extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "ctree.h"
  20 #include "extent-tree.h"
  21 #include "tree-log.h"
  22 #include "disk-io.h"
  23 #include "print-tree.h"
  24 #include "volumes.h"
  25 #include "raid56.h"
  26 #include "locking.h"
  27 #include "free-space-cache.h"
  28 #include "free-space-tree.h"
  29 #include "sysfs.h"
  30 #include "qgroup.h"
  31 #include "ref-verify.h"
  32 #include "space-info.h"
  33 #include "block-rsv.h"
  34 #include "delalloc-space.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "root-tree.h"
  42 #include "file-item.h"
  43 #include "orphan.h"
  44 #include "tree-checker.h"
  45 #include "raid-stripe-tree.h"
  46
  47 #undef SCRAMBLE_DELAYED_REFS
  48
  49
  50 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  51                                struct btrfs_delayed_ref_head *href,
  52                                struct btrfs_delayed_ref_node *node, u64 parent,
  53                                u64 root_objectid, u64 owner_objectid,
  54                                u64 owner_offset,
  55                                struct btrfs_delayed_extent_op *extra_op);
  56 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  57                                     struct extent_buffer *leaf,
  58                                     struct btrfs_extent_item *ei);
  59 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  60                                       u64 parent, u64 root_objectid,
  61                                       u64 flags, u64 owner, u64 offset,
  62                                       struct btrfs_key *ins, int ref_mod, u64 oref_root);
  63 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  64                                      struct btrfs_delayed_ref_node *node,
  65                                      struct btrfs_delayed_extent_op *extent_op);
  66 static int find_next_key(struct btrfs_path *path, int level,
  67                          struct btrfs_key *key);
  68
  69 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  70 {
  71         return (cache->flags & bits) == bits;
  72 }
  73
  74 /* simple helper to search for an existing data extent at a given offset */
  75 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  76 {
  77         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  78         int ret;
  79         struct btrfs_key key;
  80         struct btrfs_path *path;
  81
  82         path = btrfs_alloc_path();
  83         if (!path)
  84                 return -ENOMEM;
  85
  86         key.objectid = start;
  87         key.offset = len;
  88         key.type = BTRFS_EXTENT_ITEM_KEY;
  89         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  90         btrfs_free_path(path);
  91         return ret;
  92 }
  93
  94 /*
  95  * helper function to lookup reference count and flags of a tree block.
  96  *
  97  * the head node for delayed ref is used to store the sum of all the
  98  * reference count modifications queued up in the rbtree. the head
  99  * node may also store the extent flags to set. This way you can check
 100  * to see what the reference count and extent flags would be if all of
 101  * the delayed refs are not processed.
 102  */
 103 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 104                              struct btrfs_fs_info *fs_info, u64 bytenr,
 105                              u64 offset, int metadata, u64 *refs, u64 *flags,
 106                              u64 *owning_root)
 107 {
 108         struct btrfs_root *extent_root;
 109         struct btrfs_delayed_ref_head *head;
 110         struct btrfs_delayed_ref_root *delayed_refs;
 111         struct btrfs_path *path;
 112         struct btrfs_extent_item *ei;
 113         struct extent_buffer *leaf;
 114         struct btrfs_key key;
 115         u32 item_size;
 116         u64 num_refs;
 117         u64 extent_flags;
 118         u64 owner = 0;
 119         int ret;
 120
 121         /*
 122          * If we don't have skinny metadata, don't bother doing anything
 123          * different
 124          */
 125         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 126                 offset = fs_info->nodesize;
 127                 metadata = 0;
 128         }
 129
 130         path = btrfs_alloc_path();
 131         if (!path)
 132                 return -ENOMEM;
 133
 134         if (!trans) {
 135                 path->skip_locking = 1;
 136                 path->search_commit_root = 1;
 137         }
 138
 139 search_again:
 140         key.objectid = bytenr;
 141         key.offset = offset;
 142         if (metadata)
 143                 key.type = BTRFS_METADATA_ITEM_KEY;
 144         else
 145                 key.type = BTRFS_EXTENT_ITEM_KEY;
 146
 147         extent_root = btrfs_extent_root(fs_info, bytenr);
 148         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 149         if (ret < 0)
 150                 goto out_free;
 151
 152         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 153                 if (path->slots[0]) {
 154                         path->slots[0]--;
 155                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 156                                               path->slots[0]);
 157                         if (key.objectid == bytenr &&
 158                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 159                             key.offset == fs_info->nodesize)
 160                                 ret = 0;
 161                 }
 162         }
 163
 164         if (ret == 0) {
 165                 leaf = path->nodes[0];
 166                 item_size = btrfs_item_size(leaf, path->slots[0]);
 167                 if (item_size >= sizeof(*ei)) {
 168                         ei = btrfs_item_ptr(leaf, path->slots[0],
 169                                             struct btrfs_extent_item);
 170                         num_refs = btrfs_extent_refs(leaf, ei);
 171                         extent_flags = btrfs_extent_flags(leaf, ei);
 172                         owner = btrfs_get_extent_owner_root(fs_info, leaf,
 173                                                             path->slots[0]);
 174                 } else {
 175                         ret = -EUCLEAN;
 176                         btrfs_err(fs_info,
 177                         "unexpected extent item size, has %u expect >= %zu",
 178                                   item_size, sizeof(*ei));
 179                         if (trans)
 180                                 btrfs_abort_transaction(trans, ret);
 181                         else
 182                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 183
 184                         goto out_free;
 185                 }
 186
 187                 BUG_ON(num_refs == 0);
 188         } else {
 189                 num_refs = 0;
 190                 extent_flags = 0;
 191                 ret = 0;
 192         }
 193
 194         if (!trans)
 195                 goto out;
 196
 197         delayed_refs = &trans->transaction->delayed_refs;
 198         spin_lock(&delayed_refs->lock);
 199         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 200         if (head) {
 201                 if (!mutex_trylock(&head->mutex)) {
 202                         refcount_inc(&head->refs);
 203                         spin_unlock(&delayed_refs->lock);
 204
 205                         btrfs_release_path(path);
 206
 207                         /*
 208                          * Mutex was contended, block until it's released and try
 209                          * again
 210                          */
 211                         mutex_lock(&head->mutex);
 212                         mutex_unlock(&head->mutex);
 213                         btrfs_put_delayed_ref_head(head);
 214                         goto search_again;
 215                 }
 216                 spin_lock(&head->lock);
 217                 if (head->extent_op && head->extent_op->update_flags)
 218                         extent_flags |= head->extent_op->flags_to_set;
 219                 else
 220                         BUG_ON(num_refs == 0);
 221
 222                 num_refs += head->ref_mod;
 223                 spin_unlock(&head->lock);
 224                 mutex_unlock(&head->mutex);
 225         }
 226         spin_unlock(&delayed_refs->lock);
 227 out:
 228         WARN_ON(num_refs == 0);
 229         if (refs)
 230                 *refs = num_refs;
 231         if (flags)
 232                 *flags = extent_flags;
 233         if (owning_root)
 234                 *owning_root = owner;
 235 out_free:
 236         btrfs_free_path(path);
 237         return ret;
 238 }
 239
 240 /*
 241  * Back reference rules.  Back refs have three main goals:
 242  *
 243  * 1) differentiate between all holders of references to an extent so that
 244  *    when a reference is dropped we can make sure it was a valid reference
 245  *    before freeing the extent.
 246  *
 247  * 2) Provide enough information to quickly find the holders of an extent
 248  *    if we notice a given block is corrupted or bad.
 249  *
 250  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 251  *    maintenance.  This is actually the same as #2, but with a slightly
 252  *    different use case.
 253  *
 254  * There are two kinds of back refs. The implicit back refs is optimized
 255  * for pointers in non-shared tree blocks. For a given pointer in a block,
 256  * back refs of this kind provide information about the block's owner tree
 257  * and the pointer's key. These information allow us to find the block by
 258  * b-tree searching. The full back refs is for pointers in tree blocks not
 259  * referenced by their owner trees. The location of tree block is recorded
 260  * in the back refs. Actually the full back refs is generic, and can be
 261  * used in all cases the implicit back refs is used. The major shortcoming
 262  * of the full back refs is its overhead. Every time a tree block gets
 263  * COWed, we have to update back refs entry for all pointers in it.
 264  *
 265  * For a newly allocated tree block, we use implicit back refs for
 266  * pointers in it. This means most tree related operations only involve
 267  * implicit back refs. For a tree block created in old transaction, the
 268  * only way to drop a reference to it is COW it. So we can detect the
 269  * event that tree block loses its owner tree's reference and do the
 270  * back refs conversion.
 271  *
 272  * When a tree block is COWed through a tree, there are four cases:
 273  *
 274  * The reference count of the block is one and the tree is the block's
 275  * owner tree. Nothing to do in this case.
 276  *
 277  * The reference count of the block is one and the tree is not the
 278  * block's owner tree. In this case, full back refs is used for pointers
 279  * in the block. Remove these full back refs, add implicit back refs for
 280  * every pointers in the new block.
 281  *
 282  * The reference count of the block is greater than one and the tree is
 283  * the block's owner tree. In this case, implicit back refs is used for
 284  * pointers in the block. Add full back refs for every pointers in the
 285  * block, increase lower level extents' reference counts. The original
 286  * implicit back refs are entailed to the new block.
 287  *
 288  * The reference count of the block is greater than one and the tree is
 289  * not the block's owner tree. Add implicit back refs for every pointer in
 290  * the new block, increase lower level extents' reference count.
 291  *
 292  * Back Reference Key composing:
 293  *
 294  * The key objectid corresponds to the first byte in the extent,
 295  * The key type is used to differentiate between types of back refs.
 296  * There are different meanings of the key offset for different types
 297  * of back refs.
 298  *
 299  * File extents can be referenced by:
 300  *
 301  * - multiple snapshots, subvolumes, or different generations in one subvol
 302  * - different files inside a single subvolume
 303  * - different offsets inside a file (bookend extents in file.c)
 304  *
 305  * The extent ref structure for the implicit back refs has fields for:
 306  *
 307  * - Objectid of the subvolume root
 308  * - objectid of the file holding the reference
 309  * - original offset in the file
 310  * - how many bookend extents
 311  *
 312  * The key offset for the implicit back refs is hash of the first
 313  * three fields.
 314  *
 315  * The extent ref structure for the full back refs has field for:
 316  *
 317  * - number of pointers in the tree leaf
 318  *
 319  * The key offset for the implicit back refs is the first byte of
 320  * the tree leaf
 321  *
 322  * When a file extent is allocated, The implicit back refs is used.
 323  * the fields are filled in:
 324  *
 325  *     (root_key.objectid, inode objectid, offset in file, 1)
 326  *
 327  * When a file extent is removed file truncation, we find the
 328  * corresponding implicit back refs and check the following fields:
 329  *
 330  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 331  *
 332  * Btree extents can be referenced by:
 333  *
 334  * - Different subvolumes
 335  *
 336  * Both the implicit back refs and the full back refs for tree blocks
 337  * only consist of key. The key offset for the implicit back refs is
 338  * objectid of block's owner tree. The key offset for the full back refs
 339  * is the first byte of parent block.
 340  *
 341  * When implicit back refs is used, information about the lowest key and
 342  * level of the tree block are required. These information are stored in
 343  * tree block info structure.
 344  */
 345
 346 /*
 347  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 348  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 349  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 350  */
 351 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 352                                      struct btrfs_extent_inline_ref *iref,
 353                                      enum btrfs_inline_ref_type is_data)
 354 {
 355         struct btrfs_fs_info *fs_info = eb->fs_info;
 356         int type = btrfs_extent_inline_ref_type(eb, iref);
 357         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 358
 359         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 360                 ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 361                 return type;
 362         }
 363
 364         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 365             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 366             type == BTRFS_SHARED_DATA_REF_KEY ||
 367             type == BTRFS_EXTENT_DATA_REF_KEY) {
 368                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 369                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 370                                 return type;
 371                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 372                                 ASSERT(fs_info);
 373                                 /*
 374                                  * Every shared one has parent tree block,
 375                                  * which must be aligned to sector size.
 376                                  */
 377                                 if (offset && IS_ALIGNED(offset, fs_info->sectorsize))
 378                                         return type;
 379                         }
 380                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 381                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 382                                 return type;
 383                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 384                                 ASSERT(fs_info);
 385                                 /*
 386                                  * Every shared one has parent tree block,
 387                                  * which must be aligned to sector size.
 388                                  */
 389                                 if (offset &&
 390                                     IS_ALIGNED(offset, fs_info->sectorsize))
 391                                         return type;
 392                         }
 393                 } else {
 394                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 395                         return type;
 396                 }
 397         }
 398
 399         WARN_ON(1);
 400         btrfs_print_leaf(eb);
 401         btrfs_err(fs_info,
 402                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 403                   eb->start, (unsigned long)iref, type);
 404
 405         return BTRFS_REF_TYPE_INVALID;
 406 }
 407
 408 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 409 {
 410         u32 high_crc = ~(u32)0;
 411         u32 low_crc = ~(u32)0;
 412         __le64 lenum;
 413
 414         lenum = cpu_to_le64(root_objectid);
 415         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 416         lenum = cpu_to_le64(owner);
 417         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 418         lenum = cpu_to_le64(offset);
 419         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 420
 421         return ((u64)high_crc << 31) ^ (u64)low_crc;
 422 }
 423
 424 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 425                                      struct btrfs_extent_data_ref *ref)
 426 {
 427         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 428                                     btrfs_extent_data_ref_objectid(leaf, ref),
 429                                     btrfs_extent_data_ref_offset(leaf, ref));
 430 }
 431
 432 static int match_extent_data_ref(struct extent_buffer *leaf,
 433                                  struct btrfs_extent_data_ref *ref,
 434                                  u64 root_objectid, u64 owner, u64 offset)
 435 {
 436         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 437             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 438             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 439                 return 0;
 440         return 1;
 441 }
 442
 443 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 444                                            struct btrfs_path *path,
 445                                            u64 bytenr, u64 parent,
 446                                            u64 root_objectid,
 447                                            u64 owner, u64 offset)
 448 {
 449         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 450         struct btrfs_key key;
 451         struct btrfs_extent_data_ref *ref;
 452         struct extent_buffer *leaf;
 453         u32 nritems;
 454         int ret;
 455         int recow;
 456         int err = -ENOENT;
 457
 458         key.objectid = bytenr;
 459         if (parent) {
 460                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 461                 key.offset = parent;
 462         } else {
 463                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 464                 key.offset = hash_extent_data_ref(root_objectid,
 465                                                   owner, offset);
 466         }
 467 again:
 468         recow = 0;
 469         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 470         if (ret < 0) {
 471                 err = ret;
 472                 goto fail;
 473         }
 474
 475         if (parent) {
 476                 if (!ret)
 477                         return 0;
 478                 goto fail;
 479         }
 480
 481         leaf = path->nodes[0];
 482         nritems = btrfs_header_nritems(leaf);
 483         while (1) {
 484                 if (path->slots[0] >= nritems) {
 485                         ret = btrfs_next_leaf(root, path);
 486                         if (ret < 0)
 487                                 err = ret;
 488                         if (ret)
 489                                 goto fail;
 490
 491                         leaf = path->nodes[0];
 492                         nritems = btrfs_header_nritems(leaf);
 493                         recow = 1;
 494                 }
 495
 496                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 497                 if (key.objectid != bytenr ||
 498                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 499                         goto fail;
 500
 501                 ref = btrfs_item_ptr(leaf, path->slots[0],
 502                                      struct btrfs_extent_data_ref);
 503
 504                 if (match_extent_data_ref(leaf, ref, root_objectid,
 505                                           owner, offset)) {
 506                         if (recow) {
 507                                 btrfs_release_path(path);
 508                                 goto again;
 509                         }
 510                         err = 0;
 511                         break;
 512                 }
 513                 path->slots[0]++;
 514         }
 515 fail:
 516         return err;
 517 }
 518
 519 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 520                                            struct btrfs_path *path,
 521                                            u64 bytenr, u64 parent,
 522                                            u64 root_objectid, u64 owner,
 523                                            u64 offset, int refs_to_add)
 524 {
 525         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 526         struct btrfs_key key;
 527         struct extent_buffer *leaf;
 528         u32 size;
 529         u32 num_refs;
 530         int ret;
 531
 532         key.objectid = bytenr;
 533         if (parent) {
 534                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 535                 key.offset = parent;
 536                 size = sizeof(struct btrfs_shared_data_ref);
 537         } else {
 538                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 539                 key.offset = hash_extent_data_ref(root_objectid,
 540                                                   owner, offset);
 541                 size = sizeof(struct btrfs_extent_data_ref);
 542         }
 543
 544         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 545         if (ret && ret != -EEXIST)
 546                 goto fail;
 547
 548         leaf = path->nodes[0];
 549         if (parent) {
 550                 struct btrfs_shared_data_ref *ref;
 551                 ref = btrfs_item_ptr(leaf, path->slots[0],
 552                                      struct btrfs_shared_data_ref);
 553                 if (ret == 0) {
 554                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 555                 } else {
 556                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 557                         num_refs += refs_to_add;
 558                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 559                 }
 560         } else {
 561                 struct btrfs_extent_data_ref *ref;
 562                 while (ret == -EEXIST) {
 563                         ref = btrfs_item_ptr(leaf, path->slots[0],
 564                                              struct btrfs_extent_data_ref);
 565                         if (match_extent_data_ref(leaf, ref, root_objectid,
 566                                                   owner, offset))
 567                                 break;
 568                         btrfs_release_path(path);
 569                         key.offset++;
 570                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 571                                                       size);
 572                         if (ret && ret != -EEXIST)
 573                                 goto fail;
 574
 575                         leaf = path->nodes[0];
 576                 }
 577                 ref = btrfs_item_ptr(leaf, path->slots[0],
 578                                      struct btrfs_extent_data_ref);
 579                 if (ret == 0) {
 580                         btrfs_set_extent_data_ref_root(leaf, ref,
 581                                                        root_objectid);
 582                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 583                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 584                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 585                 } else {
 586                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 587                         num_refs += refs_to_add;
 588                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 589                 }
 590         }
 591         btrfs_mark_buffer_dirty(trans, leaf);
 592         ret = 0;
 593 fail:
 594         btrfs_release_path(path);
 595         return ret;
 596 }
 597
 598 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 599                                            struct btrfs_root *root,
 600                                            struct btrfs_path *path,
 601                                            int refs_to_drop)
 602 {
 603         struct btrfs_key key;
 604         struct btrfs_extent_data_ref *ref1 = NULL;
 605         struct btrfs_shared_data_ref *ref2 = NULL;
 606         struct extent_buffer *leaf;
 607         u32 num_refs = 0;
 608         int ret = 0;
 609
 610         leaf = path->nodes[0];
 611         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 612
 613         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 614                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 615                                       struct btrfs_extent_data_ref);
 616                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 617         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 618                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 619                                       struct btrfs_shared_data_ref);
 620                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 621         } else {
 622                 btrfs_err(trans->fs_info,
 623                           "unrecognized backref key (%llu %u %llu)",
 624                           key.objectid, key.type, key.offset);
 625                 btrfs_abort_transaction(trans, -EUCLEAN);
 626                 return -EUCLEAN;
 627         }
 628
 629         BUG_ON(num_refs < refs_to_drop);
 630         num_refs -= refs_to_drop;
 631
 632         if (num_refs == 0) {
 633                 ret = btrfs_del_item(trans, root, path);
 634         } else {
 635                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 636                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 637                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 638                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 639                 btrfs_mark_buffer_dirty(trans, leaf);
 640         }
 641         return ret;
 642 }
 643
 644 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 645                                           struct btrfs_extent_inline_ref *iref)
 646 {
 647         struct btrfs_key key;
 648         struct extent_buffer *leaf;
 649         struct btrfs_extent_data_ref *ref1;
 650         struct btrfs_shared_data_ref *ref2;
 651         u32 num_refs = 0;
 652         int type;
 653
 654         leaf = path->nodes[0];
 655         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 656
 657         if (iref) {
 658                 /*
 659                  * If type is invalid, we should have bailed out earlier than
 660                  * this call.
 661                  */
 662                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 663                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 664                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 665                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 666                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 667                 } else {
 668                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 669                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 670                 }
 671         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 672                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 673                                       struct btrfs_extent_data_ref);
 674                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 675         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 676                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 677                                       struct btrfs_shared_data_ref);
 678                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 679         } else {
 680                 WARN_ON(1);
 681         }
 682         return num_refs;
 683 }
 684
 685 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 686                                           struct btrfs_path *path,
 687                                           u64 bytenr, u64 parent,
 688                                           u64 root_objectid)
 689 {
 690         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 691         struct btrfs_key key;
 692         int ret;
 693
 694         key.objectid = bytenr;
 695         if (parent) {
 696                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 697                 key.offset = parent;
 698         } else {
 699                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 700                 key.offset = root_objectid;
 701         }
 702
 703         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 704         if (ret > 0)
 705                 ret = -ENOENT;
 706         return ret;
 707 }
 708
 709 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 710                                           struct btrfs_path *path,
 711                                           u64 bytenr, u64 parent,
 712                                           u64 root_objectid)
 713 {
 714         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 715         struct btrfs_key key;
 716         int ret;
 717
 718         key.objectid = bytenr;
 719         if (parent) {
 720                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 721                 key.offset = parent;
 722         } else {
 723                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 724                 key.offset = root_objectid;
 725         }
 726
 727         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 728         btrfs_release_path(path);
 729         return ret;
 730 }
 731
 732 static inline int extent_ref_type(u64 parent, u64 owner)
 733 {
 734         int type;
 735         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 736                 if (parent > 0)
 737                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 738                 else
 739                         type = BTRFS_TREE_BLOCK_REF_KEY;
 740         } else {
 741                 if (parent > 0)
 742                         type = BTRFS_SHARED_DATA_REF_KEY;
 743                 else
 744                         type = BTRFS_EXTENT_DATA_REF_KEY;
 745         }
 746         return type;
 747 }
 748
 749 static int find_next_key(struct btrfs_path *path, int level,
 750                          struct btrfs_key *key)
 751
 752 {
 753         for (; level < BTRFS_MAX_LEVEL; level++) {
 754                 if (!path->nodes[level])
 755                         break;
 756                 if (path->slots[level] + 1 >=
 757                     btrfs_header_nritems(path->nodes[level]))
 758                         continue;
 759                 if (level == 0)
 760                         btrfs_item_key_to_cpu(path->nodes[level], key,
 761                                               path->slots[level] + 1);
 762                 else
 763                         btrfs_node_key_to_cpu(path->nodes[level], key,
 764                                               path->slots[level] + 1);
 765                 return 0;
 766         }
 767         return 1;
 768 }
 769
 770 /*
 771  * look for inline back ref. if back ref is found, *ref_ret is set
 772  * to the address of inline back ref, and 0 is returned.
 773  *
 774  * if back ref isn't found, *ref_ret is set to the address where it
 775  * should be inserted, and -ENOENT is returned.
 776  *
 777  * if insert is true and there are too many inline back refs, the path
 778  * points to the extent item, and -EAGAIN is returned.
 779  *
 780  * NOTE: inline back refs are ordered in the same way that back ref
 781  *       items in the tree are ordered.
 782  */
 783 static noinline_for_stack
 784 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 785                                  struct btrfs_path *path,
 786                                  struct btrfs_extent_inline_ref **ref_ret,
 787                                  u64 bytenr, u64 num_bytes,
 788                                  u64 parent, u64 root_objectid,
 789                                  u64 owner, u64 offset, int insert)
 790 {
 791         struct btrfs_fs_info *fs_info = trans->fs_info;
 792         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 793         struct btrfs_key key;
 794         struct extent_buffer *leaf;
 795         struct btrfs_extent_item *ei;
 796         struct btrfs_extent_inline_ref *iref;
 797         u64 flags;
 798         u64 item_size;
 799         unsigned long ptr;
 800         unsigned long end;
 801         int extra_size;
 802         int type;
 803         int want;
 804         int ret;
 805         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 806         int needed;
 807
 808         key.objectid = bytenr;
 809         key.type = BTRFS_EXTENT_ITEM_KEY;
 810         key.offset = num_bytes;
 811
 812         want = extent_ref_type(parent, owner);
 813         if (insert) {
 814                 extra_size = btrfs_extent_inline_ref_size(want);
 815                 path->search_for_extension = 1;
 816                 path->keep_locks = 1;
 817         } else
 818                 extra_size = -1;
 819
 820         /*
 821          * Owner is our level, so we can just add one to get the level for the
 822          * block we are interested in.
 823          */
 824         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 825                 key.type = BTRFS_METADATA_ITEM_KEY;
 826                 key.offset = owner;
 827         }
 828
 829 again:
 830         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 831         if (ret < 0)
 832                 goto out;
 833
 834         /*
 835          * We may be a newly converted file system which still has the old fat
 836          * extent entries for metadata, so try and see if we have one of those.
