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 = !extent_op->is_data;
2074
2075         if (trans->aborted)
2076                 return 0;
2077
2078         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2079                 metadata = 0;
2080
2081         path = btrfs_alloc_path();
2082         if (!path)
2083                 return -ENOMEM;
2084
2085         key.objectid = node->bytenr;
2086
2087         if (metadata) {
2088                 key.type = BTRFS_METADATA_ITEM_KEY;
2089                 key.offset = extent_op->level;
2090         } else {
2091                 key.type = BTRFS_EXTENT_ITEM_KEY;
2092                 key.offset = node->num_bytes;
2093         }
2094
2095 again:
2096         path->reada = 1;
2097         path->leave_spinning = 1;
2098         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2099                                 path, 0, 1);
2100         if (ret < 0) {
2101                 err = ret;
2102                 goto out;
2103         }
2104         if (ret > 0) {
2105                 if (metadata) {
2106                         btrfs_release_path(path);
2107                         metadata = 0;
2108
2109                         key.offset = node->num_bytes;
2110                         key.type = BTRFS_EXTENT_ITEM_KEY;
2111                         goto again;
2112                 }
2113                 err = -EIO;
2114                 goto out;
2115         }
2116
2117         leaf = path->nodes[0];
2118         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2119 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2120         if (item_size < sizeof(*ei)) {
2121                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2122                                              path, (u64)-1, 0);
2123                 if (ret < 0) {
2124                         err = ret;
2125                         goto out;
2126                 }
2127                 leaf = path->nodes[0];
2128                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2129         }
2130 #endif
2131         BUG_ON(item_size < sizeof(*ei));
2132         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2133         __run_delayed_extent_op(extent_op, leaf, ei);
2134
2135         btrfs_mark_buffer_dirty(leaf);
2136 out:
2137         btrfs_free_path(path);
2138         return err;
2139 }
2140
2141 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2142                                 struct btrfs_root *root,
2143                                 struct btrfs_delayed_ref_node *node,
2144                                 struct btrfs_delayed_extent_op *extent_op,
2145                                 int insert_reserved)
2146 {
2147         int ret = 0;
2148         struct btrfs_delayed_tree_ref *ref;
2149         struct btrfs_key ins;
2150         u64 parent = 0;
2151         u64 ref_root = 0;
2152         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2153                                                  SKINNY_METADATA);
2154
2155         ref = btrfs_delayed_node_to_tree_ref(node);
2156         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2157                 parent = ref->parent;
2158         else
2159                 ref_root = ref->root;
2160
2161         ins.objectid = node->bytenr;
2162         if (skinny_metadata) {
2163                 ins.offset = ref->level;
2164                 ins.type = BTRFS_METADATA_ITEM_KEY;
2165         } else {
2166                 ins.offset = node->num_bytes;
2167                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2168         }
2169
2170         BUG_ON(node->ref_mod != 1);
2171         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2172                 BUG_ON(!extent_op || !extent_op->update_flags);
2173                 ret = alloc_reserved_tree_block(trans, root,
2174                                                 parent, ref_root,
2175                                                 extent_op->flags_to_set,
2176                                                 &extent_op->key,
2177                                                 ref->level, &ins);
2178         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2179                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2180                                              node->num_bytes, parent, ref_root,
2181                                              ref->level, 0, 1, extent_op);
2182         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2183                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2184                                           node->num_bytes, parent, ref_root,
2185                                           ref->level, 0, 1, extent_op);
2186         } else {
2187                 BUG();
2188         }
2189         return ret;
2190 }
2191
2192 /* helper function to actually process a single delayed ref entry */
2193 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2194                                struct btrfs_root *root,
2195                                struct btrfs_delayed_ref_node *node,
2196                                struct btrfs_delayed_extent_op *extent_op,
2197                                int insert_reserved)
2198 {
2199         int ret = 0;
2200
2201         if (trans->aborted)
2202                 return 0;
2203
2204         if (btrfs_delayed_ref_is_head(node)) {
2205                 struct btrfs_delayed_ref_head *head;
2206                 /*
2207                  * we've hit the end of the chain and we were supposed
2208                  * to insert this extent into the tree.  But, it got
2209                  * deleted before we ever needed to insert it, so all
2210                  * we have to do is clean up the accounting
2211                  */
2212                 BUG_ON(extent_op);
2213                 head = btrfs_delayed_node_to_head(node);
2214                 if (insert_reserved) {
2215                         btrfs_pin_extent(root, node->bytenr,
2216                                          node->num_bytes, 1);
2217                         if (head->is_data) {
2218                                 ret = btrfs_del_csums(trans, root,
2219                                                       node->bytenr,
2220                                                       node->num_bytes);
2221                         }
2222                 }
2223                 return ret;
2224         }
2225
2226         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2227             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2228                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2229                                            insert_reserved);
2230         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2231                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2232                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2233                                            insert_reserved);
2234         else
2235                 BUG();
2236         return ret;
2237 }
2238
2239 static noinline struct btrfs_delayed_ref_node *
2240 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2241 {
2242         struct rb_node *node;
2243         struct btrfs_delayed_ref_node *ref;
2244         int action = BTRFS_ADD_DELAYED_REF;
2245 again:
2246         /*
2247          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2248          * this prevents ref count from going down to zero when
2249          * there still are pending delayed ref.
2250          */
2251         node = rb_prev(&head->node.rb_node);
2252         while (1) {
2253                 if (!node)
2254                         break;
2255                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2256                                 rb_node);
2257                 if (ref->bytenr != head->node.bytenr)
2258                         break;
2259                 if (ref->action == action)
2260                         return ref;
2261                 node = rb_prev(node);
2262         }
2263         if (action == BTRFS_ADD_DELAYED_REF) {
2264                 action = BTRFS_DROP_DELAYED_REF;
2265                 goto again;
2266         }
2267         return NULL;
2268 }
2269
2270 /*
2271  * Returns 0 on success or if called with an already aborted transaction.
2272  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2273  */
2274 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2275                                        struct btrfs_root *root,
2276                                        struct list_head *cluster)
2277 {
2278         struct btrfs_delayed_ref_root *delayed_refs;
2279         struct btrfs_delayed_ref_node *ref;
2280         struct btrfs_delayed_ref_head *locked_ref = NULL;
2281         struct btrfs_delayed_extent_op *extent_op;
2282         struct btrfs_fs_info *fs_info = root->fs_info;
2283         int ret;
2284         int count = 0;
2285         int must_insert_reserved = 0;
2286
2287         delayed_refs = &trans->transaction->delayed_refs;
2288         while (1) {
2289                 if (!locked_ref) {
2290                         /* pick a new head ref from the cluster list */
2291                         if (list_empty(cluster))
2292                                 break;
2293
2294                         locked_ref = list_entry(cluster->next,
2295                                      struct btrfs_delayed_ref_head, cluster);
2296
2297                         /* grab the lock that says we are going to process
2298                          * all the refs for this head */
2299                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2300
2301                         /*
2302                          * we may have dropped the spin lock to get the head
2303                          * mutex lock, and that might have given someone else
2304                          * time to free the head.  If that's true, it has been
2305                          * removed from our list and we can move on.
2306                          */
2307                         if (ret == -EAGAIN) {
2308                                 locked_ref = NULL;
2309                                 count++;
2310                                 continue;
2311                         }
2312                 }
2313
2314                 /*
2315                  * We need to try and merge add/drops of the same ref since we
2316                  * can run into issues with relocate dropping the implicit ref
2317                  * and then it being added back again before the drop can
2318                  * finish.  If we merged anything we need to re-loop so we can
2319                  * get a good ref.
2320                  */
2321                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2322                                          locked_ref);
2323
2324                 /*
2325                  * locked_ref is the head node, so we have to go one
2326                  * node back for any delayed ref updates
2327                  */
2328                 ref = select_delayed_ref(locked_ref);
2329
2330                 if (ref && ref->seq &&
2331                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2332                         /*
2333                          * there are still refs with lower seq numbers in the
2334                          * process of being added. Don't run this ref yet.
2335                          */
2336                         list_del_init(&locked_ref->cluster);
2337                         btrfs_delayed_ref_unlock(locked_ref);
2338                         locked_ref = NULL;
2339                         delayed_refs->num_heads_ready++;
2340                         spin_unlock(&delayed_refs->lock);
2341                         cond_resched();
2342                         spin_lock(&delayed_refs->lock);
2343                         continue;
2344                 }
2345
2346                 /*
2347                  * record the must insert reserved flag before we
2348                  * drop the spin lock.
2349                  */
2350                 must_insert_reserved = locked_ref->must_insert_reserved;
2351                 locked_ref->must_insert_reserved = 0;
2352
2353                 extent_op = locked_ref->extent_op;
2354                 locked_ref->extent_op = NULL;
2355
2356                 if (!ref) {
2357                         /* All delayed refs have been processed, Go ahead
2358                          * and send the head node to run_one_delayed_ref,
2359                          * so that any accounting fixes can happen
2360                          */
2361                         ref = &locked_ref->node;
2362
2363                         if (extent_op && must_insert_reserved) {
2364                                 btrfs_free_delayed_extent_op(extent_op);
2365                                 extent_op = NULL;
2366                         }
2367
2368                         if (extent_op) {
2369                                 spin_unlock(&delayed_refs->lock);
2370
2371                                 ret = run_delayed_extent_op(trans, root,
2372                                                             ref, extent_op);
2373                                 btrfs_free_delayed_extent_op(extent_op);
2374
2375                                 if (ret) {
2376                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2377                                         spin_lock(&delayed_refs->lock);
2378                                         btrfs_delayed_ref_unlock(locked_ref);
2379                                         return ret;
2380                                 }
2381
2382                                 goto next;
2383                         }
2384                 }
2385
2386                 ref->in_tree = 0;
2387                 rb_erase(&ref->rb_node, &delayed_refs->root);
2388                 delayed_refs->num_entries--;
2389                 if (!btrfs_delayed_ref_is_head(ref)) {
2390                         /*
2391                          * when we play the delayed ref, also correct the
2392                          * ref_mod on head
2393                          */
2394                         switch (ref->action) {
2395                         case BTRFS_ADD_DELAYED_REF:
2396                         case BTRFS_ADD_DELAYED_EXTENT:
2397                                 locked_ref->node.ref_mod -= ref->ref_mod;
2398                                 break;
2399                         case BTRFS_DROP_DELAYED_REF:
2400                                 locked_ref->node.ref_mod += ref->ref_mod;
2401                                 break;
2402                         default:
2403                                 WARN_ON(1);
2404                         }
2405                 }
2406                 spin_unlock(&delayed_refs->lock);
2407
2408                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2409                                           must_insert_reserved);
2410
2411                 btrfs_free_delayed_extent_op(extent_op);
2412                 if (ret) {
2413                         btrfs_delayed_ref_unlock(locked_ref);
2414                         btrfs_put_delayed_ref(ref);
2415                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2416                         spin_lock(&delayed_refs->lock);
2417                         return ret;
2418                 }
2419
2420                 /*
2421                  * If this node is a head, that means all the refs in this head
2422                  * have been dealt with, and we will pick the next head to deal
2423                  * with, so we must unlock the head and drop it from the cluster
2424                  * list before we release it.
2425                  */
2426                 if (btrfs_delayed_ref_is_head(ref)) {
2427                         list_del_init(&locked_ref->cluster);
2428                         btrfs_delayed_ref_unlock(locked_ref);
2429                         locked_ref = NULL;
2430                 }
2431                 btrfs_put_delayed_ref(ref);
2432                 count++;
2433 next:
2434                 cond_resched();
2435                 spin_lock(&delayed_refs->lock);
2436         }
2437         return count;
2438 }
2439
2440 #ifdef SCRAMBLE_DELAYED_REFS
2441 /*
2442  * Normally delayed refs get processed in ascending bytenr order. This
2443  * correlates in most cases to the order added. To expose dependencies on this
2444  * order, we start to process the tree in the middle instead of the beginning
2445  */
2446 static u64 find_middle(struct rb_root *root)
2447 {
2448         struct rb_node *n = root->rb_node;
2449         struct btrfs_delayed_ref_node *entry;
2450         int alt = 1;
2451         u64 middle;
2452         u64 first = 0, last = 0;
2453
2454         n = rb_first(root);
2455         if (n) {
2456                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2457                 first = entry->bytenr;
2458         }
2459         n = rb_last(root);
2460         if (n) {
2461                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2462                 last = entry->bytenr;
2463         }
2464         n = root->rb_node;
2465
2466         while (n) {
2467                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2468                 WARN_ON(!entry->in_tree);
2469
2470                 middle = entry->bytenr;
2471
2472                 if (alt)
2473                         n = n->rb_left;
2474                 else
2475                         n = n->rb_right;
2476
2477                 alt = 1 - alt;
2478         }
2479         return middle;
2480 }
2481 #endif
2482
2483 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2484                                          struct btrfs_fs_info *fs_info)
2485 {
2486         struct qgroup_update *qgroup_update;
2487         int ret = 0;
2488
2489         if (list_empty(&trans->qgroup_ref_list) !=
2490             !trans->delayed_ref_elem.seq) {
2491                 /* list without seq or seq without list */
2492                 btrfs_err(fs_info,
2493                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2494                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2495                         (u32)(trans->delayed_ref_elem.seq >> 32),
2496                         (u32)trans->delayed_ref_elem.seq);
2497                 BUG();
2498         }
2499
2500         if (!trans->delayed_ref_elem.seq)
2501                 return 0;
2502
2503         while (!list_empty(&trans->qgroup_ref_list)) {
2504                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2505                                                  struct qgroup_update, list);
2506                 list_del(&qgroup_update->list);
2507                 if (!ret)
2508                         ret = btrfs_qgroup_account_ref(
2509                                         trans, fs_info, qgroup_update->node,
2510                                         qgroup_update->extent_op);
2511                 kfree(qgroup_update);
2512         }
2513
2514         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2515
2516         return ret;
2517 }
2518
2519 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2520                       int count)
2521 {
2522         int val = atomic_read(&delayed_refs->ref_seq);
2523
2524         if (val < seq || val >= seq + count)
2525                 return 1;
2526         return 0;
2527 }
2528
2529 /*
2530  * this starts processing the delayed reference count updates and
2531  * extent insertions we have queued up so far.  count can be
2532  * 0, which means to process everything in the tree at the start
2533  * of the run (but not newly added entries), or it can be some target
2534  * number you'd like to process.
2535  *
2536  * Returns 0 on success or if called with an aborted transaction
2537  * Returns <0 on error and aborts the transaction
2538  */
2539 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2540                            struct btrfs_root *root, unsigned long count)
2541 {
2542         struct rb_node *node;
2543         struct btrfs_delayed_ref_root *delayed_refs;
2544         struct btrfs_delayed_ref_node *ref;
2545         struct list_head cluster;
2546         int ret;
2547         u64 delayed_start;
2548         int run_all = count == (unsigned long)-1;
2549         int run_most = 0;
2550         int loops;
2551
2552         /* We'll clean this up in btrfs_cleanup_transaction */
2553         if (trans->aborted)
2554                 return 0;
2555
2556         if (root == root->fs_info->extent_root)
2557                 root = root->fs_info->tree_root;
2558
2559         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2560
2561         delayed_refs = &trans->transaction->delayed_refs;
2562         INIT_LIST_HEAD(&cluster);
2563         if (count == 0) {
2564                 count = delayed_refs->num_entries * 2;
2565                 run_most = 1;
2566         }
2567
2568         if (!run_all && !run_most) {
2569                 int old;
2570                 int seq = atomic_read(&delayed_refs->ref_seq);
2571
2572 progress:
2573                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2574                 if (old) {
2575                         DEFINE_WAIT(__wait);
2576                         if (delayed_refs->num_entries < 16348)
2577                                 return 0;
2578
2579                         prepare_to_wait(&delayed_refs->wait, &__wait,
2580                                         TASK_UNINTERRUPTIBLE);
2581
2582                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2583                         if (old) {
2584                                 schedule();
2585                                 finish_wait(&delayed_refs->wait, &__wait);
2586
2587                                 if (!refs_newer(delayed_refs, seq, 256))
2588                                         goto progress;
2589                                 else
2590                                         return 0;
2591                         } else {
2592                                 finish_wait(&delayed_refs->wait, &__wait);
2593                                 goto again;
2594                         }
2595                 }
2596
2597         } else {
2598                 atomic_inc(&delayed_refs->procs_running_refs);
2599         }
2600
2601 again:
2602         loops = 0;
2603         spin_lock(&delayed_refs->lock);
2604
2605 #ifdef SCRAMBLE_DELAYED_REFS
2606         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2607 #endif
2608
2609         while (1) {
2610                 if (!(run_all || run_most) &&
2611                     delayed_refs->num_heads_ready < 64)
2612                         break;
2613
2614                 /*
2615                  * go find something we can process in the rbtree.  We start at
2616                  * the beginning of the tree, and then build a cluster
2617                  * of refs to process starting at the first one we are able to
2618                  * lock
2619                  */
2620                 delayed_start = delayed_refs->run_delayed_start;
2621                 ret = btrfs_find_ref_cluster(trans, &cluster,
2622                                              delayed_refs->run_delayed_start);
2623                 if (ret)
2624                         break;
2625
2626                 ret = run_clustered_refs(trans, root, &cluster);
2627                 if (ret < 0) {
2628                         btrfs_release_ref_cluster(&cluster);
2629                         spin_unlock(&delayed_refs->lock);
2630                         btrfs_abort_transaction(trans, root, ret);
2631                         atomic_dec(&delayed_refs->procs_running_refs);
2632                         wake_up(&delayed_refs->wait);
2633                         return ret;
2634                 }
2635
2636                 atomic_add(ret, &delayed_refs->ref_seq);
2637
2638                 count -= min_t(unsigned long, ret, count);
2639
2640                 if (count == 0)
2641                         break;
2642
2643                 if (delayed_start >= delayed_refs->run_delayed_start) {
2644                         if (loops == 0) {
2645                                 /*
2646                                  * btrfs_find_ref_cluster looped. let's do one
2647                                  * more cycle. if we don't run any delayed ref
2648                                  * during that cycle (because we can't because
2649                                  * all of them are blocked), bail out.
2650                                  */
2651                                 loops = 1;
2652                         } else {
2653                                 /*
2654                                  * no runnable refs left, stop trying
2655                                  */
2656                                 BUG_ON(run_all);
2657                                 break;
2658                         }
2659                 }
2660                 if (ret) {
2661                         /* refs were run, let's reset staleness detection */
2662                         loops = 0;
2663                 }
2664         }
2665
2666         if (run_all) {
2667                 if (!list_empty(&trans->new_bgs)) {
2668                         spin_unlock(&delayed_refs->lock);
2669                         btrfs_create_pending_block_groups(trans, root);
2670                         spin_lock(&delayed_refs->lock);
2671                 }
2672
2673                 node = rb_first(&delayed_refs->root);
2674                 if (!node)
2675                         goto out;
2676                 count = (unsigned long)-1;
2677
2678                 while (node) {
2679                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2680                                        rb_node);
2681                         if (btrfs_delayed_ref_is_head(ref)) {
2682                                 struct btrfs_delayed_ref_head *head;
2683
2684                                 head = btrfs_delayed_node_to_head(ref);
2685                                 atomic_inc(&ref->refs);
2686
2687                                 spin_unlock(&delayed_refs->lock);
2688                                 /*
2689                                  * Mutex was contended, block until it's
2690                                  * released and try again
2691                                  */
2692                                 mutex_lock(&head->mutex);
2693                                 mutex_unlock(&head->mutex);
2694
2695                                 btrfs_put_delayed_ref(ref);
2696                                 cond_resched();
2697                                 goto again;
2698                         }
2699                         node = rb_next(node);
2700                 }
2701                 spin_unlock(&delayed_refs->lock);
2702                 schedule_timeout(1);
2703                 goto again;
2704         }
2705 out:
2706         atomic_dec(&delayed_refs->procs_running_refs);
2707         smp_mb();
2708         if (waitqueue_active(&delayed_refs->wait))
2709                 wake_up(&delayed_refs->wait);
2710
2711         spin_unlock(&delayed_refs->lock);
2712         assert_qgroups_uptodate(trans);
2713         return 0;
2714 }
2715
2716 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2717                                 struct btrfs_root *root,
2718                                 u64 bytenr, u64 num_bytes, u64 flags,
2719                                 int level, int is_data)
2720 {
2721         struct btrfs_delayed_extent_op *extent_op;
2722         int ret;
2723
2724         extent_op = btrfs_alloc_delayed_extent_op();
2725         if (!extent_op)
2726                 return -ENOMEM;
2727
2728         extent_op->flags_to_set = flags;
2729         extent_op->update_flags = 1;
2730         extent_op->update_key = 0;
2731         extent_op->is_data = is_data ? 1 : 0;
2732         extent_op->level = level;
2733
2734         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2735                                           num_bytes, extent_op);
2736         if (ret)
2737                 btrfs_free_delayed_extent_op(extent_op);
2738         return ret;
2739 }
2740
2741 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2742                                       struct btrfs_root *root,
2743                                       struct btrfs_path *path,
2744                                       u64 objectid, u64 offset, u64 bytenr)
2745 {
2746         struct btrfs_delayed_ref_head *head;
2747         struct btrfs_delayed_ref_node *ref;
2748         struct btrfs_delayed_data_ref *data_ref;
2749         struct btrfs_delayed_ref_root *delayed_refs;
2750         struct rb_node *node;
2751         int ret = 0;
2752
2753         ret = -ENOENT;
2754         delayed_refs = &trans->transaction->delayed_refs;
2755         spin_lock(&delayed_refs->lock);
2756         head = btrfs_find_delayed_ref_head(trans, bytenr);
2757         if (!head)
2758                 goto out;
2759
2760         if (!mutex_trylock(&head->mutex)) {
2761                 atomic_inc(&head->node.refs);
2762                 spin_unlock(&delayed_refs->lock);
2763
2764                 btrfs_release_path(path);
2765
2766                 /*
2767                  * Mutex was contended, block until it's released and let
2768                  * caller try again
2769                  */
2770                 mutex_lock(&head->mutex);
2771                 mutex_unlock(&head->mutex);
2772                 btrfs_put_delayed_ref(&head->node);
2773                 return -EAGAIN;
2774         }
2775
2776         node = rb_prev(&head->node.rb_node);
2777         if (!node)
2778                 goto out_unlock;
2779
2780         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2781
2782         if (ref->bytenr != bytenr)
2783                 goto out_unlock;
2784
2785         ret = 1;
2786         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2787                 goto out_unlock;
2788
2789         data_ref = btrfs_delayed_node_to_data_ref(ref);
2790
2791         node = rb_prev(node);
2792         if (node) {
2793                 int seq = ref->seq;
2794
2795                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2796                 if (ref->bytenr == bytenr && ref->seq == seq)
2797                         goto out_unlock;
2798         }
2799
2800         if (data_ref->root != root->root_key.objectid ||
2801             data_ref->objectid != objectid || data_ref->offset != offset)
2802                 goto out_unlock;
2803
2804         ret = 0;
2805 out_unlock:
2806         mutex_unlock(&head->mutex);
2807 out:
2808         spin_unlock(&delayed_refs->lock);
2809         return ret;
2810 }
2811
2812 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2813                                         struct btrfs_root *root,
2814                                         struct btrfs_path *path,
2815                                         u64 objectid, u64 offset, u64 bytenr)
2816 {
2817         struct btrfs_root *extent_root = root->fs_info->extent_root;
2818         struct extent_buffer *leaf;
2819         struct btrfs_extent_data_ref *ref;
2820         struct btrfs_extent_inline_ref *iref;
2821         struct btrfs_extent_item *ei;
2822         struct btrfs_key key;
2823         u32 item_size;
2824         int ret;
2825
2826         key.objectid = bytenr;
2827         key.offset = (u64)-1;
2828         key.type = BTRFS_EXTENT_ITEM_KEY;
2829
2830         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2831         if (ret < 0)
2832                 goto out;
2833         BUG_ON(ret == 0); /* Corruption */
2834
2835         ret = -ENOENT;
2836         if (path->slots[0] == 0)
2837                 goto out;
2838
2839         path->slots[0]--;
2840         leaf = path->nodes[0];
2841         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2842
2843         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2844                 goto out;
2845
2846         ret = 1;
2847         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2848 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2849         if (item_size < sizeof(*ei)) {
2850                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2851                 goto out;
2852         }
2853 #endif
2854         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2855
2856         if (item_size != sizeof(*ei) +
2857             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2858                 goto out;
2859
2860         if (btrfs_extent_generation(leaf, ei) <=
2861             btrfs_root_last_snapshot(&root->root_item))
2862                 goto out;
2863
2864         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2865         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2866             BTRFS_EXTENT_DATA_REF_KEY)
2867                 goto out;
2868
2869         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2870         if (btrfs_extent_refs(leaf, ei) !=
2871             btrfs_extent_data_ref_count(leaf, ref) ||
2872             btrfs_extent_data_ref_root(leaf, ref) !=
2873             root->root_key.objectid ||
2874             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2875             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2876                 goto out;
2877
2878         ret = 0;
2879 out:
2880         return ret;
2881 }
2882
2883 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2884                           struct btrfs_root *root,
2885                           u64 objectid, u64 offset, u64 bytenr)
2886 {
2887         struct btrfs_path *path;
2888         int ret;
2889         int ret2;
2890
2891         path = btrfs_alloc_path();
2892         if (!path)
2893                 return -ENOENT;
2894
2895         do {
2896                 ret = check_committed_ref(trans, root, path, objectid,
2897                                           offset, bytenr);
2898                 if (ret && ret != -ENOENT)
2899                         goto out;
2900
2901                 ret2 = check_delayed_ref(trans, root, path, objectid,
2902                                          offset, bytenr);
2903         } while (ret2 == -EAGAIN);
2904
2905         if (ret2 && ret2 != -ENOENT) {
2906                 ret = ret2;
2907                 goto out;
2908         }
2909
2910         if (ret != -ENOENT || ret2 != -ENOENT)
2911                 ret = 0;
2912 out:
2913         btrfs_free_path(path);
2914         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2915                 WARN_ON(ret > 0);
2916         return ret;
2917 }
2918
2919 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2920                            struct btrfs_root *root,
2921                            struct extent_buffer *buf,
2922                            int full_backref, int inc, int for_cow)
2923 {
2924         u64 bytenr;
2925         u64 num_bytes;
2926         u64 parent;
2927         u64 ref_root;
2928         u32 nritems;
2929         struct btrfs_key key;
2930         struct btrfs_file_extent_item *fi;
2931         int i;
2932         int level;
2933         int ret = 0;
2934         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2935                             u64, u64, u64, u64, u64, u64, int);
2936
2937         ref_root = btrfs_header_owner(buf);
2938         nritems = btrfs_header_nritems(buf);
2939         level = btrfs_header_level(buf);
2940
2941         if (!root->ref_cows && level == 0)
2942                 return 0;
2943
2944         if (inc)
2945                 process_func = btrfs_inc_extent_ref;
2946         else
2947                 process_func = btrfs_free_extent;
2948
2949         if (full_backref)
2950                 parent = buf->start;
2951         else
2952                 parent = 0;
2953
2954         for (i = 0; i < nritems; i++) {
2955                 if (level == 0) {
2956                         btrfs_item_key_to_cpu(buf, &key, i);
2957                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2958                                 continue;
2959                         fi = btrfs_item_ptr(buf, i,
2960                                             struct btrfs_file_extent_item);
2961                         if (btrfs_file_extent_type(buf, fi) ==
2962                             BTRFS_FILE_EXTENT_INLINE)
2963                                 continue;
2964                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2965                         if (bytenr == 0)
2966                                 continue;
2967
2968                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2969                         key.offset -= btrfs_file_extent_offset(buf, fi);
2970                         ret = process_func(trans, root, bytenr, num_bytes,
2971                                            parent, ref_root, key.objectid,
2972                                            key.offset, for_cow);
2973                         if (ret)
2974                                 goto fail;
2975                 } else {
2976                         bytenr = btrfs_node_blockptr(buf, i);
2977                         num_bytes = btrfs_level_size(root, level - 1);
2978                         ret = process_func(trans, root, bytenr, num_bytes,
2979                                            parent, ref_root, level - 1, 0,
2980                                            for_cow);
2981                         if (ret)
2982                                 goto fail;
2983                 }
2984         }
2985         return 0;
2986 fail:
2987         return ret;
2988 }
2989
2990 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2991                   struct extent_buffer *buf, int full_backref, int for_cow)
2992 {
2993         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2994 }
2995
2996 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2997                   struct extent_buffer *buf, int full_backref, int for_cow)
2998 {
2999         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3000 }
3001
3002 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3003                                  struct btrfs_root *root,
3004                                  struct btrfs_path *path,
3005                                  struct btrfs_block_group_cache *cache)
3006 {
3007         int ret;
3008         struct btrfs_root *extent_root = root->fs_info->extent_root;
3009         unsigned long bi;
3010         struct extent_buffer *leaf;
3011
3012         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3013         if (ret < 0)
3014                 goto fail;
3015         BUG_ON(ret); /* Corruption */
3016
3017         leaf = path->nodes[0];
3018         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3019         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3020         btrfs_mark_buffer_dirty(leaf);
3021         btrfs_release_path(path);
3022 fail:
3023         if (ret) {
3024                 btrfs_abort_transaction(trans, root, ret);
3025                 return ret;
3026         }
3027         return 0;
3028
3029 }
3030
3031 static struct btrfs_block_group_cache *
3032 next_block_group(struct btrfs_root *root,
3033                  struct btrfs_block_group_cache *cache)
3034 {
3035         struct rb_node *node;
3036         spin_lock(&root->fs_info->block_group_cache_lock);
3037         node = rb_next(&cache->cache_node);
3038         btrfs_put_block_group(cache);
3039         if (node) {
3040                 cache = rb_entry(node, struct btrfs_block_group_cache,
3041                                  cache_node);
3042                 btrfs_get_block_group(cache);
3043         } else
3044                 cache = NULL;
3045         spin_unlock(&root->fs_info->block_group_cache_lock);
3046         return cache;
3047 }
3048
3049 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3050                             struct btrfs_trans_handle *trans,
3051                             struct btrfs_path *path)
3052 {
3053         struct btrfs_root *root = block_group->fs_info->tree_root;
3054         struct inode *inode = NULL;
3055         u64 alloc_hint = 0;
3056         int dcs = BTRFS_DC_ERROR;
3057         int num_pages = 0;
3058         int retries = 0;
3059         int ret = 0;
3060
3061         /*
3062          * If this block group is smaller than 100 megs don't bother caching the
3063          * block group.
