fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "ctree.h"
  20 #include "extent-tree.h"
  21 #include "tree-log.h"
  22 #include "disk-io.h"
  23 #include "print-tree.h"
  24 #include "volumes.h"
  25 #include "raid56.h"
  26 #include "locking.h"
  27 #include "free-space-cache.h"
  28 #include "free-space-tree.h"
  29 #include "sysfs.h"
  30 #include "qgroup.h"
  31 #include "ref-verify.h"
  32 #include "space-info.h"
  33 #include "block-rsv.h"
  34 #include "delalloc-space.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "root-tree.h"
  42 #include "file-item.h"
  43 #include "orphan.h"
  44 #include "tree-checker.h"
  45 #include "raid-stripe-tree.h"
  46
  47 #undef SCRAMBLE_DELAYED_REFS
  48
  49
  50 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  51                                struct btrfs_delayed_ref_head *href,
  52                                struct btrfs_delayed_ref_node *node, u64 parent,
  53                                u64 root_objectid, u64 owner_objectid,
  54                                u64 owner_offset,
  55                                struct btrfs_delayed_extent_op *extra_op);
  56 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  57                                     struct extent_buffer *leaf,
  58                                     struct btrfs_extent_item *ei);
  59 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  60                                       u64 parent, u64 root_objectid,
  61                                       u64 flags, u64 owner, u64 offset,
  62                                       struct btrfs_key *ins, int ref_mod, u64 oref_root);
  63 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  64                                      struct btrfs_delayed_ref_node *node,
  65                                      struct btrfs_delayed_extent_op *extent_op);
  66 static int find_next_key(struct btrfs_path *path, int level,
  67                          struct btrfs_key *key);
  68
  69 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  70 {
  71         return (cache->flags & bits) == bits;
  72 }
  73
  74 /* simple helper to search for an existing data extent at a given offset */
  75 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  76 {
  77         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  78         int ret;
  79         struct btrfs_key key;
  80         struct btrfs_path *path;
  81
  82         path = btrfs_alloc_path();
  83         if (!path)
  84                 return -ENOMEM;
  85
  86         key.objectid = start;
  87         key.offset = len;
  88         key.type = BTRFS_EXTENT_ITEM_KEY;
  89         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  90         btrfs_free_path(path);
  91         return ret;
  92 }
  93
  94 /*
  95  * helper function to lookup reference count and flags of a tree block.
  96  *
  97  * the head node for delayed ref is used to store the sum of all the
  98  * reference count modifications queued up in the rbtree. the head
  99  * node may also store the extent flags to set. This way you can check
 100  * to see what the reference count and extent flags would be if all of
 101  * the delayed refs are not processed.
 102  */
 103 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 104                              struct btrfs_fs_info *fs_info, u64 bytenr,
 105                              u64 offset, int metadata, u64 *refs, u64 *flags,
 106                              u64 *owning_root)
 107 {
 108         struct btrfs_root *extent_root;
 109         struct btrfs_delayed_ref_head *head;
 110         struct btrfs_delayed_ref_root *delayed_refs;
 111         struct btrfs_path *path;
 112         struct btrfs_extent_item *ei;
 113         struct extent_buffer *leaf;
 114         struct btrfs_key key;
 115         u32 item_size;
 116         u64 num_refs;
 117         u64 extent_flags;
 118         u64 owner = 0;
 119         int ret;
 120
 121         /*
 122          * If we don't have skinny metadata, don't bother doing anything
 123          * different
 124          */
 125         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 126                 offset = fs_info->nodesize;
 127                 metadata = 0;
 128         }
 129
 130         path = btrfs_alloc_path();
 131         if (!path)
 132                 return -ENOMEM;
 133
 134         if (!trans) {
 135                 path->skip_locking = 1;
 136                 path->search_commit_root = 1;
 137         }
 138
 139 search_again:
 140         key.objectid = bytenr;
 141         key.offset = offset;
 142         if (metadata)
 143                 key.type = BTRFS_METADATA_ITEM_KEY;
 144         else
 145                 key.type = BTRFS_EXTENT_ITEM_KEY;
 146
 147         extent_root = btrfs_extent_root(fs_info, bytenr);
 148         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 149         if (ret < 0)
 150                 goto out_free;
 151
 152         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 153                 if (path->slots[0]) {
 154                         path->slots[0]--;
 155                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 156                                               path->slots[0]);
 157                         if (key.objectid == bytenr &&
 158                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 159                             key.offset == fs_info->nodesize)
 160                                 ret = 0;
 161                 }
 162         }
 163
 164         if (ret == 0) {
 165                 leaf = path->nodes[0];
 166                 item_size = btrfs_item_size(leaf, path->slots[0]);
 167                 if (item_size >= sizeof(*ei)) {
 168                         ei = btrfs_item_ptr(leaf, path->slots[0],
 169                                             struct btrfs_extent_item);
 170                         num_refs = btrfs_extent_refs(leaf, ei);
 171                         extent_flags = btrfs_extent_flags(leaf, ei);
 172                         owner = btrfs_get_extent_owner_root(fs_info, leaf,
 173                                                             path->slots[0]);
 174                 } else {
 175                         ret = -EUCLEAN;
 176                         btrfs_err(fs_info,
 177                         "unexpected extent item size, has %u expect >= %zu",
 178                                   item_size, sizeof(*ei));
 179                         if (trans)
 180                                 btrfs_abort_transaction(trans, ret);
 181                         else
 182                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 183
 184                         goto out_free;
 185                 }
 186
 187                 BUG_ON(num_refs == 0);
 188         } else {
 189                 num_refs = 0;
 190                 extent_flags = 0;
 191                 ret = 0;
 192         }
 193
 194         if (!trans)
 195                 goto out;
 196
 197         delayed_refs = &trans->transaction->delayed_refs;
 198         spin_lock(&delayed_refs->lock);
 199         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 200         if (head) {
 201                 if (!mutex_trylock(&head->mutex)) {
 202                         refcount_inc(&head->refs);
 203                         spin_unlock(&delayed_refs->lock);
 204
 205                         btrfs_release_path(path);
 206
 207                         /*
 208                          * Mutex was contended, block until it's released and try
 209                          * again
 210                          */
 211                         mutex_lock(&head->mutex);
 212                         mutex_unlock(&head->mutex);
 213                         btrfs_put_delayed_ref_head(head);
 214                         goto search_again;
 215                 }
 216                 spin_lock(&head->lock);
 217                 if (head->extent_op && head->extent_op->update_flags)
 218                         extent_flags |= head->extent_op->flags_to_set;
 219                 else
 220                         BUG_ON(num_refs == 0);
 221
 222                 num_refs += head->ref_mod;
 223                 spin_unlock(&head->lock);
 224                 mutex_unlock(&head->mutex);
 225         }
 226         spin_unlock(&delayed_refs->lock);
 227 out:
 228         WARN_ON(num_refs == 0);
 229         if (refs)
 230                 *refs = num_refs;
 231         if (flags)
 232                 *flags = extent_flags;
 233         if (owning_root)
 234                 *owning_root = owner;
 235 out_free:
 236         btrfs_free_path(path);
 237         return ret;
 238 }
 239
 240 /*
 241  * Back reference rules.  Back refs have three main goals:
 242  *
 243  * 1) differentiate between all holders of references to an extent so that
 244  *    when a reference is dropped we can make sure it was a valid reference
 245  *    before freeing the extent.
 246  *
 247  * 2) Provide enough information to quickly find the holders of an extent
 248  *    if we notice a given block is corrupted or bad.
 249  *
 250  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 251  *    maintenance.  This is actually the same as #2, but with a slightly
 252  *    different use case.
 253  *
 254  * There are two kinds of back refs. The implicit back refs is optimized
 255  * for pointers in non-shared tree blocks. For a given pointer in a block,
 256  * back refs of this kind provide information about the block's owner tree
 257  * and the pointer's key. These information allow us to find the block by
 258  * b-tree searching. The full back refs is for pointers in tree blocks not
 259  * referenced by their owner trees. The location of tree block is recorded
 260  * in the back refs. Actually the full back refs is generic, and can be
 261  * used in all cases the implicit back refs is used. The major shortcoming
 262  * of the full back refs is its overhead. Every time a tree block gets
 263  * COWed, we have to update back refs entry for all pointers in it.
 264  *
 265  * For a newly allocated tree block, we use implicit back refs for
 266  * pointers in it. This means most tree related operations only involve
 267  * implicit back refs. For a tree block created in old transaction, the
 268  * only way to drop a reference to it is COW it. So we can detect the
 269  * event that tree block loses its owner tree's reference and do the
 270  * back refs conversion.
 271  *
 272  * When a tree block is COWed through a tree, there are four cases:
 273  *
 274  * The reference count of the block is one and the tree is the block's
 275  * owner tree. Nothing to do in this case.
 276  *
 277  * The reference count of the block is one and the tree is not the
 278  * block's owner tree. In this case, full back refs is used for pointers
 279  * in the block. Remove these full back refs, add implicit back refs for
 280  * every pointers in the new block.
 281  *
 282  * The reference count of the block is greater than one and the tree is
 283  * the block's owner tree. In this case, implicit back refs is used for
 284  * pointers in the block. Add full back refs for every pointers in the
 285  * block, increase lower level extents' reference counts. The original
 286  * implicit back refs are entailed to the new block.
 287  *
 288  * The reference count of the block is greater than one and the tree is
 289  * not the block's owner tree. Add implicit back refs for every pointer in
 290  * the new block, increase lower level extents' reference count.
 291  *
 292  * Back Reference Key composing:
 293  *
 294  * The key objectid corresponds to the first byte in the extent,
 295  * The key type is used to differentiate between types of back refs.
 296  * There are different meanings of the key offset for different types
 297  * of back refs.
 298  *
 299  * File extents can be referenced by:
 300  *
 301  * - multiple snapshots, subvolumes, or different generations in one subvol
 302  * - different files inside a single subvolume
 303  * - different offsets inside a file (bookend extents in file.c)
 304  *
 305  * The extent ref structure for the implicit back refs has fields for:
 306  *
 307  * - Objectid of the subvolume root
 308  * - objectid of the file holding the reference
 309  * - original offset in the file
 310  * - how many bookend extents
 311  *
 312  * The key offset for the implicit back refs is hash of the first
 313  * three fields.
 314  *
 315  * The extent ref structure for the full back refs has field for:
 316  *
 317  * - number of pointers in the tree leaf
 318  *
 319  * The key offset for the implicit back refs is the first byte of
 320  * the tree leaf
 321  *
 322  * When a file extent is allocated, The implicit back refs is used.
 323  * the fields are filled in:
 324  *
 325  *     (root_key.objectid, inode objectid, offset in file, 1)
 326  *
 327  * When a file extent is removed file truncation, we find the
 328  * corresponding implicit back refs and check the following fields:
 329  *
 330  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 331  *
 332  * Btree extents can be referenced by:
 333  *
 334  * - Different subvolumes
 335  *
 336  * Both the implicit back refs and the full back refs for tree blocks
 337  * only consist of key. The key offset for the implicit back refs is
 338  * objectid of block's owner tree. The key offset for the full back refs
 339  * is the first byte of parent block.
 340  *
 341  * When implicit back refs is used, information about the lowest key and
 342  * level of the tree block are required. These information are stored in
 343  * tree block info structure.
 344  */
 345
 346 /*
 347  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 348  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 349  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 350  */
 351 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 352                                      struct btrfs_extent_inline_ref *iref,
 353                                      enum btrfs_inline_ref_type is_data)
 354 {
 355         struct btrfs_fs_info *fs_info = eb->fs_info;
 356         int type = btrfs_extent_inline_ref_type(eb, iref);
 357         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 358
 359         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 360                 ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 361                 return type;
 362         }
 363
 364         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 365             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 366             type == BTRFS_SHARED_DATA_REF_KEY ||
 367             type == BTRFS_EXTENT_DATA_REF_KEY) {
 368                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 369                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 370                                 return type;
 371                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 372                                 ASSERT(fs_info);
 373                                 /*
 374                                  * Every shared one has parent tree block,
 375                                  * which must be aligned to sector size.
 376                                  */
 377                                 if (offset && IS_ALIGNED(offset, fs_info->sectorsize))
 378                                         return type;
 379                         }
 380                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 381                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 382                                 return type;
 383                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 384                                 ASSERT(fs_info);
 385                                 /*
 386                                  * Every shared one has parent tree block,
 387                                  * which must be aligned to sector size.
 388                                  */
 389                                 if (offset &&
 390                                     IS_ALIGNED(offset, fs_info->sectorsize))
 391                                         return type;
 392                         }
 393                 } else {
 394                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 395                         return type;
 396                 }
 397         }
 398
 399         WARN_ON(1);
 400         btrfs_print_leaf(eb);
 401         btrfs_err(fs_info,
 402                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 403                   eb->start, (unsigned long)iref, type);
 404
 405         return BTRFS_REF_TYPE_INVALID;
 406 }
 407
 408 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 409 {
 410         u32 high_crc = ~(u32)0;
 411         u32 low_crc = ~(u32)0;
 412         __le64 lenum;
 413
 414         lenum = cpu_to_le64(root_objectid);
 415         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 416         lenum = cpu_to_le64(owner);
 417         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 418         lenum = cpu_to_le64(offset);
 419         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 420
 421         return ((u64)high_crc << 31) ^ (u64)low_crc;
 422 }
 423
 424 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 425                                      struct btrfs_extent_data_ref *ref)
 426 {
 427         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 428                                     btrfs_extent_data_ref_objectid(leaf, ref),
 429                                     btrfs_extent_data_ref_offset(leaf, ref));
 430 }
 431
 432 static int match_extent_data_ref(struct extent_buffer *leaf,
 433                                  struct btrfs_extent_data_ref *ref,
 434                                  u64 root_objectid, u64 owner, u64 offset)
 435 {
 436         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 437             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 438             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 439                 return 0;
 440         return 1;
 441 }
 442
 443 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 444                                            struct btrfs_path *path,
 445                                            u64 bytenr, u64 parent,
 446                                            u64 root_objectid,
 447                                            u64 owner, u64 offset)
 448 {
 449         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 450         struct btrfs_key key;
 451         struct btrfs_extent_data_ref *ref;
 452         struct extent_buffer *leaf;
 453         u32 nritems;
 454         int ret;
 455         int recow;
 456         int err = -ENOENT;
 457
 458         key.objectid = bytenr;
 459         if (parent) {
 460                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 461                 key.offset = parent;
 462         } else {
 463                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 464                 key.offset = hash_extent_data_ref(root_objectid,
 465                                                   owner, offset);
 466         }
 467 again:
 468         recow = 0;
 469         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 470         if (ret < 0) {
 471                 err = ret;
 472                 goto fail;
 473         }
 474
 475         if (parent) {
 476                 if (!ret)
 477                         return 0;
 478                 goto fail;
 479         }
 480
 481         leaf = path->nodes[0];
 482         nritems = btrfs_header_nritems(leaf);
 483         while (1) {
 484                 if (path->slots[0] >= nritems) {
 485                         ret = btrfs_next_leaf(root, path);
 486                         if (ret < 0)
 487                                 err = ret;
 488                         if (ret)
 489                                 goto fail;
 490
 491                         leaf = path->nodes[0];
 492                         nritems = btrfs_header_nritems(leaf);
 493                         recow = 1;
 494                 }
 495
 496                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 497                 if (key.objectid != bytenr ||
 498                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 499                         goto fail;
 500
 501                 ref = btrfs_item_ptr(leaf, path->slots[0],
 502                                      struct btrfs_extent_data_ref);
 503
 504                 if (match_extent_data_ref(leaf, ref, root_objectid,
 505                                           owner, offset)) {
 506                         if (recow) {
 507                                 btrfs_release_path(path);
 508                                 goto again;
 509                         }
 510                         err = 0;
 511                         break;
 512                 }
 513                 path->slots[0]++;
 514         }
 515 fail:
 516         return err;
 517 }
 518
 519 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 520                                            struct btrfs_path *path,
 521                                            u64 bytenr, u64 parent,
 522                                            u64 root_objectid, u64 owner,
 523                                            u64 offset, int refs_to_add)
 524 {
 525         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 526         struct btrfs_key key;
 527         struct extent_buffer *leaf;
 528         u32 size;
 529         u32 num_refs;
 530         int ret;
 531
 532         key.objectid = bytenr;
 533         if (parent) {
 534                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 535                 key.offset = parent;
 536                 size = sizeof(struct btrfs_shared_data_ref);
 537         } else {
 538                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 539                 key.offset = hash_extent_data_ref(root_objectid,
 540                                                   owner, offset);
 541                 size = sizeof(struct btrfs_extent_data_ref);
 542         }
 543
 544         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 545         if (ret && ret != -EEXIST)
 546                 goto fail;
 547
 548         leaf = path->nodes[0];
 549         if (parent) {
 550                 struct btrfs_shared_data_ref *ref;
 551                 ref = btrfs_item_ptr(leaf, path->slots[0],
 552                                      struct btrfs_shared_data_ref);
 553                 if (ret == 0) {
 554                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 555                 } else {
 556                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 557                         num_refs += refs_to_add;
 558                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 559                 }
 560         } else {
 561                 struct btrfs_extent_data_ref *ref;
 562                 while (ret == -EEXIST) {
 563                         ref = btrfs_item_ptr(leaf, path->slots[0],
 564                                              struct btrfs_extent_data_ref);
 565                         if (match_extent_data_ref(leaf, ref, root_objectid,
 566                                                   owner, offset))
 567                                 break;
 568                         btrfs_release_path(path);
 569                         key.offset++;
 570                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 571                                                       size);
 572                         if (ret && ret != -EEXIST)
 573                                 goto fail;
 574
 575                         leaf = path->nodes[0];
 576                 }
 577                 ref = btrfs_item_ptr(leaf, path->slots[0],
 578                                      struct btrfs_extent_data_ref);
 579                 if (ret == 0) {
 580                         btrfs_set_extent_data_ref_root(leaf, ref,
 581                                                        root_objectid);
 582                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 583                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 584                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 585                 } else {
 586                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 587                         num_refs += refs_to_add;
 588                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 589                 }
 590         }
 591         btrfs_mark_buffer_dirty(trans, leaf);
 592         ret = 0;
 593 fail:
 594         btrfs_release_path(path);
 595         return ret;
 596 }
 597
 598 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 599                                            struct btrfs_root *root,
 600                                            struct btrfs_path *path,
 601                                            int refs_to_drop)
 602 {
 603         struct btrfs_key key;
 604         struct btrfs_extent_data_ref *ref1 = NULL;
 605         struct btrfs_shared_data_ref *ref2 = NULL;
 606         struct extent_buffer *leaf;
 607         u32 num_refs = 0;
 608         int ret = 0;
 609
 610         leaf = path->nodes[0];
 611         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 612
 613         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 614                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 615                                       struct btrfs_extent_data_ref);
 616                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 617         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 618                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 619                                       struct btrfs_shared_data_ref);
 620                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 621         } else {
 622                 btrfs_err(trans->fs_info,
 623                           "unrecognized backref key (%llu %u %llu)",
 624                           key.objectid, key.type, key.offset);
 625                 btrfs_abort_transaction(trans, -EUCLEAN);
 626                 return -EUCLEAN;
 627         }
 628
 629         BUG_ON(num_refs < refs_to_drop);
 630         num_refs -= refs_to_drop;
 631
 632         if (num_refs == 0) {
 633                 ret = btrfs_del_item(trans, root, path);
 634         } else {
 635                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 636                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 637                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 638                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 639                 btrfs_mark_buffer_dirty(trans, leaf);
 640         }
 641         return ret;
 642 }
 643
 644 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 645                                           struct btrfs_extent_inline_ref *iref)
 646 {
 647         struct btrfs_key key;
 648         struct extent_buffer *leaf;
 649         struct btrfs_extent_data_ref *ref1;
 650         struct btrfs_shared_data_ref *ref2;
 651         u32 num_refs = 0;
 652         int type;
 653
 654         leaf = path->nodes[0];
 655         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 656
 657         if (iref) {
 658                 /*
 659                  * If type is invalid, we should have bailed out earlier than
 660                  * this call.
 661                  */
 662                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 663                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 664                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 665                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 666                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 667                 } else {
 668                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 669                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 670                 }
 671         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 672                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 673                                       struct btrfs_extent_data_ref);
 674                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 675         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 676                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 677                                       struct btrfs_shared_data_ref);
 678                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 679         } else {
 680                 WARN_ON(1);
 681         }
 682         return num_refs;
 683 }
 684
 685 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 686                                           struct btrfs_path *path,
 687                                           u64 bytenr, u64 parent,
 688                                           u64 root_objectid)
 689 {
 690         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 691         struct btrfs_key key;
 692         int ret;
 693
 694         key.objectid = bytenr;
 695         if (parent) {
 696                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 697                 key.offset = parent;
 698         } else {
 699                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 700                 key.offset = root_objectid;
 701         }
 702
 703         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 704         if (ret > 0)
 705                 ret = -ENOENT;
 706         return ret;
 707 }
 708
 709 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 710                                           struct btrfs_path *path,
 711                                           u64 bytenr, u64 parent,
 712                                           u64 root_objectid)
 713 {
 714         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 715         struct btrfs_key key;
 716         int ret;
 717
 718         key.objectid = bytenr;
 719         if (parent) {
 720                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 721                 key.offset = parent;
 722         } else {
 723                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 724                 key.offset = root_objectid;
 725         }
 726
 727         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 728         btrfs_release_path(path);
 729         return ret;
 730 }
 731
 732 static inline int extent_ref_type(u64 parent, u64 owner)
 733 {
 734         int type;
 735         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 736                 if (parent > 0)
 737                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 738                 else
 739                         type = BTRFS_TREE_BLOCK_REF_KEY;
 740         } else {
 741                 if (parent > 0)
 742                         type = BTRFS_SHARED_DATA_REF_KEY;
 743                 else
 744                         type = BTRFS_EXTENT_DATA_REF_KEY;
 745         }
 746         return type;
 747 }
 748
 749 static int find_next_key(struct btrfs_path *path, int level,
 750                          struct btrfs_key *key)
 751
 752 {
 753         for (; level < BTRFS_MAX_LEVEL; level++) {
 754                 if (!path->nodes[level])
 755                         break;
 756                 if (path->slots[level] + 1 >=
 757                     btrfs_header_nritems(path->nodes[level]))
 758                         continue;
 759                 if (level == 0)
 760                         btrfs_item_key_to_cpu(path->nodes[level], key,
 761                                               path->slots[level] + 1);
 762                 else
 763                         btrfs_node_key_to_cpu(path->nodes[level], key,
 764                                               path->slots[level] + 1);
 765                 return 0;
 766         }
 767         return 1;
 768 }
 769
 770 /*
 771  * look for inline back ref. if back ref is found, *ref_ret is set
 772  * to the address of inline back ref, and 0 is returned.
 773  *
 774  * if back ref isn't found, *ref_ret is set to the address where it
 775  * should be inserted, and -ENOENT is returned.
 776  *
 777  * if insert is true and there are too many inline back refs, the path
 778  * points to the extent item, and -EAGAIN is returned.
 779  *
 780  * NOTE: inline back refs are ordered in the same way that back ref
 781  *       items in the tree are ordered.
 782  */
 783 static noinline_for_stack
 784 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 785                                  struct btrfs_path *path,
 786                                  struct btrfs_extent_inline_ref **ref_ret,
 787                                  u64 bytenr, u64 num_bytes,
 788                                  u64 parent, u64 root_objectid,
 789                                  u64 owner, u64 offset, int insert)
 790 {
 791         struct btrfs_fs_info *fs_info = trans->fs_info;
 792         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 793         struct btrfs_key key;
 794         struct extent_buffer *leaf;
 795         struct btrfs_extent_item *ei;
 796         struct btrfs_extent_inline_ref *iref;
 797         u64 flags;
 798         u64 item_size;
 799         unsigned long ptr;
 800         unsigned long end;
 801         int extra_size;
 802         int type;
 803         int want;
 804         int ret;
 805         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 806         int needed;
 807
 808         key.objectid = bytenr;
 809         key.type = BTRFS_EXTENT_ITEM_KEY;
 810         key.offset = num_bytes;
 811
 812         want = extent_ref_type(parent, owner);
 813         if (insert) {
 814                 extra_size = btrfs_extent_inline_ref_size(want);
 815                 path->search_for_extension = 1;
 816                 path->keep_locks = 1;
 817         } else
 818                 extra_size = -1;
 819
 820         /*
 821          * Owner is our level, so we can just add one to get the level for the
 822          * block we are interested in.
 823          */
 824         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 825                 key.type = BTRFS_METADATA_ITEM_KEY;
 826                 key.offset = owner;
 827         }
 828
 829 again:
 830         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 831         if (ret < 0)
 832                 goto out;
 833
 834         /*
 835          * We may be a newly converted file system which still has the old fat
 836          * extent entries for metadata, so try and see if we have one of those.