 837          */
 838         if (ret > 0 && skinny_metadata) {
 839                 skinny_metadata = false;
 840                 if (path->slots[0]) {
 841                         path->slots[0]--;
 842                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 843                                               path->slots[0]);
 844                         if (key.objectid == bytenr &&
 845                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 846                             key.offset == num_bytes)
 847                                 ret = 0;
 848                 }
 849                 if (ret) {
 850                         key.objectid = bytenr;
 851                         key.type = BTRFS_EXTENT_ITEM_KEY;
 852                         key.offset = num_bytes;
 853                         btrfs_release_path(path);
 854                         goto again;
 855                 }
 856         }
 857
 858         if (ret && !insert) {
 859                 ret = -ENOENT;
 860                 goto out;
 861         } else if (WARN_ON(ret)) {
 862                 btrfs_print_leaf(path->nodes[0]);
 863                 btrfs_err(fs_info,
 864 "extent item not found for insert, bytenr %llu num_bytes %llu parent %llu root_objectid %llu owner %llu offset %llu",
 865                           bytenr, num_bytes, parent, root_objectid, owner,
 866                           offset);
 867                 ret = -EUCLEAN;
 868                 goto out;
 869         }
 870
 871         leaf = path->nodes[0];
 872         item_size = btrfs_item_size(leaf, path->slots[0]);
 873         if (unlikely(item_size < sizeof(*ei))) {
 874                 ret = -EUCLEAN;
 875                 btrfs_err(fs_info,
 876                           "unexpected extent item size, has %llu expect >= %zu",
 877                           item_size, sizeof(*ei));
 878                 btrfs_abort_transaction(trans, ret);
 879                 goto out;
 880         }
 881
 882         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 883         flags = btrfs_extent_flags(leaf, ei);
 884
 885         ptr = (unsigned long)(ei + 1);
 886         end = (unsigned long)ei + item_size;
 887
 888         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 889                 ptr += sizeof(struct btrfs_tree_block_info);
 890                 BUG_ON(ptr > end);
 891         }
 892
 893         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 894                 needed = BTRFS_REF_TYPE_DATA;
 895         else
 896                 needed = BTRFS_REF_TYPE_BLOCK;
 897
 898         ret = -ENOENT;
 899         while (ptr < end) {
 900                 iref = (struct btrfs_extent_inline_ref *)ptr;
 901                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 902                 if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 903                         ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 904                         ptr += btrfs_extent_inline_ref_size(type);
 905                         continue;
 906                 }
 907                 if (type == BTRFS_REF_TYPE_INVALID) {
 908                         ret = -EUCLEAN;
 909                         goto out;
 910                 }
 911
 912                 if (want < type)
 913                         break;
 914                 if (want > type) {
 915                         ptr += btrfs_extent_inline_ref_size(type);
 916                         continue;
 917                 }
 918
 919                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 920                         struct btrfs_extent_data_ref *dref;
 921                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 922                         if (match_extent_data_ref(leaf, dref, root_objectid,
 923                                                   owner, offset)) {
 924                                 ret = 0;
 925                                 break;
 926                         }
 927                         if (hash_extent_data_ref_item(leaf, dref) <
 928                             hash_extent_data_ref(root_objectid, owner, offset))
 929                                 break;
 930                 } else {
 931                         u64 ref_offset;
 932                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 933                         if (parent > 0) {
 934                                 if (parent == ref_offset) {
 935                                         ret = 0;
 936                                         break;
 937                                 }
 938                                 if (ref_offset < parent)
 939                                         break;
 940                         } else {
 941                                 if (root_objectid == ref_offset) {
 942                                         ret = 0;
 943                                         break;
 944                                 }
 945                                 if (ref_offset < root_objectid)
 946                                         break;
 947                         }
 948                 }
 949                 ptr += btrfs_extent_inline_ref_size(type);
 950         }
 951
 952         if (unlikely(ptr > end)) {
 953                 ret = -EUCLEAN;
 954                 btrfs_print_leaf(path->nodes[0]);
 955                 btrfs_crit(fs_info,
 956 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 957                            path->slots[0], root_objectid, owner, offset, parent);
 958                 goto out;
 959         }
 960
 961         if (ret == -ENOENT && insert) {
 962                 if (item_size + extra_size >=
 963                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 964                         ret = -EAGAIN;
 965                         goto out;
 966                 }
 967                 /*
 968                  * To add new inline back ref, we have to make sure
 969                  * there is no corresponding back ref item.
 970                  * For simplicity, we just do not add new inline back
 971                  * ref if there is any kind of item for this block
 972                  */
 973                 if (find_next_key(path, 0, &key) == 0 &&
 974                     key.objectid == bytenr &&
 975                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 976                         ret = -EAGAIN;
 977                         goto out;
 978                 }
 979         }
 980         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 981 out:
 982         if (insert) {
 983                 path->keep_locks = 0;
 984                 path->search_for_extension = 0;
 985                 btrfs_unlock_up_safe(path, 1);
 986         }
 987         return ret;
 988 }
 989
 990 /*
 991  * helper to add new inline back ref
 992  */
 993 static noinline_for_stack
 994 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
 995                                  struct btrfs_path *path,
 996                                  struct btrfs_extent_inline_ref *iref,
 997                                  u64 parent, u64 root_objectid,
 998                                  u64 owner, u64 offset, int refs_to_add,
 999                                  struct btrfs_delayed_extent_op *extent_op)
1000 {
1001         struct extent_buffer *leaf;
1002         struct btrfs_extent_item *ei;
1003         unsigned long ptr;
1004         unsigned long end;
1005         unsigned long item_offset;
1006         u64 refs;
1007         int size;
1008         int type;
1009
1010         leaf = path->nodes[0];
1011         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1012         item_offset = (unsigned long)iref - (unsigned long)ei;
1013
1014         type = extent_ref_type(parent, owner);
1015         size = btrfs_extent_inline_ref_size(type);
1016
1017         btrfs_extend_item(trans, path, size);
1018
1019         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1020         refs = btrfs_extent_refs(leaf, ei);
1021         refs += refs_to_add;
1022         btrfs_set_extent_refs(leaf, ei, refs);
1023         if (extent_op)
1024                 __run_delayed_extent_op(extent_op, leaf, ei);
1025
1026         ptr = (unsigned long)ei + item_offset;
1027         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1028         if (ptr < end - size)
1029                 memmove_extent_buffer(leaf, ptr + size, ptr,
1030                                       end - size - ptr);
1031
1032         iref = (struct btrfs_extent_inline_ref *)ptr;
1033         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1034         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1035                 struct btrfs_extent_data_ref *dref;
1036                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1037                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1038                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1039                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1040                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1041         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1042                 struct btrfs_shared_data_ref *sref;
1043                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1044                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1045                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1046         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1047                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1048         } else {
1049                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1050         }
1051         btrfs_mark_buffer_dirty(trans, leaf);
1052 }
1053
1054 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1055                                  struct btrfs_path *path,
1056                                  struct btrfs_extent_inline_ref **ref_ret,
1057                                  u64 bytenr, u64 num_bytes, u64 parent,
1058                                  u64 root_objectid, u64 owner, u64 offset)
1059 {
1060         int ret;
1061
1062         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1063                                            num_bytes, parent, root_objectid,
1064                                            owner, offset, 0);
1065         if (ret != -ENOENT)
1066                 return ret;
1067
1068         btrfs_release_path(path);
1069         *ref_ret = NULL;
1070
1071         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1072                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1073                                             root_objectid);
1074         } else {
1075                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1076                                              root_objectid, owner, offset);
1077         }
1078         return ret;
1079 }
1080
1081 /*
1082  * helper to update/remove inline back ref
1083  */
1084 static noinline_for_stack int update_inline_extent_backref(
1085                                   struct btrfs_trans_handle *trans,
1086                                   struct btrfs_path *path,
1087                                   struct btrfs_extent_inline_ref *iref,
1088                                   int refs_to_mod,
1089                                   struct btrfs_delayed_extent_op *extent_op)
1090 {
1091         struct extent_buffer *leaf = path->nodes[0];
1092         struct btrfs_fs_info *fs_info = leaf->fs_info;
1093         struct btrfs_extent_item *ei;
1094         struct btrfs_extent_data_ref *dref = NULL;
1095         struct btrfs_shared_data_ref *sref = NULL;
1096         unsigned long ptr;
1097         unsigned long end;
1098         u32 item_size;
1099         int size;
1100         int type;
1101         u64 refs;
1102
1103         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1104         refs = btrfs_extent_refs(leaf, ei);
1105         if (unlikely(refs_to_mod < 0 && refs + refs_to_mod <= 0)) {
1106                 struct btrfs_key key;
1107                 u32 extent_size;
1108
1109                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1110                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1111                         extent_size = fs_info->nodesize;
1112                 else
1113                         extent_size = key.offset;
1114                 btrfs_print_leaf(leaf);
1115                 btrfs_err(fs_info,
1116         "invalid refs_to_mod for extent %llu num_bytes %u, has %d expect >= -%llu",
1117                           key.objectid, extent_size, refs_to_mod, refs);
1118                 return -EUCLEAN;
1119         }
1120         refs += refs_to_mod;
1121         btrfs_set_extent_refs(leaf, ei, refs);
1122         if (extent_op)
1123                 __run_delayed_extent_op(extent_op, leaf, ei);
1124
1125         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1126         /*
1127          * Function btrfs_get_extent_inline_ref_type() has already printed
1128          * error messages.
1129          */
1130         if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1131                 return -EUCLEAN;
1132
1133         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1134                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1135                 refs = btrfs_extent_data_ref_count(leaf, dref);
1136         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1137                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1138                 refs = btrfs_shared_data_ref_count(leaf, sref);
1139         } else {
1140                 refs = 1;
1141                 /*
1142                  * For tree blocks we can only drop one ref for it, and tree
1143                  * blocks should not have refs > 1.
1144                  *
1145                  * Furthermore if we're inserting a new inline backref, we
1146                  * won't reach this path either. That would be
1147                  * setup_inline_extent_backref().
1148                  */
1149                 if (unlikely(refs_to_mod != -1)) {
1150                         struct btrfs_key key;
1151
1152                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1153
1154                         btrfs_print_leaf(leaf);
1155                         btrfs_err(fs_info,
1156                         "invalid refs_to_mod for tree block %llu, has %d expect -1",
1157                                   key.objectid, refs_to_mod);
1158                         return -EUCLEAN;
1159                 }
1160         }
1161
1162         if (unlikely(refs_to_mod < 0 && refs < -refs_to_mod)) {
1163                 struct btrfs_key key;
1164                 u32 extent_size;
1165
1166                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1167                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1168                         extent_size = fs_info->nodesize;
1169                 else
1170                         extent_size = key.offset;
1171                 btrfs_print_leaf(leaf);
1172                 btrfs_err(fs_info,
1173 "invalid refs_to_mod for backref entry, iref %lu extent %llu num_bytes %u, has %d expect >= -%llu",
1174                           (unsigned long)iref, key.objectid, extent_size,
1175                           refs_to_mod, refs);
1176                 return -EUCLEAN;
1177         }
1178         refs += refs_to_mod;
1179
1180         if (refs > 0) {
1181                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1182                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1183                 else
1184                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1185         } else {
1186                 size =  btrfs_extent_inline_ref_size(type);
1187                 item_size = btrfs_item_size(leaf, path->slots[0]);
1188                 ptr = (unsigned long)iref;
1189                 end = (unsigned long)ei + item_size;
1190                 if (ptr + size < end)
1191                         memmove_extent_buffer(leaf, ptr, ptr + size,
1192                                               end - ptr - size);
1193                 item_size -= size;
1194                 btrfs_truncate_item(trans, path, item_size, 1);
1195         }
1196         btrfs_mark_buffer_dirty(trans, leaf);
1197         return 0;
1198 }
1199
1200 static noinline_for_stack
1201 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1202                                  struct btrfs_path *path,
1203                                  u64 bytenr, u64 num_bytes, u64 parent,
1204                                  u64 root_objectid, u64 owner,
1205                                  u64 offset, int refs_to_add,
1206                                  struct btrfs_delayed_extent_op *extent_op)
1207 {
1208         struct btrfs_extent_inline_ref *iref;
1209         int ret;
1210
1211         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1212                                            num_bytes, parent, root_objectid,
1213                                            owner, offset, 1);
1214         if (ret == 0) {
1215                 /*
1216                  * We're adding refs to a tree block we already own, this
1217                  * should not happen at all.
1218                  */
1219                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1220                         btrfs_print_leaf(path->nodes[0]);
1221                         btrfs_crit(trans->fs_info,
1222 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu slot %u",
1223                                    bytenr, num_bytes, root_objectid, path->slots[0]);
1224                         return -EUCLEAN;
1225                 }
1226                 ret = update_inline_extent_backref(trans, path, iref,
1227                                                    refs_to_add, extent_op);
1228         } else if (ret == -ENOENT) {
1229                 setup_inline_extent_backref(trans, path, iref, parent,
1230                                             root_objectid, owner, offset,
1231                                             refs_to_add, extent_op);
1232                 ret = 0;
1233         }
1234         return ret;
1235 }
1236
1237 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1238                                  struct btrfs_root *root,
1239                                  struct btrfs_path *path,
1240                                  struct btrfs_extent_inline_ref *iref,
1241                                  int refs_to_drop, int is_data)
1242 {
1243         int ret = 0;
1244
1245         BUG_ON(!is_data && refs_to_drop != 1);
1246         if (iref)
1247                 ret = update_inline_extent_backref(trans, path, iref,
1248                                                    -refs_to_drop, NULL);
1249         else if (is_data)
1250                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1251         else
1252                 ret = btrfs_del_item(trans, root, path);
1253         return ret;
1254 }
1255
1256 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1257                                u64 *discarded_bytes)
1258 {
1259         int j, ret = 0;
1260         u64 bytes_left, end;
1261         u64 aligned_start = ALIGN(start, 1 << SECTOR_SHIFT);
1262
1263         /* Adjust the range to be aligned to 512B sectors if necessary. */
1264         if (start != aligned_start) {
1265                 len -= aligned_start - start;
1266                 len = round_down(len, 1 << SECTOR_SHIFT);
1267                 start = aligned_start;
1268         }
1269
1270         *discarded_bytes = 0;
1271
1272         if (!len)
1273                 return 0;
1274
1275         end = start + len;
1276         bytes_left = len;
1277
1278         /* Skip any superblocks on this device. */
1279         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1280                 u64 sb_start = btrfs_sb_offset(j);
1281                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1282                 u64 size = sb_start - start;
1283
1284                 if (!in_range(sb_start, start, bytes_left) &&
1285                     !in_range(sb_end, start, bytes_left) &&
1286                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1287                         continue;
1288
1289                 /*
1290                  * Superblock spans beginning of range.  Adjust start and
1291                  * try again.
1292                  */
1293                 if (sb_start <= start) {
1294                         start += sb_end - start;
1295                         if (start > end) {
1296                                 bytes_left = 0;
1297                                 break;
1298                         }
1299                         bytes_left = end - start;
1300                         continue;
1301                 }
1302
1303                 if (size) {
1304                         ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1305                                                    size >> SECTOR_SHIFT,
1306                                                    GFP_NOFS);
1307                         if (!ret)
1308                                 *discarded_bytes += size;
1309                         else if (ret != -EOPNOTSUPP)
1310                                 return ret;
1311                 }
1312
1313                 start = sb_end;
1314                 if (start > end) {
1315                         bytes_left = 0;
1316                         break;
1317                 }
1318                 bytes_left = end - start;
1319         }
1320
1321         if (bytes_left) {
1322                 ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1323                                            bytes_left >> SECTOR_SHIFT,
1324                                            GFP_NOFS);
1325                 if (!ret)
1326                         *discarded_bytes += bytes_left;
1327         }
1328         return ret;
1329 }
1330
1331 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1332 {
1333         struct btrfs_device *dev = stripe->dev;
1334         struct btrfs_fs_info *fs_info = dev->fs_info;
1335         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1336         u64 phys = stripe->physical;
1337         u64 len = stripe->length;
1338         u64 discarded = 0;
1339         int ret = 0;
1340
1341         /* Zone reset on a zoned filesystem */
1342         if (btrfs_can_zone_reset(dev, phys, len)) {
1343                 u64 src_disc;
1344
1345                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1346                 if (ret)
1347                         goto out;
1348
1349                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1350                     dev != dev_replace->srcdev)
1351                         goto out;
1352
1353                 src_disc = discarded;
1354
1355                 /* Send to replace target as well */
1356                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1357                                               &discarded);
1358                 discarded += src_disc;
1359         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1360                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1361         } else {
1362                 ret = 0;
1363                 *bytes = 0;
1364         }
1365
1366 out:
1367         *bytes = discarded;
1368         return ret;
1369 }
1370
1371 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1372                          u64 num_bytes, u64 *actual_bytes)
1373 {
1374         int ret = 0;
1375         u64 discarded_bytes = 0;
1376         u64 end = bytenr + num_bytes;
1377         u64 cur = bytenr;
1378
1379         /*
1380          * Avoid races with device replace and make sure the devices in the
1381          * stripes don't go away while we are discarding.
1382          */
1383         btrfs_bio_counter_inc_blocked(fs_info);
1384         while (cur < end) {
1385                 struct btrfs_discard_stripe *stripes;
1386                 unsigned int num_stripes;
1387                 int i;
1388
1389                 num_bytes = end - cur;
1390                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1391                 if (IS_ERR(stripes)) {
1392                         ret = PTR_ERR(stripes);
1393                         if (ret == -EOPNOTSUPP)
1394                                 ret = 0;
1395                         break;
1396                 }
1397
1398                 for (i = 0; i < num_stripes; i++) {
1399                         struct btrfs_discard_stripe *stripe = stripes + i;
1400                         u64 bytes;
1401
1402                         if (!stripe->dev->bdev) {
1403                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1404                                 continue;
1405                         }
1406
1407                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1408                                         &stripe->dev->dev_state))
1409                                 continue;
1410
1411                         ret = do_discard_extent(stripe, &bytes);
1412                         if (ret) {
1413                                 /*
1414                                  * Keep going if discard is not supported by the
1415                                  * device.
1416                                  */
1417                                 if (ret != -EOPNOTSUPP)
1418                                         break;
1419                                 ret = 0;
1420                         } else {
1421                                 discarded_bytes += bytes;
1422                         }
1423                 }
1424                 kfree(stripes);
1425                 if (ret)
1426                         break;
1427                 cur += num_bytes;
1428         }
1429         btrfs_bio_counter_dec(fs_info);
1430         if (actual_bytes)
1431                 *actual_bytes = discarded_bytes;
1432         return ret;
1433 }
1434
1435 /* Can return -ENOMEM */
1436 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1437                          struct btrfs_ref *generic_ref)
1438 {
1439         struct btrfs_fs_info *fs_info = trans->fs_info;
1440         int ret;
1441
1442         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1443                generic_ref->action);
1444         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1445                generic_ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID);
1446
1447         if (generic_ref->type == BTRFS_REF_METADATA)
1448                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1449         else
1450                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1451
1452         btrfs_ref_tree_mod(fs_info, generic_ref);
1453
1454         return ret;
1455 }
1456
1457 /*
1458  * Insert backreference for a given extent.
1459  *
1460  * The counterpart is in __btrfs_free_extent(), with examples and more details
1461  * how it works.
1462  *
1463  * @trans:          Handle of transaction
1464  *
1465  * @node:           The delayed ref node used to get the bytenr/length for
1466  *                  extent whose references are incremented.
1467  *
1468  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1469  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1470  *                  bytenr of the parent block. Since new extents are always
1471  *                  created with indirect references, this will only be the case
1472  *                  when relocating a shared extent. In that case, root_objectid
1473  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1474  *                  be 0
1475  *
1476  * @root_objectid:  The id of the root where this modification has originated,
1477  *                  this can be either one of the well-known metadata trees or
1478  *                  the subvolume id which references this extent.
1479  *
1480  * @owner:          For data extents it is the inode number of the owning file.
1481  *                  For metadata extents this parameter holds the level in the
1482  *                  tree of the extent.
1483  *
1484  * @offset:         For metadata extents the offset is ignored and is currently
1485  *                  always passed as 0. For data extents it is the fileoffset
1486  *                  this extent belongs to.
1487  *
1488  * @extent_op       Pointer to a structure, holding information necessary when
1489  *                  updating a tree block's flags
1490  *
1491  */
1492 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1493                                   struct btrfs_delayed_ref_node *node,
1494                                   u64 parent, u64 root_objectid,
1495                                   u64 owner, u64 offset,
1496                                   struct btrfs_delayed_extent_op *extent_op)
1497 {
1498         struct btrfs_path *path;
1499         struct extent_buffer *leaf;
1500         struct btrfs_extent_item *item;
1501         struct btrfs_key key;
1502         u64 bytenr = node->bytenr;
1503         u64 num_bytes = node->num_bytes;
1504         u64 refs;
1505         int refs_to_add = node->ref_mod;
1506         int ret;
1507
1508         path = btrfs_alloc_path();
1509         if (!path)
1510                 return -ENOMEM;
1511
1512         /* this will setup the path even if it fails to insert the back ref */
1513         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1514                                            parent, root_objectid, owner,
1515                                            offset, refs_to_add, extent_op);
1516         if ((ret < 0 && ret != -EAGAIN) || !ret)
1517                 goto out;
1518
1519         /*
1520          * Ok we had -EAGAIN which means we didn't have space to insert and
1521          * inline extent ref, so just update the reference count and add a
1522          * normal backref.
1523          */
1524         leaf = path->nodes[0];
1525         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1526         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1527         refs = btrfs_extent_refs(leaf, item);
1528         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1529         if (extent_op)
1530                 __run_delayed_extent_op(extent_op, leaf, item);
1531
1532         btrfs_mark_buffer_dirty(trans, leaf);
1533         btrfs_release_path(path);
1534
1535         /* now insert the actual backref */
1536         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1537                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1538                                             root_objectid);
1539         else
1540                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1541                                              root_objectid, owner, offset,
1542                                              refs_to_add);
1543
1544         if (ret)
1545                 btrfs_abort_transaction(trans, ret);
1546 out:
1547         btrfs_free_path(path);
1548         return ret;
1549 }
1550
1551 static void free_head_ref_squota_rsv(struct btrfs_fs_info *fs_info,
1552                                      struct btrfs_delayed_ref_head *href)
1553 {
1554         u64 root = href->owning_root;
1555
1556         /*
1557          * Don't check must_insert_reserved, as this is called from contexts
1558          * where it has already been unset.
