_block_rsv *global_rsv =
4235                         &root->fs_info->global_block_rsv;
4236
4237                 if (block_rsv != global_rsv &&
4238                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4239                         ret = 0;
4240         }
4241         if (flushing) {
4242                 spin_lock(&space_info->lock);
4243                 space_info->flush = 0;
4244                 wake_up_all(&space_info->wait);
4245                 spin_unlock(&space_info->lock);
4246         }
4247         return ret;
4248 }
4249
4250 static struct btrfs_block_rsv *get_block_rsv(
4251                                         const struct btrfs_trans_handle *trans,
4252                                         const struct btrfs_root *root)
4253 {
4254         struct btrfs_block_rsv *block_rsv = NULL;
4255
4256         if (root->ref_cows)
4257                 block_rsv = trans->block_rsv;
4258
4259         if (root == root->fs_info->csum_root && trans->adding_csums)
4260                 block_rsv = trans->block_rsv;
4261
4262         if (!block_rsv)
4263                 block_rsv = root->block_rsv;
4264
4265         if (!block_rsv)
4266                 block_rsv = &root->fs_info->empty_block_rsv;
4267
4268         return block_rsv;
4269 }
4270
4271 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4272                                u64 num_bytes)
4273 {
4274         int ret = -ENOSPC;
4275         spin_lock(&block_rsv->lock);
4276         if (block_rsv->reserved >= num_bytes) {
4277                 block_rsv->reserved -= num_bytes;
4278                 if (block_rsv->reserved < block_rsv->size)
4279                         block_rsv->full = 0;
4280                 ret = 0;
4281         }
4282         spin_unlock(&block_rsv->lock);
4283         return ret;
4284 }
4285
4286 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4287                                 u64 num_bytes, int update_size)
4288 {
4289         spin_lock(&block_rsv->lock);
4290         block_rsv->reserved += num_bytes;
4291         if (update_size)
4292                 block_rsv->size += num_bytes;
4293         else if (block_rsv->reserved >= block_rsv->size)
4294                 block_rsv->full = 1;
4295         spin_unlock(&block_rsv->lock);
4296 }
4297
4298 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4299                                     struct btrfs_block_rsv *block_rsv,
4300                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4301 {
4302         struct btrfs_space_info *space_info = block_rsv->space_info;
4303
4304         spin_lock(&block_rsv->lock);
4305         if (num_bytes == (u64)-1)
4306                 num_bytes = block_rsv->size;
4307         block_rsv->size -= num_bytes;
4308         if (block_rsv->reserved >= block_rsv->size) {
4309                 num_bytes = block_rsv->reserved - block_rsv->size;
4310                 block_rsv->reserved = block_rsv->size;
4311                 block_rsv->full = 1;
4312         } else {
4313                 num_bytes = 0;
4314         }
4315         spin_unlock(&block_rsv->lock);
4316
4317         if (num_bytes > 0) {
4318                 if (dest) {
4319                         spin_lock(&dest->lock);
4320                         if (!dest->full) {
4321                                 u64 bytes_to_add;
4322
4323                                 bytes_to_add = dest->size - dest->reserved;
4324                                 bytes_to_add = min(num_bytes, bytes_to_add);
4325                                 dest->reserved += bytes_to_add;
4326                                 if (dest->reserved >= dest->size)
4327                                         dest->full = 1;
4328                                 num_bytes -= bytes_to_add;
4329                         }
4330                         spin_unlock(&dest->lock);
4331                 }
4332                 if (num_bytes) {
4333                         spin_lock(&space_info->lock);
4334                         space_info->bytes_may_use -= num_bytes;
4335                         trace_btrfs_space_reservation(fs_info, "space_info",
4336                                         space_info->flags, num_bytes, 0);
4337                         space_info->reservation_progress++;
4338                         spin_unlock(&space_info->lock);
4339                 }
4340         }
4341 }
4342
4343 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4344                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4345 {
4346         int ret;
4347
4348         ret = block_rsv_use_bytes(src, num_bytes);
4349         if (ret)
4350                 return ret;
4351
4352         block_rsv_add_bytes(dst, num_bytes, 1);
4353         return 0;
4354 }
4355
4356 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4357 {
4358         memset(rsv, 0, sizeof(*rsv));
4359         spin_lock_init(&rsv->lock);
4360         rsv->type = type;
4361 }
4362
4363 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4364                                               unsigned short type)
4365 {
4366         struct btrfs_block_rsv *block_rsv;
4367         struct btrfs_fs_info *fs_info = root->fs_info;
4368
4369         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4370         if (!block_rsv)
4371                 return NULL;
4372
4373         btrfs_init_block_rsv(block_rsv, type);
4374         block_rsv->space_info = __find_space_info(fs_info,
4375                                                   BTRFS_BLOCK_GROUP_METADATA);
4376         return block_rsv;
4377 }
4378
4379 void btrfs_free_block_rsv(struct btrfs_root *root,
4380                           struct btrfs_block_rsv *rsv)
4381 {
4382         if (!rsv)
4383                 return;
4384         btrfs_block_rsv_release(root, rsv, (u64)-1);
4385         kfree(rsv);
4386 }
4387
4388 int btrfs_block_rsv_add(struct btrfs_root *root,
4389                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4390                         enum btrfs_reserve_flush_enum flush)
4391 {
4392         int ret;
4393
4394         if (num_bytes == 0)
4395                 return 0;
4396
4397         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4398         if (!ret) {
4399                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4400                 return 0;
4401         }
4402
4403         return ret;
4404 }
4405
4406 int btrfs_block_rsv_check(struct btrfs_root *root,
4407                           struct btrfs_block_rsv *block_rsv, int min_factor)
4408 {
4409         u64 num_bytes = 0;
4410         int ret = -ENOSPC;
4411
4412         if (!block_rsv)
4413                 return 0;
4414
4415         spin_lock(&block_rsv->lock);
4416         num_bytes = div_factor(block_rsv->size, min_factor);
4417         if (block_rsv->reserved >= num_bytes)
4418                 ret = 0;
4419         spin_unlock(&block_rsv->lock);
4420
4421         return ret;
4422 }
4423
4424 int btrfs_block_rsv_refill(struct btrfs_root *root,
4425                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4426                            enum btrfs_reserve_flush_enum flush)
4427 {
4428         u64 num_bytes = 0;
4429         int ret = -ENOSPC;
4430
4431         if (!block_rsv)
4432                 return 0;
4433
4434         spin_lock(&block_rsv->lock);
4435         num_bytes = min_reserved;
4436         if (block_rsv->reserved >= num_bytes)
4437                 ret = 0;
4438         else
4439                 num_bytes -= block_rsv->reserved;
4440         spin_unlock(&block_rsv->lock);
4441
4442         if (!ret)
4443                 return 0;
4444
4445         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4446         if (!ret) {
4447                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4448                 return 0;
4449         }
4450
4451         return ret;
4452 }
4453
4454 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4455                             struct btrfs_block_rsv *dst_rsv,
4456                             u64 num_bytes)
4457 {
4458         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4459 }
4460
4461 void btrfs_block_rsv_release(struct btrfs_root *root,
4462                              struct btrfs_block_rsv *block_rsv,
4463                              u64 num_bytes)
4464 {
4465         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4466         if (global_rsv->full || global_rsv == block_rsv ||
4467             block_rsv->space_info != global_rsv->space_info)
4468                 global_rsv = NULL;
4469         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4470                                 num_bytes);
4471 }
4472
4473 /*
4474  * helper to calculate size of global block reservation.
4475  * the desired value is sum of space used by extent tree,
4476  * checksum tree and root tree
4477  */
4478 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4479 {
4480         struct btrfs_space_info *sinfo;
4481         u64 num_bytes;
4482         u64 meta_used;
4483         u64 data_used;
4484         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4485
4486         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4487         spin_lock(&sinfo->lock);
4488         data_used = sinfo->bytes_used;
4489         spin_unlock(&sinfo->lock);
4490
4491         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4492         spin_lock(&sinfo->lock);
4493         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4494                 data_used = 0;
4495         meta_used = sinfo->bytes_used;
4496         spin_unlock(&sinfo->lock);
4497
4498         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4499                     csum_size * 2;
4500         num_bytes += div64_u64(data_used + meta_used, 50);
4501
4502         if (num_bytes * 3 > meta_used)
4503                 num_bytes = div64_u64(meta_used, 3);
4504
4505         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4506 }
4507
4508 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4509 {
4510         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4511         struct btrfs_space_info *sinfo = block_rsv->space_info;
4512         u64 num_bytes;
4513
4514         num_bytes = calc_global_metadata_size(fs_info);
4515
4516         spin_lock(&sinfo->lock);
4517         spin_lock(&block_rsv->lock);
4518
4519         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4520
4521         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4522                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4523                     sinfo->bytes_may_use;
4524
4525         if (sinfo->total_bytes > num_bytes) {
4526                 num_bytes = sinfo->total_bytes - num_bytes;
4527                 block_rsv->reserved += num_bytes;
4528                 sinfo->bytes_may_use += num_bytes;
4529                 trace_btrfs_space_reservation(fs_info, "space_info",
4530                                       sinfo->flags, num_bytes, 1);
4531         }
4532
4533         if (block_rsv->reserved >= block_rsv->size) {
4534                 num_bytes = block_rsv->reserved - block_rsv->size;
4535                 sinfo->bytes_may_use -= num_bytes;
4536                 trace_btrfs_space_reservation(fs_info, "space_info",
4537                                       sinfo->flags, num_bytes, 0);
4538                 sinfo->reservation_progress++;
4539                 block_rsv->reserved = block_rsv->size;
4540                 block_rsv->full = 1;
4541         }
4542
4543         spin_unlock(&block_rsv->lock);
4544         spin_unlock(&sinfo->lock);
4545 }
4546
4547 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4548 {
4549         struct btrfs_space_info *space_info;
4550
4551         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4552         fs_info->chunk_block_rsv.space_info = space_info;
4553
4554         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4555         fs_info->global_block_rsv.space_info = space_info;
4556         fs_info->delalloc_block_rsv.space_info = space_info;
4557         fs_info->trans_block_rsv.space_info = space_info;
4558         fs_info->empty_block_rsv.space_info = space_info;
4559         fs_info->delayed_block_rsv.space_info = space_info;
4560
4561         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4562         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4563         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4564         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4565         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4566
4567         update_global_block_rsv(fs_info);
4568 }
4569
4570 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4571 {
4572         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4573                                 (u64)-1);
4574         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4575         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4576         WARN_ON(fs_info->trans_block_rsv.size > 0);
4577         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4578         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4579         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4580         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4581         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4582 }
4583
4584 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4585                                   struct btrfs_root *root)
4586 {
4587         if (!trans->block_rsv)
4588                 return;
4589
4590         if (!trans->bytes_reserved)
4591                 return;
4592
4593         trace_btrfs_space_reservation(root->fs_info, "transaction",
4594                                       trans->transid, trans->bytes_reserved, 0);
4595         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4596         trans->bytes_reserved = 0;
4597 }
4598
4599 /* Can only return 0 or -ENOSPC */
4600 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4601                                   struct inode *inode)
4602 {
4603         struct btrfs_root *root = BTRFS_I(inode)->root;
4604         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4605         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4606
4607         /*
4608          * We need to hold space in order to delete our orphan item once we've
4609          * added it, so this takes the reservation so we can release it later
4610          * when we are truly done with the orphan item.
4611          */
4612         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4613         trace_btrfs_space_reservation(root->fs_info, "orphan",
4614                                       btrfs_ino(inode), num_bytes, 1);
4615         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4616 }
4617
4618 void btrfs_orphan_release_metadata(struct inode *inode)
4619 {
4620         struct btrfs_root *root = BTRFS_I(inode)->root;
4621         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4622         trace_btrfs_space_reservation(root->fs_info, "orphan",
4623                                       btrfs_ino(inode), num_bytes, 0);
4624         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4625 }
4626
4627 /*
4628  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4629  * root: the root of the parent directory
4630  * rsv: block reservation
4631  * items: the number of items that we need do reservation
4632  * qgroup_reserved: used to return the reserved size in qgroup
4633  *
4634  * This function is used to reserve the space for snapshot/subvolume
4635  * creation and deletion. Those operations are different with the
4636  * common file/directory operations, they change two fs/file trees
4637  * and root tree, the number of items that the qgroup reserves is
4638  * different with the free space reservation. So we can not use
4639  * the space reseravtion mechanism in start_transaction().
4640  */
4641 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4642                                      struct btrfs_block_rsv *rsv,
4643                                      int items,
4644                                      u64 *qgroup_reserved)
4645 {
4646         u64 num_bytes;
4647         int ret;
4648
4649         if (root->fs_info->quota_enabled) {
4650                 /* One for parent inode, two for dir entries */
4651                 num_bytes = 3 * root->leafsize;
4652                 ret = btrfs_qgroup_reserve(root, num_bytes);
4653                 if (ret)
4654                         return ret;
4655         } else {
4656                 num_bytes = 0;
4657         }
4658
4659         *qgroup_reserved = num_bytes;
4660
4661         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4662         rsv->space_info = __find_space_info(root->fs_info,
4663                                             BTRFS_BLOCK_GROUP_METADATA);
4664         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4665                                   BTRFS_RESERVE_FLUSH_ALL);
4666         if (ret) {
4667                 if (*qgroup_reserved)
4668                         btrfs_qgroup_free(root, *qgroup_reserved);
4669         }
4670
4671         return ret;
4672 }
4673
4674 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4675                                       struct btrfs_block_rsv *rsv,
4676                                       u64 qgroup_reserved)
4677 {
4678         btrfs_block_rsv_release(root, rsv, (u64)-1);
4679         if (qgroup_reserved)
4680                 btrfs_qgroup_free(root, qgroup_reserved);
4681 }
4682
4683 /**
4684  * drop_outstanding_extent - drop an outstanding extent
4685  * @inode: the inode we're dropping the extent for
4686  *
4687  * This is called when we are freeing up an outstanding extent, either called
4688  * after an error or after an extent is written.  This will return the number of
4689  * reserved extents that need to be freed.  This must be called with
4690  * BTRFS_I(inode)->lock held.
4691  */
4692 static unsigned drop_outstanding_extent(struct inode *inode)
4693 {
4694         unsigned drop_inode_space = 0;
4695         unsigned dropped_extents = 0;
4696
4697         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4698         BTRFS_I(inode)->outstanding_extents--;
4699
4700         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4701             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4702                                &BTRFS_I(inode)->runtime_flags))
4703                 drop_inode_space = 1;
4704
4705         /*
4706          * If we have more or the same amount of outsanding extents than we have
4707          * reserved then we need to leave the reserved extents count alone.
4708          */
4709         if (BTRFS_I(inode)->outstanding_extents >=
4710             BTRFS_I(inode)->reserved_extents)
4711                 return drop_inode_space;
4712
4713         dropped_extents = BTRFS_I(inode)->reserved_extents -
4714                 BTRFS_I(inode)->outstanding_extents;
4715         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4716         return dropped_extents + drop_inode_space;
4717 }
4718
4719 /**
4720  * calc_csum_metadata_size - return the amount of metada space that must be
4721  *      reserved/free'd for the given bytes.
4722  * @inode: the inode we're manipulating
4723  * @num_bytes: the number of bytes in question
4724  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4725  *
4726  * This adjusts the number of csum_bytes in the inode and then returns the
4727  * correct amount of metadata that must either be reserved or freed.  We
4728  * calculate how many checksums we can fit into one leaf and then divide the
4729  * number of bytes that will need to be checksumed by this value to figure out
4730  * how many checksums will be required.  If we are adding bytes then the number
4731  * may go up and we will return the number of additional bytes that must be
4732  * reserved.  If it is going down we will return the number of bytes that must
4733  * be freed.
4734  *
4735  * This must be called with BTRFS_I(inode)->lock held.
4736  */
4737 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4738                                    int reserve)
4739 {
4740         struct btrfs_root *root = BTRFS_I(inode)->root;
4741         u64 csum_size;
4742         int num_csums_per_leaf;
4743         int num_csums;
4744         int old_csums;
4745
4746         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4747             BTRFS_I(inode)->csum_bytes == 0)
4748                 return 0;
4749
4750         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4751         if (reserve)
4752                 BTRFS_I(inode)->csum_bytes += num_bytes;
4753         else
4754                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4755         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4756         num_csums_per_leaf = (int)div64_u64(csum_size,
4757                                             sizeof(struct btrfs_csum_item) +
4758                                             sizeof(struct btrfs_disk_key));
4759         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4760         num_csums = num_csums + num_csums_per_leaf - 1;
4761         num_csums = num_csums / num_csums_per_leaf;
4762
4763         old_csums = old_csums + num_csums_per_leaf - 1;
4764         old_csums = old_csums / num_csums_per_leaf;
4765
4766         /* No change, no need to reserve more */
4767         if (old_csums == num_csums)
4768                 return 0;
4769
4770         if (reserve)
4771                 return btrfs_calc_trans_metadata_size(root,
4772                                                       num_csums - old_csums);
4773
4774         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4775 }
4776
4777 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4778 {
4779         struct btrfs_root *root = BTRFS_I(inode)->root;
4780         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4781         u64 to_reserve = 0;
4782         u64 csum_bytes;
4783         unsigned nr_extents = 0;
4784         int extra_reserve = 0;
4785         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4786         int ret = 0;
4787         bool delalloc_lock = true;
4788         u64 to_free = 0;
4789         unsigned dropped;
4790
4791         /* If we are a free space inode we need to not flush since we will be in
4792          * the middle of a transaction commit.  We also don't need the delalloc
4793          * mutex since we won't race with anybody.  We need this mostly to make
4794          * lockdep shut its filthy mouth.