 837          */
 838         if (ret > 0 && skinny_metadata) {
 839                 skinny_metadata = false;
 840                 if (path->slots[0]) {
 841                         path->slots[0]--;
 842                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 843                                               path->slots[0]);
 844                         if (key.objectid == bytenr &&
 845                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 846                             key.offset == num_bytes)
 847                                 ret = 0;
 848                 }
 849                 if (ret) {
 850                         key.objectid = bytenr;
 851                         key.type = BTRFS_EXTENT_ITEM_KEY;
 852                         key.offset = num_bytes;
 853                         btrfs_release_path(path);
 854                         goto again;
 855                 }
 856         }
 857
 858         if (ret && !insert) {
 859                 ret = -ENOENT;
 860                 goto out;
 861         } else if (WARN_ON(ret)) {
 862                 btrfs_print_leaf(path->nodes[0]);
 863                 btrfs_err(fs_info,
 864 "extent item not found for insert, bytenr %llu num_bytes %llu parent %llu root_objectid %llu owner %llu offset %llu",
 865                           bytenr, num_bytes, parent, root_objectid, owner,
 866                           offset);
 867                 ret = -EUCLEAN;
 868                 goto out;
 869         }
 870
 871         leaf = path->nodes[0];
 872         item_size = btrfs_item_size(leaf, path->slots[0]);
 873         if (unlikely(item_size < sizeof(*ei))) {
 874                 ret = -EUCLEAN;
 875                 btrfs_err(fs_info,
 876                           "unexpected extent item size, has %llu expect >= %zu",
 877                           item_size, sizeof(*ei));
 878                 btrfs_abort_transaction(trans, ret);
 879                 goto out;
 880         }
 881
 882         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 883         flags = btrfs_extent_flags(leaf, ei);
 884
 885         ptr = (unsigned long)(ei + 1);
 886         end = (unsigned long)ei + item_size;
 887
 888         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 889                 ptr += sizeof(struct btrfs_tree_block_info);
 890                 BUG_ON(ptr > end);
 891         }
 892
 893         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 894                 needed = BTRFS_REF_TYPE_DATA;
 895         else
 896                 needed = BTRFS_REF_TYPE_BLOCK;
 897
 898         ret = -ENOENT;
 899         while (ptr < end) {
 900                 iref = (struct btrfs_extent_inline_ref *)ptr;
 901                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 902                 if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 903                         ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 904                         ptr += btrfs_extent_inline_ref_size(type);
 905                         continue;
 906                 }
 907                 if (type == BTRFS_REF_TYPE_INVALID) {
 908                         ret = -EUCLEAN;
 909                         goto out;
 910                 }
 911
 912                 if (want < type)
 913                         break;
 914                 if (want > type) {
 915                         ptr += btrfs_extent_inline_ref_size(type);
 916                         continue;
 917                 }
 918
 919                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 920                         struct btrfs_extent_data_ref *dref;
 921                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 922                         if (match_extent_data_ref(leaf, dref, root_objectid,
 923                                                   owner, offset)) {
 924                                 ret = 0;
 925                                 break;
 926                         }
 927                         if (hash_extent_data_ref_item(leaf, dref) <
 928                             hash_extent_data_ref(root_objectid, owner, offset))
 929                                 break;
 930                 } else {
 931                         u64 ref_offset;
 932                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 933                         if (parent > 0) {
 934                                 if (parent == ref_offset) {
 935                                         ret = 0;
 936                                         break;
 937                                 }
 938                                 if (ref_offset < parent)
 939                                         break;
 940                         } else {
 941                                 if (root_objectid == ref_offset) {
 942                                         ret = 0;
 943                                         break;
 944                                 }
 945                                 if (ref_offset < root_objectid)
 946                                         break;
 947                         }
 948                 }
 949                 ptr += btrfs_extent_inline_ref_size(type);
 950         }
 951
 952         if (unlikely(ptr > end)) {
 953                 ret = -EUCLEAN;
 954                 btrfs_print_leaf(path->nodes[0]);
 955                 btrfs_crit(fs_info,
 956 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 957                            path->slots[0], root_objectid, owner, offset, parent);
 958                 goto out;
 959         }
 960
 961         if (ret == -ENOENT && insert) {
 962                 if (item_size + extra_size >=
 963                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 964                         ret = -EAGAIN;
 965                         goto out;
 966                 }
 967                 /*
 968                  * To add new inline back ref, we have to make sure
 969                  * there is no corresponding back ref item.
 970                  * For simplicity, we just do not add new inline back
 971                  * ref if there is any kind of item for this block
 972                  */
 973                 if (find_next_key(path, 0, &key) == 0 &&
 974                     key.objectid == bytenr &&
 975                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 976                         ret = -EAGAIN;
 977                         goto out;
 978                 }
 979         }
 980         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 981 out:
 982         if (insert) {
 983                 path->keep_locks = 0;
 984                 path->search_for_extension = 0;
 985                 btrfs_unlock_up_safe(path, 1);
 986         }
 987         return ret;
 988 }
 989
 990 /*
 991  * helper to add new inline back ref
 992  */
 993 static noinline_for_stack
 994 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
 995                                  struct btrfs_path *path,
 996                                  struct btrfs_extent_inline_ref *iref,
 997                                  u64 parent, u64 root_objectid,
 998                                  u64 owner, u64 offset, int refs_to_add,
 999                                  struct btrfs_delayed_extent_op *extent_op)
1000 {
1001         struct extent_buffer *leaf;
1002         struct btrfs_extent_item *ei;
1003         unsigned long ptr;
1004         unsigned long end;
1005         unsigned long item_offset;
1006         u64 refs;
1007         int size;
1008         int type;
1009
1010         leaf = path->nodes[0];
1011         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1012         item_offset = (unsigned long)iref - (unsigned long)ei;
1013
1014         type = extent_ref_type(parent, owner);
1015         size = btrfs_extent_inline_ref_size(type);
1016
1017         btrfs_extend_item(trans, path, size);
1018
1019         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1020         refs = btrfs_extent_refs(leaf, ei);
1021         refs += refs_to_add;
1022         btrfs_set_extent_refs(leaf, ei, refs);
1023         if (extent_op)
1024                 __run_delayed_extent_op(extent_op, leaf, ei);
1025
1026         ptr = (unsigned long)ei + item_offset;
1027         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1028         if (ptr < end - size)
1029                 memmove_extent_buffer(leaf, ptr + size, ptr,
1030                                       end - size - ptr);
1031
1032         iref = (struct btrfs_extent_inline_ref *)ptr;
1033         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1034         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1035                 struct btrfs_extent_data_ref *dref;
1036                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1037                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1038                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1039                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1040                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1041         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1042                 struct btrfs_shared_data_ref *sref;
1043                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1044                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1045                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1046         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1047                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1048         } else {
1049                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1050         }
1051         btrfs_mark_buffer_dirty(trans, leaf);
1052 }
1053
1054 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1055                                  struct btrfs_path *path,
1056                                  struct btrfs_extent_inline_ref **ref_ret,
1057                                  u64 bytenr, u64 num_bytes, u64 parent,
1058                                  u64 root_objectid, u64 owner, u64 offset)
1059 {
1060         int ret;
1061
1062         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1063                                            num_bytes, parent, root_objectid,
1064                                            owner, offset, 0);
1065         if (ret != -ENOENT)
1066                 return ret;
1067
1068         btrfs_release_path(path);
1069         *ref_ret = NULL;
1070
1071         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1072                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1073                                             root_objectid);
1074         } else {
1075                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1076                                              root_objectid, owner, offset);
1077         }
1078         return ret;
1079 }
1080
1081 /*
1082  * helper to update/remove inline back ref
1083  */
1084 static noinline_for_stack int update_inline_extent_backref(
1085                                   struct btrfs_trans_handle *trans,
1086                                   struct btrfs_path *path,
1087                                   struct btrfs_extent_inline_ref *iref,
1088                                   int refs_to_mod,
1089                                   struct btrfs_delayed_extent_op *extent_op)
1090 {
1091         struct extent_buffer *leaf = path->nodes[0];
1092         struct btrfs_fs_info *fs_info = leaf->fs_info;
1093         struct btrfs_extent_item *ei;
1094         struct btrfs_extent_data_ref *dref = NULL;
1095         struct btrfs_shared_data_ref *sref = NULL;
1096         unsigned long ptr;
1097         unsigned long end;
1098         u32 item_size;
1099         int size;
1100         int type;
1101         u64 refs;
1102
1103         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1104         refs = btrfs_extent_refs(leaf, ei);
1105         if (unlikely(refs_to_mod < 0 && refs + refs_to_mod <= 0)) {
1106                 struct btrfs_key key;
1107                 u32 extent_size;
1108
1109                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1110                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1111                         extent_size = fs_info->nodesize;
1112                 else
1113                         extent_size = key.offset;
1114                 btrfs_print_leaf(leaf);
1115                 btrfs_err(fs_info,
1116         "invalid refs_to_mod for extent %llu num_bytes %u, has %d expect >= -%llu",
1117                           key.objectid, extent_size, refs_to_mod, refs);
1118                 return -EUCLEAN;
1119         }
1120         refs += refs_to_mod;
1121         btrfs_set_extent_refs(leaf, ei, refs);
1122         if (extent_op)
1123                 __run_delayed_extent_op(extent_op, leaf, ei);
1124
1125         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1126         /*
1127          * Function btrfs_get_extent_inline_ref_type() has already printed
1128          * error messages.
1129          */
1130         if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1131                 return -EUCLEAN;
1132
1133         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1134                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1135                 refs = btrfs_extent_data_ref_count(leaf, dref);
1136         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1137                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1138                 refs = btrfs_shared_data_ref_count(leaf, sref);
1139         } else {
1140                 refs = 1;
1141                 /*
1142                  * For tree blocks we can only drop one ref for it, and tree
1143                  * blocks should not have refs > 1.
1144                  *
1145                  * Furthermore if we're inserting a new inline backref, we
1146                  * won't reach this path either. That would be
1147                  * setup_inline_extent_backref().
1148                  */
1149                 if (unlikely(refs_to_mod != -1)) {
1150                         struct btrfs_key key;
1151
1152                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1153
1154                         btrfs_print_leaf(leaf);
1155                         btrfs_err(fs_info,
1156                         "invalid refs_to_mod for tree block %llu, has %d expect -1",
1157                                   key.objectid, refs_to_mod);
1158                         return -EUCLEAN;
1159                 }
1160         }
1161
1162         if (unlikely(refs_to_mod < 0 && refs < -refs_to_mod)) {
1163                 struct btrfs_key key;
1164                 u32 extent_size;
1165
1166                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1167                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1168                         extent_size = fs_info->nodesize;
1169                 else
1170                         extent_size = key.offset;
1171                 btrfs_print_leaf(leaf);
1172                 btrfs_err(fs_info,
1173 "invalid refs_to_mod for backref entry, iref %lu extent %llu num_bytes %u, has %d expect >= -%llu",
1174                           (unsigned long)iref, key.objectid, extent_size,
1175                           refs_to_mod, refs);
1176                 return -EUCLEAN;
1177         }
1178         refs += refs_to_mod;
1179
1180         if (refs > 0) {
1181                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1182                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1183                 else
1184                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1185         } else {
1186                 size =  btrfs_extent_inline_ref_size(type);
1187                 item_size = btrfs_item_size(leaf, path->slots[0]);
1188                 ptr = (unsigned long)iref;
1189                 end = (unsigned long)ei + item_size;
1190                 if (ptr + size < end)
1191                         memmove_extent_buffer(leaf, ptr, ptr + size,
1192                                               end - ptr - size);
1193                 item_size -= size;
1194                 btrfs_truncate_item(trans, path, item_size, 1);
1195         }
1196         btrfs_mark_buffer_dirty(trans, leaf);
1197         return 0;
1198 }
1199
1200 static noinline_for_stack
1201 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1202                                  struct btrfs_path *path,
1203                                  u64 bytenr, u64 num_bytes, u64 parent,
1204                                  u64 root_objectid, u64 owner,
1205                                  u64 offset, int refs_to_add,
1206                                  struct btrfs_delayed_extent_op *extent_op)
1207 {
1208         struct btrfs_extent_inline_ref *iref;
1209         int ret;
1210
1211         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1212                                            num_bytes, parent, root_objectid,
1213                                            owner, offset, 1);
1214         if (ret == 0) {
1215                 /*
1216                  * We're adding refs to a tree block we already own, this
1217                  * should not happen at all.
1218                  */
1219                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1220                         btrfs_print_leaf(path->nodes[0]);
1221                         btrfs_crit(trans->fs_info,
1222 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu slot %u",
1223                                    bytenr, num_bytes, root_objectid, path->slots[0]);
1224                         return -EUCLEAN;
1225                 }
1226                 ret = update_inline_extent_backref(trans, path, iref,
1227                                                    refs_to_add, extent_op);
1228         } else if (ret == -ENOENT) {
1229                 setup_inline_extent_backref(trans, path, iref, parent,
1230                                             root_objectid, owner, offset,
1231                                             refs_to_add, extent_op);
1232                 ret = 0;
1233         }
1234         return ret;
1235 }
1236
1237 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1238                                  struct btrfs_root *root,
1239                                  struct btrfs_path *path,
1240                                  struct btrfs_extent_inline_ref *iref,
1241                                  int refs_to_drop, int is_data)
1242 {
1243         int ret = 0;
1244
1245         BUG_ON(!is_data && refs_to_drop != 1);
1246         if (iref)
1247                 ret = update_inline_extent_backref(trans, path, iref,
1248                                                    -refs_to_drop, NULL);
1249         else if (is_data)
1250                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1251         else
1252                 ret = btrfs_del_item(trans, root, path);
1253         return ret;
1254 }
1255
1256 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1257                                u64 *discarded_bytes)
1258 {
1259         int j, ret = 0;
1260         u64 bytes_left, end;
1261         u64 aligned_start = ALIGN(start, 1 << SECTOR_SHIFT);
1262
1263         if (WARN_ON(start != aligned_start)) {
1264                 len -= aligned_start - start;
1265                 len = round_down(len, 1 << SECTOR_SHIFT);
1266                 start = aligned_start;
1267         }
1268
1269         *discarded_bytes = 0;
1270
1271         if (!len)
1272                 return 0;
1273
1274         end = start + len;
1275         bytes_left = len;
1276
1277         /* Skip any superblocks on this device. */
1278         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1279                 u64 sb_start = btrfs_sb_offset(j);
1280                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1281                 u64 size = sb_start - start;
1282
1283                 if (!in_range(sb_start, start, bytes_left) &&
1284                     !in_range(sb_end, start, bytes_left) &&
1285                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1286                         continue;
1287
1288                 /*
1289                  * Superblock spans beginning of range.  Adjust start and
1290                  * try again.
1291                  */
1292                 if (sb_start <= start) {
1293                         start += sb_end - start;
1294                         if (start > end) {
1295                                 bytes_left = 0;
1296                                 break;
1297                         }
1298                         bytes_left = end - start;
1299                         continue;
1300                 }
1301
1302                 if (size) {
1303                         ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1304                                                    size >> SECTOR_SHIFT,
1305                                                    GFP_NOFS);
1306                         if (!ret)
1307                                 *discarded_bytes += size;
1308                         else if (ret != -EOPNOTSUPP)
1309                                 return ret;
1310                 }
1311
1312                 start = sb_end;
1313                 if (start > end) {
1314                         bytes_left = 0;
1315                         break;
1316                 }
1317                 bytes_left = end - start;
1318         }
1319
1320         if (bytes_left) {
1321                 ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1322                                            bytes_left >> SECTOR_SHIFT,
1323                                            GFP_NOFS);
1324                 if (!ret)
1325                         *discarded_bytes += bytes_left;
1326         }
1327         return ret;
1328 }
1329
1330 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1331 {
1332         struct btrfs_device *dev = stripe->dev;
1333         struct btrfs_fs_info *fs_info = dev->fs_info;
1334         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1335         u64 phys = stripe->physical;
1336         u64 len = stripe->length;
1337         u64 discarded = 0;
1338         int ret = 0;
1339
1340         /* Zone reset on a zoned filesystem */
1341         if (btrfs_can_zone_reset(dev, phys, len)) {
1342                 u64 src_disc;
1343
1344                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1345                 if (ret)
1346                         goto out;
1347
1348                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1349                     dev != dev_replace->srcdev)
1350                         goto out;
1351
1352                 src_disc = discarded;
1353
1354                 /* Send to replace target as well */
1355                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1356                                               &discarded);
1357                 discarded += src_disc;
1358         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1359                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1360         } else {
1361                 ret = 0;
1362                 *bytes = 0;
1363         }
1364
1365 out:
1366         *bytes = discarded;
1367         return ret;
1368 }
1369
1370 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1371                          u64 num_bytes, u64 *actual_bytes)
1372 {
1373         int ret = 0;
1374         u64 discarded_bytes = 0;
1375         u64 end = bytenr + num_bytes;
1376         u64 cur = bytenr;
1377
1378         /*
1379          * Avoid races with device replace and make sure the devices in the
1380          * stripes don't go away while we are discarding.
1381          */
1382         btrfs_bio_counter_inc_blocked(fs_info);
1383         while (cur < end) {
1384                 struct btrfs_discard_stripe *stripes;
1385                 unsigned int num_stripes;
1386                 int i;
1387
1388                 num_bytes = end - cur;
1389                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1390                 if (IS_ERR(stripes)) {
1391                         ret = PTR_ERR(stripes);
1392                         if (ret == -EOPNOTSUPP)
1393                                 ret = 0;
1394                         break;
1395                 }
1396
1397                 for (i = 0; i < num_stripes; i++) {
1398                         struct btrfs_discard_stripe *stripe = stripes + i;
1399                         u64 bytes;
1400
1401                         if (!stripe->dev->bdev) {
1402                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1403                                 continue;
1404                         }
1405
1406                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1407                                         &stripe->dev->dev_state))
1408                                 continue;
1409
1410                         ret = do_discard_extent(stripe, &bytes);
1411                         if (ret) {
1412                                 /*
1413                                  * Keep going if discard is not supported by the
1414                                  * device.
1415                                  */
1416                                 if (ret != -EOPNOTSUPP)
1417                                         break;
1418                                 ret = 0;
1419                         } else {
1420                                 discarded_bytes += bytes;
1421                         }
1422                 }
1423                 kfree(stripes);
1424                 if (ret)
1425                         break;
1426                 cur += num_bytes;
1427         }
1428         btrfs_bio_counter_dec(fs_info);
1429         if (actual_bytes)
1430                 *actual_bytes = discarded_bytes;
1431         return ret;
1432 }
1433
1434 /* Can return -ENOMEM */
1435 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1436                          struct btrfs_ref *generic_ref)
1437 {
1438         struct btrfs_fs_info *fs_info = trans->fs_info;
1439         int ret;
1440
1441         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1442                generic_ref->action);
1443         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1444                generic_ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID);
1445
1446         if (generic_ref->type == BTRFS_REF_METADATA)
1447                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1448         else
1449                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1450
1451         btrfs_ref_tree_mod(fs_info, generic_ref);
1452
1453         return ret;
1454 }
1455
1456 /*
1457  * Insert backreference for a given extent.
1458  *
1459  * The counterpart is in __btrfs_free_extent(), with examples and more details
1460  * how it works.
1461  *
1462  * @trans:          Handle of transaction
1463  *
1464  * @node:           The delayed ref node used to get the bytenr/length for
1465  *                  extent whose references are incremented.
1466  *
1467  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1468  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1469  *                  bytenr of the parent block. Since new extents are always
1470  *                  created with indirect references, this will only be the case
1471  *                  when relocating a shared extent. In that case, root_objectid
1472  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1473  *                  be 0
1474  *
1475  * @root_objectid:  The id of the root where this modification has originated,
1476  *                  this can be either one of the well-known metadata trees or
1477  *                  the subvolume id which references this extent.
1478  *
1479  * @owner:          For data extents it is the inode number of the owning file.
1480  *                  For metadata extents this parameter holds the level in the
1481  *                  tree of the extent.
1482  *
1483  * @offset:         For metadata extents the offset is ignored and is currently
1484  *                  always passed as 0. For data extents it is the fileoffset
1485  *                  this extent belongs to.
1486  *
1487  * @extent_op       Pointer to a structure, holding information necessary when
1488  *                  updating a tree block's flags
1489  *
1490  */
1491 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1492                                   struct btrfs_delayed_ref_node *node,
1493                                   u64 parent, u64 root_objectid,
1494                                   u64 owner, u64 offset,
1495                                   struct btrfs_delayed_extent_op *extent_op)
1496 {
1497         struct btrfs_path *path;
1498         struct extent_buffer *leaf;
1499         struct btrfs_extent_item *item;
1500         struct btrfs_key key;
1501         u64 bytenr = node->bytenr;
1502         u64 num_bytes = node->num_bytes;
1503         u64 refs;
1504         int refs_to_add = node->ref_mod;
1505         int ret;
1506
1507         path = btrfs_alloc_path();
1508         if (!path)
1509                 return -ENOMEM;
1510
1511         /* this will setup the path even if it fails to insert the back ref */
1512         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1513                                            parent, root_objectid, owner,
1514                                            offset, refs_to_add, extent_op);
1515         if ((ret < 0 && ret != -EAGAIN) || !ret)
1516                 goto out;
1517
1518         /*
1519          * Ok we had -EAGAIN which means we didn't have space to insert and
1520          * inline extent ref, so just update the reference count and add a
1521          * normal backref.