1559          */
1560         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE ||
1561             !href->is_data || !is_fstree(root))
1562                 return;
1563
1564         btrfs_qgroup_free_refroot(fs_info, root, href->reserved_bytes,
1565                                   BTRFS_QGROUP_RSV_DATA);
1566 }
1567
1568 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1569                                 struct btrfs_delayed_ref_head *href,
1570                                 struct btrfs_delayed_ref_node *node,
1571                                 struct btrfs_delayed_extent_op *extent_op,
1572                                 bool insert_reserved)
1573 {
1574         int ret = 0;
1575         struct btrfs_delayed_data_ref *ref;
1576         u64 parent = 0;
1577         u64 flags = 0;
1578
1579         ref = btrfs_delayed_node_to_data_ref(node);
1580         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1581
1582         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1583                 parent = ref->parent;
1584
1585         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1586                 struct btrfs_key key;
1587                 struct btrfs_squota_delta delta = {
1588                         .root = href->owning_root,
1589                         .num_bytes = node->num_bytes,
1590                         .is_data = true,
1591                         .is_inc = true,
1592                         .generation = trans->transid,
1593                 };
1594
1595                 if (extent_op)
1596                         flags |= extent_op->flags_to_set;
1597
1598                 key.objectid = node->bytenr;
1599                 key.type = BTRFS_EXTENT_ITEM_KEY;
1600                 key.offset = node->num_bytes;
1601
1602                 ret = alloc_reserved_file_extent(trans, parent, ref->root,
1603                                                  flags, ref->objectid,
1604                                                  ref->offset, &key,
1605                                                  node->ref_mod, href->owning_root);
1606                 free_head_ref_squota_rsv(trans->fs_info, href);
1607                 if (!ret)
1608                         ret = btrfs_record_squota_delta(trans->fs_info, &delta);
1609         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1610                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref->root,
1611                                              ref->objectid, ref->offset,
1612                                              extent_op);
1613         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1614                 ret = __btrfs_free_extent(trans, href, node, parent,
1615                                           ref->root, ref->objectid,
1616                                           ref->offset, extent_op);
1617         } else {
1618                 BUG();
1619         }
1620         return ret;
1621 }
1622
1623 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1624                                     struct extent_buffer *leaf,
1625                                     struct btrfs_extent_item *ei)
1626 {
1627         u64 flags = btrfs_extent_flags(leaf, ei);
1628         if (extent_op->update_flags) {
1629                 flags |= extent_op->flags_to_set;
1630                 btrfs_set_extent_flags(leaf, ei, flags);
1631         }
1632
1633         if (extent_op->update_key) {
1634                 struct btrfs_tree_block_info *bi;
1635                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1636                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1637                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1638         }
1639 }
1640
1641 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1642                                  struct btrfs_delayed_ref_head *head,
1643                                  struct btrfs_delayed_extent_op *extent_op)
1644 {
1645         struct btrfs_fs_info *fs_info = trans->fs_info;
1646         struct btrfs_root *root;
1647         struct btrfs_key key;
1648         struct btrfs_path *path;
1649         struct btrfs_extent_item *ei;
1650         struct extent_buffer *leaf;
1651         u32 item_size;
1652         int ret;
1653         int metadata = 1;
1654
1655         if (TRANS_ABORTED(trans))
1656                 return 0;
1657
1658         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1659                 metadata = 0;
1660
1661         path = btrfs_alloc_path();
1662         if (!path)
1663                 return -ENOMEM;
1664
1665         key.objectid = head->bytenr;
1666
1667         if (metadata) {
1668                 key.type = BTRFS_METADATA_ITEM_KEY;
1669                 key.offset = extent_op->level;
1670         } else {
1671                 key.type = BTRFS_EXTENT_ITEM_KEY;
1672                 key.offset = head->num_bytes;
1673         }
1674
1675         root = btrfs_extent_root(fs_info, key.objectid);
1676 again:
1677         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1678         if (ret < 0) {
1679                 goto out;
1680         } else if (ret > 0) {
1681                 if (metadata) {
1682                         if (path->slots[0] > 0) {
1683                                 path->slots[0]--;
1684                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1685                                                       path->slots[0]);
1686                                 if (key.objectid == head->bytenr &&
1687                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1688                                     key.offset == head->num_bytes)
1689                                         ret = 0;
1690                         }
1691                         if (ret > 0) {
1692                                 btrfs_release_path(path);
1693                                 metadata = 0;
1694
1695                                 key.objectid = head->bytenr;
1696                                 key.offset = head->num_bytes;
1697                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1698                                 goto again;
1699                         }
1700                 } else {
1701                         ret = -EUCLEAN;
1702                         btrfs_err(fs_info,
1703                   "missing extent item for extent %llu num_bytes %llu level %d",
1704                                   head->bytenr, head->num_bytes, extent_op->level);
1705                         goto out;
1706                 }
1707         }
1708
1709         leaf = path->nodes[0];
1710         item_size = btrfs_item_size(leaf, path->slots[0]);
1711
1712         if (unlikely(item_size < sizeof(*ei))) {
1713                 ret = -EUCLEAN;
1714                 btrfs_err(fs_info,
1715                           "unexpected extent item size, has %u expect >= %zu",
1716                           item_size, sizeof(*ei));
1717                 btrfs_abort_transaction(trans, ret);
1718                 goto out;
1719         }
1720
1721         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1722         __run_delayed_extent_op(extent_op, leaf, ei);
1723
1724         btrfs_mark_buffer_dirty(trans, leaf);
1725 out:
1726         btrfs_free_path(path);
1727         return ret;
1728 }
1729
1730 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1731                                 struct btrfs_delayed_ref_head *href,
1732                                 struct btrfs_delayed_ref_node *node,
1733                                 struct btrfs_delayed_extent_op *extent_op,
1734                                 bool insert_reserved)
1735 {
1736         int ret = 0;
1737         struct btrfs_fs_info *fs_info = trans->fs_info;
1738         struct btrfs_delayed_tree_ref *ref;
1739         u64 parent = 0;
1740         u64 ref_root = 0;
1741
1742         ref = btrfs_delayed_node_to_tree_ref(node);
1743         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1744
1745         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1746                 parent = ref->parent;
1747         ref_root = ref->root;
1748
1749         if (unlikely(node->ref_mod != 1)) {
1750                 btrfs_err(trans->fs_info,
1751         "btree block %llu has %d references rather than 1: action %d ref_root %llu parent %llu",
1752                           node->bytenr, node->ref_mod, node->action, ref_root,
1753                           parent);
1754                 return -EUCLEAN;
1755         }
1756         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1757                 struct btrfs_squota_delta delta = {
1758                         .root = href->owning_root,
1759                         .num_bytes = fs_info->nodesize,
1760                         .is_data = false,
1761                         .is_inc = true,
1762                         .generation = trans->transid,
1763                 };
1764
1765                 BUG_ON(!extent_op || !extent_op->update_flags);
1766                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1767                 if (!ret)
1768                         btrfs_record_squota_delta(fs_info, &delta);
1769         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1770                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1771                                              ref->level, 0, extent_op);
1772         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1773                 ret = __btrfs_free_extent(trans, href, node, parent, ref_root,
1774                                           ref->level, 0, extent_op);
1775         } else {
1776                 BUG();
1777         }
1778         return ret;
1779 }
1780
1781 /* helper function to actually process a single delayed ref entry */
1782 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1783                                struct btrfs_delayed_ref_head *href,
1784                                struct btrfs_delayed_ref_node *node,
1785                                struct btrfs_delayed_extent_op *extent_op,
1786                                bool insert_reserved)
1787 {
1788         int ret = 0;
1789
1790         if (TRANS_ABORTED(trans)) {
1791                 if (insert_reserved) {
1792                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1793                         free_head_ref_squota_rsv(trans->fs_info, href);
1794                 }
1795                 return 0;
1796         }
1797
1798         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1799             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1800                 ret = run_delayed_tree_ref(trans, href, node, extent_op,
1801                                            insert_reserved);
1802         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1803                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1804                 ret = run_delayed_data_ref(trans, href, node, extent_op,
1805                                            insert_reserved);
1806         else if (node->type == BTRFS_EXTENT_OWNER_REF_KEY)
1807                 ret = 0;
1808         else
1809                 BUG();
1810         if (ret && insert_reserved)
1811                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1812         if (ret < 0)
1813                 btrfs_err(trans->fs_info,
1814 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1815                           node->bytenr, node->num_bytes, node->type,
1816                           node->action, node->ref_mod, ret);
1817         return ret;
1818 }
1819
1820 static inline struct btrfs_delayed_ref_node *
1821 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1822 {
1823         struct btrfs_delayed_ref_node *ref;
1824
1825         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1826                 return NULL;
1827
1828         /*
1829          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1830          * This is to prevent a ref count from going down to zero, which deletes
1831          * the extent item from the extent tree, when there still are references
1832          * to add, which would fail because they would not find the extent item.
1833          */
1834         if (!list_empty(&head->ref_add_list))
1835                 return list_first_entry(&head->ref_add_list,
1836                                 struct btrfs_delayed_ref_node, add_list);
1837
1838         ref = rb_entry(rb_first_cached(&head->ref_tree),
1839                        struct btrfs_delayed_ref_node, ref_node);
1840         ASSERT(list_empty(&ref->add_list));
1841         return ref;
1842 }
1843
1844 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1845                                       struct btrfs_delayed_ref_head *head)
1846 {
1847         spin_lock(&delayed_refs->lock);
1848         head->processing = false;
1849         delayed_refs->num_heads_ready++;
1850         spin_unlock(&delayed_refs->lock);
1851         btrfs_delayed_ref_unlock(head);
1852 }
1853
1854 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1855                                 struct btrfs_delayed_ref_head *head)
1856 {
1857         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1858
1859         if (!extent_op)
1860                 return NULL;
1861
1862         if (head->must_insert_reserved) {
1863                 head->extent_op = NULL;
1864                 btrfs_free_delayed_extent_op(extent_op);
1865                 return NULL;
1866         }
1867         return extent_op;
1868 }
1869
1870 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1871                                      struct btrfs_delayed_ref_head *head)
1872 {
1873         struct btrfs_delayed_extent_op *extent_op;
1874         int ret;
1875
1876         extent_op = cleanup_extent_op(head);
1877         if (!extent_op)
1878                 return 0;
1879         head->extent_op = NULL;
1880         spin_unlock(&head->lock);
1881         ret = run_delayed_extent_op(trans, head, extent_op);
1882         btrfs_free_delayed_extent_op(extent_op);
1883         return ret ? ret : 1;
1884 }
1885
1886 u64 btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1887                                   struct btrfs_delayed_ref_root *delayed_refs,
1888                                   struct btrfs_delayed_ref_head *head)
1889 {
1890         u64 ret = 0;
1891
1892         /*
1893          * We had csum deletions accounted for in our delayed refs rsv, we need
1894          * to drop the csum leaves for this update from our delayed_refs_rsv.
1895          */
1896         if (head->total_ref_mod < 0 && head->is_data) {
1897                 int nr_csums;
1898
1899                 spin_lock(&delayed_refs->lock);
1900                 delayed_refs->pending_csums -= head->num_bytes;
1901                 spin_unlock(&delayed_refs->lock);
1902                 nr_csums = btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1903
1904                 btrfs_delayed_refs_rsv_release(fs_info, 0, nr_csums);
1905
1906                 ret = btrfs_calc_delayed_ref_csum_bytes(fs_info, nr_csums);
1907         }
1908         /* must_insert_reserved can be set only if we didn't run the head ref. */
1909         if (head->must_insert_reserved)
1910                 free_head_ref_squota_rsv(fs_info, head);
1911
1912         return ret;
1913 }
1914
1915 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1916                             struct btrfs_delayed_ref_head *head,
1917                             u64 *bytes_released)
1918 {
1919
1920         struct btrfs_fs_info *fs_info = trans->fs_info;
1921         struct btrfs_delayed_ref_root *delayed_refs;
1922         int ret;
1923
1924         delayed_refs = &trans->transaction->delayed_refs;
1925
1926         ret = run_and_cleanup_extent_op(trans, head);
1927         if (ret < 0) {
1928                 unselect_delayed_ref_head(delayed_refs, head);
1929                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1930                 return ret;
1931         } else if (ret) {
1932                 return ret;
1933         }
1934
1935         /*
1936          * Need to drop our head ref lock and re-acquire the delayed ref lock
1937          * and then re-check to make sure nobody got added.
1938          */
1939         spin_unlock(&head->lock);
1940         spin_lock(&delayed_refs->lock);
1941         spin_lock(&head->lock);
1942         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1943                 spin_unlock(&head->lock);
1944                 spin_unlock(&delayed_refs->lock);
1945                 return 1;
1946         }
1947         btrfs_delete_ref_head(delayed_refs, head);
1948         spin_unlock(&head->lock);
1949         spin_unlock(&delayed_refs->lock);
1950
1951         if (head->must_insert_reserved) {
1952                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1953                 if (head->is_data) {
1954                         struct btrfs_root *csum_root;
1955
1956                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1957                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1958                                               head->num_bytes);
1959                 }
1960         }
1961
1962         *bytes_released += btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1963
1964         trace_run_delayed_ref_head(fs_info, head, 0);
1965         btrfs_delayed_ref_unlock(head);
1966         btrfs_put_delayed_ref_head(head);
1967         return ret;
1968 }
1969
1970 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1971                                         struct btrfs_trans_handle *trans)
1972 {
1973         struct btrfs_delayed_ref_root *delayed_refs =
1974                 &trans->transaction->delayed_refs;
1975         struct btrfs_delayed_ref_head *head = NULL;
1976         int ret;
1977
1978         spin_lock(&delayed_refs->lock);
1979         head = btrfs_select_ref_head(delayed_refs);
1980         if (!head) {
1981                 spin_unlock(&delayed_refs->lock);
1982                 return head;
1983         }
1984
1985         /*
1986          * Grab the lock that says we are going to process all the refs for
1987          * this head
1988          */
1989         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1990         spin_unlock(&delayed_refs->lock);
1991
1992         /*
1993          * We may have dropped the spin lock to get the head mutex lock, and
1994          * that might have given someone else time to free the head.  If that's
1995          * true, it has been removed from our list and we can move on.
1996          */
1997         if (ret == -EAGAIN)
1998                 head = ERR_PTR(-EAGAIN);
1999
2000         return head;
2001 }
2002
2003 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2004                                            struct btrfs_delayed_ref_head *locked_ref,
2005                                            u64 *bytes_released)
2006 {
2007         struct btrfs_fs_info *fs_info = trans->fs_info;
2008         struct btrfs_delayed_ref_root *delayed_refs;
2009         struct btrfs_delayed_extent_op *extent_op;
2010         struct btrfs_delayed_ref_node *ref;
2011         bool must_insert_reserved;
2012         int ret;
2013
2014         delayed_refs = &trans->transaction->delayed_refs;
2015
2016         lockdep_assert_held(&locked_ref->mutex);
2017         lockdep_assert_held(&locked_ref->lock);
2018
2019         while ((ref = select_delayed_ref(locked_ref))) {
2020                 if (ref->seq &&
2021                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2022                         spin_unlock(&locked_ref->lock);
2023                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2024                         return -EAGAIN;
2025                 }
2026
2027                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2028                 RB_CLEAR_NODE(&ref->ref_node);
2029                 if (!list_empty(&ref->add_list))
2030                         list_del(&ref->add_list);
2031                 /*
2032                  * When we play the delayed ref, also correct the ref_mod on
2033                  * head
2034                  */
2035                 switch (ref->action) {
2036                 case BTRFS_ADD_DELAYED_REF:
2037                 case BTRFS_ADD_DELAYED_EXTENT:
2038                         locked_ref->ref_mod -= ref->ref_mod;
2039                         break;
2040                 case BTRFS_DROP_DELAYED_REF:
2041                         locked_ref->ref_mod += ref->ref_mod;
2042                         break;
2043                 default:
2044                         WARN_ON(1);
2045                 }
2046                 atomic_dec(&delayed_refs->num_entries);
2047
2048                 /*
2049                  * Record the must_insert_reserved flag before we drop the
2050                  * spin lock.
2051                  */
2052                 must_insert_reserved = locked_ref->must_insert_reserved;
2053                 /*
2054                  * Unsetting this on the head ref relinquishes ownership of
2055                  * the rsv_bytes, so it is critical that every possible code
2056                  * path from here forward frees all reserves including qgroup
2057                  * reserve.
2058                  */
2059                 locked_ref->must_insert_reserved = false;
2060
2061                 extent_op = locked_ref->extent_op;
2062                 locked_ref->extent_op = NULL;
2063                 spin_unlock(&locked_ref->lock);
2064
2065                 ret = run_one_delayed_ref(trans, locked_ref, ref, extent_op,
2066                                           must_insert_reserved);
2067                 btrfs_delayed_refs_rsv_release(fs_info, 1, 0);
2068                 *bytes_released += btrfs_calc_delayed_ref_bytes(fs_info, 1);
2069
2070                 btrfs_free_delayed_extent_op(extent_op);
2071                 if (ret) {
2072                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2073                         btrfs_put_delayed_ref(ref);
2074                         return ret;
2075                 }
2076
2077                 btrfs_put_delayed_ref(ref);
2078                 cond_resched();
2079
2080                 spin_lock(&locked_ref->lock);
2081                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2082         }
2083
2084         return 0;
2085 }
2086
2087 /*
2088  * Returns 0 on success or if called with an already aborted transaction.
2089  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2090  */
2091 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2092                                              u64 min_bytes)
2093 {
2094         struct btrfs_fs_info *fs_info = trans->fs_info;
2095         struct btrfs_delayed_ref_root *delayed_refs;
2096         struct btrfs_delayed_ref_head *locked_ref = NULL;
2097         int ret;
2098         unsigned long count = 0;
2099         unsigned long max_count = 0;
2100         u64 bytes_processed = 0;
2101
2102         delayed_refs = &trans->transaction->delayed_refs;
2103         if (min_bytes == 0) {
2104                 max_count = delayed_refs->num_heads_ready;
2105                 min_bytes = U64_MAX;
2106         }
2107
2108         do {
2109                 if (!locked_ref) {
2110                         locked_ref = btrfs_obtain_ref_head(trans);
2111                         if (IS_ERR_OR_NULL(locked_ref)) {
2112                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2113                                         continue;
2114                                 } else {
2115                                         break;
2116                                 }
2117                         }
2118                         count++;
2119                 }
2120                 /*
2121                  * We need to try and merge add/drops of the same ref since we
2122                  * can run into issues with relocate dropping the implicit ref
2123                  * and then it being added back again before the drop can
2124                  * finish.  If we merged anything we need to re-loop so we can
2125                  * get a good ref.
2126                  * Or we can get node references of the same type that weren't
2127                  * merged when created due to bumps in the tree mod seq, and
2128                  * we need to merge them to prevent adding an inline extent
2129                  * backref before dropping it (triggering a BUG_ON at
2130                  * insert_inline_extent_backref()).
2131                  */
2132                 spin_lock(&locked_ref->lock);
2133                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2134
2135                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref, &bytes_processed);
2136                 if (ret < 0 && ret != -EAGAIN) {
2137                         /*
2138                          * Error, btrfs_run_delayed_refs_for_head already
2139                          * unlocked everything so just bail out
2140                          */
2141                         return ret;
2142                 } else if (!ret) {
2143                         /*
2144                          * Success, perform the usual cleanup of a processed
2145                          * head
2146                          */
2147                         ret = cleanup_ref_head(trans, locked_ref, &bytes_processed);
2148                         if (ret > 0 ) {
2149                                 /* We dropped our lock, we need to loop. */
2150                                 ret = 0;
2151                                 continue;
2152                         } else if (ret) {
2153                                 return ret;
2154                         }
2155                 }
2156
2157                 /*
2158                  * Either success case or btrfs_run_delayed_refs_for_head
2159                  * returned -EAGAIN, meaning we need to select another head
2160                  */
2161
2162                 locked_ref = NULL;
2163                 cond_resched();
2164         } while ((min_bytes != U64_MAX && bytes_processed < min_bytes) ||
2165                  (max_count > 0 && count < max_count) ||
2166                  locked_ref);
2167
2168         return 0;
2169 }
2170
2171 #ifdef SCRAMBLE_DELAYED_REFS
2172 /*
2173  * Normally delayed refs get processed in ascending bytenr order. This
2174  * correlates in most cases to the order added. To expose dependencies on this
2175  * order, we start to process the tree in the middle instead of the beginning
2176  */
2177 static u64 find_middle(struct rb_root *root)
2178 {
2179         struct rb_node *n = root->rb_node;
2180         struct btrfs_delayed_ref_node *entry;
2181         int alt = 1;
2182         u64 middle;
2183         u64 first = 0, last = 0;
2184
2185         n = rb_first(root);
2186         if (n) {
2187                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2188                 first = entry->bytenr;
2189         }
2190         n = rb_last(root);
2191         if (n) {
2192                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2193                 last = entry->bytenr;
2194         }
2195         n = root->rb_node;
2196
2197         while (n) {
2198                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2199                 WARN_ON(!entry->in_tree);
2200
2201                 middle = entry->bytenr;
2202
2203                 if (alt)
2204                         n = n->rb_left;
2205                 else
2206                         n = n->rb_right;
2207
2208                 alt = 1 - alt;
2209         }
2210         return middle;
2211 }
2212 #endif
2213
2214 /*
2215  * Start processing the delayed reference count updates and extent insertions
2216  * we have queued up so far.
2217  *
2218  * @trans:      Transaction handle.
2219  * @min_bytes:  How many bytes of delayed references to process. After this
2220  *              many bytes we stop processing delayed references if there are
2221  *              any more. If 0 it means to run all existing delayed references,
2222  *              but not new ones added after running all existing ones.
2223  *              Use (u64)-1 (U64_MAX) to run all existing delayed references
2224  *              plus any new ones that are added.
2225  *
2226  * Returns 0 on success or if called with an aborted transaction
2227  * Returns <0 on error and aborts the transaction
2228  */
2229 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, u64 min_bytes)
2230 {
2231         struct btrfs_fs_info *fs_info = trans->fs_info;
2232         struct btrfs_delayed_ref_root *delayed_refs;
2233         int ret;
2234
2235         /* We'll clean this up in btrfs_cleanup_transaction */
2236         if (TRANS_ABORTED(trans))
2237                 return 0;
2238
2239         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2240                 return 0;
2241
2242         delayed_refs = &trans->transaction->delayed_refs;
2243 again:
2244 #ifdef SCRAMBLE_DELAYED_REFS
2245         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2246 #endif
2247         ret = __btrfs_run_delayed_refs(trans, min_bytes);
2248         if (ret < 0) {
2249                 btrfs_abort_transaction(trans, ret);
2250                 return ret;
2251         }
2252
2253         if (min_bytes == U64_MAX) {
2254                 btrfs_create_pending_block_groups(trans);
2255
2256                 spin_lock(&delayed_refs->lock);
2257                 if (RB_EMPTY_ROOT(&delayed_refs->href_root.rb_root)) {
2258                         spin_unlock(&delayed_refs->lock);
2259                         return 0;
2260                 }
2261                 spin_unlock(&delayed_refs->lock);
2262
2263                 cond_resched();
2264                 goto again;
2265         }
2266
2267         return 0;
2268 }
2269
2270 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2271                                 struct extent_buffer *eb, u64 flags)
2272 {
2273         struct btrfs_delayed_extent_op *extent_op;
2274         int level = btrfs_header_level(eb);
2275         int ret;
2276
2277         extent_op = btrfs_alloc_delayed_extent_op();
2278         if (!extent_op)
2279                 return -ENOMEM;
2280
2281         extent_op->flags_to_set = flags;
2282         extent_op->update_flags = true;
2283         extent_op->update_key = false;
2284         extent_op->level = level;
2285
2286         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2287         if (ret)
2288                 btrfs_free_delayed_extent_op(extent_op);
2289         return ret;
2290 }
2291
2292 static noinline int check_delayed_ref(struct btrfs_root *root,
2293                                       struct btrfs_path *path,
2294                                       u64 objectid, u64 offset, u64 bytenr)
2295 {
2296         struct btrfs_delayed_ref_head *head;
2297         struct btrfs_delayed_ref_node *ref;
2298         struct btrfs_delayed_data_ref *data_ref;
2299         struct btrfs_delayed_ref_root *delayed_refs;
2300         struct btrfs_transaction *cur_trans;
2301         struct rb_node *node;
2302         int ret = 0;
2303
2304         spin_lock(&root->fs_info->trans_lock);
2305         cur_trans = root->fs_info->running_transaction;
2306         if (cur_trans)
2307                 refcount_inc(&cur_trans->use_count);
2308         spin_unlock(&root->fs_info->trans_lock);
2309         if (!cur_trans)
2310                 return 0;
2311
2312         delayed_refs = &cur_trans->delayed_refs;
2313         spin_lock(&delayed_refs->lock);
2314         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2315         if (!head) {
2316                 spin_unlock(&delayed_refs->lock);
2317                 btrfs_put_transaction(cur_trans);
2318                 return 0;
2319         }
2320
2321         if (!mutex_trylock(&head->mutex)) {
2322                 if (path->nowait) {
2323                         spin_unlock(&delayed_refs->lock);
2324                         btrfs_put_transaction(cur_trans);
2325                         return -EAGAIN;
2326                 }
2327
2328                 refcount_inc(&head->refs);
2329                 spin_unlock(&delayed_refs->lock);
2330
2331                 btrfs_release_path(path);
2332
2333                 /*
2334                  * Mutex was contended, block until it's released and let
2335                  * caller try again
2336                  */
2337                 mutex_lock(&head->mutex);
2338                 mutex_unlock(&head->mutex);
2339                 btrfs_put_delayed_ref_head(head);
2340                 btrfs_put_transaction(cur_trans);
2341                 return -EAGAIN;
2342         }
2343         spin_unlock(&delayed_refs->lock);
2344
2345         spin_lock(&head->lock);
2346         /*
2347          * XXX: We should replace this with a proper search function in the
2348          * future.
2349          */
2350         for (node = rb_first_cached(&head->ref_tree); node;
2351              node = rb_next(node)) {
2352                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2353                 /* If it's a shared ref we know a cross reference exists */
2354                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2355                         ret = 1;
2356                         break;
2357                 }
2358
2359                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2360
2361                 /*
2362                  * If our ref doesn't match the one we're currently looking at
2363                  * then we have a cross reference.