4795          */
4796         if (btrfs_is_free_space_inode(inode)) {
4797                 flush = BTRFS_RESERVE_NO_FLUSH;
4798                 delalloc_lock = false;
4799         }
4800
4801         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4802             btrfs_transaction_in_commit(root->fs_info))
4803                 schedule_timeout(1);
4804
4805         if (delalloc_lock)
4806                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4807
4808         num_bytes = ALIGN(num_bytes, root->sectorsize);
4809
4810         spin_lock(&BTRFS_I(inode)->lock);
4811         BTRFS_I(inode)->outstanding_extents++;
4812
4813         if (BTRFS_I(inode)->outstanding_extents >
4814             BTRFS_I(inode)->reserved_extents)
4815                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4816                         BTRFS_I(inode)->reserved_extents;
4817
4818         /*
4819          * Add an item to reserve for updating the inode when we complete the
4820          * delalloc io.
4821          */
4822         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4823                       &BTRFS_I(inode)->runtime_flags)) {
4824                 nr_extents++;
4825                 extra_reserve = 1;
4826         }
4827
4828         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4829         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4830         csum_bytes = BTRFS_I(inode)->csum_bytes;
4831         spin_unlock(&BTRFS_I(inode)->lock);
4832
4833         if (root->fs_info->quota_enabled) {
4834                 ret = btrfs_qgroup_reserve(root, num_bytes +
4835                                            nr_extents * root->leafsize);
4836                 if (ret)
4837                         goto out_fail;
4838         }
4839
4840         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4841         if (unlikely(ret)) {
4842                 if (root->fs_info->quota_enabled)
4843                         btrfs_qgroup_free(root, num_bytes +
4844                                                 nr_extents * root->leafsize);
4845                 goto out_fail;
4846         }
4847
4848         spin_lock(&BTRFS_I(inode)->lock);
4849         if (extra_reserve) {
4850                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4851                         &BTRFS_I(inode)->runtime_flags);
4852                 nr_extents--;
4853         }
4854         BTRFS_I(inode)->reserved_extents += nr_extents;
4855         spin_unlock(&BTRFS_I(inode)->lock);
4856
4857         if (delalloc_lock)
4858                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4859
4860         if (to_reserve)
4861                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4862                                               btrfs_ino(inode), to_reserve, 1);
4863         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4864
4865         return 0;
4866
4867 out_fail:
4868         spin_lock(&BTRFS_I(inode)->lock);
4869         dropped = drop_outstanding_extent(inode);
4870         /*
4871          * If the inodes csum_bytes is the same as the original
4872          * csum_bytes then we know we haven't raced with any free()ers
4873          * so we can just reduce our inodes csum bytes and carry on.
4874          */
4875         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4876                 calc_csum_metadata_size(inode, num_bytes, 0);
4877         } else {
4878                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4879                 u64 bytes;
4880
4881                 /*
4882                  * This is tricky, but first we need to figure out how much we
4883                  * free'd from any free-ers that occured during this
4884                  * reservation, so we reset ->csum_bytes to the csum_bytes
4885                  * before we dropped our lock, and then call the free for the
4886                  * number of bytes that were freed while we were trying our
4887                  * reservation.
4888                  */
4889                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4890                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4891                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4892
4893
4894                 /*
4895                  * Now we need to see how much we would have freed had we not
4896                  * been making this reservation and our ->csum_bytes were not
4897                  * artificially inflated.
4898                  */
4899                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4900                 bytes = csum_bytes - orig_csum_bytes;
4901                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4902
4903                 /*
4904                  * Now reset ->csum_bytes to what it should be.  If bytes is
4905                  * more than to_free then we would have free'd more space had we
4906                  * not had an artificially high ->csum_bytes, so we need to free
4907                  * the remainder.  If bytes is the same or less then we don't
4908                  * need to do anything, the other free-ers did the correct
4909                  * thing.
4910                  */
4911                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4912                 if (bytes > to_free)
4913                         to_free = bytes - to_free;
4914                 else
4915                         to_free = 0;
4916         }
4917         spin_unlock(&BTRFS_I(inode)->lock);
4918         if (dropped)
4919                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4920
4921         if (to_free) {
4922                 btrfs_block_rsv_release(root, block_rsv, to_free);
4923                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4924                                               btrfs_ino(inode), to_free, 0);
4925         }
4926         if (delalloc_lock)
4927                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4928         return ret;
4929 }
4930
4931 /**
4932  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4933  * @inode: the inode to release the reservation for
4934  * @num_bytes: the number of bytes we're releasing
4935  *
4936  * This will release the metadata reservation for an inode.  This can be called
4937  * once we complete IO for a given set of bytes to release their metadata
4938  * reservations.
4939  */
4940 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4941 {
4942         struct btrfs_root *root = BTRFS_I(inode)->root;
4943         u64 to_free = 0;
4944         unsigned dropped;
4945
4946         num_bytes = ALIGN(num_bytes, root->sectorsize);
4947         spin_lock(&BTRFS_I(inode)->lock);
4948         dropped = drop_outstanding_extent(inode);
4949
4950         if (num_bytes)
4951                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4952         spin_unlock(&BTRFS_I(inode)->lock);
4953         if (dropped > 0)
4954                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4955
4956         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4957                                       btrfs_ino(inode), to_free, 0);
4958         if (root->fs_info->quota_enabled) {
4959                 btrfs_qgroup_free(root, num_bytes +
4960                                         dropped * root->leafsize);
4961         }
4962
4963         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4964                                 to_free);
4965 }
4966
4967 /**
4968  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4969  * @inode: inode we're writing to
4970  * @num_bytes: the number of bytes we want to allocate
4971  *
4972  * This will do the following things
4973  *
4974  * o reserve space in the data space info for num_bytes
4975  * o reserve space in the metadata space info based on number of outstanding
4976  *   extents and how much csums will be needed
4977  * o add to the inodes ->delalloc_bytes
4978  * o add it to the fs_info's delalloc inodes list.
4979  *
4980  * This will return 0 for success and -ENOSPC if there is no space left.
4981  */
4982 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4983 {
4984         int ret;
4985
4986         ret = btrfs_check_data_free_space(inode, num_bytes);
4987         if (ret)
4988                 return ret;
4989
4990         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4991         if (ret) {
4992                 btrfs_free_reserved_data_space(inode, num_bytes);
4993                 return ret;
4994         }
4995
4996         return 0;
4997 }
4998
4999 /**
5000  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5001  * @inode: inode we're releasing space for
5002  * @num_bytes: the number of bytes we want to free up
5003  *
5004  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5005  * called in the case that we don't need the metadata AND data reservations
5006  * anymore.  So if there is an error or we insert an inline extent.
5007  *
5008  * This function will release the metadata space that was not used and will
5009  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5010  * list if there are no delalloc bytes left.
5011  */
5012 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5013 {
5014         btrfs_delalloc_release_metadata(inode, num_bytes);
5015         btrfs_free_reserved_data_space(inode, num_bytes);
5016 }
5017
5018 static int update_block_group(struct btrfs_root *root,
5019                               u64 bytenr, u64 num_bytes, int alloc)
5020 {
5021         struct btrfs_block_group_cache *cache = NULL;
5022         struct btrfs_fs_info *info = root->fs_info;
5023         u64 total = num_bytes;
5024         u64 old_val;
5025         u64 byte_in_group;
5026         int factor;
5027
5028         /* block accounting for super block */
5029         spin_lock(&info->delalloc_lock);
5030         old_val = btrfs_super_bytes_used(info->super_copy);
5031         if (alloc)
5032                 old_val += num_bytes;
5033         else
5034                 old_val -= num_bytes;
5035         btrfs_set_super_bytes_used(info->super_copy, old_val);
5036         spin_unlock(&info->delalloc_lock);
5037
5038         while (total) {
5039                 cache = btrfs_lookup_block_group(info, bytenr);
5040                 if (!cache)
5041                         return -ENOENT;
5042                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5043                                     BTRFS_BLOCK_GROUP_RAID1 |
5044                                     BTRFS_BLOCK_GROUP_RAID10))
5045                         factor = 2;
5046                 else
5047                         factor = 1;
5048                 /*
5049                  * If this block group has free space cache written out, we
5050                  * need to make sure to load it if we are removing space.  This
5051                  * is because we need the unpinning stage to actually add the
5052                  * space back to the block group, otherwise we will leak space.
5053                  */
5054                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5055                         cache_block_group(cache, 1);
5056
5057                 byte_in_group = bytenr - cache->key.objectid;
5058                 WARN_ON(byte_in_group > cache->key.offset);
5059
5060                 spin_lock(&cache->space_info->lock);
5061                 spin_lock(&cache->lock);
5062
5063                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5064                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5065                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5066
5067                 cache->dirty = 1;
5068                 old_val = btrfs_block_group_used(&cache->item);
5069                 num_bytes = min(total, cache->key.offset - byte_in_group);
5070                 if (alloc) {
5071                         old_val += num_bytes;
5072                         btrfs_set_block_group_used(&cache->item, old_val);
5073                         cache->reserved -= num_bytes;
5074                         cache->space_info->bytes_reserved -= num_bytes;
5075                         cache->space_info->bytes_used += num_bytes;
5076                         cache->space_info->disk_used += num_bytes * factor;
5077                         spin_unlock(&cache->lock);
5078                         spin_unlock(&cache->space_info->lock);
5079                 } else {
5080                         old_val -= num_bytes;
5081                         btrfs_set_block_group_used(&cache->item, old_val);
5082                         cache->pinned += num_bytes;
5083                         cache->space_info->bytes_pinned += num_bytes;
5084                         cache->space_info->bytes_used -= num_bytes;
5085                         cache->space_info->disk_used -= num_bytes * factor;
5086                         spin_unlock(&cache->lock);
5087                         spin_unlock(&cache->space_info->lock);
5088
5089                         set_extent_dirty(info->pinned_extents,
5090                                          bytenr, bytenr + num_bytes - 1,
5091                                          GFP_NOFS | __GFP_NOFAIL);
5092                 }
5093                 btrfs_put_block_group(cache);
5094                 total -= num_bytes;
5095                 bytenr += num_bytes;
5096         }
5097         return 0;
5098 }
5099
5100 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5101 {
5102         struct btrfs_block_group_cache *cache;
5103         u64 bytenr;
5104
5105         spin_lock(&root->fs_info->block_group_cache_lock);
5106         bytenr = root->fs_info->first_logical_byte;
5107         spin_unlock(&root->fs_info->block_group_cache_lock);
5108
5109         if (bytenr < (u64)-1)
5110                 return bytenr;
5111
5112         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5113         if (!cache)
5114                 return 0;
5115
5116         bytenr = cache->key.objectid;
5117         btrfs_put_block_group(cache);
5118
5119         return bytenr;
5120 }
5121
5122 static int pin_down_extent(struct btrfs_root *root,
5123                            struct btrfs_block_group_cache *cache,
5124                            u64 bytenr, u64 num_bytes, int reserved)
5125 {
5126         spin_lock(&cache->space_info->lock);
5127         spin_lock(&cache->lock);
5128         cache->pinned += num_bytes;
5129         cache->space_info->bytes_pinned += num_bytes;
5130         if (reserved) {
5131                 cache->reserved -= num_bytes;
5132                 cache->space_info->bytes_reserved -= num_bytes;
5133         }
5134         spin_unlock(&cache->lock);
5135         spin_unlock(&cache->space_info->lock);
5136
5137         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5138                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5139         return 0;
5140 }
5141
5142 /*
5143  * this function must be called within transaction
5144  */
5145 int btrfs_pin_extent(struct btrfs_root *root,
5146                      u64 bytenr, u64 num_bytes, int reserved)
5147 {
5148         struct btrfs_block_group_cache *cache;
5149
5150         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5151         BUG_ON(!cache); /* Logic error */
5152
5153         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5154
5155         btrfs_put_block_group(cache);
5156         return 0;
5157 }
5158
5159 /*
5160  * this function must be called within transaction
5161  */
5162 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5163                                     u64 bytenr, u64 num_bytes)
5164 {
5165         struct btrfs_block_group_cache *cache;
5166         int ret;
5167
5168         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5169         if (!cache)
5170                 return -EINVAL;
5171
5172         /*
5173          * pull in the free space cache (if any) so that our pin
5174          * removes the free space from the cache.  We have load_only set
5175          * to one because the slow code to read in the free extents does check
5176          * the pinned extents.
5177          */
5178         cache_block_group(cache, 1);
5179
5180         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5181
5182         /* remove us from the free space cache (if we're there at all) */
5183         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5184         btrfs_put_block_group(cache);
5185         return ret;
5186 }
5187
5188 /**
5189  * btrfs_update_reserved_bytes - update the block_group and space info counters
5190  * @cache:      The cache we are manipulating
5191  * @num_bytes:  The number of bytes in question
5192  * @reserve:    One of the reservation enums
5193  *
5194  * This is called by the allocator when it reserves space, or by somebody who is
5195  * freeing space that was never actually used on disk.  For example if you
5196  * reserve some space for a new leaf in transaction A and before transaction A
5197  * commits you free that leaf, you call this with reserve set to 0 in order to
5198  * clear the reservation.
5199  *
5200  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5201  * ENOSPC accounting.  For data we handle the reservation through clearing the
5202  * delalloc bits in the io_tree.  We have to do this since we could end up
5203  * allocating less disk space for the amount of data we have reserved in the
5204  * case of compression.
5205  *
5206  * If this is a reservation and the block group has become read only we cannot
5207  * make the reservation and return -EAGAIN, otherwise this function always
5208  * succeeds.