1522          */
1523         leaf = path->nodes[0];
1524         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1525         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1526         refs = btrfs_extent_refs(leaf, item);
1527         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1528         if (extent_op)
1529                 __run_delayed_extent_op(extent_op, leaf, item);
1530
1531         btrfs_mark_buffer_dirty(trans, leaf);
1532         btrfs_release_path(path);
1533
1534         /* now insert the actual backref */
1535         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1536                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1537                                             root_objectid);
1538         else
1539                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1540                                              root_objectid, owner, offset,
1541                                              refs_to_add);
1542
1543         if (ret)
1544                 btrfs_abort_transaction(trans, ret);
1545 out:
1546         btrfs_free_path(path);
1547         return ret;
1548 }
1549
1550 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1551                                 struct btrfs_delayed_ref_head *href,
1552                                 struct btrfs_delayed_ref_node *node,
1553                                 struct btrfs_delayed_extent_op *extent_op,
1554                                 bool insert_reserved)
1555 {
1556         int ret = 0;
1557         struct btrfs_delayed_data_ref *ref;
1558         u64 parent = 0;
1559         u64 flags = 0;
1560
1561         ref = btrfs_delayed_node_to_data_ref(node);
1562         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1563
1564         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1565                 parent = ref->parent;
1566
1567         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1568                 struct btrfs_key key;
1569                 struct btrfs_squota_delta delta = {
1570                         .root = href->owning_root,
1571                         .num_bytes = node->num_bytes,
1572                         .rsv_bytes = href->reserved_bytes,
1573                         .is_data = true,
1574                         .is_inc = true,
1575                         .generation = trans->transid,
1576                 };
1577
1578                 if (extent_op)
1579                         flags |= extent_op->flags_to_set;
1580
1581                 key.objectid = node->bytenr;
1582                 key.type = BTRFS_EXTENT_ITEM_KEY;
1583                 key.offset = node->num_bytes;
1584
1585                 ret = alloc_reserved_file_extent(trans, parent, ref->root,
1586                                                  flags, ref->objectid,
1587                                                  ref->offset, &key,
1588                                                  node->ref_mod, href->owning_root);
1589                 if (!ret)
1590                         ret = btrfs_record_squota_delta(trans->fs_info, &delta);
1591                 else
1592                         btrfs_qgroup_free_refroot(trans->fs_info, delta.root,
1593                                                   delta.rsv_bytes, BTRFS_QGROUP_RSV_DATA);
1594         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1595                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref->root,
1596                                              ref->objectid, ref->offset,
1597                                              extent_op);
1598         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1599                 ret = __btrfs_free_extent(trans, href, node, parent,
1600                                           ref->root, ref->objectid,
1601                                           ref->offset, extent_op);
1602         } else {
1603                 BUG();
1604         }
1605         return ret;
1606 }
1607
1608 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1609                                     struct extent_buffer *leaf,
1610                                     struct btrfs_extent_item *ei)
1611 {
1612         u64 flags = btrfs_extent_flags(leaf, ei);
1613         if (extent_op->update_flags) {
1614                 flags |= extent_op->flags_to_set;
1615                 btrfs_set_extent_flags(leaf, ei, flags);
1616         }
1617
1618         if (extent_op->update_key) {
1619                 struct btrfs_tree_block_info *bi;
1620                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1621                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1622                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1623         }
1624 }
1625
1626 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1627                                  struct btrfs_delayed_ref_head *head,
1628                                  struct btrfs_delayed_extent_op *extent_op)
1629 {
1630         struct btrfs_fs_info *fs_info = trans->fs_info;
1631         struct btrfs_root *root;
1632         struct btrfs_key key;
1633         struct btrfs_path *path;
1634         struct btrfs_extent_item *ei;
1635         struct extent_buffer *leaf;
1636         u32 item_size;
1637         int ret;
1638         int metadata = 1;
1639
1640         if (TRANS_ABORTED(trans))
1641                 return 0;
1642
1643         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1644                 metadata = 0;
1645
1646         path = btrfs_alloc_path();
1647         if (!path)
1648                 return -ENOMEM;
1649
1650         key.objectid = head->bytenr;
1651
1652         if (metadata) {
1653                 key.type = BTRFS_METADATA_ITEM_KEY;
1654                 key.offset = extent_op->level;
1655         } else {
1656                 key.type = BTRFS_EXTENT_ITEM_KEY;
1657                 key.offset = head->num_bytes;
1658         }
1659
1660         root = btrfs_extent_root(fs_info, key.objectid);
1661 again:
1662         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1663         if (ret < 0) {
1664                 goto out;
1665         } else if (ret > 0) {
1666                 if (metadata) {
1667                         if (path->slots[0] > 0) {
1668                                 path->slots[0]--;
1669                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1670                                                       path->slots[0]);
1671                                 if (key.objectid == head->bytenr &&
1672                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1673                                     key.offset == head->num_bytes)
1674                                         ret = 0;
1675                         }
1676                         if (ret > 0) {
1677                                 btrfs_release_path(path);
1678                                 metadata = 0;
1679
1680                                 key.objectid = head->bytenr;
1681                                 key.offset = head->num_bytes;
1682                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1683                                 goto again;
1684                         }
1685                 } else {
1686                         ret = -EUCLEAN;
1687                         btrfs_err(fs_info,
1688                   "missing extent item for extent %llu num_bytes %llu level %d",
1689                                   head->bytenr, head->num_bytes, extent_op->level);
1690                         goto out;
1691                 }
1692         }
1693
1694         leaf = path->nodes[0];
1695         item_size = btrfs_item_size(leaf, path->slots[0]);
1696
1697         if (unlikely(item_size < sizeof(*ei))) {
1698                 ret = -EUCLEAN;
1699                 btrfs_err(fs_info,
1700                           "unexpected extent item size, has %u expect >= %zu",
1701                           item_size, sizeof(*ei));
1702                 btrfs_abort_transaction(trans, ret);
1703                 goto out;
1704         }
1705
1706         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1707         __run_delayed_extent_op(extent_op, leaf, ei);
1708
1709         btrfs_mark_buffer_dirty(trans, leaf);
1710 out:
1711         btrfs_free_path(path);
1712         return ret;
1713 }
1714
1715 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1716                                 struct btrfs_delayed_ref_head *href,
1717                                 struct btrfs_delayed_ref_node *node,
1718                                 struct btrfs_delayed_extent_op *extent_op,
1719                                 bool insert_reserved)
1720 {
1721         int ret = 0;
1722         struct btrfs_fs_info *fs_info = trans->fs_info;
1723         struct btrfs_delayed_tree_ref *ref;
1724         u64 parent = 0;
1725         u64 ref_root = 0;
1726
1727         ref = btrfs_delayed_node_to_tree_ref(node);
1728         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1729
1730         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1731                 parent = ref->parent;
1732         ref_root = ref->root;
1733
1734         if (unlikely(node->ref_mod != 1)) {
1735                 btrfs_err(trans->fs_info,
1736         "btree block %llu has %d references rather than 1: action %d ref_root %llu parent %llu",
1737                           node->bytenr, node->ref_mod, node->action, ref_root,
1738                           parent);
1739                 return -EUCLEAN;
1740         }
1741         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1742                 struct btrfs_squota_delta delta = {
1743                         .root = href->owning_root,
1744                         .num_bytes = fs_info->nodesize,
1745                         .rsv_bytes = 0,
1746                         .is_data = false,
1747                         .is_inc = true,
1748                         .generation = trans->transid,
1749                 };
1750
1751                 BUG_ON(!extent_op || !extent_op->update_flags);
1752                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1753                 if (!ret)
1754                         btrfs_record_squota_delta(fs_info, &delta);
1755         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1756                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1757                                              ref->level, 0, extent_op);
1758         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1759                 ret = __btrfs_free_extent(trans, href, node, parent, ref_root,
1760                                           ref->level, 0, extent_op);
1761         } else {
1762                 BUG();
1763         }
1764         return ret;
1765 }
1766
1767 /* helper function to actually process a single delayed ref entry */
1768 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1769                                struct btrfs_delayed_ref_head *href,
1770                                struct btrfs_delayed_ref_node *node,
1771                                struct btrfs_delayed_extent_op *extent_op,
1772                                bool insert_reserved)
1773 {
1774         int ret = 0;
1775
1776         if (TRANS_ABORTED(trans)) {
1777                 if (insert_reserved)
1778                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1779                 return 0;
1780         }
1781
1782         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1783             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1784                 ret = run_delayed_tree_ref(trans, href, node, extent_op,
1785                                            insert_reserved);
1786         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1787                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1788                 ret = run_delayed_data_ref(trans, href, node, extent_op,
1789                                            insert_reserved);
1790         else if (node->type == BTRFS_EXTENT_OWNER_REF_KEY)
1791                 ret = 0;
1792         else
1793                 BUG();
1794         if (ret && insert_reserved)
1795                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1796         if (ret < 0)
1797                 btrfs_err(trans->fs_info,
1798 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1799                           node->bytenr, node->num_bytes, node->type,
1800                           node->action, node->ref_mod, ret);
1801         return ret;
1802 }
1803
1804 static inline struct btrfs_delayed_ref_node *
1805 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1806 {
1807         struct btrfs_delayed_ref_node *ref;
1808
1809         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1810                 return NULL;
1811
1812         /*
1813          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1814          * This is to prevent a ref count from going down to zero, which deletes
1815          * the extent item from the extent tree, when there still are references
1816          * to add, which would fail because they would not find the extent item.
1817          */
1818         if (!list_empty(&head->ref_add_list))
1819                 return list_first_entry(&head->ref_add_list,
1820                                 struct btrfs_delayed_ref_node, add_list);
1821
1822         ref = rb_entry(rb_first_cached(&head->ref_tree),
1823                        struct btrfs_delayed_ref_node, ref_node);
1824         ASSERT(list_empty(&ref->add_list));
1825         return ref;
1826 }
1827
1828 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1829                                       struct btrfs_delayed_ref_head *head)
1830 {
1831         spin_lock(&delayed_refs->lock);
1832         head->processing = false;
1833         delayed_refs->num_heads_ready++;
1834         spin_unlock(&delayed_refs->lock);
1835         btrfs_delayed_ref_unlock(head);
1836 }
1837
1838 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1839                                 struct btrfs_delayed_ref_head *head)
1840 {
1841         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1842
1843         if (!extent_op)
1844                 return NULL;
1845
1846         if (head->must_insert_reserved) {
1847                 head->extent_op = NULL;
1848                 btrfs_free_delayed_extent_op(extent_op);
1849                 return NULL;
1850         }
1851         return extent_op;
1852 }
1853
1854 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1855                                      struct btrfs_delayed_ref_head *head)
1856 {
1857         struct btrfs_delayed_extent_op *extent_op;
1858         int ret;
1859
1860         extent_op = cleanup_extent_op(head);
1861         if (!extent_op)
1862                 return 0;
1863         head->extent_op = NULL;
1864         spin_unlock(&head->lock);
1865         ret = run_delayed_extent_op(trans, head, extent_op);
1866         btrfs_free_delayed_extent_op(extent_op);
1867         return ret ? ret : 1;
1868 }
1869
1870 u64 btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1871                                   struct btrfs_delayed_ref_root *delayed_refs,
1872                                   struct btrfs_delayed_ref_head *head)
1873 {
1874         /*
1875          * We had csum deletions accounted for in our delayed refs rsv, we need
1876          * to drop the csum leaves for this update from our delayed_refs_rsv.
1877          */
1878         if (head->total_ref_mod < 0 && head->is_data) {
1879                 int nr_csums;
1880
1881                 spin_lock(&delayed_refs->lock);
1882                 delayed_refs->pending_csums -= head->num_bytes;
1883                 spin_unlock(&delayed_refs->lock);
1884                 nr_csums = btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1885
1886                 btrfs_delayed_refs_rsv_release(fs_info, 0, nr_csums);
1887
1888                 return btrfs_calc_delayed_ref_csum_bytes(fs_info, nr_csums);
1889         }
1890         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE &&
1891             head->must_insert_reserved && head->is_data)
1892                 btrfs_qgroup_free_refroot(fs_info, head->owning_root,
1893                                           head->reserved_bytes, BTRFS_QGROUP_RSV_DATA);
1894
1895         return 0;
1896 }
1897
1898 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1899                             struct btrfs_delayed_ref_head *head,
1900                             u64 *bytes_released)
1901 {
1902
1903         struct btrfs_fs_info *fs_info = trans->fs_info;
1904         struct btrfs_delayed_ref_root *delayed_refs;
1905         int ret;
1906
1907         delayed_refs = &trans->transaction->delayed_refs;
1908
1909         ret = run_and_cleanup_extent_op(trans, head);
1910         if (ret < 0) {
1911                 unselect_delayed_ref_head(delayed_refs, head);
1912                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1913                 return ret;
1914         } else if (ret) {
1915                 return ret;
1916         }
1917
1918         /*
1919          * Need to drop our head ref lock and re-acquire the delayed ref lock
1920          * and then re-check to make sure nobody got added.
1921          */
1922         spin_unlock(&head->lock);
1923         spin_lock(&delayed_refs->lock);
1924         spin_lock(&head->lock);
1925         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1926                 spin_unlock(&head->lock);
1927                 spin_unlock(&delayed_refs->lock);
1928                 return 1;
1929         }
1930         btrfs_delete_ref_head(delayed_refs, head);
1931         spin_unlock(&head->lock);
1932         spin_unlock(&delayed_refs->lock);
1933
1934         if (head->must_insert_reserved) {
1935                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1936                 if (head->is_data) {
1937                         struct btrfs_root *csum_root;
1938
1939                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1940                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1941                                               head->num_bytes);
1942                 }
1943         }
1944
1945         *bytes_released += btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1946
1947         trace_run_delayed_ref_head(fs_info, head, 0);
1948         btrfs_delayed_ref_unlock(head);
1949         btrfs_put_delayed_ref_head(head);
1950         return ret;
1951 }
1952
1953 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1954                                         struct btrfs_trans_handle *trans)
1955 {
1956         struct btrfs_delayed_ref_root *delayed_refs =
1957                 &trans->transaction->delayed_refs;
1958         struct btrfs_delayed_ref_head *head = NULL;
1959         int ret;
1960
1961         spin_lock(&delayed_refs->lock);
1962         head = btrfs_select_ref_head(delayed_refs);
1963         if (!head) {
1964                 spin_unlock(&delayed_refs->lock);
1965                 return head;
1966         }
1967
1968         /*
1969          * Grab the lock that says we are going to process all the refs for
1970          * this head
1971          */
1972         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1973         spin_unlock(&delayed_refs->lock);
1974
1975         /*
1976          * We may have dropped the spin lock to get the head mutex lock, and
1977          * that might have given someone else time to free the head.  If that's
1978          * true, it has been removed from our list and we can move on.
1979          */
1980         if (ret == -EAGAIN)
1981                 head = ERR_PTR(-EAGAIN);
1982
1983         return head;
1984 }
1985
1986 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1987                                            struct btrfs_delayed_ref_head *locked_ref,
1988                                            u64 *bytes_released)
1989 {
1990         struct btrfs_fs_info *fs_info = trans->fs_info;
1991         struct btrfs_delayed_ref_root *delayed_refs;
1992         struct btrfs_delayed_extent_op *extent_op;
1993         struct btrfs_delayed_ref_node *ref;
1994         bool must_insert_reserved;
1995         int ret;
1996
1997         delayed_refs = &trans->transaction->delayed_refs;
1998
1999         lockdep_assert_held(&locked_ref->mutex);
2000         lockdep_assert_held(&locked_ref->lock);
2001
2002         while ((ref = select_delayed_ref(locked_ref))) {
2003                 if (ref->seq &&
2004                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2005                         spin_unlock(&locked_ref->lock);
2006                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2007                         return -EAGAIN;
2008                 }
2009
2010                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2011                 RB_CLEAR_NODE(&ref->ref_node);
2012                 if (!list_empty(&ref->add_list))
2013                         list_del(&ref->add_list);
2014                 /*
2015                  * When we play the delayed ref, also correct the ref_mod on
2016                  * head
2017                  */
2018                 switch (ref->action) {
2019                 case BTRFS_ADD_DELAYED_REF:
2020                 case BTRFS_ADD_DELAYED_EXTENT:
2021                         locked_ref->ref_mod -= ref->ref_mod;
2022                         break;
2023                 case BTRFS_DROP_DELAYED_REF:
2024                         locked_ref->ref_mod += ref->ref_mod;
2025                         break;
2026                 default:
2027                         WARN_ON(1);
2028                 }
2029                 atomic_dec(&delayed_refs->num_entries);
2030
2031                 /*
2032                  * Record the must_insert_reserved flag before we drop the
2033                  * spin lock.
2034                  */
2035                 must_insert_reserved = locked_ref->must_insert_reserved;
2036                 locked_ref->must_insert_reserved = false;
2037
2038                 extent_op = locked_ref->extent_op;
2039                 locked_ref->extent_op = NULL;
2040                 spin_unlock(&locked_ref->lock);
2041
2042                 ret = run_one_delayed_ref(trans, locked_ref, ref, extent_op,
2043                                           must_insert_reserved);
2044                 btrfs_delayed_refs_rsv_release(fs_info, 1, 0);
2045                 *bytes_released += btrfs_calc_delayed_ref_bytes(fs_info, 1);
2046
2047                 btrfs_free_delayed_extent_op(extent_op);
2048                 if (ret) {
2049                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2050                         btrfs_put_delayed_ref(ref);
2051                         return ret;
2052                 }
2053
2054                 btrfs_put_delayed_ref(ref);
2055                 cond_resched();
2056
2057                 spin_lock(&locked_ref->lock);
2058                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2059         }
2060
2061         return 0;
2062 }
2063
2064 /*
2065  * Returns 0 on success or if called with an already aborted transaction.
2066  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2067  */
2068 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2069                                              u64 min_bytes)
2070 {
2071         struct btrfs_fs_info *fs_info = trans->fs_info;
2072         struct btrfs_delayed_ref_root *delayed_refs;
2073         struct btrfs_delayed_ref_head *locked_ref = NULL;
2074         int ret;
2075         unsigned long count = 0;
2076         unsigned long max_count = 0;
2077         u64 bytes_processed = 0;
2078
2079         delayed_refs = &trans->transaction->delayed_refs;
2080         if (min_bytes == 0) {
2081                 max_count = delayed_refs->num_heads_ready;
2082                 min_bytes = U64_MAX;
2083         }
2084
2085         do {
2086                 if (!locked_ref) {
2087                         locked_ref = btrfs_obtain_ref_head(trans);
2088                         if (IS_ERR_OR_NULL(locked_ref)) {
2089                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2090                                         continue;
2091                                 } else {
2092                                         break;
2093                                 }
2094                         }
2095                         count++;
2096                 }
2097                 /*
2098                  * We need to try and merge add/drops of the same ref since we
2099                  * can run into issues with relocate dropping the implicit ref
2100                  * and then it being added back again before the drop can
2101                  * finish.  If we merged anything we need to re-loop so we can
2102                  * get a good ref.
2103                  * Or we can get node references of the same type that weren't
2104                  * merged when created due to bumps in the tree mod seq, and
2105                  * we need to merge them to prevent adding an inline extent
2106                  * backref before dropping it (triggering a BUG_ON at
2107                  * insert_inline_extent_backref()).
2108                  */
2109                 spin_lock(&locked_ref->lock);
2110                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2111
2112                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref, &bytes_processed);
2113                 if (ret < 0 && ret != -EAGAIN) {
2114                         /*
2115                          * Error, btrfs_run_delayed_refs_for_head already
2116                          * unlocked everything so just bail out
2117                          */
2118                         return ret;
2119                 } else if (!ret) {
2120                         /*
2121                          * Success, perform the usual cleanup of a processed
2122                          * head
2123                          */
2124                         ret = cleanup_ref_head(trans, locked_ref, &bytes_processed);
2125                         if (ret > 0 ) {
2126                                 /* We dropped our lock, we need to loop. */
2127                                 ret = 0;
2128                                 continue;
2129                         } else if (ret) {
2130                                 return ret;
2131                         }
2132                 }
2133
2134                 /*
2135                  * Either success case or btrfs_run_delayed_refs_for_head
2136                  * returned -EAGAIN, meaning we need to select another head
2137                  */
2138
2139                 locked_ref = NULL;
2140                 cond_resched();
2141         } while ((min_bytes != U64_MAX && bytes_processed < min_bytes) ||
2142                  (max_count > 0 && count < max_count) ||
2143                  locked_ref);
2144
2145         return 0;
2146 }
2147
2148 #ifdef SCRAMBLE_DELAYED_REFS
2149 /*
2150  * Normally delayed refs get processed in ascending bytenr order. This
2151  * correlates in most cases to the order added. To expose dependencies on this
2152  * order, we start to process the tree in the middle instead of the beginning
2153  */
2154 static u64 find_middle(struct rb_root *root)
2155 {
2156         struct rb_node *n = root->rb_node;
2157         struct btrfs_delayed_ref_node *entry;
2158         int alt = 1;
2159         u64 middle;
2160         u64 first = 0, last = 0;
2161
2162         n = rb_first(root);
2163         if (n) {
2164                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2165                 first = entry->bytenr;
2166         }
2167         n = rb_last(root);
2168         if (n) {
2169                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2170                 last = entry->bytenr;
2171         }
2172         n = root->rb_node;
2173
2174         while (n) {
2175                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2176                 WARN_ON(!entry->in_tree);
2177
2178                 middle = entry->bytenr;
2179
2180                 if (alt)
2181                         n = n->rb_left;
2182                 else
2183                         n = n->rb_right;
2184
2185                 alt = 1 - alt;
2186         }
2187         return middle;
2188 }
2189 #endif
2190
2191 /*
2192  * Start processing the delayed reference count updates and extent insertions
2193  * we have queued up so far.
2194  *
2195  * @trans:      Transaction handle.
2196  * @min_bytes:  How many bytes of delayed references to process. After this
2197  *              many bytes we stop processing delayed references if there are
2198  *              any more. If 0 it means to run all existing delayed references,
2199  *              but not new ones added after running all existing ones.
2200  *              Use (u64)-1 (U64_MAX) to run all existing delayed references
2201  *              plus any new ones that are added.
2202  *
2203  * Returns 0 on success or if called with an aborted transaction
2204  * Returns <0 on error and aborts the transaction
2205  */
2206 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, u64 min_bytes)
2207 {
2208         struct btrfs_fs_info *fs_info = trans->fs_info;
2209         struct btrfs_delayed_ref_root *delayed_refs;
2210         int ret;
2211
2212         /* We'll clean this up in btrfs_cleanup_transaction */
2213         if (TRANS_ABORTED(trans))
2214                 return 0;
2215
2216         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2217                 return 0;
2218
2219         delayed_refs = &trans->transaction->delayed_refs;
2220 again:
2221 #ifdef SCRAMBLE_DELAYED_REFS
2222         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2223 #endif
2224         ret = __btrfs_run_delayed_refs(trans, min_bytes);
2225         if (ret < 0) {
2226                 btrfs_abort_transaction(trans, ret);
2227                 return ret;
2228         }
2229
2230         if (min_bytes == U64_MAX) {
2231                 btrfs_create_pending_block_groups(trans);
2232
2233                 spin_lock(&delayed_refs->lock);
2234                 if (RB_EMPTY_ROOT(&delayed_refs->href_root.rb_root)) {
2235                         spin_unlock(&delayed_refs->lock);
2236                         return 0;
2237                 }
2238                 spin_unlock(&delayed_refs->lock);
2239
2240                 cond_resched();
2241                 goto again;
2242         }
2243
2244         return 0;
2245 }
2246
2247 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2248                                 struct extent_buffer *eb, u64 flags)
2249 {
2250         struct btrfs_delayed_extent_op *extent_op;
2251         int level = btrfs_header_level(eb);
2252         int ret;
2253
2254         extent_op = btrfs_alloc_delayed_extent_op();
2255         if (!extent_op)
2256                 return -ENOMEM;
2257
2258         extent_op->flags_to_set = flags;
2259         extent_op->update_flags = true;
2260         extent_op->update_key = false;
2261         extent_op->level = level;
2262
2263         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2264         if (ret)
2265                 btrfs_free_delayed_extent_op(extent_op);
2266         return ret;
2267 }
2268
2269 static noinline int check_delayed_ref(struct btrfs_root *root,
2270                                       struct btrfs_path *path,
2271                                       u64 objectid, u64 offset, u64 bytenr)
2272 {
2273         struct btrfs_delayed_ref_head *head;
2274         struct btrfs_delayed_ref_node *ref;
2275         struct btrfs_delayed_data_ref *data_ref;
2276         struct btrfs_delayed_ref_root *delayed_refs;
2277         struct btrfs_transaction *cur_trans;
2278         struct rb_node *node;
2279         int ret = 0;
2280
2281         spin_lock(&root->fs_info->trans_lock);
2282         cur_trans = root->fs_info->running_transaction;
2283         if (cur_trans)
2284                 refcount_inc(&cur_trans->use_count);
2285         spin_unlock(&root->fs_info->trans_lock);
2286         if (!cur_trans)
2287                 return 0;
2288
2289         delayed_refs = &cur_trans->delayed_refs;
2290         spin_lock(&delayed_refs->lock);
2291         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2292         if (!head) {
2293                 spin_unlock(&delayed_refs->lock);
2294                 btrfs_put_transaction(cur_trans);
2295                 return 0;
2296         }
2297
2298         if (!mutex_trylock(&head->mutex)) {
2299                 if (path->nowait) {
2300                         spin_unlock(&delayed_refs->lock);
2301                         btrfs_put_transaction(cur_trans);
2302                         return -EAGAIN;
2303                 }
2304
2305                 refcount_inc(&head->refs);
2306                 spin_unlock(&delayed_refs->lock);
2307
2308                 btrfs_release_path(path);
2309
2310                 /*
2311                  * Mutex was contended, block until it's released and let
2312                  * caller try again
2313                  */
2314                 mutex_lock(&head->mutex);
2315                 mutex_unlock(&head->mutex);
2316                 btrfs_put_delayed_ref_head(head);
2317                 btrfs_put_transaction(cur_trans);
2318                 return -EAGAIN;
2319         }
2320         spin_unlock(&delayed_refs->lock);
2321
2322         spin_lock(&head->lock);
2323         /*
2324          * XXX: We should replace this with a proper search function in the
2325          * future.
2326          */
2327         for (node = rb_first_cached(&head->ref_tree); node;
2328              node = rb_next(node)) {
2329                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2330                 /* If it's a shared ref we know a cross reference exists */
2331                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2332                         ret = 1;
2333                         break;
2334                 }
2335
2336                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2337
2338                 /*
2339                  * If our ref doesn't match the one we're currently looking at
2340                  * then we have a cross reference.
2341                  */
2342                 if (data_ref->root != root->root_key.objectid ||
2343                     data_ref->objectid != objectid ||
2344                     data_ref->offset != offset) {
2345                         ret = 1;
2346                         break;
2347                 }
2348         }
2349         spin_unlock(&head->lock);
2350         mutex_unlock(&head->mutex);
2351         btrfs_put_transaction(cur_trans);
2352         return ret;
2353 }
2354
2355 static noinline int check_committed_ref(struct btrfs_root *root,
2356                                         struct btrfs_path *path,
2357                                         u64 objectid, u64 offset, u64 bytenr,
2358                                         bool strict)
2359 {
2360         struct btrfs_fs_info *fs_info = root->fs_info;
2361         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2362         struct extent_buffer *leaf;
2363         struct btrfs_extent_data_ref *ref;
2364         struct btrfs_extent_inline_ref *iref;
2365         struct btrfs_extent_item *ei;
2366         struct btrfs_key key;
2367         u32 item_size;
2368         u32 expected_size;
2369         int type;
2370         int ret;
2371
2372         key.objectid = bytenr;
2373         key.offset = (u64)-1;
2374         key.type = BTRFS_EXTENT_ITEM_KEY;
2375
2376         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2377         if (ret < 0)
2378                 goto out;
2379         BUG_ON(ret == 0); /* Corruption */
2380
2381         ret = -ENOENT;
2382         if (path->slots[0] == 0)
2383                 goto out;
2384
2385         path->slots[0]--;
2386         leaf = path->nodes[0];
2387         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2388
2389         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2390                 goto out;
2391
2392         ret = 1;
2393         item_size = btrfs_item_size(leaf, path->slots[0]);
2394         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2395         expected_size = sizeof(*ei) + btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY);
2396
2397         /* No inline refs; we need to bail before checking for owner ref. */
2398         if (item_size == sizeof(*ei))
2399                 goto out;
2400
2401         /* Check for an owner ref; skip over it to the real inline refs. */
2402         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2403         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2404         if (btrfs_fs_incompat(fs_info, SIMPLE_QUOTA) && type == BTRFS_EXTENT_OWNER_REF_KEY) {
2405                 expected_size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
2406                 iref = (struct btrfs_extent_inline_ref *)(iref + 1);
2407         }
2408
2409         /* If extent item has more than 1 inline ref then it's shared */
2410         if (item_size != expected_size)
2411                 goto out;
2412
2413         /*
2414          * If extent created before last snapshot => it's shared unless the
2415          * snapshot has been deleted. Use the heuristic if strict is false.