2364                  */
2365                 if (data_ref->root != root->root_key.objectid ||
2366                     data_ref->objectid != objectid ||
2367                     data_ref->offset != offset) {
2368                         ret = 1;
2369                         break;
2370                 }
2371         }
2372         spin_unlock(&head->lock);
2373         mutex_unlock(&head->mutex);
2374         btrfs_put_transaction(cur_trans);
2375         return ret;
2376 }
2377
2378 static noinline int check_committed_ref(struct btrfs_root *root,
2379                                         struct btrfs_path *path,
2380                                         u64 objectid, u64 offset, u64 bytenr,
2381                                         bool strict)
2382 {
2383         struct btrfs_fs_info *fs_info = root->fs_info;
2384         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2385         struct extent_buffer *leaf;
2386         struct btrfs_extent_data_ref *ref;
2387         struct btrfs_extent_inline_ref *iref;
2388         struct btrfs_extent_item *ei;
2389         struct btrfs_key key;
2390         u32 item_size;
2391         u32 expected_size;
2392         int type;
2393         int ret;
2394
2395         key.objectid = bytenr;
2396         key.offset = (u64)-1;
2397         key.type = BTRFS_EXTENT_ITEM_KEY;
2398
2399         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2400         if (ret < 0)
2401                 goto out;
2402         BUG_ON(ret == 0); /* Corruption */
2403
2404         ret = -ENOENT;
2405         if (path->slots[0] == 0)
2406                 goto out;
2407
2408         path->slots[0]--;
2409         leaf = path->nodes[0];
2410         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2411
2412         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2413                 goto out;
2414
2415         ret = 1;
2416         item_size = btrfs_item_size(leaf, path->slots[0]);
2417         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2418         expected_size = sizeof(*ei) + btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY);
2419
2420         /* No inline refs; we need to bail before checking for owner ref. */
2421         if (item_size == sizeof(*ei))
2422                 goto out;
2423
2424         /* Check for an owner ref; skip over it to the real inline refs. */
2425         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2426         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2427         if (btrfs_fs_incompat(fs_info, SIMPLE_QUOTA) && type == BTRFS_EXTENT_OWNER_REF_KEY) {
2428                 expected_size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
2429                 iref = (struct btrfs_extent_inline_ref *)(iref + 1);
2430         }
2431
2432         /* If extent item has more than 1 inline ref then it's shared */
2433         if (item_size != expected_size)
2434                 goto out;
2435
2436         /*
2437          * If extent created before last snapshot => it's shared unless the
2438          * snapshot has been deleted. Use the heuristic if strict is false.
2439          */
2440         if (!strict &&
2441             (btrfs_extent_generation(leaf, ei) <=
2442              btrfs_root_last_snapshot(&root->root_item)))
2443                 goto out;
2444
2445         /* If this extent has SHARED_DATA_REF then it's shared */
2446         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2447         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2448                 goto out;
2449
2450         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2451         if (btrfs_extent_refs(leaf, ei) !=
2452             btrfs_extent_data_ref_count(leaf, ref) ||
2453             btrfs_extent_data_ref_root(leaf, ref) !=
2454             root->root_key.objectid ||
2455             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2456             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2457                 goto out;
2458
2459         ret = 0;
2460 out:
2461         return ret;
2462 }
2463
2464 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2465                           u64 bytenr, bool strict, struct btrfs_path *path)
2466 {
2467         int ret;
2468
2469         do {
2470                 ret = check_committed_ref(root, path, objectid,
2471                                           offset, bytenr, strict);
2472                 if (ret && ret != -ENOENT)
2473                         goto out;
2474
2475                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2476         } while (ret == -EAGAIN);
2477
2478 out:
2479         btrfs_release_path(path);
2480         if (btrfs_is_data_reloc_root(root))
2481                 WARN_ON(ret > 0);
2482         return ret;
2483 }
2484
2485 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2486                            struct btrfs_root *root,
2487                            struct extent_buffer *buf,
2488                            int full_backref, int inc)
2489 {
2490         struct btrfs_fs_info *fs_info = root->fs_info;
2491         u64 bytenr;
2492         u64 num_bytes;
2493         u64 parent;
2494         u64 ref_root;
2495         u32 nritems;
2496         struct btrfs_key key;
2497         struct btrfs_file_extent_item *fi;
2498         struct btrfs_ref generic_ref = { 0 };
2499         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2500         int i;
2501         int action;
2502         int level;
2503         int ret = 0;
2504
2505         if (btrfs_is_testing(fs_info))
2506                 return 0;
2507
2508         ref_root = btrfs_header_owner(buf);
2509         nritems = btrfs_header_nritems(buf);
2510         level = btrfs_header_level(buf);
2511
2512         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2513                 return 0;
2514
2515         if (full_backref)
2516                 parent = buf->start;
2517         else
2518                 parent = 0;
2519         if (inc)
2520                 action = BTRFS_ADD_DELAYED_REF;
2521         else
2522                 action = BTRFS_DROP_DELAYED_REF;
2523
2524         for (i = 0; i < nritems; i++) {
2525                 if (level == 0) {
2526                         btrfs_item_key_to_cpu(buf, &key, i);
2527                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2528                                 continue;
2529                         fi = btrfs_item_ptr(buf, i,
2530                                             struct btrfs_file_extent_item);
2531                         if (btrfs_file_extent_type(buf, fi) ==
2532                             BTRFS_FILE_EXTENT_INLINE)
2533                                 continue;
2534                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2535                         if (bytenr == 0)
2536                                 continue;
2537
2538                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2539                         key.offset -= btrfs_file_extent_offset(buf, fi);
2540                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2541                                                num_bytes, parent, ref_root);
2542                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2543                                             key.offset, root->root_key.objectid,
2544                                             for_reloc);
2545                         if (inc)
2546                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2547                         else
2548                                 ret = btrfs_free_extent(trans, &generic_ref);
2549                         if (ret)
2550                                 goto fail;
2551                 } else {
2552                         bytenr = btrfs_node_blockptr(buf, i);
2553                         num_bytes = fs_info->nodesize;
2554                         /* We don't know the owning_root, use 0. */
2555                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2556                                                num_bytes, parent, 0);
2557                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2558                                             root->root_key.objectid, for_reloc);
2559                         if (inc)
2560                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2561                         else
2562                                 ret = btrfs_free_extent(trans, &generic_ref);
2563                         if (ret)
2564                                 goto fail;
2565                 }
2566         }
2567         return 0;
2568 fail:
2569         return ret;
2570 }
2571
2572 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2573                   struct extent_buffer *buf, int full_backref)
2574 {
2575         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2576 }
2577
2578 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2579                   struct extent_buffer *buf, int full_backref)
2580 {
2581         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2582 }
2583
2584 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2585 {
2586         struct btrfs_fs_info *fs_info = root->fs_info;
2587         u64 flags;
2588         u64 ret;
2589
2590         if (data)
2591                 flags = BTRFS_BLOCK_GROUP_DATA;
2592         else if (root == fs_info->chunk_root)
2593                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2594         else
2595                 flags = BTRFS_BLOCK_GROUP_METADATA;
2596
2597         ret = btrfs_get_alloc_profile(fs_info, flags);
2598         return ret;
2599 }
2600
2601 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2602 {
2603         struct rb_node *leftmost;
2604         u64 bytenr = 0;
2605
2606         read_lock(&fs_info->block_group_cache_lock);
2607         /* Get the block group with the lowest logical start address. */
2608         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2609         if (leftmost) {
2610                 struct btrfs_block_group *bg;
2611
2612                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2613                 bytenr = bg->start;
2614         }
2615         read_unlock(&fs_info->block_group_cache_lock);
2616
2617         return bytenr;
2618 }
2619
2620 static int pin_down_extent(struct btrfs_trans_handle *trans,
2621                            struct btrfs_block_group *cache,
2622                            u64 bytenr, u64 num_bytes, int reserved)
2623 {
2624         struct btrfs_fs_info *fs_info = cache->fs_info;
2625
2626         spin_lock(&cache->space_info->lock);
2627         spin_lock(&cache->lock);
2628         cache->pinned += num_bytes;
2629         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2630                                              num_bytes);
2631         if (reserved) {
2632                 cache->reserved -= num_bytes;
2633                 cache->space_info->bytes_reserved -= num_bytes;
2634         }
2635         spin_unlock(&cache->lock);
2636         spin_unlock(&cache->space_info->lock);
2637
2638         set_extent_bit(&trans->transaction->pinned_extents, bytenr,
2639                        bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
2640         return 0;
2641 }
2642
2643 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2644                      u64 bytenr, u64 num_bytes, int reserved)
2645 {
2646         struct btrfs_block_group *cache;
2647
2648         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2649         BUG_ON(!cache); /* Logic error */
2650
2651         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2652
2653         btrfs_put_block_group(cache);
2654         return 0;
2655 }
2656
2657 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2658                                     const struct extent_buffer *eb)
2659 {
2660         struct btrfs_block_group *cache;
2661         int ret;
2662
2663         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
2664         if (!cache)
2665                 return -EINVAL;
2666
2667         /*
2668          * Fully cache the free space first so that our pin removes the free space
2669          * from the cache.
2670          */
2671         ret = btrfs_cache_block_group(cache, true);
2672         if (ret)
2673                 goto out;
2674
2675         pin_down_extent(trans, cache, eb->start, eb->len, 0);
2676
2677         /* remove us from the free space cache (if we're there at all) */
2678         ret = btrfs_remove_free_space(cache, eb->start, eb->len);
2679 out:
2680         btrfs_put_block_group(cache);
2681         return ret;
2682 }
2683
2684 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2685                                    u64 start, u64 num_bytes)
2686 {
2687         int ret;
2688         struct btrfs_block_group *block_group;
2689
2690         block_group = btrfs_lookup_block_group(fs_info, start);
2691         if (!block_group)
2692                 return -EINVAL;
2693
2694         ret = btrfs_cache_block_group(block_group, true);
2695         if (ret)
2696                 goto out;
2697
2698         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2699 out:
2700         btrfs_put_block_group(block_group);
2701         return ret;
2702 }
2703
2704 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2705 {
2706         struct btrfs_fs_info *fs_info = eb->fs_info;
2707         struct btrfs_file_extent_item *item;
2708         struct btrfs_key key;
2709         int found_type;
2710         int i;
2711         int ret = 0;
2712
2713         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2714                 return 0;
2715
2716         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2717                 btrfs_item_key_to_cpu(eb, &key, i);
2718                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2719                         continue;
2720                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2721                 found_type = btrfs_file_extent_type(eb, item);
2722                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2723                         continue;
2724                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2725                         continue;
2726                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2727                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2728                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2729                 if (ret)
2730                         break;
2731         }
2732
2733         return ret;
2734 }
2735
2736 static void
2737 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2738 {
2739         atomic_inc(&bg->reservations);
2740 }
2741
2742 /*
2743  * Returns the free cluster for the given space info and sets empty_cluster to
2744  * what it should be based on the mount options.
2745  */
2746 static struct btrfs_free_cluster *
2747 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2748                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2749 {
2750         struct btrfs_free_cluster *ret = NULL;
2751
2752         *empty_cluster = 0;
2753         if (btrfs_mixed_space_info(space_info))
2754                 return ret;
2755
2756         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2757                 ret = &fs_info->meta_alloc_cluster;
2758                 if (btrfs_test_opt(fs_info, SSD))
2759                         *empty_cluster = SZ_2M;
2760                 else
2761                         *empty_cluster = SZ_64K;
2762         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2763                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2764                 *empty_cluster = SZ_2M;
2765                 ret = &fs_info->data_alloc_cluster;
2766         }
2767
2768         return ret;
2769 }
2770
2771 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2772                               u64 start, u64 end,
2773                               const bool return_free_space)
2774 {
2775         struct btrfs_block_group *cache = NULL;
2776         struct btrfs_space_info *space_info;
2777         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2778         struct btrfs_free_cluster *cluster = NULL;
2779         u64 len;
2780         u64 total_unpinned = 0;
2781         u64 empty_cluster = 0;
2782         bool readonly;
2783
2784         while (start <= end) {
2785                 readonly = false;
2786                 if (!cache ||
2787                     start >= cache->start + cache->length) {
2788                         if (cache)
2789                                 btrfs_put_block_group(cache);
2790                         total_unpinned = 0;
2791                         cache = btrfs_lookup_block_group(fs_info, start);
2792                         BUG_ON(!cache); /* Logic error */
2793
2794                         cluster = fetch_cluster_info(fs_info,
2795                                                      cache->space_info,
2796                                                      &empty_cluster);
2797                         empty_cluster <<= 1;
2798                 }
2799
2800                 len = cache->start + cache->length - start;
2801                 len = min(len, end + 1 - start);
2802
2803                 if (return_free_space)
2804                         btrfs_add_free_space(cache, start, len);
2805
2806                 start += len;
2807                 total_unpinned += len;
2808                 space_info = cache->space_info;
2809
2810                 /*
2811                  * If this space cluster has been marked as fragmented and we've
2812                  * unpinned enough in this block group to potentially allow a
2813                  * cluster to be created inside of it go ahead and clear the
2814                  * fragmented check.
2815                  */
2816                 if (cluster && cluster->fragmented &&
2817                     total_unpinned > empty_cluster) {
2818                         spin_lock(&cluster->lock);
2819                         cluster->fragmented = 0;
2820                         spin_unlock(&cluster->lock);
2821                 }
2822
2823                 spin_lock(&space_info->lock);
2824                 spin_lock(&cache->lock);
2825                 cache->pinned -= len;
2826                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2827                 space_info->max_extent_size = 0;
2828                 if (cache->ro) {
2829                         space_info->bytes_readonly += len;
2830                         readonly = true;
2831                 } else if (btrfs_is_zoned(fs_info)) {
2832                         /* Need reset before reusing in a zoned block group */
2833                         space_info->bytes_zone_unusable += len;
2834                         readonly = true;
2835                 }
2836                 spin_unlock(&cache->lock);
2837                 if (!readonly && return_free_space &&
2838                     global_rsv->space_info == space_info) {
2839                         spin_lock(&global_rsv->lock);
2840                         if (!global_rsv->full) {
2841                                 u64 to_add = min(len, global_rsv->size -
2842                                                       global_rsv->reserved);
2843
2844                                 global_rsv->reserved += to_add;
2845                                 btrfs_space_info_update_bytes_may_use(fs_info,
2846                                                 space_info, to_add);
2847                                 if (global_rsv->reserved >= global_rsv->size)
2848                                         global_rsv->full = 1;
2849                                 len -= to_add;
2850                         }
2851                         spin_unlock(&global_rsv->lock);
2852                 }
2853                 /* Add to any tickets we may have */
2854                 if (!readonly && return_free_space && len)
2855                         btrfs_try_granting_tickets(fs_info, space_info);
2856                 spin_unlock(&space_info->lock);
2857         }
2858
2859         if (cache)
2860                 btrfs_put_block_group(cache);
2861         return 0;
2862 }
2863
2864 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2865 {
2866         struct btrfs_fs_info *fs_info = trans->fs_info;
2867         struct btrfs_block_group *block_group, *tmp;
2868         struct list_head *deleted_bgs;
2869         struct extent_io_tree *unpin;
2870         u64 start;
2871         u64 end;
2872         int ret;
2873
2874         unpin = &trans->transaction->pinned_extents;
2875
2876         while (!TRANS_ABORTED(trans)) {
2877                 struct extent_state *cached_state = NULL;
2878
2879                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2880                 if (!find_first_extent_bit(unpin, 0, &start, &end,
2881                                            EXTENT_DIRTY, &cached_state)) {
2882                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2883                         break;
2884                 }
2885
2886                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2887                         ret = btrfs_discard_extent(fs_info, start,
2888                                                    end + 1 - start, NULL);
2889
2890                 clear_extent_dirty(unpin, start, end, &cached_state);
2891                 unpin_extent_range(fs_info, start, end, true);
2892                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2893                 free_extent_state(cached_state);
2894                 cond_resched();
2895         }
2896
2897         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2898                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2899                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2900         }
2901
2902         /*
2903          * Transaction is finished.  We don't need the lock anymore.  We
2904          * do need to clean up the block groups in case of a transaction
2905          * abort.
2906          */
2907         deleted_bgs = &trans->transaction->deleted_bgs;
2908         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2909                 u64 trimmed = 0;
2910
2911                 ret = -EROFS;
2912                 if (!TRANS_ABORTED(trans))
2913                         ret = btrfs_discard_extent(fs_info,
2914                                                    block_group->start,
2915                                                    block_group->length,
2916                                                    &trimmed);
2917
2918                 list_del_init(&block_group->bg_list);
2919                 btrfs_unfreeze_block_group(block_group);
2920                 btrfs_put_block_group(block_group);
2921
2922                 if (ret) {
2923                         const char *errstr = btrfs_decode_error(ret);
2924                         btrfs_warn(fs_info,
2925                            "discard failed while removing blockgroup: errno=%d %s",
2926                                    ret, errstr);
2927                 }
2928         }
2929
2930         return 0;
2931 }
2932
2933 /*
2934  * Parse an extent item's inline extents looking for a simple quotas owner ref.
2935  *
2936  * @fs_info:    the btrfs_fs_info for this mount
2937  * @leaf:       a leaf in the extent tree containing the extent item
2938  * @slot:       the slot in the leaf where the extent item is found
2939  *
2940  * Returns the objectid of the root that originally allocated the extent item
2941  * if the inline owner ref is expected and present, otherwise 0.
2942  *
2943  * If an extent item has an owner ref item, it will be the first inline ref
2944  * item. Therefore the logic is to check whether there are any inline ref
2945  * items, then check the type of the first one.
2946  */
2947 u64 btrfs_get_extent_owner_root(struct btrfs_fs_info *fs_info,
2948                                 struct extent_buffer *leaf, int slot)
2949 {
2950         struct btrfs_extent_item *ei;
2951         struct btrfs_extent_inline_ref *iref;
2952         struct btrfs_extent_owner_ref *oref;
2953         unsigned long ptr;
2954         unsigned long end;
2955         int type;
2956
2957         if (!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA))
2958                 return 0;
2959
2960         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
2961         ptr = (unsigned long)(ei + 1);
2962         end = (unsigned long)ei + btrfs_item_size(leaf, slot);
2963
2964         /* No inline ref items of any kind, can't check type. */
2965         if (ptr == end)
2966                 return 0;
2967
2968         iref = (struct btrfs_extent_inline_ref *)ptr;
2969         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
2970
2971         /* We found an owner ref, get the root out of it. */
2972         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
2973                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
2974                 return btrfs_extent_owner_ref_root_id(leaf, oref);
2975         }
2976
2977         /* We have inline refs, but not an owner ref. */
2978         return 0;
2979 }
2980
2981 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2982                                      u64 bytenr, struct btrfs_squota_delta *delta)
2983 {
2984         int ret;
2985         u64 num_bytes = delta->num_bytes;
2986
2987         if (delta->is_data) {
2988                 struct btrfs_root *csum_root;
2989
2990                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2991                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2992                 if (ret) {
2993                         btrfs_abort_transaction(trans, ret);
2994                         return ret;
2995                 }
2996
2997                 ret = btrfs_delete_raid_extent(trans, bytenr, num_bytes);
2998                 if (ret) {
2999                         btrfs_abort_transaction(trans, ret);
3000                         return ret;
3001                 }
3002         }
3003
3004         ret = btrfs_record_squota_delta(trans->fs_info, delta);
3005         if (ret) {
3006                 btrfs_abort_transaction(trans, ret);
3007                 return ret;
3008         }
3009
3010         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3011         if (ret) {
3012                 btrfs_abort_transaction(trans, ret);
3013                 return ret;
3014         }
3015
3016         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
3017         if (ret)
3018                 btrfs_abort_transaction(trans, ret);
3019
3020         return ret;
3021 }
3022
3023 #define abort_and_dump(trans, path, fmt, args...)       \
3024 ({                                                      \
3025         btrfs_abort_transaction(trans, -EUCLEAN);       \
3026         btrfs_print_leaf(path->nodes[0]);               \
3027         btrfs_crit(trans->fs_info, fmt, ##args);        \
3028 })
3029
3030 /*
3031  * Drop one or more refs of @node.
3032  *
3033  * 1. Locate the extent refs.
3034  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
3035  *    Locate it, then reduce the refs number or remove the ref line completely.
3036  *
3037  * 2. Update the refs count in EXTENT/METADATA_ITEM
3038  *
3039  * Inline backref case:
3040  *
3041  * in extent tree we have:
3042  *
3043  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3044  *              refs 2 gen 6 flags DATA
3045  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3046  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
3047  *
3048  * This function gets called with:
3049  *
3050  *    node->bytenr = 13631488
3051  *    node->num_bytes = 1048576
3052  *    root_objectid = FS_TREE
3053  *    owner_objectid = 257
3054  *    owner_offset = 0
3055  *    refs_to_drop = 1
3056  *
3057  * Then we should get some like:
3058  *
3059  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3060  *              refs 1 gen 6 flags DATA
3061  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3062  *
3063  * Keyed backref case:
3064  *
3065  * in extent tree we have:
3066  *
3067  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3068  *              refs 754 gen 6 flags DATA
3069  *      [...]
3070  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
3071  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
3072  *
3073  * This function get called with:
3074  *
3075  *    node->bytenr = 13631488
3076  *    node->num_bytes = 1048576
3077  *    root_objectid = FS_TREE
3078  *    owner_objectid = 866
3079  *    owner_offset = 0
3080  *    refs_to_drop = 1
3081  *
3082  * Then we should get some like:
3083  *
3084  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3085  *              refs 753 gen 6 flags DATA
3086  *
3087  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
3088  */
3089 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
3090                                struct btrfs_delayed_ref_head *href,
3091                                struct btrfs_delayed_ref_node *node, u64 parent,
3092                                u64 root_objectid, u64 owner_objectid,
3093                                u64 owner_offset,
3094                                struct btrfs_delayed_extent_op *extent_op)
3095 {
3096         struct btrfs_fs_info *info = trans->fs_info;
3097         struct btrfs_key key;
3098         struct btrfs_path *path;
3099         struct btrfs_root *extent_root;
3100         struct extent_buffer *leaf;
3101         struct btrfs_extent_item *ei;
3102         struct btrfs_extent_inline_ref *iref;
3103         int ret;
3104         int is_data;
3105         int extent_slot = 0;
3106         int found_extent = 0;
3107         int num_to_del = 1;
3108         int refs_to_drop = node->ref_mod;
3109         u32 item_size;
3110         u64 refs;
3111         u64 bytenr = node->bytenr;
3112         u64 num_bytes = node->num_bytes;
3113         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
3114         u64 delayed_ref_root = href->owning_root;
3115
3116         extent_root = btrfs_extent_root(info, bytenr);
3117         ASSERT(extent_root);
3118
3119         path = btrfs_alloc_path();
3120         if (!path)
3121                 return -ENOMEM;
3122
3123         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
3124
3125         if (!is_data && refs_to_drop != 1) {
3126                 btrfs_crit(info,
3127 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
3128                            node->bytenr, refs_to_drop);
3129                 ret = -EINVAL;
3130                 btrfs_abort_transaction(trans, ret);
3131                 goto out;
3132         }
3133
3134         if (is_data)
3135                 skinny_metadata = false;
3136
3137         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
3138                                     parent, root_objectid, owner_objectid,
3139                                     owner_offset);
3140         if (ret == 0) {
3141                 /*
3142                  * Either the inline backref or the SHARED_DATA_REF/
3143                  * SHARED_BLOCK_REF is found
3144                  *
3145                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
3146                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
3147                  */
3148                 extent_slot = path->slots[0];
3149                 while (extent_slot >= 0) {
3150                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3151                                               extent_slot);
3152                         if (key.objectid != bytenr)
3153                                 break;
3154                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3155                             key.offset == num_bytes) {
3156                                 found_extent = 1;
3157                                 break;
3158                         }
3159                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3160                             key.offset == owner_objectid) {
3161                                 found_extent = 1;
3162                                 break;
3163                         }
3164
3165                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3166                         if (path->slots[0] - extent_slot > 5)
3167                                 break;
3168                         extent_slot--;
3169                 }
3170
3171                 if (!found_extent) {
3172                         if (iref) {
3173                                 abort_and_dump(trans, path,
3174 "invalid iref slot %u, no EXTENT/METADATA_ITEM found but has inline extent ref",
3175                                            path->slots[0]);
3176                                 ret = -EUCLEAN;
3177                                 goto out;
3178                         }
3179                         /* Must be SHARED_* item, remove the backref first */
3180                         ret = remove_extent_backref(trans, extent_root, path,
3181                                                     NULL, refs_to_drop, is_data);
3182                         if (ret) {
3183                                 btrfs_abort_transaction(trans, ret);
3184                                 goto out;
3185                         }
3186                         btrfs_release_path(path);
3187
3188                         /* Slow path to locate EXTENT/METADATA_ITEM */
3189                         key.objectid = bytenr;
3190                         key.type = BTRFS_EXTENT_ITEM_KEY;
3191                         key.offset = num_bytes;
3192
3193                         if (!is_data && skinny_metadata) {
3194                                 key.type = BTRFS_METADATA_ITEM_KEY;
3195                                 key.offset = owner_objectid;
3196                         }
3197
3198                         ret = btrfs_search_slot(trans, extent_root,
3199                                                 &key, path, -1, 1);
3200                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3201                                 /*
3202                                  * Couldn't find our skinny metadata item,
3203                                  * see if we have ye olde extent item.