5209  */
5210 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5211                                        u64 num_bytes, int reserve)
5212 {
5213         struct btrfs_space_info *space_info = cache->space_info;
5214         int ret = 0;
5215
5216         spin_lock(&space_info->lock);
5217         spin_lock(&cache->lock);
5218         if (reserve != RESERVE_FREE) {
5219                 if (cache->ro) {
5220                         ret = -EAGAIN;
5221                 } else {
5222                         cache->reserved += num_bytes;
5223                         space_info->bytes_reserved += num_bytes;
5224                         if (reserve == RESERVE_ALLOC) {
5225                                 trace_btrfs_space_reservation(cache->fs_info,
5226                                                 "space_info", space_info->flags,
5227                                                 num_bytes, 0);
5228                                 space_info->bytes_may_use -= num_bytes;
5229                         }
5230                 }
5231         } else {
5232                 if (cache->ro)
5233                         space_info->bytes_readonly += num_bytes;
5234                 cache->reserved -= num_bytes;
5235                 space_info->bytes_reserved -= num_bytes;
5236                 space_info->reservation_progress++;
5237         }
5238         spin_unlock(&cache->lock);
5239         spin_unlock(&space_info->lock);
5240         return ret;
5241 }
5242
5243 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5244                                 struct btrfs_root *root)
5245 {
5246         struct btrfs_fs_info *fs_info = root->fs_info;
5247         struct btrfs_caching_control *next;
5248         struct btrfs_caching_control *caching_ctl;
5249         struct btrfs_block_group_cache *cache;
5250
5251         down_write(&fs_info->extent_commit_sem);
5252
5253         list_for_each_entry_safe(caching_ctl, next,
5254                                  &fs_info->caching_block_groups, list) {
5255                 cache = caching_ctl->block_group;
5256                 if (block_group_cache_done(cache)) {
5257                         cache->last_byte_to_unpin = (u64)-1;
5258                         list_del_init(&caching_ctl->list);
5259                         put_caching_control(caching_ctl);
5260                 } else {
5261                         cache->last_byte_to_unpin = caching_ctl->progress;
5262                 }
5263         }
5264
5265         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5266                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5267         else
5268                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5269
5270         up_write(&fs_info->extent_commit_sem);
5271
5272         update_global_block_rsv(fs_info);
5273 }
5274
5275 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5276 {
5277         struct btrfs_fs_info *fs_info = root->fs_info;
5278         struct btrfs_block_group_cache *cache = NULL;
5279         struct btrfs_space_info *space_info;
5280         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5281         u64 len;
5282         bool readonly;
5283
5284         while (start <= end) {
5285                 readonly = false;
5286                 if (!cache ||
5287                     start >= cache->key.objectid + cache->key.offset) {
5288                         if (cache)
5289                                 btrfs_put_block_group(cache);
5290                         cache = btrfs_lookup_block_group(fs_info, start);
5291                         BUG_ON(!cache); /* Logic error */
5292                 }
5293
5294                 len = cache->key.objectid + cache->key.offset - start;
5295                 len = min(len, end + 1 - start);
5296
5297                 if (start < cache->last_byte_to_unpin) {
5298                         len = min(len, cache->last_byte_to_unpin - start);
5299                         btrfs_add_free_space(cache, start, len);
5300                 }
5301
5302                 start += len;
5303                 space_info = cache->space_info;
5304
5305                 spin_lock(&space_info->lock);
5306                 spin_lock(&cache->lock);
5307                 cache->pinned -= len;
5308                 space_info->bytes_pinned -= len;
5309                 if (cache->ro) {
5310                         space_info->bytes_readonly += len;
5311                         readonly = true;
5312                 }
5313                 spin_unlock(&cache->lock);
5314                 if (!readonly && global_rsv->space_info == space_info) {
5315                         spin_lock(&global_rsv->lock);
5316                         if (!global_rsv->full) {
5317                                 len = min(len, global_rsv->size -
5318                                           global_rsv->reserved);
5319                                 global_rsv->reserved += len;
5320                                 space_info->bytes_may_use += len;
5321                                 if (global_rsv->reserved >= global_rsv->size)
5322                                         global_rsv->full = 1;
5323                         }
5324                         spin_unlock(&global_rsv->lock);
5325                 }
5326                 spin_unlock(&space_info->lock);
5327         }
5328
5329         if (cache)
5330                 btrfs_put_block_group(cache);
5331         return 0;
5332 }
5333
5334 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5335                                struct btrfs_root *root)
5336 {
5337         struct btrfs_fs_info *fs_info = root->fs_info;
5338         struct extent_io_tree *unpin;
5339         u64 start;
5340         u64 end;
5341         int ret;
5342
5343         if (trans->aborted)
5344                 return 0;
5345
5346         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5347                 unpin = &fs_info->freed_extents[1];
5348         else
5349                 unpin = &fs_info->freed_extents[0];
5350
5351         while (1) {
5352                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5353                                             EXTENT_DIRTY, NULL);
5354                 if (ret)
5355                         break;
5356
5357                 if (btrfs_test_opt(root, DISCARD))
5358                         ret = btrfs_discard_extent(root, start,
5359                                                    end + 1 - start, NULL);
5360
5361                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5362                 unpin_extent_range(root, start, end);
5363                 cond_resched();
5364         }
5365
5366         return 0;
5367 }
5368
5369 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5370                                 struct btrfs_root *root,
5371                                 u64 bytenr, u64 num_bytes, u64 parent,
5372                                 u64 root_objectid, u64 owner_objectid,
5373                                 u64 owner_offset, int refs_to_drop,
5374                                 struct btrfs_delayed_extent_op *extent_op)
5375 {
5376         struct btrfs_key key;
5377         struct btrfs_path *path;
5378         struct btrfs_fs_info *info = root->fs_info;
5379         struct btrfs_root *extent_root = info->extent_root;
5380         struct extent_buffer *leaf;
5381         struct btrfs_extent_item *ei;
5382         struct btrfs_extent_inline_ref *iref;
5383         int ret;
5384         int is_data;
5385         int extent_slot = 0;
5386         int found_extent = 0;
5387         int num_to_del = 1;
5388         u32 item_size;
5389         u64 refs;
5390         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5391                                                  SKINNY_METADATA);
5392
5393         path = btrfs_alloc_path();
5394         if (!path)
5395                 return -ENOMEM;
5396
5397         path->reada = 1;
5398         path->leave_spinning = 1;
5399
5400         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5401         BUG_ON(!is_data && refs_to_drop != 1);
5402
5403         if (is_data)
5404                 skinny_metadata = 0;
5405
5406         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5407                                     bytenr, num_bytes, parent,
5408                                     root_objectid, owner_objectid,
5409                                     owner_offset);
5410         if (ret == 0) {
5411                 extent_slot = path->slots[0];
5412                 while (extent_slot >= 0) {
5413                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5414                                               extent_slot);
5415                         if (key.objectid != bytenr)
5416                                 break;
5417                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5418                             key.offset == num_bytes) {
5419                                 found_extent = 1;
5420                                 break;
5421                         }
5422                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5423                             key.offset == owner_objectid) {
5424                                 found_extent = 1;
5425                                 break;
5426                         }
5427                         if (path->slots[0] - extent_slot > 5)
5428                                 break;
5429                         extent_slot--;
5430                 }
5431 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5432                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5433                 if (found_extent && item_size < sizeof(*ei))
5434                         found_extent = 0;
5435 #endif
5436                 if (!found_extent) {
5437                         BUG_ON(iref);
5438                         ret = remove_extent_backref(trans, extent_root, path,
5439                                                     NULL, refs_to_drop,
5440                                                     is_data);
5441                         if (ret) {
5442                                 btrfs_abort_transaction(trans, extent_root, ret);
5443                                 goto out;
5444                         }
5445                         btrfs_release_path(path);
5446                         path->leave_spinning = 1;
5447
5448                         key.objectid = bytenr;
5449                         key.type = BTRFS_EXTENT_ITEM_KEY;
5450                         key.offset = num_bytes;
5451
5452                         if (!is_data && skinny_metadata) {
5453                                 key.type = BTRFS_METADATA_ITEM_KEY;
5454                                 key.offset = owner_objectid;
5455                         }
5456
5457                         ret = btrfs_search_slot(trans, extent_root,
5458                                                 &key, path, -1, 1);
5459                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5460                                 /*
5461                                  * Couldn't find our skinny metadata item,
5462                                  * see if we have ye olde extent item.
5463                                  */
5464                                 path->slots[0]--;
5465                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5466                                                       path->slots[0]);
5467                                 if (key.objectid == bytenr &&
5468                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5469                                     key.offset == num_bytes)
5470                                         ret = 0;
5471                         }
5472
5473                         if (ret > 0 && skinny_metadata) {
5474                                 skinny_metadata = false;
5475                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5476                                 key.offset = num_bytes;
5477                                 btrfs_release_path(path);
5478                                 ret = btrfs_search_slot(trans, extent_root,
5479                                                         &key, path, -1, 1);
5480                         }
5481
5482                         if (ret) {
5483                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5484                                         ret, (unsigned long long)bytenr);
5485                                 if (ret > 0)
5486                                         btrfs_print_leaf(extent_root,
5487                                                          path->nodes[0]);
5488                         }
5489                         if (ret < 0) {
5490                                 btrfs_abort_transaction(trans, extent_root, ret);
5491                                 goto out;
5492                         }
5493                         extent_slot = path->slots[0];
5494                 }
5495         } else if (ret == -ENOENT) {
5496                 btrfs_print_leaf(extent_root, path->nodes[0]);
5497                 WARN_ON(1);
5498                 btrfs_err(info,
5499                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5500                         (unsigned long long)bytenr,
5501                         (unsigned long long)parent,
5502                         (unsigned long long)root_objectid,
5503                         (unsigned long long)owner_objectid,
5504                         (unsigned long long)owner_offset);
5505         } else {
5506                 btrfs_abort_transaction(trans, extent_root, ret);
5507                 goto out;
5508         }
5509
5510         leaf = path->nodes[0];
5511         item_size = btrfs_item_size_nr(leaf, extent_slot);
5512 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5513         if (item_size < sizeof(*ei)) {
5514                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5515                 ret = convert_extent_item_v0(trans, extent_root, path,
5516                                              owner_objectid, 0);
5517                 if (ret < 0) {
5518                         btrfs_abort_transaction(trans, extent_root, ret);
5519                         goto out;
5520                 }
5521
5522                 btrfs_release_path(path);
5523                 path->leave_spinning = 1;
5524
5525                 key.objectid = bytenr;
5526                 key.type = BTRFS_EXTENT_ITEM_KEY;
5527                 key.offset = num_bytes;
5528
5529                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5530                                         -1, 1);
5531                 if (ret) {
5532                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5533                                 ret, (unsigned long long)bytenr);
5534                         btrfs_print_leaf(extent_root, path->nodes[0]);
5535                 }
5536                 if (ret < 0) {
5537                         btrfs_abort_transaction(trans, extent_root, ret);
5538                         goto out;
5539                 }
5540
5541                 extent_slot = path->slots[0];
5542                 leaf = path->nodes[0];
5543                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5544         }
5545 #endif
5546         BUG_ON(item_size < sizeof(*ei));
5547         ei = btrfs_item_ptr(leaf, extent_slot,
5548                             struct btrfs_extent_item);
5549         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5550             key.type == BTRFS_EXTENT_ITEM_KEY) {
5551                 struct btrfs_tree_block_info *bi;
5552                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5553                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5554                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5555         }
5556
5557         refs = btrfs_extent_refs(leaf, ei);
5558         if (refs < refs_to_drop) {
5559                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5560                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5561                 ret = -EINVAL;
5562                 btrfs_abort_transaction(trans, extent_root, ret);
5563                 goto out;
5564         }
5565         refs -= refs_to_drop;
5566
5567         if (refs > 0) {
5568                 if (extent_op)
5569                         __run_delayed_extent_op(extent_op, leaf, ei);
5570                 /*
5571                  * In the case of inline back ref, reference count will
5572                  * be updated by remove_extent_backref
5573                  */
5574                 if (iref) {
5575                         BUG_ON(!found_extent);
5576                 } else {
5577                         btrfs_set_extent_refs(leaf, ei, refs);
5578                         btrfs_mark_buffer_dirty(leaf);
5579                 }
5580                 if (found_extent) {
5581                         ret = remove_extent_backref(trans, extent_root, path,
5582                                                     iref, refs_to_drop,
5583                                                     is_data);
5584                         if (ret) {
5585                                 btrfs_abort_transaction(trans, extent_root, ret);
5586                                 goto out;
5587                         }
5588                 }
5589         } else {
5590                 if (found_extent) {
5591                         BUG_ON(is_data && refs_to_drop !=
5592                                extent_data_ref_count(root, path, iref));
5593                         if (iref) {
5594                                 BUG_ON(path->slots[0] != extent_slot);
5595                         } else {
5596                                 BUG_ON(path->slots[0] != extent_slot + 1);
5597                                 path->slots[0] = extent_slot;
5598                                 num_to_del = 2;
5599                         }
5600                 }
5601
5602                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5603                                       num_to_del);
5604                 if (ret) {
5605                         btrfs_abort_transaction(trans, extent_root, ret);
5606                         goto out;
5607                 }
5608                 btrfs_release_path(path);
5609
5610                 if (is_data) {
5611                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5612                         if (ret) {
5613                                 btrfs_abort_transaction(trans, extent_root, ret);
5614                                 goto out;
5615                         }
5616                 }
5617
5618                 ret = update_block_group(root, bytenr, num_bytes, 0);
5619                 if (ret) {
5620                         btrfs_abort_transaction(trans, extent_root, ret);
5621                         goto out;
5622                 }
5623         }
5624 out:
5625         btrfs_free_path(path);
5626         return ret;
5627 }
5628
5629 /*
5630  * when we free an block, it is possible (and likely) that we free the last
5631  * delayed ref for that extent as well.  This searches the delayed ref tree for
5632  * a given extent, and if there are no other delayed refs to be processed, it
5633  * removes it from the tree.
5634  */
5635 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5636                                       struct btrfs_root *root, u64 bytenr)
5637 {
5638         struct btrfs_delayed_ref_head *head;
5639         struct btrfs_delayed_ref_root *delayed_refs;
5640         struct btrfs_delayed_ref_node *ref;
5641         struct rb_node *node;
5642         int ret = 0;
5643
5644         delayed_refs = &trans->transaction->delayed_refs;
5645         spin_lock(&delayed_refs->lock);
5646         head = btrfs_find_delayed_ref_head(trans, bytenr);
5647         if (!head)
5648                 goto out;
5649
5650         node = rb_prev(&head->node.rb_node);
5651         if (!node)
5652                 goto out;
5653
5654         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5655
5656         /* there are still entries for this ref, we can't drop it */
5657         if (ref->bytenr == bytenr)
5658                 goto out;
5659
5660         if (head->extent_op) {
5661                 if (!head->must_insert_reserved)
5662                         goto out;
5663                 btrfs_free_delayed_extent_op(head->extent_op);
5664                 head->extent_op = NULL;
5665         }
5666
5667         /*
5668          * waiting for the lock here would deadlock.  If someone else has it
5669          * locked they are already in the process of dropping it anyway
5670          */
5671         if (!mutex_trylock(&head->mutex))
5672                 goto out;
5673
5674         /*
5675          * at this point we have a head with no other entries.  Go
5676          * ahead and process it.
5677          */
5678         head->node.in_tree = 0;
5679         rb_erase(&head->node.rb_node, &delayed_refs->root);
5680
5681         delayed_refs->num_entries--;
5682
5683         /*
5684          * we don't take a ref on the node because we're removing it from the
5685          * tree, so we just steal the ref the tree was holding.
5686          */
5687         delayed_refs->num_heads--;
5688         if (list_empty(&head->cluster))
5689                 delayed_refs->num_heads_ready--;
5690
5691         list_del_init(&head->cluster);
5692         spin_unlock(&delayed_refs->lock);
5693
5694         BUG_ON(head->extent_op);
5695         if (head->must_insert_reserved)
5696                 ret = 1;
5697
5698         mutex_unlock(&head->mutex);
5699         btrfs_put_delayed_ref(&head->node);
5700         return ret;
5701 out:
5702         spin_unlock(&delayed_refs->lock);
5703         return 0;
5704 }
5705
5706 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5707                            struct btrfs_root *root,
5708                            struct extent_buffer *buf,
5709                            u64 parent, int last_ref)
5710 {
5711         struct btrfs_block_group_cache *cache = NULL;
5712         int ret;
5713
5714         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5715                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5716                                         buf->start, buf->len,
5717                                         parent, root->root_key.objectid,
5718                                         btrfs_header_level(buf),
5719                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5720                 BUG_ON(ret); /* -ENOMEM */
5721         }
5722
5723         if (!last_ref)
5724                 return;
5725
5726         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5727
5728         if (btrfs_header_generation(buf) == trans->transid) {
5729                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5730                         ret = check_ref_cleanup(trans, root, buf->start);
5731                         if (!ret)
5732                                 goto out;
5733                 }
5734
5735                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5736                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5737                         goto out;
5738                 }
5739
5740                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5741
5742                 btrfs_add_free_space(cache, buf->start, buf->len);
5743                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5744         }
5745 out:
5746         /*
5747          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5748          * anymore.
5749          */
5750         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5751         btrfs_put_block_group(cache);
5752 }
5753
5754 /* Can return -ENOMEM */
5755 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5756                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5757                       u64 owner, u64 offset, int for_cow)
5758 {
5759         int ret;
5760         struct btrfs_fs_info *fs_info = root->fs_info;
5761
5762         /*
5763          * tree log blocks never actually go into the extent allocation
5764          * tree, just update pinning info and exit early.
5765          */
5766         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5767                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5768                 /* unlocks the pinned mutex */
5769                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5770                 ret = 0;
5771         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5772                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5773                                         num_bytes,
5774                                         parent, root_objectid, (int)owner,
5775                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5776         } else {
5777                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5778                                                 num_bytes,
5779                                                 parent, root_objectid, owner,
5780                                                 offset, BTRFS_DROP_DELAYED_REF,
5781                                                 NULL, for_cow);
5782         }
5783         return ret;
5784 }
5785
5786 static u64 stripe_align(struct btrfs_root *root,
5787                         struct btrfs_block_group_cache *cache,
5788                         u64 val, u64 num_bytes)
5789 {
5790         u64 ret = ALIGN(val, root->stripesize);
5791         return ret;
5792 }
5793
5794 /*
5795  * when we wait for progress in the block group caching, its because
5796  * our allocation attempt failed at least once.  So, we must sleep
5797  * and let some progress happen before we try again.
5798  *
5799  * This function will sleep at least once waiting for new free space to
5800  * show up, and then it will check the block group free space numbers
5801  * for our min num_bytes.  Another option is to have it go ahead
5802  * and look in the rbtree for a free extent of a given size, but this
5803  * is a good start.
5804  */
5805 static noinline int
5806 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5807                                 u64 num_bytes)
5808 {
5809         struct btrfs_caching_control *caching_ctl;
5810
5811         caching_ctl = get_caching_control(cache);
5812         if (!caching_ctl)
5813                 return 0;
5814
5815         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5816                    (cache->free_space_ctl->free_space >= num_bytes));
5817
5818         put_caching_control(caching_ctl);
5819         return 0;
5820 }
5821
5822 static noinline int
5823 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5824 {
5825         struct btrfs_caching_control *caching_ctl;
5826
5827         caching_ctl = get_caching_control(cache);
5828         if (!caching_ctl)
5829                 return 0;
5830
5831         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5832
5833         put_caching_control(caching_ctl);
5834         return 0;
5835 }
5836
5837 int __get_raid_index(u64 flags)
5838 {
5839         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5840                 return BTRFS_RAID_RAID10;
5841         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5842                 return BTRFS_RAID_RAID1;
5843         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5844                 return BTRFS_RAID_DUP;
5845         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5846                 return BTRFS_RAID_RAID0;
5847         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5848                 return BTRFS_RAID_RAID5;
5849         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5850                 return BTRFS_RAID_RAID6;
5851
5852         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5853 }
5854
5855 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5856 {
5857         return __get_raid_index(cache->flags);
5858 }
5859
5860 enum btrfs_loop_type {
5861         LOOP_CACHING_NOWAIT = 0,
5862         LOOP_CACHING_WAIT = 1,
5863         LOOP_ALLOC_CHUNK = 2,
5864         LOOP_NO_EMPTY_SIZE = 3,
5865 };
5866
5867 /*
5868  * walks the btree of allocated extents and find a hole of a given size.