2416          */
2417         if (!strict &&
2418             (btrfs_extent_generation(leaf, ei) <=
2419              btrfs_root_last_snapshot(&root->root_item)))
2420                 goto out;
2421
2422         /* If this extent has SHARED_DATA_REF then it's shared */
2423         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2424         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2425                 goto out;
2426
2427         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2428         if (btrfs_extent_refs(leaf, ei) !=
2429             btrfs_extent_data_ref_count(leaf, ref) ||
2430             btrfs_extent_data_ref_root(leaf, ref) !=
2431             root->root_key.objectid ||
2432             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2433             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2434                 goto out;
2435
2436         ret = 0;
2437 out:
2438         return ret;
2439 }
2440
2441 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2442                           u64 bytenr, bool strict, struct btrfs_path *path)
2443 {
2444         int ret;
2445
2446         do {
2447                 ret = check_committed_ref(root, path, objectid,
2448                                           offset, bytenr, strict);
2449                 if (ret && ret != -ENOENT)
2450                         goto out;
2451
2452                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2453         } while (ret == -EAGAIN);
2454
2455 out:
2456         btrfs_release_path(path);
2457         if (btrfs_is_data_reloc_root(root))
2458                 WARN_ON(ret > 0);
2459         return ret;
2460 }
2461
2462 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2463                            struct btrfs_root *root,
2464                            struct extent_buffer *buf,
2465                            int full_backref, int inc)
2466 {
2467         struct btrfs_fs_info *fs_info = root->fs_info;
2468         u64 bytenr;
2469         u64 num_bytes;
2470         u64 parent;
2471         u64 ref_root;
2472         u32 nritems;
2473         struct btrfs_key key;
2474         struct btrfs_file_extent_item *fi;
2475         struct btrfs_ref generic_ref = { 0 };
2476         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2477         int i;
2478         int action;
2479         int level;
2480         int ret = 0;
2481
2482         if (btrfs_is_testing(fs_info))
2483                 return 0;
2484
2485         ref_root = btrfs_header_owner(buf);
2486         nritems = btrfs_header_nritems(buf);
2487         level = btrfs_header_level(buf);
2488
2489         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2490                 return 0;
2491
2492         if (full_backref)
2493                 parent = buf->start;
2494         else
2495                 parent = 0;
2496         if (inc)
2497                 action = BTRFS_ADD_DELAYED_REF;
2498         else
2499                 action = BTRFS_DROP_DELAYED_REF;
2500
2501         for (i = 0; i < nritems; i++) {
2502                 if (level == 0) {
2503                         btrfs_item_key_to_cpu(buf, &key, i);
2504                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2505                                 continue;
2506                         fi = btrfs_item_ptr(buf, i,
2507                                             struct btrfs_file_extent_item);
2508                         if (btrfs_file_extent_type(buf, fi) ==
2509                             BTRFS_FILE_EXTENT_INLINE)
2510                                 continue;
2511                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2512                         if (bytenr == 0)
2513                                 continue;
2514
2515                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2516                         key.offset -= btrfs_file_extent_offset(buf, fi);
2517                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2518                                                num_bytes, parent, ref_root);
2519                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2520                                             key.offset, root->root_key.objectid,
2521                                             for_reloc);
2522                         if (inc)
2523                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2524                         else
2525                                 ret = btrfs_free_extent(trans, &generic_ref);
2526                         if (ret)
2527                                 goto fail;
2528                 } else {
2529                         bytenr = btrfs_node_blockptr(buf, i);
2530                         num_bytes = fs_info->nodesize;
2531                         /* We don't know the owning_root, use 0. */
2532                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2533                                                num_bytes, parent, 0);
2534                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2535                                             root->root_key.objectid, for_reloc);
2536                         if (inc)
2537                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2538                         else
2539                                 ret = btrfs_free_extent(trans, &generic_ref);
2540                         if (ret)
2541                                 goto fail;
2542                 }
2543         }
2544         return 0;
2545 fail:
2546         return ret;
2547 }
2548
2549 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2550                   struct extent_buffer *buf, int full_backref)
2551 {
2552         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2553 }
2554
2555 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2556                   struct extent_buffer *buf, int full_backref)
2557 {
2558         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2559 }
2560
2561 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2562 {
2563         struct btrfs_fs_info *fs_info = root->fs_info;
2564         u64 flags;
2565         u64 ret;
2566
2567         if (data)
2568                 flags = BTRFS_BLOCK_GROUP_DATA;
2569         else if (root == fs_info->chunk_root)
2570                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2571         else
2572                 flags = BTRFS_BLOCK_GROUP_METADATA;
2573
2574         ret = btrfs_get_alloc_profile(fs_info, flags);
2575         return ret;
2576 }
2577
2578 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2579 {
2580         struct rb_node *leftmost;
2581         u64 bytenr = 0;
2582
2583         read_lock(&fs_info->block_group_cache_lock);
2584         /* Get the block group with the lowest logical start address. */
2585         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2586         if (leftmost) {
2587                 struct btrfs_block_group *bg;
2588
2589                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2590                 bytenr = bg->start;
2591         }
2592         read_unlock(&fs_info->block_group_cache_lock);
2593
2594         return bytenr;
2595 }
2596
2597 static int pin_down_extent(struct btrfs_trans_handle *trans,
2598                            struct btrfs_block_group *cache,
2599                            u64 bytenr, u64 num_bytes, int reserved)
2600 {
2601         struct btrfs_fs_info *fs_info = cache->fs_info;
2602
2603         spin_lock(&cache->space_info->lock);
2604         spin_lock(&cache->lock);
2605         cache->pinned += num_bytes;
2606         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2607                                              num_bytes);
2608         if (reserved) {
2609                 cache->reserved -= num_bytes;
2610                 cache->space_info->bytes_reserved -= num_bytes;
2611         }
2612         spin_unlock(&cache->lock);
2613         spin_unlock(&cache->space_info->lock);
2614
2615         set_extent_bit(&trans->transaction->pinned_extents, bytenr,
2616                        bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
2617         return 0;
2618 }
2619
2620 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2621                      u64 bytenr, u64 num_bytes, int reserved)
2622 {
2623         struct btrfs_block_group *cache;
2624
2625         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2626         BUG_ON(!cache); /* Logic error */
2627
2628         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2629
2630         btrfs_put_block_group(cache);
2631         return 0;
2632 }
2633
2634 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2635                                     const struct extent_buffer *eb)
2636 {
2637         struct btrfs_block_group *cache;
2638         int ret;
2639
2640         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
2641         if (!cache)
2642                 return -EINVAL;
2643
2644         /*
2645          * Fully cache the free space first so that our pin removes the free space
2646          * from the cache.
2647          */
2648         ret = btrfs_cache_block_group(cache, true);
2649         if (ret)
2650                 goto out;
2651
2652         pin_down_extent(trans, cache, eb->start, eb->len, 0);
2653
2654         /* remove us from the free space cache (if we're there at all) */
2655         ret = btrfs_remove_free_space(cache, eb->start, eb->len);
2656 out:
2657         btrfs_put_block_group(cache);
2658         return ret;
2659 }
2660
2661 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2662                                    u64 start, u64 num_bytes)
2663 {
2664         int ret;
2665         struct btrfs_block_group *block_group;
2666
2667         block_group = btrfs_lookup_block_group(fs_info, start);
2668         if (!block_group)
2669                 return -EINVAL;
2670
2671         ret = btrfs_cache_block_group(block_group, true);
2672         if (ret)
2673                 goto out;
2674
2675         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2676 out:
2677         btrfs_put_block_group(block_group);
2678         return ret;
2679 }
2680
2681 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2682 {
2683         struct btrfs_fs_info *fs_info = eb->fs_info;
2684         struct btrfs_file_extent_item *item;
2685         struct btrfs_key key;
2686         int found_type;
2687         int i;
2688         int ret = 0;
2689
2690         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2691                 return 0;
2692
2693         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2694                 btrfs_item_key_to_cpu(eb, &key, i);
2695                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2696                         continue;
2697                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2698                 found_type = btrfs_file_extent_type(eb, item);
2699                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2700                         continue;
2701                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2702                         continue;
2703                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2704                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2705                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2706                 if (ret)
2707                         break;
2708         }
2709
2710         return ret;
2711 }
2712
2713 static void
2714 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2715 {
2716         atomic_inc(&bg->reservations);
2717 }
2718
2719 /*
2720  * Returns the free cluster for the given space info and sets empty_cluster to
2721  * what it should be based on the mount options.
2722  */
2723 static struct btrfs_free_cluster *
2724 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2725                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2726 {
2727         struct btrfs_free_cluster *ret = NULL;
2728
2729         *empty_cluster = 0;
2730         if (btrfs_mixed_space_info(space_info))
2731                 return ret;
2732
2733         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2734                 ret = &fs_info->meta_alloc_cluster;
2735                 if (btrfs_test_opt(fs_info, SSD))
2736                         *empty_cluster = SZ_2M;
2737                 else
2738                         *empty_cluster = SZ_64K;
2739         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2740                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2741                 *empty_cluster = SZ_2M;
2742                 ret = &fs_info->data_alloc_cluster;
2743         }
2744
2745         return ret;
2746 }
2747
2748 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2749                               u64 start, u64 end,
2750                               const bool return_free_space)
2751 {
2752         struct btrfs_block_group *cache = NULL;
2753         struct btrfs_space_info *space_info;
2754         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2755         struct btrfs_free_cluster *cluster = NULL;
2756         u64 len;
2757         u64 total_unpinned = 0;
2758         u64 empty_cluster = 0;
2759         bool readonly;
2760
2761         while (start <= end) {
2762                 readonly = false;
2763                 if (!cache ||
2764                     start >= cache->start + cache->length) {
2765                         if (cache)
2766                                 btrfs_put_block_group(cache);
2767                         total_unpinned = 0;
2768                         cache = btrfs_lookup_block_group(fs_info, start);
2769                         BUG_ON(!cache); /* Logic error */
2770
2771                         cluster = fetch_cluster_info(fs_info,
2772                                                      cache->space_info,
2773                                                      &empty_cluster);
2774                         empty_cluster <<= 1;
2775                 }
2776
2777                 len = cache->start + cache->length - start;
2778                 len = min(len, end + 1 - start);
2779
2780                 if (return_free_space)
2781                         btrfs_add_free_space(cache, start, len);
2782
2783                 start += len;
2784                 total_unpinned += len;
2785                 space_info = cache->space_info;
2786
2787                 /*
2788                  * If this space cluster has been marked as fragmented and we've
2789                  * unpinned enough in this block group to potentially allow a
2790                  * cluster to be created inside of it go ahead and clear the
2791                  * fragmented check.
2792                  */
2793                 if (cluster && cluster->fragmented &&
2794                     total_unpinned > empty_cluster) {
2795                         spin_lock(&cluster->lock);
2796                         cluster->fragmented = 0;
2797                         spin_unlock(&cluster->lock);
2798                 }
2799
2800                 spin_lock(&space_info->lock);
2801                 spin_lock(&cache->lock);
2802                 cache->pinned -= len;
2803                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2804                 space_info->max_extent_size = 0;
2805                 if (cache->ro) {
2806                         space_info->bytes_readonly += len;
2807                         readonly = true;
2808                 } else if (btrfs_is_zoned(fs_info)) {
2809                         /* Need reset before reusing in a zoned block group */
2810                         space_info->bytes_zone_unusable += len;
2811                         readonly = true;
2812                 }
2813                 spin_unlock(&cache->lock);
2814                 if (!readonly && return_free_space &&
2815                     global_rsv->space_info == space_info) {
2816                         spin_lock(&global_rsv->lock);
2817                         if (!global_rsv->full) {
2818                                 u64 to_add = min(len, global_rsv->size -
2819                                                       global_rsv->reserved);
2820
2821                                 global_rsv->reserved += to_add;
2822                                 btrfs_space_info_update_bytes_may_use(fs_info,
2823                                                 space_info, to_add);
2824                                 if (global_rsv->reserved >= global_rsv->size)
2825                                         global_rsv->full = 1;
2826                                 len -= to_add;
2827                         }
2828                         spin_unlock(&global_rsv->lock);
2829                 }
2830                 /* Add to any tickets we may have */
2831                 if (!readonly && return_free_space && len)
2832                         btrfs_try_granting_tickets(fs_info, space_info);
2833                 spin_unlock(&space_info->lock);
2834         }
2835
2836         if (cache)
2837                 btrfs_put_block_group(cache);
2838         return 0;
2839 }
2840
2841 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2842 {
2843         struct btrfs_fs_info *fs_info = trans->fs_info;
2844         struct btrfs_block_group *block_group, *tmp;
2845         struct list_head *deleted_bgs;
2846         struct extent_io_tree *unpin;
2847         u64 start;
2848         u64 end;
2849         int ret;
2850
2851         unpin = &trans->transaction->pinned_extents;
2852
2853         while (!TRANS_ABORTED(trans)) {
2854                 struct extent_state *cached_state = NULL;
2855
2856                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2857                 if (!find_first_extent_bit(unpin, 0, &start, &end,
2858                                            EXTENT_DIRTY, &cached_state)) {
2859                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2860                         break;
2861                 }
2862
2863                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2864                         ret = btrfs_discard_extent(fs_info, start,
2865                                                    end + 1 - start, NULL);
2866
2867                 clear_extent_dirty(unpin, start, end, &cached_state);
2868                 unpin_extent_range(fs_info, start, end, true);
2869                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2870                 free_extent_state(cached_state);
2871                 cond_resched();
2872         }
2873
2874         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2875                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2876                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2877         }
2878
2879         /*
2880          * Transaction is finished.  We don't need the lock anymore.  We
2881          * do need to clean up the block groups in case of a transaction
2882          * abort.
2883          */
2884         deleted_bgs = &trans->transaction->deleted_bgs;
2885         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2886                 u64 trimmed = 0;
2887
2888                 ret = -EROFS;
2889                 if (!TRANS_ABORTED(trans))
2890                         ret = btrfs_discard_extent(fs_info,
2891                                                    block_group->start,
2892                                                    block_group->length,
2893                                                    &trimmed);
2894
2895                 list_del_init(&block_group->bg_list);
2896                 btrfs_unfreeze_block_group(block_group);
2897                 btrfs_put_block_group(block_group);
2898
2899                 if (ret) {
2900                         const char *errstr = btrfs_decode_error(ret);
2901                         btrfs_warn(fs_info,
2902                            "discard failed while removing blockgroup: errno=%d %s",
2903                                    ret, errstr);
2904                 }
2905         }
2906
2907         return 0;
2908 }
2909
2910 /*
2911  * Parse an extent item's inline extents looking for a simple quotas owner ref.
2912  *
2913  * @fs_info:    the btrfs_fs_info for this mount
2914  * @leaf:       a leaf in the extent tree containing the extent item
2915  * @slot:       the slot in the leaf where the extent item is found
2916  *
2917  * Returns the objectid of the root that originally allocated the extent item
2918  * if the inline owner ref is expected and present, otherwise 0.
2919  *
2920  * If an extent item has an owner ref item, it will be the first inline ref
2921  * item. Therefore the logic is to check whether there are any inline ref
2922  * items, then check the type of the first one.
2923  */
2924 u64 btrfs_get_extent_owner_root(struct btrfs_fs_info *fs_info,
2925                                 struct extent_buffer *leaf, int slot)
2926 {
2927         struct btrfs_extent_item *ei;
2928         struct btrfs_extent_inline_ref *iref;
2929         struct btrfs_extent_owner_ref *oref;
2930         unsigned long ptr;
2931         unsigned long end;
2932         int type;
2933
2934         if (!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA))
2935                 return 0;
2936
2937         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
2938         ptr = (unsigned long)(ei + 1);
2939         end = (unsigned long)ei + btrfs_item_size(leaf, slot);
2940
2941         /* No inline ref items of any kind, can't check type. */
2942         if (ptr == end)
2943                 return 0;
2944
2945         iref = (struct btrfs_extent_inline_ref *)ptr;
2946         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
2947
2948         /* We found an owner ref, get the root out of it. */
2949         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
2950                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
2951                 return btrfs_extent_owner_ref_root_id(leaf, oref);
2952         }
2953
2954         /* We have inline refs, but not an owner ref. */
2955         return 0;
2956 }
2957
2958 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2959                                      u64 bytenr, struct btrfs_squota_delta *delta)
2960 {
2961         int ret;
2962         u64 num_bytes = delta->num_bytes;
2963
2964         if (delta->is_data) {
2965                 struct btrfs_root *csum_root;
2966
2967                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2968                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2969                 if (ret) {
2970                         btrfs_abort_transaction(trans, ret);
2971                         return ret;
2972                 }
2973
2974                 ret = btrfs_delete_raid_extent(trans, bytenr, num_bytes);
2975                 if (ret) {
2976                         btrfs_abort_transaction(trans, ret);
2977                         return ret;
2978                 }
2979         }
2980
2981         ret = btrfs_record_squota_delta(trans->fs_info, delta);
2982         if (ret) {
2983                 btrfs_abort_transaction(trans, ret);
2984                 return ret;
2985         }
2986
2987         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2988         if (ret) {
2989                 btrfs_abort_transaction(trans, ret);
2990                 return ret;
2991         }
2992
2993         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2994         if (ret)
2995                 btrfs_abort_transaction(trans, ret);
2996
2997         return ret;
2998 }
2999
3000 #define abort_and_dump(trans, path, fmt, args...)       \
3001 ({                                                      \
3002         btrfs_abort_transaction(trans, -EUCLEAN);       \
3003         btrfs_print_leaf(path->nodes[0]);               \
3004         btrfs_crit(trans->fs_info, fmt, ##args);        \
3005 })
3006
3007 /*
3008  * Drop one or more refs of @node.
3009  *
3010  * 1. Locate the extent refs.
3011  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
3012  *    Locate it, then reduce the refs number or remove the ref line completely.
3013  *
3014  * 2. Update the refs count in EXTENT/METADATA_ITEM
3015  *
3016  * Inline backref case:
3017  *
3018  * in extent tree we have:
3019  *
3020  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3021  *              refs 2 gen 6 flags DATA
3022  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3023  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
3024  *
3025  * This function gets called with:
3026  *
3027  *    node->bytenr = 13631488
3028  *    node->num_bytes = 1048576
3029  *    root_objectid = FS_TREE
3030  *    owner_objectid = 257
3031  *    owner_offset = 0
3032  *    refs_to_drop = 1
3033  *
3034  * Then we should get some like:
3035  *
3036  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3037  *              refs 1 gen 6 flags DATA
3038  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3039  *
3040  * Keyed backref case:
3041  *
3042  * in extent tree we have:
3043  *
3044  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3045  *              refs 754 gen 6 flags DATA
3046  *      [...]
3047  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
3048  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
3049  *
3050  * This function get called with:
3051  *
3052  *    node->bytenr = 13631488
3053  *    node->num_bytes = 1048576
3054  *    root_objectid = FS_TREE
3055  *    owner_objectid = 866
3056  *    owner_offset = 0
3057  *    refs_to_drop = 1
3058  *
3059  * Then we should get some like:
3060  *
3061  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3062  *              refs 753 gen 6 flags DATA
3063  *
3064  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
3065  */
3066 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
3067                                struct btrfs_delayed_ref_head *href,
3068                                struct btrfs_delayed_ref_node *node, u64 parent,
3069                                u64 root_objectid, u64 owner_objectid,
3070                                u64 owner_offset,
3071                                struct btrfs_delayed_extent_op *extent_op)
3072 {
3073         struct btrfs_fs_info *info = trans->fs_info;
3074         struct btrfs_key key;
3075         struct btrfs_path *path;
3076         struct btrfs_root *extent_root;
3077         struct extent_buffer *leaf;
3078         struct btrfs_extent_item *ei;
3079         struct btrfs_extent_inline_ref *iref;
3080         int ret;
3081         int is_data;
3082         int extent_slot = 0;
3083         int found_extent = 0;
3084         int num_to_del = 1;
3085         int refs_to_drop = node->ref_mod;
3086         u32 item_size;
3087         u64 refs;
3088         u64 bytenr = node->bytenr;
3089         u64 num_bytes = node->num_bytes;
3090         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
3091         u64 delayed_ref_root = href->owning_root;
3092
3093         extent_root = btrfs_extent_root(info, bytenr);
3094         ASSERT(extent_root);
3095
3096         path = btrfs_alloc_path();
3097         if (!path)
3098                 return -ENOMEM;
3099
3100         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
3101
3102         if (!is_data && refs_to_drop != 1) {
3103                 btrfs_crit(info,
3104 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
3105                            node->bytenr, refs_to_drop);
3106                 ret = -EINVAL;
3107                 btrfs_abort_transaction(trans, ret);
3108                 goto out;
3109         }
3110
3111         if (is_data)
3112                 skinny_metadata = false;
3113
3114         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
3115                                     parent, root_objectid, owner_objectid,
3116                                     owner_offset);
3117         if (ret == 0) {
3118                 /*
3119                  * Either the inline backref or the SHARED_DATA_REF/
3120                  * SHARED_BLOCK_REF is found
3121                  *
3122                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
3123                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
3124                  */
3125                 extent_slot = path->slots[0];
3126                 while (extent_slot >= 0) {
3127                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3128                                               extent_slot);
3129                         if (key.objectid != bytenr)
3130                                 break;
3131                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3132                             key.offset == num_bytes) {
3133                                 found_extent = 1;
3134                                 break;
3135                         }
3136                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3137                             key.offset == owner_objectid) {
3138                                 found_extent = 1;
3139                                 break;
3140                         }
3141
3142                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3143                         if (path->slots[0] - extent_slot > 5)
3144                                 break;
3145                         extent_slot--;
3146                 }
3147
3148                 if (!found_extent) {
3149                         if (iref) {
3150                                 abort_and_dump(trans, path,
3151 "invalid iref slot %u, no EXTENT/METADATA_ITEM found but has inline extent ref",
3152                                            path->slots[0]);
3153                                 ret = -EUCLEAN;
3154                                 goto out;
3155                         }
3156                         /* Must be SHARED_* item, remove the backref first */
3157                         ret = remove_extent_backref(trans, extent_root, path,
3158                                                     NULL, refs_to_drop, is_data);
3159                         if (ret) {
3160                                 btrfs_abort_transaction(trans, ret);
3161                                 goto out;
3162                         }
3163                         btrfs_release_path(path);
3164
3165                         /* Slow path to locate EXTENT/METADATA_ITEM */
3166                         key.objectid = bytenr;
3167                         key.type = BTRFS_EXTENT_ITEM_KEY;
3168                         key.offset = num_bytes;
3169
3170                         if (!is_data && skinny_metadata) {
3171                                 key.type = BTRFS_METADATA_ITEM_KEY;
3172                                 key.offset = owner_objectid;
3173                         }
3174
3175                         ret = btrfs_search_slot(trans, extent_root,
3176                                                 &key, path, -1, 1);
3177                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3178                                 /*
3179                                  * Couldn't find our skinny metadata item,
3180                                  * see if we have ye olde extent item.