3064          */
3065         if (block_group->key.offset < (100 * 1024 * 1024)) {
3066                 spin_lock(&block_group->lock);
3067                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3068                 spin_unlock(&block_group->lock);
3069                 return 0;
3070         }
3071
3072 again:
3073         inode = lookup_free_space_inode(root, block_group, path);
3074         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3075                 ret = PTR_ERR(inode);
3076                 btrfs_release_path(path);
3077                 goto out;
3078         }
3079
3080         if (IS_ERR(inode)) {
3081                 BUG_ON(retries);
3082                 retries++;
3083
3084                 if (block_group->ro)
3085                         goto out_free;
3086
3087                 ret = create_free_space_inode(root, trans, block_group, path);
3088                 if (ret)
3089                         goto out_free;
3090                 goto again;
3091         }
3092
3093         /* We've already setup this transaction, go ahead and exit */
3094         if (block_group->cache_generation == trans->transid &&
3095             i_size_read(inode)) {
3096                 dcs = BTRFS_DC_SETUP;
3097                 goto out_put;
3098         }
3099
3100         /*
3101          * We want to set the generation to 0, that way if anything goes wrong
3102          * from here on out we know not to trust this cache when we load up next
3103          * time.
3104          */
3105         BTRFS_I(inode)->generation = 0;
3106         ret = btrfs_update_inode(trans, root, inode);
3107         WARN_ON(ret);
3108
3109         if (i_size_read(inode) > 0) {
3110                 ret = btrfs_check_trunc_cache_free_space(root,
3111                                         &root->fs_info->global_block_rsv);
3112                 if (ret)
3113                         goto out_put;
3114
3115                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3116                                                       inode);
3117                 if (ret)
3118                         goto out_put;
3119         }
3120
3121         spin_lock(&block_group->lock);
3122         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3123             !btrfs_test_opt(root, SPACE_CACHE)) {
3124                 /*
3125                  * don't bother trying to write stuff out _if_
3126                  * a) we're not cached,
3127                  * b) we're with nospace_cache mount option.
3128                  */
3129                 dcs = BTRFS_DC_WRITTEN;
3130                 spin_unlock(&block_group->lock);
3131                 goto out_put;
3132         }
3133         spin_unlock(&block_group->lock);
3134
3135         /*
3136          * Try to preallocate enough space based on how big the block group is.
3137          * Keep in mind this has to include any pinned space which could end up
3138          * taking up quite a bit since it's not folded into the other space
3139          * cache.
3140          */
3141         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3142         if (!num_pages)
3143                 num_pages = 1;
3144
3145         num_pages *= 16;
3146         num_pages *= PAGE_CACHE_SIZE;
3147
3148         ret = btrfs_check_data_free_space(inode, num_pages);
3149         if (ret)
3150                 goto out_put;
3151
3152         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3153                                               num_pages, num_pages,
3154                                               &alloc_hint);
3155         if (!ret)
3156                 dcs = BTRFS_DC_SETUP;
3157         btrfs_free_reserved_data_space(inode, num_pages);
3158
3159 out_put:
3160         iput(inode);
3161 out_free:
3162         btrfs_release_path(path);
3163 out:
3164         spin_lock(&block_group->lock);
3165         if (!ret && dcs == BTRFS_DC_SETUP)
3166                 block_group->cache_generation = trans->transid;
3167         block_group->disk_cache_state = dcs;
3168         spin_unlock(&block_group->lock);
3169
3170         return ret;
3171 }
3172
3173 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3174                                    struct btrfs_root *root)
3175 {
3176         struct btrfs_block_group_cache *cache;
3177         int err = 0;
3178         struct btrfs_path *path;
3179         u64 last = 0;
3180
3181         path = btrfs_alloc_path();
3182         if (!path)
3183                 return -ENOMEM;
3184
3185 again:
3186         while (1) {
3187                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3188                 while (cache) {
3189                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3190                                 break;
3191                         cache = next_block_group(root, cache);
3192                 }
3193                 if (!cache) {
3194                         if (last == 0)
3195                                 break;
3196                         last = 0;
3197                         continue;
3198                 }
3199                 err = cache_save_setup(cache, trans, path);
3200                 last = cache->key.objectid + cache->key.offset;
3201                 btrfs_put_block_group(cache);
3202         }
3203
3204         while (1) {
3205                 if (last == 0) {
3206                         err = btrfs_run_delayed_refs(trans, root,
3207                                                      (unsigned long)-1);
3208                         if (err) /* File system offline */
3209                                 goto out;
3210                 }
3211
3212                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3213                 while (cache) {
3214                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3215                                 btrfs_put_block_group(cache);
3216                                 goto again;
3217                         }
3218
3219                         if (cache->dirty)
3220                                 break;
3221                         cache = next_block_group(root, cache);
3222                 }
3223                 if (!cache) {
3224                         if (last == 0)
3225                                 break;
3226                         last = 0;
3227                         continue;
3228                 }
3229
3230                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3231                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3232                 cache->dirty = 0;
3233                 last = cache->key.objectid + cache->key.offset;
3234
3235                 err = write_one_cache_group(trans, root, path, cache);
3236                 if (err) /* File system offline */
3237                         goto out;
3238
3239                 btrfs_put_block_group(cache);
3240         }
3241
3242         while (1) {
3243                 /*
3244                  * I don't think this is needed since we're just marking our
3245                  * preallocated extent as written, but just in case it can't
3246                  * hurt.
3247                  */
3248                 if (last == 0) {
3249                         err = btrfs_run_delayed_refs(trans, root,
3250                                                      (unsigned long)-1);
3251                         if (err) /* File system offline */
3252                                 goto out;
3253                 }
3254
3255                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3256                 while (cache) {
3257                         /*
3258                          * Really this shouldn't happen, but it could if we
3259                          * couldn't write the entire preallocated extent and
3260                          * splitting the extent resulted in a new block.
3261                          */
3262                         if (cache->dirty) {
3263                                 btrfs_put_block_group(cache);
3264                                 goto again;
3265                         }
3266                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3267                                 break;
3268                         cache = next_block_group(root, cache);
3269                 }
3270                 if (!cache) {
3271                         if (last == 0)
3272                                 break;
3273                         last = 0;
3274                         continue;
3275                 }
3276
3277                 err = btrfs_write_out_cache(root, trans, cache, path);
3278
3279                 /*
3280                  * If we didn't have an error then the cache state is still
3281                  * NEED_WRITE, so we can set it to WRITTEN.
3282                  */
3283                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3284                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3285                 last = cache->key.objectid + cache->key.offset;
3286                 btrfs_put_block_group(cache);
3287         }
3288 out:
3289
3290         btrfs_free_path(path);
3291         return err;
3292 }
3293
3294 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3295 {
3296         struct btrfs_block_group_cache *block_group;
3297         int readonly = 0;
3298
3299         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3300         if (!block_group || block_group->ro)
3301                 readonly = 1;
3302         if (block_group)
3303                 btrfs_put_block_group(block_group);
3304         return readonly;
3305 }
3306
3307 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3308                              u64 total_bytes, u64 bytes_used,
3309                              struct btrfs_space_info **space_info)
3310 {
3311         struct btrfs_space_info *found;
3312         int i;
3313         int factor;
3314
3315         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3316                      BTRFS_BLOCK_GROUP_RAID10))
3317                 factor = 2;
3318         else
3319                 factor = 1;
3320
3321         found = __find_space_info(info, flags);
3322         if (found) {
3323                 spin_lock(&found->lock);
3324                 found->total_bytes += total_bytes;
3325                 found->disk_total += total_bytes * factor;
3326                 found->bytes_used += bytes_used;
3327                 found->disk_used += bytes_used * factor;
3328                 found->full = 0;
3329                 spin_unlock(&found->lock);
3330                 *space_info = found;
3331                 return 0;
3332         }
3333         found = kzalloc(sizeof(*found), GFP_NOFS);
3334         if (!found)
3335                 return -ENOMEM;
3336
3337         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3338                 INIT_LIST_HEAD(&found->block_groups[i]);
3339         init_rwsem(&found->groups_sem);
3340         spin_lock_init(&found->lock);
3341         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3342         found->total_bytes = total_bytes;
3343         found->disk_total = total_bytes * factor;
3344         found->bytes_used = bytes_used;
3345         found->disk_used = bytes_used * factor;
3346         found->bytes_pinned = 0;
3347         found->bytes_reserved = 0;
3348         found->bytes_readonly = 0;
3349         found->bytes_may_use = 0;
3350         found->full = 0;
3351         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3352         found->chunk_alloc = 0;
3353         found->flush = 0;
3354         init_waitqueue_head(&found->wait);
3355         *space_info = found;
3356         list_add_rcu(&found->list, &info->space_info);
3357         if (flags & BTRFS_BLOCK_GROUP_DATA)
3358                 info->data_sinfo = found;
3359         return 0;
3360 }
3361
3362 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3363 {
3364         u64 extra_flags = chunk_to_extended(flags) &
3365                                 BTRFS_EXTENDED_PROFILE_MASK;
3366
3367         write_seqlock(&fs_info->profiles_lock);
3368         if (flags & BTRFS_BLOCK_GROUP_DATA)
3369                 fs_info->avail_data_alloc_bits |= extra_flags;
3370         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3371                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3372         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3373                 fs_info->avail_system_alloc_bits |= extra_flags;
3374         write_sequnlock(&fs_info->profiles_lock);
3375 }
3376
3377 /*
3378  * returns target flags in extended format or 0 if restripe for this
3379  * chunk_type is not in progress
3380  *
3381  * should be called with either volume_mutex or balance_lock held
3382  */
3383 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3384 {
3385         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3386         u64 target = 0;
3387
3388         if (!bctl)
3389                 return 0;
3390
3391         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3392             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3393                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3394         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3395                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3396                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3397         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3398                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3399                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3400         }
3401
3402         return target;
3403 }
3404
3405 /*
3406  * @flags: available profiles in extended format (see ctree.h)
3407  *
3408  * Returns reduced profile in chunk format.  If profile changing is in
3409  * progress (either running or paused) picks the target profile (if it's
3410  * already available), otherwise falls back to plain reducing.
3411  */
3412 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3413 {
3414         /*
3415          * we add in the count of missing devices because we want
3416          * to make sure that any RAID levels on a degraded FS
3417          * continue to be honored.
3418          */
3419         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3420                 root->fs_info->fs_devices->missing_devices;
3421         u64 target;
3422         u64 tmp;
3423
3424         /*
3425          * see if restripe for this chunk_type is in progress, if so
3426          * try to reduce to the target profile
3427          */
3428         spin_lock(&root->fs_info->balance_lock);
3429         target = get_restripe_target(root->fs_info, flags);
3430         if (target) {
3431                 /* pick target profile only if it's already available */
3432                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3433                         spin_unlock(&root->fs_info->balance_lock);
3434                         return extended_to_chunk(target);
3435                 }
3436         }
3437         spin_unlock(&root->fs_info->balance_lock);
3438
3439         /* First, mask out the RAID levels which aren't possible */
3440         if (num_devices == 1)
3441                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3442                            BTRFS_BLOCK_GROUP_RAID5);
3443         if (num_devices < 3)
3444                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3445         if (num_devices < 4)
3446                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3447
3448         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3449                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3450                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3451         flags &= ~tmp;
3452
3453         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3454                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3455         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3456                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3457         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3458                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3459         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3460                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3461         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3462                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3463
3464         return extended_to_chunk(flags | tmp);
3465 }
3466
3467 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3468 {
3469         unsigned seq;
3470
3471         do {
3472                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3473
3474                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3475                         flags |= root->fs_info->avail_data_alloc_bits;
3476                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3477                         flags |= root->fs_info->avail_system_alloc_bits;
3478                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3479                         flags |= root->fs_info->avail_metadata_alloc_bits;
3480         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3481
3482         return btrfs_reduce_alloc_profile(root, flags);
3483 }
3484
3485 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3486 {
3487         u64 flags;
3488         u64 ret;
3489
3490         if (data)
3491                 flags = BTRFS_BLOCK_GROUP_DATA;
3492         else if (root == root->fs_info->chunk_root)
3493                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3494         else
3495                 flags = BTRFS_BLOCK_GROUP_METADATA;
3496
3497         ret = get_alloc_profile(root, flags);
3498         return ret;
3499 }
3500
3501 /*
3502  * This will check the space that the inode allocates from to make sure we have
3503  * enough space for bytes.
3504  */
3505 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3506 {
3507         struct btrfs_space_info *data_sinfo;
3508         struct btrfs_root *root = BTRFS_I(inode)->root;
3509         struct btrfs_fs_info *fs_info = root->fs_info;
3510         u64 used;
3511         int ret = 0, committed = 0, alloc_chunk = 1;
3512
3513         /* make sure bytes are sectorsize aligned */
3514         bytes = ALIGN(bytes, root->sectorsize);
3515
3516         if (root == root->fs_info->tree_root ||
3517             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3518                 alloc_chunk = 0;
3519                 committed = 1;
3520         }
3521
3522         data_sinfo = fs_info->data_sinfo;
3523         if (!data_sinfo)
3524                 goto alloc;
3525
3526 again:
3527         /* make sure we have enough space to handle the data first */
3528         spin_lock(&data_sinfo->lock);
3529         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3530                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3531                 data_sinfo->bytes_may_use;
3532
3533         if (used + bytes > data_sinfo->total_bytes) {
3534                 struct btrfs_trans_handle *trans;
3535
3536                 /*
3537                  * if we don't have enough free bytes in this space then we need
3538                  * to alloc a new chunk.
3539                  */
3540                 if (!data_sinfo->full && alloc_chunk) {
3541                         u64 alloc_target;
3542
3543                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3544                         spin_unlock(&data_sinfo->lock);
3545 alloc:
3546                         alloc_target = btrfs_get_alloc_profile(root, 1);
3547                         trans = btrfs_join_transaction(root);
3548                         if (IS_ERR(trans))
3549                                 return PTR_ERR(trans);
3550
3551                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3552                                              alloc_target,
3553                                              CHUNK_ALLOC_NO_FORCE);
3554                         btrfs_end_transaction(trans, root);
3555                         if (ret < 0) {
3556                                 if (ret != -ENOSPC)
3557                                         return ret;
3558                                 else
3559                                         goto commit_trans;
3560                         }
3561
3562                         if (!data_sinfo)
3563                                 data_sinfo = fs_info->data_sinfo;
3564
3565                         goto again;
3566                 }
3567
3568                 /*
3569                  * If we have less pinned bytes than we want to allocate then
3570                  * don't bother committing the transaction, it won't help us.
3571                  */
3572                 if (data_sinfo->bytes_pinned < bytes)
3573                         committed = 1;
3574                 spin_unlock(&data_sinfo->lock);
3575
3576                 /* commit the current transaction and try again */
3577 commit_trans:
3578                 if (!committed &&
3579                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3580                         committed = 1;
3581                         trans = btrfs_join_transaction(root);
3582                         if (IS_ERR(trans))
3583                                 return PTR_ERR(trans);
3584                         ret = btrfs_commit_transaction(trans, root);
3585                         if (ret)
3586                                 return ret;
3587                         goto again;
3588                 }
3589
3590                 return -ENOSPC;
3591         }
3592         data_sinfo->bytes_may_use += bytes;
3593         trace_btrfs_space_reservation(root->fs_info, "space_info",
3594                                       data_sinfo->flags, bytes, 1);
3595         spin_unlock(&data_sinfo->lock);
3596
3597         return 0;
3598 }
3599
3600 /*
3601  * Called if we need to clear a data reservation for this inode.
3602  */
3603 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3604 {
3605         struct btrfs_root *root = BTRFS_I(inode)->root;
3606         struct btrfs_space_info *data_sinfo;
3607
3608         /* make sure bytes are sectorsize aligned */
3609         bytes = ALIGN(bytes, root->sectorsize);
3610
3611         data_sinfo = root->fs_info->data_sinfo;
3612         spin_lock(&data_sinfo->lock);
3613         data_sinfo->bytes_may_use -= bytes;
3614         trace_btrfs_space_reservation(root->fs_info, "space_info",
3615                                       data_sinfo->flags, bytes, 0);
3616         spin_unlock(&data_sinfo->lock);
3617 }
3618
3619 static void force_metadata_allocation(struct btrfs_fs_info *info)
3620 {
3621         struct list_head *head = &info->space_info;
3622         struct btrfs_space_info *found;
3623
3624         rcu_read_lock();
3625         list_for_each_entry_rcu(found, head, list) {
3626                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3627                         found->force_alloc = CHUNK_ALLOC_FORCE;
3628         }
3629         rcu_read_unlock();
3630 }
3631
3632 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3633 {
3634         return (global->size << 1);
3635 }
3636
3637 static int should_alloc_chunk(struct btrfs_root *root,
3638                               struct btrfs_space_info *sinfo, int force)
3639 {
3640         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3641         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3642         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3643         u64 thresh;
3644
3645         if (force == CHUNK_ALLOC_FORCE)
3646                 return 1;
3647
3648         /*
3649          * We need to take into account the global rsv because for all intents
3650          * and purposes it's used space.  Don't worry about locking the
3651          * global_rsv, it doesn't change except when the transaction commits.
3652          */
3653         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3654                 num_allocated += calc_global_rsv_need_space(global_rsv);
3655
3656         /*
3657          * in limited mode, we want to have some free space up to
3658          * about 1% of the FS size.
3659          */
3660         if (force == CHUNK_ALLOC_LIMITED) {
3661                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3662                 thresh = max_t(u64, 64 * 1024 * 1024,
3663                                div_factor_fine(thresh, 1));
3664
3665                 if (num_bytes - num_allocated < thresh)
3666                         return 1;
3667         }
3668
3669         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3670                 return 0;
3671         return 1;
3672 }
3673
3674 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3675 {
3676         u64 num_dev;
3677
3678         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3679                     BTRFS_BLOCK_GROUP_RAID0 |
3680                     BTRFS_BLOCK_GROUP_RAID5 |
3681                     BTRFS_BLOCK_GROUP_RAID6))
3682                 num_dev = root->fs_info->fs_devices->rw_devices;
3683         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3684                 num_dev = 2;
3685         else
3686                 num_dev = 1;    /* DUP or single */
3687
3688         /* metadata for updaing devices and chunk tree */
3689         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3690 }
3691
3692 static void check_system_chunk(struct btrfs_trans_handle *trans,
3693                                struct btrfs_root *root, u64 type)
3694 {
3695         struct btrfs_space_info *info;
3696         u64 left;
3697         u64 thresh;
3698
3699         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3700         spin_lock(&info->lock);
3701         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3702                 info->bytes_reserved - info->bytes_readonly;
3703         spin_unlock(&info->lock);
3704
3705         thresh = get_system_chunk_thresh(root, type);
3706         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3707                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3708                         left, thresh, type);
3709                 dump_space_info(info, 0, 0);
3710         }
3711
3712         if (left < thresh) {
3713                 u64 flags;
3714
3715                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3716                 btrfs_alloc_chunk(trans, root, flags);
3717         }
3718 }
3719
3720 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3721                           struct btrfs_root *extent_root, u64 flags, int force)
3722 {
3723         struct btrfs_space_info *space_info;
3724         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3725         int wait_for_alloc = 0;
3726         int ret = 0;
3727
3728         /* Don't re-enter if we're already allocating a chunk */
3729         if (trans->allocating_chunk)
3730                 return -ENOSPC;
3731
3732         space_info = __find_space_info(extent_root->fs_info, flags);
3733         if (!space_info) {
3734                 ret = update_space_info(extent_root->fs_info, flags,
3735                                         0, 0, &space_info);
3736                 BUG_ON(ret); /* -ENOMEM */
3737         }
3738         BUG_ON(!space_info); /* Logic error */
3739
3740 again:
3741         spin_lock(&space_info->lock);
3742         if (force < space_info->force_alloc)
3743                 force = space_info->force_alloc;
3744         if (space_info->full) {
3745                 spin_unlock(&space_info->lock);
3746                 return 0;
3747         }
3748
3749         if (!should_alloc_chunk(extent_root, space_info, force)) {
3750                 spin_unlock(&space_info->lock);
3751                 return 0;
3752         } else if (space_info->chunk_alloc) {
3753                 wait_for_alloc = 1;
3754         } else {
3755                 space_info->chunk_alloc = 1;
3756         }
3757
3758         spin_unlock(&space_info->lock);
3759
3760         mutex_lock(&fs_info->chunk_mutex);
3761
3762         /*
3763          * The chunk_mutex is held throughout the entirety of a chunk
3764          * allocation, so once we've acquired the chunk_mutex we know that the
3765          * other guy is done and we need to recheck and see if we should
3766          * allocate.
3767          */
3768         if (wait_for_alloc) {
3769                 mutex_unlock(&fs_info->chunk_mutex);
3770                 wait_for_alloc = 0;
3771                 goto again;
3772         }
3773
3774         trans->allocating_chunk = true;
3775
3776         /*
3777          * If we have mixed data/metadata chunks we want to make sure we keep
3778          * allocating mixed chunks instead of individual chunks.