3204                                  */
3205                                 path->slots[0]--;
3206                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3207                                                       path->slots[0]);
3208                                 if (key.objectid == bytenr &&
3209                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3210                                     key.offset == num_bytes)
3211                                         ret = 0;
3212                         }
3213
3214                         if (ret > 0 && skinny_metadata) {
3215                                 skinny_metadata = false;
3216                                 key.objectid = bytenr;
3217                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3218                                 key.offset = num_bytes;
3219                                 btrfs_release_path(path);
3220                                 ret = btrfs_search_slot(trans, extent_root,
3221                                                         &key, path, -1, 1);
3222                         }
3223
3224                         if (ret) {
3225                                 if (ret > 0)
3226                                         btrfs_print_leaf(path->nodes[0]);
3227                                 btrfs_err(info,
3228                         "umm, got %d back from search, was looking for %llu, slot %d",
3229                                           ret, bytenr, path->slots[0]);
3230                         }
3231                         if (ret < 0) {
3232                                 btrfs_abort_transaction(trans, ret);
3233                                 goto out;
3234                         }
3235                         extent_slot = path->slots[0];
3236                 }
3237         } else if (WARN_ON(ret == -ENOENT)) {
3238                 abort_and_dump(trans, path,
3239 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu slot %d",
3240                                bytenr, parent, root_objectid, owner_objectid,
3241                                owner_offset, path->slots[0]);
3242                 goto out;
3243         } else {
3244                 btrfs_abort_transaction(trans, ret);
3245                 goto out;
3246         }
3247
3248         leaf = path->nodes[0];
3249         item_size = btrfs_item_size(leaf, extent_slot);
3250         if (unlikely(item_size < sizeof(*ei))) {
3251                 ret = -EUCLEAN;
3252                 btrfs_err(trans->fs_info,
3253                           "unexpected extent item size, has %u expect >= %zu",
3254                           item_size, sizeof(*ei));
3255                 btrfs_abort_transaction(trans, ret);
3256                 goto out;
3257         }
3258         ei = btrfs_item_ptr(leaf, extent_slot,
3259                             struct btrfs_extent_item);
3260         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3261             key.type == BTRFS_EXTENT_ITEM_KEY) {
3262                 struct btrfs_tree_block_info *bi;
3263
3264                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3265                         abort_and_dump(trans, path,
3266 "invalid extent item size for key (%llu, %u, %llu) slot %u owner %llu, has %u expect >= %zu",
3267                                        key.objectid, key.type, key.offset,
3268                                        path->slots[0], owner_objectid, item_size,
3269                                        sizeof(*ei) + sizeof(*bi));
3270                         ret = -EUCLEAN;
3271                         goto out;
3272                 }
3273                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3274                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3275         }
3276
3277         refs = btrfs_extent_refs(leaf, ei);
3278         if (refs < refs_to_drop) {
3279                 abort_and_dump(trans, path,
3280                 "trying to drop %d refs but we only have %llu for bytenr %llu slot %u",
3281                                refs_to_drop, refs, bytenr, path->slots[0]);
3282                 ret = -EUCLEAN;
3283                 goto out;
3284         }
3285         refs -= refs_to_drop;
3286
3287         if (refs > 0) {
3288                 if (extent_op)
3289                         __run_delayed_extent_op(extent_op, leaf, ei);
3290                 /*
3291                  * In the case of inline back ref, reference count will
3292                  * be updated by remove_extent_backref
3293                  */
3294                 if (iref) {
3295                         if (!found_extent) {
3296                                 abort_and_dump(trans, path,
3297 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found, slot %u",
3298                                                path->slots[0]);
3299                                 ret = -EUCLEAN;
3300                                 goto out;
3301                         }
3302                 } else {
3303                         btrfs_set_extent_refs(leaf, ei, refs);
3304                         btrfs_mark_buffer_dirty(trans, leaf);
3305                 }
3306                 if (found_extent) {
3307                         ret = remove_extent_backref(trans, extent_root, path,
3308                                                     iref, refs_to_drop, is_data);
3309                         if (ret) {
3310                                 btrfs_abort_transaction(trans, ret);
3311                                 goto out;
3312                         }
3313                 }
3314         } else {
3315                 struct btrfs_squota_delta delta = {
3316                         .root = delayed_ref_root,
3317                         .num_bytes = num_bytes,
3318                         .is_data = is_data,
3319                         .is_inc = false,
3320                         .generation = btrfs_extent_generation(leaf, ei),
3321                 };
3322
3323                 /* In this branch refs == 1 */
3324                 if (found_extent) {
3325                         if (is_data && refs_to_drop !=
3326                             extent_data_ref_count(path, iref)) {
3327                                 abort_and_dump(trans, path,
3328                 "invalid refs_to_drop, current refs %u refs_to_drop %u slot %u",
3329                                                extent_data_ref_count(path, iref),
3330                                                refs_to_drop, path->slots[0]);
3331                                 ret = -EUCLEAN;
3332                                 goto out;
3333                         }
3334                         if (iref) {
3335                                 if (path->slots[0] != extent_slot) {
3336                                         abort_and_dump(trans, path,
3337 "invalid iref, extent item key (%llu %u %llu) slot %u doesn't have wanted iref",
3338                                                        key.objectid, key.type,
3339                                                        key.offset, path->slots[0]);
3340                                         ret = -EUCLEAN;
3341                                         goto out;
3342                                 }
3343                         } else {
3344                                 /*
3345                                  * No inline ref, we must be at SHARED_* item,
3346                                  * And it's single ref, it must be:
3347                                  * |    extent_slot       ||extent_slot + 1|
3348                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3349                                  */
3350                                 if (path->slots[0] != extent_slot + 1) {
3351                                         abort_and_dump(trans, path,
3352         "invalid SHARED_* item slot %u, previous item is not EXTENT/METADATA_ITEM",
3353                                                        path->slots[0]);
3354                                         ret = -EUCLEAN;
3355                                         goto out;
3356                                 }
3357                                 path->slots[0] = extent_slot;
3358                                 num_to_del = 2;
3359                         }
3360                 }
3361                 /*
3362                  * We can't infer the data owner from the delayed ref, so we need
3363                  * to try to get it from the owning ref item.
3364                  *
3365                  * If it is not present, then that extent was not written under
3366                  * simple quotas mode, so we don't need to account for its deletion.
3367                  */
3368                 if (is_data)
3369                         delta.root = btrfs_get_extent_owner_root(trans->fs_info,
3370                                                                  leaf, extent_slot);
3371
3372                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3373                                       num_to_del);
3374                 if (ret) {
3375                         btrfs_abort_transaction(trans, ret);
3376                         goto out;
3377                 }
3378                 btrfs_release_path(path);
3379
3380                 ret = do_free_extent_accounting(trans, bytenr, &delta);
3381         }
3382         btrfs_release_path(path);
3383
3384 out:
3385         btrfs_free_path(path);
3386         return ret;
3387 }
3388
3389 /*
3390  * when we free an block, it is possible (and likely) that we free the last
3391  * delayed ref for that extent as well.  This searches the delayed ref tree for
3392  * a given extent, and if there are no other delayed refs to be processed, it
3393  * removes it from the tree.
3394  */
3395 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3396                                       u64 bytenr)
3397 {
3398         struct btrfs_delayed_ref_head *head;
3399         struct btrfs_delayed_ref_root *delayed_refs;
3400         int ret = 0;
3401
3402         delayed_refs = &trans->transaction->delayed_refs;
3403         spin_lock(&delayed_refs->lock);
3404         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3405         if (!head)
3406                 goto out_delayed_unlock;
3407
3408         spin_lock(&head->lock);
3409         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3410                 goto out;
3411
3412         if (cleanup_extent_op(head) != NULL)
3413                 goto out;
3414
3415         /*
3416          * waiting for the lock here would deadlock.  If someone else has it
3417          * locked they are already in the process of dropping it anyway
3418          */
3419         if (!mutex_trylock(&head->mutex))
3420                 goto out;
3421
3422         btrfs_delete_ref_head(delayed_refs, head);
3423         head->processing = false;
3424
3425         spin_unlock(&head->lock);
3426         spin_unlock(&delayed_refs->lock);
3427
3428         BUG_ON(head->extent_op);
3429         if (head->must_insert_reserved)
3430                 ret = 1;
3431
3432         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3433         mutex_unlock(&head->mutex);
3434         btrfs_put_delayed_ref_head(head);
3435         return ret;
3436 out:
3437         spin_unlock(&head->lock);
3438
3439 out_delayed_unlock:
3440         spin_unlock(&delayed_refs->lock);
3441         return 0;
3442 }
3443
3444 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3445                            u64 root_id,
3446                            struct extent_buffer *buf,
3447                            u64 parent, int last_ref)
3448 {
3449         struct btrfs_fs_info *fs_info = trans->fs_info;
3450         struct btrfs_ref generic_ref = { 0 };
3451         struct btrfs_block_group *bg;
3452         int ret;
3453
3454         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3455                                buf->start, buf->len, parent, btrfs_header_owner(buf));
3456         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3457                             root_id, 0, false);
3458
3459         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3460                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3461                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3462                 BUG_ON(ret); /* -ENOMEM */
3463         }
3464
3465         if (!last_ref)
3466                 return;
3467
3468         if (btrfs_header_generation(buf) != trans->transid)
3469                 goto out;
3470
3471         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3472                 ret = check_ref_cleanup(trans, buf->start);
3473                 if (!ret)
3474                         goto out;
3475         }
3476
3477         bg = btrfs_lookup_block_group(fs_info, buf->start);
3478
3479         if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3480                 pin_down_extent(trans, bg, buf->start, buf->len, 1);
3481                 btrfs_put_block_group(bg);
3482                 goto out;
3483         }
3484
3485         /*
3486          * If there are tree mod log users we may have recorded mod log
3487          * operations for this node.  If we re-allocate this node we
3488          * could replay operations on this node that happened when it
3489          * existed in a completely different root.  For example if it
3490          * was part of root A, then was reallocated to root B, and we
3491          * are doing a btrfs_old_search_slot(root b), we could replay
3492          * operations that happened when the block was part of root A,
3493          * giving us an inconsistent view of the btree.
3494          *
3495          * We are safe from races here because at this point no other
3496          * node or root points to this extent buffer, so if after this
3497          * check a new tree mod log user joins we will not have an
3498          * existing log of operations on this node that we have to
3499          * contend with.
3500          */
3501
3502         if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags)
3503                      || btrfs_is_zoned(fs_info)) {
3504                 pin_down_extent(trans, bg, buf->start, buf->len, 1);
3505                 btrfs_put_block_group(bg);
3506                 goto out;
3507         }
3508
3509         WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3510
3511         btrfs_add_free_space(bg, buf->start, buf->len);
3512         btrfs_free_reserved_bytes(bg, buf->len, 0);
3513         btrfs_put_block_group(bg);
3514         trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3515
3516 out:
3517
3518         /*
3519          * Deleting the buffer, clear the corrupt flag since it doesn't
3520          * matter anymore.
3521          */
3522         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3523 }
3524
3525 /* Can return -ENOMEM */
3526 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3527 {
3528         struct btrfs_fs_info *fs_info = trans->fs_info;
3529         int ret;
3530
3531         if (btrfs_is_testing(fs_info))
3532                 return 0;
3533
3534         /*
3535          * tree log blocks never actually go into the extent allocation
3536          * tree, just update pinning info and exit early.
3537          */
3538         if ((ref->type == BTRFS_REF_METADATA &&
3539              ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID) ||
3540             (ref->type == BTRFS_REF_DATA &&
3541              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3542                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3543                 ret = 0;
3544         } else if (ref->type == BTRFS_REF_METADATA) {
3545                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3546         } else {
3547                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3548         }
3549
3550         if (!((ref->type == BTRFS_REF_METADATA &&
3551                ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID) ||
3552               (ref->type == BTRFS_REF_DATA &&
3553                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3554                 btrfs_ref_tree_mod(fs_info, ref);
3555
3556         return ret;
3557 }
3558
3559 enum btrfs_loop_type {
3560         /*
3561          * Start caching block groups but do not wait for progress or for them
3562          * to be done.
3563          */
3564         LOOP_CACHING_NOWAIT,
3565
3566         /*
3567          * Wait for the block group free_space >= the space we're waiting for if
3568          * the block group isn't cached.
3569          */
3570         LOOP_CACHING_WAIT,
3571
3572         /*
3573          * Allow allocations to happen from block groups that do not yet have a
3574          * size classification.
3575          */
3576         LOOP_UNSET_SIZE_CLASS,
3577
3578         /*
3579          * Allocate a chunk and then retry the allocation.
3580          */
3581         LOOP_ALLOC_CHUNK,
3582
3583         /*
3584          * Ignore the size class restrictions for this allocation.
3585          */
3586         LOOP_WRONG_SIZE_CLASS,
3587
3588         /*
3589          * Ignore the empty size, only try to allocate the number of bytes
3590          * needed for this allocation.
3591          */
3592         LOOP_NO_EMPTY_SIZE,
3593 };
3594
3595 static inline void
3596 btrfs_lock_block_group(struct btrfs_block_group *cache,
3597                        int delalloc)
3598 {
3599         if (delalloc)
3600                 down_read(&cache->data_rwsem);
3601 }
3602
3603 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3604                        int delalloc)
3605 {
3606         btrfs_get_block_group(cache);
3607         if (delalloc)
3608                 down_read(&cache->data_rwsem);
3609 }
3610
3611 static struct btrfs_block_group *btrfs_lock_cluster(
3612                    struct btrfs_block_group *block_group,
3613                    struct btrfs_free_cluster *cluster,
3614                    int delalloc)
3615         __acquires(&cluster->refill_lock)
3616 {
3617         struct btrfs_block_group *used_bg = NULL;
3618
3619         spin_lock(&cluster->refill_lock);
3620         while (1) {
3621                 used_bg = cluster->block_group;
3622                 if (!used_bg)
3623                         return NULL;
3624
3625                 if (used_bg == block_group)
3626                         return used_bg;
3627
3628                 btrfs_get_block_group(used_bg);
3629
3630                 if (!delalloc)
3631                         return used_bg;
3632
3633                 if (down_read_trylock(&used_bg->data_rwsem))
3634                         return used_bg;
3635
3636                 spin_unlock(&cluster->refill_lock);
3637
3638                 /* We should only have one-level nested. */
3639                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3640
3641                 spin_lock(&cluster->refill_lock);
3642                 if (used_bg == cluster->block_group)
3643                         return used_bg;
3644
3645                 up_read(&used_bg->data_rwsem);
3646                 btrfs_put_block_group(used_bg);
3647         }
3648 }
3649
3650 static inline void
3651 btrfs_release_block_group(struct btrfs_block_group *cache,
3652                          int delalloc)
3653 {
3654         if (delalloc)
3655                 up_read(&cache->data_rwsem);
3656         btrfs_put_block_group(cache);
3657 }
3658
3659 /*
3660  * Helper function for find_free_extent().
3661  *
3662  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3663  * Return >0 to inform caller that we find nothing
3664  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3665  */
3666 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3667                                       struct find_free_extent_ctl *ffe_ctl,
3668                                       struct btrfs_block_group **cluster_bg_ret)
3669 {
3670         struct btrfs_block_group *cluster_bg;
3671         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3672         u64 aligned_cluster;
3673         u64 offset;
3674         int ret;
3675
3676         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3677         if (!cluster_bg)
3678                 goto refill_cluster;
3679         if (cluster_bg != bg && (cluster_bg->ro ||
3680             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3681                 goto release_cluster;
3682
3683         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3684                         ffe_ctl->num_bytes, cluster_bg->start,
3685                         &ffe_ctl->max_extent_size);
3686         if (offset) {
3687                 /* We have a block, we're done */
3688                 spin_unlock(&last_ptr->refill_lock);
3689                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3690                 *cluster_bg_ret = cluster_bg;
3691                 ffe_ctl->found_offset = offset;
3692                 return 0;
3693         }
3694         WARN_ON(last_ptr->block_group != cluster_bg);
3695
3696 release_cluster:
3697         /*
3698          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3699          * lets just skip it and let the allocator find whatever block it can
3700          * find. If we reach this point, we will have tried the cluster
3701          * allocator plenty of times and not have found anything, so we are
3702          * likely way too fragmented for the clustering stuff to find anything.
3703          *
3704          * However, if the cluster is taken from the current block group,
3705          * release the cluster first, so that we stand a better chance of
3706          * succeeding in the unclustered allocation.
3707          */
3708         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3709                 spin_unlock(&last_ptr->refill_lock);
3710                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3711                 return -ENOENT;
3712         }
3713
3714         /* This cluster didn't work out, free it and start over */
3715         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3716
3717         if (cluster_bg != bg)
3718                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3719
3720 refill_cluster:
3721         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3722                 spin_unlock(&last_ptr->refill_lock);
3723                 return -ENOENT;
3724         }
3725
3726         aligned_cluster = max_t(u64,
3727                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3728                         bg->full_stripe_len);
3729         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3730                         ffe_ctl->num_bytes, aligned_cluster);
3731         if (ret == 0) {
3732                 /* Now pull our allocation out of this cluster */
3733                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3734                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3735                                 &ffe_ctl->max_extent_size);
3736                 if (offset) {
3737                         /* We found one, proceed */
3738                         spin_unlock(&last_ptr->refill_lock);
3739                         ffe_ctl->found_offset = offset;
3740                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3741                         return 0;
3742                 }
3743         }
3744         /*
3745          * At this point we either didn't find a cluster or we weren't able to
3746          * allocate a block from our cluster.  Free the cluster we've been
3747          * trying to use, and go to the next block group.
3748          */
3749         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3750         spin_unlock(&last_ptr->refill_lock);
3751         return 1;
3752 }
3753
3754 /*
3755  * Return >0 to inform caller that we find nothing
3756  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3757  */
3758 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3759                                         struct find_free_extent_ctl *ffe_ctl)
3760 {
3761         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3762         u64 offset;
3763
3764         /*
3765          * We are doing an unclustered allocation, set the fragmented flag so
3766          * we don't bother trying to setup a cluster again until we get more
3767          * space.
3768          */
3769         if (unlikely(last_ptr)) {
3770                 spin_lock(&last_ptr->lock);
3771                 last_ptr->fragmented = 1;
3772                 spin_unlock(&last_ptr->lock);
3773         }
3774         if (ffe_ctl->cached) {
3775                 struct btrfs_free_space_ctl *free_space_ctl;
3776
3777                 free_space_ctl = bg->free_space_ctl;
3778                 spin_lock(&free_space_ctl->tree_lock);
3779                 if (free_space_ctl->free_space <
3780                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3781                     ffe_ctl->empty_size) {
3782                         ffe_ctl->total_free_space = max_t(u64,
3783                                         ffe_ctl->total_free_space,
3784                                         free_space_ctl->free_space);
3785                         spin_unlock(&free_space_ctl->tree_lock);
3786                         return 1;
3787                 }
3788                 spin_unlock(&free_space_ctl->tree_lock);
3789         }
3790
3791         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3792                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3793                         &ffe_ctl->max_extent_size);
3794         if (!offset)
3795                 return 1;
3796         ffe_ctl->found_offset = offset;
3797         return 0;
3798 }
3799
3800 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3801                                    struct find_free_extent_ctl *ffe_ctl,
3802                                    struct btrfs_block_group **bg_ret)
3803 {
3804         int ret;
3805
3806         /* We want to try and use the cluster allocator, so lets look there */
3807         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3808                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3809                 if (ret >= 0)
3810                         return ret;
3811                 /* ret == -ENOENT case falls through */
3812         }
3813
3814         return find_free_extent_unclustered(block_group, ffe_ctl);
3815 }
3816
3817 /*
3818  * Tree-log block group locking
3819  * ============================
3820  *
3821  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3822  * indicates the starting address of a block group, which is reserved only
3823  * for tree-log metadata.
3824  *
3825  * Lock nesting
3826  * ============
3827  *
3828  * space_info::lock
3829  *   block_group::lock
3830  *     fs_info::treelog_bg_lock
3831  */
3832
3833 /*
3834  * Simple allocator for sequential-only block group. It only allows sequential
3835  * allocation. No need to play with trees. This function also reserves the
3836  * bytes as in btrfs_add_reserved_bytes.
3837  */
3838 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3839                                struct find_free_extent_ctl *ffe_ctl,
3840                                struct btrfs_block_group **bg_ret)
3841 {
3842         struct btrfs_fs_info *fs_info = block_group->fs_info;
3843         struct btrfs_space_info *space_info = block_group->space_info;
3844         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3845         u64 start = block_group->start;
3846         u64 num_bytes = ffe_ctl->num_bytes;
3847         u64 avail;
3848         u64 bytenr = block_group->start;
3849         u64 log_bytenr;
3850         u64 data_reloc_bytenr;
3851         int ret = 0;
3852         bool skip = false;
3853
3854         ASSERT(btrfs_is_zoned(block_group->fs_info));
3855
3856         /*
3857          * Do not allow non-tree-log blocks in the dedicated tree-log block
3858          * group, and vice versa.
3859          */
3860         spin_lock(&fs_info->treelog_bg_lock);
3861         log_bytenr = fs_info->treelog_bg;
3862         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3863                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3864                 skip = true;
3865         spin_unlock(&fs_info->treelog_bg_lock);
3866         if (skip)
3867                 return 1;
3868
3869         /*
3870          * Do not allow non-relocation blocks in the dedicated relocation block
3871          * group, and vice versa.
3872          */
3873         spin_lock(&fs_info->relocation_bg_lock);
3874         data_reloc_bytenr = fs_info->data_reloc_bg;
3875         if (data_reloc_bytenr &&
3876             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3877              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3878                 skip = true;
3879         spin_unlock(&fs_info->relocation_bg_lock);
3880         if (skip)
3881                 return 1;
3882
3883         /* Check RO and no space case before trying to activate it */
3884         spin_lock(&block_group->lock);
3885         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3886                 ret = 1;
3887                 /*
3888                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3889                  * Return the error after taking the locks.
3890                  */
3891         }
3892         spin_unlock(&block_group->lock);
3893
3894         /* Metadata block group is activated at write time. */
3895         if (!ret && (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
3896             !btrfs_zone_activate(block_group)) {
3897                 ret = 1;
3898                 /*
3899                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3900                  * Return the error after taking the locks.
3901                  */
3902         }
3903
3904         spin_lock(&space_info->lock);
3905         spin_lock(&block_group->lock);
3906         spin_lock(&fs_info->treelog_bg_lock);
3907         spin_lock(&fs_info->relocation_bg_lock);
3908
3909         if (ret)
3910                 goto out;
3911
3912         ASSERT(!ffe_ctl->for_treelog ||
3913                block_group->start == fs_info->treelog_bg ||
3914                fs_info->treelog_bg == 0);
3915         ASSERT(!ffe_ctl->for_data_reloc ||
3916                block_group->start == fs_info->data_reloc_bg ||
3917                fs_info->data_reloc_bg == 0);
3918
3919         if (block_group->ro ||
3920             (!ffe_ctl->for_data_reloc &&
3921              test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))) {
3922                 ret = 1;
3923                 goto out;
3924         }
3925
3926         /*
3927          * Do not allow currently using block group to be tree-log dedicated
3928          * block group.
3929          */
3930         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3931             (block_group->used || block_group->reserved)) {
3932                 ret = 1;
3933                 goto out;
3934         }
3935
3936         /*
3937          * Do not allow currently used block group to be the data relocation
3938          * dedicated block group.
3939          */
3940         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3941             (block_group->used || block_group->reserved)) {
3942                 ret = 1;
3943                 goto out;
3944         }
3945
3946         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3947         avail = block_group->zone_capacity - block_group->alloc_offset;
3948         if (avail < num_bytes) {
3949                 if (ffe_ctl->max_extent_size < avail) {
3950                         /*
3951                          * With sequential allocator, free space is always
3952                          * contiguous
3953                          */
3954                         ffe_ctl->max_extent_size = avail;
3955                         ffe_ctl->total_free_space = avail;
3956                 }
3957                 ret = 1;
3958                 goto out;
3959         }
3960
3961         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3962                 fs_info->treelog_bg = block_group->start;
3963
3964         if (ffe_ctl->for_data_reloc) {
3965                 if (!fs_info->data_reloc_bg)
3966                         fs_info->data_reloc_bg = block_group->start;
3967                 /*
3968                  * Do not allow allocations from this block group, unless it is
3969                  * for data relocation. Compared to increasing the ->ro, setting
3970                  * the ->zoned_data_reloc_ongoing flag still allows nocow
3971                  * writers to come in. See btrfs_inc_nocow_writers().
3972                  *
3973                  * We need to disable an allocation to avoid an allocation of
3974                  * regular (non-relocation data) extent. With mix of relocation
3975                  * extents and regular extents, we can dispatch WRITE commands
3976                  * (for relocation extents) and ZONE APPEND commands (for
3977                  * regular extents) at the same time to the same zone, which
3978                  * easily break the write pointer.
3979                  *
3980                  * Also, this flag avoids this block group to be zone finished.
3981                  */
3982                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3983         }
3984
3985         ffe_ctl->found_offset = start + block_group->alloc_offset;
3986         block_group->alloc_offset += num_bytes;
3987         spin_lock(&ctl->tree_lock);
3988         ctl->free_space -= num_bytes;
3989         spin_unlock(&ctl->tree_lock);
3990
3991         /*
3992          * We do not check if found_offset is aligned to stripesize. The
3993          * address is anyway rewritten when using zone append writing.