3181                                  */
3182                                 path->slots[0]--;
3183                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3184                                                       path->slots[0]);
3185                                 if (key.objectid == bytenr &&
3186                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3187                                     key.offset == num_bytes)
3188                                         ret = 0;
3189                         }
3190
3191                         if (ret > 0 && skinny_metadata) {
3192                                 skinny_metadata = false;
3193                                 key.objectid = bytenr;
3194                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3195                                 key.offset = num_bytes;
3196                                 btrfs_release_path(path);
3197                                 ret = btrfs_search_slot(trans, extent_root,
3198                                                         &key, path, -1, 1);
3199                         }
3200
3201                         if (ret) {
3202                                 if (ret > 0)
3203                                         btrfs_print_leaf(path->nodes[0]);
3204                                 btrfs_err(info,
3205                         "umm, got %d back from search, was looking for %llu, slot %d",
3206                                           ret, bytenr, path->slots[0]);
3207                         }
3208                         if (ret < 0) {
3209                                 btrfs_abort_transaction(trans, ret);
3210                                 goto out;
3211                         }
3212                         extent_slot = path->slots[0];
3213                 }
3214         } else if (WARN_ON(ret == -ENOENT)) {
3215                 abort_and_dump(trans, path,
3216 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu slot %d",
3217                                bytenr, parent, root_objectid, owner_objectid,
3218                                owner_offset, path->slots[0]);
3219                 goto out;
3220         } else {
3221                 btrfs_abort_transaction(trans, ret);
3222                 goto out;
3223         }
3224
3225         leaf = path->nodes[0];
3226         item_size = btrfs_item_size(leaf, extent_slot);
3227         if (unlikely(item_size < sizeof(*ei))) {
3228                 ret = -EUCLEAN;
3229                 btrfs_err(trans->fs_info,
3230                           "unexpected extent item size, has %u expect >= %zu",
3231                           item_size, sizeof(*ei));
3232                 btrfs_abort_transaction(trans, ret);
3233                 goto out;
3234         }
3235         ei = btrfs_item_ptr(leaf, extent_slot,
3236                             struct btrfs_extent_item);
3237         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3238             key.type == BTRFS_EXTENT_ITEM_KEY) {
3239                 struct btrfs_tree_block_info *bi;
3240
3241                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3242                         abort_and_dump(trans, path,
3243 "invalid extent item size for key (%llu, %u, %llu) slot %u owner %llu, has %u expect >= %zu",
3244                                        key.objectid, key.type, key.offset,
3245                                        path->slots[0], owner_objectid, item_size,
3246                                        sizeof(*ei) + sizeof(*bi));
3247                         ret = -EUCLEAN;
3248                         goto out;
3249                 }
3250                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3251                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3252         }
3253
3254         refs = btrfs_extent_refs(leaf, ei);
3255         if (refs < refs_to_drop) {
3256                 abort_and_dump(trans, path,
3257                 "trying to drop %d refs but we only have %llu for bytenr %llu slot %u",
3258                                refs_to_drop, refs, bytenr, path->slots[0]);
3259                 ret = -EUCLEAN;
3260                 goto out;
3261         }
3262         refs -= refs_to_drop;
3263
3264         if (refs > 0) {
3265                 if (extent_op)
3266                         __run_delayed_extent_op(extent_op, leaf, ei);
3267                 /*
3268                  * In the case of inline back ref, reference count will
3269                  * be updated by remove_extent_backref
3270                  */
3271                 if (iref) {
3272                         if (!found_extent) {
3273                                 abort_and_dump(trans, path,
3274 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found, slot %u",
3275                                                path->slots[0]);
3276                                 ret = -EUCLEAN;
3277                                 goto out;
3278                         }
3279                 } else {
3280                         btrfs_set_extent_refs(leaf, ei, refs);
3281                         btrfs_mark_buffer_dirty(trans, leaf);
3282                 }
3283                 if (found_extent) {
3284                         ret = remove_extent_backref(trans, extent_root, path,
3285                                                     iref, refs_to_drop, is_data);
3286                         if (ret) {
3287                                 btrfs_abort_transaction(trans, ret);
3288                                 goto out;
3289                         }
3290                 }
3291         } else {
3292                 struct btrfs_squota_delta delta = {
3293                         .root = delayed_ref_root,
3294                         .num_bytes = num_bytes,
3295                         .rsv_bytes = 0,
3296                         .is_data = is_data,
3297                         .is_inc = false,
3298                         .generation = btrfs_extent_generation(leaf, ei),
3299                 };
3300
3301                 /* In this branch refs == 1 */
3302                 if (found_extent) {
3303                         if (is_data && refs_to_drop !=
3304                             extent_data_ref_count(path, iref)) {
3305                                 abort_and_dump(trans, path,
3306                 "invalid refs_to_drop, current refs %u refs_to_drop %u slot %u",
3307                                                extent_data_ref_count(path, iref),
3308                                                refs_to_drop, path->slots[0]);
3309                                 ret = -EUCLEAN;
3310                                 goto out;
3311                         }
3312                         if (iref) {
3313                                 if (path->slots[0] != extent_slot) {
3314                                         abort_and_dump(trans, path,
3315 "invalid iref, extent item key (%llu %u %llu) slot %u doesn't have wanted iref",
3316                                                        key.objectid, key.type,
3317                                                        key.offset, path->slots[0]);
3318                                         ret = -EUCLEAN;
3319                                         goto out;
3320                                 }
3321                         } else {
3322                                 /*
3323                                  * No inline ref, we must be at SHARED_* item,
3324                                  * And it's single ref, it must be:
3325                                  * |    extent_slot       ||extent_slot + 1|
3326                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3327                                  */
3328                                 if (path->slots[0] != extent_slot + 1) {
3329                                         abort_and_dump(trans, path,
3330         "invalid SHARED_* item slot %u, previous item is not EXTENT/METADATA_ITEM",
3331                                                        path->slots[0]);
3332                                         ret = -EUCLEAN;
3333                                         goto out;
3334                                 }
3335                                 path->slots[0] = extent_slot;
3336                                 num_to_del = 2;
3337                         }
3338                 }
3339                 /*
3340                  * We can't infer the data owner from the delayed ref, so we need
3341                  * to try to get it from the owning ref item.
3342                  *
3343                  * If it is not present, then that extent was not written under
3344                  * simple quotas mode, so we don't need to account for its deletion.
3345                  */
3346                 if (is_data)
3347                         delta.root = btrfs_get_extent_owner_root(trans->fs_info,
3348                                                                  leaf, extent_slot);
3349
3350                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3351                                       num_to_del);
3352                 if (ret) {
3353                         btrfs_abort_transaction(trans, ret);
3354                         goto out;
3355                 }
3356                 btrfs_release_path(path);
3357
3358                 ret = do_free_extent_accounting(trans, bytenr, &delta);
3359         }
3360         btrfs_release_path(path);
3361
3362 out:
3363         btrfs_free_path(path);
3364         return ret;
3365 }
3366
3367 /*
3368  * when we free an block, it is possible (and likely) that we free the last
3369  * delayed ref for that extent as well.  This searches the delayed ref tree for
3370  * a given extent, and if there are no other delayed refs to be processed, it
3371  * removes it from the tree.
3372  */
3373 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3374                                       u64 bytenr)
3375 {
3376         struct btrfs_delayed_ref_head *head;
3377         struct btrfs_delayed_ref_root *delayed_refs;
3378         int ret = 0;
3379
3380         delayed_refs = &trans->transaction->delayed_refs;
3381         spin_lock(&delayed_refs->lock);
3382         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3383         if (!head)
3384                 goto out_delayed_unlock;
3385
3386         spin_lock(&head->lock);
3387         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3388                 goto out;
3389
3390         if (cleanup_extent_op(head) != NULL)
3391                 goto out;
3392
3393         /*
3394          * waiting for the lock here would deadlock.  If someone else has it
3395          * locked they are already in the process of dropping it anyway
3396          */
3397         if (!mutex_trylock(&head->mutex))
3398                 goto out;
3399
3400         btrfs_delete_ref_head(delayed_refs, head);
3401         head->processing = false;
3402
3403         spin_unlock(&head->lock);
3404         spin_unlock(&delayed_refs->lock);
3405
3406         BUG_ON(head->extent_op);
3407         if (head->must_insert_reserved)
3408                 ret = 1;
3409
3410         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3411         mutex_unlock(&head->mutex);
3412         btrfs_put_delayed_ref_head(head);
3413         return ret;
3414 out:
3415         spin_unlock(&head->lock);
3416
3417 out_delayed_unlock:
3418         spin_unlock(&delayed_refs->lock);
3419         return 0;
3420 }
3421
3422 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3423                            u64 root_id,
3424                            struct extent_buffer *buf,
3425                            u64 parent, int last_ref)
3426 {
3427         struct btrfs_fs_info *fs_info = trans->fs_info;
3428         struct btrfs_ref generic_ref = { 0 };
3429         int ret;
3430
3431         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3432                                buf->start, buf->len, parent, btrfs_header_owner(buf));
3433         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3434                             root_id, 0, false);
3435
3436         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3437                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3438                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3439                 BUG_ON(ret); /* -ENOMEM */
3440         }
3441
3442         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3443                 struct btrfs_block_group *cache;
3444                 bool must_pin = false;
3445
3446                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3447                         ret = check_ref_cleanup(trans, buf->start);
3448                         if (!ret) {
3449                                 btrfs_redirty_list_add(trans->transaction, buf);
3450                                 goto out;
3451                         }
3452                 }
3453
3454                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3455
3456                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3457                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3458                         btrfs_put_block_group(cache);
3459                         goto out;
3460                 }
3461
3462                 /*
3463                  * If there are tree mod log users we may have recorded mod log
3464                  * operations for this node.  If we re-allocate this node we
3465                  * could replay operations on this node that happened when it
3466                  * existed in a completely different root.  For example if it
3467                  * was part of root A, then was reallocated to root B, and we
3468                  * are doing a btrfs_old_search_slot(root b), we could replay
3469                  * operations that happened when the block was part of root A,
3470                  * giving us an inconsistent view of the btree.
3471                  *
3472                  * We are safe from races here because at this point no other
3473                  * node or root points to this extent buffer, so if after this
3474                  * check a new tree mod log user joins we will not have an
3475                  * existing log of operations on this node that we have to
3476                  * contend with.
3477                  */
3478                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3479                         must_pin = true;
3480
3481                 if (must_pin || btrfs_is_zoned(fs_info)) {
3482                         btrfs_redirty_list_add(trans->transaction, buf);
3483                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3484                         btrfs_put_block_group(cache);
3485                         goto out;
3486                 }
3487
3488                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3489
3490                 btrfs_add_free_space(cache, buf->start, buf->len);
3491                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3492                 btrfs_put_block_group(cache);
3493                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3494         }
3495 out:
3496         if (last_ref) {
3497                 /*
3498                  * Deleting the buffer, clear the corrupt flag since it doesn't
3499                  * matter anymore.
3500                  */
3501                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3502         }
3503 }
3504
3505 /* Can return -ENOMEM */
3506 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3507 {
3508         struct btrfs_fs_info *fs_info = trans->fs_info;
3509         int ret;
3510
3511         if (btrfs_is_testing(fs_info))
3512                 return 0;
3513
3514         /*
3515          * tree log blocks never actually go into the extent allocation
3516          * tree, just update pinning info and exit early.
3517          */
3518         if ((ref->type == BTRFS_REF_METADATA &&
3519              ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID) ||
3520             (ref->type == BTRFS_REF_DATA &&
3521              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3522                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3523                 ret = 0;
3524         } else if (ref->type == BTRFS_REF_METADATA) {
3525                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3526         } else {
3527                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3528         }
3529
3530         if (!((ref->type == BTRFS_REF_METADATA &&
3531                ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID) ||
3532               (ref->type == BTRFS_REF_DATA &&
3533                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3534                 btrfs_ref_tree_mod(fs_info, ref);
3535
3536         return ret;
3537 }
3538
3539 enum btrfs_loop_type {
3540         /*
3541          * Start caching block groups but do not wait for progress or for them
3542          * to be done.
3543          */
3544         LOOP_CACHING_NOWAIT,
3545
3546         /*
3547          * Wait for the block group free_space >= the space we're waiting for if
3548          * the block group isn't cached.
3549          */
3550         LOOP_CACHING_WAIT,
3551
3552         /*
3553          * Allow allocations to happen from block groups that do not yet have a
3554          * size classification.
3555          */
3556         LOOP_UNSET_SIZE_CLASS,
3557
3558         /*
3559          * Allocate a chunk and then retry the allocation.
3560          */
3561         LOOP_ALLOC_CHUNK,
3562
3563         /*
3564          * Ignore the size class restrictions for this allocation.
3565          */
3566         LOOP_WRONG_SIZE_CLASS,
3567
3568         /*
3569          * Ignore the empty size, only try to allocate the number of bytes
3570          * needed for this allocation.
3571          */
3572         LOOP_NO_EMPTY_SIZE,
3573 };
3574
3575 static inline void
3576 btrfs_lock_block_group(struct btrfs_block_group *cache,
3577                        int delalloc)
3578 {
3579         if (delalloc)
3580                 down_read(&cache->data_rwsem);
3581 }
3582
3583 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3584                        int delalloc)
3585 {
3586         btrfs_get_block_group(cache);
3587         if (delalloc)
3588                 down_read(&cache->data_rwsem);
3589 }
3590
3591 static struct btrfs_block_group *btrfs_lock_cluster(
3592                    struct btrfs_block_group *block_group,
3593                    struct btrfs_free_cluster *cluster,
3594                    int delalloc)
3595         __acquires(&cluster->refill_lock)
3596 {
3597         struct btrfs_block_group *used_bg = NULL;
3598
3599         spin_lock(&cluster->refill_lock);
3600         while (1) {
3601                 used_bg = cluster->block_group;
3602                 if (!used_bg)
3603                         return NULL;
3604
3605                 if (used_bg == block_group)
3606                         return used_bg;
3607
3608                 btrfs_get_block_group(used_bg);
3609
3610                 if (!delalloc)
3611                         return used_bg;
3612
3613                 if (down_read_trylock(&used_bg->data_rwsem))
3614                         return used_bg;
3615
3616                 spin_unlock(&cluster->refill_lock);
3617
3618                 /* We should only have one-level nested. */
3619                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3620
3621                 spin_lock(&cluster->refill_lock);
3622                 if (used_bg == cluster->block_group)
3623                         return used_bg;
3624
3625                 up_read(&used_bg->data_rwsem);
3626                 btrfs_put_block_group(used_bg);
3627         }
3628 }
3629
3630 static inline void
3631 btrfs_release_block_group(struct btrfs_block_group *cache,
3632                          int delalloc)
3633 {
3634         if (delalloc)
3635                 up_read(&cache->data_rwsem);
3636         btrfs_put_block_group(cache);
3637 }
3638
3639 /*
3640  * Helper function for find_free_extent().
3641  *
3642  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3643  * Return >0 to inform caller that we find nothing
3644  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3645  */
3646 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3647                                       struct find_free_extent_ctl *ffe_ctl,
3648                                       struct btrfs_block_group **cluster_bg_ret)
3649 {
3650         struct btrfs_block_group *cluster_bg;
3651         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3652         u64 aligned_cluster;
3653         u64 offset;
3654         int ret;
3655
3656         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3657         if (!cluster_bg)
3658                 goto refill_cluster;
3659         if (cluster_bg != bg && (cluster_bg->ro ||
3660             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3661                 goto release_cluster;
3662
3663         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3664                         ffe_ctl->num_bytes, cluster_bg->start,
3665                         &ffe_ctl->max_extent_size);
3666         if (offset) {
3667                 /* We have a block, we're done */
3668                 spin_unlock(&last_ptr->refill_lock);
3669                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3670                 *cluster_bg_ret = cluster_bg;
3671                 ffe_ctl->found_offset = offset;
3672                 return 0;
3673         }
3674         WARN_ON(last_ptr->block_group != cluster_bg);
3675
3676 release_cluster:
3677         /*
3678          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3679          * lets just skip it and let the allocator find whatever block it can
3680          * find. If we reach this point, we will have tried the cluster
3681          * allocator plenty of times and not have found anything, so we are
3682          * likely way too fragmented for the clustering stuff to find anything.
3683          *
3684          * However, if the cluster is taken from the current block group,
3685          * release the cluster first, so that we stand a better chance of
3686          * succeeding in the unclustered allocation.
3687          */
3688         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3689                 spin_unlock(&last_ptr->refill_lock);
3690                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3691                 return -ENOENT;
3692         }
3693
3694         /* This cluster didn't work out, free it and start over */
3695         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3696
3697         if (cluster_bg != bg)
3698                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3699
3700 refill_cluster:
3701         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3702                 spin_unlock(&last_ptr->refill_lock);
3703                 return -ENOENT;
3704         }
3705
3706         aligned_cluster = max_t(u64,
3707                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3708                         bg->full_stripe_len);
3709         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3710                         ffe_ctl->num_bytes, aligned_cluster);
3711         if (ret == 0) {
3712                 /* Now pull our allocation out of this cluster */
3713                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3714                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3715                                 &ffe_ctl->max_extent_size);
3716                 if (offset) {
3717                         /* We found one, proceed */
3718                         spin_unlock(&last_ptr->refill_lock);
3719                         ffe_ctl->found_offset = offset;
3720                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3721                         return 0;
3722                 }
3723         }
3724         /*
3725          * At this point we either didn't find a cluster or we weren't able to
3726          * allocate a block from our cluster.  Free the cluster we've been
3727          * trying to use, and go to the next block group.
3728          */
3729         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3730         spin_unlock(&last_ptr->refill_lock);
3731         return 1;
3732 }
3733
3734 /*
3735  * Return >0 to inform caller that we find nothing
3736  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3737  */
3738 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3739                                         struct find_free_extent_ctl *ffe_ctl)
3740 {
3741         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3742         u64 offset;
3743
3744         /*
3745          * We are doing an unclustered allocation, set the fragmented flag so
3746          * we don't bother trying to setup a cluster again until we get more
3747          * space.
3748          */
3749         if (unlikely(last_ptr)) {
3750                 spin_lock(&last_ptr->lock);
3751                 last_ptr->fragmented = 1;
3752                 spin_unlock(&last_ptr->lock);
3753         }
3754         if (ffe_ctl->cached) {
3755                 struct btrfs_free_space_ctl *free_space_ctl;
3756
3757                 free_space_ctl = bg->free_space_ctl;
3758                 spin_lock(&free_space_ctl->tree_lock);
3759                 if (free_space_ctl->free_space <
3760                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3761                     ffe_ctl->empty_size) {
3762                         ffe_ctl->total_free_space = max_t(u64,
3763                                         ffe_ctl->total_free_space,
3764                                         free_space_ctl->free_space);
3765                         spin_unlock(&free_space_ctl->tree_lock);
3766                         return 1;
3767                 }
3768                 spin_unlock(&free_space_ctl->tree_lock);
3769         }
3770
3771         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3772                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3773                         &ffe_ctl->max_extent_size);
3774         if (!offset)
3775                 return 1;
3776         ffe_ctl->found_offset = offset;
3777         return 0;
3778 }
3779
3780 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3781                                    struct find_free_extent_ctl *ffe_ctl,
3782                                    struct btrfs_block_group **bg_ret)
3783 {
3784         int ret;
3785
3786         /* We want to try and use the cluster allocator, so lets look there */
3787         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3788                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3789                 if (ret >= 0)
3790                         return ret;
3791                 /* ret == -ENOENT case falls through */
3792         }
3793
3794         return find_free_extent_unclustered(block_group, ffe_ctl);
3795 }
3796
3797 /*
3798  * Tree-log block group locking
3799  * ============================
3800  *
3801  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3802  * indicates the starting address of a block group, which is reserved only
3803  * for tree-log metadata.
3804  *
3805  * Lock nesting
3806  * ============
3807  *
3808  * space_info::lock
3809  *   block_group::lock
3810  *     fs_info::treelog_bg_lock
3811  */
3812
3813 /*
3814  * Simple allocator for sequential-only block group. It only allows sequential
3815  * allocation. No need to play with trees. This function also reserves the
3816  * bytes as in btrfs_add_reserved_bytes.
3817  */
3818 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3819                                struct find_free_extent_ctl *ffe_ctl,
3820                                struct btrfs_block_group **bg_ret)
3821 {
3822         struct btrfs_fs_info *fs_info = block_group->fs_info;
3823         struct btrfs_space_info *space_info = block_group->space_info;
3824         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3825         u64 start = block_group->start;
3826         u64 num_bytes = ffe_ctl->num_bytes;
3827         u64 avail;
3828         u64 bytenr = block_group->start;
3829         u64 log_bytenr;
3830         u64 data_reloc_bytenr;
3831         int ret = 0;
3832         bool skip = false;
3833
3834         ASSERT(btrfs_is_zoned(block_group->fs_info));
3835
3836         /*
3837          * Do not allow non-tree-log blocks in the dedicated tree-log block
3838          * group, and vice versa.
3839          */
3840         spin_lock(&fs_info->treelog_bg_lock);
3841         log_bytenr = fs_info->treelog_bg;
3842         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3843                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3844                 skip = true;
3845         spin_unlock(&fs_info->treelog_bg_lock);
3846         if (skip)
3847                 return 1;
3848
3849         /*
3850          * Do not allow non-relocation blocks in the dedicated relocation block
3851          * group, and vice versa.
3852          */
3853         spin_lock(&fs_info->relocation_bg_lock);
3854         data_reloc_bytenr = fs_info->data_reloc_bg;
3855         if (data_reloc_bytenr &&
3856             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3857              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3858                 skip = true;
3859         spin_unlock(&fs_info->relocation_bg_lock);
3860         if (skip)
3861                 return 1;
3862
3863         /* Check RO and no space case before trying to activate it */
3864         spin_lock(&block_group->lock);
3865         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3866                 ret = 1;
3867                 /*
3868                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3869                  * Return the error after taking the locks.
3870                  */
3871         }
3872         spin_unlock(&block_group->lock);
3873
3874         /* Metadata block group is activated at write time. */
3875         if (!ret && (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
3876             !btrfs_zone_activate(block_group)) {
3877                 ret = 1;
3878                 /*
3879                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3880                  * Return the error after taking the locks.
3881                  */
3882         }
3883
3884         spin_lock(&space_info->lock);
3885         spin_lock(&block_group->lock);
3886         spin_lock(&fs_info->treelog_bg_lock);
3887         spin_lock(&fs_info->relocation_bg_lock);
3888
3889         if (ret)
3890                 goto out;
3891
3892         ASSERT(!ffe_ctl->for_treelog ||
3893                block_group->start == fs_info->treelog_bg ||
3894                fs_info->treelog_bg == 0);
3895         ASSERT(!ffe_ctl->for_data_reloc ||
3896                block_group->start == fs_info->data_reloc_bg ||
3897                fs_info->data_reloc_bg == 0);
3898
3899         if (block_group->ro ||
3900             (!ffe_ctl->for_data_reloc &&
3901              test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))) {
3902                 ret = 1;
3903                 goto out;
3904         }
3905
3906         /*
3907          * Do not allow currently using block group to be tree-log dedicated
3908          * block group.
3909          */
3910         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3911             (block_group->used || block_group->reserved)) {
3912                 ret = 1;
3913                 goto out;
3914         }
3915
3916         /*
3917          * Do not allow currently used block group to be the data relocation
3918          * dedicated block group.
3919          */
3920         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3921             (block_group->used || block_group->reserved)) {
3922                 ret = 1;
3923                 goto out;
3924         }
3925
3926         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3927         avail = block_group->zone_capacity - block_group->alloc_offset;
3928         if (avail < num_bytes) {
3929                 if (ffe_ctl->max_extent_size < avail) {
3930                         /*
3931                          * With sequential allocator, free space is always
3932                          * contiguous
3933                          */
3934                         ffe_ctl->max_extent_size = avail;
3935                         ffe_ctl->total_free_space = avail;
3936                 }
3937                 ret = 1;
3938                 goto out;
3939         }
3940
3941         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3942                 fs_info->treelog_bg = block_group->start;
3943
3944         if (ffe_ctl->for_data_reloc) {
3945                 if (!fs_info->data_reloc_bg)
3946                         fs_info->data_reloc_bg = block_group->start;
3947                 /*
3948                  * Do not allow allocations from this block group, unless it is
3949                  * for data relocation. Compared to increasing the ->ro, setting
3950                  * the ->zoned_data_reloc_ongoing flag still allows nocow
3951                  * writers to come in. See btrfs_inc_nocow_writers().
3952                  *
3953                  * We need to disable an allocation to avoid an allocation of
3954                  * regular (non-relocation data) extent. With mix of relocation
3955                  * extents and regular extents, we can dispatch WRITE commands
3956                  * (for relocation extents) and ZONE APPEND commands (for
3957                  * regular extents) at the same time to the same zone, which
3958                  * easily break the write pointer.
3959                  *
3960                  * Also, this flag avoids this block group to be zone finished.
3961                  */
3962                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3963         }
3964
3965         ffe_ctl->found_offset = start + block_group->alloc_offset;
3966         block_group->alloc_offset += num_bytes;
3967         spin_lock(&ctl->tree_lock);
3968         ctl->free_space -= num_bytes;
3969         spin_unlock(&ctl->tree_lock);
3970
3971         /*
3972          * We do not check if found_offset is aligned to stripesize. The
3973          * address is anyway rewritten when using zone append writing.