3779          */
3780         if (btrfs_mixed_space_info(space_info))
3781                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3782
3783         /*
3784          * if we're doing a data chunk, go ahead and make sure that
3785          * we keep a reasonable number of metadata chunks allocated in the
3786          * FS as well.
3787          */
3788         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3789                 fs_info->data_chunk_allocations++;
3790                 if (!(fs_info->data_chunk_allocations %
3791                       fs_info->metadata_ratio))
3792                         force_metadata_allocation(fs_info);
3793         }
3794
3795         /*
3796          * Check if we have enough space in SYSTEM chunk because we may need
3797          * to update devices.
3798          */
3799         check_system_chunk(trans, extent_root, flags);
3800
3801         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3802         trans->allocating_chunk = false;
3803
3804         spin_lock(&space_info->lock);
3805         if (ret < 0 && ret != -ENOSPC)
3806                 goto out;
3807         if (ret)
3808                 space_info->full = 1;
3809         else
3810                 ret = 1;
3811
3812         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3813 out:
3814         space_info->chunk_alloc = 0;
3815         spin_unlock(&space_info->lock);
3816         mutex_unlock(&fs_info->chunk_mutex);
3817         return ret;
3818 }
3819
3820 static int can_overcommit(struct btrfs_root *root,
3821                           struct btrfs_space_info *space_info, u64 bytes,
3822                           enum btrfs_reserve_flush_enum flush)
3823 {
3824         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3825         u64 profile = btrfs_get_alloc_profile(root, 0);
3826         u64 space_size;
3827         u64 avail;
3828         u64 used;
3829         u64 to_add;
3830
3831         used = space_info->bytes_used + space_info->bytes_reserved +
3832                 space_info->bytes_pinned + space_info->bytes_readonly;
3833
3834         /*
3835          * We only want to allow over committing if we have lots of actual space
3836          * free, but if we don't have enough space to handle the global reserve
3837          * space then we could end up having a real enospc problem when trying
3838          * to allocate a chunk or some other such important allocation.
3839          */
3840         spin_lock(&global_rsv->lock);
3841         space_size = calc_global_rsv_need_space(global_rsv);
3842         spin_unlock(&global_rsv->lock);
3843         if (used + space_size >= space_info->total_bytes)
3844                 return 0;
3845
3846         used += space_info->bytes_may_use;
3847
3848         spin_lock(&root->fs_info->free_chunk_lock);
3849         avail = root->fs_info->free_chunk_space;
3850         spin_unlock(&root->fs_info->free_chunk_lock);
3851
3852         /*
3853          * If we have dup, raid1 or raid10 then only half of the free
3854          * space is actually useable.  For raid56, the space info used
3855          * doesn't include the parity drive, so we don't have to
3856          * change the math
3857          */
3858         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3859                        BTRFS_BLOCK_GROUP_RAID1 |
3860                        BTRFS_BLOCK_GROUP_RAID10))
3861                 avail >>= 1;
3862
3863         to_add = space_info->total_bytes;
3864
3865         /*
3866          * If we aren't flushing all things, let us overcommit up to
3867          * 1/2th of the space. If we can flush, don't let us overcommit
3868          * too much, let it overcommit up to 1/8 of the space.
3869          */
3870         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3871                 to_add >>= 3;
3872         else
3873                 to_add >>= 1;
3874
3875         /*
3876          * Limit the overcommit to the amount of free space we could possibly
3877          * allocate for chunks.
3878          */
3879         to_add = min(avail, to_add);
3880
3881         if (used + bytes < space_info->total_bytes + to_add)
3882                 return 1;
3883         return 0;
3884 }
3885
3886 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3887                                          unsigned long nr_pages)
3888 {
3889         struct super_block *sb = root->fs_info->sb;
3890         int started;
3891
3892         /* If we can not start writeback, just sync all the delalloc file. */
3893         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3894                                                       WB_REASON_FS_FREE_SPACE);
3895         if (!started) {
3896                 /*
3897                  * We needn't worry the filesystem going from r/w to r/o though
3898                  * we don't acquire ->s_umount mutex, because the filesystem
3899                  * should guarantee the delalloc inodes list be empty after
3900                  * the filesystem is readonly(all dirty pages are written to
3901                  * the disk).
3902                  */
3903                 btrfs_start_all_delalloc_inodes(root->fs_info, 0);
3904                 if (!current->journal_info)
3905                         btrfs_wait_all_ordered_extents(root->fs_info, 0);
3906         }
3907 }
3908
3909 /*
3910  * shrink metadata reservation for delalloc
3911  */
3912 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3913                             bool wait_ordered)
3914 {
3915         struct btrfs_block_rsv *block_rsv;
3916         struct btrfs_space_info *space_info;
3917         struct btrfs_trans_handle *trans;
3918         u64 delalloc_bytes;
3919         u64 max_reclaim;
3920         long time_left;
3921         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3922         int loops = 0;
3923         enum btrfs_reserve_flush_enum flush;
3924
3925         trans = (struct btrfs_trans_handle *)current->journal_info;
3926         block_rsv = &root->fs_info->delalloc_block_rsv;
3927         space_info = block_rsv->space_info;
3928
3929         smp_mb();
3930         delalloc_bytes = percpu_counter_sum_positive(
3931                                                 &root->fs_info->delalloc_bytes);
3932         if (delalloc_bytes == 0) {
3933                 if (trans)
3934                         return;
3935                 btrfs_wait_all_ordered_extents(root->fs_info, 0);
3936                 return;
3937         }
3938
3939         while (delalloc_bytes && loops < 3) {
3940                 max_reclaim = min(delalloc_bytes, to_reclaim);
3941                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3942                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3943                 /*
3944                  * We need to wait for the async pages to actually start before
3945                  * we do anything.
3946                  */
3947                 wait_event(root->fs_info->async_submit_wait,
3948                            !atomic_read(&root->fs_info->async_delalloc_pages));
3949
3950                 if (!trans)
3951                         flush = BTRFS_RESERVE_FLUSH_ALL;
3952                 else
3953                         flush = BTRFS_RESERVE_NO_FLUSH;
3954                 spin_lock(&space_info->lock);
3955                 if (can_overcommit(root, space_info, orig, flush)) {
3956                         spin_unlock(&space_info->lock);
3957                         break;
3958                 }
3959                 spin_unlock(&space_info->lock);
3960
3961                 loops++;
3962                 if (wait_ordered && !trans) {
3963                         btrfs_wait_all_ordered_extents(root->fs_info, 0);
3964                 } else {
3965                         time_left = schedule_timeout_killable(1);
3966                         if (time_left)
3967                                 break;
3968                 }
3969                 smp_mb();
3970                 delalloc_bytes = percpu_counter_sum_positive(
3971                                                 &root->fs_info->delalloc_bytes);
3972         }
3973 }
3974
3975 /**
3976  * maybe_commit_transaction - possibly commit the transaction if its ok to
3977  * @root - the root we're allocating for
3978  * @bytes - the number of bytes we want to reserve
3979  * @force - force the commit
3980  *
3981  * This will check to make sure that committing the transaction will actually
3982  * get us somewhere and then commit the transaction if it does.  Otherwise it
3983  * will return -ENOSPC.
3984  */
3985 static int may_commit_transaction(struct btrfs_root *root,
3986                                   struct btrfs_space_info *space_info,
3987                                   u64 bytes, int force)
3988 {
3989         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3990         struct btrfs_trans_handle *trans;
3991
3992         trans = (struct btrfs_trans_handle *)current->journal_info;
3993         if (trans)
3994                 return -EAGAIN;
3995
3996         if (force)
3997                 goto commit;
3998
3999         /* See if there is enough pinned space to make this reservation */
4000         spin_lock(&space_info->lock);
4001         if (space_info->bytes_pinned >= bytes) {
4002                 spin_unlock(&space_info->lock);
4003                 goto commit;
4004         }
4005         spin_unlock(&space_info->lock);
4006
4007         /*
4008          * See if there is some space in the delayed insertion reservation for
4009          * this reservation.
4010          */
4011         if (space_info != delayed_rsv->space_info)
4012                 return -ENOSPC;
4013
4014         spin_lock(&space_info->lock);
4015         spin_lock(&delayed_rsv->lock);
4016         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
4017                 spin_unlock(&delayed_rsv->lock);
4018                 spin_unlock(&space_info->lock);
4019                 return -ENOSPC;
4020         }
4021         spin_unlock(&delayed_rsv->lock);
4022         spin_unlock(&space_info->lock);
4023
4024 commit:
4025         trans = btrfs_join_transaction(root);
4026         if (IS_ERR(trans))
4027                 return -ENOSPC;
4028
4029         return btrfs_commit_transaction(trans, root);
4030 }
4031
4032 enum flush_state {
4033         FLUSH_DELAYED_ITEMS_NR  =       1,
4034         FLUSH_DELAYED_ITEMS     =       2,
4035         FLUSH_DELALLOC          =       3,
4036         FLUSH_DELALLOC_WAIT     =       4,
4037         ALLOC_CHUNK             =       5,
4038         COMMIT_TRANS            =       6,
4039 };
4040
4041 static int flush_space(struct btrfs_root *root,
4042                        struct btrfs_space_info *space_info, u64 num_bytes,
4043                        u64 orig_bytes, int state)
4044 {
4045         struct btrfs_trans_handle *trans;
4046         int nr;
4047         int ret = 0;
4048
4049         switch (state) {
4050         case FLUSH_DELAYED_ITEMS_NR:
4051         case FLUSH_DELAYED_ITEMS:
4052                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4053                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4054
4055                         nr = (int)div64_u64(num_bytes, bytes);
4056                         if (!nr)
4057                                 nr = 1;
4058                         nr *= 2;
4059                 } else {
4060                         nr = -1;
4061                 }
4062                 trans = btrfs_join_transaction(root);
4063                 if (IS_ERR(trans)) {
4064                         ret = PTR_ERR(trans);
4065                         break;
4066                 }
4067                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4068                 btrfs_end_transaction(trans, root);
4069                 break;
4070         case FLUSH_DELALLOC:
4071         case FLUSH_DELALLOC_WAIT:
4072                 shrink_delalloc(root, num_bytes, orig_bytes,
4073                                 state == FLUSH_DELALLOC_WAIT);
4074                 break;
4075         case ALLOC_CHUNK:
4076                 trans = btrfs_join_transaction(root);
4077                 if (IS_ERR(trans)) {
4078                         ret = PTR_ERR(trans);
4079                         break;
4080                 }
4081                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4082                                      btrfs_get_alloc_profile(root, 0),
4083                                      CHUNK_ALLOC_NO_FORCE);
4084                 btrfs_end_transaction(trans, root);
4085                 if (ret == -ENOSPC)
4086                         ret = 0;
4087                 break;
4088         case COMMIT_TRANS:
4089                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4090                 break;
4091         default:
4092                 ret = -ENOSPC;
4093                 break;
4094         }
4095
4096         return ret;
4097 }
4098 /**
4099  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4100  * @root - the root we're allocating for
4101  * @block_rsv - the block_rsv we're allocating for
4102  * @orig_bytes - the number of bytes we want
4103  * @flush - whether or not we can flush to make our reservation
4104  *
4105  * This will reserve orgi_bytes number of bytes from the space info associated
4106  * with the block_rsv.  If there is not enough space it will make an attempt to
4107  * flush out space to make room.  It will do this by flushing delalloc if
4108  * possible or committing the transaction.  If flush is 0 then no attempts to
4109  * regain reservations will be made and this will fail if there is not enough
4110  * space already.
4111  */
4112 static int reserve_metadata_bytes(struct btrfs_root *root,
4113                                   struct btrfs_block_rsv *block_rsv,
4114                                   u64 orig_bytes,
4115                                   enum btrfs_reserve_flush_enum flush)
4116 {
4117         struct btrfs_space_info *space_info = block_rsv->space_info;
4118         u64 used;
4119         u64 num_bytes = orig_bytes;
4120         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4121         int ret = 0;
4122         bool flushing = false;
4123
4124 again:
4125         ret = 0;
4126         spin_lock(&space_info->lock);
4127         /*
4128          * We only want to wait if somebody other than us is flushing and we
4129          * are actually allowed to flush all things.
4130          */
4131         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4132                space_info->flush) {
4133                 spin_unlock(&space_info->lock);
4134                 /*
4135                  * If we have a trans handle we can't wait because the flusher
4136                  * may have to commit the transaction, which would mean we would
4137                  * deadlock since we are waiting for the flusher to finish, but
4138                  * hold the current transaction open.
4139                  */
4140                 if (current->journal_info)
4141                         return -EAGAIN;
4142                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4143                 /* Must have been killed, return */
4144                 if (ret)
4145                         return -EINTR;
4146
4147                 spin_lock(&space_info->lock);
4148         }
4149
4150         ret = -ENOSPC;
4151         used = space_info->bytes_used + space_info->bytes_reserved +
4152                 space_info->bytes_pinned + space_info->bytes_readonly +
4153                 space_info->bytes_may_use;
4154
4155         /*
4156          * The idea here is that we've not already over-reserved the block group
4157          * then we can go ahead and save our reservation first and then start
4158          * flushing if we need to.  Otherwise if we've already overcommitted
4159          * lets start flushing stuff first and then come back and try to make
4160          * our reservation.
4161          */
4162         if (used <= space_info->total_bytes) {
4163                 if (used + orig_bytes <= space_info->total_bytes) {
4164                         space_info->bytes_may_use += orig_bytes;
4165                         trace_btrfs_space_reservation(root->fs_info,
4166                                 "space_info", space_info->flags, orig_bytes, 1);
4167                         ret = 0;
4168                 } else {
4169                         /*
4170                          * Ok set num_bytes to orig_bytes since we aren't
4171                          * overocmmitted, this way we only try and reclaim what
4172                          * we need.
4173                          */
4174                         num_bytes = orig_bytes;
4175                 }
4176         } else {
4177                 /*
4178                  * Ok we're over committed, set num_bytes to the overcommitted
4179                  * amount plus the amount of bytes that we need for this
4180                  * reservation.
4181                  */
4182                 num_bytes = used - space_info->total_bytes +
4183                         (orig_bytes * 2);
4184         }
4185
4186         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4187                 space_info->bytes_may_use += orig_bytes;
4188                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4189                                               space_info->flags, orig_bytes,
4190                                               1);
4191                 ret = 0;
4192         }
4193
4194         /*
4195          * Couldn't make our reservation, save our place so while we're trying
4196          * to reclaim space we can actually use it instead of somebody else
4197          * stealing it from us.
4198          *
4199          * We make the other tasks wait for the flush only when we can flush
4200          * all things.
4201          */
4202         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4203                 flushing = true;
4204                 space_info->flush = 1;
4205         }
4206
4207         spin_unlock(&space_info->lock);
4208
4209         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4210                 goto out;
4211
4212         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4213                           flush_state);
4214         flush_state++;
4215
4216         /*
4217          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4218          * would happen. So skip delalloc flush.
4219          */
4220         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4221             (flush_state == FLUSH_DELALLOC ||
4222              flush_state == FLUSH_DELALLOC_WAIT))
4223                 flush_state = ALLOC_CHUNK;
4224
4225         if (!ret)
4226                 goto again;
4227         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4228                  flush_state < COMMIT_TRANS)
4229                 goto again;
4230         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4231                  flush_state <= COMMIT_TRANS)
4232                 goto again;
4233
4234 out:
4235         if (ret == -ENOSPC &&
4236             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4237                 struct btrfs_block_rsv *global_rsv =
4238                         &root->fs_info->global_block_rsv;
4239
4240                 if (block_rsv != global_rsv &&
4241                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4242                         ret = 0;
4243         }
4244         if (flushing) {
4245                 spin_lock(&space_info->lock);
4246                 space_info->flush = 0;
4247                 wake_up_all(&space_info->wait);
4248                 spin_unlock(&space_info->lock);
4249         }
4250         return ret;
4251 }
4252
4253 static struct btrfs_block_rsv *get_block_rsv(
4254                                         const struct btrfs_trans_handle *trans,
4255                                         const struct btrfs_root *root)
4256 {
4257         struct btrfs_block_rsv *block_rsv = NULL;
4258
4259         if (root->ref_cows)
4260                 block_rsv = trans->block_rsv;
4261
4262         if (root == root->fs_info->csum_root && trans->adding_csums)
4263                 block_rsv = trans->block_rsv;
4264
4265         if (!block_rsv)
4266                 block_rsv = root->block_rsv;
4267
4268         if (!block_rsv)
4269                 block_rsv = &root->fs_info->empty_block_rsv;
4270
4271         return block_rsv;
4272 }
4273
4274 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4275                                u64 num_bytes)
4276 {
4277         int ret = -ENOSPC;
4278         spin_lock(&block_rsv->lock);
4279         if (block_rsv->reserved >= num_bytes) {
4280                 block_rsv->reserved -= num_bytes;
4281                 if (block_rsv->reserved < block_rsv->size)
4282                         block_rsv->full = 0;
4283                 ret = 0;
4284         }
4285         spin_unlock(&block_rsv->lock);
4286         return ret;
4287 }
4288
4289 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4290                                 u64 num_bytes, int update_size)
4291 {
4292         spin_lock(&block_rsv->lock);
4293         block_rsv->reserved += num_bytes;
4294         if (update_size)
4295                 block_rsv->size += num_bytes;
4296         else if (block_rsv->reserved >= block_rsv->size)
4297                 block_rsv->full = 1;
4298         spin_unlock(&block_rsv->lock);
4299 }
4300
4301 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4302                              struct btrfs_block_rsv *dest, u64 num_bytes,
4303                              int min_factor)
4304 {
4305         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4306         u64 min_bytes;
4307
4308         if (global_rsv->space_info != dest->space_info)
4309                 return -ENOSPC;
4310
4311         spin_lock(&global_rsv->lock);
4312         min_bytes = div_factor(global_rsv->size, min_factor);
4313         if (global_rsv->reserved < min_bytes + num_bytes) {
4314                 spin_unlock(&global_rsv->lock);
4315                 return -ENOSPC;
4316         }
4317         global_rsv->reserved -= num_bytes;
4318         if (global_rsv->reserved < global_rsv->size)
4319                 global_rsv->full = 0;
4320         spin_unlock(&global_rsv->lock);
4321
4322         block_rsv_add_bytes(dest, num_bytes, 1);
4323         return 0;
4324 }
4325
4326 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4327                                     struct btrfs_block_rsv *block_rsv,
4328                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4329 {
4330         struct btrfs_space_info *space_info = block_rsv->space_info;
4331
4332         spin_lock(&block_rsv->lock);
4333         if (num_bytes == (u64)-1)
4334                 num_bytes = block_rsv->size;
4335         block_rsv->size -= num_bytes;
4336         if (block_rsv->reserved >= block_rsv->size) {
4337                 num_bytes = block_rsv->reserved - block_rsv->size;
4338                 block_rsv->reserved = block_rsv->size;
4339                 block_rsv->full = 1;
4340         } else {
4341                 num_bytes = 0;
4342         }
4343         spin_unlock(&block_rsv->lock);
4344
4345         if (num_bytes > 0) {
4346                 if (dest) {
4347                         spin_lock(&dest->lock);
4348                         if (!dest->full) {
4349                                 u64 bytes_to_add;
4350
4351                                 bytes_to_add = dest->size - dest->reserved;
4352                                 bytes_to_add = min(num_bytes, bytes_to_add);
4353                                 dest->reserved += bytes_to_add;
4354                                 if (dest->reserved >= dest->size)
4355                                         dest->full = 1;
4356                                 num_bytes -= bytes_to_add;
4357                         }
4358                         spin_unlock(&dest->lock);
4359                 }
4360                 if (num_bytes) {
4361                         spin_lock(&space_info->lock);
4362                         space_info->bytes_may_use -= num_bytes;
4363                         trace_btrfs_space_reservation(fs_info, "space_info",
4364                                         space_info->flags, num_bytes, 0);
4365                         space_info->reservation_progress++;
4366                         spin_unlock(&space_info->lock);
4367                 }
4368         }
4369 }
4370
4371 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4372                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4373 {
4374         int ret;
4375
4376         ret = block_rsv_use_bytes(src, num_bytes);
4377         if (ret)
4378                 return ret;
4379
4380         block_rsv_add_bytes(dst, num_bytes, 1);
4381         return 0;
4382 }
4383
4384 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4385 {
4386         memset(rsv, 0, sizeof(*rsv));
4387         spin_lock_init(&rsv->lock);
4388         rsv->type = type;
4389 }
4390
4391 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4392                                               unsigned short type)
4393 {
4394         struct btrfs_block_rsv *block_rsv;
4395         struct btrfs_fs_info *fs_info = root->fs_info;
4396
4397         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4398         if (!block_rsv)
4399                 return NULL;
4400
4401         btrfs_init_block_rsv(block_rsv, type);
4402         block_rsv->space_info = __find_space_info(fs_info,
4403                                                   BTRFS_BLOCK_GROUP_METADATA);
4404         return block_rsv;
4405 }
4406
4407 void btrfs_free_block_rsv(struct btrfs_root *root,
4408                           struct btrfs_block_rsv *rsv)
4409 {
4410         if (!rsv)
4411                 return;
4412         btrfs_block_rsv_release(root, rsv, (u64)-1);
4413         kfree(rsv);
4414 }
4415
4416 int btrfs_block_rsv_add(struct btrfs_root *root,
4417                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4418                         enum btrfs_reserve_flush_enum flush)
4419 {
4420         int ret;
4421
4422         if (num_bytes == 0)
4423                 return 0;
4424
4425         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4426         if (!ret) {
4427                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4428                 return 0;
4429         }
4430
4431         return ret;
4432 }
4433
4434 int btrfs_block_rsv_check(struct btrfs_root *root,
4435                           struct btrfs_block_rsv *block_rsv, int min_factor)
4436 {
4437         u64 num_bytes = 0;
4438         int ret = -ENOSPC;
4439
4440         if (!block_rsv)
4441                 return 0;
4442
4443         spin_lock(&block_rsv->lock);
4444         num_bytes = div_factor(block_rsv->size, min_factor);
4445         if (block_rsv->reserved >= num_bytes)
4446                 ret = 0;
4447         spin_unlock(&block_rsv->lock);
4448
4449         return ret;
4450 }
4451
4452 int btrfs_block_rsv_refill(struct btrfs_root *root,
4453                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4454                            enum btrfs_reserve_flush_enum flush)
4455 {
4456         u64 num_bytes = 0;
4457         int ret = -ENOSPC;
4458
4459         if (!block_rsv)
4460                 return 0;
4461
4462         spin_lock(&block_rsv->lock);
4463         num_bytes = min_reserved;
4464         if (block_rsv->reserved >= num_bytes)
4465                 ret = 0;
4466         else
4467                 num_bytes -= block_rsv->reserved;
4468         spin_unlock(&block_rsv->lock);
4469
4470         if (!ret)
4471                 return 0;
4472
4473         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4474         if (!ret) {
4475                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4476                 return 0;
4477         }
4478
4479         return ret;
4480 }
4481
4482 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4483                             struct btrfs_block_rsv *dst_rsv,
4484                             u64 num_bytes)
4485 {
4486         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4487 }
4488
4489 void btrfs_block_rsv_release(struct btrfs_root *root,
4490                              struct btrfs_block_rsv *block_rsv,
4491                              u64 num_bytes)
4492 {
4493         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4494         if (global_rsv->full || global_rsv == block_rsv ||
4495             block_rsv->space_info != global_rsv->space_info)
4496                 global_rsv = NULL;
4497         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4498                                 num_bytes);
4499 }
4500
4501 /*
4502  * helper to calculate size of global block reservation.
4503  * the desired value is sum of space used by extent tree,
4504  * checksum tree and root tree
4505  */
4506 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4507 {
4508         struct btrfs_space_info *sinfo;
4509         u64 num_bytes;
4510         u64 meta_used;
4511         u64 data_used;
4512         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4513
4514         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4515         spin_lock(&sinfo->lock);
4516         data_used = sinfo->bytes_used;
4517         spin_unlock(&sinfo->lock);
4518
4519         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4520         spin_lock(&sinfo->lock);
4521         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4522                 data_used = 0;
4523         meta_used = sinfo->bytes_used;
4524         spin_unlock(&sinfo->lock);
4525
4526         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4527                     csum_size * 2;
4528         num_bytes += div64_u64(data_used + meta_used, 50);
4529
4530         if (num_bytes * 3 > meta_used)
4531                 num_bytes = div64_u64(meta_used, 3);
4532
4533         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4534 }
4535
4536 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4537 {
4538         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4539         struct btrfs_space_info *sinfo = block_rsv->space_info;
4540         u64 num_bytes;
4541
4542         num_bytes = calc_global_metadata_size(fs_info);
4543
4544         spin_lock(&sinfo->lock);
4545         spin_lock(&block_rsv->lock);
4546
4547         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4548
4549         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4550                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4551                     sinfo->bytes_may_use;
4552
4553         if (sinfo->total_bytes > num_bytes) {
4554                 num_bytes = sinfo->total_bytes - num_bytes;
4555                 block_rsv->reserved += num_bytes;
4556                 sinfo->bytes_may_use += num_bytes;
4557                 trace_btrfs_space_reservation(fs_info, "space_info",
4558                                       sinfo->flags, num_bytes, 1);
4559         }
4560
4561         if (block_rsv->reserved >= block_rsv->size) {
4562                 num_bytes = block_rsv->reserved - block_rsv->size;
4563                 sinfo->bytes_may_use -= num_bytes;
4564                 trace_btrfs_space_reservation(fs_info, "space_info",
4565                                       sinfo->flags, num_bytes, 0);
4566                 sinfo->reservation_progress++;
4567                 block_rsv->reserved = block_rsv->size;
4568                 block_rsv->full = 1;
4569         }
4570
4571         spin_unlock(&block_rsv->lock);
4572         spin_unlock(&sinfo->lock);
4573 }
4574
4575 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4576 {
4577         struct btrfs_space_info *space_info;
4578
4579         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4580         fs_info->chunk_block_rsv.space_info = space_info;
4581
4582         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4583         fs_info->global_block_rsv.space_info = space_info;
4584         fs_info->delalloc_block_rsv.space_info = space_info;
4585         fs_info->trans_block_rsv.space_info = space_info;
4586         fs_info->empty_block_rsv.space_info = space_info;
4587         fs_info->delayed_block_rsv.space_info = space_info;
4588
4589         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4590         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4591         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4592         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4593         if (fs_info->quota_root)
4594                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4595         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4596
4597         update_global_block_rsv(fs_info);
4598 }
4599
4600 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4601 {
4602         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4603                                 (u64)-1);
4604         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4605         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4606         WARN_ON(fs_info->trans_block_rsv.size > 0);
4607         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4608         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4609         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4610         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4611         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4612 }
4613
4614 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4615                                   struct btrfs_root *root)
4616 {
4617         if (!trans->block_rsv)
4618                 return;
4619
4620         if (!trans->bytes_reserved)
4621                 return;
4622
4623         trace_btrfs_space_reservation(root->fs_info, "transaction",
4624                                       trans->transid, trans->bytes_reserved, 0);
4625         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4626         trans->bytes_reserved = 0;
4627 }
4628
4629 /* Can only return 0 or -ENOSPC */
4630 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4631                                   struct inode *inode)
4632 {
4633         struct btrfs_root *root = BTRFS_I(inode)->root;
4634         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4635         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4636
4637         /*
4638          * We need to hold space in order to delete our orphan item once we've
4639          * added it, so this takes the reservation so we can release it later
4640          * when we are truly done with the orphan item.
