fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "misc.h"
  20 #include "tree-log.h"
  21 #include "disk-io.h"
  22 #include "print-tree.h"
  23 #include "volumes.h"
  24 #include "raid56.h"
  25 #include "locking.h"
  26 #include "free-space-cache.h"
  27 #include "free-space-tree.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31 #include "space-info.h"
  32 #include "block-rsv.h"
  33 #include "delalloc-space.h"
  34 #include "block-group.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42
  43 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  44                                struct btrfs_delayed_ref_node *node, u64 parent,
  45                                u64 root_objectid, u64 owner_objectid,
  46                                u64 owner_offset, int refs_to_drop,
  47                                struct btrfs_delayed_extent_op *extra_op);
  48 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  49                                     struct extent_buffer *leaf,
  50                                     struct btrfs_extent_item *ei);
  51 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  52                                       u64 parent, u64 root_objectid,
  53                                       u64 flags, u64 owner, u64 offset,
  54                                       struct btrfs_key *ins, int ref_mod);
  55 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  56                                      struct btrfs_delayed_ref_node *node,
  57                                      struct btrfs_delayed_extent_op *extent_op);
  58 static int find_next_key(struct btrfs_path *path, int level,
  59                          struct btrfs_key *key);
  60
  61 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  62 {
  63         return (cache->flags & bits) == bits;
  64 }
  65
  66 int btrfs_add_excluded_extent(struct btrfs_fs_info *fs_info,
  67                               u64 start, u64 num_bytes)
  68 {
  69         u64 end = start + num_bytes - 1;
  70         set_extent_bits(&fs_info->excluded_extents, start, end,
  71                         EXTENT_UPTODATE);
  72         return 0;
  73 }
  74
  75 void btrfs_free_excluded_extents(struct btrfs_block_group *cache)
  76 {
  77         struct btrfs_fs_info *fs_info = cache->fs_info;
  78         u64 start, end;
  79
  80         start = cache->start;
  81         end = start + cache->length - 1;
  82
  83         clear_extent_bits(&fs_info->excluded_extents, start, end,
  84                           EXTENT_UPTODATE);
  85 }
  86
  87 /* simple helper to search for an existing data extent at a given offset */
  88 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  89 {
  90         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  91         int ret;
  92         struct btrfs_key key;
  93         struct btrfs_path *path;
  94
  95         path = btrfs_alloc_path();
  96         if (!path)
  97                 return -ENOMEM;
  98
  99         key.objectid = start;
 100         key.offset = len;
 101         key.type = BTRFS_EXTENT_ITEM_KEY;
 102         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 103         btrfs_free_path(path);
 104         return ret;
 105 }
 106
 107 /*
 108  * helper function to lookup reference count and flags of a tree block.
 109  *
 110  * the head node for delayed ref is used to store the sum of all the
 111  * reference count modifications queued up in the rbtree. the head
 112  * node may also store the extent flags to set. This way you can check
 113  * to see what the reference count and extent flags would be if all of
 114  * the delayed refs are not processed.
 115  */
 116 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 117                              struct btrfs_fs_info *fs_info, u64 bytenr,
 118                              u64 offset, int metadata, u64 *refs, u64 *flags)
 119 {
 120         struct btrfs_root *extent_root;
 121         struct btrfs_delayed_ref_head *head;
 122         struct btrfs_delayed_ref_root *delayed_refs;
 123         struct btrfs_path *path;
 124         struct btrfs_extent_item *ei;
 125         struct extent_buffer *leaf;
 126         struct btrfs_key key;
 127         u32 item_size;
 128         u64 num_refs;
 129         u64 extent_flags;
 130         int ret;
 131
 132         /*
 133          * If we don't have skinny metadata, don't bother doing anything
 134          * different
 135          */
 136         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 137                 offset = fs_info->nodesize;
 138                 metadata = 0;
 139         }
 140
 141         path = btrfs_alloc_path();
 142         if (!path)
 143                 return -ENOMEM;
 144
 145         if (!trans) {
 146                 path->skip_locking = 1;
 147                 path->search_commit_root = 1;
 148         }
 149
 150 search_again:
 151         key.objectid = bytenr;
 152         key.offset = offset;
 153         if (metadata)
 154                 key.type = BTRFS_METADATA_ITEM_KEY;
 155         else
 156                 key.type = BTRFS_EXTENT_ITEM_KEY;
 157
 158         extent_root = btrfs_extent_root(fs_info, bytenr);
 159         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 160         if (ret < 0)
 161                 goto out_free;
 162
 163         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 164                 if (path->slots[0]) {
 165                         path->slots[0]--;
 166                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 167                                               path->slots[0]);
 168                         if (key.objectid == bytenr &&
 169                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 170                             key.offset == fs_info->nodesize)
 171                                 ret = 0;
 172                 }
 173         }
 174
 175         if (ret == 0) {
 176                 leaf = path->nodes[0];
 177                 item_size = btrfs_item_size(leaf, path->slots[0]);
 178                 if (item_size >= sizeof(*ei)) {
 179                         ei = btrfs_item_ptr(leaf, path->slots[0],
 180                                             struct btrfs_extent_item);
 181                         num_refs = btrfs_extent_refs(leaf, ei);
 182                         extent_flags = btrfs_extent_flags(leaf, ei);
 183                 } else {
 184                         ret = -EINVAL;
 185                         btrfs_print_v0_err(fs_info);
 186                         if (trans)
 187                                 btrfs_abort_transaction(trans, ret);
 188                         else
 189                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 190
 191                         goto out_free;
 192                 }
 193
 194                 BUG_ON(num_refs == 0);
 195         } else {
 196                 num_refs = 0;
 197                 extent_flags = 0;
 198                 ret = 0;
 199         }
 200
 201         if (!trans)
 202                 goto out;
 203
 204         delayed_refs = &trans->transaction->delayed_refs;
 205         spin_lock(&delayed_refs->lock);
 206         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 207         if (head) {
 208                 if (!mutex_trylock(&head->mutex)) {
 209                         refcount_inc(&head->refs);
 210                         spin_unlock(&delayed_refs->lock);
 211
 212                         btrfs_release_path(path);
 213
 214                         /*
 215                          * Mutex was contended, block until it's released and try
 216                          * again
 217                          */
 218                         mutex_lock(&head->mutex);
 219                         mutex_unlock(&head->mutex);
 220                         btrfs_put_delayed_ref_head(head);
 221                         goto search_again;
 222                 }
 223                 spin_lock(&head->lock);
 224                 if (head->extent_op && head->extent_op->update_flags)
 225                         extent_flags |= head->extent_op->flags_to_set;
 226                 else
 227                         BUG_ON(num_refs == 0);
 228
 229                 num_refs += head->ref_mod;
 230                 spin_unlock(&head->lock);
 231                 mutex_unlock(&head->mutex);
 232         }
 233         spin_unlock(&delayed_refs->lock);
 234 out:
 235         WARN_ON(num_refs == 0);
 236         if (refs)
 237                 *refs = num_refs;
 238         if (flags)
 239                 *flags = extent_flags;
 240 out_free:
 241         btrfs_free_path(path);
 242         return ret;
 243 }
 244
 245 /*
 246  * Back reference rules.  Back refs have three main goals:
 247  *
 248  * 1) differentiate between all holders of references to an extent so that
 249  *    when a reference is dropped we can make sure it was a valid reference
 250  *    before freeing the extent.
 251  *
 252  * 2) Provide enough information to quickly find the holders of an extent
 253  *    if we notice a given block is corrupted or bad.
 254  *
 255  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 256  *    maintenance.  This is actually the same as #2, but with a slightly
 257  *    different use case.
 258  *
 259  * There are two kinds of back refs. The implicit back refs is optimized
 260  * for pointers in non-shared tree blocks. For a given pointer in a block,
 261  * back refs of this kind provide information about the block's owner tree
 262  * and the pointer's key. These information allow us to find the block by
 263  * b-tree searching. The full back refs is for pointers in tree blocks not
 264  * referenced by their owner trees. The location of tree block is recorded
 265  * in the back refs. Actually the full back refs is generic, and can be
 266  * used in all cases the implicit back refs is used. The major shortcoming
 267  * of the full back refs is its overhead. Every time a tree block gets
 268  * COWed, we have to update back refs entry for all pointers in it.
 269  *
 270  * For a newly allocated tree block, we use implicit back refs for
 271  * pointers in it. This means most tree related operations only involve
 272  * implicit back refs. For a tree block created in old transaction, the
 273  * only way to drop a reference to it is COW it. So we can detect the
 274  * event that tree block loses its owner tree's reference and do the
 275  * back refs conversion.
 276  *
 277  * When a tree block is COWed through a tree, there are four cases:
 278  *
 279  * The reference count of the block is one and the tree is the block's
 280  * owner tree. Nothing to do in this case.
 281  *
 282  * The reference count of the block is one and the tree is not the
 283  * block's owner tree. In this case, full back refs is used for pointers
 284  * in the block. Remove these full back refs, add implicit back refs for
 285  * every pointers in the new block.
 286  *
 287  * The reference count of the block is greater than one and the tree is
 288  * the block's owner tree. In this case, implicit back refs is used for
 289  * pointers in the block. Add full back refs for every pointers in the
 290  * block, increase lower level extents' reference counts. The original
 291  * implicit back refs are entailed to the new block.
 292  *
 293  * The reference count of the block is greater than one and the tree is
 294  * not the block's owner tree. Add implicit back refs for every pointer in
 295  * the new block, increase lower level extents' reference count.
 296  *
 297  * Back Reference Key composing:
 298  *
 299  * The key objectid corresponds to the first byte in the extent,
 300  * The key type is used to differentiate between types of back refs.
 301  * There are different meanings of the key offset for different types
 302  * of back refs.