5869  * The key ins is changed to record the hole:
5870  * ins->objectid == block start
5871  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5872  * ins->offset == number of blocks
5873  * Any available blocks before search_start are skipped.
5874  */
5875 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5876                                      struct btrfs_root *orig_root,
5877                                      u64 num_bytes, u64 empty_size,
5878                                      u64 hint_byte, struct btrfs_key *ins,
5879                                      u64 flags)
5880 {
5881         int ret = 0;
5882         struct btrfs_root *root = orig_root->fs_info->extent_root;
5883         struct btrfs_free_cluster *last_ptr = NULL;
5884         struct btrfs_block_group_cache *block_group = NULL;
5885         struct btrfs_block_group_cache *used_block_group;
5886         u64 search_start = 0;
5887         int empty_cluster = 2 * 1024 * 1024;
5888         struct btrfs_space_info *space_info;
5889         int loop = 0;
5890         int index = __get_raid_index(flags);
5891         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
5892                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5893         bool found_uncached_bg = false;
5894         bool failed_cluster_refill = false;
5895         bool failed_alloc = false;
5896         bool use_cluster = true;
5897         bool have_caching_bg = false;
5898
5899         WARN_ON(num_bytes < root->sectorsize);
5900         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5901         ins->objectid = 0;
5902         ins->offset = 0;
5903
5904         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
5905
5906         space_info = __find_space_info(root->fs_info, flags);
5907         if (!space_info) {
5908                 btrfs_err(root->fs_info, "No space info for %llu", flags);
5909                 return -ENOSPC;
5910         }
5911
5912         /*
5913          * If the space info is for both data and metadata it means we have a
5914          * small filesystem and we can't use the clustering stuff.
5915          */
5916         if (btrfs_mixed_space_info(space_info))
5917                 use_cluster = false;
5918
5919         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5920                 last_ptr = &root->fs_info->meta_alloc_cluster;
5921                 if (!btrfs_test_opt(root, SSD))
5922                         empty_cluster = 64 * 1024;
5923         }
5924
5925         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5926             btrfs_test_opt(root, SSD)) {
5927                 last_ptr = &root->fs_info->data_alloc_cluster;
5928         }
5929
5930         if (last_ptr) {
5931                 spin_lock(&last_ptr->lock);
5932                 if (last_ptr->block_group)
5933                         hint_byte = last_ptr->window_start;
5934                 spin_unlock(&last_ptr->lock);
5935         }
5936
5937         search_start = max(search_start, first_logical_byte(root, 0));
5938         search_start = max(search_start, hint_byte);
5939
5940         if (!last_ptr)
5941                 empty_cluster = 0;
5942
5943         if (search_start == hint_byte) {
5944                 block_group = btrfs_lookup_block_group(root->fs_info,
5945                                                        search_start);
5946                 used_block_group = block_group;
5947                 /*
5948                  * we don't want to use the block group if it doesn't match our
5949                  * allocation bits, or if its not cached.
5950                  *
5951                  * However if we are re-searching with an ideal block group
5952                  * picked out then we don't care that the block group is cached.
5953                  */
5954                 if (block_group && block_group_bits(block_group, flags) &&
5955                     block_group->cached != BTRFS_CACHE_NO) {
5956                         down_read(&space_info->groups_sem);
5957                         if (list_empty(&block_group->list) ||
5958                             block_group->ro) {
5959                                 /*
5960                                  * someone is removing this block group,
5961                                  * we can't jump into the have_block_group
5962                                  * target because our list pointers are not
5963                                  * valid
5964                                  */
5965                                 btrfs_put_block_group(block_group);
5966                                 up_read(&space_info->groups_sem);
5967                         } else {
5968                                 index = get_block_group_index(block_group);
5969                                 goto have_block_group;
5970                         }
5971                 } else if (block_group) {
5972                         btrfs_put_block_group(block_group);
5973                 }
5974         }
5975 search:
5976         have_caching_bg = false;
5977         down_read(&space_info->groups_sem);
5978         list_for_each_entry(block_group, &space_info->block_groups[index],
5979                             list) {
5980                 u64 offset;
5981                 int cached;
5982
5983                 used_block_group = block_group;
5984                 btrfs_get_block_group(block_group);
5985                 search_start = block_group->key.objectid;
5986
5987                 /*
5988                  * this can happen if we end up cycling through all the
5989                  * raid types, but we want to make sure we only allocate
5990                  * for the proper type.
5991                  */
5992                 if (!block_group_bits(block_group, flags)) {
5993                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5994                                 BTRFS_BLOCK_GROUP_RAID1 |
5995                                 BTRFS_BLOCK_GROUP_RAID5 |
5996                                 BTRFS_BLOCK_GROUP_RAID6 |
5997                                 BTRFS_BLOCK_GROUP_RAID10;
5998
5999                         /*
6000                          * if they asked for extra copies and this block group
6001                          * doesn't provide them, bail.  This does allow us to
6002                          * fill raid0 from raid1.
6003                          */
6004                         if ((flags & extra) && !(block_group->flags & extra))
6005                                 goto loop;
6006                 }
6007
6008 have_block_group:
6009                 cached = block_group_cache_done(block_group);
6010                 if (unlikely(!cached)) {
6011                         found_uncached_bg = true;
6012                         ret = cache_block_group(block_group, 0);
6013                         BUG_ON(ret < 0);
6014                         ret = 0;
6015                 }
6016
6017                 if (unlikely(block_group->ro))
6018                         goto loop;
6019
6020                 /*
6021                  * Ok we want to try and use the cluster allocator, so
6022                  * lets look there
6023                  */
6024                 if (last_ptr) {
6025                         unsigned long aligned_cluster;
6026                         /*
6027                          * the refill lock keeps out other
6028                          * people trying to start a new cluster
6029                          */
6030                         spin_lock(&last_ptr->refill_lock);
6031                         used_block_group = last_ptr->block_group;
6032                         if (used_block_group != block_group &&
6033                             (!used_block_group ||
6034                              used_block_group->ro ||
6035                              !block_group_bits(used_block_group, flags))) {
6036                                 used_block_group = block_group;
6037                                 goto refill_cluster;
6038                         }
6039
6040                         if (used_block_group != block_group)
6041                                 btrfs_get_block_group(used_block_group);
6042
6043                         offset = btrfs_alloc_from_cluster(used_block_group,
6044                           last_ptr, num_bytes, used_block_group->key.objectid);
6045                         if (offset) {
6046                                 /* we have a block, we're done */
6047                                 spin_unlock(&last_ptr->refill_lock);
6048                                 trace_btrfs_reserve_extent_cluster(root,
6049                                         block_group, search_start, num_bytes);
6050                                 goto checks;
6051                         }
6052
6053                         WARN_ON(last_ptr->block_group != used_block_group);
6054                         if (used_block_group != block_group) {
6055                                 btrfs_put_block_group(used_block_group);
6056                                 used_block_group = block_group;
6057                         }
6058 refill_cluster:
6059                         BUG_ON(used_block_group != block_group);
6060                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6061                          * set up a new clusters, so lets just skip it
6062                          * and let the allocator find whatever block
6063                          * it can find.  If we reach this point, we
6064                          * will have tried the cluster allocator
6065                          * plenty of times and not have found
6066                          * anything, so we are likely way too
6067                          * fragmented for the clustering stuff to find
6068                          * anything.
6069                          *
6070                          * However, if the cluster is taken from the
6071                          * current block group, release the cluster
6072                          * first, so that we stand a better chance of
6073                          * succeeding in the unclustered
6074                          * allocation.  */
6075                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6076                             last_ptr->block_group != block_group) {
6077                                 spin_unlock(&last_ptr->refill_lock);
6078                                 goto unclustered_alloc;
6079                         }
6080
6081                         /*
6082                          * this cluster didn't work out, free it and
6083                          * start over
6084                          */
6085                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6086
6087                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6088                                 spin_unlock(&last_ptr->refill_lock);
6089                                 goto unclustered_alloc;
6090                         }
6091
6092                         aligned_cluster = max_t(unsigned long,
6093                                                 empty_cluster + empty_size,
6094                                               block_group->full_stripe_len);
6095
6096                         /* allocate a cluster in this block group */
6097                         ret = btrfs_find_space_cluster(trans, root,
6098                                                block_group, last_ptr,
6099                                                search_start, num_bytes,
6100                                                aligned_cluster);
6101                         if (ret == 0) {
6102                                 /*
6103                                  * now pull our allocation out of this
6104                                  * cluster
6105                                  */
6106                                 offset = btrfs_alloc_from_cluster(block_group,
6107                                                   last_ptr, num_bytes,
6108                                                   search_start);
6109                                 if (offset) {
6110                                         /* we found one, proceed */
6111                                         spin_unlock(&last_ptr->refill_lock);
6112                                         trace_btrfs_reserve_extent_cluster(root,
6113                                                 block_group, search_start,
6114                                                 num_bytes);
6115                                         goto checks;
6116                                 }
6117                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6118                                    && !failed_cluster_refill) {
6119                                 spin_unlock(&last_ptr->refill_lock);
6120
6121                                 failed_cluster_refill = true;
6122                                 wait_block_group_cache_progress(block_group,
6123                                        num_bytes + empty_cluster + empty_size);
6124                                 goto have_block_group;
6125                         }
6126
6127                         /*
6128                          * at this point we either didn't find a cluster
6129                          * or we weren't able to allocate a block from our
6130                          * cluster.  Free the cluster we've been trying
6131                          * to use, and go to the next block group
6132                          */
6133                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6134                         spin_unlock(&last_ptr->refill_lock);
6135                         goto loop;
6136                 }
6137
6138 unclustered_alloc:
6139                 spin_lock(&block_group->free_space_ctl->tree_lock);
6140                 if (cached &&
6141                     block_group->free_space_ctl->free_space <
6142                     num_bytes + empty_cluster + empty_size) {
6143                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6144                         goto loop;
6145                 }
6146                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6147
6148                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6149                                                     num_bytes, empty_size);
6150                 /*
6151                  * If we didn't find a chunk, and we haven't failed on this
6152                  * block group before, and this block group is in the middle of
6153                  * caching and we are ok with waiting, then go ahead and wait
6154                  * for progress to be made, and set failed_alloc to true.
6155                  *
6156                  * If failed_alloc is true then we've already waited on this
6157                  * block group once and should move on to the next block group.
6158                  */
6159                 if (!offset && !failed_alloc && !cached &&
6160                     loop > LOOP_CACHING_NOWAIT) {
6161                         wait_block_group_cache_progress(block_group,
6162                                                 num_bytes + empty_size);
6163                         failed_alloc = true;
6164                         goto have_block_group;
6165                 } else if (!offset) {
6166                         if (!cached)
6167                                 have_caching_bg = true;
6168                         goto loop;
6169                 }
6170 checks:
6171                 search_start = stripe_align(root, used_block_group,
6172                                             offset, num_bytes);
6173
6174                 /* move on to the next group */
6175                 if (search_start + num_bytes >
6176                     used_block_group->key.objectid + used_block_group->key.offset) {
6177                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6178                         goto loop;
6179                 }
6180
6181                 if (offset < search_start)
6182                         btrfs_add_free_space(used_block_group, offset,
6183                                              search_start - offset);
6184                 BUG_ON(offset > search_start);
6185
6186                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6187                                                   alloc_type);
6188                 if (ret == -EAGAIN) {
6189                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6190                         goto loop;
6191                 }
6192
6193                 /* we are all good, lets return */
6194                 ins->objectid = search_start;
6195                 ins->offset = num_bytes;
6196
6197                 trace_btrfs_reserve_extent(orig_root, block_group,
6198                                            search_start, num_bytes);
6199                 if (used_block_group != block_group)
6200                         btrfs_put_block_group(used_block_group);
6201                 btrfs_put_block_group(block_group);
6202                 break;
6203 loop:
6204                 failed_cluster_refill = false;
6205                 failed_alloc = false;
6206                 BUG_ON(index != get_block_group_index(block_group));
6207                 if (used_block_group != block_group)
6208                         btrfs_put_block_group(used_block_group);
6209                 btrfs_put_block_group(block_group);
6210         }
6211         up_read(&space_info->groups_sem);
6212
6213         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6214                 goto search;
6215
6216         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6217                 goto search;
6218
6219         /*
6220          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6221          *                      caching kthreads as we move along
6222          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6223          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6224          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6225          *                      again
6226          */
6227         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6228                 index = 0;
6229                 loop++;
6230                 if (loop == LOOP_ALLOC_CHUNK) {
6231                         ret = do_chunk_alloc(trans, root, flags,
6232                                              CHUNK_ALLOC_FORCE);
6233                         /*
6234                          * Do not bail out on ENOSPC since we
6235                          * can do more things.