4641          */
4642         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4643         trace_btrfs_space_reservation(root->fs_info, "orphan",
4644                                       btrfs_ino(inode), num_bytes, 1);
4645         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4646 }
4647
4648 void btrfs_orphan_release_metadata(struct inode *inode)
4649 {
4650         struct btrfs_root *root = BTRFS_I(inode)->root;
4651         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4652         trace_btrfs_space_reservation(root->fs_info, "orphan",
4653                                       btrfs_ino(inode), num_bytes, 0);
4654         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4655 }
4656
4657 /*
4658  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4659  * root: the root of the parent directory
4660  * rsv: block reservation
4661  * items: the number of items that we need do reservation
4662  * qgroup_reserved: used to return the reserved size in qgroup
4663  *
4664  * This function is used to reserve the space for snapshot/subvolume
4665  * creation and deletion. Those operations are different with the
4666  * common file/directory operations, they change two fs/file trees
4667  * and root tree, the number of items that the qgroup reserves is
4668  * different with the free space reservation. So we can not use
4669  * the space reseravtion mechanism in start_transaction().
4670  */
4671 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4672                                      struct btrfs_block_rsv *rsv,
4673                                      int items,
4674                                      u64 *qgroup_reserved)
4675 {
4676         u64 num_bytes;
4677         int ret;
4678
4679         if (root->fs_info->quota_enabled) {
4680                 /* One for parent inode, two for dir entries */
4681                 num_bytes = 3 * root->leafsize;
4682                 ret = btrfs_qgroup_reserve(root, num_bytes);
4683                 if (ret)
4684                         return ret;
4685         } else {
4686                 num_bytes = 0;
4687         }
4688
4689         *qgroup_reserved = num_bytes;
4690
4691         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4692         rsv->space_info = __find_space_info(root->fs_info,
4693                                             BTRFS_BLOCK_GROUP_METADATA);
4694         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4695                                   BTRFS_RESERVE_FLUSH_ALL);
4696         if (ret) {
4697                 if (*qgroup_reserved)
4698                         btrfs_qgroup_free(root, *qgroup_reserved);
4699         }
4700
4701         return ret;
4702 }
4703
4704 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4705                                       struct btrfs_block_rsv *rsv,
4706                                       u64 qgroup_reserved)
4707 {
4708         btrfs_block_rsv_release(root, rsv, (u64)-1);
4709         if (qgroup_reserved)
4710                 btrfs_qgroup_free(root, qgroup_reserved);
4711 }
4712
4713 /**
4714  * drop_outstanding_extent - drop an outstanding extent
4715  * @inode: the inode we're dropping the extent for
4716  *
4717  * This is called when we are freeing up an outstanding extent, either called
4718  * after an error or after an extent is written.  This will return the number of
4719  * reserved extents that need to be freed.  This must be called with
4720  * BTRFS_I(inode)->lock held.
4721  */
4722 static unsigned drop_outstanding_extent(struct inode *inode)
4723 {
4724         unsigned drop_inode_space = 0;
4725         unsigned dropped_extents = 0;
4726
4727         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4728         BTRFS_I(inode)->outstanding_extents--;
4729
4730         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4731             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4732                                &BTRFS_I(inode)->runtime_flags))
4733                 drop_inode_space = 1;
4734
4735         /*
4736          * If we have more or the same amount of outsanding extents than we have
4737          * reserved then we need to leave the reserved extents count alone.
4738          */
4739         if (BTRFS_I(inode)->outstanding_extents >=
4740             BTRFS_I(inode)->reserved_extents)
4741                 return drop_inode_space;
4742
4743         dropped_extents = BTRFS_I(inode)->reserved_extents -
4744                 BTRFS_I(inode)->outstanding_extents;
4745         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4746         return dropped_extents + drop_inode_space;
4747 }
4748
4749 /**
4750  * calc_csum_metadata_size - return the amount of metada space that must be
4751  *      reserved/free'd for the given bytes.
4752  * @inode: the inode we're manipulating
4753  * @num_bytes: the number of bytes in question
4754  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4755  *
4756  * This adjusts the number of csum_bytes in the inode and then returns the
4757  * correct amount of metadata that must either be reserved or freed.  We
4758  * calculate how many checksums we can fit into one leaf and then divide the
4759  * number of bytes that will need to be checksumed by this value to figure out
4760  * how many checksums will be required.  If we are adding bytes then the number
4761  * may go up and we will return the number of additional bytes that must be
4762  * reserved.  If it is going down we will return the number of bytes that must
4763  * be freed.
4764  *
4765  * This must be called with BTRFS_I(inode)->lock held.
4766  */
4767 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4768                                    int reserve)
4769 {
4770         struct btrfs_root *root = BTRFS_I(inode)->root;
4771         u64 csum_size;
4772         int num_csums_per_leaf;
4773         int num_csums;
4774         int old_csums;
4775
4776         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4777             BTRFS_I(inode)->csum_bytes == 0)
4778                 return 0;
4779
4780         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4781         if (reserve)
4782                 BTRFS_I(inode)->csum_bytes += num_bytes;
4783         else
4784                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4785         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4786         num_csums_per_leaf = (int)div64_u64(csum_size,
4787                                             sizeof(struct btrfs_csum_item) +
4788                                             sizeof(struct btrfs_disk_key));
4789         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4790         num_csums = num_csums + num_csums_per_leaf - 1;
4791         num_csums = num_csums / num_csums_per_leaf;
4792
4793         old_csums = old_csums + num_csums_per_leaf - 1;
4794         old_csums = old_csums / num_csums_per_leaf;
4795
4796         /* No change, no need to reserve more */
4797         if (old_csums == num_csums)
4798                 return 0;
4799
4800         if (reserve)
4801                 return btrfs_calc_trans_metadata_size(root,
4802                                                       num_csums - old_csums);
4803
4804         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4805 }
4806
4807 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4808 {
4809         struct btrfs_root *root = BTRFS_I(inode)->root;
4810         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4811         u64 to_reserve = 0;
4812         u64 csum_bytes;
4813         unsigned nr_extents = 0;
4814         int extra_reserve = 0;
4815         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4816         int ret = 0;
4817         bool delalloc_lock = true;
4818         u64 to_free = 0;
4819         unsigned dropped;
4820
4821         /* If we are a free space inode we need to not flush since we will be in
4822          * the middle of a transaction commit.  We also don't need the delalloc
4823          * mutex since we won't race with anybody.  We need this mostly to make
4824          * lockdep shut its filthy mouth.
4825          */
4826         if (btrfs_is_free_space_inode(inode)) {
4827                 flush = BTRFS_RESERVE_NO_FLUSH;
4828                 delalloc_lock = false;
4829         }
4830
4831         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4832             btrfs_transaction_in_commit(root->fs_info))
4833                 schedule_timeout(1);
4834
4835         if (delalloc_lock)
4836                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4837
4838         num_bytes = ALIGN(num_bytes, root->sectorsize);
4839
4840         spin_lock(&BTRFS_I(inode)->lock);
4841         BTRFS_I(inode)->outstanding_extents++;
4842
4843         if (BTRFS_I(inode)->outstanding_extents >
4844             BTRFS_I(inode)->reserved_extents)
4845                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4846                         BTRFS_I(inode)->reserved_extents;
4847
4848         /*
4849          * Add an item to reserve for updating the inode when we complete the
4850          * delalloc io.
4851          */
4852         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4853                       &BTRFS_I(inode)->runtime_flags)) {
4854                 nr_extents++;
4855                 extra_reserve = 1;
4856         }
4857
4858         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4859         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4860         csum_bytes = BTRFS_I(inode)->csum_bytes;
4861         spin_unlock(&BTRFS_I(inode)->lock);
4862
4863         if (root->fs_info->quota_enabled) {
4864                 ret = btrfs_qgroup_reserve(root, num_bytes +
4865                                            nr_extents * root->leafsize);
4866                 if (ret)
4867                         goto out_fail;
4868         }
4869
4870         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4871         if (unlikely(ret)) {
4872                 if (root->fs_info->quota_enabled)
4873                         btrfs_qgroup_free(root, num_bytes +
4874                                                 nr_extents * root->leafsize);
4875                 goto out_fail;
4876         }
4877
4878         spin_lock(&BTRFS_I(inode)->lock);
4879         if (extra_reserve) {
4880                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4881                         &BTRFS_I(inode)->runtime_flags);
4882                 nr_extents--;
4883         }
4884         BTRFS_I(inode)->reserved_extents += nr_extents;
4885         spin_unlock(&BTRFS_I(inode)->lock);
4886
4887         if (delalloc_lock)
4888                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4889
4890         if (to_reserve)
4891                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4892                                               btrfs_ino(inode), to_reserve, 1);
4893         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4894
4895         return 0;
4896
4897 out_fail:
4898         spin_lock(&BTRFS_I(inode)->lock);
4899         dropped = drop_outstanding_extent(inode);
4900         /*
4901          * If the inodes csum_bytes is the same as the original
4902          * csum_bytes then we know we haven't raced with any free()ers
4903          * so we can just reduce our inodes csum bytes and carry on.
4904          */
4905         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4906                 calc_csum_metadata_size(inode, num_bytes, 0);
4907         } else {
4908                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4909                 u64 bytes;
4910
4911                 /*
4912                  * This is tricky, but first we need to figure out how much we
4913                  * free'd from any free-ers that occured during this
4914                  * reservation, so we reset ->csum_bytes to the csum_bytes
4915                  * before we dropped our lock, and then call the free for the
4916                  * number of bytes that were freed while we were trying our
4917                  * reservation.
4918                  */
4919                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4920                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4921                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4922
4923
4924                 /*
4925                  * Now we need to see how much we would have freed had we not
4926                  * been making this reservation and our ->csum_bytes were not
4927                  * artificially inflated.
4928                  */
4929                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4930                 bytes = csum_bytes - orig_csum_bytes;
4931                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4932
4933                 /*
4934                  * Now reset ->csum_bytes to what it should be.  If bytes is
4935                  * more than to_free then we would have free'd more space had we
4936                  * not had an artificially high ->csum_bytes, so we need to free
4937                  * the remainder.  If bytes is the same or less then we don't
4938                  * need to do anything, the other free-ers did the correct
4939                  * thing.
4940                  */
4941                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4942                 if (bytes > to_free)
4943                         to_free = bytes - to_free;
4944                 else
4945                         to_free = 0;
4946         }
4947         spin_unlock(&BTRFS_I(inode)->lock);
4948         if (dropped)
4949                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4950
4951         if (to_free) {
4952                 btrfs_block_rsv_release(root, block_rsv, to_free);
4953                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4954                                               btrfs_ino(inode), to_free, 0);
4955         }
4956         if (delalloc_lock)
4957                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4958         return ret;
4959 }
4960
4961 /**
4962  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4963  * @inode: the inode to release the reservation for
4964  * @num_bytes: the number of bytes we're releasing
4965  *
4966  * This will release the metadata reservation for an inode.  This can be called
4967  * once we complete IO for a given set of bytes to release their metadata
4968  * reservations.
4969  */
4970 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4971 {
4972         struct btrfs_root *root = BTRFS_I(inode)->root;
4973         u64 to_free = 0;
4974         unsigned dropped;
4975
4976         num_bytes = ALIGN(num_bytes, root->sectorsize);
4977         spin_lock(&BTRFS_I(inode)->lock);
4978         dropped = drop_outstanding_extent(inode);
4979
4980         if (num_bytes)
4981                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4982         spin_unlock(&BTRFS_I(inode)->lock);
4983         if (dropped > 0)
4984                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4985
4986         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4987                                       btrfs_ino(inode), to_free, 0);
4988         if (root->fs_info->quota_enabled) {
4989                 btrfs_qgroup_free(root, num_bytes +
4990                                         dropped * root->leafsize);
4991         }
4992
4993         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4994                                 to_free);
4995 }
4996
4997 /**
4998  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4999  * @inode: inode we're writing to
5000  * @num_bytes: the number of bytes we want to allocate
5001  *
5002  * This will do the following things
5003  *
5004  * o reserve space in the data space info for num_bytes
5005  * o reserve space in the metadata space info based on number of outstanding
5006  *   extents and how much csums will be needed
5007  * o add to the inodes ->delalloc_bytes
5008  * o add it to the fs_info's delalloc inodes list.
5009  *
5010  * This will return 0 for success and -ENOSPC if there is no space left.
5011  */
5012 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5013 {
5014         int ret;
5015
5016         ret = btrfs_check_data_free_space(inode, num_bytes);
5017         if (ret)
5018                 return ret;
5019
5020         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5021         if (ret) {
5022                 btrfs_free_reserved_data_space(inode, num_bytes);
5023                 return ret;
5024         }
5025
5026         return 0;
5027 }
5028
5029 /**
5030  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5031  * @inode: inode we're releasing space for
5032  * @num_bytes: the number of bytes we want to free up
5033  *
5034  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5035  * called in the case that we don't need the metadata AND data reservations
5036  * anymore.  So if there is an error or we insert an inline extent.
5037  *
5038  * This function will release the metadata space that was not used and will
5039  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5040  * list if there are no delalloc bytes left.
5041  */
5042 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5043 {
5044         btrfs_delalloc_release_metadata(inode, num_bytes);
5045         btrfs_free_reserved_data_space(inode, num_bytes);
5046 }
5047
5048 static int update_block_group(struct btrfs_root *root,
5049                               u64 bytenr, u64 num_bytes, int alloc)
5050 {
5051         struct btrfs_block_group_cache *cache = NULL;
5052         struct btrfs_fs_info *info = root->fs_info;
5053         u64 total = num_bytes;
5054         u64 old_val;
5055         u64 byte_in_group;
5056         int factor;
5057
5058         /* block accounting for super block */
5059         spin_lock(&info->delalloc_root_lock);
5060         old_val = btrfs_super_bytes_used(info->super_copy);
5061         if (alloc)
5062                 old_val += num_bytes;
5063         else
5064                 old_val -= num_bytes;
5065         btrfs_set_super_bytes_used(info->super_copy, old_val);
5066         spin_unlock(&info->delalloc_root_lock);
5067
5068         while (total) {
5069                 cache = btrfs_lookup_block_group(info, bytenr);
5070                 if (!cache)
5071                         return -ENOENT;
5072                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5073                                     BTRFS_BLOCK_GROUP_RAID1 |
5074                                     BTRFS_BLOCK_GROUP_RAID10))
5075                         factor = 2;
5076                 else
5077                         factor = 1;
5078                 /*
5079                  * If this block group has free space cache written out, we
5080                  * need to make sure to load it if we are removing space.  This
5081                  * is because we need the unpinning stage to actually add the
5082                  * space back to the block group, otherwise we will leak space.
5083                  */
5084                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5085                         cache_block_group(cache, 1);
5086
5087                 byte_in_group = bytenr - cache->key.objectid;
5088                 WARN_ON(byte_in_group > cache->key.offset);
5089
5090                 spin_lock(&cache->space_info->lock);
5091                 spin_lock(&cache->lock);
5092
5093                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5094                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5095                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5096
5097                 cache->dirty = 1;
5098                 old_val = btrfs_block_group_used(&cache->item);
5099                 num_bytes = min(total, cache->key.offset - byte_in_group);
5100                 if (alloc) {
5101                         old_val += num_bytes;
5102                         btrfs_set_block_group_used(&cache->item, old_val);
5103                         cache->reserved -= num_bytes;
5104                         cache->space_info->bytes_reserved -= num_bytes;
5105                         cache->space_info->bytes_used += num_bytes;
5106                         cache->space_info->disk_used += num_bytes * factor;
5107                         spin_unlock(&cache->lock);
5108                         spin_unlock(&cache->space_info->lock);
5109                 } else {
5110                         old_val -= num_bytes;
5111                         btrfs_set_block_group_used(&cache->item, old_val);
5112                         cache->pinned += num_bytes;
5113                         cache->space_info->bytes_pinned += num_bytes;
5114                         cache->space_info->bytes_used -= num_bytes;
5115                         cache->space_info->disk_used -= num_bytes * factor;
5116                         spin_unlock(&cache->lock);
5117                         spin_unlock(&cache->space_info->lock);
5118
5119                         set_extent_dirty(info->pinned_extents,
5120                                          bytenr, bytenr + num_bytes - 1,
5121                                          GFP_NOFS | __GFP_NOFAIL);
5122                 }
5123                 btrfs_put_block_group(cache);
5124                 total -= num_bytes;
5125                 bytenr += num_bytes;
5126         }
5127         return 0;
5128 }
5129
5130 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5131 {
5132         struct btrfs_block_group_cache *cache;
5133         u64 bytenr;
5134
5135         spin_lock(&root->fs_info->block_group_cache_lock);
5136         bytenr = root->fs_info->first_logical_byte;
5137         spin_unlock(&root->fs_info->block_group_cache_lock);
5138
5139         if (bytenr < (u64)-1)
5140                 return bytenr;
5141
5142         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5143         if (!cache)
5144                 return 0;
5145
5146         bytenr = cache->key.objectid;
5147         btrfs_put_block_group(cache);
5148
5149         return bytenr;
5150 }
5151
5152 static int pin_down_extent(struct btrfs_root *root,
5153                            struct btrfs_block_group_cache *cache,
5154                            u64 bytenr, u64 num_bytes, int reserved)
5155 {
5156         spin_lock(&cache->space_info->lock);
5157         spin_lock(&cache->lock);
5158         cache->pinned += num_bytes;
5159         cache->space_info->bytes_pinned += num_bytes;
5160         if (reserved) {
5161                 cache->reserved -= num_bytes;
5162                 cache->space_info->bytes_reserved -= num_bytes;
5163         }
5164         spin_unlock(&cache->lock);
5165         spin_unlock(&cache->space_info->lock);
5166
5167         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5168                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5169         return 0;
5170 }
5171
5172 /*
5173  * this function must be called within transaction
5174  */
5175 int btrfs_pin_extent(struct btrfs_root *root,
5176                      u64 bytenr, u64 num_bytes, int reserved)
5177 {
5178         struct btrfs_block_group_cache *cache;
5179
5180         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5181         BUG_ON(!cache); /* Logic error */
5182
5183         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5184
5185         btrfs_put_block_group(cache);
5186         return 0;
5187 }
5188
5189 /*
5190  * this function must be called within transaction
5191  */
5192 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5193                                     u64 bytenr, u64 num_bytes)
5194 {
5195         struct btrfs_block_group_cache *cache;
5196         int ret;
5197
5198         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5199         if (!cache)
5200                 return -EINVAL;
5201
5202         /*
5203          * pull in the free space cache (if any) so that our pin
5204          * removes the free space from the cache.  We have load_only set
5205          * to one because the slow code to read in the free extents does check
5206          * the pinned extents.
5207          */
5208         cache_block_group(cache, 1);
5209
5210         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5211
5212         /* remove us from the free space cache (if we're there at all) */
5213         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5214         btrfs_put_block_group(cache);
5215         return ret;
5216 }
5217
5218 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5219 {
5220         int ret;
5221         struct btrfs_block_group_cache *block_group;
5222         struct btrfs_caching_control *caching_ctl;
5223
5224         block_group = btrfs_lookup_block_group(root->fs_info, start);
5225         if (!block_group)
5226                 return -EINVAL;
5227
5228         cache_block_group(block_group, 0);
5229         caching_ctl = get_caching_control(block_group);
5230
5231         if (!caching_ctl) {
5232                 /* Logic error */
5233                 BUG_ON(!block_group_cache_done(block_group));
5234                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5235         } else {
5236                 mutex_lock(&caching_ctl->mutex);
5237
5238                 if (start >= caching_ctl->progress) {
5239                         ret = add_excluded_extent(root, start, num_bytes);
5240                 } else if (start + num_bytes <= caching_ctl->progress) {
5241                         ret = btrfs_remove_free_space(block_group,
5242                                                       start, num_bytes);
5243                 } else {
5244                         num_bytes = caching_ctl->progress - start;
5245                         ret = btrfs_remove_free_space(block_group,
5246                                                       start, num_bytes);
5247                         if (ret)
5248                                 goto out_lock;
5249
5250                         num_bytes = (start + num_bytes) -
5251                                 caching_ctl->progress;
5252                         start = caching_ctl->progress;
5253                         ret = add_excluded_extent(root, start, num_bytes);
5254                 }
5255 out_lock:
5256                 mutex_unlock(&caching_ctl->mutex);
5257                 put_caching_control(caching_ctl);
5258         }
5259         btrfs_put_block_group(block_group);
5260         return ret;
5261 }
5262
5263 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5264                                  struct extent_buffer *eb)
5265 {
5266         struct btrfs_file_extent_item *item;
5267         struct btrfs_key key;
5268         int found_type;
5269         int i;
5270
5271         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5272                 return 0;
5273
5274         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5275                 btrfs_item_key_to_cpu(eb, &key, i);
5276                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5277                         continue;
5278                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5279                 found_type = btrfs_file_extent_type(eb, item);
5280                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5281                         continue;
5282                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5283                         continue;
5284                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5285                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5286                 __exclude_logged_extent(log, key.objectid, key.offset);
5287         }
5288
5289         return 0;
5290 }
5291
5292 /**
5293  * btrfs_update_reserved_bytes - update the block_group and space info counters
5294  * @cache:      The cache we are manipulating
5295  * @num_bytes:  The number of bytes in question
5296  * @reserve:    One of the reservation enums
5297  *
5298  * This is called by the allocator when it reserves space, or by somebody who is
5299  * freeing space that was never actually used on disk.  For example if you
5300  * reserve some space for a new leaf in transaction A and before transaction A
5301  * commits you free that leaf, you call this with reserve set to 0 in order to
5302  * clear the reservation.
5303  *
5304  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5305  * ENOSPC accounting.  For data we handle the reservation through clearing the
5306  * delalloc bits in the io_tree.  We have to do this since we could end up
5307  * allocating less disk space for the amount of data we have reserved in the
5308  * case of compression.
5309  *
5310  * If this is a reservation and the block group has become read only we cannot
5311  * make the reservation and return -EAGAIN, otherwise this function always
5312  * succeeds.