 303  *
 304  * File extents can be referenced by:
 305  *
 306  * - multiple snapshots, subvolumes, or different generations in one subvol
 307  * - different files inside a single subvolume
 308  * - different offsets inside a file (bookend extents in file.c)
 309  *
 310  * The extent ref structure for the implicit back refs has fields for:
 311  *
 312  * - Objectid of the subvolume root
 313  * - objectid of the file holding the reference
 314  * - original offset in the file
 315  * - how many bookend extents
 316  *
 317  * The key offset for the implicit back refs is hash of the first
 318  * three fields.
 319  *
 320  * The extent ref structure for the full back refs has field for:
 321  *
 322  * - number of pointers in the tree leaf
 323  *
 324  * The key offset for the implicit back refs is the first byte of
 325  * the tree leaf
 326  *
 327  * When a file extent is allocated, The implicit back refs is used.
 328  * the fields are filled in:
 329  *
 330  *     (root_key.objectid, inode objectid, offset in file, 1)
 331  *
 332  * When a file extent is removed file truncation, we find the
 333  * corresponding implicit back refs and check the following fields:
 334  *
 335  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 336  *
 337  * Btree extents can be referenced by:
 338  *
 339  * - Different subvolumes
 340  *
 341  * Both the implicit back refs and the full back refs for tree blocks
 342  * only consist of key. The key offset for the implicit back refs is
 343  * objectid of block's owner tree. The key offset for the full back refs
 344  * is the first byte of parent block.
 345  *
 346  * When implicit back refs is used, information about the lowest key and
 347  * level of the tree block are required. These information are stored in
 348  * tree block info structure.
 349  */
 350
 351 /*
 352  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 353  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 354  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 355  */
 356 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 357                                      struct btrfs_extent_inline_ref *iref,
 358                                      enum btrfs_inline_ref_type is_data)
 359 {
 360         int type = btrfs_extent_inline_ref_type(eb, iref);
 361         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 362
 363         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 364             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 365             type == BTRFS_SHARED_DATA_REF_KEY ||
 366             type == BTRFS_EXTENT_DATA_REF_KEY) {
 367                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 368                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 369                                 return type;
 370                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 371                                 ASSERT(eb->fs_info);
 372                                 /*
 373                                  * Every shared one has parent tree block,
 374                                  * which must be aligned to sector size.
 375                                  */
 376                                 if (offset &&
 377                                     IS_ALIGNED(offset, eb->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(eb->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, eb->fs_info->sectorsize))
 391                                         return type;
 392                         }
 393                 } else {
 394                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 395                         return type;
 396                 }
 397         }
 398
 399         btrfs_print_leaf((struct extent_buffer *)eb);
 400         btrfs_err(eb->fs_info,
 401                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 402                   eb->start, (unsigned long)iref, type);
 403         WARN_ON(1);
 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 = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 416         lenum = cpu_to_le64(owner);
 417         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 418         lenum = cpu_to_le64(offset);
 419         low_crc = btrfs_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(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 if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 622                 btrfs_print_v0_err(trans->fs_info);
 623                 btrfs_abort_transaction(trans, -EINVAL);
 624                 return -EINVAL;
 625         } else {
 626                 BUG();
 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(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         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 658         if (iref) {
 659                 /*
 660                  * If type is invalid, we should have bailed out earlier than
 661                  * this call.
 662                  */
 663                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 664                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 665                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 666                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 667                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 668                 } else {
 669                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 670                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 671                 }
 672         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 673                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 674                                       struct btrfs_extent_data_ref);
 675                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 676         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 677                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 678                                       struct btrfs_shared_data_ref);
 679                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 680         } else {
 681                 WARN_ON(1);
 682         }
 683         return num_refs;
 684 }
 685
 686 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 687                                           struct btrfs_path *path,
 688                                           u64 bytenr, u64 parent,
 689                                           u64 root_objectid)
 690 {
 691         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 692         struct btrfs_key key;
 693         int ret;
 694
 695         key.objectid = bytenr;
 696         if (parent) {
 697                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 698                 key.offset = parent;
 699         } else {
 700                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 701                 key.offset = root_objectid;
 702         }
 703
 704         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 705         if (ret > 0)
 706                 ret = -ENOENT;
 707         return ret;
 708 }
 709
 710 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 711                                           struct btrfs_path *path,
 712                                           u64 bytenr, u64 parent,
 713                                           u64 root_objectid)
 714 {
 715         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 716         struct btrfs_key key;
 717         int ret;
 718
 719         key.objectid = bytenr;
 720         if (parent) {
 721                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 722                 key.offset = parent;
 723         } else {
 724                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 725                 key.offset = root_objectid;
 726         }
 727
 728         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 729         btrfs_release_path(path);
 730         return ret;
 731 }
 732
 733 static inline int extent_ref_type(u64 parent, u64 owner)
 734 {
 735         int type;
 736         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 737                 if (parent > 0)
 738                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 739                 else
 740                         type = BTRFS_TREE_BLOCK_REF_KEY;
 741         } else {
 742                 if (parent > 0)
 743                         type = BTRFS_SHARED_DATA_REF_KEY;
 744                 else
 745                         type = BTRFS_EXTENT_DATA_REF_KEY;
 746         }
 747         return type;
 748 }
 749
 750 static int find_next_key(struct btrfs_path *path, int level,
 751                          struct btrfs_key *key)
 752
 753 {
 754         for (; level < BTRFS_MAX_LEVEL; level++) {
 755                 if (!path->nodes[level])
 756                         break;
 757                 if (path->slots[level] + 1 >=
 758                     btrfs_header_nritems(path->nodes[level]))
 759                         continue;
 760                 if (level == 0)
 761                         btrfs_item_key_to_cpu(path->nodes[level], key,
 762                                               path->slots[level] + 1);
 763                 else
 764                         btrfs_node_key_to_cpu(path->nodes[level], key,
 765                                               path->slots[level] + 1);
 766                 return 0;
 767         }
 768         return 1;
 769 }
 770
 771 /*
 772  * look for inline back ref. if back ref is found, *ref_ret is set
 773  * to the address of inline back ref, and 0 is returned.
 774  *
 775  * if back ref isn't found, *ref_ret is set to the address where it
 776  * should be inserted, and -ENOENT is returned.
 777  *
 778  * if insert is true and there are too many inline back refs, the path
 779  * points to the extent item, and -EAGAIN is returned.
 780  *
 781  * NOTE: inline back refs are ordered in the same way that back ref
 782  *       items in the tree are ordered.
 783  */
 784 static noinline_for_stack
 785 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 786                                  struct btrfs_path *path,
 787                                  struct btrfs_extent_inline_ref **ref_ret,
 788                                  u64 bytenr, u64 num_bytes,
 789                                  u64 parent, u64 root_objectid,
 790                                  u64 owner, u64 offset, int insert)
 791 {
 792         struct btrfs_fs_info *fs_info = trans->fs_info;
 793         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 794         struct btrfs_key key;
 795         struct extent_buffer *leaf;
 796         struct btrfs_extent_item *ei;
 797         struct btrfs_extent_inline_ref *iref;
 798         u64 flags;
 799         u64 item_size;
 800         unsigned long ptr;
 801         unsigned long end;
 802         int extra_size;
 803         int type;
 804         int want;
 805         int ret;
 806         int err = 0;
 807         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 808         int needed;
 809
 810         key.objectid = bytenr;
 811         key.type = BTRFS_EXTENT_ITEM_KEY;
 812         key.offset = num_bytes;
 813
 814         want = extent_ref_type(parent, owner);
 815         if (insert) {
 816                 extra_size = btrfs_extent_inline_ref_size(want);
 817                 path->search_for_extension = 1;
 818                 path->keep_locks = 1;
 819         } else
 820                 extra_size = -1;
 821
 822         /*
 823          * Owner is our level, so we can just add one to get the level for the
 824          * block we are interested in.
 825          */
 826         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 827                 key.type = BTRFS_METADATA_ITEM_KEY;
 828                 key.offset = owner;
 829         }
 830
 831 again:
 832         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 833         if (ret < 0) {
 834                 err = ret;
 835                 goto out;
 836         }
 837
 838         /*
 839          * We may be a newly converted file system which still has the old fat
 840          * extent entries for metadata, so try and see if we have one of those.
 841          */
 842         if (ret > 0 && skinny_metadata) {
 843                 skinny_metadata = false;
 844                 if (path->slots[0]) {
 845                         path->slots[0]--;
 846                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 847                                               path->slots[0]);
 848                         if (key.objectid == bytenr &&
 849                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 850                             key.offset == num_bytes)
 851                                 ret = 0;
 852                 }
 853                 if (ret) {
 854                         key.objectid = bytenr;
 855                         key.type = BTRFS_EXTENT_ITEM_KEY;
 856                         key.offset = num_bytes;
 857                         btrfs_release_path(path);
 858                         goto again;
 859                 }
 860         }
 861
 862         if (ret && !insert) {
 863                 err = -ENOENT;
 864                 goto out;
 865         } else if (WARN_ON(ret)) {
 866                 err = -EIO;
 867                 goto out;
 868         }
 869
 870         leaf = path->nodes[0];
 871         item_size = btrfs_item_size(leaf, path->slots[0]);
 872         if (unlikely(item_size < sizeof(*ei))) {
 873                 err = -EINVAL;
 874                 btrfs_print_v0_err(fs_info);
 875                 btrfs_abort_transaction(trans, err);
 876                 goto out;
 877         }
 878
 879         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 880         flags = btrfs_extent_flags(leaf, ei);
 881
 882         ptr = (unsigned long)(ei + 1);
 883         end = (unsigned long)ei + item_size;
 884
 885         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 886                 ptr += sizeof(struct btrfs_tree_block_info);
 887                 BUG_ON(ptr > end);
 888         }
 889
 890         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 891                 needed = BTRFS_REF_TYPE_DATA;
 892         else
 893                 needed = BTRFS_REF_TYPE_BLOCK;
 894
 895         err = -ENOENT;
 896         while (1) {
 897                 if (ptr >= end) {
 898                         if (ptr > end) {
 899                                 err = -EUCLEAN;
 900                                 btrfs_print_leaf(path->nodes[0]);
 901                                 btrfs_crit(fs_info,
 902 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 903                                         path->slots[0], root_objectid, owner, offset, parent);
 904                         }
 905                         break;
 906                 }
 907                 iref = (struct btrfs_extent_inline_ref *)ptr;
 908                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 909                 if (type == BTRFS_REF_TYPE_INVALID) {
 910                         err = -EUCLEAN;
 911                         goto out;
 912                 }
 913
 914                 if (want < type)
 915                         break;
 916                 if (want > type) {
 917                         ptr += btrfs_extent_inline_ref_size(type);
 918                         continue;
 919                 }
 920
 921                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 922                         struct btrfs_extent_data_ref *dref;
 923                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 924                         if (match_extent_data_ref(leaf, dref, root_objectid,
 925                                                   owner, offset)) {
 926                                 err = 0;
 927                                 break;
 928                         }
 929                         if (hash_extent_data_ref_item(leaf, dref) <
 930                             hash_extent_data_ref(root_objectid, owner, offset))
 931                                 break;
 932                 } else {
 933                         u64 ref_offset;
 934                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 935                         if (parent > 0) {
 936                                 if (parent == ref_offset) {
 937                                         err = 0;
 938                                         break;
 939                                 }
 940                                 if (ref_offset < parent)
 941                                         break;
 942                         } else {
 943                                 if (root_objectid == ref_offset) {
 944                                         err = 0;
 945                                         break;
 946                                 }
 947                                 if (ref_offset < root_objectid)
 948                                         break;
 949                         }
 950                 }
 951                 ptr += btrfs_extent_inline_ref_size(type);
 952         }
 953         if (err == -ENOENT && insert) {
 954                 if (item_size + extra_size >=
 955                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 956                         err = -EAGAIN;
 957                         goto out;
 958                 }
 959                 /*
 960                  * To add new inline back ref, we have to make sure
 961                  * there is no corresponding back ref item.