6236                          */
6237                         if (ret < 0 && ret != -ENOSPC) {
6238                                 btrfs_abort_transaction(trans,
6239                                                         root, ret);
6240                                 goto out;
6241                         }
6242                 }
6243
6244                 if (loop == LOOP_NO_EMPTY_SIZE) {
6245                         empty_size = 0;
6246                         empty_cluster = 0;
6247                 }
6248
6249                 goto search;
6250         } else if (!ins->objectid) {
6251                 ret = -ENOSPC;
6252         } else if (ins->objectid) {
6253                 ret = 0;
6254         }
6255 out:
6256
6257         return ret;
6258 }
6259
6260 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6261                             int dump_block_groups)
6262 {
6263         struct btrfs_block_group_cache *cache;
6264         int index = 0;
6265
6266         spin_lock(&info->lock);
6267         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6268                (unsigned long long)info->flags,
6269                (unsigned long long)(info->total_bytes - info->bytes_used -
6270                                     info->bytes_pinned - info->bytes_reserved -
6271                                     info->bytes_readonly),
6272                (info->full) ? "" : "not ");
6273         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6274                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6275                (unsigned long long)info->total_bytes,
6276                (unsigned long long)info->bytes_used,
6277                (unsigned long long)info->bytes_pinned,
6278                (unsigned long long)info->bytes_reserved,
6279                (unsigned long long)info->bytes_may_use,
6280                (unsigned long long)info->bytes_readonly);
6281         spin_unlock(&info->lock);
6282
6283         if (!dump_block_groups)
6284                 return;
6285
6286         down_read(&info->groups_sem);
6287 again:
6288         list_for_each_entry(cache, &info->block_groups[index], list) {
6289                 spin_lock(&cache->lock);
6290                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6291                        (unsigned long long)cache->key.objectid,
6292                        (unsigned long long)cache->key.offset,
6293                        (unsigned long long)btrfs_block_group_used(&cache->item),
6294                        (unsigned long long)cache->pinned,
6295                        (unsigned long long)cache->reserved,
6296                        cache->ro ? "[readonly]" : "");
6297                 btrfs_dump_free_space(cache, bytes);
6298                 spin_unlock(&cache->lock);
6299         }
6300         if (++index < BTRFS_NR_RAID_TYPES)
6301                 goto again;
6302         up_read(&info->groups_sem);
6303 }
6304
6305 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6306                          struct btrfs_root *root,
6307                          u64 num_bytes, u64 min_alloc_size,
6308                          u64 empty_size, u64 hint_byte,
6309                          struct btrfs_key *ins, int is_data)
6310 {
6311         bool final_tried = false;
6312         u64 flags;
6313         int ret;
6314
6315         flags = btrfs_get_alloc_profile(root, is_data);
6316 again:
6317         WARN_ON(num_bytes < root->sectorsize);
6318         ret = find_free_extent(trans, root, num_bytes, empty_size,
6319                                hint_byte, ins, flags);
6320
6321         if (ret == -ENOSPC) {
6322                 if (!final_tried) {
6323                         num_bytes = num_bytes >> 1;
6324                         num_bytes = round_down(num_bytes, root->sectorsize);
6325                         num_bytes = max(num_bytes, min_alloc_size);
6326                         if (num_bytes == min_alloc_size)
6327                                 final_tried = true;
6328                         goto again;
6329                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6330                         struct btrfs_space_info *sinfo;
6331
6332                         sinfo = __find_space_info(root->fs_info, flags);
6333                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6334                                 (unsigned long long)flags,
6335                                 (unsigned long long)num_bytes);
6336                         if (sinfo)
6337                                 dump_space_info(sinfo, num_bytes, 1);
6338                 }
6339         }
6340
6341         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6342
6343         return ret;
6344 }
6345
6346 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6347                                         u64 start, u64 len, int pin)
6348 {
6349         struct btrfs_block_group_cache *cache;
6350         int ret = 0;
6351
6352         cache = btrfs_lookup_block_group(root->fs_info, start);
6353         if (!cache) {
6354                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6355                         (unsigned long long)start);
6356                 return -ENOSPC;
6357         }
6358
6359         if (btrfs_test_opt(root, DISCARD))
6360                 ret = btrfs_discard_extent(root, start, len, NULL);
6361
6362         if (pin)
6363                 pin_down_extent(root, cache, start, len, 1);
6364         else {
6365                 btrfs_add_free_space(cache, start, len);
6366                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6367         }
6368         btrfs_put_block_group(cache);
6369
6370         trace_btrfs_reserved_extent_free(root, start, len);
6371
6372         return ret;
6373 }
6374
6375 int btrfs_free_reserved_extent(struct btrfs_root *root,
6376                                         u64 start, u64 len)
6377 {
6378         return __btrfs_free_reserved_extent(root, start, len, 0);
6379 }
6380
6381 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6382                                        u64 start, u64 len)
6383 {
6384         return __btrfs_free_reserved_extent(root, start, len, 1);
6385 }
6386
6387 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6388                                       struct btrfs_root *root,
6389                                       u64 parent, u64 root_objectid,
6390                                       u64 flags, u64 owner, u64 offset,
6391                                       struct btrfs_key *ins, int ref_mod)
6392 {
6393         int ret;
6394         struct btrfs_fs_info *fs_info = root->fs_info;
6395         struct btrfs_extent_item *extent_item;
6396         struct btrfs_extent_inline_ref *iref;
6397         struct btrfs_path *path;
6398         struct extent_buffer *leaf;
6399         int type;
6400         u32 size;
6401
6402         if (parent > 0)
6403                 type = BTRFS_SHARED_DATA_REF_KEY;
6404         else
6405                 type = BTRFS_EXTENT_DATA_REF_KEY;
6406
6407         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6408
6409         path = btrfs_alloc_path();
6410         if (!path)
6411                 return -ENOMEM;
6412
6413         path->leave_spinning = 1;
6414         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6415                                       ins, size);
6416         if (ret) {
6417                 btrfs_free_path(path);
6418                 return ret;
6419         }
6420
6421         leaf = path->nodes[0];
6422         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6423                                      struct btrfs_extent_item);
6424         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6425         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6426         btrfs_set_extent_flags(leaf, extent_item,
6427                                flags | BTRFS_EXTENT_FLAG_DATA);
6428
6429         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6430         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6431         if (parent > 0) {
6432                 struct btrfs_shared_data_ref *ref;
6433                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6434                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6435                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6436         } else {
6437                 struct btrfs_extent_data_ref *ref;
6438                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6439                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6440                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6441                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6442                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6443         }
6444
6445         btrfs_mark_buffer_dirty(path->nodes[0]);
6446         btrfs_free_path(path);
6447
6448         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6449         if (ret) { /* -ENOENT, logic error */
6450                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6451                         (unsigned long long)ins->objectid,
6452                         (unsigned long long)ins->offset);
6453                 BUG();
6454         }
6455         return ret;
6456 }
6457
6458 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6459                                      struct btrfs_root *root,
6460                                      u64 parent, u64 root_objectid,
6461                                      u64 flags, struct btrfs_disk_key *key,
6462                                      int level, struct btrfs_key *ins)
6463 {
6464         int ret;
6465         struct btrfs_fs_info *fs_info = root->fs_info;
6466         struct btrfs_extent_item *extent_item;
6467         struct btrfs_tree_block_info *block_info;
6468         struct btrfs_extent_inline_ref *iref;
6469         struct btrfs_path *path;
6470         struct extent_buffer *leaf;
6471         u32 size = sizeof(*extent_item) + sizeof(*iref);
6472         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6473                                                  SKINNY_METADATA);
6474
6475         if (!skinny_metadata)
6476                 size += sizeof(*block_info);
6477
6478         path = btrfs_alloc_path();
6479         if (!path)
6480                 return -ENOMEM;
6481
6482         path->leave_spinning = 1;
6483         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6484                                       ins, size);
6485         if (ret) {
6486                 btrfs_free_path(path);
6487                 return ret;
6488         }
6489
6490         leaf = path->nodes[0];
6491         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6492                                      struct btrfs_extent_item);
6493         btrfs_set_extent_refs(leaf, extent_item, 1);
6494         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6495         btrfs_set_extent_flags(leaf, extent_item,
6496                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6497
6498         if (skinny_metadata) {
6499                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6500         } else {
6501                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6502                 btrfs_set_tree_block_key(leaf, block_info, key);
6503                 btrfs_set_tree_block_level(leaf, block_info, level);
6504                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6505         }
6506
6507         if (parent > 0) {
6508                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6509                 btrfs_set_extent_inline_ref_type(leaf, iref,
6510                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6511                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6512         } else {
6513                 btrfs_set_extent_inline_ref_type(leaf, iref,
6514                                                  BTRFS_TREE_BLOCK_REF_KEY);
6515                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6516         }
6517
6518         btrfs_mark_buffer_dirty(leaf);
6519         btrfs_free_path(path);
6520
6521         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6522         if (ret) { /* -ENOENT, logic error */
6523                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6524                         (unsigned long long)ins->objectid,
6525                         (unsigned long long)ins->offset);
6526                 BUG();
6527         }
6528         return ret;
6529 }
6530
6531 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6532                                      struct btrfs_root *root,
6533                                      u64 root_objectid, u64 owner,
6534                                      u64 offset, struct btrfs_key *ins)
6535 {
6536         int ret;
6537
6538         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6539
6540         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6541                                          ins->offset, 0,
6542                                          root_objectid, owner, offset,
6543                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6544         return ret;
6545 }
6546
6547 /*
6548  * this is used by the tree logging recovery code.  It records that
6549  * an extent has been allocated and makes sure to clear the free
6550  * space cache bits as well
6551  */
6552 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6553                                    struct btrfs_root *root,
6554                                    u64 root_objectid, u64 owner, u64 offset,
6555                                    struct btrfs_key *ins)
6556 {
6557         int ret;
6558         struct btrfs_block_group_cache *block_group;
6559         struct btrfs_caching_control *caching_ctl;
6560         u64 start = ins->objectid;
6561         u64 num_bytes = ins->offset;
6562
6563         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6564         cache_block_group(block_group, 0);
6565         caching_ctl = get_caching_control(block_group);
6566
6567         if (!caching_ctl) {
6568                 BUG_ON(!block_group_cache_done(block_group));
6569                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6570                 if (ret)
6571                         goto out;
6572         } else {
6573                 mutex_lock(&caching_ctl->mutex);
6574
6575                 if (start >= caching_ctl->progress) {
6576                         ret = add_excluded_extent(root, start, num_bytes);
6577                 } else if (start + num_bytes <= caching_ctl->progress) {
6578                         ret = btrfs_remove_free_space(block_group,
6579                                                       start, num_bytes);
6580                 } else {
6581                         num_bytes = caching_ctl->progress - start;
6582                         ret = btrfs_remove_free_space(block_group,
6583                                                       start, num_bytes);
6584                         if (ret)
6585                                 goto out_lock;
6586
6587                         start = caching_ctl->progress;
6588                         num_bytes = ins->objectid + ins->offset -
6589                                     caching_ctl->progress;
6590                         ret = add_excluded_extent(root, start, num_bytes);
6591                 }
6592 out_lock:
6593                 mutex_unlock(&caching_ctl->mutex);
6594                 put_caching_control(caching_ctl);
6595                 if (ret)
6596                         goto out;
6597         }
6598
6599         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6600                                           RESERVE_ALLOC_NO_ACCOUNT);
6601         BUG_ON(ret); /* logic error */
6602         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6603                                          0, owner, offset, ins, 1);
6604 out:
6605         btrfs_put_block_group(block_group);
6606         return ret;
6607 }
6608
6609 static struct extent_buffer *
6610 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6611                       u64 bytenr, u32 blocksize, int level)
6612 {
6613         struct extent_buffer *buf;
6614
6615         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6616         if (!buf)
6617                 return ERR_PTR(-ENOMEM);
6618         btrfs_set_header_generation(buf, trans->transid);
6619         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6620         btrfs_tree_lock(buf);
6621         clean_tree_block(trans, root, buf);
6622         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6623
6624         btrfs_set_lock_blocking(buf);
6625         btrfs_set_buffer_uptodate(buf);
6626
6627         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6628                 /*
6629                  * we allow two log transactions at a time, use different
6630                  * EXENT bit to differentiate dirty pages.
6631                  */
6632                 if (root->log_transid % 2 == 0)
6633                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6634                                         buf->start + buf->len - 1, GFP_NOFS);
6635                 else
6636                         set_extent_new(&root->dirty_log_pages, buf->start,
6637                                         buf->start + buf->len - 1, GFP_NOFS);
6638         } else {
6639                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6640                          buf->start + buf->len - 1, GFP_NOFS);
6641         }
6642         trans->blocks_used++;
6643         /* this returns a buffer locked for blocking */
6644         return buf;
6645 }
6646
6647 static struct btrfs_block_rsv *
6648 use_block_rsv(struct btrfs_trans_handle *trans,
6649               struct btrfs_root *root, u32 blocksize)
6650 {
6651         struct btrfs_block_rsv *block_rsv;
6652         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6653         int ret;
6654
6655         block_rsv = get_block_rsv(trans, root);
6656
6657         if (block_rsv->size == 0) {
6658                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6659                                              BTRFS_RESERVE_NO_FLUSH);
6660                 /*
6661                  * If we couldn't reserve metadata bytes try and use some from
6662                  * the global reserve.
6663                  */
6664                 if (ret && block_rsv != global_rsv) {
6665                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6666                         if (!ret)
6667                                 return global_rsv;
6668                         return ERR_PTR(ret);
6669                 } else if (ret) {
6670                         return ERR_PTR(ret);
6671                 }
6672                 return block_rsv;
6673         }
6674
6675         ret = block_rsv_use_bytes(block_rsv, blocksize);
6676         if (!ret)
6677                 return block_rsv;
6678         if (ret && !block_rsv->failfast) {
6679                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6680                         static DEFINE_RATELIMIT_STATE(_rs,
6681                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6682                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6683                         if (__ratelimit(&_rs))
6684                                 WARN(1, KERN_DEBUG
6685                                         "btrfs: block rsv returned %d\n", ret);
6686                 }
6687                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6688                                              BTRFS_RESERVE_NO_FLUSH);
6689                 if (!ret) {
6690                         return block_rsv;
6691                 } else if (ret && block_rsv != global_rsv) {
6692                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6693                         if (!ret)
6694                                 return global_rsv;
6695                 }
6696         }
6697
6698         return ERR_PTR(-ENOSPC);
6699 }
6700
6701 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6702                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6703 {
6704         block_rsv_add_bytes(block_rsv, blocksize, 0);
6705         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6706 }
6707
6708 /*
6709  * finds a free extent and does all the dirty work required for allocation
6710  * returns the key for the extent through ins, and a tree buffer for
6711  * the first block of the extent through buf.
6712  *
6713  * returns the tree buffer or NULL.
6714  */
6715 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6716                                         struct btrfs_root *root, u32 blocksize,
6717                                         u64 parent, u64 root_objectid,
6718                                         struct btrfs_disk_key *key, int level,
6719                                         u64 hint, u64 empty_size)
6720 {
6721         struct btrfs_key ins;
6722         struct btrfs_block_rsv *block_rsv;
6723         struct extent_buffer *buf;
6724         u64 flags = 0;
6725         int ret;
6726         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6727                                                  SKINNY_METADATA);
6728
6729         block_rsv = use_block_rsv(trans, root, blocksize);
6730         if (IS_ERR(block_rsv))
6731                 return ERR_CAST(block_rsv);
6732
6733         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6734                                    empty_size, hint, &ins, 0);
6735         if (ret) {
6736                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6737                 return ERR_PTR(ret);
6738         }
6739
6740         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6741                                     blocksize, level);
6742         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6743
6744         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6745                 if (parent == 0)
6746                         parent = ins.objectid;
6747                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6748         } else
6749                 BUG_ON(parent > 0);
6750
6751         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6752                 struct btrfs_delayed_extent_op *extent_op;
6753                 extent_op = btrfs_alloc_delayed_extent_op();
6754                 BUG_ON(!extent_op); /* -ENOMEM */
6755                 if (key)
6756                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6757                 else
6758                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6759                 extent_op->flags_to_set = flags;
6760                 if (skinny_metadata)
6761                         extent_op->update_key = 0;
6762                 else
6763                         extent_op->update_key = 1;
6764                 extent_op->update_flags = 1;
6765                 extent_op->is_data = 0;
6766
6767                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6768                                         ins.objectid,
6769                                         ins.offset, parent, root_objectid,
6770                                         level, BTRFS_ADD_DELAYED_EXTENT,
6771                                         extent_op, 0);
6772                 BUG_ON(ret); /* -ENOMEM */
6773         }
6774         return buf;
6775 }
6776
6777 struct walk_control {
6778         u64 refs[BTRFS_MAX_LEVEL];
6779         u64 flags[BTRFS_MAX_LEVEL];
6780         struct btrfs_key update_progress;
6781         int stage;
6782         int level;
6783         int shared_level;
6784         int update_ref;
6785         int keep_locks;
6786         int reada_slot;
6787         int reada_count;
6788         int for_reloc;
6789 };
6790
6791 #define DROP_REFERENCE  1
6792 #define UPDATE_BACKREF  2
6793
6794 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6795                                      struct btrfs_root *root,
6796                                      struct walk_control *wc,
6797                                      struct btrfs_path *path)
6798 {
6799         u64 bytenr;
6800         u64 generation;
6801         u64 refs;
6802         u64 flags;
6803         u32 nritems;
6804         u32 blocksize;
6805         struct btrfs_key key;
6806         struct extent_buffer *eb;
6807         int ret;
6808         int slot;
6809         int nread = 0;
6810
6811         if (path->slots[wc->level] < wc->reada_slot) {
6812                 wc->reada_count = wc->reada_count * 2 / 3;
6813                 wc->reada_count = max(wc->reada_count, 2);
6814         } else {
6815                 wc->reada_count = wc->reada_count * 3 / 2;
6816                 wc->reada_count = min_t(int, wc->reada_count,
6817                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6818         }
6819
6820         eb = path->nodes[wc->level];
6821         nritems = btrfs_header_nritems(eb);
6822         blocksize = btrfs_level_size(root, wc->level - 1);
6823
6824         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6825                 if (nread >= wc->reada_count)
6826                         break;
6827
6828                 cond_resched();
6829                 bytenr = btrfs_node_blockptr(eb, slot);
6830                 generation = btrfs_node_ptr_generation(eb, slot);
6831
6832                 if (slot == path->slots[wc->level])
6833                         goto reada;
6834
6835                 if (wc->stage == UPDATE_BACKREF &&
6836                     generation <= root->root_key.offset)
6837                         continue;
6838
6839                 /* We don't lock the tree block, it's OK to be racy here */
6840                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
6841                                                wc->level - 1, 1, &refs,
6842                                                &flags);
6843                 /* We don't care about errors in readahead. */
6844                 if (ret < 0)
6845                         continue;
6846                 BUG_ON(refs == 0);
6847
6848                 if (wc->stage == DROP_REFERENCE) {
6849                         if (refs == 1)
6850                                 goto reada;
6851
6852                         if (wc->level == 1 &&
6853                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6854                                 continue;
6855                         if (!wc->update_ref ||
6856                             generation <= root->root_key.offset)
6857                                 continue;
6858                         btrfs_node_key_to_cpu(eb, &key, slot);
6859                         ret = btrfs_comp_cpu_keys(&key,
6860                                                   &wc->update_progress);
6861                         if (ret < 0)
6862                                 continue;
6863                 } else {
6864                         if (wc->level == 1 &&
6865                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6866                                 continue;
6867                 }
6868 reada:
6869                 ret = readahead_tree_block(root, bytenr, blocksize,
6870                                            generation);
6871                 if (ret)
6872                         break;
6873                 nread++;
6874         }
6875         wc->reada_slot = slot;
6876 }
6877
6878 /*
6879  * helper to process tree block while walking down the tree.
6880  *
6881  * when wc->stage == UPDATE_BACKREF, this function updates
6882  * back refs for pointers in the block.
6883  *
6884  * NOTE: return value 1 means we should stop walking down.
6885  */
6886 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6887                                    struct btrfs_root *root,
6888                                    struct btrfs_path *path,
6889                                    struct walk_control *wc, int lookup_info)
6890 {
6891         int level = wc->level;
6892         struct extent_buffer *eb = path->nodes[level];
6893         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6894         int ret;
6895
6896         if (wc->stage == UPDATE_BACKREF &&
6897             btrfs_header_owner(eb) != root->root_key.objectid)
6898                 return 1;
6899
6900         /*
6901          * when reference count of tree block is 1, it won't increase
6902          * again. once full backref flag is set, we never clear it.
6903          */
6904         if (lookup_info &&
6905             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6906              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6907                 BUG_ON(!path->locks[level]);
6908                 ret = btrfs_lookup_extent_info(trans, root,
6909                                                eb->start, level, 1,
6910                                                &wc->refs[level],
6911                                                &wc->flags[level]);
6912                 BUG_ON(ret == -ENOMEM);
6913                 if (ret)
6914                         return ret;
6915                 BUG_ON(wc->refs[level] == 0);
6916         }
6917
6918         if (wc->stage == DROP_REFERENCE) {
6919                 if (wc->refs[level] > 1)
6920                         return 1;
6921
6922                 if (path->locks[level] && !wc->keep_locks) {
6923                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6924                         path->locks[level] = 0;
6925                 }
6926                 return 0;
6927         }
6928
6929         /* wc->stage == UPDATE_BACKREF */
6930         if (!(wc->flags[level] & flag)) {
6931                 BUG_ON(!path->locks[level]);
6932                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6933                 BUG_ON(ret); /* -ENOMEM */
6934                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6935                 BUG_ON(ret); /* -ENOMEM */
6936                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6937                                                   eb->len, flag, 0);
6938                 BUG_ON(ret); /* -ENOMEM */
6939                 wc->flags[level] |= flag;
6940         }
6941
6942         /*
6943          * the block is shared by multiple trees, so it's not good to
6944          * keep the tree lock
6945          */
6946         if (path->locks[level] && level > 0) {
6947                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6948                 path->locks[level] = 0;
6949         }
6950         return 0;
6951 }
6952
6953 /*
6954  * helper to process tree block pointer.