5313  */
5314 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5315                                        u64 num_bytes, int reserve)
5316 {
5317         struct btrfs_space_info *space_info = cache->space_info;
5318         int ret = 0;
5319
5320         spin_lock(&space_info->lock);
5321         spin_lock(&cache->lock);
5322         if (reserve != RESERVE_FREE) {
5323                 if (cache->ro) {
5324                         ret = -EAGAIN;
5325                 } else {
5326                         cache->reserved += num_bytes;
5327                         space_info->bytes_reserved += num_bytes;
5328                         if (reserve == RESERVE_ALLOC) {
5329                                 trace_btrfs_space_reservation(cache->fs_info,
5330                                                 "space_info", space_info->flags,
5331                                                 num_bytes, 0);
5332                                 space_info->bytes_may_use -= num_bytes;
5333                         }
5334                 }
5335         } else {
5336                 if (cache->ro)
5337                         space_info->bytes_readonly += num_bytes;
5338                 cache->reserved -= num_bytes;
5339                 space_info->bytes_reserved -= num_bytes;
5340                 space_info->reservation_progress++;
5341         }
5342         spin_unlock(&cache->lock);
5343         spin_unlock(&space_info->lock);
5344         return ret;
5345 }
5346
5347 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5348                                 struct btrfs_root *root)
5349 {
5350         struct btrfs_fs_info *fs_info = root->fs_info;
5351         struct btrfs_caching_control *next;
5352         struct btrfs_caching_control *caching_ctl;
5353         struct btrfs_block_group_cache *cache;
5354
5355         down_write(&fs_info->extent_commit_sem);
5356
5357         list_for_each_entry_safe(caching_ctl, next,
5358                                  &fs_info->caching_block_groups, list) {
5359                 cache = caching_ctl->block_group;
5360                 if (block_group_cache_done(cache)) {
5361                         cache->last_byte_to_unpin = (u64)-1;
5362                         list_del_init(&caching_ctl->list);
5363                         put_caching_control(caching_ctl);
5364                 } else {
5365                         cache->last_byte_to_unpin = caching_ctl->progress;
5366                 }
5367         }
5368
5369         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5370                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5371         else
5372                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5373
5374         up_write(&fs_info->extent_commit_sem);
5375
5376         update_global_block_rsv(fs_info);
5377 }
5378
5379 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5380 {
5381         struct btrfs_fs_info *fs_info = root->fs_info;
5382         struct btrfs_block_group_cache *cache = NULL;
5383         struct btrfs_space_info *space_info;
5384         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5385         u64 len;
5386         bool readonly;
5387
5388         while (start <= end) {
5389                 readonly = false;
5390                 if (!cache ||
5391                     start >= cache->key.objectid + cache->key.offset) {
5392                         if (cache)
5393                                 btrfs_put_block_group(cache);
5394                         cache = btrfs_lookup_block_group(fs_info, start);
5395                         BUG_ON(!cache); /* Logic error */
5396                 }
5397
5398                 len = cache->key.objectid + cache->key.offset - start;
5399                 len = min(len, end + 1 - start);
5400
5401                 if (start < cache->last_byte_to_unpin) {
5402                         len = min(len, cache->last_byte_to_unpin - start);
5403                         btrfs_add_free_space(cache, start, len);
5404                 }
5405
5406                 start += len;
5407                 space_info = cache->space_info;
5408
5409                 spin_lock(&space_info->lock);
5410                 spin_lock(&cache->lock);
5411                 cache->pinned -= len;
5412                 space_info->bytes_pinned -= len;
5413                 if (cache->ro) {
5414                         space_info->bytes_readonly += len;
5415                         readonly = true;
5416                 }
5417                 spin_unlock(&cache->lock);
5418                 if (!readonly && global_rsv->space_info == space_info) {
5419                         spin_lock(&global_rsv->lock);
5420                         if (!global_rsv->full) {
5421                                 len = min(len, global_rsv->size -
5422                                           global_rsv->reserved);
5423                                 global_rsv->reserved += len;
5424                                 space_info->bytes_may_use += len;
5425                                 if (global_rsv->reserved >= global_rsv->size)
5426                                         global_rsv->full = 1;
5427                         }
5428                         spin_unlock(&global_rsv->lock);
5429                 }
5430                 spin_unlock(&space_info->lock);
5431         }
5432
5433         if (cache)
5434                 btrfs_put_block_group(cache);
5435         return 0;
5436 }
5437
5438 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5439                                struct btrfs_root *root)
5440 {
5441         struct btrfs_fs_info *fs_info = root->fs_info;
5442         struct extent_io_tree *unpin;
5443         u64 start;
5444         u64 end;
5445         int ret;
5446
5447         if (trans->aborted)
5448                 return 0;
5449
5450         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5451                 unpin = &fs_info->freed_extents[1];
5452         else
5453                 unpin = &fs_info->freed_extents[0];
5454
5455         while (1) {
5456                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5457                                             EXTENT_DIRTY, NULL);
5458                 if (ret)
5459                         break;
5460
5461                 if (btrfs_test_opt(root, DISCARD))
5462                         ret = btrfs_discard_extent(root, start,
5463                                                    end + 1 - start, NULL);
5464
5465                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5466                 unpin_extent_range(root, start, end);
5467                 cond_resched();
5468         }
5469
5470         return 0;
5471 }
5472
5473 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5474                                 struct btrfs_root *root,
5475                                 u64 bytenr, u64 num_bytes, u64 parent,
5476                                 u64 root_objectid, u64 owner_objectid,
5477                                 u64 owner_offset, int refs_to_drop,
5478                                 struct btrfs_delayed_extent_op *extent_op)
5479 {
5480         struct btrfs_key key;
5481         struct btrfs_path *path;
5482         struct btrfs_fs_info *info = root->fs_info;
5483         struct btrfs_root *extent_root = info->extent_root;
5484         struct extent_buffer *leaf;
5485         struct btrfs_extent_item *ei;
5486         struct btrfs_extent_inline_ref *iref;
5487         int ret;
5488         int is_data;
5489         int extent_slot = 0;
5490         int found_extent = 0;
5491         int num_to_del = 1;
5492         u32 item_size;
5493         u64 refs;
5494         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5495                                                  SKINNY_METADATA);
5496
5497         path = btrfs_alloc_path();
5498         if (!path)
5499                 return -ENOMEM;
5500
5501         path->reada = 1;
5502         path->leave_spinning = 1;
5503
5504         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5505         BUG_ON(!is_data && refs_to_drop != 1);
5506
5507         if (is_data)
5508                 skinny_metadata = 0;
5509
5510         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5511                                     bytenr, num_bytes, parent,
5512                                     root_objectid, owner_objectid,
5513                                     owner_offset);
5514         if (ret == 0) {
5515                 extent_slot = path->slots[0];
5516                 while (extent_slot >= 0) {
5517                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5518                                               extent_slot);
5519                         if (key.objectid != bytenr)
5520                                 break;
5521                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5522                             key.offset == num_bytes) {
5523                                 found_extent = 1;
5524                                 break;
5525                         }
5526                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5527                             key.offset == owner_objectid) {
5528                                 found_extent = 1;
5529                                 break;
5530                         }
5531                         if (path->slots[0] - extent_slot > 5)
5532                                 break;
5533                         extent_slot--;
5534                 }
5535 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5536                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5537                 if (found_extent && item_size < sizeof(*ei))
5538                         found_extent = 0;
5539 #endif
5540                 if (!found_extent) {
5541                         BUG_ON(iref);
5542                         ret = remove_extent_backref(trans, extent_root, path,
5543                                                     NULL, refs_to_drop,
5544                                                     is_data);
5545                         if (ret) {
5546                                 btrfs_abort_transaction(trans, extent_root, ret);
5547                                 goto out;
5548                         }
5549                         btrfs_release_path(path);
5550                         path->leave_spinning = 1;
5551
5552                         key.objectid = bytenr;
5553                         key.type = BTRFS_EXTENT_ITEM_KEY;
5554                         key.offset = num_bytes;
5555
5556                         if (!is_data && skinny_metadata) {
5557                                 key.type = BTRFS_METADATA_ITEM_KEY;
5558                                 key.offset = owner_objectid;
5559                         }
5560
5561                         ret = btrfs_search_slot(trans, extent_root,
5562                                                 &key, path, -1, 1);
5563                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5564                                 /*
5565                                  * Couldn't find our skinny metadata item,
5566                                  * see if we have ye olde extent item.
5567                                  */
5568                                 path->slots[0]--;
5569                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5570                                                       path->slots[0]);
5571                                 if (key.objectid == bytenr &&
5572                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5573                                     key.offset == num_bytes)
5574                                         ret = 0;
5575                         }
5576
5577                         if (ret > 0 && skinny_metadata) {
5578                                 skinny_metadata = false;
5579                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5580                                 key.offset = num_bytes;
5581                                 btrfs_release_path(path);
5582                                 ret = btrfs_search_slot(trans, extent_root,
5583                                                         &key, path, -1, 1);
5584                         }
5585
5586                         if (ret) {
5587                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5588                                         ret, (unsigned long long)bytenr);
5589                                 if (ret > 0)
5590                                         btrfs_print_leaf(extent_root,
5591                                                          path->nodes[0]);
5592                         }
5593                         if (ret < 0) {
5594                                 btrfs_abort_transaction(trans, extent_root, ret);
5595                                 goto out;
5596                         }
5597                         extent_slot = path->slots[0];
5598                 }
5599         } else if (ret == -ENOENT) {
5600                 btrfs_print_leaf(extent_root, path->nodes[0]);
5601                 WARN_ON(1);
5602                 btrfs_err(info,
5603                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5604                         (unsigned long long)bytenr,
5605                         (unsigned long long)parent,
5606                         (unsigned long long)root_objectid,
5607                         (unsigned long long)owner_objectid,
5608                         (unsigned long long)owner_offset);
5609         } else {
5610                 btrfs_abort_transaction(trans, extent_root, ret);
5611                 goto out;
5612         }
5613
5614         leaf = path->nodes[0];
5615         item_size = btrfs_item_size_nr(leaf, extent_slot);
5616 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5617         if (item_size < sizeof(*ei)) {
5618                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5619                 ret = convert_extent_item_v0(trans, extent_root, path,
5620                                              owner_objectid, 0);
5621                 if (ret < 0) {
5622                         btrfs_abort_transaction(trans, extent_root, ret);
5623                         goto out;
5624                 }
5625
5626                 btrfs_release_path(path);
5627                 path->leave_spinning = 1;
5628
5629                 key.objectid = bytenr;
5630                 key.type = BTRFS_EXTENT_ITEM_KEY;
5631                 key.offset = num_bytes;
5632
5633                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5634                                         -1, 1);
5635                 if (ret) {
5636                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5637                                 ret, (unsigned long long)bytenr);
5638                         btrfs_print_leaf(extent_root, path->nodes[0]);
5639                 }
5640                 if (ret < 0) {
5641                         btrfs_abort_transaction(trans, extent_root, ret);
5642                         goto out;
5643                 }
5644
5645                 extent_slot = path->slots[0];
5646                 leaf = path->nodes[0];
5647                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5648         }
5649 #endif
5650         BUG_ON(item_size < sizeof(*ei));
5651         ei = btrfs_item_ptr(leaf, extent_slot,
5652                             struct btrfs_extent_item);
5653         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5654             key.type == BTRFS_EXTENT_ITEM_KEY) {
5655                 struct btrfs_tree_block_info *bi;
5656                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5657                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5658                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5659         }
5660
5661         refs = btrfs_extent_refs(leaf, ei);
5662         if (refs < refs_to_drop) {
5663                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5664                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5665                 ret = -EINVAL;
5666                 btrfs_abort_transaction(trans, extent_root, ret);
5667                 goto out;
5668         }
5669         refs -= refs_to_drop;
5670
5671         if (refs > 0) {
5672                 if (extent_op)
5673                         __run_delayed_extent_op(extent_op, leaf, ei);
5674                 /*
5675                  * In the case of inline back ref, reference count will
5676                  * be updated by remove_extent_backref
5677                  */
5678                 if (iref) {
5679                         BUG_ON(!found_extent);
5680                 } else {
5681                         btrfs_set_extent_refs(leaf, ei, refs);
5682                         btrfs_mark_buffer_dirty(leaf);
5683                 }
5684                 if (found_extent) {
5685                         ret = remove_extent_backref(trans, extent_root, path,
5686                                                     iref, refs_to_drop,
5687                                                     is_data);
5688                         if (ret) {
5689                                 btrfs_abort_transaction(trans, extent_root, ret);
5690                                 goto out;
5691                         }
5692                 }
5693         } else {
5694                 if (found_extent) {
5695                         BUG_ON(is_data && refs_to_drop !=
5696                                extent_data_ref_count(root, path, iref));
5697                         if (iref) {
5698                                 BUG_ON(path->slots[0] != extent_slot);
5699                         } else {
5700                                 BUG_ON(path->slots[0] != extent_slot + 1);
5701                                 path->slots[0] = extent_slot;
5702                                 num_to_del = 2;
5703                         }
5704                 }
5705
5706                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5707                                       num_to_del);
5708                 if (ret) {
5709                         btrfs_abort_transaction(trans, extent_root, ret);
5710                         goto out;
5711                 }
5712                 btrfs_release_path(path);
5713
5714                 if (is_data) {
5715                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5716                         if (ret) {
5717                                 btrfs_abort_transaction(trans, extent_root, ret);
5718                                 goto out;
5719                         }
5720                 }
5721
5722                 ret = update_block_group(root, bytenr, num_bytes, 0);
5723                 if (ret) {
5724                         btrfs_abort_transaction(trans, extent_root, ret);
5725                         goto out;
5726                 }
5727         }
5728 out:
5729         btrfs_free_path(path);
5730         return ret;
5731 }
5732
5733 /*
5734  * when we free an block, it is possible (and likely) that we free the last
5735  * delayed ref for that extent as well.  This searches the delayed ref tree for
5736  * a given extent, and if there are no other delayed refs to be processed, it
5737  * removes it from the tree.
5738  */
5739 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5740                                       struct btrfs_root *root, u64 bytenr)
5741 {
5742         struct btrfs_delayed_ref_head *head;
5743         struct btrfs_delayed_ref_root *delayed_refs;
5744         struct btrfs_delayed_ref_node *ref;
5745         struct rb_node *node;
5746         int ret = 0;
5747
5748         delayed_refs = &trans->transaction->delayed_refs;
5749         spin_lock(&delayed_refs->lock);
5750         head = btrfs_find_delayed_ref_head(trans, bytenr);
5751         if (!head)
5752                 goto out;
5753
5754         node = rb_prev(&head->node.rb_node);
5755         if (!node)
5756                 goto out;
5757
5758         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5759
5760         /* there are still entries for this ref, we can't drop it */
5761         if (ref->bytenr == bytenr)
5762                 goto out;
5763
5764         if (head->extent_op) {
5765                 if (!head->must_insert_reserved)
5766                         goto out;
5767                 btrfs_free_delayed_extent_op(head->extent_op);
5768                 head->extent_op = NULL;
5769         }
5770
5771         /*
5772          * waiting for the lock here would deadlock.  If someone else has it
5773          * locked they are already in the process of dropping it anyway
5774          */
5775         if (!mutex_trylock(&head->mutex))
5776                 goto out;
5777
5778         /*
5779          * at this point we have a head with no other entries.  Go
5780          * ahead and process it.
5781          */
5782         head->node.in_tree = 0;
5783         rb_erase(&head->node.rb_node, &delayed_refs->root);
5784
5785         delayed_refs->num_entries--;
5786
5787         /*
5788          * we don't take a ref on the node because we're removing it from the
5789          * tree, so we just steal the ref the tree was holding.
5790          */
5791         delayed_refs->num_heads--;
5792         if (list_empty(&head->cluster))
5793                 delayed_refs->num_heads_ready--;
5794
5795         list_del_init(&head->cluster);
5796         spin_unlock(&delayed_refs->lock);
5797
5798         BUG_ON(head->extent_op);
5799         if (head->must_insert_reserved)
5800                 ret = 1;
5801
5802         mutex_unlock(&head->mutex);
5803         btrfs_put_delayed_ref(&head->node);
5804         return ret;
5805 out:
5806         spin_unlock(&delayed_refs->lock);
5807         return 0;
5808 }
5809
5810 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5811                            struct btrfs_root *root,
5812                            struct extent_buffer *buf,
5813                            u64 parent, int last_ref)
5814 {
5815         struct btrfs_block_group_cache *cache = NULL;
5816         int ret;
5817
5818         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5819                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5820                                         buf->start, buf->len,
5821                                         parent, root->root_key.objectid,
5822                                         btrfs_header_level(buf),
5823                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5824                 BUG_ON(ret); /* -ENOMEM */
5825         }
5826
5827         if (!last_ref)
5828                 return;
5829
5830         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5831
5832         if (btrfs_header_generation(buf) == trans->transid) {
5833                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5834                         ret = check_ref_cleanup(trans, root, buf->start);
5835                         if (!ret)
5836                                 goto out;
5837                 }
5838
5839                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5840                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5841                         goto out;
5842                 }
5843
5844                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5845
5846                 btrfs_add_free_space(cache, buf->start, buf->len);
5847                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5848         }
5849 out:
5850         /*
5851          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5852          * anymore.
5853          */
5854         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5855         btrfs_put_block_group(cache);
5856 }
5857
5858 /* Can return -ENOMEM */
5859 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5860                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5861                       u64 owner, u64 offset, int for_cow)
5862 {
5863         int ret;
5864         struct btrfs_fs_info *fs_info = root->fs_info;
5865
5866         /*
5867          * tree log blocks never actually go into the extent allocation
5868          * tree, just update pinning info and exit early.
5869          */
5870         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5871                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5872                 /* unlocks the pinned mutex */
5873                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5874                 ret = 0;
5875         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5876                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5877                                         num_bytes,
5878                                         parent, root_objectid, (int)owner,
5879                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5880         } else {
5881                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5882                                                 num_bytes,
5883                                                 parent, root_objectid, owner,
5884                                                 offset, BTRFS_DROP_DELAYED_REF,
5885                                                 NULL, for_cow);
5886         }
5887         return ret;
5888 }
5889
5890 static u64 stripe_align(struct btrfs_root *root,
5891                         struct btrfs_block_group_cache *cache,
5892                         u64 val, u64 num_bytes)
5893 {
5894         u64 ret = ALIGN(val, root->stripesize);
5895         return ret;
5896 }
5897
5898 /*
5899  * when we wait for progress in the block group caching, its because
5900  * our allocation attempt failed at least once.  So, we must sleep
5901  * and let some progress happen before we try again.
5902  *
5903  * This function will sleep at least once waiting for new free space to
5904  * show up, and then it will check the block group free space numbers
5905  * for our min num_bytes.  Another option is to have it go ahead
5906  * and look in the rbtree for a free extent of a given size, but this
5907  * is a good start.
5908  */
5909 static noinline int
5910 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5911                                 u64 num_bytes)
5912 {
5913         struct btrfs_caching_control *caching_ctl;
5914
5915         caching_ctl = get_caching_control(cache);
5916         if (!caching_ctl)
5917                 return 0;
5918
5919         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5920                    (cache->free_space_ctl->free_space >= num_bytes));
5921
5922         put_caching_control(caching_ctl);
5923         return 0;
5924 }
5925
5926 static noinline int
5927 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5928 {
5929         struct btrfs_caching_control *caching_ctl;
5930
5931         caching_ctl = get_caching_control(cache);
5932         if (!caching_ctl)
5933                 return 0;
5934
5935         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5936
5937         put_caching_control(caching_ctl);
5938         return 0;
5939 }
5940
5941 int __get_raid_index(u64 flags)
5942 {
5943         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5944                 return BTRFS_RAID_RAID10;
5945         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5946                 return BTRFS_RAID_RAID1;
5947         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5948                 return BTRFS_RAID_DUP;
5949         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5950                 return BTRFS_RAID_RAID0;
5951         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5952                 return BTRFS_RAID_RAID5;
5953         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5954                 return BTRFS_RAID_RAID6;
5955
5956         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5957 }
5958
5959 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5960 {
5961         return __get_raid_index(cache->flags);
5962 }
5963
5964 enum btrfs_loop_type {
5965         LOOP_CACHING_NOWAIT = 0,
5966         LOOP_CACHING_WAIT = 1,
5967         LOOP_ALLOC_CHUNK = 2,
5968         LOOP_NO_EMPTY_SIZE = 3,
5969 };
5970
5971 /*
5972  * walks the btree of allocated extents and find a hole of a given size.
5973  * The key ins is changed to record the hole:
5974  * ins->objectid == block start
5975  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5976  * ins->offset == number of blocks
5977  * Any available blocks before search_start are skipped.
5978  */
5979 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5980                                      struct btrfs_root *orig_root,
5981                                      u64 num_bytes, u64 empty_size,
5982                                      u64 hint_byte, struct btrfs_key *ins,
5983                                      u64 flags)
5984 {
5985         int ret = 0;
5986         struct btrfs_root *root = orig_root->fs_info->extent_root;
5987         struct btrfs_free_cluster *last_ptr = NULL;
5988         struct btrfs_block_group_cache *block_group = NULL;
5989         struct btrfs_block_group_cache *used_block_group;
5990         u64 search_start = 0;
5991         int empty_cluster = 2 * 1024 * 1024;
5992         struct btrfs_space_info *space_info;
5993         int loop = 0;
5994         int index = __get_raid_index(flags);
5995         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
5996                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5997         bool found_uncached_bg = false;
5998         bool failed_cluster_refill = false;
5999         bool failed_alloc = false;
6000         bool use_cluster = true;
6001         bool have_caching_bg = false;
6002
6003         WARN_ON(num_bytes < root->sectorsize);
6004         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6005         ins->objectid = 0;
6006         ins->offset = 0;
6007
6008         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6009
6010         space_info = __find_space_info(root->fs_info, flags);
6011         if (!space_info) {
6012                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6013                 return -ENOSPC;
6014         }
6015
6016         /*
6017          * If the space info is for both data and metadata it means we have a
6018          * small filesystem and we can't use the clustering stuff.
6019          */
6020         if (btrfs_mixed_space_info(space_info))
6021                 use_cluster = false;
6022
6023         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6024                 last_ptr = &root->fs_info->meta_alloc_cluster;
6025                 if (!btrfs_test_opt(root, SSD))
6026                         empty_cluster = 64 * 1024;
6027         }
6028
6029         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6030             btrfs_test_opt(root, SSD)) {
6031                 last_ptr = &root->fs_info->data_alloc_cluster;
6032         }
6033
6034         if (last_ptr) {
6035                 spin_lock(&last_ptr->lock);
6036                 if (last_ptr->block_group)
6037                         hint_byte = last_ptr->window_start;
6038                 spin_unlock(&last_ptr->lock);
6039         }
6040
6041         search_start = max(search_start, first_logical_byte(root, 0));
6042         search_start = max(search_start, hint_byte);
6043
6044         if (!last_ptr)
6045                 empty_cluster = 0;
6046
6047         if (search_start == hint_byte) {
6048                 block_group = btrfs_lookup_block_group(root->fs_info,
6049                                                        search_start);
6050                 used_block_group = block_group;
6051                 /*
6052                  * we don't want to use the block group if it doesn't match our
6053                  * allocation bits, or if its not cached.
6054                  *
6055                  * However if we are re-searching with an ideal block group
6056                  * picked out then we don't care that the block group is cached.
6057                  */
6058                 if (block_group && block_group_bits(block_group, flags) &&
6059                     block_group->cached != BTRFS_CACHE_NO) {
6060                         down_read(&space_info->groups_sem);
6061                         if (list_empty(&block_group->list) ||
6062                             block_group->ro) {
6063                                 /*
6064                                  * someone is removing this block group,
6065                                  * we can't jump into the have_block_group
6066                                  * target because our list pointers are not
6067                                  * valid
6068                                  */
6069                                 btrfs_put_block_group(block_group);
6070                                 up_read(&space_info->groups_sem);
6071                         } else {
6072                                 index = get_block_group_index(block_group);
6073                                 goto have_block_group;
6074                         }
6075                 } else if (block_group) {
6076                         btrfs_put_block_group(block_group);
6077                 }
6078         }
6079 search:
6080         have_caching_bg = false;
6081         down_read(&space_info->groups_sem);
6082         list_for_each_entry(block_group, &space_info->block_groups[index],
6083                             list) {
6084                 u64 offset;
6085                 int cached;
6086
6087                 used_block_group = block_group;
6088                 btrfs_get_block_group(block_group);
6089                 search_start = block_group->key.objectid;
6090
6091                 /*
6092                  * this can happen if we end up cycling through all the
6093                  * raid types, but we want to make sure we only allocate
6094                  * for the proper type.
6095                  */
6096                 if (!block_group_bits(block_group, flags)) {
6097                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6098                                 BTRFS_BLOCK_GROUP_RAID1 |
6099                                 BTRFS_BLOCK_GROUP_RAID5 |
6100                                 BTRFS_BLOCK_GROUP_RAID6 |
6101                                 BTRFS_BLOCK_GROUP_RAID10;
6102
6103                         /*
6104                          * if they asked for extra copies and this block group
6105                          * doesn't provide them, bail.  This does allow us to
6106                          * fill raid0 from raid1.
6107                          */
6108                         if ((flags & extra) && !(block_group->flags & extra))
6109                                 goto loop;
6110                 }
6111
6112 have_block_group:
6113                 cached = block_group_cache_done(block_group);
6114                 if (unlikely(!cached)) {
6115                         found_uncached_bg = true;
6116                         ret = cache_block_group(block_group, 0);
6117                         BUG_ON(ret < 0);
6118                         ret = 0;
6119                 }
6120
6121                 if (unlikely(block_group->ro))
6122                         goto loop;
6123
6124                 /*
6125                  * Ok we want to try and use the cluster allocator, so
6126                  * lets look there
6127                  */
6128                 if (last_ptr) {
6129                         unsigned long aligned_cluster;
6130                         /*
6131                          * the refill lock keeps out other
6132                          * people trying to start a new cluster
6133                          */
6134                         spin_lock(&last_ptr->refill_lock);
6135                         used_block_group = last_ptr->block_group;
6136                         if (used_block_group != block_group &&
6137                             (!used_block_group ||
6138                              used_block_group->ro ||
6139                              !block_group_bits(used_block_group, flags))) {
6140                                 used_block_group = block_group;
6141                                 goto refill_cluster;
6142                         }
6143
6144                         if (used_block_group != block_group)
6145                                 btrfs_get_block_group(used_block_group);
6146
6147                         offset = btrfs_alloc_from_cluster(used_block_group,
6148                           last_ptr, num_bytes, used_block_group->key.objectid);
6149                         if (offset) {
6150                                 /* we have a block, we're done */
6151                                 spin_unlock(&last_ptr->refill_lock);
6152                                 trace_btrfs_reserve_extent_cluster(root,
6153                                         block_group, search_start, num_bytes);
6154                                 goto checks;
6155                         }
6156
6157                         WARN_ON(last_ptr->block_group != used_block_group);
6158                         if (used_block_group != block_group) {
6159                                 btrfs_put_block_group(used_block_group);
6160                                 used_block_group = block_group;
6161                         }
6162 refill_cluster:
6163                         BUG_ON(used_block_group != block_group);
6164                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6165                          * set up a new clusters, so lets just skip it
6166                          * and let the allocator find whatever block
6167                          * it can find.  If we reach this point, we
6168                          * will have tried the cluster allocator
6169                          * plenty of times and not have found
6170                          * anything, so we are likely way too
6171                          * fragmented for the clustering stuff to find
6172                          * anything.