 962                  * For simplicity, we just do not add new inline back
 963                  * ref if there is any kind of item for this block
 964                  */
 965                 if (find_next_key(path, 0, &key) == 0 &&
 966                     key.objectid == bytenr &&
 967                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 968                         err = -EAGAIN;
 969                         goto out;
 970                 }
 971         }
 972         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 973 out:
 974         if (insert) {
 975                 path->keep_locks = 0;
 976                 path->search_for_extension = 0;
 977                 btrfs_unlock_up_safe(path, 1);
 978         }
 979         return err;
 980 }
 981
 982 /*
 983  * helper to add new inline back ref
 984  */
 985 static noinline_for_stack
 986 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
 987                                  struct btrfs_path *path,
 988                                  struct btrfs_extent_inline_ref *iref,
 989                                  u64 parent, u64 root_objectid,
 990                                  u64 owner, u64 offset, int refs_to_add,
 991                                  struct btrfs_delayed_extent_op *extent_op)
 992 {
 993         struct extent_buffer *leaf;
 994         struct btrfs_extent_item *ei;
 995         unsigned long ptr;
 996         unsigned long end;
 997         unsigned long item_offset;
 998         u64 refs;
 999         int size;
1000         int type;
1001
1002         leaf = path->nodes[0];
1003         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1004         item_offset = (unsigned long)iref - (unsigned long)ei;
1005
1006         type = extent_ref_type(parent, owner);
1007         size = btrfs_extent_inline_ref_size(type);
1008
1009         btrfs_extend_item(path, size);
1010
1011         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1012         refs = btrfs_extent_refs(leaf, ei);
1013         refs += refs_to_add;
1014         btrfs_set_extent_refs(leaf, ei, refs);
1015         if (extent_op)
1016                 __run_delayed_extent_op(extent_op, leaf, ei);
1017
1018         ptr = (unsigned long)ei + item_offset;
1019         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1020         if (ptr < end - size)
1021                 memmove_extent_buffer(leaf, ptr + size, ptr,
1022                                       end - size - ptr);
1023
1024         iref = (struct btrfs_extent_inline_ref *)ptr;
1025         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1026         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1027                 struct btrfs_extent_data_ref *dref;
1028                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1029                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1030                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1031                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1032                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1033         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1034                 struct btrfs_shared_data_ref *sref;
1035                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1036                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1037                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1038         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1039                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1040         } else {
1041                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1042         }
1043         btrfs_mark_buffer_dirty(leaf);
1044 }
1045
1046 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1047                                  struct btrfs_path *path,
1048                                  struct btrfs_extent_inline_ref **ref_ret,
1049                                  u64 bytenr, u64 num_bytes, u64 parent,
1050                                  u64 root_objectid, u64 owner, u64 offset)
1051 {
1052         int ret;
1053
1054         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1055                                            num_bytes, parent, root_objectid,
1056                                            owner, offset, 0);
1057         if (ret != -ENOENT)
1058                 return ret;
1059
1060         btrfs_release_path(path);
1061         *ref_ret = NULL;
1062
1063         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1064                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1065                                             root_objectid);
1066         } else {
1067                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1068                                              root_objectid, owner, offset);
1069         }
1070         return ret;
1071 }
1072
1073 /*
1074  * helper to update/remove inline back ref
1075  */
1076 static noinline_for_stack
1077 void update_inline_extent_backref(struct btrfs_path *path,
1078                                   struct btrfs_extent_inline_ref *iref,
1079                                   int refs_to_mod,
1080                                   struct btrfs_delayed_extent_op *extent_op)
1081 {
1082         struct extent_buffer *leaf = path->nodes[0];
1083         struct btrfs_extent_item *ei;
1084         struct btrfs_extent_data_ref *dref = NULL;
1085         struct btrfs_shared_data_ref *sref = NULL;
1086         unsigned long ptr;
1087         unsigned long end;
1088         u32 item_size;
1089         int size;
1090         int type;
1091         u64 refs;
1092
1093         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1094         refs = btrfs_extent_refs(leaf, ei);
1095         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1096         refs += refs_to_mod;
1097         btrfs_set_extent_refs(leaf, ei, refs);
1098         if (extent_op)
1099                 __run_delayed_extent_op(extent_op, leaf, ei);
1100
1101         /*
1102          * If type is invalid, we should have bailed out after
1103          * lookup_inline_extent_backref().
1104          */
1105         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1106         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1107
1108         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1109                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1110                 refs = btrfs_extent_data_ref_count(leaf, dref);
1111         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1112                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1113                 refs = btrfs_shared_data_ref_count(leaf, sref);
1114         } else {
1115                 refs = 1;
1116                 BUG_ON(refs_to_mod != -1);
1117         }
1118
1119         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1120         refs += refs_to_mod;
1121
1122         if (refs > 0) {
1123                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1124                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1125                 else
1126                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1127         } else {
1128                 size =  btrfs_extent_inline_ref_size(type);
1129                 item_size = btrfs_item_size(leaf, path->slots[0]);
1130                 ptr = (unsigned long)iref;
1131                 end = (unsigned long)ei + item_size;
1132                 if (ptr + size < end)
1133                         memmove_extent_buffer(leaf, ptr, ptr + size,
1134                                               end - ptr - size);
1135                 item_size -= size;
1136                 btrfs_truncate_item(path, item_size, 1);
1137         }
1138         btrfs_mark_buffer_dirty(leaf);
1139 }
1140
1141 static noinline_for_stack
1142 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1143                                  struct btrfs_path *path,
1144                                  u64 bytenr, u64 num_bytes, u64 parent,
1145                                  u64 root_objectid, u64 owner,
1146                                  u64 offset, int refs_to_add,
1147                                  struct btrfs_delayed_extent_op *extent_op)
1148 {
1149         struct btrfs_extent_inline_ref *iref;
1150         int ret;
1151
1152         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1153                                            num_bytes, parent, root_objectid,
1154                                            owner, offset, 1);
1155         if (ret == 0) {
1156                 /*
1157                  * We're adding refs to a tree block we already own, this
1158                  * should not happen at all.
1159                  */
1160                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1161                         btrfs_crit(trans->fs_info,
1162 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu",
1163                                    bytenr, num_bytes, root_objectid);
1164                         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
1165                                 WARN_ON(1);
1166                                 btrfs_crit(trans->fs_info,
1167                         "path->slots[0]=%d path->nodes[0]:", path->slots[0]);
1168                                 btrfs_print_leaf(path->nodes[0]);
1169                         }
1170                         return -EUCLEAN;
1171                 }
1172                 update_inline_extent_backref(path, iref, refs_to_add, extent_op);
1173         } else if (ret == -ENOENT) {
1174                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1175                                             root_objectid, owner, offset,
1176                                             refs_to_add, extent_op);
1177                 ret = 0;
1178         }
1179         return ret;
1180 }
1181
1182 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1183                                  struct btrfs_root *root,
1184                                  struct btrfs_path *path,
1185                                  struct btrfs_extent_inline_ref *iref,
1186                                  int refs_to_drop, int is_data)
1187 {
1188         int ret = 0;
1189
1190         BUG_ON(!is_data && refs_to_drop != 1);
1191         if (iref)
1192                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL);
1193         else if (is_data)
1194                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1195         else
1196                 ret = btrfs_del_item(trans, root, path);
1197         return ret;
1198 }
1199
1200 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1201                                u64 *discarded_bytes)
1202 {
1203         int j, ret = 0;
1204         u64 bytes_left, end;
1205         u64 aligned_start = ALIGN(start, 1 << 9);
1206
1207         if (WARN_ON(start != aligned_start)) {
1208                 len -= aligned_start - start;
1209                 len = round_down(len, 1 << 9);
1210                 start = aligned_start;
1211         }
1212
1213         *discarded_bytes = 0;
1214
1215         if (!len)
1216                 return 0;
1217
1218         end = start + len;
1219         bytes_left = len;
1220
1221         /* Skip any superblocks on this device. */
1222         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1223                 u64 sb_start = btrfs_sb_offset(j);
1224                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1225                 u64 size = sb_start - start;
1226
1227                 if (!in_range(sb_start, start, bytes_left) &&
1228                     !in_range(sb_end, start, bytes_left) &&
1229                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1230                         continue;
1231
1232                 /*
1233                  * Superblock spans beginning of range.  Adjust start and
1234                  * try again.