3994          */
3995
3996         ffe_ctl->search_start = ffe_ctl->found_offset;
3997
3998 out:
3999         if (ret && ffe_ctl->for_treelog)
4000                 fs_info->treelog_bg = 0;
4001         if (ret && ffe_ctl->for_data_reloc)
4002                 fs_info->data_reloc_bg = 0;
4003         spin_unlock(&fs_info->relocation_bg_lock);
4004         spin_unlock(&fs_info->treelog_bg_lock);
4005         spin_unlock(&block_group->lock);
4006         spin_unlock(&space_info->lock);
4007         return ret;
4008 }
4009
4010 static int do_allocation(struct btrfs_block_group *block_group,
4011                          struct find_free_extent_ctl *ffe_ctl,
4012                          struct btrfs_block_group **bg_ret)
4013 {
4014         switch (ffe_ctl->policy) {
4015         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4016                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
4017         case BTRFS_EXTENT_ALLOC_ZONED:
4018                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
4019         default:
4020                 BUG();
4021         }
4022 }
4023
4024 static void release_block_group(struct btrfs_block_group *block_group,
4025                                 struct find_free_extent_ctl *ffe_ctl,
4026                                 int delalloc)
4027 {
4028         switch (ffe_ctl->policy) {
4029         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4030                 ffe_ctl->retry_uncached = false;
4031                 break;
4032         case BTRFS_EXTENT_ALLOC_ZONED:
4033                 /* Nothing to do */
4034                 break;
4035         default:
4036                 BUG();
4037         }
4038
4039         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
4040                ffe_ctl->index);
4041         btrfs_release_block_group(block_group, delalloc);
4042 }
4043
4044 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
4045                                    struct btrfs_key *ins)
4046 {
4047         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4048
4049         if (!ffe_ctl->use_cluster && last_ptr) {
4050                 spin_lock(&last_ptr->lock);
4051                 last_ptr->window_start = ins->objectid;
4052                 spin_unlock(&last_ptr->lock);
4053         }
4054 }
4055
4056 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
4057                          struct btrfs_key *ins)
4058 {
4059         switch (ffe_ctl->policy) {
4060         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4061                 found_extent_clustered(ffe_ctl, ins);
4062                 break;
4063         case BTRFS_EXTENT_ALLOC_ZONED:
4064                 /* Nothing to do */
4065                 break;
4066         default:
4067                 BUG();
4068         }
4069 }
4070
4071 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
4072                                     struct find_free_extent_ctl *ffe_ctl)
4073 {
4074         /* Block group's activeness is not a requirement for METADATA block groups. */
4075         if (!(ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA))
4076                 return 0;
4077
4078         /* If we can activate new zone, just allocate a chunk and use it */
4079         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
4080                 return 0;
4081
4082         /*
4083          * We already reached the max active zones. Try to finish one block
4084          * group to make a room for a new block group. This is only possible
4085          * for a data block group because btrfs_zone_finish() may need to wait
4086          * for a running transaction which can cause a deadlock for metadata
4087          * allocation.
4088          */
4089         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4090                 int ret = btrfs_zone_finish_one_bg(fs_info);
4091
4092                 if (ret == 1)
4093                         return 0;
4094                 else if (ret < 0)
4095                         return ret;
4096         }
4097
4098         /*
4099          * If we have enough free space left in an already active block group
4100          * and we can't activate any other zone now, do not allow allocating a
4101          * new chunk and let find_free_extent() retry with a smaller size.
4102          */
4103         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
4104                 return -ENOSPC;
4105
4106         /*
4107          * Even min_alloc_size is not left in any block groups. Since we cannot
4108          * activate a new block group, allocating it may not help. Let's tell a
4109          * caller to try again and hope it progress something by writing some
4110          * parts of the region. That is only possible for data block groups,
4111          * where a part of the region can be written.
4112          */
4113         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
4114                 return -EAGAIN;
4115
4116         /*
4117          * We cannot activate a new block group and no enough space left in any
4118          * block groups. So, allocating a new block group may not help. But,
4119          * there is nothing to do anyway, so let's go with it.
4120          */
4121         return 0;
4122 }
4123
4124 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4125                               struct find_free_extent_ctl *ffe_ctl)
4126 {
4127         switch (ffe_ctl->policy) {
4128         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4129                 return 0;
4130         case BTRFS_EXTENT_ALLOC_ZONED:
4131                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4132         default:
4133                 BUG();
4134         }
4135 }
4136
4137 /*
4138  * Return >0 means caller needs to re-search for free extent
4139  * Return 0 means we have the needed free extent.
4140  * Return <0 means we failed to locate any free extent.
4141  */
4142 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4143                                         struct btrfs_key *ins,
4144                                         struct find_free_extent_ctl *ffe_ctl,
4145                                         bool full_search)
4146 {
4147         struct btrfs_root *root = fs_info->chunk_root;
4148         int ret;
4149
4150         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4151             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4152                 ffe_ctl->orig_have_caching_bg = true;
4153
4154         if (ins->objectid) {
4155                 found_extent(ffe_ctl, ins);
4156                 return 0;
4157         }
4158
4159         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4160                 return 1;
4161
4162         ffe_ctl->index++;
4163         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4164                 return 1;
4165
4166         /* See the comments for btrfs_loop_type for an explanation of the phases. */
4167         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4168                 ffe_ctl->index = 0;
4169                 /*
4170                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
4171                  * any uncached bgs and we've already done a full search
4172                  * through.
4173                  */
4174                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
4175                     (!ffe_ctl->orig_have_caching_bg && full_search))
4176                         ffe_ctl->loop++;
4177                 ffe_ctl->loop++;
4178
4179                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4180                         struct btrfs_trans_handle *trans;
4181                         int exist = 0;
4182
4183                         /* Check if allocation policy allows to create a new chunk */
4184                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4185                         if (ret)
4186                                 return ret;
4187
4188                         trans = current->journal_info;
4189                         if (trans)
4190                                 exist = 1;
4191                         else
4192                                 trans = btrfs_join_transaction(root);
4193
4194                         if (IS_ERR(trans)) {
4195                                 ret = PTR_ERR(trans);
4196                                 return ret;
4197                         }
4198
4199                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4200                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4201
4202                         /* Do not bail out on ENOSPC since we can do more. */
4203                         if (ret == -ENOSPC) {
4204                                 ret = 0;
4205                                 ffe_ctl->loop++;
4206                         }
4207                         else if (ret < 0)
4208                                 btrfs_abort_transaction(trans, ret);
4209                         else
4210                                 ret = 0;
4211                         if (!exist)
4212                                 btrfs_end_transaction(trans);
4213                         if (ret)
4214                                 return ret;
4215                 }
4216
4217                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4218                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4219                                 return -ENOSPC;
4220
4221                         /*
4222                          * Don't loop again if we already have no empty_size and
4223                          * no empty_cluster.
4224                          */
4225                         if (ffe_ctl->empty_size == 0 &&
4226                             ffe_ctl->empty_cluster == 0)
4227                                 return -ENOSPC;
4228                         ffe_ctl->empty_size = 0;
4229                         ffe_ctl->empty_cluster = 0;
4230                 }
4231                 return 1;
4232         }
4233         return -ENOSPC;
4234 }
4235
4236 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4237                                               struct btrfs_block_group *bg)
4238 {
4239         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4240                 return true;
4241         if (!btrfs_block_group_should_use_size_class(bg))
4242                 return true;
4243         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4244                 return true;
4245         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4246             bg->size_class == BTRFS_BG_SZ_NONE)
4247                 return true;
4248         return ffe_ctl->size_class == bg->size_class;
4249 }
4250
4251 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4252                                         struct find_free_extent_ctl *ffe_ctl,
4253                                         struct btrfs_space_info *space_info,
4254                                         struct btrfs_key *ins)
4255 {
4256         /*
4257          * If our free space is heavily fragmented we may not be able to make
4258          * big contiguous allocations, so instead of doing the expensive search
4259          * for free space, simply return ENOSPC with our max_extent_size so we
4260          * can go ahead and search for a more manageable chunk.
4261          *
4262          * If our max_extent_size is large enough for our allocation simply
4263          * disable clustering since we will likely not be able to find enough
4264          * space to create a cluster and induce latency trying.
4265          */
4266         if (space_info->max_extent_size) {
4267                 spin_lock(&space_info->lock);
4268                 if (space_info->max_extent_size &&
4269                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4270                         ins->offset = space_info->max_extent_size;
4271                         spin_unlock(&space_info->lock);
4272                         return -ENOSPC;
4273                 } else if (space_info->max_extent_size) {
4274                         ffe_ctl->use_cluster = false;
4275                 }
4276                 spin_unlock(&space_info->lock);
4277         }
4278
4279         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4280                                                &ffe_ctl->empty_cluster);
4281         if (ffe_ctl->last_ptr) {
4282                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4283
4284                 spin_lock(&last_ptr->lock);
4285                 if (last_ptr->block_group)
4286                         ffe_ctl->hint_byte = last_ptr->window_start;
4287                 if (last_ptr->fragmented) {
4288                         /*
4289                          * We still set window_start so we can keep track of the
4290                          * last place we found an allocation to try and save
4291                          * some time.
4292                          */
4293                         ffe_ctl->hint_byte = last_ptr->window_start;
4294                         ffe_ctl->use_cluster = false;
4295                 }
4296                 spin_unlock(&last_ptr->lock);
4297         }
4298
4299         return 0;
4300 }
4301
4302 static int prepare_allocation_zoned(struct btrfs_fs_info *fs_info,
4303                                     struct find_free_extent_ctl *ffe_ctl)
4304 {
4305         if (ffe_ctl->for_treelog) {
4306                 spin_lock(&fs_info->treelog_bg_lock);
4307                 if (fs_info->treelog_bg)
4308                         ffe_ctl->hint_byte = fs_info->treelog_bg;
4309                 spin_unlock(&fs_info->treelog_bg_lock);
4310         } else if (ffe_ctl->for_data_reloc) {
4311                 spin_lock(&fs_info->relocation_bg_lock);
4312                 if (fs_info->data_reloc_bg)
4313                         ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4314                 spin_unlock(&fs_info->relocation_bg_lock);
4315         } else if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4316                 struct btrfs_block_group *block_group;
4317
4318                 spin_lock(&fs_info->zone_active_bgs_lock);
4319                 list_for_each_entry(block_group, &fs_info->zone_active_bgs, active_bg_list) {
4320                         /*
4321                          * No lock is OK here because avail is monotinically
4322                          * decreasing, and this is just a hint.
4323                          */
4324                         u64 avail = block_group->zone_capacity - block_group->alloc_offset;
4325
4326                         if (block_group_bits(block_group, ffe_ctl->flags) &&
4327                             avail >= ffe_ctl->num_bytes) {
4328                                 ffe_ctl->hint_byte = block_group->start;
4329                                 break;
4330                         }
4331                 }
4332                 spin_unlock(&fs_info->zone_active_bgs_lock);
4333         }
4334
4335         return 0;
4336 }
4337
4338 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4339                               struct find_free_extent_ctl *ffe_ctl,
4340                               struct btrfs_space_info *space_info,
4341                               struct btrfs_key *ins)
4342 {
4343         switch (ffe_ctl->policy) {
4344         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4345                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4346                                                     space_info, ins);
4347         case BTRFS_EXTENT_ALLOC_ZONED:
4348                 return prepare_allocation_zoned(fs_info, ffe_ctl);
4349         default:
4350                 BUG();
4351         }
4352 }
4353
4354 /*
4355  * walks the btree of allocated extents and find a hole of a given size.
4356  * The key ins is changed to record the hole:
4357  * ins->objectid == start position
4358  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4359  * ins->offset == the size of the hole.
4360  * Any available blocks before search_start are skipped.
4361  *
4362  * If there is no suitable free space, we will record the max size of
4363  * the free space extent currently.
4364  *
4365  * The overall logic and call chain:
4366  *
4367  * find_free_extent()
4368  * |- Iterate through all block groups
4369  * |  |- Get a valid block group
4370  * |  |- Try to do clustered allocation in that block group
4371  * |  |- Try to do unclustered allocation in that block group
4372  * |  |- Check if the result is valid
4373  * |  |  |- If valid, then exit
4374  * |  |- Jump to next block group
4375  * |
4376  * |- Push harder to find free extents
4377  *    |- If not found, re-iterate all block groups
4378  */
4379 static noinline int find_free_extent(struct btrfs_root *root,
4380                                      struct btrfs_key *ins,
4381                                      struct find_free_extent_ctl *ffe_ctl)
4382 {
4383         struct btrfs_fs_info *fs_info = root->fs_info;
4384         int ret = 0;
4385         int cache_block_group_error = 0;
4386         struct btrfs_block_group *block_group = NULL;
4387         struct btrfs_space_info *space_info;
4388         bool full_search = false;
4389
4390         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4391
4392         ffe_ctl->search_start = 0;
4393         /* For clustered allocation */
4394         ffe_ctl->empty_cluster = 0;
4395         ffe_ctl->last_ptr = NULL;
4396         ffe_ctl->use_cluster = true;
4397         ffe_ctl->have_caching_bg = false;
4398         ffe_ctl->orig_have_caching_bg = false;
4399         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4400         ffe_ctl->loop = 0;
4401         ffe_ctl->retry_uncached = false;
4402         ffe_ctl->cached = 0;
4403         ffe_ctl->max_extent_size = 0;
4404         ffe_ctl->total_free_space = 0;
4405         ffe_ctl->found_offset = 0;
4406         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4407         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4408
4409         if (btrfs_is_zoned(fs_info))
4410                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4411
4412         ins->type = BTRFS_EXTENT_ITEM_KEY;
4413         ins->objectid = 0;
4414         ins->offset = 0;
4415
4416         trace_find_free_extent(root, ffe_ctl);
4417
4418         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4419         if (!space_info) {
4420                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4421                 return -ENOSPC;
4422         }
4423
4424         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4425         if (ret < 0)
4426                 return ret;
4427
4428         ffe_ctl->search_start = max(ffe_ctl->search_start,
4429                                     first_logical_byte(fs_info));
4430         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4431         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4432                 block_group = btrfs_lookup_block_group(fs_info,
4433                                                        ffe_ctl->search_start);
4434                 /*
4435                  * we don't want to use the block group if it doesn't match our
4436                  * allocation bits, or if its not cached.
4437                  *
4438                  * However if we are re-searching with an ideal block group
4439                  * picked out then we don't care that the block group is cached.
4440                  */
4441                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4442                     block_group->cached != BTRFS_CACHE_NO) {
4443                         down_read(&space_info->groups_sem);
4444                         if (list_empty(&block_group->list) ||
4445                             block_group->ro) {
4446                                 /*
4447                                  * someone is removing this block group,
4448                                  * we can't jump into the have_block_group
4449                                  * target because our list pointers are not
4450                                  * valid
4451                                  */
4452                                 btrfs_put_block_group(block_group);
4453                                 up_read(&space_info->groups_sem);
4454                         } else {
4455                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4456                                                         block_group->flags);
4457                                 btrfs_lock_block_group(block_group,
4458                                                        ffe_ctl->delalloc);
4459                                 ffe_ctl->hinted = true;
4460                                 goto have_block_group;
4461                         }
4462                 } else if (block_group) {
4463                         btrfs_put_block_group(block_group);
4464                 }
4465         }
4466 search:
4467         trace_find_free_extent_search_loop(root, ffe_ctl);
4468         ffe_ctl->have_caching_bg = false;
4469         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4470             ffe_ctl->index == 0)
4471                 full_search = true;
4472         down_read(&space_info->groups_sem);
4473         list_for_each_entry(block_group,
4474                             &space_info->block_groups[ffe_ctl->index], list) {
4475                 struct btrfs_block_group *bg_ret;
4476
4477                 ffe_ctl->hinted = false;
4478                 /* If the block group is read-only, we can skip it entirely. */
4479                 if (unlikely(block_group->ro)) {
4480                         if (ffe_ctl->for_treelog)
4481                                 btrfs_clear_treelog_bg(block_group);
4482                         if (ffe_ctl->for_data_reloc)
4483                                 btrfs_clear_data_reloc_bg(block_group);
4484                         continue;
4485                 }
4486
4487                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4488                 ffe_ctl->search_start = block_group->start;
4489
4490                 /*
4491                  * this can happen if we end up cycling through all the
4492                  * raid types, but we want to make sure we only allocate
4493                  * for the proper type.
4494                  */
4495                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4496                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4497                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4498                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4499                                 BTRFS_BLOCK_GROUP_RAID10;
4500
4501                         /*
4502                          * if they asked for extra copies and this block group
4503                          * doesn't provide them, bail.  This does allow us to
4504                          * fill raid0 from raid1.
4505                          */
4506                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4507                                 goto loop;
4508
4509                         /*
4510                          * This block group has different flags than we want.
4511                          * It's possible that we have MIXED_GROUP flag but no
4512                          * block group is mixed.  Just skip such block group.
4513                          */
4514                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4515                         continue;
4516                 }
4517
4518 have_block_group:
4519                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4520                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4521                 if (unlikely(!ffe_ctl->cached)) {
4522                         ffe_ctl->have_caching_bg = true;
4523                         ret = btrfs_cache_block_group(block_group, false);
4524
4525                         /*
4526                          * If we get ENOMEM here or something else we want to
4527                          * try other block groups, because it may not be fatal.
4528                          * However if we can't find anything else we need to
4529                          * save our return here so that we return the actual
4530                          * error that caused problems, not ENOSPC.
4531                          */
4532                         if (ret < 0) {
4533                                 if (!cache_block_group_error)
4534                                         cache_block_group_error = ret;
4535                                 ret = 0;
4536                                 goto loop;
4537                         }
4538                         ret = 0;
4539                 }
4540
4541                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) {
4542                         if (!cache_block_group_error)
4543                                 cache_block_group_error = -EIO;
4544                         goto loop;
4545                 }
4546
4547                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4548                         goto loop;
4549
4550                 bg_ret = NULL;
4551                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4552                 if (ret > 0)
4553                         goto loop;
4554
4555                 if (bg_ret && bg_ret != block_group) {
4556                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4557                         block_group = bg_ret;
4558                 }
4559
4560                 /* Checks */
4561                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4562                                                  fs_info->stripesize);
4563
4564                 /* move on to the next group */
4565                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4566                     block_group->start + block_group->length) {
4567                         btrfs_add_free_space_unused(block_group,
4568                                             ffe_ctl->found_offset,
4569                                             ffe_ctl->num_bytes);
4570                         goto loop;
4571                 }
4572
4573                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4574                         btrfs_add_free_space_unused(block_group,
4575                                         ffe_ctl->found_offset,
4576                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4577
4578                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4579                                                ffe_ctl->num_bytes,
4580                                                ffe_ctl->delalloc,
4581                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4582                 if (ret == -EAGAIN) {
4583                         btrfs_add_free_space_unused(block_group,
4584                                         ffe_ctl->found_offset,
4585                                         ffe_ctl->num_bytes);
4586                         goto loop;
4587                 }
4588                 btrfs_inc_block_group_reservations(block_group);
4589
4590                 /* we are all good, lets return */
4591                 ins->objectid = ffe_ctl->search_start;
4592                 ins->offset = ffe_ctl->num_bytes;
4593
4594                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4595                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4596                 break;
4597 loop:
4598                 if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
4599                     !ffe_ctl->retry_uncached) {
4600                         ffe_ctl->retry_uncached = true;
4601                         btrfs_wait_block_group_cache_progress(block_group,
4602                                                 ffe_ctl->num_bytes +
4603                                                 ffe_ctl->empty_cluster +
4604                                                 ffe_ctl->empty_size);
4605                         goto have_block_group;
4606                 }
4607                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4608                 cond_resched();
4609         }
4610         up_read(&space_info->groups_sem);
4611
4612         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4613         if (ret > 0)
4614                 goto search;
4615
4616         if (ret == -ENOSPC && !cache_block_group_error) {
4617                 /*
4618                  * Use ffe_ctl->total_free_space as fallback if we can't find
4619                  * any contiguous hole.
4620                  */
4621                 if (!ffe_ctl->max_extent_size)
4622                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4623                 spin_lock(&space_info->lock);
4624                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4625                 spin_unlock(&space_info->lock);
4626                 ins->offset = ffe_ctl->max_extent_size;
4627         } else if (ret == -ENOSPC) {
4628                 ret = cache_block_group_error;
4629         }
4630         return ret;
4631 }
4632
4633 /*
4634  * Entry point to the extent allocator. Tries to find a hole that is at least
4635  * as big as @num_bytes.
4636  *
4637  * @root           -    The root that will contain this extent
4638  *
4639  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4640  *                      is used for accounting purposes. This value differs
4641  *                      from @num_bytes only in the case of compressed extents.
4642  *
4643  * @num_bytes      -    Number of bytes to allocate on-disk.
4644  *
4645  * @min_alloc_size -    Indicates the minimum amount of space that the
4646  *                      allocator should try to satisfy. In some cases
4647  *                      @num_bytes may be larger than what is required and if
4648  *                      the filesystem is fragmented then allocation fails.
4649  *                      However, the presence of @min_alloc_size gives a
4650  *                      chance to try and satisfy the smaller allocation.
4651  *
4652  * @empty_size     -    A hint that you plan on doing more COW. This is the
4653  *                      size in bytes the allocator should try to find free
4654  *                      next to the block it returns.  This is just a hint and
4655  *                      may be ignored by the allocator.
4656  *
4657  * @hint_byte      -    Hint to the allocator to start searching above the byte
4658  *                      address passed. It might be ignored.
4659  *
4660  * @ins            -    This key is modified to record the found hole. It will
4661  *                      have the following values:
4662  *                      ins->objectid == start position
4663  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4664  *                      ins->offset == the size of the hole.
4665  *
4666  * @is_data        -    Boolean flag indicating whether an extent is
4667  *                      allocated for data (true) or metadata (false)
4668  *
4669  * @delalloc       -    Boolean flag indicating whether this allocation is for
4670  *                      delalloc or not. If 'true' data_rwsem of block groups
4671  *                      is going to be acquired.
4672  *
4673  *
4674  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4675  * case -ENOSPC is returned then @ins->offset will contain the size of the
4676  * largest available hole the allocator managed to find.