6173                          *
6174                          * However, if the cluster is taken from the
6175                          * current block group, release the cluster
6176                          * first, so that we stand a better chance of
6177                          * succeeding in the unclustered
6178                          * allocation.  */
6179                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6180                             last_ptr->block_group != block_group) {
6181                                 spin_unlock(&last_ptr->refill_lock);
6182                                 goto unclustered_alloc;
6183                         }
6184
6185                         /*
6186                          * this cluster didn't work out, free it and
6187                          * start over
6188                          */
6189                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6190
6191                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6192                                 spin_unlock(&last_ptr->refill_lock);
6193                                 goto unclustered_alloc;
6194                         }
6195
6196                         aligned_cluster = max_t(unsigned long,
6197                                                 empty_cluster + empty_size,
6198                                               block_group->full_stripe_len);
6199
6200                         /* allocate a cluster in this block group */
6201                         ret = btrfs_find_space_cluster(trans, root,
6202                                                block_group, last_ptr,
6203                                                search_start, num_bytes,
6204                                                aligned_cluster);
6205                         if (ret == 0) {
6206                                 /*
6207                                  * now pull our allocation out of this
6208                                  * cluster
6209                                  */
6210                                 offset = btrfs_alloc_from_cluster(block_group,
6211                                                   last_ptr, num_bytes,
6212                                                   search_start);
6213                                 if (offset) {
6214                                         /* we found one, proceed */
6215                                         spin_unlock(&last_ptr->refill_lock);
6216                                         trace_btrfs_reserve_extent_cluster(root,
6217                                                 block_group, search_start,
6218                                                 num_bytes);
6219                                         goto checks;
6220                                 }
6221                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6222                                    && !failed_cluster_refill) {
6223                                 spin_unlock(&last_ptr->refill_lock);
6224
6225                                 failed_cluster_refill = true;
6226                                 wait_block_group_cache_progress(block_group,
6227                                        num_bytes + empty_cluster + empty_size);
6228                                 goto have_block_group;
6229                         }
6230
6231                         /*
6232                          * at this point we either didn't find a cluster
6233                          * or we weren't able to allocate a block from our
6234                          * cluster.  Free the cluster we've been trying
6235                          * to use, and go to the next block group
6236                          */
6237                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6238                         spin_unlock(&last_ptr->refill_lock);
6239                         goto loop;
6240                 }
6241
6242 unclustered_alloc:
6243                 spin_lock(&block_group->free_space_ctl->tree_lock);
6244                 if (cached &&
6245                     block_group->free_space_ctl->free_space <
6246                     num_bytes + empty_cluster + empty_size) {
6247                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6248                         goto loop;
6249                 }
6250                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6251
6252                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6253                                                     num_bytes, empty_size);
6254                 /*
6255                  * If we didn't find a chunk, and we haven't failed on this
6256                  * block group before, and this block group is in the middle of
6257                  * caching and we are ok with waiting, then go ahead and wait
6258                  * for progress to be made, and set failed_alloc to true.
6259                  *
6260                  * If failed_alloc is true then we've already waited on this
6261                  * block group once and should move on to the next block group.
6262                  */
6263                 if (!offset && !failed_alloc && !cached &&
6264                     loop > LOOP_CACHING_NOWAIT) {
6265                         wait_block_group_cache_progress(block_group,
6266                                                 num_bytes + empty_size);
6267                         failed_alloc = true;
6268                         goto have_block_group;
6269                 } else if (!offset) {
6270                         if (!cached)
6271                                 have_caching_bg = true;
6272                         goto loop;
6273                 }
6274 checks:
6275                 search_start = stripe_align(root, used_block_group,
6276                                             offset, num_bytes);
6277
6278                 /* move on to the next group */
6279                 if (search_start + num_bytes >
6280                     used_block_group->key.objectid + used_block_group->key.offset) {
6281                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6282                         goto loop;
6283                 }
6284
6285                 if (offset < search_start)
6286                         btrfs_add_free_space(used_block_group, offset,
6287                                              search_start - offset);
6288                 BUG_ON(offset > search_start);
6289
6290                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6291                                                   alloc_type);
6292                 if (ret == -EAGAIN) {
6293                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6294                         goto loop;
6295                 }
6296
6297                 /* we are all good, lets return */
6298                 ins->objectid = search_start;
6299                 ins->offset = num_bytes;
6300
6301                 trace_btrfs_reserve_extent(orig_root, block_group,
6302                                            search_start, num_bytes);
6303                 if (used_block_group != block_group)
6304                         btrfs_put_block_group(used_block_group);
6305                 btrfs_put_block_group(block_group);
6306                 break;
6307 loop:
6308                 failed_cluster_refill = false;
6309                 failed_alloc = false;
6310                 BUG_ON(index != get_block_group_index(block_group));
6311                 if (used_block_group != block_group)
6312                         btrfs_put_block_group(used_block_group);
6313                 btrfs_put_block_group(block_group);
6314         }
6315         up_read(&space_info->groups_sem);
6316
6317         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6318                 goto search;
6319
6320         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6321                 goto search;
6322
6323         /*
6324          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6325          *                      caching kthreads as we move along
6326          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6327          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6328          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6329          *                      again
6330          */
6331         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6332                 index = 0;
6333                 loop++;
6334                 if (loop == LOOP_ALLOC_CHUNK) {
6335                         ret = do_chunk_alloc(trans, root, flags,
6336                                              CHUNK_ALLOC_FORCE);
6337                         /*
6338                          * Do not bail out on ENOSPC since we
6339                          * can do more things.
6340                          */
6341                         if (ret < 0 && ret != -ENOSPC) {
6342                                 btrfs_abort_transaction(trans,
6343                                                         root, ret);
6344                                 goto out;
6345                         }
6346                 }
6347
6348                 if (loop == LOOP_NO_EMPTY_SIZE) {
6349                         empty_size = 0;
6350                         empty_cluster = 0;
6351                 }
6352
6353                 goto search;
6354         } else if (!ins->objectid) {
6355                 ret = -ENOSPC;
6356         } else if (ins->objectid) {
6357                 ret = 0;
6358         }
6359 out:
6360
6361         return ret;
6362 }
6363
6364 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6365                             int dump_block_groups)
6366 {
6367         struct btrfs_block_group_cache *cache;
6368         int index = 0;
6369
6370         spin_lock(&info->lock);
6371         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6372                (unsigned long long)info->flags,
6373                (unsigned long long)(info->total_bytes - info->bytes_used -
6374                                     info->bytes_pinned - info->bytes_reserved -
6375                                     info->bytes_readonly),
6376                (info->full) ? "" : "not ");
6377         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6378                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6379                (unsigned long long)info->total_bytes,
6380                (unsigned long long)info->bytes_used,
6381                (unsigned long long)info->bytes_pinned,
6382                (unsigned long long)info->bytes_reserved,
6383                (unsigned long long)info->bytes_may_use,
6384                (unsigned long long)info->bytes_readonly);
6385         spin_unlock(&info->lock);
6386
6387         if (!dump_block_groups)
6388                 return;
6389
6390         down_read(&info->groups_sem);
6391 again:
6392         list_for_each_entry(cache, &info->block_groups[index], list) {
6393                 spin_lock(&cache->lock);
6394                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6395                        (unsigned long long)cache->key.objectid,
6396                        (unsigned long long)cache->key.offset,
6397                        (unsigned long long)btrfs_block_group_used(&cache->item),
6398                        (unsigned long long)cache->pinned,
6399                        (unsigned long long)cache->reserved,
6400                        cache->ro ? "[readonly]" : "");
6401                 btrfs_dump_free_space(cache, bytes);
6402                 spin_unlock(&cache->lock);
6403         }
6404         if (++index < BTRFS_NR_RAID_TYPES)
6405                 goto again;
6406         up_read(&info->groups_sem);
6407 }
6408
6409 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6410                          struct btrfs_root *root,
6411                          u64 num_bytes, u64 min_alloc_size,
6412                          u64 empty_size, u64 hint_byte,
6413                          struct btrfs_key *ins, int is_data)
6414 {
6415         bool final_tried = false;
6416         u64 flags;
6417         int ret;
6418
6419         flags = btrfs_get_alloc_profile(root, is_data);
6420 again:
6421         WARN_ON(num_bytes < root->sectorsize);
6422         ret = find_free_extent(trans, root, num_bytes, empty_size,
6423                                hint_byte, ins, flags);
6424
6425         if (ret == -ENOSPC) {
6426                 if (!final_tried) {
6427                         num_bytes = num_bytes >> 1;
6428                         num_bytes = round_down(num_bytes, root->sectorsize);
6429                         num_bytes = max(num_bytes, min_alloc_size);
6430                         if (num_bytes == min_alloc_size)
6431                                 final_tried = true;
6432                         goto again;
6433                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6434                         struct btrfs_space_info *sinfo;
6435
6436                         sinfo = __find_space_info(root->fs_info, flags);
6437                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6438                                 (unsigned long long)flags,
6439                                 (unsigned long long)num_bytes);
6440                         if (sinfo)
6441                                 dump_space_info(sinfo, num_bytes, 1);
6442                 }
6443         }
6444
6445         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6446
6447         return ret;
6448 }
6449
6450 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6451                                         u64 start, u64 len, int pin)
6452 {
6453         struct btrfs_block_group_cache *cache;
6454         int ret = 0;
6455
6456         cache = btrfs_lookup_block_group(root->fs_info, start);
6457         if (!cache) {
6458                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6459                         (unsigned long long)start);
6460                 return -ENOSPC;
6461         }
6462
6463         if (btrfs_test_opt(root, DISCARD))
6464                 ret = btrfs_discard_extent(root, start, len, NULL);
6465
6466         if (pin)
6467                 pin_down_extent(root, cache, start, len, 1);
6468         else {
6469                 btrfs_add_free_space(cache, start, len);
6470                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6471         }
6472         btrfs_put_block_group(cache);
6473
6474         trace_btrfs_reserved_extent_free(root, start, len);
6475
6476         return ret;
6477 }
6478
6479 int btrfs_free_reserved_extent(struct btrfs_root *root,
6480                                         u64 start, u64 len)
6481 {
6482         return __btrfs_free_reserved_extent(root, start, len, 0);
6483 }
6484
6485 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6486                                        u64 start, u64 len)
6487 {
6488         return __btrfs_free_reserved_extent(root, start, len, 1);
6489 }
6490
6491 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6492                                       struct btrfs_root *root,
6493                                       u64 parent, u64 root_objectid,
6494                                       u64 flags, u64 owner, u64 offset,
6495                                       struct btrfs_key *ins, int ref_mod)
6496 {
6497         int ret;
6498         struct btrfs_fs_info *fs_info = root->fs_info;
6499         struct btrfs_extent_item *extent_item;
6500         struct btrfs_extent_inline_ref *iref;
6501         struct btrfs_path *path;
6502         struct extent_buffer *leaf;
6503         int type;
6504         u32 size;
6505
6506         if (parent > 0)
6507                 type = BTRFS_SHARED_DATA_REF_KEY;
6508         else
6509                 type = BTRFS_EXTENT_DATA_REF_KEY;
6510
6511         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6512
6513         path = btrfs_alloc_path();
6514         if (!path)
6515                 return -ENOMEM;
6516
6517         path->leave_spinning = 1;
6518         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6519                                       ins, size);
6520         if (ret) {
6521                 btrfs_free_path(path);
6522                 return ret;
6523         }
6524
6525         leaf = path->nodes[0];
6526         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6527                                      struct btrfs_extent_item);
6528         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6529         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6530         btrfs_set_extent_flags(leaf, extent_item,
6531                                flags | BTRFS_EXTENT_FLAG_DATA);
6532
6533         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6534         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6535         if (parent > 0) {
6536                 struct btrfs_shared_data_ref *ref;
6537                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6538                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6539                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6540         } else {
6541                 struct btrfs_extent_data_ref *ref;
6542                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6543                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6544                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6545                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6546                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6547         }
6548
6549         btrfs_mark_buffer_dirty(path->nodes[0]);
6550         btrfs_free_path(path);
6551
6552         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6553         if (ret) { /* -ENOENT, logic error */
6554                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6555                         (unsigned long long)ins->objectid,
6556                         (unsigned long long)ins->offset);
6557                 BUG();
6558         }
6559         return ret;
6560 }
6561
6562 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6563                                      struct btrfs_root *root,
6564                                      u64 parent, u64 root_objectid,
6565                                      u64 flags, struct btrfs_disk_key *key,
6566                                      int level, struct btrfs_key *ins)
6567 {
6568         int ret;
6569         struct btrfs_fs_info *fs_info = root->fs_info;
6570         struct btrfs_extent_item *extent_item;
6571         struct btrfs_tree_block_info *block_info;
6572         struct btrfs_extent_inline_ref *iref;
6573         struct btrfs_path *path;
6574         struct extent_buffer *leaf;
6575         u32 size = sizeof(*extent_item) + sizeof(*iref);
6576         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6577                                                  SKINNY_METADATA);
6578
6579         if (!skinny_metadata)
6580                 size += sizeof(*block_info);
6581
6582         path = btrfs_alloc_path();
6583         if (!path)
6584                 return -ENOMEM;
6585
6586         path->leave_spinning = 1;
6587         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6588                                       ins, size);
6589         if (ret) {
6590                 btrfs_free_path(path);
6591                 return ret;
6592         }
6593
6594         leaf = path->nodes[0];
6595         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6596                                      struct btrfs_extent_item);
6597         btrfs_set_extent_refs(leaf, extent_item, 1);
6598         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6599         btrfs_set_extent_flags(leaf, extent_item,
6600                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6601
6602         if (skinny_metadata) {
6603                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6604         } else {
6605                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6606                 btrfs_set_tree_block_key(leaf, block_info, key);
6607                 btrfs_set_tree_block_level(leaf, block_info, level);
6608                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6609         }
6610
6611         if (parent > 0) {
6612                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6613                 btrfs_set_extent_inline_ref_type(leaf, iref,
6614                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6615                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6616         } else {
6617                 btrfs_set_extent_inline_ref_type(leaf, iref,
6618                                                  BTRFS_TREE_BLOCK_REF_KEY);
6619                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6620         }
6621
6622         btrfs_mark_buffer_dirty(leaf);
6623         btrfs_free_path(path);
6624
6625         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6626         if (ret) { /* -ENOENT, logic error */
6627                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6628                         (unsigned long long)ins->objectid,
6629                         (unsigned long long)ins->offset);
6630                 BUG();
6631         }
6632         return ret;
6633 }
6634
6635 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6636                                      struct btrfs_root *root,
6637                                      u64 root_objectid, u64 owner,
6638                                      u64 offset, struct btrfs_key *ins)
6639 {
6640         int ret;
6641
6642         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6643
6644         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6645                                          ins->offset, 0,
6646                                          root_objectid, owner, offset,
6647                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6648         return ret;
6649 }
6650
6651 /*
6652  * this is used by the tree logging recovery code.  It records that
6653  * an extent has been allocated and makes sure to clear the free
6654  * space cache bits as well
6655  */
6656 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6657                                    struct btrfs_root *root,
6658                                    u64 root_objectid, u64 owner, u64 offset,
6659                                    struct btrfs_key *ins)
6660 {
6661         int ret;
6662         struct btrfs_block_group_cache *block_group;
6663
6664         /*
6665          * Mixed block groups will exclude before processing the log so we only
6666          * need to do the exlude dance if this fs isn't mixed.
6667          */
6668         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
6669                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
6670                 if (ret)
6671                         return ret;
6672         }
6673
6674         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6675         if (!block_group)
6676                 return -EINVAL;
6677
6678         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6679                                           RESERVE_ALLOC_NO_ACCOUNT);
6680         BUG_ON(ret); /* logic error */
6681         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6682                                          0, owner, offset, ins, 1);
6683         btrfs_put_block_group(block_group);
6684         return ret;
6685 }
6686
6687 static struct extent_buffer *
6688 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6689                       u64 bytenr, u32 blocksize, int level)
6690 {
6691         struct extent_buffer *buf;
6692
6693         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6694         if (!buf)
6695                 return ERR_PTR(-ENOMEM);
6696         btrfs_set_header_generation(buf, trans->transid);
6697         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6698         btrfs_tree_lock(buf);
6699         clean_tree_block(trans, root, buf);
6700         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6701
6702         btrfs_set_lock_blocking(buf);
6703         btrfs_set_buffer_uptodate(buf);
6704
6705         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6706                 /*
6707                  * we allow two log transactions at a time, use different
6708                  * EXENT bit to differentiate dirty pages.
6709                  */
6710                 if (root->log_transid % 2 == 0)
6711                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6712                                         buf->start + buf->len - 1, GFP_NOFS);
6713                 else
6714                         set_extent_new(&root->dirty_log_pages, buf->start,
6715                                         buf->start + buf->len - 1, GFP_NOFS);
6716         } else {
6717                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6718                          buf->start + buf->len - 1, GFP_NOFS);
6719         }
6720         trans->blocks_used++;
6721         /* this returns a buffer locked for blocking */
6722         return buf;
6723 }
6724
6725 static struct btrfs_block_rsv *
6726 use_block_rsv(struct btrfs_trans_handle *trans,
6727               struct btrfs_root *root, u32 blocksize)
6728 {
6729         struct btrfs_block_rsv *block_rsv;
6730         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6731         int ret;
6732         bool global_updated = false;
6733
6734         block_rsv = get_block_rsv(trans, root);
6735
6736         if (unlikely(block_rsv->size == 0))
6737                 goto try_reserve;
6738 again:
6739         ret = block_rsv_use_bytes(block_rsv, blocksize);
6740         if (!ret)
6741                 return block_rsv;
6742
6743         if (block_rsv->failfast)
6744                 return ERR_PTR(ret);
6745
6746         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
6747                 global_updated = true;
6748                 update_global_block_rsv(root->fs_info);
6749                 goto again;
6750         }
6751
6752         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6753                 static DEFINE_RATELIMIT_STATE(_rs,
6754                                 DEFAULT_RATELIMIT_INTERVAL * 10,
6755                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
6756                 if (__ratelimit(&_rs))
6757                         WARN(1, KERN_DEBUG
6758                                 "btrfs: block rsv returned %d\n", ret);
6759         }
6760 try_reserve:
6761         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6762                                      BTRFS_RESERVE_NO_FLUSH);
6763         if (!ret)
6764                 return block_rsv;
6765         /*
6766          * If we couldn't reserve metadata bytes try and use some from
6767          * the global reserve if its space type is the same as the global
6768          * reservation.
6769          */
6770         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
6771             block_rsv->space_info == global_rsv->space_info) {
6772                 ret = block_rsv_use_bytes(global_rsv, blocksize);
6773                 if (!ret)
6774                         return global_rsv;
6775         }
6776         return ERR_PTR(ret);
6777 }
6778
6779 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6780                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6781 {
6782         block_rsv_add_bytes(block_rsv, blocksize, 0);
6783         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6784 }
6785
6786 /*
6787  * finds a free extent and does all the dirty work required for allocation
6788  * returns the key for the extent through ins, and a tree buffer for
6789  * the first block of the extent through buf.
6790  *
6791  * returns the tree buffer or NULL.
6792  */
6793 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6794                                         struct btrfs_root *root, u32 blocksize,
6795                                         u64 parent, u64 root_objectid,
6796                                         struct btrfs_disk_key *key, int level,
6797                                         u64 hint, u64 empty_size)
6798 {
6799         struct btrfs_key ins;
6800         struct btrfs_block_rsv *block_rsv;
6801         struct extent_buffer *buf;
6802         u64 flags = 0;
6803         int ret;
6804         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6805                                                  SKINNY_METADATA);
6806
6807         block_rsv = use_block_rsv(trans, root, blocksize);
6808         if (IS_ERR(block_rsv))
6809                 return ERR_CAST(block_rsv);
6810
6811         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6812                                    empty_size, hint, &ins, 0);
6813         if (ret) {
6814                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6815                 return ERR_PTR(ret);
6816         }
6817
6818         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6819                                     blocksize, level);
6820         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6821
6822         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6823                 if (parent == 0)
6824                         parent = ins.objectid;
6825                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6826         } else
6827                 BUG_ON(parent > 0);
6828
6829         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6830                 struct btrfs_delayed_extent_op *extent_op;
6831                 extent_op = btrfs_alloc_delayed_extent_op();
6832                 BUG_ON(!extent_op); /* -ENOMEM */
6833                 if (key)
6834                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6835                 else
6836                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6837                 extent_op->flags_to_set = flags;
6838                 if (skinny_metadata)
6839                         extent_op->update_key = 0;
6840                 else
6841                         extent_op->update_key = 1;
6842                 extent_op->update_flags = 1;
6843                 extent_op->is_data = 0;
6844                 extent_op->level = level;
6845
6846                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6847                                         ins.objectid,
6848                                         ins.offset, parent, root_objectid,
6849                                         level, BTRFS_ADD_DELAYED_EXTENT,
6850                                         extent_op, 0);
6851                 BUG_ON(ret); /* -ENOMEM */
6852         }
6853         return buf;
6854 }
6855
6856 struct walk_control {
6857         u64 refs[BTRFS_MAX_LEVEL];
6858         u64 flags[BTRFS_MAX_LEVEL];
6859         struct btrfs_key update_progress;
6860         int stage;
6861         int level;
6862         int shared_level;
6863         int update_ref;
6864         int keep_locks;
6865         int reada_slot;
6866         int reada_count;
6867         int for_reloc;
6868 };
6869
6870 #define DROP_REFERENCE  1
6871 #define UPDATE_BACKREF  2
6872
6873 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6874                                      struct btrfs_root *root,
6875                                      struct walk_control *wc,
6876                                      struct btrfs_path *path)
6877 {
6878         u64 bytenr;
6879         u64 generation;
6880         u64 refs;
6881         u64 flags;
6882         u32 nritems;
6883         u32 blocksize;
6884         struct btrfs_key key;
6885         struct extent_buffer *eb;
6886         int ret;
6887         int slot;
6888         int nread = 0;
6889
6890         if (path->slots[wc->level] < wc->reada_slot) {
6891                 wc->reada_count = wc->reada_count * 2 / 3;
6892                 wc->reada_count = max(wc->reada_count, 2);
6893         } else {
6894                 wc->reada_count = wc->reada_count * 3 / 2;
6895                 wc->reada_count = min_t(int, wc->reada_count,
6896                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6897         }
6898
6899         eb = path->nodes[wc->level];
6900         nritems = btrfs_header_nritems(eb);
6901         blocksize = btrfs_level_size(root, wc->level - 1);
6902
6903         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6904                 if (nread >= wc->reada_count)
6905                         break;
6906
6907                 cond_resched();
6908                 bytenr = btrfs_node_blockptr(eb, slot);
6909                 generation = btrfs_node_ptr_generation(eb, slot);
6910
6911                 if (slot == path->slots[wc->level])
6912                         goto reada;
6913
6914                 if (wc->stage == UPDATE_BACKREF &&
6915                     generation <= root->root_key.offset)
6916                         continue;
6917
6918                 /* We don't lock the tree block, it's OK to be racy here */
6919                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
6920                                                wc->level - 1, 1, &refs,
6921                                                &flags);
6922                 /* We don't care about errors in readahead. */
6923                 if (ret < 0)
6924                         continue;
6925                 BUG_ON(refs == 0);
6926
6927                 if (wc->stage == DROP_REFERENCE) {
6928                         if (refs == 1)
6929                                 goto reada;
6930
6931                         if (wc->level == 1 &&
6932                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6933                                 continue;
6934                         if (!wc->update_ref ||
6935                             generation <= root->root_key.offset)
6936                                 continue;
6937                         btrfs_node_key_to_cpu(eb, &key, slot);
6938                         ret = btrfs_comp_cpu_keys(&key,
6939                                                   &wc->update_progress);
6940                         if (ret < 0)
6941                                 continue;
6942                 } else {
6943                         if (wc->level == 1 &&
6944                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6945                                 continue;
6946                 }
6947 reada:
6948                 ret = readahead_tree_block(root, bytenr, blocksize,
6949                                            generation);
6950                 if (ret)
6951                         break;
6952                 nread++;
6953         }
6954         wc->reada_slot = slot;
6955 }
6956
6957 /*
6958  * helper to process tree block while walking down the tree.
6959  *
6960  * when wc->stage == UPDATE_BACKREF, this function updates
6961  * back refs for pointers in the block.
6962  *
6963  * NOTE: return value 1 means we should stop walking down.
6964  */
6965 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6966                                    struct btrfs_root *root,
6967                                    struct btrfs_path *path,
6968                                    struct walk_control *wc, int lookup_info)
6969 {
6970         int level = wc->level;
6971         struct extent_buffer *eb = path->nodes[level];
6972         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6973         int ret;
6974
6975         if (wc->stage == UPDATE_BACKREF &&
6976             btrfs_header_owner(eb) != root->root_key.objectid)
6977                 return 1;
6978
6979         /*
6980          * when reference count of tree block is 1, it won't increase
6981          * again. once full backref flag is set, we never clear it.
6982          */
6983         if (lookup_info &&
6984             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6985              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6986                 BUG_ON(!path->locks[level]);
6987                 ret = btrfs_lookup_extent_info(trans, root,
6988                                                eb->start, level, 1,
6989                                                &wc->refs[level],
6990                                                &wc->flags[level]);
6991                 BUG_ON(ret == -ENOMEM);
6992                 if (ret)
6993                         return ret;
6994                 BUG_ON(wc->refs[level] == 0);
6995         }
6996
6997         if (wc->stage == DROP_REFERENCE) {
6998                 if (wc->refs[level] > 1)
6999                         return 1;
7000
7001                 if (path->locks[level] && !wc->keep_locks) {
7002                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7003                         path->locks[level] = 0;
7004                 }
7005                 return 0;
7006         }
7007
7008         /* wc->stage == UPDATE_BACKREF */
7009         if (!(wc->flags[level] & flag)) {
7010                 BUG_ON(!path->locks[level]);
7011                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7012                 BUG_ON(ret); /* -ENOMEM */
7013                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7014                 BUG_ON(ret); /* -ENOMEM */
7015                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7016                                                   eb->len, flag,
7017                                                   btrfs_header_level(eb), 0);
7018                 BUG_ON(ret); /* -ENOMEM */
7019                 wc->flags[level] |= flag;
7020         }
7021
7022         /*
7023          * the block is shared by multiple trees, so it's not good to
7024          * keep the tree lock
7025          */
7026         if (path->locks[level] && level > 0) {
7027                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7028                 path->locks[level] = 0;
7029         }
7030         return 0;
7031 }
7032
7033 /*
7034  * helper to process tree block pointer.