1235                  */
1236                 if (sb_start <= start) {
1237                         start += sb_end - start;
1238                         if (start > end) {
1239                                 bytes_left = 0;
1240                                 break;
1241                         }
1242                         bytes_left = end - start;
1243                         continue;
1244                 }
1245
1246                 if (size) {
1247                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1248                                                    GFP_NOFS);
1249                         if (!ret)
1250                                 *discarded_bytes += size;
1251                         else if (ret != -EOPNOTSUPP)
1252                                 return ret;
1253                 }
1254
1255                 start = sb_end;
1256                 if (start > end) {
1257                         bytes_left = 0;
1258                         break;
1259                 }
1260                 bytes_left = end - start;
1261         }
1262
1263         if (bytes_left) {
1264                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1265                                            GFP_NOFS);
1266                 if (!ret)
1267                         *discarded_bytes += bytes_left;
1268         }
1269         return ret;
1270 }
1271
1272 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1273 {
1274         struct btrfs_device *dev = stripe->dev;
1275         struct btrfs_fs_info *fs_info = dev->fs_info;
1276         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1277         u64 phys = stripe->physical;
1278         u64 len = stripe->length;
1279         u64 discarded = 0;
1280         int ret = 0;
1281
1282         /* Zone reset on a zoned filesystem */
1283         if (btrfs_can_zone_reset(dev, phys, len)) {
1284                 u64 src_disc;
1285
1286                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1287                 if (ret)
1288                         goto out;
1289
1290                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1291                     dev != dev_replace->srcdev)
1292                         goto out;
1293
1294                 src_disc = discarded;
1295
1296                 /* Send to replace target as well */
1297                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1298                                               &discarded);
1299                 discarded += src_disc;
1300         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1301                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1302         } else {
1303                 ret = 0;
1304                 *bytes = 0;
1305         }
1306
1307 out:
1308         *bytes = discarded;
1309         return ret;
1310 }
1311
1312 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1313                          u64 num_bytes, u64 *actual_bytes)
1314 {
1315         int ret = 0;
1316         u64 discarded_bytes = 0;
1317         u64 end = bytenr + num_bytes;
1318         u64 cur = bytenr;
1319
1320         /*
1321          * Avoid races with device replace and make sure the devices in the
1322          * stripes don't go away while we are discarding.
1323          */
1324         btrfs_bio_counter_inc_blocked(fs_info);
1325         while (cur < end) {
1326                 struct btrfs_discard_stripe *stripes;
1327                 unsigned int num_stripes;
1328                 int i;
1329
1330                 num_bytes = end - cur;
1331                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1332                 if (IS_ERR(stripes)) {
1333                         ret = PTR_ERR(stripes);
1334                         if (ret == -EOPNOTSUPP)
1335                                 ret = 0;
1336                         break;
1337                 }
1338
1339                 for (i = 0; i < num_stripes; i++) {
1340                         struct btrfs_discard_stripe *stripe = stripes + i;
1341                         u64 bytes;
1342
1343                         if (!stripe->dev->bdev) {
1344                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1345                                 continue;
1346                         }
1347
1348                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1349                                         &stripe->dev->dev_state))
1350                                 continue;
1351
1352                         ret = do_discard_extent(stripe, &bytes);
1353                         if (ret) {
1354                                 /*
1355                                  * Keep going if discard is not supported by the
1356                                  * device.
1357                                  */
1358                                 if (ret != -EOPNOTSUPP)
1359                                         break;
1360                                 ret = 0;
1361                         } else {
1362                                 discarded_bytes += bytes;
1363                         }
1364                 }
1365                 kfree(stripes);
1366                 if (ret)
1367                         break;
1368                 cur += num_bytes;
1369         }
1370         btrfs_bio_counter_dec(fs_info);
1371         if (actual_bytes)
1372                 *actual_bytes = discarded_bytes;
1373         return ret;
1374 }
1375
1376 /* Can return -ENOMEM */
1377 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1378                          struct btrfs_ref *generic_ref)
1379 {
1380         struct btrfs_fs_info *fs_info = trans->fs_info;
1381         int ret;
1382
1383         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1384                generic_ref->action);
1385         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1386                generic_ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID);
1387
1388         if (generic_ref->type == BTRFS_REF_METADATA)
1389                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1390         else
1391                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1392
1393         btrfs_ref_tree_mod(fs_info, generic_ref);
1394
1395         return ret;
1396 }
1397
1398 /*
1399  * __btrfs_inc_extent_ref - insert backreference for a given extent
1400  *
1401  * The counterpart is in __btrfs_free_extent(), with examples and more details
1402  * how it works.
1403  *
1404  * @trans:          Handle of transaction
1405  *
1406  * @node:           The delayed ref node used to get the bytenr/length for
1407  *                  extent whose references are incremented.
1408  *
1409  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1410  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1411  *                  bytenr of the parent block. Since new extents are always
1412  *                  created with indirect references, this will only be the case
1413  *                  when relocating a shared extent. In that case, root_objectid
1414  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1415  *                  be 0
1416  *
1417  * @root_objectid:  The id of the root where this modification has originated,
1418  *                  this can be either one of the well-known metadata trees or
1419  *                  the subvolume id which references this extent.
1420  *
1421  * @owner:          For data extents it is the inode number of the owning file.
1422  *                  For metadata extents this parameter holds the level in the
1423  *                  tree of the extent.
1424  *
1425  * @offset:         For metadata extents the offset is ignored and is currently
1426  *                  always passed as 0. For data extents it is the fileoffset
1427  *                  this extent belongs to.
1428  *
1429  * @refs_to_add     Number of references to add
1430  *
1431  * @extent_op       Pointer to a structure, holding information necessary when
1432  *                  updating a tree block's flags
1433  *
1434  */
1435 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1436                                   struct btrfs_delayed_ref_node *node,
1437                                   u64 parent, u64 root_objectid,
1438                                   u64 owner, u64 offset, int refs_to_add,
1439                                   struct btrfs_delayed_extent_op *extent_op)
1440 {
1441         struct btrfs_path *path;
1442         struct extent_buffer *leaf;
1443         struct btrfs_extent_item *item;
1444         struct btrfs_key key;
1445         u64 bytenr = node->bytenr;
1446         u64 num_bytes = node->num_bytes;
1447         u64 refs;
1448         int ret;
1449
1450         path = btrfs_alloc_path();
1451         if (!path)
1452                 return -ENOMEM;
1453
1454         /* this will setup the path even if it fails to insert the back ref */
1455         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1456                                            parent, root_objectid, owner,
1457                                            offset, refs_to_add, extent_op);
1458         if ((ret < 0 && ret != -EAGAIN) || !ret)
1459                 goto out;
1460
1461         /*
1462          * Ok we had -EAGAIN which means we didn't have space to insert and
1463          * inline extent ref, so just update the reference count and add a
1464          * normal backref.
1465          */
1466         leaf = path->nodes[0];
1467         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1468         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1469         refs = btrfs_extent_refs(leaf, item);
1470         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1471         if (extent_op)
1472                 __run_delayed_extent_op(extent_op, leaf, item);
1473
1474         btrfs_mark_buffer_dirty(leaf);
1475         btrfs_release_path(path);
1476
1477         /* now insert the actual backref */
1478         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1479                 BUG_ON(refs_to_add != 1);
1480                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1481                                             root_objectid);
1482         } else {
1483                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1484                                              root_objectid, owner, offset,
1485                                              refs_to_add);
1486         }
1487         if (ret)
1488                 btrfs_abort_transaction(trans, ret);
1489 out:
1490         btrfs_free_path(path);
1491         return ret;
1492 }
1493
1494 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1495                                 struct btrfs_delayed_ref_node *node,
1496                                 struct btrfs_delayed_extent_op *extent_op,
1497                                 int insert_reserved)
1498 {
1499         int ret = 0;
1500         struct btrfs_delayed_data_ref *ref;
1501         struct btrfs_key ins;
1502         u64 parent = 0;
1503         u64 ref_root = 0;
1504         u64 flags = 0;
1505
1506         ins.objectid = node->bytenr;
1507         ins.offset = node->num_bytes;
1508         ins.type = BTRFS_EXTENT_ITEM_KEY;
1509
1510         ref = btrfs_delayed_node_to_data_ref(node);
1511         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1512
1513         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1514                 parent = ref->parent;
1515         ref_root = ref->root;
1516
1517         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1518                 if (extent_op)
1519                         flags |= extent_op->flags_to_set;
1520                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1521                                                  flags, ref->objectid,
1522                                                  ref->offset, &ins,
1523                                                  node->ref_mod);
1524         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1525                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1526                                              ref->objectid, ref->offset,
1527                                              node->ref_mod, extent_op);
1528         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1529                 ret = __btrfs_free_extent(trans, node, parent,
1530                                           ref_root, ref->objectid,
1531                                           ref->offset, node->ref_mod,
1532                                           extent_op);
1533         } else {
1534                 BUG();
1535         }
1536         return ret;
1537 }
1538
1539 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1540                                     struct extent_buffer *leaf,
1541                                     struct btrfs_extent_item *ei)
1542 {
1543         u64 flags = btrfs_extent_flags(leaf, ei);
1544         if (extent_op->update_flags) {
1545                 flags |= extent_op->flags_to_set;
1546                 btrfs_set_extent_flags(leaf, ei, flags);
1547         }
1548
1549         if (extent_op->update_key) {
1550                 struct btrfs_tree_block_info *bi;