4677  */
4678 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4679                          u64 num_bytes, u64 min_alloc_size,
4680                          u64 empty_size, u64 hint_byte,
4681                          struct btrfs_key *ins, int is_data, int delalloc)
4682 {
4683         struct btrfs_fs_info *fs_info = root->fs_info;
4684         struct find_free_extent_ctl ffe_ctl = {};
4685         bool final_tried = num_bytes == min_alloc_size;
4686         u64 flags;
4687         int ret;
4688         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4689         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4690
4691         flags = get_alloc_profile_by_root(root, is_data);
4692 again:
4693         WARN_ON(num_bytes < fs_info->sectorsize);
4694
4695         ffe_ctl.ram_bytes = ram_bytes;
4696         ffe_ctl.num_bytes = num_bytes;
4697         ffe_ctl.min_alloc_size = min_alloc_size;
4698         ffe_ctl.empty_size = empty_size;
4699         ffe_ctl.flags = flags;
4700         ffe_ctl.delalloc = delalloc;
4701         ffe_ctl.hint_byte = hint_byte;
4702         ffe_ctl.for_treelog = for_treelog;
4703         ffe_ctl.for_data_reloc = for_data_reloc;
4704
4705         ret = find_free_extent(root, ins, &ffe_ctl);
4706         if (!ret && !is_data) {
4707                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4708         } else if (ret == -ENOSPC) {
4709                 if (!final_tried && ins->offset) {
4710                         num_bytes = min(num_bytes >> 1, ins->offset);
4711                         num_bytes = round_down(num_bytes,
4712                                                fs_info->sectorsize);
4713                         num_bytes = max(num_bytes, min_alloc_size);
4714                         ram_bytes = num_bytes;
4715                         if (num_bytes == min_alloc_size)
4716                                 final_tried = true;
4717                         goto again;
4718                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4719                         struct btrfs_space_info *sinfo;
4720
4721                         sinfo = btrfs_find_space_info(fs_info, flags);
4722                         btrfs_err(fs_info,
4723         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4724                                   flags, num_bytes, for_treelog, for_data_reloc);
4725                         if (sinfo)
4726                                 btrfs_dump_space_info(fs_info, sinfo,
4727                                                       num_bytes, 1);
4728                 }
4729         }
4730
4731         return ret;
4732 }
4733
4734 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4735                                u64 start, u64 len, int delalloc)
4736 {
4737         struct btrfs_block_group *cache;
4738
4739         cache = btrfs_lookup_block_group(fs_info, start);
4740         if (!cache) {
4741                 btrfs_err(fs_info, "Unable to find block group for %llu",
4742                           start);
4743                 return -ENOSPC;
4744         }
4745
4746         btrfs_add_free_space(cache, start, len);
4747         btrfs_free_reserved_bytes(cache, len, delalloc);
4748         trace_btrfs_reserved_extent_free(fs_info, start, len);
4749
4750         btrfs_put_block_group(cache);
4751         return 0;
4752 }
4753
4754 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans,
4755                               const struct extent_buffer *eb)
4756 {
4757         struct btrfs_block_group *cache;
4758         int ret = 0;
4759
4760         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
4761         if (!cache) {
4762                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4763                           eb->start);
4764                 return -ENOSPC;
4765         }
4766
4767         ret = pin_down_extent(trans, cache, eb->start, eb->len, 1);
4768         btrfs_put_block_group(cache);
4769         return ret;
4770 }
4771
4772 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4773                                  u64 num_bytes)
4774 {
4775         struct btrfs_fs_info *fs_info = trans->fs_info;
4776         int ret;
4777
4778         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4779         if (ret)
4780                 return ret;
4781
4782         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4783         if (ret) {
4784                 ASSERT(!ret);
4785                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4786                           bytenr, num_bytes);
4787                 return ret;
4788         }
4789
4790         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4791         return 0;
4792 }
4793
4794 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4795                                       u64 parent, u64 root_objectid,
4796                                       u64 flags, u64 owner, u64 offset,
4797                                       struct btrfs_key *ins, int ref_mod, u64 oref_root)
4798 {
4799         struct btrfs_fs_info *fs_info = trans->fs_info;
4800         struct btrfs_root *extent_root;
4801         int ret;
4802         struct btrfs_extent_item *extent_item;
4803         struct btrfs_extent_owner_ref *oref;
4804         struct btrfs_extent_inline_ref *iref;
4805         struct btrfs_path *path;
4806         struct extent_buffer *leaf;
4807         int type;
4808         u32 size;
4809         const bool simple_quota = (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE);
4810
4811         if (parent > 0)
4812                 type = BTRFS_SHARED_DATA_REF_KEY;
4813         else
4814                 type = BTRFS_EXTENT_DATA_REF_KEY;
4815
4816         size = sizeof(*extent_item);
4817         if (simple_quota)
4818                 size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
4819         size += btrfs_extent_inline_ref_size(type);
4820
4821         path = btrfs_alloc_path();
4822         if (!path)
4823                 return -ENOMEM;
4824
4825         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4826         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4827         if (ret) {
4828                 btrfs_free_path(path);
4829                 return ret;
4830         }
4831
4832         leaf = path->nodes[0];
4833         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4834                                      struct btrfs_extent_item);
4835         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4836         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4837         btrfs_set_extent_flags(leaf, extent_item,
4838                                flags | BTRFS_EXTENT_FLAG_DATA);
4839
4840         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4841         if (simple_quota) {
4842                 btrfs_set_extent_inline_ref_type(leaf, iref, BTRFS_EXTENT_OWNER_REF_KEY);
4843                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
4844                 btrfs_set_extent_owner_ref_root_id(leaf, oref, oref_root);
4845                 iref = (struct btrfs_extent_inline_ref *)(oref + 1);
4846         }
4847         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4848
4849         if (parent > 0) {
4850                 struct btrfs_shared_data_ref *ref;
4851                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4852                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4853                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4854         } else {
4855                 struct btrfs_extent_data_ref *ref;
4856                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4857                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4858                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4859                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4860                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4861         }
4862
4863         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
4864         btrfs_free_path(path);
4865
4866         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4867 }
4868
4869 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4870                                      struct btrfs_delayed_ref_node *node,
4871                                      struct btrfs_delayed_extent_op *extent_op)
4872 {
4873         struct btrfs_fs_info *fs_info = trans->fs_info;
4874         struct btrfs_root *extent_root;
4875         int ret;
4876         struct btrfs_extent_item *extent_item;
4877         struct btrfs_key extent_key;
4878         struct btrfs_tree_block_info *block_info;
4879         struct btrfs_extent_inline_ref *iref;
4880         struct btrfs_path *path;
4881         struct extent_buffer *leaf;
4882         struct btrfs_delayed_tree_ref *ref;
4883         u32 size = sizeof(*extent_item) + sizeof(*iref);
4884         u64 flags = extent_op->flags_to_set;
4885         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4886
4887         ref = btrfs_delayed_node_to_tree_ref(node);
4888
4889         extent_key.objectid = node->bytenr;
4890         if (skinny_metadata) {
4891                 extent_key.offset = ref->level;
4892                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4893         } else {
4894                 extent_key.offset = node->num_bytes;
4895                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4896                 size += sizeof(*block_info);
4897         }
4898
4899         path = btrfs_alloc_path();
4900         if (!path)
4901                 return -ENOMEM;
4902
4903         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4904         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4905                                       size);
4906         if (ret) {
4907                 btrfs_free_path(path);
4908                 return ret;
4909         }
4910
4911         leaf = path->nodes[0];
4912         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4913                                      struct btrfs_extent_item);
4914         btrfs_set_extent_refs(leaf, extent_item, 1);
4915         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4916         btrfs_set_extent_flags(leaf, extent_item,
4917                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4918
4919         if (skinny_metadata) {
4920                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4921         } else {
4922                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4923                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4924                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4925                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4926         }
4927
4928         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4929                 btrfs_set_extent_inline_ref_type(leaf, iref,
4930                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4931                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4932         } else {
4933                 btrfs_set_extent_inline_ref_type(leaf, iref,
4934                                                  BTRFS_TREE_BLOCK_REF_KEY);
4935                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4936         }
4937
4938         btrfs_mark_buffer_dirty(trans, leaf);
4939         btrfs_free_path(path);
4940
4941         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4942 }
4943
4944 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4945                                      struct btrfs_root *root, u64 owner,
4946                                      u64 offset, u64 ram_bytes,
4947                                      struct btrfs_key *ins)
4948 {
4949         struct btrfs_ref generic_ref = { 0 };
4950         u64 root_objectid = root->root_key.objectid;
4951         u64 owning_root = root_objectid;
4952
4953         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
4954
4955         if (btrfs_is_data_reloc_root(root) && is_fstree(root->relocation_src_root))
4956                 owning_root = root->relocation_src_root;
4957
4958         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4959                                ins->objectid, ins->offset, 0, owning_root);
4960         btrfs_init_data_ref(&generic_ref, root_objectid, owner,
4961                             offset, 0, false);
4962         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4963
4964         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4965 }
4966
4967 /*
4968  * this is used by the tree logging recovery code.  It records that
4969  * an extent has been allocated and makes sure to clear the free
4970  * space cache bits as well
4971  */
4972 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4973                                    u64 root_objectid, u64 owner, u64 offset,
4974                                    struct btrfs_key *ins)
4975 {
4976         struct btrfs_fs_info *fs_info = trans->fs_info;
4977         int ret;
4978         struct btrfs_block_group *block_group;
4979         struct btrfs_space_info *space_info;
4980         struct btrfs_squota_delta delta = {
4981                 .root = root_objectid,
4982                 .num_bytes = ins->offset,
4983                 .generation = trans->transid,
4984                 .is_data = true,
4985                 .is_inc = true,
4986         };
4987
4988         /*
4989          * Mixed block groups will exclude before processing the log so we only
4990          * need to do the exclude dance if this fs isn't mixed.
4991          */
4992         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4993                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4994                                               ins->offset);
4995                 if (ret)
4996                         return ret;
4997         }
4998
4999         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
5000         if (!block_group)
5001                 return -EINVAL;
5002
5003         space_info = block_group->space_info;
5004         spin_lock(&space_info->lock);
5005         spin_lock(&block_group->lock);
5006         space_info->bytes_reserved += ins->offset;
5007         block_group->reserved += ins->offset;
5008         spin_unlock(&block_group->lock);
5009         spin_unlock(&space_info->lock);
5010
5011         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
5012                                          offset, ins, 1, root_objectid);
5013         if (ret)
5014                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
5015         ret = btrfs_record_squota_delta(fs_info, &delta);
5016         btrfs_put_block_group(block_group);
5017         return ret;
5018 }
5019
5020 #ifdef CONFIG_BTRFS_DEBUG
5021 /*
5022  * Extra safety check in case the extent tree is corrupted and extent allocator
5023  * chooses to use a tree block which is already used and locked.
5024  */
5025 static bool check_eb_lock_owner(const struct extent_buffer *eb)
5026 {
5027         if (eb->lock_owner == current->pid) {
5028                 btrfs_err_rl(eb->fs_info,
5029 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
5030                              eb->start, btrfs_header_owner(eb), current->pid);
5031                 return true;
5032         }
5033         return false;
5034 }
5035 #else
5036 static bool check_eb_lock_owner(struct extent_buffer *eb)
5037 {
5038         return false;
5039 }
5040 #endif
5041
5042 static struct extent_buffer *
5043 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5044                       u64 bytenr, int level, u64 owner,
5045                       enum btrfs_lock_nesting nest)
5046 {
5047         struct btrfs_fs_info *fs_info = root->fs_info;
5048         struct extent_buffer *buf;
5049         u64 lockdep_owner = owner;
5050
5051         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
5052         if (IS_ERR(buf))
5053                 return buf;
5054
5055         if (check_eb_lock_owner(buf)) {
5056                 free_extent_buffer(buf);
5057                 return ERR_PTR(-EUCLEAN);
5058         }
5059
5060         /*
5061          * The reloc trees are just snapshots, so we need them to appear to be
5062          * just like any other fs tree WRT lockdep.
5063          *
5064          * The exception however is in replace_path() in relocation, where we
5065          * hold the lock on the original fs root and then search for the reloc
5066          * root.  At that point we need to make sure any reloc root buffers are
5067          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
5068          * lockdep happy.
5069          */
5070         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
5071             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
5072                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
5073
5074         /* btrfs_clear_buffer_dirty() accesses generation field. */
5075         btrfs_set_header_generation(buf, trans->transid);
5076
5077         /*
5078          * This needs to stay, because we could allocate a freed block from an
5079          * old tree into a new tree, so we need to make sure this new block is
5080          * set to the appropriate level and owner.
5081          */
5082         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
5083
5084         __btrfs_tree_lock(buf, nest);
5085         btrfs_clear_buffer_dirty(trans, buf);
5086         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
5087         clear_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &buf->bflags);
5088
5089         set_extent_buffer_uptodate(buf);
5090
5091         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
5092         btrfs_set_header_level(buf, level);
5093         btrfs_set_header_bytenr(buf, buf->start);
5094         btrfs_set_header_generation(buf, trans->transid);
5095         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
5096         btrfs_set_header_owner(buf, owner);
5097         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
5098         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
5099         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5100                 buf->log_index = root->log_transid % 2;
5101                 /*
5102                  * we allow two log transactions at a time, use different
5103                  * EXTENT bit to differentiate dirty pages.
5104                  */
5105                 if (buf->log_index == 0)
5106                         set_extent_bit(&root->dirty_log_pages, buf->start,
5107                                        buf->start + buf->len - 1,
5108                                        EXTENT_DIRTY, NULL);
5109                 else
5110                         set_extent_bit(&root->dirty_log_pages, buf->start,
5111                                        buf->start + buf->len - 1,
5112                                        EXTENT_NEW, NULL);
5113         } else {
5114                 buf->log_index = -1;
5115                 set_extent_bit(&trans->transaction->dirty_pages, buf->start,
5116                                buf->start + buf->len - 1, EXTENT_DIRTY, NULL);
5117         }
5118         /* this returns a buffer locked for blocking */
5119         return buf;
5120 }
5121
5122 /*
5123  * finds a free extent and does all the dirty work required for allocation
5124  * returns the tree buffer or an ERR_PTR on error.
5125  */
5126 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
5127                                              struct btrfs_root *root,
5128                                              u64 parent, u64 root_objectid,
5129                                              const struct btrfs_disk_key *key,
5130                                              int level, u64 hint,
5131                                              u64 empty_size,
5132                                              u64 reloc_src_root,
5133                                              enum btrfs_lock_nesting nest)
5134 {
5135         struct btrfs_fs_info *fs_info = root->fs_info;
5136         struct btrfs_key ins;
5137         struct btrfs_block_rsv *block_rsv;
5138         struct extent_buffer *buf;
5139         struct btrfs_delayed_extent_op *extent_op;
5140         struct btrfs_ref generic_ref = { 0 };
5141         u64 flags = 0;
5142         int ret;
5143         u32 blocksize = fs_info->nodesize;
5144         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
5145         u64 owning_root;
5146
5147 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5148         if (btrfs_is_testing(fs_info)) {
5149                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
5150                                             level, root_objectid, nest);
5151                 if (!IS_ERR(buf))
5152                         root->alloc_bytenr += blocksize;
5153                 return buf;
5154         }
5155 #endif
5156
5157         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
5158         if (IS_ERR(block_rsv))
5159                 return ERR_CAST(block_rsv);
5160
5161         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
5162                                    empty_size, hint, &ins, 0, 0);
5163         if (ret)
5164                 goto out_unuse;
5165
5166         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
5167                                     root_objectid, nest);
5168         if (IS_ERR(buf)) {
5169                 ret = PTR_ERR(buf);
5170                 goto out_free_reserved;
5171         }
5172         owning_root = btrfs_header_owner(buf);
5173
5174         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5175                 if (parent == 0)
5176                         parent = ins.objectid;
5177                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5178                 owning_root = reloc_src_root;
5179         } else
5180                 BUG_ON(parent > 0);
5181
5182         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5183                 extent_op = btrfs_alloc_delayed_extent_op();
5184                 if (!extent_op) {
5185                         ret = -ENOMEM;
5186                         goto out_free_buf;
5187                 }
5188                 if (key)
5189                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5190                 else
5191                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5192                 extent_op->flags_to_set = flags;
5193                 extent_op->update_key = skinny_metadata ? false : true;
5194                 extent_op->update_flags = true;
5195                 extent_op->level = level;
5196
5197                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
5198                                        ins.objectid, ins.offset, parent, owning_root);
5199                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
5200                                     root->root_key.objectid, false);
5201                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5202                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5203                 if (ret)
5204                         goto out_free_delayed;
5205         }
5206         return buf;
5207
5208 out_free_delayed:
5209         btrfs_free_delayed_extent_op(extent_op);
5210 out_free_buf:
5211         btrfs_tree_unlock(buf);
5212         free_extent_buffer(buf);
5213 out_free_reserved:
5214         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5215 out_unuse:
5216         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5217         return ERR_PTR(ret);
5218 }
5219
5220 struct walk_control {
5221         u64 refs[BTRFS_MAX_LEVEL];
5222         u64 flags[BTRFS_MAX_LEVEL];
5223         struct btrfs_key update_progress;
5224         struct btrfs_key drop_progress;
5225         int drop_level;
5226         int stage;
5227         int level;
5228         int shared_level;
5229         int update_ref;
5230         int keep_locks;
5231         int reada_slot;
5232         int reada_count;
5233         int restarted;
5234 };
5235
5236 #define DROP_REFERENCE  1
5237 #define UPDATE_BACKREF  2
5238
5239 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5240                                      struct btrfs_root *root,
5241                                      struct walk_control *wc,
5242                                      struct btrfs_path *path)
5243 {
5244         struct btrfs_fs_info *fs_info = root->fs_info;
5245         u64 bytenr;
5246         u64 generation;
5247         u64 refs;
5248         u64 flags;
5249         u32 nritems;
5250         struct btrfs_key key;
5251         struct extent_buffer *eb;
5252         int ret;
5253         int slot;
5254         int nread = 0;
5255
5256         if (path->slots[wc->level] < wc->reada_slot) {
5257                 wc->reada_count = wc->reada_count * 2 / 3;
5258                 wc->reada_count = max(wc->reada_count, 2);
5259         } else {
5260                 wc->reada_count = wc->reada_count * 3 / 2;
5261                 wc->reada_count = min_t(int, wc->reada_count,
5262                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5263         }
5264
5265         eb = path->nodes[wc->level];
5266         nritems = btrfs_header_nritems(eb);
5267
5268         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5269                 if (nread >= wc->reada_count)
5270                         break;
5271
5272                 cond_resched();
5273                 bytenr = btrfs_node_blockptr(eb, slot);
5274                 generation = btrfs_node_ptr_generation(eb, slot);
5275
5276                 if (slot == path->slots[wc->level])
5277                         goto reada;
5278
5279                 if (wc->stage == UPDATE_BACKREF &&
5280                     generation <= root->root_key.offset)
5281                         continue;
5282
5283                 /* We don't lock the tree block, it's OK to be racy here */
5284                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5285                                                wc->level - 1, 1, &refs,
5286                                                &flags, NULL);
5287                 /* We don't care about errors in readahead. */
5288                 if (ret < 0)
5289                         continue;
5290                 BUG_ON(refs == 0);
5291
5292                 if (wc->stage == DROP_REFERENCE) {
5293                         if (refs == 1)
5294                                 goto reada;
5295
5296                         if (wc->level == 1 &&
5297                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5298                                 continue;
5299                         if (!wc->update_ref ||
5300                             generation <= root->root_key.offset)
5301                                 continue;
5302                         btrfs_node_key_to_cpu(eb, &key, slot);
5303                         ret = btrfs_comp_cpu_keys(&key,
5304                                                   &wc->update_progress);
5305                         if (ret < 0)
5306                                 continue;
5307                 } else {
5308                         if (wc->level == 1 &&
5309                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5310                                 continue;
5311                 }
5312 reada:
5313                 btrfs_readahead_node_child(eb, slot);
5314                 nread++;
5315         }
5316         wc->reada_slot = slot;
5317 }
5318
5319 /*
5320  * helper to process tree block while walking down the tree.
5321  *
5322  * when wc->stage == UPDATE_BACKREF, this function updates
5323  * back refs for pointers in the block.
5324  *
5325  * NOTE: return value 1 means we should stop walking down.
5326  */
5327 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5328                                    struct btrfs_root *root,
5329                                    struct btrfs_path *path,
5330                                    struct walk_control *wc, int lookup_info)
5331 {
5332         struct btrfs_fs_info *fs_info = root->fs_info;
5333         int level = wc->level;
5334         struct extent_buffer *eb = path->nodes[level];
5335         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5336         int ret;
5337
5338         if (wc->stage == UPDATE_BACKREF &&
5339             btrfs_header_owner(eb) != root->root_key.objectid)
5340                 return 1;
5341
5342         /*
5343          * when reference count of tree block is 1, it won't increase
5344          * again. once full backref flag is set, we never clear it.
5345          */
5346         if (lookup_info &&
5347             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5348              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5349                 BUG_ON(!path->locks[level]);
5350                 ret = btrfs_lookup_extent_info(trans, fs_info,
5351                                                eb->start, level, 1,
5352                                                &wc->refs[level],
5353                                                &wc->flags[level],
5354                                                NULL);
5355                 BUG_ON(ret == -ENOMEM);
5356                 if (ret)
5357                         return ret;
5358                 BUG_ON(wc->refs[level] == 0);
5359         }
5360
5361         if (wc->stage == DROP_REFERENCE) {
5362                 if (wc->refs[level] > 1)
5363                         return 1;
5364
5365                 if (path->locks[level] && !wc->keep_locks) {
5366                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5367                         path->locks[level] = 0;
5368                 }
5369                 return 0;
5370         }
5371
5372         /* wc->stage == UPDATE_BACKREF */
5373         if (!(wc->flags[level] & flag)) {
5374                 BUG_ON(!path->locks[level]);
5375                 ret = btrfs_inc_ref(trans, root, eb, 1);
5376                 BUG_ON(ret); /* -ENOMEM */
5377                 ret = btrfs_dec_ref(trans, root, eb, 0);
5378                 BUG_ON(ret); /* -ENOMEM */
5379                 ret = btrfs_set_disk_extent_flags(trans, eb, flag);
5380                 BUG_ON(ret); /* -ENOMEM */
5381                 wc->flags[level] |= flag;
5382         }
5383
5384         /*
5385          * the block is shared by multiple trees, so it's not good to
5386          * keep the tree lock
5387          */
5388         if (path->locks[level] && level > 0) {
5389                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5390                 path->locks[level] = 0;
5391         }
5392         return 0;
5393 }
5394
5395 /*
5396  * This is used to verify a ref exists for this root to deal with a bug where we
5397  * would have a drop_progress key that hadn't been updated properly.
5398  */
5399 static int check_ref_exists(struct btrfs_trans_handle *trans,
5400                             struct btrfs_root *root, u64 bytenr, u64 parent,
5401                             int level)
5402 {
5403         struct btrfs_path *path;
5404         struct btrfs_extent_inline_ref *iref;
5405         int ret;
5406
5407         path = btrfs_alloc_path();
5408         if (!path)
5409                 return -ENOMEM;
5410
5411         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5412                                     root->fs_info->nodesize, parent,
5413                                     root->root_key.objectid, level, 0);
5414         btrfs_free_path(path);
5415         if (ret == -ENOENT)
5416                 return 0;
5417         if (ret < 0)
5418                 return ret;
5419         return 1;
5420 }
5421
5422 /*
5423  * helper to process tree block pointer.
5424  *
5425  * when wc->stage == DROP_REFERENCE, this function checks
5426  * reference count of the block pointed to. if the block
5427  * is shared and we need update back refs for the subtree
5428  * rooted at the block, this function changes wc->stage to
5429  * UPDATE_BACKREF. if the block is shared and there is no
5430  * need to update back, this function drops the reference
5431  * to the block.
5432  *
5433  * NOTE: return value 1 means we should stop walking down.
5434  */
5435 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5436                                  struct btrfs_root *root,
5437                                  struct btrfs_path *path,
5438                                  struct walk_control *wc, int *lookup_info)
5439 {
5440         struct btrfs_fs_info *fs_info = root->fs_info;
5441         u64 bytenr;
5442         u64 generation;
5443         u64 parent;
5444         u64 owner_root = 0;
5445         struct btrfs_tree_parent_check check = { 0 };
5446         struct btrfs_key key;
5447         struct btrfs_ref ref = { 0 };
5448         struct extent_buffer *next;
5449         int level = wc->level;
5450         int reada = 0;
5451         int ret = 0;
5452         bool need_account = false;
5453
5454         generation = btrfs_node_ptr_generation(path->nodes[level],
5455                                                path->slots[level]);
5456         /*
5457          * if the lower level block was created before the snapshot
5458          * was created, we know there is no need to update back refs
5459          * for the subtree
5460          */
5461         if (wc->stage == UPDATE_BACKREF &&
5462             generation <= root->root_key.offset) {
5463                 *lookup_info = 1;
5464                 return 1;
5465         }
5466
5467         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5468
5469         check.level = level - 1;
5470         check.transid = generation;
5471         check.owner_root = root->root_key.objectid;
5472         check.has_first_key = true;
5473         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5474                               path->slots[level]);
5475
5476         next = find_extent_buffer(fs_info, bytenr);
5477         if (!next) {
5478                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5479                                 root->root_key.objectid, level - 1);
5480                 if (IS_ERR(next))
5481                         return PTR_ERR(next);
5482                 reada = 1;
5483         }
5484         btrfs_tree_lock(next);
5485
5486         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5487                                        &wc->refs[level - 1],
5488                                        &wc->flags[level - 1],
5489                                        &owner_root);
5490         if (ret < 0)
5491                 goto out_unlock;
5492
5493         if (unlikely(wc->refs[level - 1] == 0)) {
5494                 btrfs_err(fs_info, "Missing references.");
5495                 ret = -EIO;
5496                 goto out_unlock;
5497         }
5498         *lookup_info = 0;
5499
5500         if (wc->stage == DROP_REFERENCE) {
5501                 if (wc->refs[level - 1] > 1) {
5502                         need_account = true;
5503                         if (level == 1 &&
5504                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5505                                 goto skip;
5506
5507                         if (!wc->update_ref ||
5508                             generation <= root->root_key.offset)
5509                                 goto skip;
5510
5511                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5512                                               path->slots[level]);
5513                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5514                         if (ret < 0)
5515                                 goto skip;
5516
5517                         wc->stage = UPDATE_BACKREF;
5518                         wc->shared_level = level - 1;
5519                 }
5520         } else {
5521                 if (level == 1 &&
5522                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5523                         goto skip;
5524         }
5525
5526         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5527                 btrfs_tree_unlock(next);
5528                 free_extent_buffer(next);
5529                 next = NULL;
5530                 *lookup_info = 1;
5531         }
5532
5533         if (!next) {
5534                 if (reada && level == 1)
5535                         reada_walk_down(trans, root, wc, path);
5536                 next = read_tree_block(fs_info, bytenr, &check);
5537                 if (IS_ERR(next)) {
5538                         return PTR_ERR(next);
5539                 } else if (!extent_buffer_uptodate(next)) {
5540                         free_extent_buffer(next);
5541                         return -EIO;
5542                 }
5543                 btrfs_tree_lock(next);
5544         }
5545
5546         level--;
5547         ASSERT(level == btrfs_header_level(next));
5548         if (level != btrfs_header_level(next)) {
5549                 btrfs_err(root->fs_info, "mismatched level");
5550                 ret = -EIO;
5551                 goto out_unlock;
5552         }
5553         path->nodes[level] = next;
5554         path->slots[level] = 0;
5555         path->locks[level] = BTRFS_WRITE_LOCK;
5556         wc->level = level;
5557         if (wc->level == 1)
5558                 wc->reada_slot = 0;
5559         return 0;
5560 skip:
5561         wc->refs[level - 1] = 0;
5562         wc->flags[level - 1] = 0;
5563         if (wc->stage == DROP_REFERENCE) {
5564                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5565                         parent = path->nodes[level]->start;
5566                 } else {
5567                         ASSERT(root->root_key.objectid ==
5568                                btrfs_header_owner(path->nodes[level]));
5569                         if (root->root_key.objectid !=
5570                             btrfs_header_owner(path->nodes[level])) {
5571                                 btrfs_err(root->fs_info,
5572                                                 "mismatched block owner");
5573                                 ret = -EIO;
5574                                 goto out_unlock;
5575                         }
5576                         parent = 0;
5577                 }
5578
5579                 /*
5580                  * If we had a drop_progress we need to verify the refs are set
5581                  * as expected.  If we find our ref then we know that from here
5582                  * on out everything should be correct, and we can clear the
5583                  * ->restarted flag.