3974          */
3975
3976         ffe_ctl->search_start = ffe_ctl->found_offset;
3977
3978 out:
3979         if (ret && ffe_ctl->for_treelog)
3980                 fs_info->treelog_bg = 0;
3981         if (ret && ffe_ctl->for_data_reloc)
3982                 fs_info->data_reloc_bg = 0;
3983         spin_unlock(&fs_info->relocation_bg_lock);
3984         spin_unlock(&fs_info->treelog_bg_lock);
3985         spin_unlock(&block_group->lock);
3986         spin_unlock(&space_info->lock);
3987         return ret;
3988 }
3989
3990 static int do_allocation(struct btrfs_block_group *block_group,
3991                          struct find_free_extent_ctl *ffe_ctl,
3992                          struct btrfs_block_group **bg_ret)
3993 {
3994         switch (ffe_ctl->policy) {
3995         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3996                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3997         case BTRFS_EXTENT_ALLOC_ZONED:
3998                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3999         default:
4000                 BUG();
4001         }
4002 }
4003
4004 static void release_block_group(struct btrfs_block_group *block_group,
4005                                 struct find_free_extent_ctl *ffe_ctl,
4006                                 int delalloc)
4007 {
4008         switch (ffe_ctl->policy) {
4009         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4010                 ffe_ctl->retry_uncached = false;
4011                 break;
4012         case BTRFS_EXTENT_ALLOC_ZONED:
4013                 /* Nothing to do */
4014                 break;
4015         default:
4016                 BUG();
4017         }
4018
4019         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
4020                ffe_ctl->index);
4021         btrfs_release_block_group(block_group, delalloc);
4022 }
4023
4024 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
4025                                    struct btrfs_key *ins)
4026 {
4027         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4028
4029         if (!ffe_ctl->use_cluster && last_ptr) {
4030                 spin_lock(&last_ptr->lock);
4031                 last_ptr->window_start = ins->objectid;
4032                 spin_unlock(&last_ptr->lock);
4033         }
4034 }
4035
4036 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
4037                          struct btrfs_key *ins)
4038 {
4039         switch (ffe_ctl->policy) {
4040         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4041                 found_extent_clustered(ffe_ctl, ins);
4042                 break;
4043         case BTRFS_EXTENT_ALLOC_ZONED:
4044                 /* Nothing to do */
4045                 break;
4046         default:
4047                 BUG();
4048         }
4049 }
4050
4051 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
4052                                     struct find_free_extent_ctl *ffe_ctl)
4053 {
4054         /* Block group's activeness is not a requirement for METADATA block groups. */
4055         if (!(ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA))
4056                 return 0;
4057
4058         /* If we can activate new zone, just allocate a chunk and use it */
4059         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
4060                 return 0;
4061
4062         /*
4063          * We already reached the max active zones. Try to finish one block
4064          * group to make a room for a new block group. This is only possible
4065          * for a data block group because btrfs_zone_finish() may need to wait
4066          * for a running transaction which can cause a deadlock for metadata
4067          * allocation.
4068          */
4069         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4070                 int ret = btrfs_zone_finish_one_bg(fs_info);
4071
4072                 if (ret == 1)
4073                         return 0;
4074                 else if (ret < 0)
4075                         return ret;
4076         }
4077
4078         /*
4079          * If we have enough free space left in an already active block group
4080          * and we can't activate any other zone now, do not allow allocating a
4081          * new chunk and let find_free_extent() retry with a smaller size.
4082          */
4083         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
4084                 return -ENOSPC;
4085
4086         /*
4087          * Even min_alloc_size is not left in any block groups. Since we cannot
4088          * activate a new block group, allocating it may not help. Let's tell a
4089          * caller to try again and hope it progress something by writing some
4090          * parts of the region. That is only possible for data block groups,
4091          * where a part of the region can be written.
4092          */
4093         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
4094                 return -EAGAIN;
4095
4096         /*
4097          * We cannot activate a new block group and no enough space left in any
4098          * block groups. So, allocating a new block group may not help. But,
4099          * there is nothing to do anyway, so let's go with it.
4100          */
4101         return 0;
4102 }
4103
4104 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4105                               struct find_free_extent_ctl *ffe_ctl)
4106 {
4107         switch (ffe_ctl->policy) {
4108         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4109                 return 0;
4110         case BTRFS_EXTENT_ALLOC_ZONED:
4111                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4112         default:
4113                 BUG();
4114         }
4115 }
4116
4117 /*
4118  * Return >0 means caller needs to re-search for free extent
4119  * Return 0 means we have the needed free extent.
4120  * Return <0 means we failed to locate any free extent.
4121  */
4122 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4123                                         struct btrfs_key *ins,
4124                                         struct find_free_extent_ctl *ffe_ctl,
4125                                         bool full_search)
4126 {
4127         struct btrfs_root *root = fs_info->chunk_root;
4128         int ret;
4129
4130         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4131             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4132                 ffe_ctl->orig_have_caching_bg = true;
4133
4134         if (ins->objectid) {
4135                 found_extent(ffe_ctl, ins);
4136                 return 0;
4137         }
4138
4139         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4140                 return 1;
4141
4142         ffe_ctl->index++;
4143         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4144                 return 1;
4145
4146         /* See the comments for btrfs_loop_type for an explanation of the phases. */
4147         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4148                 ffe_ctl->index = 0;
4149                 /*
4150                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
4151                  * any uncached bgs and we've already done a full search
4152                  * through.
4153                  */
4154                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
4155                     (!ffe_ctl->orig_have_caching_bg && full_search))
4156                         ffe_ctl->loop++;
4157                 ffe_ctl->loop++;
4158
4159                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4160                         struct btrfs_trans_handle *trans;
4161                         int exist = 0;
4162
4163                         /* Check if allocation policy allows to create a new chunk */
4164                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4165                         if (ret)
4166                                 return ret;
4167
4168                         trans = current->journal_info;
4169                         if (trans)
4170                                 exist = 1;
4171                         else
4172                                 trans = btrfs_join_transaction(root);
4173
4174                         if (IS_ERR(trans)) {
4175                                 ret = PTR_ERR(trans);
4176                                 return ret;
4177                         }
4178
4179                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4180                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4181
4182                         /* Do not bail out on ENOSPC since we can do more. */
4183                         if (ret == -ENOSPC) {
4184                                 ret = 0;
4185                                 ffe_ctl->loop++;
4186                         }
4187                         else if (ret < 0)
4188                                 btrfs_abort_transaction(trans, ret);
4189                         else
4190                                 ret = 0;
4191                         if (!exist)
4192                                 btrfs_end_transaction(trans);
4193                         if (ret)
4194                                 return ret;
4195                 }
4196
4197                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4198                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4199                                 return -ENOSPC;
4200
4201                         /*
4202                          * Don't loop again if we already have no empty_size and
4203                          * no empty_cluster.
4204                          */
4205                         if (ffe_ctl->empty_size == 0 &&
4206                             ffe_ctl->empty_cluster == 0)
4207                                 return -ENOSPC;
4208                         ffe_ctl->empty_size = 0;
4209                         ffe_ctl->empty_cluster = 0;
4210                 }
4211                 return 1;
4212         }
4213         return -ENOSPC;
4214 }
4215
4216 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4217                                               struct btrfs_block_group *bg)
4218 {
4219         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4220                 return true;
4221         if (!btrfs_block_group_should_use_size_class(bg))
4222                 return true;
4223         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4224                 return true;
4225         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4226             bg->size_class == BTRFS_BG_SZ_NONE)
4227                 return true;
4228         return ffe_ctl->size_class == bg->size_class;
4229 }
4230
4231 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4232                                         struct find_free_extent_ctl *ffe_ctl,
4233                                         struct btrfs_space_info *space_info,
4234                                         struct btrfs_key *ins)
4235 {
4236         /*
4237          * If our free space is heavily fragmented we may not be able to make
4238          * big contiguous allocations, so instead of doing the expensive search
4239          * for free space, simply return ENOSPC with our max_extent_size so we
4240          * can go ahead and search for a more manageable chunk.
4241          *
4242          * If our max_extent_size is large enough for our allocation simply
4243          * disable clustering since we will likely not be able to find enough
4244          * space to create a cluster and induce latency trying.
4245          */
4246         if (space_info->max_extent_size) {
4247                 spin_lock(&space_info->lock);
4248                 if (space_info->max_extent_size &&
4249                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4250                         ins->offset = space_info->max_extent_size;
4251                         spin_unlock(&space_info->lock);
4252                         return -ENOSPC;
4253                 } else if (space_info->max_extent_size) {
4254                         ffe_ctl->use_cluster = false;
4255                 }
4256                 spin_unlock(&space_info->lock);
4257         }
4258
4259         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4260                                                &ffe_ctl->empty_cluster);
4261         if (ffe_ctl->last_ptr) {
4262                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4263
4264                 spin_lock(&last_ptr->lock);
4265                 if (last_ptr->block_group)
4266                         ffe_ctl->hint_byte = last_ptr->window_start;
4267                 if (last_ptr->fragmented) {
4268                         /*
4269                          * We still set window_start so we can keep track of the
4270                          * last place we found an allocation to try and save
4271                          * some time.
4272                          */
4273                         ffe_ctl->hint_byte = last_ptr->window_start;
4274                         ffe_ctl->use_cluster = false;
4275                 }
4276                 spin_unlock(&last_ptr->lock);
4277         }
4278
4279         return 0;
4280 }
4281
4282 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4283                               struct find_free_extent_ctl *ffe_ctl,
4284                               struct btrfs_space_info *space_info,
4285                               struct btrfs_key *ins)
4286 {
4287         switch (ffe_ctl->policy) {
4288         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4289                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4290                                                     space_info, ins);
4291         case BTRFS_EXTENT_ALLOC_ZONED:
4292                 if (ffe_ctl->for_treelog) {
4293                         spin_lock(&fs_info->treelog_bg_lock);
4294                         if (fs_info->treelog_bg)
4295                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4296                         spin_unlock(&fs_info->treelog_bg_lock);
4297                 }
4298                 if (ffe_ctl->for_data_reloc) {
4299                         spin_lock(&fs_info->relocation_bg_lock);
4300                         if (fs_info->data_reloc_bg)
4301                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4302                         spin_unlock(&fs_info->relocation_bg_lock);
4303                 }
4304                 return 0;
4305         default:
4306                 BUG();
4307         }
4308 }
4309
4310 /*
4311  * walks the btree of allocated extents and find a hole of a given size.
4312  * The key ins is changed to record the hole:
4313  * ins->objectid == start position
4314  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4315  * ins->offset == the size of the hole.
4316  * Any available blocks before search_start are skipped.
4317  *
4318  * If there is no suitable free space, we will record the max size of
4319  * the free space extent currently.
4320  *
4321  * The overall logic and call chain:
4322  *
4323  * find_free_extent()
4324  * |- Iterate through all block groups
4325  * |  |- Get a valid block group
4326  * |  |- Try to do clustered allocation in that block group
4327  * |  |- Try to do unclustered allocation in that block group
4328  * |  |- Check if the result is valid
4329  * |  |  |- If valid, then exit
4330  * |  |- Jump to next block group
4331  * |
4332  * |- Push harder to find free extents
4333  *    |- If not found, re-iterate all block groups
4334  */
4335 static noinline int find_free_extent(struct btrfs_root *root,
4336                                      struct btrfs_key *ins,
4337                                      struct find_free_extent_ctl *ffe_ctl)
4338 {
4339         struct btrfs_fs_info *fs_info = root->fs_info;
4340         int ret = 0;
4341         int cache_block_group_error = 0;
4342         struct btrfs_block_group *block_group = NULL;
4343         struct btrfs_space_info *space_info;
4344         bool full_search = false;
4345
4346         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4347
4348         ffe_ctl->search_start = 0;
4349         /* For clustered allocation */
4350         ffe_ctl->empty_cluster = 0;
4351         ffe_ctl->last_ptr = NULL;
4352         ffe_ctl->use_cluster = true;
4353         ffe_ctl->have_caching_bg = false;
4354         ffe_ctl->orig_have_caching_bg = false;
4355         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4356         ffe_ctl->loop = 0;
4357         ffe_ctl->retry_uncached = false;
4358         ffe_ctl->cached = 0;
4359         ffe_ctl->max_extent_size = 0;
4360         ffe_ctl->total_free_space = 0;
4361         ffe_ctl->found_offset = 0;
4362         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4363         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4364
4365         if (btrfs_is_zoned(fs_info))
4366                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4367
4368         ins->type = BTRFS_EXTENT_ITEM_KEY;
4369         ins->objectid = 0;
4370         ins->offset = 0;
4371
4372         trace_find_free_extent(root, ffe_ctl);
4373
4374         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4375         if (!space_info) {
4376                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4377                 return -ENOSPC;
4378         }
4379
4380         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4381         if (ret < 0)
4382                 return ret;
4383
4384         ffe_ctl->search_start = max(ffe_ctl->search_start,
4385                                     first_logical_byte(fs_info));
4386         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4387         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4388                 block_group = btrfs_lookup_block_group(fs_info,
4389                                                        ffe_ctl->search_start);
4390                 /*
4391                  * we don't want to use the block group if it doesn't match our
4392                  * allocation bits, or if its not cached.
4393                  *
4394                  * However if we are re-searching with an ideal block group
4395                  * picked out then we don't care that the block group is cached.
4396                  */
4397                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4398                     block_group->cached != BTRFS_CACHE_NO) {
4399                         down_read(&space_info->groups_sem);
4400                         if (list_empty(&block_group->list) ||
4401                             block_group->ro) {
4402                                 /*
4403                                  * someone is removing this block group,
4404                                  * we can't jump into the have_block_group
4405                                  * target because our list pointers are not
4406                                  * valid
4407                                  */
4408                                 btrfs_put_block_group(block_group);
4409                                 up_read(&space_info->groups_sem);
4410                         } else {
4411                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4412                                                         block_group->flags);
4413                                 btrfs_lock_block_group(block_group,
4414                                                        ffe_ctl->delalloc);
4415                                 ffe_ctl->hinted = true;
4416                                 goto have_block_group;
4417                         }
4418                 } else if (block_group) {
4419                         btrfs_put_block_group(block_group);
4420                 }
4421         }
4422 search:
4423         trace_find_free_extent_search_loop(root, ffe_ctl);
4424         ffe_ctl->have_caching_bg = false;
4425         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4426             ffe_ctl->index == 0)
4427                 full_search = true;
4428         down_read(&space_info->groups_sem);
4429         list_for_each_entry(block_group,
4430                             &space_info->block_groups[ffe_ctl->index], list) {
4431                 struct btrfs_block_group *bg_ret;
4432
4433                 ffe_ctl->hinted = false;
4434                 /* If the block group is read-only, we can skip it entirely. */
4435                 if (unlikely(block_group->ro)) {
4436                         if (ffe_ctl->for_treelog)
4437                                 btrfs_clear_treelog_bg(block_group);
4438                         if (ffe_ctl->for_data_reloc)
4439                                 btrfs_clear_data_reloc_bg(block_group);
4440                         continue;
4441                 }
4442
4443                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4444                 ffe_ctl->search_start = block_group->start;
4445
4446                 /*
4447                  * this can happen if we end up cycling through all the
4448                  * raid types, but we want to make sure we only allocate
4449                  * for the proper type.
4450                  */
4451                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4452                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4453                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4454                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4455                                 BTRFS_BLOCK_GROUP_RAID10;
4456
4457                         /*
4458                          * if they asked for extra copies and this block group
4459                          * doesn't provide them, bail.  This does allow us to
4460                          * fill raid0 from raid1.
4461                          */
4462                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4463                                 goto loop;
4464
4465                         /*
4466                          * This block group has different flags than we want.
4467                          * It's possible that we have MIXED_GROUP flag but no
4468                          * block group is mixed.  Just skip such block group.
4469                          */
4470                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4471                         continue;
4472                 }
4473
4474 have_block_group:
4475                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4476                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4477                 if (unlikely(!ffe_ctl->cached)) {
4478                         ffe_ctl->have_caching_bg = true;
4479                         ret = btrfs_cache_block_group(block_group, false);
4480
4481                         /*
4482                          * If we get ENOMEM here or something else we want to
4483                          * try other block groups, because it may not be fatal.
4484                          * However if we can't find anything else we need to
4485                          * save our return here so that we return the actual
4486                          * error that caused problems, not ENOSPC.
4487                          */
4488                         if (ret < 0) {
4489                                 if (!cache_block_group_error)
4490                                         cache_block_group_error = ret;
4491                                 ret = 0;
4492                                 goto loop;
4493                         }
4494                         ret = 0;
4495                 }
4496
4497                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) {
4498                         if (!cache_block_group_error)
4499                                 cache_block_group_error = -EIO;
4500                         goto loop;
4501                 }
4502
4503                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4504                         goto loop;
4505
4506                 bg_ret = NULL;
4507                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4508                 if (ret > 0)
4509                         goto loop;
4510
4511                 if (bg_ret && bg_ret != block_group) {
4512                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4513                         block_group = bg_ret;
4514                 }
4515
4516                 /* Checks */
4517                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4518                                                  fs_info->stripesize);
4519
4520                 /* move on to the next group */
4521                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4522                     block_group->start + block_group->length) {
4523                         btrfs_add_free_space_unused(block_group,
4524                                             ffe_ctl->found_offset,
4525                                             ffe_ctl->num_bytes);
4526                         goto loop;
4527                 }
4528
4529                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4530                         btrfs_add_free_space_unused(block_group,
4531                                         ffe_ctl->found_offset,
4532                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4533
4534                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4535                                                ffe_ctl->num_bytes,
4536                                                ffe_ctl->delalloc,
4537                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4538                 if (ret == -EAGAIN) {
4539                         btrfs_add_free_space_unused(block_group,
4540                                         ffe_ctl->found_offset,
4541                                         ffe_ctl->num_bytes);
4542                         goto loop;
4543                 }
4544                 btrfs_inc_block_group_reservations(block_group);
4545
4546                 /* we are all good, lets return */
4547                 ins->objectid = ffe_ctl->search_start;
4548                 ins->offset = ffe_ctl->num_bytes;
4549
4550                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4551                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4552                 break;
4553 loop:
4554                 if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
4555                     !ffe_ctl->retry_uncached) {
4556                         ffe_ctl->retry_uncached = true;
4557                         btrfs_wait_block_group_cache_progress(block_group,
4558                                                 ffe_ctl->num_bytes +
4559                                                 ffe_ctl->empty_cluster +
4560                                                 ffe_ctl->empty_size);
4561                         goto have_block_group;
4562                 }
4563                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4564                 cond_resched();
4565         }
4566         up_read(&space_info->groups_sem);
4567
4568         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4569         if (ret > 0)
4570                 goto search;
4571
4572         if (ret == -ENOSPC && !cache_block_group_error) {
4573                 /*
4574                  * Use ffe_ctl->total_free_space as fallback if we can't find
4575                  * any contiguous hole.
4576                  */
4577                 if (!ffe_ctl->max_extent_size)
4578                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4579                 spin_lock(&space_info->lock);
4580                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4581                 spin_unlock(&space_info->lock);
4582                 ins->offset = ffe_ctl->max_extent_size;
4583         } else if (ret == -ENOSPC) {
4584                 ret = cache_block_group_error;
4585         }
4586         return ret;
4587 }
4588
4589 /*
4590  * Entry point to the extent allocator. Tries to find a hole that is at least
4591  * as big as @num_bytes.
4592  *
4593  * @root           -    The root that will contain this extent
4594  *
4595  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4596  *                      is used for accounting purposes. This value differs
4597  *                      from @num_bytes only in the case of compressed extents.
4598  *
4599  * @num_bytes      -    Number of bytes to allocate on-disk.
4600  *
4601  * @min_alloc_size -    Indicates the minimum amount of space that the
4602  *                      allocator should try to satisfy. In some cases
4603  *                      @num_bytes may be larger than what is required and if
4604  *                      the filesystem is fragmented then allocation fails.
4605  *                      However, the presence of @min_alloc_size gives a
4606  *                      chance to try and satisfy the smaller allocation.
4607  *
4608  * @empty_size     -    A hint that you plan on doing more COW. This is the
4609  *                      size in bytes the allocator should try to find free
4610  *                      next to the block it returns.  This is just a hint and
4611  *                      may be ignored by the allocator.
4612  *
4613  * @hint_byte      -    Hint to the allocator to start searching above the byte
4614  *                      address passed. It might be ignored.
4615  *
4616  * @ins            -    This key is modified to record the found hole. It will
4617  *                      have the following values:
4618  *                      ins->objectid == start position
4619  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4620  *                      ins->offset == the size of the hole.
4621  *
4622  * @is_data        -    Boolean flag indicating whether an extent is
4623  *                      allocated for data (true) or metadata (false)
4624  *
4625  * @delalloc       -    Boolean flag indicating whether this allocation is for
4626  *                      delalloc or not. If 'true' data_rwsem of block groups
4627  *                      is going to be acquired.
4628  *
4629  *
4630  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4631  * case -ENOSPC is returned then @ins->offset will contain the size of the
4632  * largest available hole the allocator managed to find.