1551                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1552                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1553                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1554         }
1555 }
1556
1557 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1558                                  struct btrfs_delayed_ref_head *head,
1559                                  struct btrfs_delayed_extent_op *extent_op)
1560 {
1561         struct btrfs_fs_info *fs_info = trans->fs_info;
1562         struct btrfs_root *root;
1563         struct btrfs_key key;
1564         struct btrfs_path *path;
1565         struct btrfs_extent_item *ei;
1566         struct extent_buffer *leaf;
1567         u32 item_size;
1568         int ret;
1569         int err = 0;
1570         int metadata = 1;
1571
1572         if (TRANS_ABORTED(trans))
1573                 return 0;
1574
1575         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1576                 metadata = 0;
1577
1578         path = btrfs_alloc_path();
1579         if (!path)
1580                 return -ENOMEM;
1581
1582         key.objectid = head->bytenr;
1583
1584         if (metadata) {
1585                 key.type = BTRFS_METADATA_ITEM_KEY;
1586                 key.offset = extent_op->level;
1587         } else {
1588                 key.type = BTRFS_EXTENT_ITEM_KEY;
1589                 key.offset = head->num_bytes;
1590         }
1591
1592         root = btrfs_extent_root(fs_info, key.objectid);
1593 again:
1594         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1595         if (ret < 0) {
1596                 err = ret;
1597                 goto out;
1598         }
1599         if (ret > 0) {
1600                 if (metadata) {
1601                         if (path->slots[0] > 0) {
1602                                 path->slots[0]--;
1603                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1604                                                       path->slots[0]);
1605                                 if (key.objectid == head->bytenr &&
1606                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1607                                     key.offset == head->num_bytes)
1608                                         ret = 0;
1609                         }
1610                         if (ret > 0) {
1611                                 btrfs_release_path(path);
1612                                 metadata = 0;
1613
1614                                 key.objectid = head->bytenr;
1615                                 key.offset = head->num_bytes;
1616                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1617                                 goto again;
1618                         }
1619                 } else {
1620                         err = -EIO;
1621                         goto out;
1622                 }
1623         }
1624
1625         leaf = path->nodes[0];
1626         item_size = btrfs_item_size(leaf, path->slots[0]);
1627
1628         if (unlikely(item_size < sizeof(*ei))) {
1629                 err = -EINVAL;
1630                 btrfs_print_v0_err(fs_info);
1631                 btrfs_abort_transaction(trans, err);
1632                 goto out;
1633         }
1634
1635         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1636         __run_delayed_extent_op(extent_op, leaf, ei);
1637
1638         btrfs_mark_buffer_dirty(leaf);
1639 out:
1640         btrfs_free_path(path);
1641         return err;
1642 }
1643
1644 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1645                                 struct btrfs_delayed_ref_node *node,
1646                                 struct btrfs_delayed_extent_op *extent_op,
1647                                 int insert_reserved)
1648 {
1649         int ret = 0;
1650         struct btrfs_delayed_tree_ref *ref;
1651         u64 parent = 0;
1652         u64 ref_root = 0;
1653
1654         ref = btrfs_delayed_node_to_tree_ref(node);
1655         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1656
1657         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1658                 parent = ref->parent;
1659         ref_root = ref->root;
1660
1661         if (node->ref_mod != 1) {
1662                 btrfs_err(trans->fs_info,
1663         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1664                           node->bytenr, node->ref_mod, node->action, ref_root,
1665                           parent);
1666                 return -EIO;
1667         }
1668         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1669                 BUG_ON(!extent_op || !extent_op->update_flags);
1670                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1671         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1672                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1673                                              ref->level, 0, 1, extent_op);
1674         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1675                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1676                                           ref->level, 0, 1, extent_op);
1677         } else {
1678                 BUG();
1679         }
1680         return ret;
1681 }
1682
1683 /* helper function to actually process a single delayed ref entry */
1684 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1685                                struct btrfs_delayed_ref_node *node,
1686                                struct btrfs_delayed_extent_op *extent_op,
1687                                int insert_reserved)
1688 {
1689         int ret = 0;
1690
1691         if (TRANS_ABORTED(trans)) {
1692                 if (insert_reserved)
1693                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1694                 return 0;
1695         }
1696
1697         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1698             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1699                 ret = run_delayed_tree_ref(trans, node, extent_op,
1700                                            insert_reserved);
1701         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1702                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1703                 ret = run_delayed_data_ref(trans, node, extent_op,
1704                                            insert_reserved);
1705         else
1706                 BUG();
1707         if (ret && insert_reserved)
1708                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1709         return ret;
1710 }
1711
1712 static inline struct btrfs_delayed_ref_node *
1713 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1714 {
1715         struct btrfs_delayed_ref_node *ref;
1716
1717         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1718                 return NULL;
1719
1720         /*
1721          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1722          * This is to prevent a ref count from going down to zero, which deletes
1723          * the extent item from the extent tree, when there still are references
1724          * to add, which would fail because they would not find the extent item.
1725          */
1726         if (!list_empty(&head->ref_add_list))
1727                 return list_first_entry(&head->ref_add_list,
1728                                 struct btrfs_delayed_ref_node, add_list);
1729
1730         ref = rb_entry(rb_first_cached(&head->ref_tree),
1731                        struct btrfs_delayed_ref_node, ref_node);
1732         ASSERT(list_empty(&ref->add_list));
1733         return ref;
1734 }
1735
1736 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1737                                       struct btrfs_delayed_ref_head *head)
1738 {
1739         spin_lock(&delayed_refs->lock);
1740         head->processing = 0;
1741         delayed_refs->num_heads_ready++;
1742         spin_unlock(&delayed_refs->lock);
1743         btrfs_delayed_ref_unlock(head);
1744 }
1745
1746 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1747                                 struct btrfs_delayed_ref_head *head)
1748 {
1749         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1750
1751         if (!extent_op)
1752                 return NULL;
1753
1754         if (head->must_insert_reserved) {
1755                 head->extent_op = NULL;
1756                 btrfs_free_delayed_extent_op(extent_op);
1757                 return NULL;
1758         }
1759         return extent_op;
1760 }
1761
1762 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1763                                      struct btrfs_delayed_ref_head *head)
1764 {
1765         struct btrfs_delayed_extent_op *extent_op;
1766         int ret;
1767
1768         extent_op = cleanup_extent_op(head);
1769         if (!extent_op)
1770                 return 0;
1771         head->extent_op = NULL;
1772         spin_unlock(&head->lock);
1773         ret = run_delayed_extent_op(trans, head, extent_op);
1774         btrfs_free_delayed_extent_op(extent_op);
1775         return ret ? ret : 1;
1776 }
1777
1778 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1779                                   struct btrfs_delayed_ref_root *delayed_refs,
1780                                   struct btrfs_delayed_ref_head *head)
1781 {
1782         int nr_items = 1;       /* Dropping this ref head update. */
1783
1784         /*
1785          * We had csum deletions accounted for in our delayed refs rsv, we need
1786          * to drop the csum leaves for this update from our delayed_refs_rsv.
1787          */
1788         if (head->total_ref_mod < 0 && head->is_data) {
1789                 spin_lock(&delayed_refs->lock);
1790                 delayed_refs->pending_csums -= head->num_bytes;
1791                 spin_unlock(&delayed_refs->lock);
1792                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1793         }
1794
1795         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1796 }
1797
1798 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1799                             struct btrfs_delayed_ref_head *head)
1800 {
1801
1802         struct btrfs_fs_info *fs_info = trans->fs_info;
1803         struct btrfs_delayed_ref_root *delayed_refs;
1804         int ret;
1805
1806         delayed_refs = &trans->transaction->delayed_refs;
1807
1808         ret = run_and_cleanup_extent_op(trans, head);
1809         if (ret < 0) {
1810                 unselect_delayed_ref_head(delayed_refs, head);
1811                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1812                 return ret;
1813         } else if (ret) {
1814                 return ret;
1815         }
1816
1817         /*
1818          * Need to drop our head ref lock and re-acquire the delayed ref lock
1819          * and then re-check to make sure nobody got added.
1820          */
1821         spin_unlock(&head->lock);
1822         spin_lock(&delayed_refs->lock);
1823         spin_lock(&head->lock);
1824         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1825                 spin_unlock(&head->lock);
1826                 spin_unlock(&delayed_refs->lock);
1827                 return 1;
1828         }
1829         btrfs_delete_ref_head(delayed_refs, head);
1830         spin_unlock(&head->lock);
1831         spin_unlock(&delayed_refs->lock);
1832
1833         if (head->must_insert_reserved) {
1834                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1835                 if (head->is_data) {
1836                         struct btrfs_root *csum_root;
1837
1838                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1839                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1840                                               head->num_bytes);
1841                 }
1842         }
1843
1844         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1845
1846         trace_run_delayed_ref_head(fs_info, head, 0);
1847         btrfs_delayed_ref_unlock(head);
1848         btrfs_put_delayed_ref_head(head);
1849         return ret;
1850 }
1851
1852 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1853                                         struct btrfs_trans_handle *trans)
1854 {
1855         struct btrfs_delayed_ref_root *delayed_refs =
1856                 &trans->transaction->delayed_refs;
1857         struct btrfs_delayed_ref_head *head = NULL;
1858         int ret;
1859
1860         spin_lock(&delayed_refs->lock);
1861         head = btrfs_select_ref_head(delayed_refs);
1862         if (!head) {
1863                 spin_unlock(&delayed_refs->lock);
1864                 return head;
1865         }
1866
1867         /*
1868          * Grab the lock that says we are going to process all the refs for
1869          * this head
1870          */
1871         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1872         spin_unlock(&delayed_refs->lock);
1873
1874         /*
1875          * We may have dropped the spin lock to get the head mutex lock, and
1876          * that might have given someone else time to free the head.  If that's
1877          * true, it has been removed from our list and we can move on.