7035  *
7036  * when wc->stage == DROP_REFERENCE, this function checks
7037  * reference count of the block pointed to. if the block
7038  * is shared and we need update back refs for the subtree
7039  * rooted at the block, this function changes wc->stage to
7040  * UPDATE_BACKREF. if the block is shared and there is no
7041  * need to update back, this function drops the reference
7042  * to the block.
7043  *
7044  * NOTE: return value 1 means we should stop walking down.
7045  */
7046 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7047                                  struct btrfs_root *root,
7048                                  struct btrfs_path *path,
7049                                  struct walk_control *wc, int *lookup_info)
7050 {
7051         u64 bytenr;
7052         u64 generation;
7053         u64 parent;
7054         u32 blocksize;
7055         struct btrfs_key key;
7056         struct extent_buffer *next;
7057         int level = wc->level;
7058         int reada = 0;
7059         int ret = 0;
7060
7061         generation = btrfs_node_ptr_generation(path->nodes[level],
7062                                                path->slots[level]);
7063         /*
7064          * if the lower level block was created before the snapshot
7065          * was created, we know there is no need to update back refs
7066          * for the subtree
7067          */
7068         if (wc->stage == UPDATE_BACKREF &&
7069             generation <= root->root_key.offset) {
7070                 *lookup_info = 1;
7071                 return 1;
7072         }
7073
7074         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7075         blocksize = btrfs_level_size(root, level - 1);
7076
7077         next = btrfs_find_tree_block(root, bytenr, blocksize);
7078         if (!next) {
7079                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7080                 if (!next)
7081                         return -ENOMEM;
7082                 reada = 1;
7083         }
7084         btrfs_tree_lock(next);
7085         btrfs_set_lock_blocking(next);
7086
7087         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7088                                        &wc->refs[level - 1],
7089                                        &wc->flags[level - 1]);
7090         if (ret < 0) {
7091                 btrfs_tree_unlock(next);
7092                 return ret;
7093         }
7094
7095         if (unlikely(wc->refs[level - 1] == 0)) {
7096                 btrfs_err(root->fs_info, "Missing references.");
7097                 BUG();
7098         }
7099         *lookup_info = 0;
7100
7101         if (wc->stage == DROP_REFERENCE) {
7102                 if (wc->refs[level - 1] > 1) {
7103                         if (level == 1 &&
7104                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7105                                 goto skip;
7106
7107                         if (!wc->update_ref ||
7108                             generation <= root->root_key.offset)
7109                                 goto skip;
7110
7111                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7112                                               path->slots[level]);
7113                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7114                         if (ret < 0)
7115                                 goto skip;
7116
7117                         wc->stage = UPDATE_BACKREF;
7118                         wc->shared_level = level - 1;
7119                 }
7120         } else {
7121                 if (level == 1 &&
7122                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7123                         goto skip;
7124         }
7125
7126         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7127                 btrfs_tree_unlock(next);
7128                 free_extent_buffer(next);
7129                 next = NULL;
7130                 *lookup_info = 1;
7131         }
7132
7133         if (!next) {
7134                 if (reada && level == 1)
7135                         reada_walk_down(trans, root, wc, path);
7136                 next = read_tree_block(root, bytenr, blocksize, generation);
7137                 if (!next || !extent_buffer_uptodate(next)) {
7138                         free_extent_buffer(next);
7139                         return -EIO;
7140                 }
7141                 btrfs_tree_lock(next);
7142                 btrfs_set_lock_blocking(next);
7143         }
7144
7145         level--;
7146         BUG_ON(level != btrfs_header_level(next));
7147         path->nodes[level] = next;
7148         path->slots[level] = 0;
7149         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7150         wc->level = level;
7151         if (wc->level == 1)
7152                 wc->reada_slot = 0;
7153         return 0;
7154 skip:
7155         wc->refs[level - 1] = 0;
7156         wc->flags[level - 1] = 0;
7157         if (wc->stage == DROP_REFERENCE) {
7158                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7159                         parent = path->nodes[level]->start;
7160                 } else {
7161                         BUG_ON(root->root_key.objectid !=
7162                                btrfs_header_owner(path->nodes[level]));
7163                         parent = 0;
7164                 }
7165
7166                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7167                                 root->root_key.objectid, level - 1, 0, 0);
7168                 BUG_ON(ret); /* -ENOMEM */
7169         }
7170         btrfs_tree_unlock(next);
7171         free_extent_buffer(next);
7172         *lookup_info = 1;
7173         return 1;
7174 }
7175
7176 /*
7177  * helper to process tree block while walking up the tree.
7178  *
7179  * when wc->stage == DROP_REFERENCE, this function drops
7180  * reference count on the block.
7181  *
7182  * when wc->stage == UPDATE_BACKREF, this function changes
7183  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7184  * to UPDATE_BACKREF previously while processing the block.
7185  *
7186  * NOTE: return value 1 means we should stop walking up.
7187  */
7188 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7189                                  struct btrfs_root *root,
7190                                  struct btrfs_path *path,
7191                                  struct walk_control *wc)
7192 {
7193         int ret;
7194         int level = wc->level;
7195         struct extent_buffer *eb = path->nodes[level];
7196         u64 parent = 0;
7197
7198         if (wc->stage == UPDATE_BACKREF) {
7199                 BUG_ON(wc->shared_level < level);
7200                 if (level < wc->shared_level)
7201                         goto out;
7202
7203                 ret = find_next_key(path, level + 1, &wc->update_progress);
7204                 if (ret > 0)
7205                         wc->update_ref = 0;
7206
7207                 wc->stage = DROP_REFERENCE;
7208                 wc->shared_level = -1;
7209                 path->slots[level] = 0;
7210
7211                 /*
7212                  * check reference count again if the block isn't locked.
7213                  * we should start walking down the tree again if reference
7214                  * count is one.
7215                  */
7216                 if (!path->locks[level]) {
7217                         BUG_ON(level == 0);
7218                         btrfs_tree_lock(eb);
7219                         btrfs_set_lock_blocking(eb);
7220                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7221
7222                         ret = btrfs_lookup_extent_info(trans, root,
7223                                                        eb->start, level, 1,
7224                                                        &wc->refs[level],
7225                                                        &wc->flags[level]);
7226                         if (ret < 0) {
7227                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7228                                 path->locks[level] = 0;
7229                                 return ret;
7230                         }
7231                         BUG_ON(wc->refs[level] == 0);
7232                         if (wc->refs[level] == 1) {
7233                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7234                                 path->locks[level] = 0;
7235                                 return 1;
7236                         }
7237                 }
7238         }
7239
7240         /* wc->stage == DROP_REFERENCE */
7241         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7242
7243         if (wc->refs[level] == 1) {
7244                 if (level == 0) {
7245                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7246                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7247                                                     wc->for_reloc);
7248                         else
7249                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7250                                                     wc->for_reloc);
7251                         BUG_ON(ret); /* -ENOMEM */
7252                 }
7253                 /* make block locked assertion in clean_tree_block happy */
7254                 if (!path->locks[level] &&
7255                     btrfs_header_generation(eb) == trans->transid) {
7256                         btrfs_tree_lock(eb);
7257                         btrfs_set_lock_blocking(eb);
7258                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7259                 }
7260                 clean_tree_block(trans, root, eb);
7261         }
7262
7263         if (eb == root->node) {
7264                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7265                         parent = eb->start;
7266                 else
7267                         BUG_ON(root->root_key.objectid !=
7268                                btrfs_header_owner(eb));
7269         } else {
7270                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7271                         parent = path->nodes[level + 1]->start;
7272                 else
7273                         BUG_ON(root->root_key.objectid !=
7274                                btrfs_header_owner(path->nodes[level + 1]));
7275         }
7276
7277         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7278 out:
7279         wc->refs[level] = 0;
7280         wc->flags[level] = 0;
7281         return 0;
7282 }
7283
7284 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7285                                    struct btrfs_root *root,
7286                                    struct btrfs_path *path,
7287                                    struct walk_control *wc)
7288 {
7289         int level = wc->level;
7290         int lookup_info = 1;
7291         int ret;
7292
7293         while (level >= 0) {
7294                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7295                 if (ret > 0)
7296                         break;
7297
7298                 if (level == 0)
7299                         break;
7300
7301                 if (path->slots[level] >=
7302                     btrfs_header_nritems(path->nodes[level]))
7303                         break;
7304
7305                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7306                 if (ret > 0) {
7307                         path->slots[level]++;
7308                         continue;
7309                 } else if (ret < 0)
7310                         return ret;
7311                 level = wc->level;
7312         }
7313         return 0;
7314 }
7315
7316 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7317                                  struct btrfs_root *root,
7318                                  struct btrfs_path *path,
7319                                  struct walk_control *wc, int max_level)
7320 {
7321         int level = wc->level;
7322         int ret;
7323
7324         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7325         while (level < max_level && path->nodes[level]) {
7326                 wc->level = level;
7327                 if (path->slots[level] + 1 <
7328                     btrfs_header_nritems(path->nodes[level])) {
7329                         path->slots[level]++;
7330                         return 0;
7331                 } else {
7332                         ret = walk_up_proc(trans, root, path, wc);
7333                         if (ret > 0)
7334                                 return 0;
7335
7336                         if (path->locks[level]) {
7337                                 btrfs_tree_unlock_rw(path->nodes[level],
7338                                                      path->locks[level]);
7339                                 path->locks[level] = 0;
7340                         }
7341                         free_extent_buffer(path->nodes[level]);
7342                         path->nodes[level] = NULL;
7343                         level++;
7344                 }
7345         }
7346         return 1;
7347 }
7348
7349 /*
7350  * drop a subvolume tree.
7351  *
7352  * this function traverses the tree freeing any blocks that only
7353  * referenced by the tree.
7354  *
7355  * when a shared tree block is found. this function decreases its
7356  * reference count by one. if update_ref is true, this function
7357  * also make sure backrefs for the shared block and all lower level
7358  * blocks are properly updated.
7359  *
7360  * If called with for_reloc == 0, may exit early with -EAGAIN
7361  */
7362 int btrfs_drop_snapshot(struct btrfs_root *root,
7363                          struct btrfs_block_rsv *block_rsv, int update_ref,
7364                          int for_reloc)
7365 {
7366         struct btrfs_path *path;
7367         struct btrfs_trans_handle *trans;
7368         struct btrfs_root *tree_root = root->fs_info->tree_root;
7369         struct btrfs_root_item *root_item = &root->root_item;
7370         struct walk_control *wc;
7371         struct btrfs_key key;
7372         int err = 0;
7373         int ret;
7374         int level;
7375
7376         path = btrfs_alloc_path();
7377         if (!path) {
7378                 err = -ENOMEM;
7379                 goto out;
7380         }
7381
7382         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7383         if (!wc) {
7384                 btrfs_free_path(path);
7385                 err = -ENOMEM;
7386                 goto out;
7387         }
7388
7389         trans = btrfs_start_transaction(tree_root, 0);
7390         if (IS_ERR(trans)) {
7391                 err = PTR_ERR(trans);
7392                 goto out_free;
7393         }
7394
7395         if (block_rsv)
7396                 trans->block_rsv = block_rsv;
7397
7398         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7399                 level = btrfs_header_level(root->node);
7400                 path->nodes[level] = btrfs_lock_root_node(root);
7401                 btrfs_set_lock_blocking(path->nodes[level]);
7402                 path->slots[level] = 0;
7403                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7404                 memset(&wc->update_progress, 0,
7405                        sizeof(wc->update_progress));
7406         } else {
7407                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7408                 memcpy(&wc->update_progress, &key,
7409                        sizeof(wc->update_progress));
7410
7411                 level = root_item->drop_level;
7412                 BUG_ON(level == 0);
7413                 path->lowest_level = level;
7414                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7415                 path->lowest_level = 0;
7416                 if (ret < 0) {
7417                         err = ret;
7418                         goto out_end_trans;
7419                 }
7420                 WARN_ON(ret > 0);
7421
7422                 /*
7423                  * unlock our path, this is safe because only this
7424                  * function is allowed to delete this snapshot
7425                  */
7426                 btrfs_unlock_up_safe(path, 0);
7427
7428                 level = btrfs_header_level(root->node);
7429                 while (1) {
7430                         btrfs_tree_lock(path->nodes[level]);
7431                         btrfs_set_lock_blocking(path->nodes[level]);
7432
7433                         ret = btrfs_lookup_extent_info(trans, root,
7434                                                 path->nodes[level]->start,
7435                                                 level, 1, &wc->refs[level],
7436                                                 &wc->flags[level]);
7437                         if (ret < 0) {
7438                                 err = ret;
7439                                 goto out_end_trans;
7440                         }
7441                         BUG_ON(wc->refs[level] == 0);
7442
7443                         if (level == root_item->drop_level)
7444                                 break;
7445
7446                         btrfs_tree_unlock(path->nodes[level]);
7447                         WARN_ON(wc->refs[level] != 1);
7448                         level--;
7449                 }
7450         }
7451
7452         wc->level = level;
7453         wc->shared_level = -1;
7454         wc->stage = DROP_REFERENCE;
7455         wc->update_ref = update_ref;
7456         wc->keep_locks = 0;
7457         wc->for_reloc = for_reloc;
7458         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7459
7460         while (1) {
7461                 if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7462                         pr_debug("btrfs: drop snapshot early exit\n");
7463                         err = -EAGAIN;
7464                         goto out_end_trans;
7465                 }
7466
7467                 ret = walk_down_tree(trans, root, path, wc);
7468                 if (ret < 0) {
7469                         err = ret;
7470                         break;
7471                 }
7472
7473                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7474                 if (ret < 0) {
7475                         err = ret;
7476                         break;
7477                 }
7478
7479                 if (ret > 0) {
7480                         BUG_ON(wc->stage != DROP_REFERENCE);
7481                         break;
7482                 }
7483
7484                 if (wc->stage == DROP_REFERENCE) {
7485                         level = wc->level;
7486                         btrfs_node_key(path->nodes[level],
7487                                        &root_item->drop_progress,
7488                                        path->slots[level]);
7489                         root_item->drop_level = level;
7490                 }
7491
7492                 BUG_ON(wc->level == 0);
7493                 if (btrfs_should_end_transaction(trans, tree_root)) {
7494                         ret = btrfs_update_root(trans, tree_root,
7495                                                 &root->root_key,
7496                                                 root_item);
7497                         if (ret) {
7498                                 btrfs_abort_transaction(trans, tree_root, ret);
7499                                 err = ret;
7500                                 goto out_end_trans;
7501                         }
7502
7503                         btrfs_end_transaction_throttle(trans, tree_root);
7504                         trans = btrfs_start_transaction(tree_root, 0);
7505                         if (IS_ERR(trans)) {
7506                                 err = PTR_ERR(trans);
7507                                 goto out_free;
7508                         }
7509                         if (block_rsv)
7510                                 trans->block_rsv = block_rsv;
7511                 }
7512         }
7513         btrfs_release_path(path);
7514         if (err)
7515                 goto out_end_trans;
7516
7517         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7518         if (ret) {
7519                 btrfs_abort_transaction(trans, tree_root, ret);
7520                 goto out_end_trans;
7521         }
7522
7523         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7524                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7525                                       NULL, NULL);
7526                 if (ret < 0) {
7527                         btrfs_abort_transaction(trans, tree_root, ret);
7528                         err = ret;
7529                         goto out_end_trans;
7530                 } else if (ret > 0) {
7531                         /* if we fail to delete the orphan item this time
7532                          * around, it'll get picked up the next time.
7533                          *
7534                          * The most common failure here is just -ENOENT.
7535                          */
7536                         btrfs_del_orphan_item(trans, tree_root,
7537                                               root->root_key.objectid);
7538                 }
7539         }
7540
7541         if (root->in_radix) {
7542                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7543         } else {
7544                 free_extent_buffer(root->node);
7545                 free_extent_buffer(root->commit_root);
7546                 btrfs_put_fs_root(root);
7547         }
7548 out_end_trans:
7549         btrfs_end_transaction_throttle(trans, tree_root);
7550 out_free:
7551         kfree(wc);
7552         btrfs_free_path(path);
7553 out:
7554         if (err)
7555                 btrfs_std_error(root->fs_info, err);
7556         return err;
7557 }
7558
7559 /*
7560  * drop subtree rooted at tree block 'node'.
7561  *
7562  * NOTE: this function will unlock and release tree block 'node'
7563  * only used by relocation code
7564  */
7565 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7566                         struct btrfs_root *root,
7567                         struct extent_buffer *node,
7568                         struct extent_buffer *parent)
7569 {
7570         struct btrfs_path *path;
7571         struct walk_control *wc;
7572         int level;
7573         int parent_level;
7574         int ret = 0;
7575         int wret;
7576
7577         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7578
7579         path = btrfs_alloc_path();
7580         if (!path)
7581                 return -ENOMEM;
7582
7583         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7584         if (!wc) {
7585                 btrfs_free_path(path);
7586                 return -ENOMEM;
7587         }
7588
7589         btrfs_assert_tree_locked(parent);
7590         parent_level = btrfs_header_level(parent);
7591         extent_buffer_get(parent);
7592         path->nodes[parent_level] = parent;
7593         path->slots[parent_level] = btrfs_header_nritems(parent);
7594
7595         btrfs_assert_tree_locked(node);
7596         level = btrfs_header_level(node);
7597         path->nodes[level] = node;
7598         path->slots[level] = 0;
7599         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7600
7601         wc->refs[parent_level] = 1;
7602         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7603         wc->level = level;
7604         wc->shared_level = -1;
7605         wc->stage = DROP_REFERENCE;
7606         wc->update_ref = 0;
7607         wc->keep_locks = 1;
7608         wc->for_reloc = 1;
7609         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7610
7611         while (1) {
7612                 wret = walk_down_tree(trans, root, path, wc);
7613                 if (wret < 0) {
7614                         ret = wret;
7615                         break;
7616                 }
7617
7618                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7619                 if (wret < 0)
7620                         ret = wret;
7621                 if (wret != 0)
7622                         break;
7623         }
7624
7625         kfree(wc);
7626         btrfs_free_path(path);
7627         return ret;
7628 }
7629
7630 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7631 {
7632         u64 num_devices;
7633         u64 stripped;
7634
7635         /*
7636          * if restripe for this chunk_type is on pick target profile and
7637          * return, otherwise do the usual balance
7638          */
7639         stripped = get_restripe_target(root->fs_info, flags);
7640         if (stripped)
7641                 return extended_to_chunk(stripped);
7642
7643         /*
7644          * we add in the count of missing devices because we want
7645          * to make sure that any RAID levels on a degraded FS
7646          * continue to be honored.
7647          */
7648         num_devices = root->fs_info->fs_devices->rw_devices +
7649                 root->fs_info->fs_devices->missing_devices;
7650
7651         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7652                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7653                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7654
7655         if (num_devices == 1) {
7656                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7657                 stripped = flags & ~stripped;
7658
7659                 /* turn raid0 into single device chunks */
7660                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7661                         return stripped;
7662
7663                 /* turn mirroring into duplication */
7664                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7665                              BTRFS_BLOCK_GROUP_RAID10))
7666                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7667         } else {
7668                 /* they already had raid on here, just return */
7669                 if (flags & stripped)
7670                         return flags;
7671
7672                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7673                 stripped = flags & ~stripped;
7674
7675                 /* switch duplicated blocks with raid1 */
7676                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7677                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7678
7679                 /* this is drive concat, leave it alone */
7680         }
7681
7682         return flags;
7683 }
7684
7685 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7686 {
7687         struct btrfs_space_info *sinfo = cache->space_info;
7688         u64 num_bytes;
7689         u64 min_allocable_bytes;
7690         int ret = -ENOSPC;
7691
7692
7693         /*
7694          * We need some metadata space and system metadata space for
7695          * allocating chunks in some corner cases until we force to set
7696          * it to be readonly.
7697          */
7698         if ((sinfo->flags &
7699              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7700             !force)
7701                 min_allocable_bytes = 1 * 1024 * 1024;
7702         else
7703                 min_allocable_bytes = 0;
7704
7705         spin_lock(&sinfo->lock);
7706         spin_lock(&cache->lock);
7707
7708         if (cache->ro) {
7709                 ret = 0;
7710                 goto out;
7711         }
7712
7713         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7714                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7715
7716         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7717             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7718             min_allocable_bytes <= sinfo->total_bytes) {
7719                 sinfo->bytes_readonly += num_bytes;
7720                 cache->ro = 1;
7721                 ret = 0;
7722         }
7723 out:
7724         spin_unlock(&cache->lock);
7725         spin_unlock(&sinfo->lock);
7726         return ret;
7727 }
7728
7729 int btrfs_set_block_group_ro(struct btrfs_root *root,
7730                              struct btrfs_block_group_cache *cache)
7731
7732 {
7733         struct btrfs_trans_handle *trans;
7734         u64 alloc_flags;
7735         int ret;
7736
7737         BUG_ON(cache->ro);
7738
7739         trans = btrfs_join_transaction(root);
7740         if (IS_ERR(trans))
7741                 return PTR_ERR(trans);
7742
7743         alloc_flags = update_block_group_flags(root, cache->flags);
7744         if (alloc_flags != cache->flags) {
7745                 ret = do_chunk_alloc(trans, root, alloc_flags,
7746                                      CHUNK_ALLOC_FORCE);
7747                 if (ret < 0)
7748                         goto out;
7749         }
7750
7751         ret = set_block_group_ro(cache, 0);
7752         if (!ret)
7753                 goto out;
7754         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7755         ret = do_chunk_alloc(trans, root, alloc_flags,
7756                              CHUNK_ALLOC_FORCE);
7757         if (ret < 0)
7758                 goto out;
7759         ret = set_block_group_ro(cache, 0);
7760 out:
7761         btrfs_end_transaction(trans, root);
7762         return ret;
7763 }
7764
7765 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7766                             struct btrfs_root *root, u64 type)
7767 {
7768         u64 alloc_flags = get_alloc_profile(root, type);
7769         return do_chunk_alloc(trans, root, alloc_flags,
7770                               CHUNK_ALLOC_FORCE);
7771 }
7772
7773 /*
7774  * helper to account the unused space of all the readonly block group in the
7775  * list. takes mirrors into account.
7776  */
7777 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7778 {
7779         struct btrfs_block_group_cache *block_group;
7780         u64 free_bytes = 0;
7781         int factor;
7782
7783         list_for_each_entry(block_group, groups_list, list) {
7784                 spin_lock(&block_group->lock);
7785
7786                 if (!block_group->ro) {
7787                         spin_unlock(&block_group->lock);
7788                         continue;
7789                 }
7790
7791                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7792                                           BTRFS_BLOCK_GROUP_RAID10 |
7793                                           BTRFS_BLOCK_GROUP_DUP))
7794                         factor = 2;
7795                 else
7796                         factor = 1;
7797
7798                 free_bytes += (block_group->key.offset -
7799                                btrfs_block_group_used(&block_group->item)) *
7800                                factor;
7801
7802                 spin_unlock(&block_group->lock);
7803         }
7804
7805         return free_bytes;
7806 }
7807
7808 /*
7809  * helper to account the unused space of all the readonly block group in the
7810  * space_info. takes mirrors into account.
7811  */
7812 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7813 {
7814         int i;
7815         u64 free_bytes = 0;
7816
7817         spin_lock(&sinfo->lock);
7818
7819         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7820                 if (!list_empty(&sinfo->block_groups[i]))
7821                         free_bytes += __btrfs_get_ro_block_group_free_space(
7822                                                 &sinfo->block_groups[i]);
7823
7824         spin_unlock(&sinfo->lock);
7825
7826         return free_bytes;
7827 }
7828
7829 void btrfs_set_block_group_rw(struct btrfs_root *root,
7830                               struct btrfs_block_group_cache *cache)
7831 {
7832         struct btrfs_space_info *sinfo = cache->space_info;
7833         u64 num_bytes;
7834
7835         BUG_ON(!cache->ro);
7836
7837         spin_lock(&sinfo->lock);
7838         spin_lock(&cache->lock);
7839         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7840                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7841         sinfo->bytes_readonly -= num_bytes;
7842         cache->ro = 0;
7843         spin_unlock(&cache->lock);
7844         spin_unlock(&sinfo->lock);
7845 }
7846
7847 /*
7848  * checks to see if its even possible to relocate this block group.
7849  *
7850  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7851  * ok to go ahead and try.
7852  */
7853 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7854 {
7855         struct btrfs_block_group_cache *block_group;
7856         struct btrfs_space_info *space_info;
7857         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7858         struct btrfs_device *device;
7859         u64 min_free;
7860         u64 dev_min = 1;
7861         u64 dev_nr = 0;
7862         u64 target;
7863         int index;
7864         int full = 0;
7865         int ret = 0;
7866
7867         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7868
7869         /* odd, couldn't find the block group, leave it alone */
7870         if (!block_group)
7871                 return -1;
7872
7873         min_free = btrfs_block_group_used(&block_group->item);
7874
7875         /* no bytes used, we're good */
7876         if (!min_free)
7877                 goto out;
7878
7879         space_info = block_group->space_info;
7880         spin_lock(&space_info->lock);
7881
7882         full = space_info->full;
7883
7884         /*
7885          * if this is the last block group we have in this space, we can't
7886          * relocate it unless we're able to allocate a new chunk below.
