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