1878          */
1879         if (ret == -EAGAIN)
1880                 head = ERR_PTR(-EAGAIN);
1881
1882         return head;
1883 }
1884
1885 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1886                                     struct btrfs_delayed_ref_head *locked_ref,
1887                                     unsigned long *run_refs)
1888 {
1889         struct btrfs_fs_info *fs_info = trans->fs_info;
1890         struct btrfs_delayed_ref_root *delayed_refs;
1891         struct btrfs_delayed_extent_op *extent_op;
1892         struct btrfs_delayed_ref_node *ref;
1893         int must_insert_reserved = 0;
1894         int ret;
1895
1896         delayed_refs = &trans->transaction->delayed_refs;
1897
1898         lockdep_assert_held(&locked_ref->mutex);
1899         lockdep_assert_held(&locked_ref->lock);
1900
1901         while ((ref = select_delayed_ref(locked_ref))) {
1902                 if (ref->seq &&
1903                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1904                         spin_unlock(&locked_ref->lock);
1905                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1906                         return -EAGAIN;
1907                 }
1908
1909                 (*run_refs)++;
1910                 ref->in_tree = 0;
1911                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1912                 RB_CLEAR_NODE(&ref->ref_node);
1913                 if (!list_empty(&ref->add_list))
1914                         list_del(&ref->add_list);
1915                 /*
1916                  * When we play the delayed ref, also correct the ref_mod on
1917                  * head
1918                  */
1919                 switch (ref->action) {
1920                 case BTRFS_ADD_DELAYED_REF:
1921                 case BTRFS_ADD_DELAYED_EXTENT:
1922                         locked_ref->ref_mod -= ref->ref_mod;
1923                         break;
1924                 case BTRFS_DROP_DELAYED_REF:
1925                         locked_ref->ref_mod += ref->ref_mod;
1926                         break;
1927                 default:
1928                         WARN_ON(1);
1929                 }
1930                 atomic_dec(&delayed_refs->num_entries);
1931
1932                 /*
1933                  * Record the must_insert_reserved flag before we drop the
1934                  * spin lock.
1935                  */
1936                 must_insert_reserved = locked_ref->must_insert_reserved;
1937                 locked_ref->must_insert_reserved = 0;
1938
1939                 extent_op = locked_ref->extent_op;
1940                 locked_ref->extent_op = NULL;
1941                 spin_unlock(&locked_ref->lock);
1942
1943                 ret = run_one_delayed_ref(trans, ref, extent_op,
1944                                           must_insert_reserved);
1945
1946                 btrfs_free_delayed_extent_op(extent_op);
1947                 if (ret) {
1948                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1949                         btrfs_put_delayed_ref(ref);
1950                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
1951                                     ret);
1952                         return ret;
1953                 }
1954
1955                 btrfs_put_delayed_ref(ref);
1956                 cond_resched();
1957
1958                 spin_lock(&locked_ref->lock);
1959                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1960         }
1961
1962         return 0;
1963 }
1964
1965 /*
1966  * Returns 0 on success or if called with an already aborted transaction.
1967  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1968  */
1969 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1970                                              unsigned long nr)
1971 {
1972         struct btrfs_fs_info *fs_info = trans->fs_info;
1973         struct btrfs_delayed_ref_root *delayed_refs;
1974         struct btrfs_delayed_ref_head *locked_ref = NULL;
1975         ktime_t start = ktime_get();
1976         int ret;
1977         unsigned long count = 0;
1978         unsigned long actual_count = 0;
1979
1980         delayed_refs = &trans->transaction->delayed_refs;
1981         do {
1982                 if (!locked_ref) {
1983                         locked_ref = btrfs_obtain_ref_head(trans);
1984                         if (IS_ERR_OR_NULL(locked_ref)) {
1985                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
1986                                         continue;
1987                                 } else {
1988                                         break;
1989                                 }
1990                         }
1991                         count++;
1992                 }
1993                 /*
1994                  * We need to try and merge add/drops of the same ref since we
1995                  * can run into issues with relocate dropping the implicit ref
1996                  * and then it being added back again before the drop can
1997                  * finish.  If we merged anything we need to re-loop so we can
1998                  * get a good ref.
1999                  * Or we can get node references of the same type that weren't
2000                  * merged when created due to bumps in the tree mod seq, and
2001                  * we need to merge them to prevent adding an inline extent
2002                  * backref before dropping it (triggering a BUG_ON at
2003                  * insert_inline_extent_backref()).
2004                  */
2005                 spin_lock(&locked_ref->lock);
2006                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2007
2008                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2009                                                       &actual_count);
2010                 if (ret < 0 && ret != -EAGAIN) {
2011                         /*
2012                          * Error, btrfs_run_delayed_refs_for_head already
2013                          * unlocked everything so just bail out
2014                          */
2015                         return ret;
2016                 } else if (!ret) {
2017                         /*
2018                          * Success, perform the usual cleanup of a processed
2019                          * head
2020                          */
2021                         ret = cleanup_ref_head(trans, locked_ref);
2022                         if (ret > 0 ) {
2023                                 /* We dropped our lock, we need to loop. */
2024                                 ret = 0;
2025                                 continue;
2026                         } else if (ret) {
2027                                 return ret;
2028                         }
2029                 }
2030
2031                 /*
2032                  * Either success case or btrfs_run_delayed_refs_for_head
2033                  * returned -EAGAIN, meaning we need to select another head
2034                  */
2035
2036                 locked_ref = NULL;
2037                 cond_resched();
2038         } while ((nr != -1 && count < nr) || locked_ref);
2039
2040         /*
2041          * We don't want to include ref heads since we can have empty ref heads
2042          * and those will drastically skew our runtime down since we just do
2043          * accounting, no actual extent tree updates.
2044          */
2045         if (actual_count > 0) {
2046                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2047                 u64 avg;
2048
2049                 /*
2050                  * We weigh the current average higher than our current runtime
2051                  * to avoid large swings in the average.
2052                  */
2053                 spin_lock(&delayed_refs->lock);
2054                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2055                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2056                 spin_unlock(&delayed_refs->lock);
2057         }
2058         return 0;
2059 }
2060
2061 #ifdef SCRAMBLE_DELAYED_REFS
2062 /*
2063  * Normally delayed refs get processed in ascending bytenr order. This
2064  * correlates in most cases to the order added. To expose dependencies on this
2065  * order, we start to process the tree in the middle instead of the beginning
2066  */
2067 static u64 find_middle(struct rb_root *root)
2068 {
2069         struct rb_node *n = root->rb_node;
2070         struct btrfs_delayed_ref_node *entry;
2071         int alt = 1;
2072         u64 middle;
2073         u64 first = 0, last = 0;
2074
2075         n = rb_first(root);
2076         if (n) {
2077                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2078                 first = entry->bytenr;
2079         }
2080         n = rb_last(root);
2081         if (n) {
2082                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2083                 last = entry->bytenr;
2084         }
2085         n = root->rb_node;
2086
2087         while (n) {
2088                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2089                 WARN_ON(!entry->in_tree);
2090
2091                 middle = entry->bytenr;
2092
2093                 if (alt)
2094                         n = n->rb_left;
2095                 else
2096                         n = n->rb_right;
2097
2098                 alt = 1 - alt;
2099         }
2100         return middle;
2101 }
2102 #endif
2103
2104 /*
2105  * this starts processing the delayed reference count updates and
2106  * extent insertions we have queued up so far.  count can be
2107  * 0, which means to process everything in the tree at the start
2108  * of the run (but not newly added entries), or it can be some target
2109  * number you'd like to process.
2110  *
2111  * Returns 0 on success or if called with an aborted transaction
2112  * Returns <0 on error and aborts the transaction
2113  */
2114 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2115                            unsigned long count)
2116 {
2117         struct btrfs_fs_info *fs_info = trans->fs_info;
2118         struct rb_node *node;
2119         struct btrfs_delayed_ref_root *delayed_refs;
2120         struct btrfs_delayed_ref_head *head;
2121         int ret;
2122         int run_all = count == (unsigned long)-1;
2123
2124         /* We'll clean this up in btrfs_cleanup_transaction */
2125         if (TRANS_ABORTED(trans))
2126                 return 0;
2127
2128         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2129                 return 0;
2130
2131         delayed_refs = &trans->transaction->delayed_refs;
2132         if (count == 0)
2133                 count = delayed_refs->num_heads_ready;
2134
2135 again:
2136 #ifdef SCRAMBLE_DELAYED_REFS
2137         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2138 #endif
2139         ret = __btrfs_run_delayed_refs(trans, count);
2140         if (ret < 0) {
2141                 btrfs_abort_transaction(trans, ret);
2142                 return ret;
2143         }
2144
2145         if (run_all) {
2146                 btrfs_create_pending_block_groups(trans);
2147
2148                 spin_lock(&delayed_refs->lock);
2149                 node = rb_first_cached(&delayed_refs->href_root);
2150                 if (!node) {
2151                         spin_unlock(&delayed_refs->lock);
2152                         goto out;
2153                 }
2154                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2155                                 href_node);
2156                 refcount_inc(&head->refs);
2157                 spin_unlock(&delayed_refs->lock);
2158
2159                 /* Mutex was contended, block until it's released and retry. */
2160                 mutex_lock(&head->mutex);
2161                 mutex_unlock(&head->mutex);
2162
2163                 btrfs_put_delayed_ref_head(head);
2164                 cond_resched();
2165                 goto again;
2166         }
2167 out:
2168         return 0;
2169 }
2170
2171 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2172                                 struct extent_buffer *eb, u64 flags,
2173                                 int level)
2174 {
2175         struct btrfs_delayed_extent_op *extent_op;
2176         int ret;
2177
2178         extent_op = btrfs_alloc_delayed_extent_op();
2179         if (!extent_op)
2180                 return -ENOMEM;
2181
2182         extent_op->flags_to_set = flags;
2183         extent_op->update_flags = true;
2184         extent_op->update_key = false;
2185         extent_op->level = level;
2186
2187         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2188         if (ret)
2189                 btrfs_free_delayed_extent_op(extent_op);
2190         return ret;
2191 }
2192
2193 static noinline int check_delayed_ref(struct btrfs_root *root,
2194                                       struct btrfs_path *path,
2195                                       u64 objectid, u64 offset, u64 bytenr)
2196 {
2197         struct btrfs_delayed_ref_head *head;
2198         struct btrfs_delayed_ref_node *ref;
2199         struct btrfs_delayed_data_ref *data_ref;
2200         struct btrfs_delayed_ref_root *delayed_refs;
2201         struct btrfs_transaction *cur_trans;
2202         struct rb_node *node;
2203         int ret = 0;
2204
2205         spin_lock(&root->fs_info->trans_lock);
2206         cur_trans = root->fs_info->running_transaction;
2207         if (cur_trans)
2208                 refcount_inc(&cur_trans->use_count);
2209         spin_unlock(&root->fs_info->trans_lock);
2210         if (!cur_trans)
2211                 return 0;
2212
2213         delayed_refs = &cur_trans->delayed_refs;
2214         spin_lock(&delayed_refs->lock);
2215         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2216         if (!head) {
2217                 spin_unlock(&delayed_refs->lock);
2218                 btrfs_put_transaction(cur_trans);
2219                 return 0;
2220         }
2221
2222         if (!mutex_trylock(&head->mutex)) {
2223                 if (path->nowait) {
2224                         spin_unlock(&delayed_refs->lock);
2225                         btrfs_put_transaction(cur_trans);
2226                         return -EAGAIN;
2227                 }
2228
2229                 refcount_inc(&head->refs);
2230                 spin_unlock(&delayed_refs->lock);
2231
2232                 btrfs_release_path(path);
2233
2234                 /*
2235                  * Mutex was contended, block until it's released and let
2236                  * caller try again
2237                  */
2238                 mutex_lock(&head->mutex);
2239                 mutex_unlock(&head->mutex);
2240                 btrfs_put_delayed_ref_head(head);
2241                 btrfs_put_transaction(cur_trans);
2242                 return -EAGAIN;
2243         }
2244         spin_unlock(&delayed_refs->lock);
2245
2246         spin_lock(&head->lock);
2247         /*
2248          * XXX: We should replace this with a proper search function in the
2249          * future.