6955  *
6956  * when wc->stage == DROP_REFERENCE, this function checks
6957  * reference count of the block pointed to. if the block
6958  * is shared and we need update back refs for the subtree
6959  * rooted at the block, this function changes wc->stage to
6960  * UPDATE_BACKREF. if the block is shared and there is no
6961  * need to update back, this function drops the reference
6962  * to the block.
6963  *
6964  * NOTE: return value 1 means we should stop walking down.
6965  */
6966 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6967                                  struct btrfs_root *root,
6968                                  struct btrfs_path *path,
6969                                  struct walk_control *wc, int *lookup_info)
6970 {
6971         u64 bytenr;
6972         u64 generation;
6973         u64 parent;
6974         u32 blocksize;
6975         struct btrfs_key key;
6976         struct extent_buffer *next;
6977         int level = wc->level;
6978         int reada = 0;
6979         int ret = 0;
6980
6981         generation = btrfs_node_ptr_generation(path->nodes[level],
6982                                                path->slots[level]);
6983         /*
6984          * if the lower level block was created before the snapshot
6985          * was created, we know there is no need to update back refs
6986          * for the subtree
6987          */
6988         if (wc->stage == UPDATE_BACKREF &&
6989             generation <= root->root_key.offset) {
6990                 *lookup_info = 1;
6991                 return 1;
6992         }
6993
6994         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6995         blocksize = btrfs_level_size(root, level - 1);
6996
6997         next = btrfs_find_tree_block(root, bytenr, blocksize);
6998         if (!next) {
6999                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7000                 if (!next)
7001                         return -ENOMEM;
7002                 reada = 1;
7003         }
7004         btrfs_tree_lock(next);
7005         btrfs_set_lock_blocking(next);
7006
7007         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7008                                        &wc->refs[level - 1],
7009                                        &wc->flags[level - 1]);
7010         if (ret < 0) {
7011                 btrfs_tree_unlock(next);
7012                 return ret;
7013         }
7014
7015         if (unlikely(wc->refs[level - 1] == 0)) {
7016                 btrfs_err(root->fs_info, "Missing references.");
7017                 BUG();
7018         }
7019         *lookup_info = 0;
7020
7021         if (wc->stage == DROP_REFERENCE) {
7022                 if (wc->refs[level - 1] > 1) {
7023                         if (level == 1 &&
7024                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7025                                 goto skip;
7026
7027                         if (!wc->update_ref ||
7028                             generation <= root->root_key.offset)
7029                                 goto skip;
7030
7031                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7032                                               path->slots[level]);
7033                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7034                         if (ret < 0)
7035                                 goto skip;
7036
7037                         wc->stage = UPDATE_BACKREF;
7038                         wc->shared_level = level - 1;
7039                 }
7040         } else {
7041                 if (level == 1 &&
7042                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7043                         goto skip;
7044         }
7045
7046         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7047                 btrfs_tree_unlock(next);
7048                 free_extent_buffer(next);
7049                 next = NULL;
7050                 *lookup_info = 1;
7051         }
7052
7053         if (!next) {
7054                 if (reada && level == 1)
7055                         reada_walk_down(trans, root, wc, path);
7056                 next = read_tree_block(root, bytenr, blocksize, generation);
7057                 if (!next || !extent_buffer_uptodate(next)) {
7058                         free_extent_buffer(next);
7059                         return -EIO;
7060                 }
7061                 btrfs_tree_lock(next);
7062                 btrfs_set_lock_blocking(next);
7063         }
7064
7065         level--;
7066         BUG_ON(level != btrfs_header_level(next));
7067         path->nodes[level] = next;
7068         path->slots[level] = 0;
7069         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7070         wc->level = level;
7071         if (wc->level == 1)
7072                 wc->reada_slot = 0;
7073         return 0;
7074 skip:
7075         wc->refs[level - 1] = 0;
7076         wc->flags[level - 1] = 0;
7077         if (wc->stage == DROP_REFERENCE) {
7078                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7079                         parent = path->nodes[level]->start;
7080                 } else {
7081                         BUG_ON(root->root_key.objectid !=
7082                                btrfs_header_owner(path->nodes[level]));
7083                         parent = 0;
7084                 }
7085
7086                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7087                                 root->root_key.objectid, level - 1, 0, 0);
7088                 BUG_ON(ret); /* -ENOMEM */
7089         }
7090         btrfs_tree_unlock(next);
7091         free_extent_buffer(next);
7092         *lookup_info = 1;
7093         return 1;
7094 }
7095
7096 /*
7097  * helper to process tree block while walking up the tree.
7098  *
7099  * when wc->stage == DROP_REFERENCE, this function drops
7100  * reference count on the block.
7101  *
7102  * when wc->stage == UPDATE_BACKREF, this function changes
7103  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7104  * to UPDATE_BACKREF previously while processing the block.
7105  *
7106  * NOTE: return value 1 means we should stop walking up.
7107  */
7108 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7109                                  struct btrfs_root *root,
7110                                  struct btrfs_path *path,
7111                                  struct walk_control *wc)
7112 {
7113         int ret;
7114         int level = wc->level;
7115         struct extent_buffer *eb = path->nodes[level];
7116         u64 parent = 0;
7117
7118         if (wc->stage == UPDATE_BACKREF) {
7119                 BUG_ON(wc->shared_level < level);
7120                 if (level < wc->shared_level)
7121                         goto out;
7122
7123                 ret = find_next_key(path, level + 1, &wc->update_progress);
7124                 if (ret > 0)
7125                         wc->update_ref = 0;
7126
7127                 wc->stage = DROP_REFERENCE;
7128                 wc->shared_level = -1;
7129                 path->slots[level] = 0;
7130
7131                 /*
7132                  * check reference count again if the block isn't locked.
7133                  * we should start walking down the tree again if reference
7134                  * count is one.
7135                  */
7136                 if (!path->locks[level]) {
7137                         BUG_ON(level == 0);
7138                         btrfs_tree_lock(eb);
7139                         btrfs_set_lock_blocking(eb);
7140                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7141
7142                         ret = btrfs_lookup_extent_info(trans, root,
7143                                                        eb->start, level, 1,
7144                                                        &wc->refs[level],
7145                                                        &wc->flags[level]);
7146                         if (ret < 0) {
7147                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7148                                 path->locks[level] = 0;
7149                                 return ret;
7150                         }
7151                         BUG_ON(wc->refs[level] == 0);
7152                         if (wc->refs[level] == 1) {
7153                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7154                                 path->locks[level] = 0;
7155                                 return 1;
7156                         }
7157                 }
7158         }
7159
7160         /* wc->stage == DROP_REFERENCE */
7161         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7162
7163         if (wc->refs[level] == 1) {
7164                 if (level == 0) {
7165                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7166                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7167                                                     wc->for_reloc);
7168                         else
7169                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7170                                                     wc->for_reloc);
7171                         BUG_ON(ret); /* -ENOMEM */
7172                 }
7173                 /* make block locked assertion in clean_tree_block happy */
7174                 if (!path->locks[level] &&
7175                     btrfs_header_generation(eb) == trans->transid) {
7176                         btrfs_tree_lock(eb);
7177                         btrfs_set_lock_blocking(eb);
7178                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7179                 }
7180                 clean_tree_block(trans, root, eb);
7181         }
7182
7183         if (eb == root->node) {
7184                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7185                         parent = eb->start;
7186                 else
7187                         BUG_ON(root->root_key.objectid !=
7188                                btrfs_header_owner(eb));
7189         } else {
7190                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7191                         parent = path->nodes[level + 1]->start;
7192                 else
7193                         BUG_ON(root->root_key.objectid !=
7194                                btrfs_header_owner(path->nodes[level + 1]));
7195         }
7196
7197         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7198 out:
7199         wc->refs[level] = 0;
7200         wc->flags[level] = 0;
7201         return 0;
7202 }
7203
7204 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7205                                    struct btrfs_root *root,
7206                                    struct btrfs_path *path,
7207                                    struct walk_control *wc)
7208 {
7209         int level = wc->level;
7210         int lookup_info = 1;
7211         int ret;
7212
7213         while (level >= 0) {
7214                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7215                 if (ret > 0)
7216                         break;
7217
7218                 if (level == 0)
7219                         break;
7220
7221                 if (path->slots[level] >=
7222                     btrfs_header_nritems(path->nodes[level]))
7223                         break;
7224
7225                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7226                 if (ret > 0) {
7227                         path->slots[level]++;
7228                         continue;
7229                 } else if (ret < 0)
7230                         return ret;
7231                 level = wc->level;
7232         }
7233         return 0;
7234 }
7235
7236 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7237                                  struct btrfs_root *root,
7238                                  struct btrfs_path *path,
7239                                  struct walk_control *wc, int max_level)
7240 {
7241         int level = wc->level;
7242         int ret;
7243
7244         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7245         while (level < max_level && path->nodes[level]) {
7246                 wc->level = level;
7247                 if (path->slots[level] + 1 <
7248                     btrfs_header_nritems(path->nodes[level])) {
7249                         path->slots[level]++;
7250                         return 0;
7251                 } else {
7252                         ret = walk_up_proc(trans, root, path, wc);
7253                         if (ret > 0)
7254                                 return 0;
7255
7256                         if (path->locks[level]) {
7257                                 btrfs_tree_unlock_rw(path->nodes[level],
7258                                                      path->locks[level]);
7259                                 path->locks[level] = 0;
7260                         }
7261                         free_extent_buffer(path->nodes[level]);
7262                         path->nodes[level] = NULL;
7263                         level++;
7264                 }
7265         }
7266         return 1;
7267 }
7268
7269 /*
7270  * drop a subvolume tree.
7271  *
7272  * this function traverses the tree freeing any blocks that only
7273  * referenced by the tree.
7274  *
7275  * when a shared tree block is found. this function decreases its
7276  * reference count by one. if update_ref is true, this function
7277  * also make sure backrefs for the shared block and all lower level
7278  * blocks are properly updated.
7279  *
7280  * If called with for_reloc == 0, may exit early with -EAGAIN
7281  */
7282 int btrfs_drop_snapshot(struct btrfs_root *root,
7283                          struct btrfs_block_rsv *block_rsv, int update_ref,
7284                          int for_reloc)
7285 {
7286         struct btrfs_path *path;
7287         struct btrfs_trans_handle *trans;
7288         struct btrfs_root *tree_root = root->fs_info->tree_root;
7289         struct btrfs_root_item *root_item = &root->root_item;
7290         struct walk_control *wc;
7291         struct btrfs_key key;
7292         int err = 0;
7293         int ret;
7294         int level;
7295
7296         path = btrfs_alloc_path();
7297         if (!path) {
7298                 err = -ENOMEM;
7299                 goto out;
7300         }
7301
7302         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7303         if (!wc) {
7304                 btrfs_free_path(path);
7305                 err = -ENOMEM;
7306                 goto out;
7307         }
7308
7309         trans = btrfs_start_transaction(tree_root, 0);
7310         if (IS_ERR(trans)) {
7311                 err = PTR_ERR(trans);
7312                 goto out_free;
7313         }
7314
7315         if (block_rsv)
7316                 trans->block_rsv = block_rsv;
7317
7318         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7319                 level = btrfs_header_level(root->node);
7320                 path->nodes[level] = btrfs_lock_root_node(root);
7321                 btrfs_set_lock_blocking(path->nodes[level]);
7322                 path->slots[level] = 0;
7323                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7324                 memset(&wc->update_progress, 0,
7325                        sizeof(wc->update_progress));
7326         } else {
7327                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7328                 memcpy(&wc->update_progress, &key,
7329                        sizeof(wc->update_progress));
7330
7331                 level = root_item->drop_level;
7332                 BUG_ON(level == 0);
7333                 path->lowest_level = level;
7334                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7335                 path->lowest_level = 0;
7336                 if (ret < 0) {
7337                         err = ret;
7338                         goto out_end_trans;
7339                 }
7340                 WARN_ON(ret > 0);
7341
7342                 /*
7343                  * unlock our path, this is safe because only this
7344                  * function is allowed to delete this snapshot
7345                  */
7346                 btrfs_unlock_up_safe(path, 0);
7347
7348                 level = btrfs_header_level(root->node);
7349                 while (1) {
7350                         btrfs_tree_lock(path->nodes[level]);
7351                         btrfs_set_lock_blocking(path->nodes[level]);
7352
7353                         ret = btrfs_lookup_extent_info(trans, root,
7354                                                 path->nodes[level]->start,
7355                                                 level, 1, &wc->refs[level],
7356                                                 &wc->flags[level]);
7357                         if (ret < 0) {
7358                                 err = ret;
7359                                 goto out_end_trans;
7360                         }
7361                         BUG_ON(wc->refs[level] == 0);
7362
7363                         if (level == root_item->drop_level)
7364                                 break;
7365
7366                         btrfs_tree_unlock(path->nodes[level]);
7367                         WARN_ON(wc->refs[level] != 1);
7368                         level--;
7369                 }
7370         }
7371
7372         wc->level = level;
7373         wc->shared_level = -1;
7374         wc->stage = DROP_REFERENCE;
7375         wc->update_ref = update_ref;
7376         wc->keep_locks = 0;
7377         wc->for_reloc = for_reloc;
7378         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7379
7380         while (1) {
7381                 if (!for_reloc && btrfs_fs_closing(root->fs_info)) {
7382                         pr_debug("btrfs: drop snapshot early exit\n");
7383                         err = -EAGAIN;
7384                         goto out_end_trans;
7385                 }
7386
7387                 ret = walk_down_tree(trans, root, path, wc);
7388                 if (ret < 0) {
7389                         err = ret;
7390                         break;
7391                 }
7392
7393                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7394                 if (ret < 0) {
7395                         err = ret;
7396                         break;
7397                 }
7398
7399                 if (ret > 0) {
7400                         BUG_ON(wc->stage != DROP_REFERENCE);
7401                         break;
7402                 }
7403
7404                 if (wc->stage == DROP_REFERENCE) {
7405                         level = wc->level;
7406                         btrfs_node_key(path->nodes[level],
7407                                        &root_item->drop_progress,
7408                                        path->slots[level]);
7409                         root_item->drop_level = level;
7410                 }
7411
7412                 BUG_ON(wc->level == 0);
7413                 if (btrfs_should_end_transaction(trans, tree_root)) {
7414                         ret = btrfs_update_root(trans, tree_root,
7415                                                 &root->root_key,
7416                                                 root_item);
7417                         if (ret) {
7418                                 btrfs_abort_transaction(trans, tree_root, ret);
7419                                 err = ret;
7420                                 goto out_end_trans;
7421                         }
7422
7423                         btrfs_end_transaction_throttle(trans, tree_root);
7424                         trans = btrfs_start_transaction(tree_root, 0);
7425                         if (IS_ERR(trans)) {
7426                                 err = PTR_ERR(trans);
7427                                 goto out_free;
7428                         }
7429                         if (block_rsv)
7430                                 trans->block_rsv = block_rsv;
7431                 }
7432         }
7433         btrfs_release_path(path);
7434         if (err)
7435                 goto out_end_trans;
7436
7437         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7438         if (ret) {
7439                 btrfs_abort_transaction(trans, tree_root, ret);
7440                 goto out_end_trans;
7441         }
7442
7443         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7444                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7445                                            NULL, NULL);
7446                 if (ret < 0) {
7447                         btrfs_abort_transaction(trans, tree_root, ret);
7448                         err = ret;
7449                         goto out_end_trans;
7450                 } else if (ret > 0) {
7451                         /* if we fail to delete the orphan item this time
7452                          * around, it'll get picked up the next time.
7453                          *
7454                          * The most common failure here is just -ENOENT.
7455                          */
7456                         btrfs_del_orphan_item(trans, tree_root,
7457                                               root->root_key.objectid);
7458                 }
7459         }
7460
7461         if (root->in_radix) {
7462                 btrfs_free_fs_root(tree_root->fs_info, root);
7463         } else {
7464                 free_extent_buffer(root->node);
7465                 free_extent_buffer(root->commit_root);
7466                 kfree(root);
7467         }
7468 out_end_trans:
7469         btrfs_end_transaction_throttle(trans, tree_root);
7470 out_free:
7471         kfree(wc);
7472         btrfs_free_path(path);
7473 out:
7474         if (err)
7475                 btrfs_std_error(root->fs_info, err);
7476         return err;
7477 }
7478
7479 /*
7480  * drop subtree rooted at tree block 'node'.