4633  */
4634 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4635                          u64 num_bytes, u64 min_alloc_size,
4636                          u64 empty_size, u64 hint_byte,
4637                          struct btrfs_key *ins, int is_data, int delalloc)
4638 {
4639         struct btrfs_fs_info *fs_info = root->fs_info;
4640         struct find_free_extent_ctl ffe_ctl = {};
4641         bool final_tried = num_bytes == min_alloc_size;
4642         u64 flags;
4643         int ret;
4644         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4645         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4646
4647         flags = get_alloc_profile_by_root(root, is_data);
4648 again:
4649         WARN_ON(num_bytes < fs_info->sectorsize);
4650
4651         ffe_ctl.ram_bytes = ram_bytes;
4652         ffe_ctl.num_bytes = num_bytes;
4653         ffe_ctl.min_alloc_size = min_alloc_size;
4654         ffe_ctl.empty_size = empty_size;
4655         ffe_ctl.flags = flags;
4656         ffe_ctl.delalloc = delalloc;
4657         ffe_ctl.hint_byte = hint_byte;
4658         ffe_ctl.for_treelog = for_treelog;
4659         ffe_ctl.for_data_reloc = for_data_reloc;
4660
4661         ret = find_free_extent(root, ins, &ffe_ctl);
4662         if (!ret && !is_data) {
4663                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4664         } else if (ret == -ENOSPC) {
4665                 if (!final_tried && ins->offset) {
4666                         num_bytes = min(num_bytes >> 1, ins->offset);
4667                         num_bytes = round_down(num_bytes,
4668                                                fs_info->sectorsize);
4669                         num_bytes = max(num_bytes, min_alloc_size);
4670                         ram_bytes = num_bytes;
4671                         if (num_bytes == min_alloc_size)
4672                                 final_tried = true;
4673                         goto again;
4674                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4675                         struct btrfs_space_info *sinfo;
4676
4677                         sinfo = btrfs_find_space_info(fs_info, flags);
4678                         btrfs_err(fs_info,
4679         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4680                                   flags, num_bytes, for_treelog, for_data_reloc);
4681                         if (sinfo)
4682                                 btrfs_dump_space_info(fs_info, sinfo,
4683                                                       num_bytes, 1);
4684                 }
4685         }
4686
4687         return ret;
4688 }
4689
4690 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4691                                u64 start, u64 len, int delalloc)
4692 {
4693         struct btrfs_block_group *cache;
4694
4695         cache = btrfs_lookup_block_group(fs_info, start);
4696         if (!cache) {
4697                 btrfs_err(fs_info, "Unable to find block group for %llu",
4698                           start);
4699                 return -ENOSPC;
4700         }
4701
4702         btrfs_add_free_space(cache, start, len);
4703         btrfs_free_reserved_bytes(cache, len, delalloc);
4704         trace_btrfs_reserved_extent_free(fs_info, start, len);
4705
4706         btrfs_put_block_group(cache);
4707         return 0;
4708 }
4709
4710 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans,
4711                               const struct extent_buffer *eb)
4712 {
4713         struct btrfs_block_group *cache;
4714         int ret = 0;
4715
4716         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
4717         if (!cache) {
4718                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4719                           eb->start);
4720                 return -ENOSPC;
4721         }
4722
4723         ret = pin_down_extent(trans, cache, eb->start, eb->len, 1);
4724         btrfs_put_block_group(cache);
4725         return ret;
4726 }
4727
4728 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4729                                  u64 num_bytes)
4730 {
4731         struct btrfs_fs_info *fs_info = trans->fs_info;
4732         int ret;
4733
4734         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4735         if (ret)
4736                 return ret;
4737
4738         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4739         if (ret) {
4740                 ASSERT(!ret);
4741                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4742                           bytenr, num_bytes);
4743                 return ret;
4744         }
4745
4746         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4747         return 0;
4748 }
4749
4750 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4751                                       u64 parent, u64 root_objectid,
4752                                       u64 flags, u64 owner, u64 offset,
4753                                       struct btrfs_key *ins, int ref_mod, u64 oref_root)
4754 {
4755         struct btrfs_fs_info *fs_info = trans->fs_info;
4756         struct btrfs_root *extent_root;
4757         int ret;
4758         struct btrfs_extent_item *extent_item;
4759         struct btrfs_extent_owner_ref *oref;
4760         struct btrfs_extent_inline_ref *iref;
4761         struct btrfs_path *path;
4762         struct extent_buffer *leaf;
4763         int type;
4764         u32 size;
4765         const bool simple_quota = (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE);
4766
4767         if (parent > 0)
4768                 type = BTRFS_SHARED_DATA_REF_KEY;
4769         else
4770                 type = BTRFS_EXTENT_DATA_REF_KEY;
4771
4772         size = sizeof(*extent_item);
4773         if (simple_quota)
4774                 size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
4775         size += btrfs_extent_inline_ref_size(type);
4776
4777         path = btrfs_alloc_path();
4778         if (!path)
4779                 return -ENOMEM;
4780
4781         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4782         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4783         if (ret) {
4784                 btrfs_free_path(path);
4785                 return ret;
4786         }
4787
4788         leaf = path->nodes[0];
4789         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4790                                      struct btrfs_extent_item);
4791         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4792         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4793         btrfs_set_extent_flags(leaf, extent_item,
4794                                flags | BTRFS_EXTENT_FLAG_DATA);
4795
4796         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4797         if (simple_quota) {
4798                 btrfs_set_extent_inline_ref_type(leaf, iref, BTRFS_EXTENT_OWNER_REF_KEY);
4799                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
4800                 btrfs_set_extent_owner_ref_root_id(leaf, oref, oref_root);
4801                 iref = (struct btrfs_extent_inline_ref *)(oref + 1);
4802         }
4803         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4804
4805         if (parent > 0) {
4806                 struct btrfs_shared_data_ref *ref;
4807                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4808                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4809                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4810         } else {
4811                 struct btrfs_extent_data_ref *ref;
4812                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4813                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4814                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4815                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4816                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4817         }
4818
4819         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
4820         btrfs_free_path(path);
4821
4822         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4823 }
4824
4825 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4826                                      struct btrfs_delayed_ref_node *node,
4827                                      struct btrfs_delayed_extent_op *extent_op)
4828 {
4829         struct btrfs_fs_info *fs_info = trans->fs_info;
4830         struct btrfs_root *extent_root;
4831         int ret;
4832         struct btrfs_extent_item *extent_item;
4833         struct btrfs_key extent_key;
4834         struct btrfs_tree_block_info *block_info;
4835         struct btrfs_extent_inline_ref *iref;
4836         struct btrfs_path *path;
4837         struct extent_buffer *leaf;
4838         struct btrfs_delayed_tree_ref *ref;
4839         u32 size = sizeof(*extent_item) + sizeof(*iref);
4840         u64 flags = extent_op->flags_to_set;
4841         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4842
4843         ref = btrfs_delayed_node_to_tree_ref(node);
4844
4845         extent_key.objectid = node->bytenr;
4846         if (skinny_metadata) {
4847                 extent_key.offset = ref->level;
4848                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4849         } else {
4850                 extent_key.offset = node->num_bytes;
4851                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4852                 size += sizeof(*block_info);
4853         }
4854
4855         path = btrfs_alloc_path();
4856         if (!path)
4857                 return -ENOMEM;
4858
4859         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4860         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4861                                       size);
4862         if (ret) {
4863                 btrfs_free_path(path);
4864                 return ret;
4865         }
4866
4867         leaf = path->nodes[0];
4868         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4869                                      struct btrfs_extent_item);
4870         btrfs_set_extent_refs(leaf, extent_item, 1);
4871         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4872         btrfs_set_extent_flags(leaf, extent_item,
4873                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4874
4875         if (skinny_metadata) {
4876                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4877         } else {
4878                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4879                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4880                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4881                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4882         }
4883
4884         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4885                 btrfs_set_extent_inline_ref_type(leaf, iref,
4886                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4887                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4888         } else {
4889                 btrfs_set_extent_inline_ref_type(leaf, iref,
4890                                                  BTRFS_TREE_BLOCK_REF_KEY);
4891                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4892         }
4893
4894         btrfs_mark_buffer_dirty(trans, leaf);
4895         btrfs_free_path(path);
4896
4897         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4898 }
4899
4900 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4901                                      struct btrfs_root *root, u64 owner,
4902                                      u64 offset, u64 ram_bytes,
4903                                      struct btrfs_key *ins)
4904 {
4905         struct btrfs_ref generic_ref = { 0 };
4906         u64 root_objectid = root->root_key.objectid;
4907         u64 owning_root = root_objectid;
4908
4909         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
4910
4911         if (btrfs_is_data_reloc_root(root) && is_fstree(root->relocation_src_root))
4912                 owning_root = root->relocation_src_root;
4913
4914         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4915                                ins->objectid, ins->offset, 0, owning_root);
4916         btrfs_init_data_ref(&generic_ref, root_objectid, owner,
4917                             offset, 0, false);
4918         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4919
4920         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4921 }
4922
4923 /*
4924  * this is used by the tree logging recovery code.  It records that
4925  * an extent has been allocated and makes sure to clear the free
4926  * space cache bits as well
4927  */
4928 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4929                                    u64 root_objectid, u64 owner, u64 offset,
4930                                    struct btrfs_key *ins)
4931 {
4932         struct btrfs_fs_info *fs_info = trans->fs_info;
4933         int ret;
4934         struct btrfs_block_group *block_group;
4935         struct btrfs_space_info *space_info;
4936         struct btrfs_squota_delta delta = {
4937                 .root = root_objectid,
4938                 .num_bytes = ins->offset,
4939                 .generation = trans->transid,
4940                 .rsv_bytes = 0,
4941                 .is_data = true,
4942                 .is_inc = true,
4943         };
4944
4945         /*
4946          * Mixed block groups will exclude before processing the log so we only
4947          * need to do the exclude dance if this fs isn't mixed.
4948          */
4949         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4950                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4951                                               ins->offset);
4952                 if (ret)
4953                         return ret;
4954         }
4955
4956         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4957         if (!block_group)
4958                 return -EINVAL;
4959
4960         space_info = block_group->space_info;
4961         spin_lock(&space_info->lock);
4962         spin_lock(&block_group->lock);
4963         space_info->bytes_reserved += ins->offset;
4964         block_group->reserved += ins->offset;
4965         spin_unlock(&block_group->lock);
4966         spin_unlock(&space_info->lock);
4967
4968         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4969                                          offset, ins, 1, root_objectid);
4970         if (ret)
4971                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4972         ret = btrfs_record_squota_delta(fs_info, &delta);
4973         btrfs_put_block_group(block_group);
4974         return ret;
4975 }
4976
4977 #ifdef CONFIG_BTRFS_DEBUG
4978 /*
4979  * Extra safety check in case the extent tree is corrupted and extent allocator
4980  * chooses to use a tree block which is already used and locked.
4981  */
4982 static bool check_eb_lock_owner(const struct extent_buffer *eb)
4983 {
4984         if (eb->lock_owner == current->pid) {
4985                 btrfs_err_rl(eb->fs_info,
4986 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4987                              eb->start, btrfs_header_owner(eb), current->pid);
4988                 return true;
4989         }
4990         return false;
4991 }
4992 #else
4993 static bool check_eb_lock_owner(struct extent_buffer *eb)
4994 {
4995         return false;
4996 }
4997 #endif
4998
4999 static struct extent_buffer *
5000 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5001                       u64 bytenr, int level, u64 owner,
5002                       enum btrfs_lock_nesting nest)
5003 {
5004         struct btrfs_fs_info *fs_info = root->fs_info;
5005         struct extent_buffer *buf;
5006         u64 lockdep_owner = owner;
5007
5008         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
5009         if (IS_ERR(buf))
5010                 return buf;
5011
5012         if (check_eb_lock_owner(buf)) {
5013                 free_extent_buffer(buf);
5014                 return ERR_PTR(-EUCLEAN);
5015         }
5016
5017         /*
5018          * The reloc trees are just snapshots, so we need them to appear to be
5019          * just like any other fs tree WRT lockdep.
5020          *
5021          * The exception however is in replace_path() in relocation, where we
5022          * hold the lock on the original fs root and then search for the reloc
5023          * root.  At that point we need to make sure any reloc root buffers are
5024          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
5025          * lockdep happy.
5026          */
5027         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
5028             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
5029                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
5030
5031         /* btrfs_clear_buffer_dirty() accesses generation field. */
5032         btrfs_set_header_generation(buf, trans->transid);
5033
5034         /*
5035          * This needs to stay, because we could allocate a freed block from an
5036          * old tree into a new tree, so we need to make sure this new block is
5037          * set to the appropriate level and owner.
5038          */
5039         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
5040
5041         __btrfs_tree_lock(buf, nest);
5042         btrfs_clear_buffer_dirty(trans, buf);
5043         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
5044         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
5045
5046         set_extent_buffer_uptodate(buf);
5047
5048         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
5049         btrfs_set_header_level(buf, level);
5050         btrfs_set_header_bytenr(buf, buf->start);
5051         btrfs_set_header_generation(buf, trans->transid);
5052         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
5053         btrfs_set_header_owner(buf, owner);
5054         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
5055         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
5056         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5057                 buf->log_index = root->log_transid % 2;
5058                 /*
5059                  * we allow two log transactions at a time, use different
5060                  * EXTENT bit to differentiate dirty pages.
5061                  */
5062                 if (buf->log_index == 0)
5063                         set_extent_bit(&root->dirty_log_pages, buf->start,
5064                                        buf->start + buf->len - 1,
5065                                        EXTENT_DIRTY, NULL);
5066                 else
5067                         set_extent_bit(&root->dirty_log_pages, buf->start,
5068                                        buf->start + buf->len - 1,
5069                                        EXTENT_NEW, NULL);
5070         } else {
5071                 buf->log_index = -1;
5072                 set_extent_bit(&trans->transaction->dirty_pages, buf->start,
5073                                buf->start + buf->len - 1, EXTENT_DIRTY, NULL);
5074         }
5075         /* this returns a buffer locked for blocking */
5076         return buf;
5077 }
5078
5079 /*
5080  * finds a free extent and does all the dirty work required for allocation
5081  * returns the tree buffer or an ERR_PTR on error.
5082  */
5083 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
5084                                              struct btrfs_root *root,
5085                                              u64 parent, u64 root_objectid,
5086                                              const struct btrfs_disk_key *key,
5087                                              int level, u64 hint,
5088                                              u64 empty_size,
5089                                              u64 reloc_src_root,
5090                                              enum btrfs_lock_nesting nest)
5091 {
5092         struct btrfs_fs_info *fs_info = root->fs_info;
5093         struct btrfs_key ins;
5094         struct btrfs_block_rsv *block_rsv;
5095         struct extent_buffer *buf;
5096         struct btrfs_delayed_extent_op *extent_op;
5097         struct btrfs_ref generic_ref = { 0 };
5098         u64 flags = 0;
5099         int ret;
5100         u32 blocksize = fs_info->nodesize;
5101         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
5102         u64 owning_root;
5103
5104 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5105         if (btrfs_is_testing(fs_info)) {
5106                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
5107                                             level, root_objectid, nest);
5108                 if (!IS_ERR(buf))
5109                         root->alloc_bytenr += blocksize;
5110                 return buf;
5111         }
5112 #endif
5113
5114         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
5115         if (IS_ERR(block_rsv))
5116                 return ERR_CAST(block_rsv);
5117
5118         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
5119                                    empty_size, hint, &ins, 0, 0);
5120         if (ret)
5121                 goto out_unuse;
5122
5123         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
5124                                     root_objectid, nest);
5125         if (IS_ERR(buf)) {
5126                 ret = PTR_ERR(buf);
5127                 goto out_free_reserved;
5128         }
5129         owning_root = btrfs_header_owner(buf);
5130
5131         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5132                 if (parent == 0)
5133                         parent = ins.objectid;
5134                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5135                 owning_root = reloc_src_root;
5136         } else
5137                 BUG_ON(parent > 0);
5138
5139         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5140                 extent_op = btrfs_alloc_delayed_extent_op();
5141                 if (!extent_op) {
5142                         ret = -ENOMEM;
5143                         goto out_free_buf;
5144                 }
5145                 if (key)
5146                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5147                 else
5148                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5149                 extent_op->flags_to_set = flags;
5150                 extent_op->update_key = skinny_metadata ? false : true;
5151                 extent_op->update_flags = true;
5152                 extent_op->level = level;
5153
5154                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
5155                                        ins.objectid, ins.offset, parent, owning_root);
5156                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
5157                                     root->root_key.objectid, false);
5158                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5159                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5160                 if (ret)
5161                         goto out_free_delayed;
5162         }
5163         return buf;
5164
5165 out_free_delayed:
5166         btrfs_free_delayed_extent_op(extent_op);
5167 out_free_buf:
5168         btrfs_tree_unlock(buf);
5169         free_extent_buffer(buf);
5170 out_free_reserved:
5171         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5172 out_unuse:
5173         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5174         return ERR_PTR(ret);
5175 }
5176
5177 struct walk_control {
5178         u64 refs[BTRFS_MAX_LEVEL];
5179         u64 flags[BTRFS_MAX_LEVEL];
5180         struct btrfs_key update_progress;
5181         struct btrfs_key drop_progress;
5182         int drop_level;
5183         int stage;
5184         int level;
5185         int shared_level;
5186         int update_ref;
5187         int keep_locks;
5188         int reada_slot;
5189         int reada_count;
5190         int restarted;
5191 };
5192
5193 #define DROP_REFERENCE  1
5194 #define UPDATE_BACKREF  2
5195
5196 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5197                                      struct btrfs_root *root,
5198                                      struct walk_control *wc,
5199                                      struct btrfs_path *path)
5200 {
5201         struct btrfs_fs_info *fs_info = root->fs_info;
5202         u64 bytenr;
5203         u64 generation;
5204         u64 refs;
5205         u64 flags;
5206         u32 nritems;
5207         struct btrfs_key key;
5208         struct extent_buffer *eb;
5209         int ret;
5210         int slot;
5211         int nread = 0;
5212
5213         if (path->slots[wc->level] < wc->reada_slot) {
5214                 wc->reada_count = wc->reada_count * 2 / 3;
5215                 wc->reada_count = max(wc->reada_count, 2);
5216         } else {
5217                 wc->reada_count = wc->reada_count * 3 / 2;
5218                 wc->reada_count = min_t(int, wc->reada_count,
5219                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5220         }
5221
5222         eb = path->nodes[wc->level];
5223         nritems = btrfs_header_nritems(eb);
5224
5225         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5226                 if (nread >= wc->reada_count)
5227                         break;
5228
5229                 cond_resched();
5230                 bytenr = btrfs_node_blockptr(eb, slot);
5231                 generation = btrfs_node_ptr_generation(eb, slot);
5232
5233                 if (slot == path->slots[wc->level])
5234                         goto reada;
5235
5236                 if (wc->stage == UPDATE_BACKREF &&
5237                     generation <= root->root_key.offset)
5238                         continue;
5239
5240                 /* We don't lock the tree block, it's OK to be racy here */
5241                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5242                                                wc->level - 1, 1, &refs,
5243                                                &flags, NULL);
5244                 /* We don't care about errors in readahead. */
5245                 if (ret < 0)
5246                         continue;
5247                 BUG_ON(refs == 0);
5248
5249                 if (wc->stage == DROP_REFERENCE) {
5250                         if (refs == 1)
5251                                 goto reada;
5252
5253                         if (wc->level == 1 &&
5254                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5255                                 continue;
5256                         if (!wc->update_ref ||
5257                             generation <= root->root_key.offset)
5258                                 continue;
5259                         btrfs_node_key_to_cpu(eb, &key, slot);
5260                         ret = btrfs_comp_cpu_keys(&key,
5261                                                   &wc->update_progress);
5262                         if (ret < 0)
5263                                 continue;
5264                 } else {
5265                         if (wc->level == 1 &&
5266                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5267                                 continue;
5268                 }
5269 reada:
5270                 btrfs_readahead_node_child(eb, slot);
5271                 nread++;
5272         }
5273         wc->reada_slot = slot;
5274 }
5275
5276 /*
5277  * helper to process tree block while walking down the tree.
5278  *
5279  * when wc->stage == UPDATE_BACKREF, this function updates
5280  * back refs for pointers in the block.
5281  *
5282  * NOTE: return value 1 means we should stop walking down.
5283  */
5284 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5285                                    struct btrfs_root *root,
5286                                    struct btrfs_path *path,
5287                                    struct walk_control *wc, int lookup_info)
5288 {
5289         struct btrfs_fs_info *fs_info = root->fs_info;
5290         int level = wc->level;
5291         struct extent_buffer *eb = path->nodes[level];
5292         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5293         int ret;
5294
5295         if (wc->stage == UPDATE_BACKREF &&
5296             btrfs_header_owner(eb) != root->root_key.objectid)
5297                 return 1;
5298
5299         /*
5300          * when reference count of tree block is 1, it won't increase
5301          * again. once full backref flag is set, we never clear it.
5302          */
5303         if (lookup_info &&
5304             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5305              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5306                 BUG_ON(!path->locks[level]);
5307                 ret = btrfs_lookup_extent_info(trans, fs_info,
5308                                                eb->start, level, 1,
5309                                                &wc->refs[level],
5310                                                &wc->flags[level],
5311                                                NULL);
5312                 BUG_ON(ret == -ENOMEM);
5313                 if (ret)
5314                         return ret;
5315                 BUG_ON(wc->refs[level] == 0);
5316         }
5317
5318         if (wc->stage == DROP_REFERENCE) {
5319                 if (wc->refs[level] > 1)
5320                         return 1;
5321
5322                 if (path->locks[level] && !wc->keep_locks) {
5323                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5324                         path->locks[level] = 0;
5325                 }
5326                 return 0;
5327         }
5328
5329         /* wc->stage == UPDATE_BACKREF */
5330         if (!(wc->flags[level] & flag)) {
5331                 BUG_ON(!path->locks[level]);
5332                 ret = btrfs_inc_ref(trans, root, eb, 1);
5333                 BUG_ON(ret); /* -ENOMEM */
5334                 ret = btrfs_dec_ref(trans, root, eb, 0);
5335                 BUG_ON(ret); /* -ENOMEM */
5336                 ret = btrfs_set_disk_extent_flags(trans, eb, flag);
5337                 BUG_ON(ret); /* -ENOMEM */
5338                 wc->flags[level] |= flag;
5339         }
5340
5341         /*
5342          * the block is shared by multiple trees, so it's not good to
5343          * keep the tree lock
5344          */
5345         if (path->locks[level] && level > 0) {
5346                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5347                 path->locks[level] = 0;
5348         }
5349         return 0;
5350 }
5351
5352 /*
5353  * This is used to verify a ref exists for this root to deal with a bug where we
5354  * would have a drop_progress key that hadn't been updated properly.
5355  */
5356 static int check_ref_exists(struct btrfs_trans_handle *trans,
5357                             struct btrfs_root *root, u64 bytenr, u64 parent,
5358                             int level)
5359 {
5360         struct btrfs_path *path;
5361         struct btrfs_extent_inline_ref *iref;
5362         int ret;
5363
5364         path = btrfs_alloc_path();
5365         if (!path)
5366                 return -ENOMEM;
5367
5368         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5369                                     root->fs_info->nodesize, parent,
5370                                     root->root_key.objectid, level, 0);
5371         btrfs_free_path(path);
5372         if (ret == -ENOENT)
5373                 return 0;
5374         if (ret < 0)
5375                 return ret;
5376         return 1;
5377 }
5378
5379 /*
5380  * helper to process tree block pointer.
5381  *
5382  * when wc->stage == DROP_REFERENCE, this function checks
5383  * reference count of the block pointed to. if the block
5384  * is shared and we need update back refs for the subtree
5385  * rooted at the block, this function changes wc->stage to
5386  * UPDATE_BACKREF. if the block is shared and there is no
5387  * need to update back, this function drops the reference
5388  * to the block.
5389  *
5390  * NOTE: return value 1 means we should stop walking down.
5391  */
5392 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5393                                  struct btrfs_root *root,
5394                                  struct btrfs_path *path,
5395                                  struct walk_control *wc, int *lookup_info)
5396 {
5397         struct btrfs_fs_info *fs_info = root->fs_info;
5398         u64 bytenr;
5399         u64 generation;
5400         u64 parent;
5401         u64 owner_root = 0;
5402         struct btrfs_tree_parent_check check = { 0 };
5403         struct btrfs_key key;
5404         struct btrfs_ref ref = { 0 };
5405         struct extent_buffer *next;
5406         int level = wc->level;
5407         int reada = 0;
5408         int ret = 0;
5409         bool need_account = false;
5410
5411         generation = btrfs_node_ptr_generation(path->nodes[level],
5412                                                path->slots[level]);
5413         /*
5414          * if the lower level block was created before the snapshot
5415          * was created, we know there is no need to update back refs
5416          * for the subtree
5417          */
5418         if (wc->stage == UPDATE_BACKREF &&
5419             generation <= root->root_key.offset) {
5420                 *lookup_info = 1;
5421                 return 1;
5422         }
5423
5424         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5425
5426         check.level = level - 1;
5427         check.transid = generation;
5428         check.owner_root = root->root_key.objectid;
5429         check.has_first_key = true;
5430         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5431                               path->slots[level]);
5432
5433         next = find_extent_buffer(fs_info, bytenr);
5434         if (!next) {
5435                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5436                                 root->root_key.objectid, level - 1);
5437                 if (IS_ERR(next))
5438                         return PTR_ERR(next);
5439                 reada = 1;
5440         }
5441         btrfs_tree_lock(next);
5442
5443         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5444                                        &wc->refs[level - 1],
5445                                        &wc->flags[level - 1],
5446                                        &owner_root);
5447         if (ret < 0)
5448                 goto out_unlock;
5449
5450         if (unlikely(wc->refs[level - 1] == 0)) {
5451                 btrfs_err(fs_info, "Missing references.");
5452                 ret = -EIO;
5453                 goto out_unlock;
5454         }
5455         *lookup_info = 0;
5456
5457         if (wc->stage == DROP_REFERENCE) {
5458                 if (wc->refs[level - 1] > 1) {
5459                         need_account = true;
5460                         if (level == 1 &&
5461                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5462                                 goto skip;
5463
5464                         if (!wc->update_ref ||
5465                             generation <= root->root_key.offset)
5466                                 goto skip;
5467
5468                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5469                                               path->slots[level]);
5470                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5471                         if (ret < 0)
5472                                 goto skip;
5473
5474                         wc->stage = UPDATE_BACKREF;
5475                         wc->shared_level = level - 1;
5476                 }
5477         } else {
5478                 if (level == 1 &&
5479                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5480                         goto skip;
5481         }
5482
5483         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5484                 btrfs_tree_unlock(next);
5485                 free_extent_buffer(next);
5486                 next = NULL;
5487                 *lookup_info = 1;
5488         }
5489
5490         if (!next) {
5491                 if (reada && level == 1)
5492                         reada_walk_down(trans, root, wc, path);
5493                 next = read_tree_block(fs_info, bytenr, &check);
5494                 if (IS_ERR(next)) {
5495                         return PTR_ERR(next);
5496                 } else if (!extent_buffer_uptodate(next)) {
5497                         free_extent_buffer(next);
5498                         return -EIO;
5499                 }
5500                 btrfs_tree_lock(next);
5501         }
5502
5503         level--;
5504         ASSERT(level == btrfs_header_level(next));
5505         if (level != btrfs_header_level(next)) {
5506                 btrfs_err(root->fs_info, "mismatched level");
5507                 ret = -EIO;
5508                 goto out_unlock;
5509         }
5510         path->nodes[level] = next;
5511         path->slots[level] = 0;
5512         path->locks[level] = BTRFS_WRITE_LOCK;
5513         wc->level = level;
5514         if (wc->level == 1)
5515                 wc->reada_slot = 0;
5516         return 0;
5517 skip:
5518         wc->refs[level - 1] = 0;
5519         wc->flags[level - 1] = 0;
5520         if (wc->stage == DROP_REFERENCE) {
5521                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5522                         parent = path->nodes[level]->start;
5523                 } else {
5524                         ASSERT(root->root_key.objectid ==
5525                                btrfs_header_owner(path->nodes[level]));
5526                         if (root->root_key.objectid !=
5527                             btrfs_header_owner(path->nodes[level])) {
5528                                 btrfs_err(root->fs_info,
5529                                                 "mismatched block owner");
5530                                 ret = -EIO;
5531                                 goto out_unlock;
5532                         }
5533                         parent = 0;
5534                 }
5535
5536                 /*
5537                  * If we had a drop_progress we need to verify the refs are set
5538                  * as expected.  If we find our ref then we know that from here
5539                  * on out everything should be correct, and we can clear the
5540                  * ->restarted flag.