2250          */
2251         for (node = rb_first_cached(&head->ref_tree); node;
2252              node = rb_next(node)) {
2253                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2254                 /* If it's a shared ref we know a cross reference exists */
2255                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2256                         ret = 1;
2257                         break;
2258                 }
2259
2260                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2261
2262                 /*
2263                  * If our ref doesn't match the one we're currently looking at
2264                  * then we have a cross reference.
2265                  */
2266                 if (data_ref->root != root->root_key.objectid ||
2267                     data_ref->objectid != objectid ||
2268                     data_ref->offset != offset) {
2269                         ret = 1;
2270                         break;
2271                 }
2272         }
2273         spin_unlock(&head->lock);
2274         mutex_unlock(&head->mutex);
2275         btrfs_put_transaction(cur_trans);
2276         return ret;
2277 }
2278
2279 static noinline int check_committed_ref(struct btrfs_root *root,
2280                                         struct btrfs_path *path,
2281                                         u64 objectid, u64 offset, u64 bytenr,
2282                                         bool strict)
2283 {
2284         struct btrfs_fs_info *fs_info = root->fs_info;
2285         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2286         struct extent_buffer *leaf;
2287         struct btrfs_extent_data_ref *ref;
2288         struct btrfs_extent_inline_ref *iref;
2289         struct btrfs_extent_item *ei;
2290         struct btrfs_key key;
2291         u32 item_size;
2292         int type;
2293         int ret;
2294
2295         key.objectid = bytenr;
2296         key.offset = (u64)-1;
2297         key.type = BTRFS_EXTENT_ITEM_KEY;
2298
2299         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2300         if (ret < 0)
2301                 goto out;
2302         BUG_ON(ret == 0); /* Corruption */
2303
2304         ret = -ENOENT;
2305         if (path->slots[0] == 0)
2306                 goto out;
2307
2308         path->slots[0]--;
2309         leaf = path->nodes[0];
2310         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2311
2312         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2313                 goto out;
2314
2315         ret = 1;
2316         item_size = btrfs_item_size(leaf, path->slots[0]);
2317         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2318
2319         /* If extent item has more than 1 inline ref then it's shared */
2320         if (item_size != sizeof(*ei) +
2321             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2322                 goto out;
2323
2324         /*
2325          * If extent created before last snapshot => it's shared unless the
2326          * snapshot has been deleted. Use the heuristic if strict is false.
2327          */
2328         if (!strict &&
2329             (btrfs_extent_generation(leaf, ei) <=
2330              btrfs_root_last_snapshot(&root->root_item)))
2331                 goto out;
2332
2333         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2334
2335         /* If this extent has SHARED_DATA_REF then it's shared */
2336         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2337         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2338                 goto out;
2339
2340         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2341         if (btrfs_extent_refs(leaf, ei) !=
2342             btrfs_extent_data_ref_count(leaf, ref) ||
2343             btrfs_extent_data_ref_root(leaf, ref) !=
2344             root->root_key.objectid ||
2345             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2346             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2347                 goto out;
2348
2349         ret = 0;
2350 out:
2351         return ret;
2352 }
2353
2354 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2355                           u64 bytenr, bool strict, struct btrfs_path *path)
2356 {
2357         int ret;
2358
2359         do {
2360                 ret = check_committed_ref(root, path, objectid,
2361                                           offset, bytenr, strict);
2362                 if (ret && ret != -ENOENT)
2363                         goto out;
2364
2365                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2366         } while (ret == -EAGAIN);
2367
2368 out:
2369         btrfs_release_path(path);
2370         if (btrfs_is_data_reloc_root(root))
2371                 WARN_ON(ret > 0);
2372         return ret;
2373 }
2374
2375 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2376                            struct btrfs_root *root,
2377                            struct extent_buffer *buf,
2378                            int full_backref, int inc)
2379 {
2380         struct btrfs_fs_info *fs_info = root->fs_info;
2381         u64 bytenr;
2382         u64 num_bytes;
2383         u64 parent;
2384         u64 ref_root;
2385         u32 nritems;
2386         struct btrfs_key key;
2387         struct btrfs_file_extent_item *fi;
2388         struct btrfs_ref generic_ref = { 0 };
2389         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2390         int i;
2391         int action;
2392         int level;
2393         int ret = 0;
2394
2395         if (btrfs_is_testing(fs_info))
2396                 return 0;
2397
2398         ref_root = btrfs_header_owner(buf);
2399         nritems = btrfs_header_nritems(buf);
2400         level = btrfs_header_level(buf);
2401
2402         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2403                 return 0;
2404
2405         if (full_backref)
2406                 parent = buf->start;
2407         else
2408                 parent = 0;
2409         if (inc)
2410                 action = BTRFS_ADD_DELAYED_REF;
2411         else
2412                 action = BTRFS_DROP_DELAYED_REF;
2413
2414         for (i = 0; i < nritems; i++) {
2415                 if (level == 0) {
2416                         btrfs_item_key_to_cpu(buf, &key, i);
2417                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2418                                 continue;
2419                         fi = btrfs_item_ptr(buf, i,
2420                                             struct btrfs_file_extent_item);
2421                         if (btrfs_file_extent_type(buf, fi) ==
2422                             BTRFS_FILE_EXTENT_INLINE)
2423                                 continue;
2424                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2425                         if (bytenr == 0)
2426                                 continue;
2427
2428                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2429                         key.offset -= btrfs_file_extent_offset(buf, fi);
2430                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2431                                                num_bytes, parent);
2432                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2433                                             key.offset, root->root_key.objectid,
2434                                             for_reloc);
2435                         if (inc)
2436                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2437                         else
2438                                 ret = btrfs_free_extent(trans, &generic_ref);
2439                         if (ret)
2440                                 goto fail;
2441                 } else {
2442                         bytenr = btrfs_node_blockptr(buf, i);
2443                         num_bytes = fs_info->nodesize;
2444                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2445                                                num_bytes, parent);
2446                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2447                                             root->root_key.objectid, for_reloc);
2448                         if (inc)
2449                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2450                         else
2451                                 ret = btrfs_free_extent(trans, &generic_ref);
2452                         if (ret)
2453                                 goto fail;
2454                 }
2455         }
2456         return 0;
2457 fail:
2458         return ret;
2459 }
2460
2461 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2462                   struct extent_buffer *buf, int full_backref)
2463 {
2464         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2465 }
2466
2467 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2468                   struct extent_buffer *buf, int full_backref)
2469 {
2470         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2471 }
2472
2473 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2474 {
2475         struct btrfs_fs_info *fs_info = root->fs_info;
2476         u64 flags;
2477         u64 ret;
2478
2479         if (data)
2480                 flags = BTRFS_BLOCK_GROUP_DATA;
2481         else if (root == fs_info->chunk_root)
2482                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2483         else
2484                 flags = BTRFS_BLOCK_GROUP_METADATA;
2485
2486         ret = btrfs_get_alloc_profile(fs_info, flags);
2487         return ret;
2488 }
2489
2490 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2491 {
2492         struct rb_node *leftmost;
2493         u64 bytenr = 0;
2494
2495         read_lock(&fs_info->block_group_cache_lock);
2496         /* Get the block group with the lowest logical start address. */
2497         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2498         if (leftmost) {
2499                 struct btrfs_block_group *bg;
2500
2501                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2502                 bytenr = bg->start;
2503         }
2504         read_unlock(&fs_info->block_group_cache_lock);
2505
2506         return bytenr;
2507 }
2508
2509 static int pin_down_extent(struct btrfs_trans_handle *trans,
2510                            struct btrfs_block_group *cache,
2511                            u64 bytenr, u64 num_bytes, int reserved)
2512 {
2513         struct btrfs_fs_info *fs_info = cache->fs_info;
2514
2515         spin_lock(&cache->space_info->lock);
2516         spin_lock(&cache->lock);
2517         cache->pinned += num_bytes;
2518         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2519                                              num_bytes);
2520         if (reserved) {
2521                 cache->reserved -= num_bytes;
2522                 cache->space_info->bytes_reserved -= num_bytes;
2523         }
2524         spin_unlock(&cache->lock);
2525         spin_unlock(&cache->space_info->lock);
2526
2527         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2528                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2529         return 0;
2530 }
2531
2532 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2533                      u64 bytenr, u64 num_bytes, int reserved)
2534 {
2535         struct btrfs_block_group *cache;
2536
2537         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2538         BUG_ON(!cache); /* Logic error */
2539
2540         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2541
2542         btrfs_put_block_group(cache);
2543         return 0;
2544 }
2545
2546 /*
2547  * this function must be called within transaction
2548  */
2549 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2550                                     u64 bytenr, u64 num_bytes)
2551 {
2552         struct btrfs_block_group *cache;
2553         int ret;
2554
2555         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2556         if (!cache)
2557                 return -EINVAL;
2558
2559         /*
2560          * Fully cache the free space first so that our pin removes the free space
2561          * from the cache.