7481  *
7482  * NOTE: this function will unlock and release tree block 'node'
7483  * only used by relocation code
7484  */
7485 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7486                         struct btrfs_root *root,
7487                         struct extent_buffer *node,
7488                         struct extent_buffer *parent)
7489 {
7490         struct btrfs_path *path;
7491         struct walk_control *wc;
7492         int level;
7493         int parent_level;
7494         int ret = 0;
7495         int wret;
7496
7497         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7498
7499         path = btrfs_alloc_path();
7500         if (!path)
7501                 return -ENOMEM;
7502
7503         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7504         if (!wc) {
7505                 btrfs_free_path(path);
7506                 return -ENOMEM;
7507         }
7508
7509         btrfs_assert_tree_locked(parent);
7510         parent_level = btrfs_header_level(parent);
7511         extent_buffer_get(parent);
7512         path->nodes[parent_level] = parent;
7513         path->slots[parent_level] = btrfs_header_nritems(parent);
7514
7515         btrfs_assert_tree_locked(node);
7516         level = btrfs_header_level(node);
7517         path->nodes[level] = node;
7518         path->slots[level] = 0;
7519         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7520
7521         wc->refs[parent_level] = 1;
7522         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7523         wc->level = level;
7524         wc->shared_level = -1;
7525         wc->stage = DROP_REFERENCE;
7526         wc->update_ref = 0;
7527         wc->keep_locks = 1;
7528         wc->for_reloc = 1;
7529         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7530
7531         while (1) {
7532                 wret = walk_down_tree(trans, root, path, wc);
7533                 if (wret < 0) {
7534                         ret = wret;
7535                         break;
7536                 }
7537
7538                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7539                 if (wret < 0)
7540                         ret = wret;
7541                 if (wret != 0)
7542                         break;
7543         }
7544
7545         kfree(wc);
7546         btrfs_free_path(path);
7547         return ret;
7548 }
7549
7550 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7551 {
7552         u64 num_devices;
7553         u64 stripped;
7554
7555         /*
7556          * if restripe for this chunk_type is on pick target profile and
7557          * return, otherwise do the usual balance
7558          */
7559         stripped = get_restripe_target(root->fs_info, flags);
7560         if (stripped)
7561                 return extended_to_chunk(stripped);
7562
7563         /*
7564          * we add in the count of missing devices because we want
7565          * to make sure that any RAID levels on a degraded FS
7566          * continue to be honored.
7567          */
7568         num_devices = root->fs_info->fs_devices->rw_devices +
7569                 root->fs_info->fs_devices->missing_devices;
7570
7571         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7572                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7573                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7574
7575         if (num_devices == 1) {
7576                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7577                 stripped = flags & ~stripped;
7578
7579                 /* turn raid0 into single device chunks */
7580                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7581                         return stripped;
7582
7583                 /* turn mirroring into duplication */
7584                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7585                              BTRFS_BLOCK_GROUP_RAID10))
7586                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7587         } else {
7588                 /* they already had raid on here, just return */
7589                 if (flags & stripped)
7590                         return flags;
7591
7592                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7593                 stripped = flags & ~stripped;
7594
7595                 /* switch duplicated blocks with raid1 */
7596                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7597                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7598
7599                 /* this is drive concat, leave it alone */
7600         }
7601
7602         return flags;
7603 }
7604
7605 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7606 {
7607         struct btrfs_space_info *sinfo = cache->space_info;
7608         u64 num_bytes;
7609         u64 min_allocable_bytes;
7610         int ret = -ENOSPC;
7611
7612
7613         /*
7614          * We need some metadata space and system metadata space for
7615          * allocating chunks in some corner cases until we force to set
7616          * it to be readonly.
7617          */
7618         if ((sinfo->flags &
7619              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7620             !force)
7621                 min_allocable_bytes = 1 * 1024 * 1024;
7622         else
7623                 min_allocable_bytes = 0;
7624
7625         spin_lock(&sinfo->lock);
7626         spin_lock(&cache->lock);
7627
7628         if (cache->ro) {
7629                 ret = 0;
7630                 goto out;
7631         }
7632
7633         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7634                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7635
7636         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7637             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7638             min_allocable_bytes <= sinfo->total_bytes) {
7639                 sinfo->bytes_readonly += num_bytes;
7640                 cache->ro = 1;
7641                 ret = 0;
7642         }
7643 out:
7644         spin_unlock(&cache->lock);
7645         spin_unlock(&sinfo->lock);
7646         return ret;
7647 }
7648
7649 int btrfs_set_block_group_ro(struct btrfs_root *root,
7650                              struct btrfs_block_group_cache *cache)
7651
7652 {
7653         struct btrfs_trans_handle *trans;
7654         u64 alloc_flags;
7655         int ret;
7656
7657         BUG_ON(cache->ro);
7658
7659         trans = btrfs_join_transaction(root);
7660         if (IS_ERR(trans))
7661                 return PTR_ERR(trans);
7662
7663         alloc_flags = update_block_group_flags(root, cache->flags);
7664         if (alloc_flags != cache->flags) {
7665                 ret = do_chunk_alloc(trans, root, alloc_flags,
7666                                      CHUNK_ALLOC_FORCE);
7667                 if (ret < 0)
7668                         goto out;
7669         }
7670
7671         ret = set_block_group_ro(cache, 0);
7672         if (!ret)
7673                 goto out;
7674         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7675         ret = do_chunk_alloc(trans, root, alloc_flags,
7676                              CHUNK_ALLOC_FORCE);
7677         if (ret < 0)
7678                 goto out;
7679         ret = set_block_group_ro(cache, 0);
7680 out:
7681         btrfs_end_transaction(trans, root);
7682         return ret;
7683 }
7684
7685 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7686                             struct btrfs_root *root, u64 type)
7687 {
7688         u64 alloc_flags = get_alloc_profile(root, type);
7689         return do_chunk_alloc(trans, root, alloc_flags,
7690                               CHUNK_ALLOC_FORCE);
7691 }
7692
7693 /*
7694  * helper to account the unused space of all the readonly block group in the
7695  * list. takes mirrors into account.
7696  */
7697 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7698 {
7699         struct btrfs_block_group_cache *block_group;
7700         u64 free_bytes = 0;
7701         int factor;
7702
7703         list_for_each_entry(block_group, groups_list, list) {
7704                 spin_lock(&block_group->lock);
7705
7706                 if (!block_group->ro) {
7707                         spin_unlock(&block_group->lock);
7708                         continue;
7709                 }
7710
7711                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7712                                           BTRFS_BLOCK_GROUP_RAID10 |
7713                                           BTRFS_BLOCK_GROUP_DUP))
7714                         factor = 2;
7715                 else
7716                         factor = 1;
7717
7718                 free_bytes += (block_group->key.offset -
7719                                btrfs_block_group_used(&block_group->item)) *
7720                                factor;
7721
7722                 spin_unlock(&block_group->lock);
7723         }
7724
7725         return free_bytes;
7726 }
7727
7728 /*
7729  * helper to account the unused space of all the readonly block group in the
7730  * space_info. takes mirrors into account.
7731  */
7732 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7733 {
7734         int i;
7735         u64 free_bytes = 0;
7736
7737         spin_lock(&sinfo->lock);
7738
7739         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7740                 if (!list_empty(&sinfo->block_groups[i]))
7741                         free_bytes += __btrfs_get_ro_block_group_free_space(
7742                                                 &sinfo->block_groups[i]);
7743
7744         spin_unlock(&sinfo->lock);
7745
7746         return free_bytes;
7747 }
7748
7749 void btrfs_set_block_group_rw(struct btrfs_root *root,
7750                               struct btrfs_block_group_cache *cache)
7751 {
7752         struct btrfs_space_info *sinfo = cache->space_info;
7753         u64 num_bytes;
7754
7755         BUG_ON(!cache->ro);
7756
7757         spin_lock(&sinfo->lock);
7758         spin_lock(&cache->lock);
7759         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7760                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7761         sinfo->bytes_readonly -= num_bytes;
7762         cache->ro = 0;
7763         spin_unlock(&cache->lock);
7764         spin_unlock(&sinfo->lock);
7765 }
7766
7767 /*
7768  * checks to see if its even possible to relocate this block group.
7769  *
7770  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7771  * ok to go ahead and try.
7772  */
7773 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7774 {
7775         struct btrfs_block_group_cache *block_group;
7776         struct btrfs_space_info *space_info;
7777         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7778         struct btrfs_device *device;
7779         u64 min_free;
7780         u64 dev_min = 1;
7781         u64 dev_nr = 0;
7782         u64 target;
7783         int index;
7784         int full = 0;
7785         int ret = 0;
7786
7787         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7788
7789         /* odd, couldn't find the block group, leave it alone */
7790         if (!block_group)
7791                 return -1;
7792
7793         min_free = btrfs_block_group_used(&block_group->item);
7794
7795         /* no bytes used, we're good */
7796         if (!min_free)
7797                 goto out;
7798
7799         space_info = block_group->space_info;
7800         spin_lock(&space_info->lock);
7801
7802         full = space_info->full;
7803
7804         /*
7805          * if this is the last block group we have in this space, we can't
7806          * relocate it unless we're able to allocate a new chunk below.
7807          *
7808          * Otherwise, we need to make sure we have room in the space to handle
7809          * all of the extents from this block group.  If we can, we're good
7810          */
7811         if ((space_info->total_bytes != block_group->key.offset) &&
7812             (space_info->bytes_used + space_info->bytes_reserved +
7813              space_info->bytes_pinned + space_info->bytes_readonly +
7814              min_free < space_info->total_bytes)) {
7815                 spin_unlock(&space_info->lock);
7816                 goto out;
7817         }
7818         spin_unlock(&space_info->lock);
7819
7820         /*
7821          * ok we don't have enough space, but maybe we have free space on our
7822          * devices to allocate new chunks for relocation, so loop through our
7823          * alloc devices and guess if we have enough space.  if this block
7824          * group is going to be restriped, run checks against the target
7825          * profile instead of the current one.
7826          */
7827         ret = -1;
7828
7829         /*
7830          * index:
7831          *      0: raid10
7832          *      1: raid1
7833          *      2: dup
7834          *      3: raid0
7835          *      4: single
7836          */
7837         target = get_restripe_target(root->fs_info, block_group->flags);
7838         if (target) {
7839                 index = __get_raid_index(extended_to_chunk(target));
7840         } else {
7841                 /*
7842                  * this is just a balance, so if we were marked as full
7843                  * we know there is no space for a new chunk
7844                  */
7845                 if (full)
7846                         goto out;
7847
7848                 index = get_block_group_index(block_group);
7849         }
7850
7851         if (index == BTRFS_RAID_RAID10) {
7852                 dev_min = 4;
7853                 /* Divide by 2 */
7854                 min_free >>= 1;
7855         } else if (index == BTRFS_RAID_RAID1) {
7856                 dev_min = 2;
7857         } else if (index == BTRFS_RAID_DUP) {
7858                 /* Multiply by 2 */
7859                 min_free <<= 1;
7860         } else if (index == BTRFS_RAID_RAID0) {
7861                 dev_min = fs_devices->rw_devices;
7862                 do_div(min_free, dev_min);
7863         }
7864
7865         mutex_lock(&root->fs_info->chunk_mutex);
7866         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7867                 u64 dev_offset;
7868
7869                 /*
7870                  * check to make sure we can actually find a chunk with enough
7871                  * space to fit our block group in.
7872                  */
7873                 if (device->total_bytes > device->bytes_used + min_free &&
7874                     !device->is_tgtdev_for_dev_replace) {
7875                         ret = find_free_dev_extent(device, min_free,
7876                                                    &dev_offset, NULL);
7877                         if (!ret)
7878                                 dev_nr++;
7879
7880                         if (dev_nr >= dev_min)
7881                                 break;
7882
7883                         ret = -1;
7884                 }
7885         }
7886         mutex_unlock(&root->fs_info->chunk_mutex);
7887 out:
7888         btrfs_put_block_group(block_group);
7889         return ret;
7890 }
7891
7892 static int find_first_block_group(struct btrfs_root *root,
7893                 struct btrfs_path *path, struct btrfs_key *key)
7894 {
7895         int ret = 0;
7896         struct btrfs_key found_key;
7897         struct extent_buffer *leaf;
7898         int slot;
7899
7900         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7901         if (ret < 0)
7902                 goto out;
7903
7904         while (1) {
7905                 slot = path->slots[0];
7906                 leaf = path->nodes[0];
7907                 if (slot >= btrfs_header_nritems(leaf)) {
7908                         ret = btrfs_next_leaf(root, path);
7909                         if (ret == 0)
7910                                 continue;
7911                         if (ret < 0)
7912                                 goto out;
7913                         break;
7914                 }
7915                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7916
7917                 if (found_key.objectid >= key->objectid &&
7918                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7919                         ret = 0;
7920                         goto out;
7921                 }
7922                 path->slots[0]++;
7923         }
7924 out:
7925         return ret;
7926 }
7927
7928 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7929 {
7930         struct btrfs_block_group_cache *block_group;
7931         u64 last = 0;
7932
7933         while (1) {
7934                 struct inode *inode;
7935
7936                 block_group = btrfs_lookup_first_block_group(info, last);
7937                 while (block_group) {
7938                         spin_lock(&block_group->lock);
7939                         if (block_group->iref)
7940                                 break;
7941                         spin_unlock(&block_group->lock);
7942                         block_group = next_block_group(info->tree_root,
7943                                                        block_group);
7944                 }
7945                 if (!block_group) {
7946                         if (last == 0)
7947                                 break;
7948                         last = 0;
7949                         continue;
7950                 }
7951
7952                 inode = block_group->inode;
7953                 block_group->iref = 0;
7954                 block_group->inode = NULL;
7955                 spin_unlock(&block_group->lock);
7956                 iput(inode);
7957                 last = block_group->key.objectid + block_group->key.offset;
7958                 btrfs_put_block_group(block_group);
7959         }
7960 }
7961
7962 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7963 {
7964         struct btrfs_block_group_cache *block_group;
7965         struct btrfs_space_info *space_info;
7966         struct btrfs_caching_control *caching_ctl;
7967         struct rb_node *n;
7968
7969         down_write(&info->extent_commit_sem);
7970         while (!list_empty(&info->caching_block_groups)) {
7971                 caching_ctl = list_entry(info->caching_block_groups.next,
7972                                          struct btrfs_caching_control, list);
7973                 list_del(&caching_ctl->list);
7974                 put_caching_control(caching_ctl);
7975         }
7976         up_write(&info->extent_commit_sem);
7977
7978         spin_lock(&info->block_group_cache_lock);
7979         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7980                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7981                                        cache_node);
7982                 rb_erase(&block_group->cache_node,
7983                          &info->block_group_cache_tree);
7984                 spin_unlock(&info->block_group_cache_lock);
7985
7986                 down_write(&block_group->space_info->groups_sem);
7987                 list_del(&block_group->list);
7988                 up_write(&block_group->space_info->groups_sem);
7989
7990                 if (block_group->cached == BTRFS_CACHE_STARTED)
7991                         wait_block_group_cache_done(block_group);
7992
7993                 /*
7994                  * We haven't cached this block group, which means we could
7995                  * possibly have excluded extents on this block group.
7996                  */
7997                 if (block_group->cached == BTRFS_CACHE_NO)
7998                         free_excluded_extents(info->extent_root, block_group);
7999
8000                 btrfs_remove_free_space_cache(block_group);
8001                 btrfs_put_block_group(block_group);
8002
8003                 spin_lock(&info->block_group_cache_lock);
8004         }
8005         spin_unlock(&info->block_group_cache_lock);
8006
8007         /* now that all the block groups are freed, go through and
8008          * free all the space_info structs.  This is only called during
8009          * the final stages of unmount, and so we know nobody is
8010          * using them.  We call synchronize_rcu() once before we start,
8011          * just to be on the safe side.
8012          */
8013         synchronize_rcu();
8014
8015         release_global_block_rsv(info);
8016
8017         while(!list_empty(&info->space_info)) {
8018                 space_info = list_entry(info->space_info.next,
8019                                         struct btrfs_space_info,
8020                                         list);
8021                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8022                         if (space_info->bytes_pinned > 0 ||
8023                             space_info->bytes_reserved > 0 ||
8024                             space_info->bytes_may_use > 0) {
8025                                 WARN_ON(1);
8026                                 dump_space_info(space_info, 0, 0);
8027                         }
8028                 }
8029                 list_del(&space_info->list);
8030                 kfree(space_info);
8031         }
8032         return 0;
8033 }
8034
8035 static void __link_block_group(struct btrfs_space_info *space_info,
8036                                struct btrfs_block_group_cache *cache)
8037 {
8038         int index = get_block_group_index(cache);
8039
8040         down_write(&space_info->groups_sem);
8041         list_add_tail(&cache->list, &space_info->block_groups[index]);
8042         up_write(&space_info->groups_sem);
8043 }
8044
8045 int btrfs_read_block_groups(struct btrfs_root *root)
8046 {
8047         struct btrfs_path *path;
8048         int ret;
8049         struct btrfs_block_group_cache *cache;
8050         struct btrfs_fs_info *info = root->fs_info;
8051         struct btrfs_space_info *space_info;
8052         struct btrfs_key key;
8053         struct btrfs_key found_key;
8054         struct extent_buffer *leaf;
8055         int need_clear = 0;
8056         u64 cache_gen;
8057
8058         root = info->extent_root;
8059         key.objectid = 0;
8060         key.offset = 0;
8061         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8062         path = btrfs_alloc_path();
8063         if (!path)
8064                 return -ENOMEM;
8065         path->reada = 1;
8066
8067         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8068         if (btrfs_test_opt(root, SPACE_CACHE) &&
8069             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8070                 need_clear = 1;
8071         if (btrfs_test_opt(root, CLEAR_CACHE))
8072                 need_clear = 1;
8073
8074         while (1) {
8075                 ret = find_first_block_group(root, path, &key);
8076                 if (ret > 0)
8077                         break;
8078                 if (ret != 0)
8079                         goto error;
8080                 leaf = path->nodes[0];
8081                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8082                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8083                 if (!cache) {
8084                         ret = -ENOMEM;
8085                         goto error;
8086                 }
8087                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8088                                                 GFP_NOFS);
8089                 if (!cache->free_space_ctl) {
8090                         kfree(cache);
8091                         ret = -ENOMEM;
8092                         goto error;
8093                 }
8094
8095                 atomic_set(&cache->count, 1);
8096                 spin_lock_init(&cache->lock);
8097                 cache->fs_info = info;
8098                 INIT_LIST_HEAD(&cache->list);
8099                 INIT_LIST_HEAD(&cache->cluster_list);
8100
8101                 if (need_clear) {
8102                         /*
8103                          * When we mount with old space cache, we need to
8104                          * set BTRFS_DC_CLEAR and set dirty flag.