5541                  */
5542                 if (wc->restarted) {
5543                         ret = check_ref_exists(trans, root, bytenr, parent,
5544                                                level - 1);
5545                         if (ret < 0)
5546                                 goto out_unlock;
5547                         if (ret == 0)
5548                                 goto no_delete;
5549                         ret = 0;
5550                         wc->restarted = 0;
5551                 }
5552
5553                 /*
5554                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5555                  * already accounted them at merge time (replace_path),
5556                  * thus we could skip expensive subtree trace here.
5557                  */
5558                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5559                     need_account) {
5560                         ret = btrfs_qgroup_trace_subtree(trans, next,
5561                                                          generation, level - 1);
5562                         if (ret) {
5563                                 btrfs_err_rl(fs_info,
5564                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5565                                              ret);
5566                         }
5567                 }
5568
5569                 /*
5570                  * We need to update the next key in our walk control so we can
5571                  * update the drop_progress key accordingly.  We don't care if
5572                  * find_next_key doesn't find a key because that means we're at
5573                  * the end and are going to clean up now.
5574                  */
5575                 wc->drop_level = level;
5576                 find_next_key(path, level, &wc->drop_progress);
5577
5578                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5579                                        fs_info->nodesize, parent, owner_root);
5580                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5581                                     0, false);
5582                 ret = btrfs_free_extent(trans, &ref);
5583                 if (ret)
5584                         goto out_unlock;
5585         }
5586 no_delete:
5587         *lookup_info = 1;
5588         ret = 1;
5589
5590 out_unlock:
5591         btrfs_tree_unlock(next);
5592         free_extent_buffer(next);
5593
5594         return ret;
5595 }
5596
5597 /*
5598  * helper to process tree block while walking up the tree.
5599  *
5600  * when wc->stage == DROP_REFERENCE, this function drops
5601  * reference count on the block.
5602  *
5603  * when wc->stage == UPDATE_BACKREF, this function changes
5604  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5605  * to UPDATE_BACKREF previously while processing the block.
5606  *
5607  * NOTE: return value 1 means we should stop walking up.
5608  */
5609 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5610                                  struct btrfs_root *root,
5611                                  struct btrfs_path *path,
5612                                  struct walk_control *wc)
5613 {
5614         struct btrfs_fs_info *fs_info = root->fs_info;
5615         int ret;
5616         int level = wc->level;
5617         struct extent_buffer *eb = path->nodes[level];
5618         u64 parent = 0;
5619
5620         if (wc->stage == UPDATE_BACKREF) {
5621                 BUG_ON(wc->shared_level < level);
5622                 if (level < wc->shared_level)
5623                         goto out;
5624
5625                 ret = find_next_key(path, level + 1, &wc->update_progress);
5626                 if (ret > 0)
5627                         wc->update_ref = 0;
5628
5629                 wc->stage = DROP_REFERENCE;
5630                 wc->shared_level = -1;
5631                 path->slots[level] = 0;
5632
5633                 /*
5634                  * check reference count again if the block isn't locked.
5635                  * we should start walking down the tree again if reference
5636                  * count is one.
5637                  */
5638                 if (!path->locks[level]) {
5639                         BUG_ON(level == 0);
5640                         btrfs_tree_lock(eb);
5641                         path->locks[level] = BTRFS_WRITE_LOCK;
5642
5643                         ret = btrfs_lookup_extent_info(trans, fs_info,
5644                                                        eb->start, level, 1,
5645                                                        &wc->refs[level],
5646                                                        &wc->flags[level],
5647                                                        NULL);
5648                         if (ret < 0) {
5649                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5650                                 path->locks[level] = 0;
5651                                 return ret;
5652                         }
5653                         BUG_ON(wc->refs[level] == 0);
5654                         if (wc->refs[level] == 1) {
5655                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5656                                 path->locks[level] = 0;
5657                                 return 1;
5658                         }
5659                 }
5660         }
5661
5662         /* wc->stage == DROP_REFERENCE */
5663         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5664
5665         if (wc->refs[level] == 1) {
5666                 if (level == 0) {
5667                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5668                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5669                         else
5670                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5671                         BUG_ON(ret); /* -ENOMEM */
5672                         if (is_fstree(root->root_key.objectid)) {
5673                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5674                                 if (ret) {
5675                                         btrfs_err_rl(fs_info,
5676         "error %d accounting leaf items, quota is out of sync, rescan required",
5677                                              ret);
5678                                 }
5679                         }
5680                 }
5681                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5682                 if (!path->locks[level]) {
5683                         btrfs_tree_lock(eb);
5684                         path->locks[level] = BTRFS_WRITE_LOCK;
5685                 }
5686                 btrfs_clear_buffer_dirty(trans, eb);
5687         }
5688
5689         if (eb == root->node) {
5690                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5691                         parent = eb->start;
5692                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5693                         goto owner_mismatch;
5694         } else {
5695                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5696                         parent = path->nodes[level + 1]->start;
5697                 else if (root->root_key.objectid !=
5698                          btrfs_header_owner(path->nodes[level + 1]))
5699                         goto owner_mismatch;
5700         }
5701
5702         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5703                               wc->refs[level] == 1);
5704 out:
5705         wc->refs[level] = 0;
5706         wc->flags[level] = 0;
5707         return 0;
5708
5709 owner_mismatch:
5710         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5711                      btrfs_header_owner(eb), root->root_key.objectid);
5712         return -EUCLEAN;
5713 }
5714
5715 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5716                                    struct btrfs_root *root,
5717                                    struct btrfs_path *path,
5718                                    struct walk_control *wc)
5719 {
5720         int level = wc->level;
5721         int lookup_info = 1;
5722         int ret = 0;
5723
5724         while (level >= 0) {
5725                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5726                 if (ret)
5727                         break;
5728
5729                 if (level == 0)
5730                         break;
5731
5732                 if (path->slots[level] >=
5733                     btrfs_header_nritems(path->nodes[level]))
5734                         break;
5735
5736                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5737                 if (ret > 0) {
5738                         path->slots[level]++;
5739                         continue;
5740                 } else if (ret < 0)
5741                         break;
5742                 level = wc->level;
5743         }
5744         return (ret == 1) ? 0 : ret;
5745 }
5746
5747 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5748                                  struct btrfs_root *root,
5749                                  struct btrfs_path *path,
5750                                  struct walk_control *wc, int max_level)
5751 {
5752         int level = wc->level;
5753         int ret;
5754
5755         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5756         while (level < max_level && path->nodes[level]) {
5757                 wc->level = level;
5758                 if (path->slots[level] + 1 <
5759                     btrfs_header_nritems(path->nodes[level])) {
5760                         path->slots[level]++;
5761                         return 0;
5762                 } else {
5763                         ret = walk_up_proc(trans, root, path, wc);
5764                         if (ret > 0)
5765                                 return 0;
5766                         if (ret < 0)
5767                                 return ret;
5768
5769                         if (path->locks[level]) {
5770                                 btrfs_tree_unlock_rw(path->nodes[level],
5771                                                      path->locks[level]);
5772                                 path->locks[level] = 0;
5773                         }
5774                         free_extent_buffer(path->nodes[level]);
5775                         path->nodes[level] = NULL;
5776                         level++;
5777                 }
5778         }
5779         return 1;
5780 }
5781
5782 /*
5783  * drop a subvolume tree.
5784  *
5785  * this function traverses the tree freeing any blocks that only
5786  * referenced by the tree.
5787  *
5788  * when a shared tree block is found. this function decreases its
5789  * reference count by one. if update_ref is true, this function
5790  * also make sure backrefs for the shared block and all lower level
5791  * blocks are properly updated.
5792  *
5793  * If called with for_reloc == 0, may exit early with -EAGAIN
5794  */
5795 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5796 {
5797         const bool is_reloc_root = (root->root_key.objectid ==
5798                                     BTRFS_TREE_RELOC_OBJECTID);
5799         struct btrfs_fs_info *fs_info = root->fs_info;
5800         struct btrfs_path *path;
5801         struct btrfs_trans_handle *trans;
5802         struct btrfs_root *tree_root = fs_info->tree_root;
5803         struct btrfs_root_item *root_item = &root->root_item;
5804         struct walk_control *wc;
5805         struct btrfs_key key;
5806         int err = 0;
5807         int ret;
5808         int level;
5809         bool root_dropped = false;
5810         bool unfinished_drop = false;
5811
5812         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5813
5814         path = btrfs_alloc_path();
5815         if (!path) {
5816                 err = -ENOMEM;
5817                 goto out;
5818         }
5819
5820         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5821         if (!wc) {
5822                 btrfs_free_path(path);
5823                 err = -ENOMEM;
5824                 goto out;
5825         }
5826
5827         /*
5828          * Use join to avoid potential EINTR from transaction start. See
5829          * wait_reserve_ticket and the whole reservation callchain.
5830          */
5831         if (for_reloc)
5832                 trans = btrfs_join_transaction(tree_root);
5833         else
5834                 trans = btrfs_start_transaction(tree_root, 0);
5835         if (IS_ERR(trans)) {
5836                 err = PTR_ERR(trans);
5837                 goto out_free;
5838         }
5839
5840         err = btrfs_run_delayed_items(trans);
5841         if (err)
5842                 goto out_end_trans;
5843
5844         /*
5845          * This will help us catch people modifying the fs tree while we're
5846          * dropping it.  It is unsafe to mess with the fs tree while it's being
5847          * dropped as we unlock the root node and parent nodes as we walk down
5848          * the tree, assuming nothing will change.  If something does change
5849          * then we'll have stale information and drop references to blocks we've
5850          * already dropped.
5851          */
5852         set_bit(BTRFS_ROOT_DELETING, &root->state);
5853         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5854
5855         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5856                 level = btrfs_header_level(root->node);
5857                 path->nodes[level] = btrfs_lock_root_node(root);
5858                 path->slots[level] = 0;
5859                 path->locks[level] = BTRFS_WRITE_LOCK;
5860                 memset(&wc->update_progress, 0,
5861                        sizeof(wc->update_progress));
5862         } else {
5863                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5864                 memcpy(&wc->update_progress, &key,
5865                        sizeof(wc->update_progress));
5866
5867                 level = btrfs_root_drop_level(root_item);
5868                 BUG_ON(level == 0);
5869                 path->lowest_level = level;
5870                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5871                 path->lowest_level = 0;
5872                 if (ret < 0) {
5873                         err = ret;
5874                         goto out_end_trans;
5875                 }
5876                 WARN_ON(ret > 0);
5877
5878                 /*
5879                  * unlock our path, this is safe because only this
5880                  * function is allowed to delete this snapshot
5881                  */
5882                 btrfs_unlock_up_safe(path, 0);
5883
5884                 level = btrfs_header_level(root->node);
5885                 while (1) {
5886                         btrfs_tree_lock(path->nodes[level]);
5887                         path->locks[level] = BTRFS_WRITE_LOCK;
5888
5889                         ret = btrfs_lookup_extent_info(trans, fs_info,
5890                                                 path->nodes[level]->start,
5891                                                 level, 1, &wc->refs[level],
5892                                                 &wc->flags[level], NULL);
5893                         if (ret < 0) {
5894                                 err = ret;
5895                                 goto out_end_trans;
5896                         }
5897                         BUG_ON(wc->refs[level] == 0);
5898
5899                         if (level == btrfs_root_drop_level(root_item))
5900                                 break;
5901
5902                         btrfs_tree_unlock(path->nodes[level]);
5903                         path->locks[level] = 0;
5904                         WARN_ON(wc->refs[level] != 1);
5905                         level--;
5906                 }
5907         }
5908
5909         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5910         wc->level = level;
5911         wc->shared_level = -1;
5912         wc->stage = DROP_REFERENCE;
5913         wc->update_ref = update_ref;
5914         wc->keep_locks = 0;
5915         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5916
5917         while (1) {
5918
5919                 ret = walk_down_tree(trans, root, path, wc);
5920                 if (ret < 0) {
5921                         btrfs_abort_transaction(trans, ret);
5922                         err = ret;
5923                         break;
5924                 }
5925
5926                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5927                 if (ret < 0) {
5928                         btrfs_abort_transaction(trans, ret);
5929                         err = ret;
5930                         break;
5931                 }
5932
5933                 if (ret > 0) {
5934                         BUG_ON(wc->stage != DROP_REFERENCE);
5935                         break;
5936                 }
5937
5938                 if (wc->stage == DROP_REFERENCE) {
5939                         wc->drop_level = wc->level;
5940                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5941                                               &wc->drop_progress,
5942                                               path->slots[wc->drop_level]);
5943                 }
5944                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5945                                       &wc->drop_progress);
5946                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5947
5948                 BUG_ON(wc->level == 0);
5949                 if (btrfs_should_end_transaction(trans) ||
5950                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5951                         ret = btrfs_update_root(trans, tree_root,
5952                                                 &root->root_key,
5953                                                 root_item);
5954                         if (ret) {
5955                                 btrfs_abort_transaction(trans, ret);
5956                                 err = ret;
5957                                 goto out_end_trans;
5958                         }
5959
5960                         if (!is_reloc_root)
5961                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5962
5963                         btrfs_end_transaction_throttle(trans);
5964                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5965                                 btrfs_debug(fs_info,
5966                                             "drop snapshot early exit");
5967                                 err = -EAGAIN;
5968                                 goto out_free;
5969                         }
5970
5971                        /*
5972                         * Use join to avoid potential EINTR from transaction
5973                         * start. See wait_reserve_ticket and the whole
5974                         * reservation callchain.
5975                         */
5976                         if (for_reloc)
5977                                 trans = btrfs_join_transaction(tree_root);
5978                         else
5979                                 trans = btrfs_start_transaction(tree_root, 0);
5980                         if (IS_ERR(trans)) {
5981                                 err = PTR_ERR(trans);
5982                                 goto out_free;
5983                         }
5984                 }
5985         }
5986         btrfs_release_path(path);
5987         if (err)
5988                 goto out_end_trans;
5989
5990         ret = btrfs_del_root(trans, &root->root_key);
5991         if (ret) {
5992                 btrfs_abort_transaction(trans, ret);
5993                 err = ret;
5994                 goto out_end_trans;
5995         }
5996
5997         if (!is_reloc_root) {
5998                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5999                                       NULL, NULL);
6000                 if (ret < 0) {
6001                         btrfs_abort_transaction(trans, ret);
6002                         err = ret;
6003                         goto out_end_trans;
6004                 } else if (ret > 0) {
6005                         /* if we fail to delete the orphan item this time
6006                          * around, it'll get picked up the next time.
6007                          *
6008                          * The most common failure here is just -ENOENT.
6009                          */
6010                         btrfs_del_orphan_item(trans, tree_root,
6011                                               root->root_key.objectid);
6012                 }
6013         }
6014
6015         /*
6016          * This subvolume is going to be completely dropped, and won't be
6017          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
6018          * commit transaction time.  So free it here manually.
6019          */
6020         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
6021         btrfs_qgroup_free_meta_all_pertrans(root);
6022
6023         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
6024                 btrfs_add_dropped_root(trans, root);
6025         else
6026                 btrfs_put_root(root);
6027         root_dropped = true;
6028 out_end_trans:
6029         if (!is_reloc_root)
6030                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6031
6032         btrfs_end_transaction_throttle(trans);
6033 out_free:
6034         kfree(wc);
6035         btrfs_free_path(path);
6036 out:
6037         /*
6038          * We were an unfinished drop root, check to see if there are any
6039          * pending, and if not clear and wake up any waiters.
6040          */
6041         if (!err && unfinished_drop)
6042                 btrfs_maybe_wake_unfinished_drop(fs_info);
6043
6044         /*
6045          * So if we need to stop dropping the snapshot for whatever reason we
6046          * need to make sure to add it back to the dead root list so that we
6047          * keep trying to do the work later.  This also cleans up roots if we
6048          * don't have it in the radix (like when we recover after a power fail
6049          * or unmount) so we don't leak memory.
6050          */
6051         if (!for_reloc && !root_dropped)
6052                 btrfs_add_dead_root(root);
6053         return err;
6054 }
6055
6056 /*
6057  * drop subtree rooted at tree block 'node'.
6058  *
6059  * NOTE: this function will unlock and release tree block 'node'
6060  * only used by relocation code
6061  */
6062 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6063                         struct btrfs_root *root,
6064                         struct extent_buffer *node,
6065                         struct extent_buffer *parent)
6066 {
6067         struct btrfs_fs_info *fs_info = root->fs_info;
6068         struct btrfs_path *path;
6069         struct walk_control *wc;
6070         int level;
6071         int parent_level;
6072         int ret = 0;
6073         int wret;
6074
6075         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6076
6077         path = btrfs_alloc_path();
6078         if (!path)
6079                 return -ENOMEM;
6080
6081         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6082         if (!wc) {
6083                 btrfs_free_path(path);
6084                 return -ENOMEM;
6085         }
6086
6087         btrfs_assert_tree_write_locked(parent);
6088         parent_level = btrfs_header_level(parent);
6089         atomic_inc(&parent->refs);
6090         path->nodes[parent_level] = parent;
6091         path->slots[parent_level] = btrfs_header_nritems(parent);
6092
6093         btrfs_assert_tree_write_locked(node);
6094         level = btrfs_header_level(node);
6095         path->nodes[level] = node;
6096         path->slots[level] = 0;
6097         path->locks[level] = BTRFS_WRITE_LOCK;
6098
6099         wc->refs[parent_level] = 1;
6100         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6101         wc->level = level;
6102         wc->shared_level = -1;
6103         wc->stage = DROP_REFERENCE;
6104         wc->update_ref = 0;
6105         wc->keep_locks = 1;
6106         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
6107
6108         while (1) {
6109                 wret = walk_down_tree(trans, root, path, wc);
6110                 if (wret < 0) {
6111                         ret = wret;
6112                         break;
6113                 }
6114
6115                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6116                 if (wret < 0)
6117                         ret = wret;
6118                 if (wret != 0)
6119                         break;
6120         }
6121
6122         kfree(wc);
6123         btrfs_free_path(path);
6124         return ret;
6125 }
6126
6127 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
6128                                    u64 start, u64 end)
6129 {
6130         return unpin_extent_range(fs_info, start, end, false);
6131 }
6132
6133 /*
6134  * It used to be that old block groups would be left around forever.
6135  * Iterating over them would be enough to trim unused space.  Since we
6136  * now automatically remove them, we also need to iterate over unallocated
6137  * space.
6138  *
6139  * We don't want a transaction for this since the discard may take a
6140  * substantial amount of time.  We don't require that a transaction be
6141  * running, but we do need to take a running transaction into account
6142  * to ensure that we're not discarding chunks that were released or
6143  * allocated in the current transaction.
6144  *
6145  * Holding the chunks lock will prevent other threads from allocating
6146  * or releasing chunks, but it won't prevent a running transaction
6147  * from committing and releasing the memory that the pending chunks
6148  * list head uses.  For that, we need to take a reference to the
6149  * transaction and hold the commit root sem.  We only need to hold
6150  * it while performing the free space search since we have already
6151  * held back allocations.
6152  */
6153 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6154 {
6155         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6156         int ret;
6157
6158         *trimmed = 0;
6159
6160         /* Discard not supported = nothing to do. */
6161         if (!bdev_max_discard_sectors(device->bdev))
6162                 return 0;
6163
6164         /* Not writable = nothing to do. */
6165         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6166                 return 0;
6167
6168         /* No free space = nothing to do. */
6169         if (device->total_bytes <= device->bytes_used)
6170                 return 0;
6171
6172         ret = 0;
6173
6174         while (1) {
6175                 struct btrfs_fs_info *fs_info = device->fs_info;
6176                 u64 bytes;
6177
6178                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6179                 if (ret)
6180                         break;
6181
6182                 find_first_clear_extent_bit(&device->alloc_state, start,
6183                                             &start, &end,
6184                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6185
6186                 /* Check if there are any CHUNK_* bits left */
6187                 if (start > device->total_bytes) {
6188                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6189                         btrfs_warn_in_rcu(fs_info,
6190 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6191                                           start, end - start + 1,
6192                                           btrfs_dev_name(device),
6193                                           device->total_bytes);
6194                         mutex_unlock(&fs_info->chunk_mutex);
6195                         ret = 0;
6196                         break;
6197                 }
6198
6199                 /* Ensure we skip the reserved space on each device. */
6200                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6201
6202                 /*
6203                  * If find_first_clear_extent_bit find a range that spans the
6204                  * end of the device it will set end to -1, in this case it's up
6205                  * to the caller to trim the value to the size of the device.
6206                  */
6207                 end = min(end, device->total_bytes - 1);
6208
6209                 len = end - start + 1;
6210
6211                 /* We didn't find any extents */
6212                 if (!len) {
6213                         mutex_unlock(&fs_info->chunk_mutex);
6214                         ret = 0;
6215                         break;
6216                 }
6217
6218                 ret = btrfs_issue_discard(device->bdev, start, len,
6219                                           &bytes);
6220                 if (!ret)
6221                         set_extent_bit(&device->alloc_state, start,
6222                                        start + bytes - 1, CHUNK_TRIMMED, NULL);
6223                 mutex_unlock(&fs_info->chunk_mutex);
6224
6225                 if (ret)
6226                         break;
6227
6228                 start += len;
6229                 *trimmed += bytes;
6230
6231                 if (fatal_signal_pending(current)) {
6232                         ret = -ERESTARTSYS;
6233                         break;
6234                 }
6235
6236                 cond_resched();
6237         }
6238
6239         return ret;
6240 }
6241
6242 /*
6243  * Trim the whole filesystem by:
6244  * 1) trimming the free space in each block group
6245  * 2) trimming the unallocated space on each device
6246  *
6247  * This will also continue trimming even if a block group or device encounters
6248  * an error.  The return value will be the last error, or 0 if nothing bad
6249  * happens.
6250  */
6251 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6252 {
6253         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6254         struct btrfs_block_group *cache = NULL;
6255         struct btrfs_device *device;
6256         u64 group_trimmed;
6257         u64 range_end = U64_MAX;
6258         u64 start;
6259         u64 end;
6260         u64 trimmed = 0;
6261         u64 bg_failed = 0;
6262         u64 dev_failed = 0;
6263         int bg_ret = 0;
6264         int dev_ret = 0;
6265         int ret = 0;
6266
6267         if (range->start == U64_MAX)
6268                 return -EINVAL;
6269
6270         /*
6271          * Check range overflow if range->len is set.
6272          * The default range->len is U64_MAX.
6273          */
6274         if (range->len != U64_MAX &&
6275             check_add_overflow(range->start, range->len, &range_end))
6276                 return -EINVAL;
6277
6278         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6279         for (; cache; cache = btrfs_next_block_group(cache)) {
6280                 if (cache->start >= range_end) {
6281                         btrfs_put_block_group(cache);
6282                         break;
6283                 }
6284
6285                 start = max(range->start, cache->start);
6286                 end = min(range_end, cache->start + cache->length);
6287
6288                 if (end - start >= range->minlen) {
6289                         if (!btrfs_block_group_done(cache)) {
6290                                 ret = btrfs_cache_block_group(cache, true);
6291                                 if (ret) {
6292                                         bg_failed++;
6293                                         bg_ret = ret;
6294                                         continue;
6295                                 }
6296                         }
6297                         ret = btrfs_trim_block_group(cache,
6298                                                      &group_trimmed,
6299                                                      start,
6300                                                      end,
6301                                                      range->minlen);
6302
6303                         trimmed += group_trimmed;
6304                         if (ret) {
6305                                 bg_failed++;
6306                                 bg_ret = ret;
6307                                 continue;
6308                         }
6309                 }
6310         }
6311
6312         if (bg_failed)
6313                 btrfs_warn(fs_info,
6314                         "failed to trim %llu block group(s), last error %d",
6315                         bg_failed, bg_ret);
6316
6317         mutex_lock(&fs_devices->device_list_mutex);
6318         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6319                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6320                         continue;
6321
6322                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6323                 if (ret) {
6324                         dev_failed++;
6325                         dev_ret = ret;
6326                         break;
6327                 }
6328
6329                 trimmed += group_trimmed;
6330         }
6331         mutex_unlock(&fs_devices->device_list_mutex);
6332
6333         if (dev_failed)
6334                 btrfs_warn(fs_info,
6335                         "failed to trim %llu device(s), last error %d",
6336                         dev_failed, dev_ret);
6337         range->len = trimmed;
6338         if (bg_ret)
6339                 return bg_ret;
6340         return dev_ret;
6341 }