7887          *
7888          * Otherwise, we need to make sure we have room in the space to handle
7889          * all of the extents from this block group.  If we can, we're good
7890          */
7891         if ((space_info->total_bytes != block_group->key.offset) &&
7892             (space_info->bytes_used + space_info->bytes_reserved +
7893              space_info->bytes_pinned + space_info->bytes_readonly +
7894              min_free < space_info->total_bytes)) {
7895                 spin_unlock(&space_info->lock);
7896                 goto out;
7897         }
7898         spin_unlock(&space_info->lock);
7899
7900         /*
7901          * ok we don't have enough space, but maybe we have free space on our
7902          * devices to allocate new chunks for relocation, so loop through our
7903          * alloc devices and guess if we have enough space.  if this block
7904          * group is going to be restriped, run checks against the target
7905          * profile instead of the current one.
7906          */
7907         ret = -1;
7908
7909         /*
7910          * index:
7911          *      0: raid10
7912          *      1: raid1
7913          *      2: dup
7914          *      3: raid0
7915          *      4: single
7916          */
7917         target = get_restripe_target(root->fs_info, block_group->flags);
7918         if (target) {
7919                 index = __get_raid_index(extended_to_chunk(target));
7920         } else {
7921                 /*
7922                  * this is just a balance, so if we were marked as full
7923                  * we know there is no space for a new chunk
7924                  */
7925                 if (full)
7926                         goto out;
7927
7928                 index = get_block_group_index(block_group);
7929         }
7930
7931         if (index == BTRFS_RAID_RAID10) {
7932                 dev_min = 4;
7933                 /* Divide by 2 */
7934                 min_free >>= 1;
7935         } else if (index == BTRFS_RAID_RAID1) {
7936                 dev_min = 2;
7937         } else if (index == BTRFS_RAID_DUP) {
7938                 /* Multiply by 2 */
7939                 min_free <<= 1;
7940         } else if (index == BTRFS_RAID_RAID0) {
7941                 dev_min = fs_devices->rw_devices;
7942                 do_div(min_free, dev_min);
7943         }
7944
7945         mutex_lock(&root->fs_info->chunk_mutex);
7946         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7947                 u64 dev_offset;
7948
7949                 /*
7950                  * check to make sure we can actually find a chunk with enough
7951                  * space to fit our block group in.
7952                  */
7953                 if (device->total_bytes > device->bytes_used + min_free &&
7954                     !device->is_tgtdev_for_dev_replace) {
7955                         ret = find_free_dev_extent(device, min_free,
7956                                                    &dev_offset, NULL);
7957                         if (!ret)
7958                                 dev_nr++;
7959
7960                         if (dev_nr >= dev_min)
7961                                 break;
7962
7963                         ret = -1;
7964                 }
7965         }
7966         mutex_unlock(&root->fs_info->chunk_mutex);
7967 out:
7968         btrfs_put_block_group(block_group);
7969         return ret;
7970 }
7971
7972 static int find_first_block_group(struct btrfs_root *root,
7973                 struct btrfs_path *path, struct btrfs_key *key)
7974 {
7975         int ret = 0;
7976         struct btrfs_key found_key;
7977         struct extent_buffer *leaf;
7978         int slot;
7979
7980         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7981         if (ret < 0)
7982                 goto out;
7983
7984         while (1) {
7985                 slot = path->slots[0];
7986                 leaf = path->nodes[0];
7987                 if (slot >= btrfs_header_nritems(leaf)) {
7988                         ret = btrfs_next_leaf(root, path);
7989                         if (ret == 0)
7990                                 continue;
7991                         if (ret < 0)
7992                                 goto out;
7993                         break;
7994                 }
7995                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7996
7997                 if (found_key.objectid >= key->objectid &&
7998                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7999                         ret = 0;
8000                         goto out;
8001                 }
8002                 path->slots[0]++;
8003         }
8004 out:
8005         return ret;
8006 }
8007
8008 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8009 {
8010         struct btrfs_block_group_cache *block_group;
8011         u64 last = 0;
8012
8013         while (1) {
8014                 struct inode *inode;
8015
8016                 block_group = btrfs_lookup_first_block_group(info, last);
8017                 while (block_group) {
8018                         spin_lock(&block_group->lock);
8019                         if (block_group->iref)
8020                                 break;
8021                         spin_unlock(&block_group->lock);
8022                         block_group = next_block_group(info->tree_root,
8023                                                        block_group);
8024                 }
8025                 if (!block_group) {
8026                         if (last == 0)
8027                                 break;
8028                         last = 0;
8029                         continue;
8030                 }
8031
8032                 inode = block_group->inode;
8033                 block_group->iref = 0;
8034                 block_group->inode = NULL;
8035                 spin_unlock(&block_group->lock);
8036                 iput(inode);
8037                 last = block_group->key.objectid + block_group->key.offset;
8038                 btrfs_put_block_group(block_group);
8039         }
8040 }
8041
8042 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8043 {
8044         struct btrfs_block_group_cache *block_group;
8045         struct btrfs_space_info *space_info;
8046         struct btrfs_caching_control *caching_ctl;
8047         struct rb_node *n;
8048
8049         down_write(&info->extent_commit_sem);
8050         while (!list_empty(&info->caching_block_groups)) {
8051                 caching_ctl = list_entry(info->caching_block_groups.next,
8052                                          struct btrfs_caching_control, list);
8053                 list_del(&caching_ctl->list);
8054                 put_caching_control(caching_ctl);
8055         }
8056         up_write(&info->extent_commit_sem);
8057
8058         spin_lock(&info->block_group_cache_lock);
8059         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8060                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8061                                        cache_node);
8062                 rb_erase(&block_group->cache_node,
8063                          &info->block_group_cache_tree);
8064                 spin_unlock(&info->block_group_cache_lock);
8065
8066                 down_write(&block_group->space_info->groups_sem);
8067                 list_del(&block_group->list);
8068                 up_write(&block_group->space_info->groups_sem);
8069
8070                 if (block_group->cached == BTRFS_CACHE_STARTED)
8071                         wait_block_group_cache_done(block_group);
8072
8073                 /*
8074                  * We haven't cached this block group, which means we could
8075                  * possibly have excluded extents on this block group.
8076                  */
8077                 if (block_group->cached == BTRFS_CACHE_NO)
8078                         free_excluded_extents(info->extent_root, block_group);
8079
8080                 btrfs_remove_free_space_cache(block_group);
8081                 btrfs_put_block_group(block_group);
8082
8083                 spin_lock(&info->block_group_cache_lock);
8084         }
8085         spin_unlock(&info->block_group_cache_lock);
8086
8087         /* now that all the block groups are freed, go through and
8088          * free all the space_info structs.  This is only called during
8089          * the final stages of unmount, and so we know nobody is
8090          * using them.  We call synchronize_rcu() once before we start,
8091          * just to be on the safe side.
8092          */
8093         synchronize_rcu();
8094
8095         release_global_block_rsv(info);
8096
8097         while(!list_empty(&info->space_info)) {
8098                 space_info = list_entry(info->space_info.next,
8099                                         struct btrfs_space_info,
8100                                         list);
8101                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8102                         if (space_info->bytes_pinned > 0 ||
8103                             space_info->bytes_reserved > 0 ||
8104                             space_info->bytes_may_use > 0) {
8105                                 WARN_ON(1);
8106                                 dump_space_info(space_info, 0, 0);
8107                         }
8108                 }
8109                 list_del(&space_info->list);
8110                 kfree(space_info);
8111         }
8112         return 0;
8113 }
8114
8115 static void __link_block_group(struct btrfs_space_info *space_info,
8116                                struct btrfs_block_group_cache *cache)
8117 {
8118         int index = get_block_group_index(cache);
8119
8120         down_write(&space_info->groups_sem);
8121         list_add_tail(&cache->list, &space_info->block_groups[index]);
8122         up_write(&space_info->groups_sem);
8123 }
8124
8125 int btrfs_read_block_groups(struct btrfs_root *root)
8126 {
8127         struct btrfs_path *path;
8128         int ret;
8129         struct btrfs_block_group_cache *cache;
8130         struct btrfs_fs_info *info = root->fs_info;
8131         struct btrfs_space_info *space_info;
8132         struct btrfs_key key;
8133         struct btrfs_key found_key;
8134         struct extent_buffer *leaf;
8135         int need_clear = 0;
8136         u64 cache_gen;
8137
8138         root = info->extent_root;
8139         key.objectid = 0;
8140         key.offset = 0;
8141         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8142         path = btrfs_alloc_path();
8143         if (!path)
8144                 return -ENOMEM;
8145         path->reada = 1;
8146
8147         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8148         if (btrfs_test_opt(root, SPACE_CACHE) &&
8149             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8150                 need_clear = 1;
8151         if (btrfs_test_opt(root, CLEAR_CACHE))
8152                 need_clear = 1;
8153
8154         while (1) {
8155                 ret = find_first_block_group(root, path, &key);
8156                 if (ret > 0)
8157                         break;
8158                 if (ret != 0)
8159                         goto error;
8160                 leaf = path->nodes[0];
8161                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8162                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8163                 if (!cache) {
8164                         ret = -ENOMEM;
8165                         goto error;
8166                 }
8167                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8168                                                 GFP_NOFS);
8169                 if (!cache->free_space_ctl) {
8170                         kfree(cache);
8171                         ret = -ENOMEM;
8172                         goto error;
8173                 }
8174
8175                 atomic_set(&cache->count, 1);
8176                 spin_lock_init(&cache->lock);
8177                 cache->fs_info = info;
8178                 INIT_LIST_HEAD(&cache->list);
8179                 INIT_LIST_HEAD(&cache->cluster_list);
8180
8181                 if (need_clear) {
8182                         /*
8183                          * When we mount with old space cache, we need to
8184                          * set BTRFS_DC_CLEAR and set dirty flag.
8185                          *
8186                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8187                          *    truncate the old free space cache inode and
8188                          *    setup a new one.
8189                          * b) Setting 'dirty flag' makes sure that we flush
8190                          *    the new space cache info onto disk.
8191                          */
8192                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8193                         if (btrfs_test_opt(root, SPACE_CACHE))
8194                                 cache->dirty = 1;
8195                 }
8196
8197                 read_extent_buffer(leaf, &cache->item,
8198                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8199                                    sizeof(cache->item));
8200                 memcpy(&cache->key, &found_key, sizeof(found_key));
8201
8202                 key.objectid = found_key.objectid + found_key.offset;
8203                 btrfs_release_path(path);
8204                 cache->flags = btrfs_block_group_flags(&cache->item);
8205                 cache->sectorsize = root->sectorsize;
8206                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8207                                                &root->fs_info->mapping_tree,
8208                                                found_key.objectid);
8209                 btrfs_init_free_space_ctl(cache);
8210
8211                 /*
8212                  * We need to exclude the super stripes now so that the space
8213                  * info has super bytes accounted for, otherwise we'll think
8214                  * we have more space than we actually do.
8215                  */
8216                 ret = exclude_super_stripes(root, cache);
8217                 if (ret) {
8218                         /*
8219                          * We may have excluded something, so call this just in
8220                          * case.
8221                          */
8222                         free_excluded_extents(root, cache);
8223                         kfree(cache->free_space_ctl);
8224                         kfree(cache);
8225                         goto error;
8226                 }
8227
8228                 /*
8229                  * check for two cases, either we are full, and therefore
8230                  * don't need to bother with the caching work since we won't
8231                  * find any space, or we are empty, and we can just add all
8232                  * the space in and be done with it.  This saves us _alot_ of
8233                  * time, particularly in the full case.
8234                  */
8235                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8236                         cache->last_byte_to_unpin = (u64)-1;
8237                         cache->cached = BTRFS_CACHE_FINISHED;
8238                         free_excluded_extents(root, cache);
8239                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8240                         cache->last_byte_to_unpin = (u64)-1;
8241                         cache->cached = BTRFS_CACHE_FINISHED;
8242                         add_new_free_space(cache, root->fs_info,
8243                                            found_key.objectid,
8244                                            found_key.objectid +
8245                                            found_key.offset);
8246                         free_excluded_extents(root, cache);
8247                 }
8248
8249                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8250                 if (ret) {
8251                         btrfs_remove_free_space_cache(cache);
8252                         btrfs_put_block_group(cache);
8253                         goto error;
8254                 }
8255
8256                 ret = update_space_info(info, cache->flags, found_key.offset,
8257                                         btrfs_block_group_used(&cache->item),
8258                                         &space_info);
8259                 if (ret) {
8260                         btrfs_remove_free_space_cache(cache);
8261                         spin_lock(&info->block_group_cache_lock);
8262                         rb_erase(&cache->cache_node,
8263                                  &info->block_group_cache_tree);
8264                         spin_unlock(&info->block_group_cache_lock);
8265                         btrfs_put_block_group(cache);
8266                         goto error;
8267                 }
8268
8269                 cache->space_info = space_info;
8270                 spin_lock(&cache->space_info->lock);
8271                 cache->space_info->bytes_readonly += cache->bytes_super;
8272                 spin_unlock(&cache->space_info->lock);
8273
8274                 __link_block_group(space_info, cache);
8275
8276                 set_avail_alloc_bits(root->fs_info, cache->flags);
8277                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8278                         set_block_group_ro(cache, 1);
8279         }
8280
8281         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8282                 if (!(get_alloc_profile(root, space_info->flags) &
8283                       (BTRFS_BLOCK_GROUP_RAID10 |
8284                        BTRFS_BLOCK_GROUP_RAID1 |
8285                        BTRFS_BLOCK_GROUP_RAID5 |
8286                        BTRFS_BLOCK_GROUP_RAID6 |
8287                        BTRFS_BLOCK_GROUP_DUP)))
8288                         continue;
8289                 /*
8290                  * avoid allocating from un-mirrored block group if there are
8291                  * mirrored block groups.
8292                  */
8293                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8294                         set_block_group_ro(cache, 1);
8295                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8296                         set_block_group_ro(cache, 1);
8297         }
8298
8299         init_global_block_rsv(info);
8300         ret = 0;
8301 error:
8302         btrfs_free_path(path);
8303         return ret;
8304 }
8305
8306 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8307                                        struct btrfs_root *root)
8308 {
8309         struct btrfs_block_group_cache *block_group, *tmp;
8310         struct btrfs_root *extent_root = root->fs_info->extent_root;
8311         struct btrfs_block_group_item item;
8312         struct btrfs_key key;
8313         int ret = 0;
8314
8315         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8316                                  new_bg_list) {
8317                 list_del_init(&block_group->new_bg_list);
8318
8319                 if (ret)
8320                         continue;
8321
8322                 spin_lock(&block_group->lock);
8323                 memcpy(&item, &block_group->item, sizeof(item));
8324                 memcpy(&key, &block_group->key, sizeof(key));
8325                 spin_unlock(&block_group->lock);
8326
8327                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8328                                         sizeof(item));
8329                 if (ret)
8330                         btrfs_abort_transaction(trans, extent_root, ret);
8331         }
8332 }
8333
8334 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8335                            struct btrfs_root *root, u64 bytes_used,
8336                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8337                            u64 size)
8338 {
8339         int ret;
8340         struct btrfs_root *extent_root;
8341         struct btrfs_block_group_cache *cache;
8342
8343         extent_root = root->fs_info->extent_root;
8344
8345         root->fs_info->last_trans_log_full_commit = trans->transid;
8346
8347         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8348         if (!cache)
8349                 return -ENOMEM;
8350         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8351                                         GFP_NOFS);
8352         if (!cache->free_space_ctl) {
8353                 kfree(cache);
8354                 return -ENOMEM;
8355         }
8356
8357         cache->key.objectid = chunk_offset;
8358         cache->key.offset = size;
8359         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8360         cache->sectorsize = root->sectorsize;
8361         cache->fs_info = root->fs_info;
8362         cache->full_stripe_len = btrfs_full_stripe_len(root,
8363                                                &root->fs_info->mapping_tree,
8364                                                chunk_offset);
8365
8366         atomic_set(&cache->count, 1);
8367         spin_lock_init(&cache->lock);
8368         INIT_LIST_HEAD(&cache->list);
8369         INIT_LIST_HEAD(&cache->cluster_list);
8370         INIT_LIST_HEAD(&cache->new_bg_list);
8371
8372         btrfs_init_free_space_ctl(cache);
8373
8374         btrfs_set_block_group_used(&cache->item, bytes_used);
8375         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8376         cache->flags = type;
8377         btrfs_set_block_group_flags(&cache->item, type);
8378
8379         cache->last_byte_to_unpin = (u64)-1;
8380         cache->cached = BTRFS_CACHE_FINISHED;
8381         ret = exclude_super_stripes(root, cache);
8382         if (ret) {
8383                 /*
8384                  * We may have excluded something, so call this just in
8385                  * case.
8386                  */
8387                 free_excluded_extents(root, cache);
8388                 kfree(cache->free_space_ctl);
8389                 kfree(cache);
8390                 return ret;
8391         }
8392
8393         add_new_free_space(cache, root->fs_info, chunk_offset,
8394                            chunk_offset + size);
8395
8396         free_excluded_extents(root, cache);
8397
8398         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8399         if (ret) {
8400                 btrfs_remove_free_space_cache(cache);
8401                 btrfs_put_block_group(cache);
8402                 return ret;
8403         }
8404
8405         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8406                                 &cache->space_info);
8407         if (ret) {
8408                 btrfs_remove_free_space_cache(cache);
8409                 spin_lock(&root->fs_info->block_group_cache_lock);
8410                 rb_erase(&cache->cache_node,
8411                          &root->fs_info->block_group_cache_tree);
8412                 spin_unlock(&root->fs_info->block_group_cache_lock);
8413                 btrfs_put_block_group(cache);
8414                 return ret;
8415         }
8416         update_global_block_rsv(root->fs_info);
8417
8418         spin_lock(&cache->space_info->lock);
8419         cache->space_info->bytes_readonly += cache->bytes_super;
8420         spin_unlock(&cache->space_info->lock);
8421
8422         __link_block_group(cache->space_info, cache);
8423
8424         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8425
8426         set_avail_alloc_bits(extent_root->fs_info, type);
8427
8428         return 0;
8429 }
8430
8431 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8432 {
8433         u64 extra_flags = chunk_to_extended(flags) &
8434                                 BTRFS_EXTENDED_PROFILE_MASK;
8435
8436         write_seqlock(&fs_info->profiles_lock);
8437         if (flags & BTRFS_BLOCK_GROUP_DATA)
8438                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8439         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8440                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8441         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8442                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8443         write_sequnlock(&fs_info->profiles_lock);
8444 }
8445
8446 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8447                              struct btrfs_root *root, u64 group_start)
8448 {
8449         struct btrfs_path *path;
8450         struct btrfs_block_group_cache *block_group;
8451         struct btrfs_free_cluster *cluster;
8452         struct btrfs_root *tree_root = root->fs_info->tree_root;
8453         struct btrfs_key key;
8454         struct inode *inode;
8455         int ret;
8456         int index;
8457         int factor;
8458
8459         root = root->fs_info->extent_root;
8460
8461         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8462         BUG_ON(!block_group);
8463         BUG_ON(!block_group->ro);
8464
8465         /*
8466          * Free the reserved super bytes from this block group before
8467          * remove it.
8468          */
8469         free_excluded_extents(root, block_group);
8470
8471         memcpy(&key, &block_group->key, sizeof(key));
8472         index = get_block_group_index(block_group);
8473         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8474                                   BTRFS_BLOCK_GROUP_RAID1 |
8475                                   BTRFS_BLOCK_GROUP_RAID10))
8476                 factor = 2;
8477         else
8478                 factor = 1;
8479
8480         /* make sure this block group isn't part of an allocation cluster */
8481         cluster = &root->fs_info->data_alloc_cluster;
8482         spin_lock(&cluster->refill_lock);
8483         btrfs_return_cluster_to_free_space(block_group, cluster);
8484         spin_unlock(&cluster->refill_lock);
8485
8486         /*
8487          * make sure this block group isn't part of a metadata
8488          * allocation cluster
8489          */
8490         cluster = &root->fs_info->meta_alloc_cluster;
8491         spin_lock(&cluster->refill_lock);
8492         btrfs_return_cluster_to_free_space(block_group, cluster);
8493         spin_unlock(&cluster->refill_lock);
8494
8495         path = btrfs_alloc_path();
8496         if (!path) {
8497                 ret = -ENOMEM;
8498                 goto out;
8499         }
8500
8501         inode = lookup_free_space_inode(tree_root, block_group, path);
8502         if (!IS_ERR(inode)) {
8503                 ret = btrfs_orphan_add(trans, inode);
8504                 if (ret) {
8505                         btrfs_add_delayed_iput(inode);
8506                         goto out;
8507                 }
8508                 clear_nlink(inode);
8509                 /* One for the block groups ref */
8510                 spin_lock(&block_group->lock);
8511                 if (block_group->iref) {
8512                         block_group->iref = 0;
8513                         block_group->inode = NULL;
8514                         spin_unlock(&block_group->lock);
8515                         iput(inode);
8516                 } else {
8517                         spin_unlock(&block_group->lock);
8518                 }
8519                 /* One for our lookup ref */
8520                 btrfs_add_delayed_iput(inode);
8521         }
8522
8523         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8524         key.offset = block_group->key.objectid;
8525         key.type = 0;
8526
8527         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8528         if (ret < 0)
8529                 goto out;
8530         if (ret > 0)
8531                 btrfs_release_path(path);
8532         if (ret == 0) {
8533                 ret = btrfs_del_item(trans, tree_root, path);
8534                 if (ret)
8535                         goto out;
8536                 btrfs_release_path(path);
8537         }
8538
8539         spin_lock(&root->fs_info->block_group_cache_lock);
8540         rb_erase(&block_group->cache_node,
8541                  &root->fs_info->block_group_cache_tree);
8542
8543         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8544                 root->fs_info->first_logical_byte = (u64)-1;
8545         spin_unlock(&root->fs_info->block_group_cache_lock);
8546
8547         down_write(&block_group->space_info->groups_sem);
8548         /*
8549          * we must use list_del_init so people can check to see if they
8550          * are still on the list after taking the semaphore
8551          */
8552         list_del_init(&block_group->list);
8553         if (list_empty(&block_group->space_info->block_groups[index]))
8554                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8555         up_write(&block_group->space_info->groups_sem);
8556
8557         if (block_group->cached == BTRFS_CACHE_STARTED)
8558                 wait_block_group_cache_done(block_group);
8559
8560         btrfs_remove_free_space_cache(block_group);
8561
8562         spin_lock(&block_group->space_info->lock);
8563         block_group->space_info->total_bytes -= block_group->key.offset;
8564         block_group->space_info->bytes_readonly -= block_group->key.offset;
8565         block_group->space_info->disk_total -= block_group->key.offset * factor;
8566         spin_unlock(&block_group->space_info->lock);
8567
8568         memcpy(&key, &block_group->key, sizeof(key));
8569
8570         btrfs_clear_space_info_full(root->fs_info);
8571
8572         btrfs_put_block_group(block_group);
8573         btrfs_put_block_group(block_group);
8574
8575         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8576         if (ret > 0)
8577                 ret = -EIO;
8578         if (ret < 0)
8579                 goto out;
8580
8581         ret = btrfs_del_item(trans, root, path);
8582 out:
8583         btrfs_free_path(path);
8584         return ret;
8585 }
8586
8587 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8588 {
8589         struct btrfs_space_info *space_info;
8590         struct btrfs_super_block *disk_super;
8591         u64 features;
8592         u64 flags;
8593         int mixed = 0;
8594         int ret;
8595
8596         disk_super = fs_info->super_copy;
8597         if (!btrfs_super_root(disk_super))
8598                 return 1;
8599
8600         features = btrfs_super_incompat_flags(disk_super);
8601         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8602                 mixed = 1;
8603
8604         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8605         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8606         if (ret)
8607                 goto out;
8608
8609         if (mixed) {
8610                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8611                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8612         } else {
8613                 flags = BTRFS_BLOCK_GROUP_METADATA;
8614                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8615                 if (ret)
8616                         goto out;
8617
8618                 flags = BTRFS_BLOCK_GROUP_DATA;
8619                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8620         }
8621 out:
8622         return ret;
8623 }
8624
8625 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8626 {
8627         return unpin_extent_range(root, start, end);
8628 }
8629
8630 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8631                                u64 num_bytes, u64 *actual_bytes)
8632 {
8633         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8634 }
8635
8636 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8637 {
8638         struct btrfs_fs_info *fs_info = root->fs_info;
8639         struct btrfs_block_group_cache *cache = NULL;
8640         u64 group_trimmed;
8641         u64 start;
8642         u64 end;
8643         u64 trimmed = 0;
8644         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8645         int ret = 0;
8646
8647         /*
8648          * try to trim all FS space, our block group may start from non-zero.
8649          */
8650         if (range->len == total_bytes)
8651                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8652         else
8653                 cache = btrfs_lookup_block_group(fs_info, range->start);
8654
8655         while (cache) {
8656                 if (cache->key.objectid >= (range->start + range->len)) {
8657                         btrfs_put_block_group(cache);
8658                         break;
8659                 }
8660
8661                 start = max(range->start, cache->key.objectid);
8662                 end = min(range->start + range->len,
8663                                 cache->key.objectid + cache->key.offset);
8664
8665                 if (end - start >= range->minlen) {
8666                         if (!block_group_cache_done(cache)) {
8667                                 ret = cache_block_group(cache, 0);
8668                                 if (!ret)
8669                                         wait_block_group_cache_done(cache);
8670                         }
8671                         ret = btrfs_trim_block_group(cache,
8672                                                      &group_trimmed,
8673                                                      start,
8674                                                      end,
8675                                                      range->minlen);
8676
8677                         trimmed += group_trimmed;
8678                         if (ret) {
8679                                 btrfs_put_block_group(cache);
8680                                 break;
8681                         }
8682                 }
8683
8684                 cache = next_block_group(fs_info->tree_root, cache);
8685         }
8686
8687         range->len = trimmed;
8688         return ret;
8689 }