8105                          *
8106                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8107                          *    truncate the old free space cache inode and
8108                          *    setup a new one.
8109                          * b) Setting 'dirty flag' makes sure that we flush
8110                          *    the new space cache info onto disk.
8111                          */
8112                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8113                         if (btrfs_test_opt(root, SPACE_CACHE))
8114                                 cache->dirty = 1;
8115                 }
8116
8117                 read_extent_buffer(leaf, &cache->item,
8118                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8119                                    sizeof(cache->item));
8120                 memcpy(&cache->key, &found_key, sizeof(found_key));
8121
8122                 key.objectid = found_key.objectid + found_key.offset;
8123                 btrfs_release_path(path);
8124                 cache->flags = btrfs_block_group_flags(&cache->item);
8125                 cache->sectorsize = root->sectorsize;
8126                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8127                                                &root->fs_info->mapping_tree,
8128                                                found_key.objectid);
8129                 btrfs_init_free_space_ctl(cache);
8130
8131                 /*
8132                  * We need to exclude the super stripes now so that the space
8133                  * info has super bytes accounted for, otherwise we'll think
8134                  * we have more space than we actually do.
8135                  */
8136                 ret = exclude_super_stripes(root, cache);
8137                 if (ret) {
8138                         /*
8139                          * We may have excluded something, so call this just in
8140                          * case.
8141                          */
8142                         free_excluded_extents(root, cache);
8143                         kfree(cache->free_space_ctl);
8144                         kfree(cache);
8145                         goto error;
8146                 }
8147
8148                 /*
8149                  * check for two cases, either we are full, and therefore
8150                  * don't need to bother with the caching work since we won't
8151                  * find any space, or we are empty, and we can just add all
8152                  * the space in and be done with it.  This saves us _alot_ of
8153                  * time, particularly in the full case.
8154                  */
8155                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8156                         cache->last_byte_to_unpin = (u64)-1;
8157                         cache->cached = BTRFS_CACHE_FINISHED;
8158                         free_excluded_extents(root, cache);
8159                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8160                         cache->last_byte_to_unpin = (u64)-1;
8161                         cache->cached = BTRFS_CACHE_FINISHED;
8162                         add_new_free_space(cache, root->fs_info,
8163                                            found_key.objectid,
8164                                            found_key.objectid +
8165                                            found_key.offset);
8166                         free_excluded_extents(root, cache);
8167                 }
8168
8169                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8170                 if (ret) {
8171                         btrfs_remove_free_space_cache(cache);
8172                         btrfs_put_block_group(cache);
8173                         goto error;
8174                 }
8175
8176                 ret = update_space_info(info, cache->flags, found_key.offset,
8177                                         btrfs_block_group_used(&cache->item),
8178                                         &space_info);
8179                 if (ret) {
8180                         btrfs_remove_free_space_cache(cache);
8181                         spin_lock(&info->block_group_cache_lock);
8182                         rb_erase(&cache->cache_node,
8183                                  &info->block_group_cache_tree);
8184                         spin_unlock(&info->block_group_cache_lock);
8185                         btrfs_put_block_group(cache);
8186                         goto error;
8187                 }
8188
8189                 cache->space_info = space_info;
8190                 spin_lock(&cache->space_info->lock);
8191                 cache->space_info->bytes_readonly += cache->bytes_super;
8192                 spin_unlock(&cache->space_info->lock);
8193
8194                 __link_block_group(space_info, cache);
8195
8196                 set_avail_alloc_bits(root->fs_info, cache->flags);
8197                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8198                         set_block_group_ro(cache, 1);
8199         }
8200
8201         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8202                 if (!(get_alloc_profile(root, space_info->flags) &
8203                       (BTRFS_BLOCK_GROUP_RAID10 |
8204                        BTRFS_BLOCK_GROUP_RAID1 |
8205                        BTRFS_BLOCK_GROUP_RAID5 |
8206                        BTRFS_BLOCK_GROUP_RAID6 |
8207                        BTRFS_BLOCK_GROUP_DUP)))
8208                         continue;
8209                 /*
8210                  * avoid allocating from un-mirrored block group if there are
8211                  * mirrored block groups.
8212                  */
8213                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8214                         set_block_group_ro(cache, 1);
8215                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8216                         set_block_group_ro(cache, 1);
8217         }
8218
8219         init_global_block_rsv(info);
8220         ret = 0;
8221 error:
8222         btrfs_free_path(path);
8223         return ret;
8224 }
8225
8226 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8227                                        struct btrfs_root *root)
8228 {
8229         struct btrfs_block_group_cache *block_group, *tmp;
8230         struct btrfs_root *extent_root = root->fs_info->extent_root;
8231         struct btrfs_block_group_item item;
8232         struct btrfs_key key;
8233         int ret = 0;
8234
8235         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8236                                  new_bg_list) {
8237                 list_del_init(&block_group->new_bg_list);
8238
8239                 if (ret)
8240                         continue;
8241
8242                 spin_lock(&block_group->lock);
8243                 memcpy(&item, &block_group->item, sizeof(item));
8244                 memcpy(&key, &block_group->key, sizeof(key));
8245                 spin_unlock(&block_group->lock);
8246
8247                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8248                                         sizeof(item));
8249                 if (ret)
8250                         btrfs_abort_transaction(trans, extent_root, ret);
8251         }
8252 }
8253
8254 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8255                            struct btrfs_root *root, u64 bytes_used,
8256                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8257                            u64 size)
8258 {
8259         int ret;
8260         struct btrfs_root *extent_root;
8261         struct btrfs_block_group_cache *cache;
8262
8263         extent_root = root->fs_info->extent_root;
8264
8265         root->fs_info->last_trans_log_full_commit = trans->transid;
8266
8267         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8268         if (!cache)
8269                 return -ENOMEM;
8270         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8271                                         GFP_NOFS);
8272         if (!cache->free_space_ctl) {
8273                 kfree(cache);
8274                 return -ENOMEM;
8275         }
8276
8277         cache->key.objectid = chunk_offset;
8278         cache->key.offset = size;
8279         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8280         cache->sectorsize = root->sectorsize;
8281         cache->fs_info = root->fs_info;
8282         cache->full_stripe_len = btrfs_full_stripe_len(root,
8283                                                &root->fs_info->mapping_tree,
8284                                                chunk_offset);
8285
8286         atomic_set(&cache->count, 1);
8287         spin_lock_init(&cache->lock);
8288         INIT_LIST_HEAD(&cache->list);
8289         INIT_LIST_HEAD(&cache->cluster_list);
8290         INIT_LIST_HEAD(&cache->new_bg_list);
8291
8292         btrfs_init_free_space_ctl(cache);
8293
8294         btrfs_set_block_group_used(&cache->item, bytes_used);
8295         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8296         cache->flags = type;
8297         btrfs_set_block_group_flags(&cache->item, type);
8298
8299         cache->last_byte_to_unpin = (u64)-1;
8300         cache->cached = BTRFS_CACHE_FINISHED;
8301         ret = exclude_super_stripes(root, cache);
8302         if (ret) {
8303                 /*
8304                  * We may have excluded something, so call this just in
8305                  * case.
8306                  */
8307                 free_excluded_extents(root, cache);
8308                 kfree(cache->free_space_ctl);
8309                 kfree(cache);
8310                 return ret;
8311         }
8312
8313         add_new_free_space(cache, root->fs_info, chunk_offset,
8314                            chunk_offset + size);
8315
8316         free_excluded_extents(root, cache);
8317
8318         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8319         if (ret) {
8320                 btrfs_remove_free_space_cache(cache);
8321                 btrfs_put_block_group(cache);
8322                 return ret;
8323         }
8324
8325         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8326                                 &cache->space_info);
8327         if (ret) {
8328                 btrfs_remove_free_space_cache(cache);
8329                 spin_lock(&root->fs_info->block_group_cache_lock);
8330                 rb_erase(&cache->cache_node,
8331                          &root->fs_info->block_group_cache_tree);
8332                 spin_unlock(&root->fs_info->block_group_cache_lock);
8333                 btrfs_put_block_group(cache);
8334                 return ret;
8335         }
8336         update_global_block_rsv(root->fs_info);
8337
8338         spin_lock(&cache->space_info->lock);
8339         cache->space_info->bytes_readonly += cache->bytes_super;
8340         spin_unlock(&cache->space_info->lock);
8341
8342         __link_block_group(cache->space_info, cache);
8343
8344         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8345
8346         set_avail_alloc_bits(extent_root->fs_info, type);
8347
8348         return 0;
8349 }
8350
8351 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8352 {
8353         u64 extra_flags = chunk_to_extended(flags) &
8354                                 BTRFS_EXTENDED_PROFILE_MASK;
8355
8356         write_seqlock(&fs_info->profiles_lock);
8357         if (flags & BTRFS_BLOCK_GROUP_DATA)
8358                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8359         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8360                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8361         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8362                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8363         write_sequnlock(&fs_info->profiles_lock);
8364 }
8365
8366 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8367                              struct btrfs_root *root, u64 group_start)
8368 {
8369         struct btrfs_path *path;
8370         struct btrfs_block_group_cache *block_group;
8371         struct btrfs_free_cluster *cluster;
8372         struct btrfs_root *tree_root = root->fs_info->tree_root;
8373         struct btrfs_key key;
8374         struct inode *inode;
8375         int ret;
8376         int index;
8377         int factor;
8378
8379         root = root->fs_info->extent_root;
8380
8381         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8382         BUG_ON(!block_group);
8383         BUG_ON(!block_group->ro);
8384
8385         /*
8386          * Free the reserved super bytes from this block group before
8387          * remove it.
8388          */
8389         free_excluded_extents(root, block_group);
8390
8391         memcpy(&key, &block_group->key, sizeof(key));
8392         index = get_block_group_index(block_group);
8393         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8394                                   BTRFS_BLOCK_GROUP_RAID1 |
8395                                   BTRFS_BLOCK_GROUP_RAID10))
8396                 factor = 2;
8397         else
8398                 factor = 1;
8399
8400         /* make sure this block group isn't part of an allocation cluster */
8401         cluster = &root->fs_info->data_alloc_cluster;
8402         spin_lock(&cluster->refill_lock);
8403         btrfs_return_cluster_to_free_space(block_group, cluster);
8404         spin_unlock(&cluster->refill_lock);
8405
8406         /*
8407          * make sure this block group isn't part of a metadata
8408          * allocation cluster
8409          */
8410         cluster = &root->fs_info->meta_alloc_cluster;
8411         spin_lock(&cluster->refill_lock);
8412         btrfs_return_cluster_to_free_space(block_group, cluster);
8413         spin_unlock(&cluster->refill_lock);
8414
8415         path = btrfs_alloc_path();
8416         if (!path) {
8417                 ret = -ENOMEM;
8418                 goto out;
8419         }
8420
8421         inode = lookup_free_space_inode(tree_root, block_group, path);
8422         if (!IS_ERR(inode)) {
8423                 ret = btrfs_orphan_add(trans, inode);
8424                 if (ret) {
8425                         btrfs_add_delayed_iput(inode);
8426                         goto out;
8427                 }
8428                 clear_nlink(inode);
8429                 /* One for the block groups ref */
8430                 spin_lock(&block_group->lock);
8431                 if (block_group->iref) {
8432                         block_group->iref = 0;
8433                         block_group->inode = NULL;
8434                         spin_unlock(&block_group->lock);
8435                         iput(inode);
8436                 } else {
8437                         spin_unlock(&block_group->lock);
8438                 }
8439                 /* One for our lookup ref */
8440                 btrfs_add_delayed_iput(inode);
8441         }
8442
8443         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8444         key.offset = block_group->key.objectid;
8445         key.type = 0;
8446
8447         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8448         if (ret < 0)
8449                 goto out;
8450         if (ret > 0)
8451                 btrfs_release_path(path);
8452         if (ret == 0) {
8453                 ret = btrfs_del_item(trans, tree_root, path);
8454                 if (ret)
8455                         goto out;
8456                 btrfs_release_path(path);
8457         }
8458
8459         spin_lock(&root->fs_info->block_group_cache_lock);
8460         rb_erase(&block_group->cache_node,
8461                  &root->fs_info->block_group_cache_tree);
8462
8463         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8464                 root->fs_info->first_logical_byte = (u64)-1;
8465         spin_unlock(&root->fs_info->block_group_cache_lock);
8466
8467         down_write(&block_group->space_info->groups_sem);
8468         /*
8469          * we must use list_del_init so people can check to see if they
8470          * are still on the list after taking the semaphore
8471          */
8472         list_del_init(&block_group->list);
8473         if (list_empty(&block_group->space_info->block_groups[index]))
8474                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8475         up_write(&block_group->space_info->groups_sem);
8476
8477         if (block_group->cached == BTRFS_CACHE_STARTED)
8478                 wait_block_group_cache_done(block_group);
8479
8480         btrfs_remove_free_space_cache(block_group);
8481
8482         spin_lock(&block_group->space_info->lock);
8483         block_group->space_info->total_bytes -= block_group->key.offset;
8484         block_group->space_info->bytes_readonly -= block_group->key.offset;
8485         block_group->space_info->disk_total -= block_group->key.offset * factor;
8486         spin_unlock(&block_group->space_info->lock);
8487
8488         memcpy(&key, &block_group->key, sizeof(key));
8489
8490         btrfs_clear_space_info_full(root->fs_info);
8491
8492         btrfs_put_block_group(block_group);
8493         btrfs_put_block_group(block_group);
8494
8495         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8496         if (ret > 0)
8497                 ret = -EIO;
8498         if (ret < 0)
8499                 goto out;
8500
8501         ret = btrfs_del_item(trans, root, path);
8502 out:
8503         btrfs_free_path(path);
8504         return ret;
8505 }
8506
8507 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8508 {
8509         struct btrfs_space_info *space_info;
8510         struct btrfs_super_block *disk_super;
8511         u64 features;
8512         u64 flags;
8513         int mixed = 0;
8514         int ret;
8515
8516         disk_super = fs_info->super_copy;
8517         if (!btrfs_super_root(disk_super))
8518                 return 1;
8519
8520         features = btrfs_super_incompat_flags(disk_super);
8521         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8522                 mixed = 1;
8523
8524         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8525         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8526         if (ret)
8527                 goto out;
8528
8529         if (mixed) {
8530                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8531                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8532         } else {
8533                 flags = BTRFS_BLOCK_GROUP_METADATA;
8534                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8535                 if (ret)
8536                         goto out;
8537
8538                 flags = BTRFS_BLOCK_GROUP_DATA;
8539                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8540         }
8541 out:
8542         return ret;
8543 }
8544
8545 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8546 {
8547         return unpin_extent_range(root, start, end);
8548 }
8549
8550 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8551                                u64 num_bytes, u64 *actual_bytes)
8552 {
8553         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8554 }
8555
8556 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8557 {
8558         struct btrfs_fs_info *fs_info = root->fs_info;
8559         struct btrfs_block_group_cache *cache = NULL;
8560         u64 group_trimmed;
8561         u64 start;
8562         u64 end;
8563         u64 trimmed = 0;
8564         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8565         int ret = 0;
8566
8567         /*
8568          * try to trim all FS space, our block group may start from non-zero.
8569          */
8570         if (range->len == total_bytes)
8571                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8572         else
8573                 cache = btrfs_lookup_block_group(fs_info, range->start);
8574
8575         while (cache) {
8576                 if (cache->key.objectid >= (range->start + range->len)) {
8577                         btrfs_put_block_group(cache);
8578                         break;
8579                 }
8580
8581                 start = max(range->start, cache->key.objectid);
8582                 end = min(range->start + range->len,
8583                                 cache->key.objectid + cache->key.offset);
8584
8585                 if (end - start >= range->minlen) {
8586                         if (!block_group_cache_done(cache)) {
8587                                 ret = cache_block_group(cache, 0);
8588                                 if (!ret)
8589                                         wait_block_group_cache_done(cache);
8590                         }
8591                         ret = btrfs_trim_block_group(cache,
8592                                                      &group_trimmed,
8593                                                      start,
8594                                                      end,
8595                                                      range->minlen);
8596
8597                         trimmed += group_trimmed;
8598                         if (ret) {
8599                                 btrfs_put_block_group(cache);
8600                                 break;
8601                         }
8602                 }
8603
8604                 cache = next_block_group(fs_info->tree_root, cache);
8605         }
8606
8607         range->len = trimmed;
8608         return ret;
8609 }