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