5584                  */
5585                 if (wc->restarted) {
5586                         ret = check_ref_exists(trans, root, bytenr, parent,
5587                                                level - 1);
5588                         if (ret < 0)
5589                                 goto out_unlock;
5590                         if (ret == 0)
5591                                 goto no_delete;
5592                         ret = 0;
5593                         wc->restarted = 0;
5594                 }
5595
5596                 /*
5597                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5598                  * already accounted them at merge time (replace_path),
5599                  * thus we could skip expensive subtree trace here.
5600                  */
5601                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5602                     need_account) {
5603                         ret = btrfs_qgroup_trace_subtree(trans, next,
5604                                                          generation, level - 1);
5605                         if (ret) {
5606                                 btrfs_err_rl(fs_info,
5607                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5608                                              ret);
5609                         }
5610                 }
5611
5612                 /*
5613                  * We need to update the next key in our walk control so we can
5614                  * update the drop_progress key accordingly.  We don't care if
5615                  * find_next_key doesn't find a key because that means we're at
5616                  * the end and are going to clean up now.
5617                  */
5618                 wc->drop_level = level;
5619                 find_next_key(path, level, &wc->drop_progress);
5620
5621                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5622                                        fs_info->nodesize, parent, owner_root);
5623                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5624                                     0, false);
5625                 ret = btrfs_free_extent(trans, &ref);
5626                 if (ret)
5627                         goto out_unlock;
5628         }
5629 no_delete:
5630         *lookup_info = 1;
5631         ret = 1;
5632
5633 out_unlock:
5634         btrfs_tree_unlock(next);
5635         free_extent_buffer(next);
5636
5637         return ret;
5638 }
5639
5640 /*
5641  * helper to process tree block while walking up the tree.
5642  *
5643  * when wc->stage == DROP_REFERENCE, this function drops
5644  * reference count on the block.
5645  *
5646  * when wc->stage == UPDATE_BACKREF, this function changes
5647  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5648  * to UPDATE_BACKREF previously while processing the block.
5649  *
5650  * NOTE: return value 1 means we should stop walking up.
5651  */
5652 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5653                                  struct btrfs_root *root,
5654                                  struct btrfs_path *path,
5655                                  struct walk_control *wc)
5656 {
5657         struct btrfs_fs_info *fs_info = root->fs_info;
5658         int ret;
5659         int level = wc->level;
5660         struct extent_buffer *eb = path->nodes[level];
5661         u64 parent = 0;
5662
5663         if (wc->stage == UPDATE_BACKREF) {
5664                 BUG_ON(wc->shared_level < level);
5665                 if (level < wc->shared_level)
5666                         goto out;
5667
5668                 ret = find_next_key(path, level + 1, &wc->update_progress);
5669                 if (ret > 0)
5670                         wc->update_ref = 0;
5671
5672                 wc->stage = DROP_REFERENCE;
5673                 wc->shared_level = -1;
5674                 path->slots[level] = 0;
5675
5676                 /*
5677                  * check reference count again if the block isn't locked.
5678                  * we should start walking down the tree again if reference
5679                  * count is one.
5680                  */
5681                 if (!path->locks[level]) {
5682                         BUG_ON(level == 0);
5683                         btrfs_tree_lock(eb);
5684                         path->locks[level] = BTRFS_WRITE_LOCK;
5685
5686                         ret = btrfs_lookup_extent_info(trans, fs_info,
5687                                                        eb->start, level, 1,
5688                                                        &wc->refs[level],
5689                                                        &wc->flags[level],
5690                                                        NULL);
5691                         if (ret < 0) {
5692                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5693                                 path->locks[level] = 0;
5694                                 return ret;
5695                         }
5696                         BUG_ON(wc->refs[level] == 0);
5697                         if (wc->refs[level] == 1) {
5698                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5699                                 path->locks[level] = 0;
5700                                 return 1;
5701                         }
5702                 }
5703         }
5704
5705         /* wc->stage == DROP_REFERENCE */
5706         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5707
5708         if (wc->refs[level] == 1) {
5709                 if (level == 0) {
5710                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5711                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5712                         else
5713                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5714                         BUG_ON(ret); /* -ENOMEM */
5715                         if (is_fstree(root->root_key.objectid)) {
5716                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5717                                 if (ret) {
5718                                         btrfs_err_rl(fs_info,
5719         "error %d accounting leaf items, quota is out of sync, rescan required",
5720                                              ret);
5721                                 }
5722                         }
5723                 }
5724                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5725                 if (!path->locks[level]) {
5726                         btrfs_tree_lock(eb);
5727                         path->locks[level] = BTRFS_WRITE_LOCK;
5728                 }
5729                 btrfs_clear_buffer_dirty(trans, eb);
5730         }
5731
5732         if (eb == root->node) {
5733                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5734                         parent = eb->start;
5735                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5736                         goto owner_mismatch;
5737         } else {
5738                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5739                         parent = path->nodes[level + 1]->start;
5740                 else if (root->root_key.objectid !=
5741                          btrfs_header_owner(path->nodes[level + 1]))
5742                         goto owner_mismatch;
5743         }
5744
5745         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5746                               wc->refs[level] == 1);
5747 out:
5748         wc->refs[level] = 0;
5749         wc->flags[level] = 0;
5750         return 0;
5751
5752 owner_mismatch:
5753         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5754                      btrfs_header_owner(eb), root->root_key.objectid);
5755         return -EUCLEAN;
5756 }
5757
5758 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5759                                    struct btrfs_root *root,
5760                                    struct btrfs_path *path,
5761                                    struct walk_control *wc)
5762 {
5763         int level = wc->level;
5764         int lookup_info = 1;
5765         int ret = 0;
5766
5767         while (level >= 0) {
5768                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5769                 if (ret)
5770                         break;
5771
5772                 if (level == 0)
5773                         break;
5774
5775                 if (path->slots[level] >=
5776                     btrfs_header_nritems(path->nodes[level]))
5777                         break;
5778
5779                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5780                 if (ret > 0) {
5781                         path->slots[level]++;
5782                         continue;
5783                 } else if (ret < 0)
5784                         break;
5785                 level = wc->level;
5786         }
5787         return (ret == 1) ? 0 : ret;
5788 }
5789
5790 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5791                                  struct btrfs_root *root,
5792                                  struct btrfs_path *path,
5793                                  struct walk_control *wc, int max_level)
5794 {
5795         int level = wc->level;
5796         int ret;
5797
5798         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5799         while (level < max_level && path->nodes[level]) {
5800                 wc->level = level;
5801                 if (path->slots[level] + 1 <
5802                     btrfs_header_nritems(path->nodes[level])) {
5803                         path->slots[level]++;
5804                         return 0;
5805                 } else {
5806                         ret = walk_up_proc(trans, root, path, wc);
5807                         if (ret > 0)
5808                                 return 0;
5809                         if (ret < 0)
5810                                 return ret;
5811
5812                         if (path->locks[level]) {
5813                                 btrfs_tree_unlock_rw(path->nodes[level],
5814                                                      path->locks[level]);
5815                                 path->locks[level] = 0;
5816                         }
5817                         free_extent_buffer(path->nodes[level]);
5818                         path->nodes[level] = NULL;
5819                         level++;
5820                 }
5821         }
5822         return 1;
5823 }
5824
5825 /*
5826  * drop a subvolume tree.
5827  *
5828  * this function traverses the tree freeing any blocks that only
5829  * referenced by the tree.
5830  *
5831  * when a shared tree block is found. this function decreases its
5832  * reference count by one. if update_ref is true, this function
5833  * also make sure backrefs for the shared block and all lower level
5834  * blocks are properly updated.
5835  *
5836  * If called with for_reloc == 0, may exit early with -EAGAIN
5837  */
5838 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5839 {
5840         const bool is_reloc_root = (root->root_key.objectid ==
5841                                     BTRFS_TREE_RELOC_OBJECTID);
5842         struct btrfs_fs_info *fs_info = root->fs_info;
5843         struct btrfs_path *path;
5844         struct btrfs_trans_handle *trans;
5845         struct btrfs_root *tree_root = fs_info->tree_root;
5846         struct btrfs_root_item *root_item = &root->root_item;
5847         struct walk_control *wc;
5848         struct btrfs_key key;
5849         int err = 0;
5850         int ret;
5851         int level;
5852         bool root_dropped = false;
5853         bool unfinished_drop = false;
5854
5855         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5856
5857         path = btrfs_alloc_path();
5858         if (!path) {
5859                 err = -ENOMEM;
5860                 goto out;
5861         }
5862
5863         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5864         if (!wc) {
5865                 btrfs_free_path(path);
5866                 err = -ENOMEM;
5867                 goto out;
5868         }
5869
5870         /*
5871          * Use join to avoid potential EINTR from transaction start. See
5872          * wait_reserve_ticket and the whole reservation callchain.
5873          */
5874         if (for_reloc)
5875                 trans = btrfs_join_transaction(tree_root);
5876         else
5877                 trans = btrfs_start_transaction(tree_root, 0);
5878         if (IS_ERR(trans)) {
5879                 err = PTR_ERR(trans);
5880                 goto out_free;
5881         }
5882
5883         err = btrfs_run_delayed_items(trans);
5884         if (err)
5885                 goto out_end_trans;
5886
5887         /*
5888          * This will help us catch people modifying the fs tree while we're
5889          * dropping it.  It is unsafe to mess with the fs tree while it's being
5890          * dropped as we unlock the root node and parent nodes as we walk down
5891          * the tree, assuming nothing will change.  If something does change
5892          * then we'll have stale information and drop references to blocks we've
5893          * already dropped.
5894          */
5895         set_bit(BTRFS_ROOT_DELETING, &root->state);
5896         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5897
5898         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5899                 level = btrfs_header_level(root->node);
5900                 path->nodes[level] = btrfs_lock_root_node(root);
5901                 path->slots[level] = 0;
5902                 path->locks[level] = BTRFS_WRITE_LOCK;
5903                 memset(&wc->update_progress, 0,
5904                        sizeof(wc->update_progress));
5905         } else {
5906                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5907                 memcpy(&wc->update_progress, &key,
5908                        sizeof(wc->update_progress));
5909
5910                 level = btrfs_root_drop_level(root_item);
5911                 BUG_ON(level == 0);
5912                 path->lowest_level = level;
5913                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5914                 path->lowest_level = 0;
5915                 if (ret < 0) {
5916                         err = ret;
5917                         goto out_end_trans;
5918                 }
5919                 WARN_ON(ret > 0);
5920
5921                 /*
5922                  * unlock our path, this is safe because only this
5923                  * function is allowed to delete this snapshot
5924                  */
5925                 btrfs_unlock_up_safe(path, 0);
5926
5927                 level = btrfs_header_level(root->node);
5928                 while (1) {
5929                         btrfs_tree_lock(path->nodes[level]);
5930                         path->locks[level] = BTRFS_WRITE_LOCK;
5931
5932                         ret = btrfs_lookup_extent_info(trans, fs_info,
5933                                                 path->nodes[level]->start,
5934                                                 level, 1, &wc->refs[level],
5935                                                 &wc->flags[level], NULL);
5936                         if (ret < 0) {
5937                                 err = ret;
5938                                 goto out_end_trans;
5939                         }
5940                         BUG_ON(wc->refs[level] == 0);
5941
5942                         if (level == btrfs_root_drop_level(root_item))
5943                                 break;
5944
5945                         btrfs_tree_unlock(path->nodes[level]);
5946                         path->locks[level] = 0;
5947                         WARN_ON(wc->refs[level] != 1);
5948                         level--;
5949                 }
5950         }
5951
5952         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5953         wc->level = level;
5954         wc->shared_level = -1;
5955         wc->stage = DROP_REFERENCE;
5956         wc->update_ref = update_ref;
5957         wc->keep_locks = 0;
5958         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5959
5960         while (1) {
5961
5962                 ret = walk_down_tree(trans, root, path, wc);
5963                 if (ret < 0) {
5964                         btrfs_abort_transaction(trans, ret);
5965                         err = ret;
5966                         break;
5967                 }
5968
5969                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5970                 if (ret < 0) {
5971                         btrfs_abort_transaction(trans, ret);
5972                         err = ret;
5973                         break;
5974                 }
5975
5976                 if (ret > 0) {
5977                         BUG_ON(wc->stage != DROP_REFERENCE);
5978                         break;
5979                 }
5980
5981                 if (wc->stage == DROP_REFERENCE) {
5982                         wc->drop_level = wc->level;
5983                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5984                                               &wc->drop_progress,
5985                                               path->slots[wc->drop_level]);
5986                 }
5987                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5988                                       &wc->drop_progress);
5989                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5990
5991                 BUG_ON(wc->level == 0);
5992                 if (btrfs_should_end_transaction(trans) ||
5993                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5994                         ret = btrfs_update_root(trans, tree_root,
5995                                                 &root->root_key,
5996                                                 root_item);
5997                         if (ret) {
5998                                 btrfs_abort_transaction(trans, ret);
5999                                 err = ret;
6000                                 goto out_end_trans;
6001                         }
6002
6003                         if (!is_reloc_root)
6004                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6005
6006                         btrfs_end_transaction_throttle(trans);
6007                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
6008                                 btrfs_debug(fs_info,
6009                                             "drop snapshot early exit");
6010                                 err = -EAGAIN;
6011                                 goto out_free;
6012                         }
6013
6014                        /*
6015                         * Use join to avoid potential EINTR from transaction
6016                         * start. See wait_reserve_ticket and the whole
6017                         * reservation callchain.
6018                         */
6019                         if (for_reloc)
6020                                 trans = btrfs_join_transaction(tree_root);
6021                         else
6022                                 trans = btrfs_start_transaction(tree_root, 0);
6023                         if (IS_ERR(trans)) {
6024                                 err = PTR_ERR(trans);
6025                                 goto out_free;
6026                         }
6027                 }
6028         }
6029         btrfs_release_path(path);
6030         if (err)
6031                 goto out_end_trans;
6032
6033         ret = btrfs_del_root(trans, &root->root_key);
6034         if (ret) {
6035                 btrfs_abort_transaction(trans, ret);
6036                 err = ret;
6037                 goto out_end_trans;
6038         }
6039
6040         if (!is_reloc_root) {
6041                 ret = btrfs_find_root(tree_root, &root->root_key, path,
6042                                       NULL, NULL);
6043                 if (ret < 0) {
6044                         btrfs_abort_transaction(trans, ret);
6045                         err = ret;
6046                         goto out_end_trans;
6047                 } else if (ret > 0) {
6048                         /* if we fail to delete the orphan item this time
6049                          * around, it'll get picked up the next time.
6050                          *
6051                          * The most common failure here is just -ENOENT.
6052                          */
6053                         btrfs_del_orphan_item(trans, tree_root,
6054                                               root->root_key.objectid);
6055                 }
6056         }
6057
6058         /*
6059          * This subvolume is going to be completely dropped, and won't be
6060          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
6061          * commit transaction time.  So free it here manually.
6062          */
6063         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
6064         btrfs_qgroup_free_meta_all_pertrans(root);
6065
6066         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
6067                 btrfs_add_dropped_root(trans, root);
6068         else
6069                 btrfs_put_root(root);
6070         root_dropped = true;
6071 out_end_trans:
6072         if (!is_reloc_root)
6073                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6074
6075         btrfs_end_transaction_throttle(trans);
6076 out_free:
6077         kfree(wc);
6078         btrfs_free_path(path);
6079 out:
6080         /*
6081          * We were an unfinished drop root, check to see if there are any
6082          * pending, and if not clear and wake up any waiters.
6083          */
6084         if (!err && unfinished_drop)
6085                 btrfs_maybe_wake_unfinished_drop(fs_info);
6086
6087         /*
6088          * So if we need to stop dropping the snapshot for whatever reason we
6089          * need to make sure to add it back to the dead root list so that we
6090          * keep trying to do the work later.  This also cleans up roots if we
6091          * don't have it in the radix (like when we recover after a power fail
6092          * or unmount) so we don't leak memory.
6093          */
6094         if (!for_reloc && !root_dropped)
6095                 btrfs_add_dead_root(root);
6096         return err;
6097 }
6098
6099 /*
6100  * drop subtree rooted at tree block 'node'.
6101  *
6102  * NOTE: this function will unlock and release tree block 'node'
6103  * only used by relocation code
6104  */
6105 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6106                         struct btrfs_root *root,
6107                         struct extent_buffer *node,
6108                         struct extent_buffer *parent)
6109 {
6110         struct btrfs_fs_info *fs_info = root->fs_info;
6111         struct btrfs_path *path;
6112         struct walk_control *wc;
6113         int level;
6114         int parent_level;
6115         int ret = 0;
6116         int wret;
6117
6118         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6119
6120         path = btrfs_alloc_path();
6121         if (!path)
6122                 return -ENOMEM;
6123
6124         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6125         if (!wc) {
6126                 btrfs_free_path(path);
6127                 return -ENOMEM;
6128         }
6129
6130         btrfs_assert_tree_write_locked(parent);
6131         parent_level = btrfs_header_level(parent);
6132         atomic_inc(&parent->refs);
6133         path->nodes[parent_level] = parent;
6134         path->slots[parent_level] = btrfs_header_nritems(parent);
6135
6136         btrfs_assert_tree_write_locked(node);
6137         level = btrfs_header_level(node);
6138         path->nodes[level] = node;
6139         path->slots[level] = 0;
6140         path->locks[level] = BTRFS_WRITE_LOCK;
6141
6142         wc->refs[parent_level] = 1;
6143         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6144         wc->level = level;
6145         wc->shared_level = -1;
6146         wc->stage = DROP_REFERENCE;
6147         wc->update_ref = 0;
6148         wc->keep_locks = 1;
6149         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
6150
6151         while (1) {
6152                 wret = walk_down_tree(trans, root, path, wc);
6153                 if (wret < 0) {
6154                         ret = wret;
6155                         break;
6156                 }
6157
6158                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6159                 if (wret < 0)
6160                         ret = wret;
6161                 if (wret != 0)
6162                         break;
6163         }
6164
6165         kfree(wc);
6166         btrfs_free_path(path);
6167         return ret;
6168 }
6169
6170 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
6171                                    u64 start, u64 end)
6172 {
6173         return unpin_extent_range(fs_info, start, end, false);
6174 }
6175
6176 /*
6177  * It used to be that old block groups would be left around forever.
6178  * Iterating over them would be enough to trim unused space.  Since we
6179  * now automatically remove them, we also need to iterate over unallocated
6180  * space.
6181  *
6182  * We don't want a transaction for this since the discard may take a
6183  * substantial amount of time.  We don't require that a transaction be
6184  * running, but we do need to take a running transaction into account
6185  * to ensure that we're not discarding chunks that were released or
6186  * allocated in the current transaction.
6187  *
6188  * Holding the chunks lock will prevent other threads from allocating
6189  * or releasing chunks, but it won't prevent a running transaction
6190  * from committing and releasing the memory that the pending chunks
6191  * list head uses.  For that, we need to take a reference to the
6192  * transaction and hold the commit root sem.  We only need to hold
6193  * it while performing the free space search since we have already
6194  * held back allocations.
6195  */
6196 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6197 {
6198         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6199         int ret;
6200
6201         *trimmed = 0;
6202
6203         /* Discard not supported = nothing to do. */
6204         if (!bdev_max_discard_sectors(device->bdev))
6205                 return 0;
6206
6207         /* Not writable = nothing to do. */
6208         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6209                 return 0;
6210
6211         /* No free space = nothing to do. */
6212         if (device->total_bytes <= device->bytes_used)
6213                 return 0;
6214
6215         ret = 0;
6216
6217         while (1) {
6218                 struct btrfs_fs_info *fs_info = device->fs_info;
6219                 u64 bytes;
6220
6221                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6222                 if (ret)
6223                         break;
6224
6225                 find_first_clear_extent_bit(&device->alloc_state, start,
6226                                             &start, &end,
6227                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6228
6229                 /* Check if there are any CHUNK_* bits left */
6230                 if (start > device->total_bytes) {
6231                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6232                         btrfs_warn_in_rcu(fs_info,
6233 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6234                                           start, end - start + 1,
6235                                           btrfs_dev_name(device),
6236                                           device->total_bytes);
6237                         mutex_unlock(&fs_info->chunk_mutex);
6238                         ret = 0;
6239                         break;
6240                 }
6241
6242                 /* Ensure we skip the reserved space on each device. */
6243                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6244
6245                 /*
6246                  * If find_first_clear_extent_bit find a range that spans the
6247                  * end of the device it will set end to -1, in this case it's up
6248                  * to the caller to trim the value to the size of the device.
6249                  */
6250                 end = min(end, device->total_bytes - 1);
6251
6252                 len = end - start + 1;
6253
6254                 /* We didn't find any extents */
6255                 if (!len) {
6256                         mutex_unlock(&fs_info->chunk_mutex);
6257                         ret = 0;
6258                         break;
6259                 }
6260
6261                 ret = btrfs_issue_discard(device->bdev, start, len,
6262                                           &bytes);
6263                 if (!ret)
6264                         set_extent_bit(&device->alloc_state, start,
6265                                        start + bytes - 1, CHUNK_TRIMMED, NULL);
6266                 mutex_unlock(&fs_info->chunk_mutex);
6267
6268                 if (ret)
6269                         break;
6270
6271                 start += len;
6272                 *trimmed += bytes;
6273
6274                 if (fatal_signal_pending(current)) {
6275                         ret = -ERESTARTSYS;
6276                         break;
6277                 }
6278
6279                 cond_resched();
6280         }
6281
6282         return ret;
6283 }
6284
6285 /*
6286  * Trim the whole filesystem by:
6287  * 1) trimming the free space in each block group
6288  * 2) trimming the unallocated space on each device
6289  *
6290  * This will also continue trimming even if a block group or device encounters
6291  * an error.  The return value will be the last error, or 0 if nothing bad
6292  * happens.
6293  */
6294 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6295 {
6296         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6297         struct btrfs_block_group *cache = NULL;
6298         struct btrfs_device *device;
6299         u64 group_trimmed;
6300         u64 range_end = U64_MAX;
6301         u64 start;
6302         u64 end;
6303         u64 trimmed = 0;
6304         u64 bg_failed = 0;
6305         u64 dev_failed = 0;
6306         int bg_ret = 0;
6307         int dev_ret = 0;
6308         int ret = 0;
6309
6310         if (range->start == U64_MAX)
6311                 return -EINVAL;
6312
6313         /*
6314          * Check range overflow if range->len is set.
6315          * The default range->len is U64_MAX.
6316          */
6317         if (range->len != U64_MAX &&
6318             check_add_overflow(range->start, range->len, &range_end))
6319                 return -EINVAL;
6320
6321         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6322         for (; cache; cache = btrfs_next_block_group(cache)) {
6323                 if (cache->start >= range_end) {
6324                         btrfs_put_block_group(cache);
6325                         break;
6326                 }
6327
6328                 start = max(range->start, cache->start);
6329                 end = min(range_end, cache->start + cache->length);
6330
6331                 if (end - start >= range->minlen) {
6332                         if (!btrfs_block_group_done(cache)) {
6333                                 ret = btrfs_cache_block_group(cache, true);
6334                                 if (ret) {
6335                                         bg_failed++;
6336                                         bg_ret = ret;
6337                                         continue;
6338                                 }
6339                         }
6340                         ret = btrfs_trim_block_group(cache,
6341                                                      &group_trimmed,
6342                                                      start,
6343                                                      end,
6344                                                      range->minlen);
6345
6346                         trimmed += group_trimmed;
6347                         if (ret) {
6348                                 bg_failed++;
6349                                 bg_ret = ret;
6350                                 continue;
6351                         }
6352                 }
6353         }
6354
6355         if (bg_failed)
6356                 btrfs_warn(fs_info,
6357                         "failed to trim %llu block group(s), last error %d",
6358                         bg_failed, bg_ret);
6359
6360         mutex_lock(&fs_devices->device_list_mutex);
6361         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6362                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6363                         continue;
6364
6365                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6366                 if (ret) {
6367                         dev_failed++;
6368                         dev_ret = ret;
6369                         break;
6370                 }
6371
6372                 trimmed += group_trimmed;
6373         }
6374         mutex_unlock(&fs_devices->device_list_mutex);
6375
6376         if (dev_failed)
6377                 btrfs_warn(fs_info,
6378                         "failed to trim %llu device(s), last error %d",
6379                         dev_failed, dev_ret);
6380         range->len = trimmed;
6381         if (bg_ret)
6382                 return bg_ret;
6383         return dev_ret;
6384 }