2562          */
2563         ret = btrfs_cache_block_group(cache, true);
2564         if (ret)
2565                 goto out;
2566
2567         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2568
2569         /* remove us from the free space cache (if we're there at all) */
2570         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2571 out:
2572         btrfs_put_block_group(cache);
2573         return ret;
2574 }
2575
2576 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2577                                    u64 start, u64 num_bytes)
2578 {
2579         int ret;
2580         struct btrfs_block_group *block_group;
2581
2582         block_group = btrfs_lookup_block_group(fs_info, start);
2583         if (!block_group)
2584                 return -EINVAL;
2585
2586         ret = btrfs_cache_block_group(block_group, true);
2587         if (ret)
2588                 goto out;
2589
2590         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2591 out:
2592         btrfs_put_block_group(block_group);
2593         return ret;
2594 }
2595
2596 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2597 {
2598         struct btrfs_fs_info *fs_info = eb->fs_info;
2599         struct btrfs_file_extent_item *item;
2600         struct btrfs_key key;
2601         int found_type;
2602         int i;
2603         int ret = 0;
2604
2605         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2606                 return 0;
2607
2608         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2609                 btrfs_item_key_to_cpu(eb, &key, i);
2610                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2611                         continue;
2612                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2613                 found_type = btrfs_file_extent_type(eb, item);
2614                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2615                         continue;
2616                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2617                         continue;
2618                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2619                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2620                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2621                 if (ret)
2622                         break;
2623         }
2624
2625         return ret;
2626 }
2627
2628 static void
2629 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2630 {
2631         atomic_inc(&bg->reservations);
2632 }
2633
2634 /*
2635  * Returns the free cluster for the given space info and sets empty_cluster to
2636  * what it should be based on the mount options.
2637  */
2638 static struct btrfs_free_cluster *
2639 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2640                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2641 {
2642         struct btrfs_free_cluster *ret = NULL;
2643
2644         *empty_cluster = 0;
2645         if (btrfs_mixed_space_info(space_info))
2646                 return ret;
2647
2648         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2649                 ret = &fs_info->meta_alloc_cluster;
2650                 if (btrfs_test_opt(fs_info, SSD))
2651                         *empty_cluster = SZ_2M;
2652                 else
2653                         *empty_cluster = SZ_64K;
2654         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2655                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2656                 *empty_cluster = SZ_2M;
2657                 ret = &fs_info->data_alloc_cluster;
2658         }
2659
2660         return ret;
2661 }
2662
2663 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2664                               u64 start, u64 end,
2665                               const bool return_free_space)
2666 {
2667         struct btrfs_block_group *cache = NULL;
2668         struct btrfs_space_info *space_info;
2669         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2670         struct btrfs_free_cluster *cluster = NULL;
2671         u64 len;
2672         u64 total_unpinned = 0;
2673         u64 empty_cluster = 0;
2674         bool readonly;
2675
2676         while (start <= end) {
2677                 readonly = false;
2678                 if (!cache ||
2679                     start >= cache->start + cache->length) {
2680                         if (cache)
2681                                 btrfs_put_block_group(cache);
2682                         total_unpinned = 0;
2683                         cache = btrfs_lookup_block_group(fs_info, start);
2684                         BUG_ON(!cache); /* Logic error */
2685
2686                         cluster = fetch_cluster_info(fs_info,
2687                                                      cache->space_info,
2688                                                      &empty_cluster);
2689                         empty_cluster <<= 1;
2690                 }
2691
2692                 len = cache->start + cache->length - start;
2693                 len = min(len, end + 1 - start);
2694
2695                 if (return_free_space)
2696                         btrfs_add_free_space(cache, start, len);
2697
2698                 start += len;
2699                 total_unpinned += len;
2700                 space_info = cache->space_info;
2701
2702                 /*
2703                  * If this space cluster has been marked as fragmented and we've
2704                  * unpinned enough in this block group to potentially allow a
2705                  * cluster to be created inside of it go ahead and clear the
2706                  * fragmented check.
2707                  */
2708                 if (cluster && cluster->fragmented &&
2709                     total_unpinned > empty_cluster) {
2710                         spin_lock(&cluster->lock);
2711                         cluster->fragmented = 0;
2712                         spin_unlock(&cluster->lock);
2713                 }
2714
2715                 spin_lock(&space_info->lock);
2716                 spin_lock(&cache->lock);
2717                 cache->pinned -= len;
2718                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2719                 space_info->max_extent_size = 0;
2720                 if (cache->ro) {
2721                         space_info->bytes_readonly += len;
2722                         readonly = true;
2723                 } else if (btrfs_is_zoned(fs_info)) {
2724                         /* Need reset before reusing in a zoned block group */
2725                         space_info->bytes_zone_unusable += len;
2726                         readonly = true;
2727                 }
2728                 spin_unlock(&cache->lock);
2729                 if (!readonly && return_free_space &&
2730                     global_rsv->space_info == space_info) {
2731                         spin_lock(&global_rsv->lock);
2732                         if (!global_rsv->full) {
2733                                 u64 to_add = min(len, global_rsv->size -
2734                                                       global_rsv->reserved);
2735
2736                                 global_rsv->reserved += to_add;
2737                                 btrfs_space_info_update_bytes_may_use(fs_info,
2738                                                 space_info, to_add);
2739                                 if (global_rsv->reserved >= global_rsv->size)
2740                                         global_rsv->full = 1;
2741                                 len -= to_add;
2742                         }
2743                         spin_unlock(&global_rsv->lock);
2744                 }
2745                 /* Add to any tickets we may have */
2746                 if (!readonly && return_free_space && len)
2747                         btrfs_try_granting_tickets(fs_info, space_info);
2748                 spin_unlock(&space_info->lock);
2749         }
2750
2751         if (cache)
2752                 btrfs_put_block_group(cache);
2753         return 0;
2754 }
2755
2756 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2757 {
2758         struct btrfs_fs_info *fs_info = trans->fs_info;
2759         struct btrfs_block_group *block_group, *tmp;
2760         struct list_head *deleted_bgs;
2761         struct extent_io_tree *unpin;
2762         u64 start;
2763         u64 end;
2764         int ret;
2765
2766         unpin = &trans->transaction->pinned_extents;
2767
2768         while (!TRANS_ABORTED(trans)) {
2769                 struct extent_state *cached_state = NULL;
2770
2771                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2772                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2773                                             EXTENT_DIRTY, &cached_state);
2774                 if (ret) {
2775                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2776                         break;
2777                 }
2778
2779                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2780                         ret = btrfs_discard_extent(fs_info, start,
2781                                                    end + 1 - start, NULL);
2782
2783                 clear_extent_dirty(unpin, start, end, &cached_state);
2784                 unpin_extent_range(fs_info, start, end, true);
2785                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2786                 free_extent_state(cached_state);
2787                 cond_resched();
2788         }
2789
2790         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2791                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2792                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2793         }
2794
2795         /*
2796          * Transaction is finished.  We don't need the lock anymore.  We
2797          * do need to clean up the block groups in case of a transaction
2798          * abort.
2799          */
2800         deleted_bgs = &trans->transaction->deleted_bgs;
2801         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2802                 u64 trimmed = 0;
2803
2804                 ret = -EROFS;
2805                 if (!TRANS_ABORTED(trans))
2806                         ret = btrfs_discard_extent(fs_info,
2807                                                    block_group->start,
2808                                                    block_group->length,
2809                                                    &trimmed);
2810
2811                 list_del_init(&block_group->bg_list);
2812                 btrfs_unfreeze_block_group(block_group);
2813                 btrfs_put_block_group(block_group);
2814
2815                 if (ret) {
2816                         const char *errstr = btrfs_decode_error(ret);
2817                         btrfs_warn(fs_info,
2818                            "discard failed while removing blockgroup: errno=%d %s",
2819                                    ret, errstr);
2820                 }
2821         }
2822
2823         return 0;
2824 }
2825
2826 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2827                                      u64 bytenr, u64 num_bytes, bool is_data)
2828 {
2829         int ret;
2830
2831         if (is_data) {
2832                 struct btrfs_root *csum_root;
2833
2834                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2835                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2836                 if (ret) {
2837                         btrfs_abort_transaction(trans, ret);
2838                         return ret;
2839                 }
2840         }
2841
2842         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2843         if (ret) {
2844                 btrfs_abort_transaction(trans, ret);
2845                 return ret;
2846         }
2847
2848         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2849         if (ret)
2850                 btrfs_abort_transaction(trans, ret);
2851
2852         return ret;
2853 }
2854
2855 /*
2856  * Drop one or more refs of @node.
2857  *
2858  * 1. Locate the extent refs.
2859  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2860  *    Locate it, then reduce the refs number or remove the ref line completely.
2861  *
2862  * 2. Update the refs count in EXTENT/METADATA_ITEM
2863  *
2864  * Inline backref case:
2865  *
2866  * in extent tree we have:
2867  *
2868  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2869  *              refs 2 gen 6 flags DATA
2870  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2871  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2872  *
2873  * This function gets called with:
2874  *
2875  *    node->bytenr = 13631488
2876  *    node->num_bytes = 1048576
2877  *    root_objectid = FS_TREE
2878  *    owner_objectid = 257
2879  *    owner_offset = 0
2880  *    refs_to_drop = 1
2881  *
2882  * Then we should get some like:
2883  *
2884  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2885  *              refs 1 gen 6 flags DATA
2886  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2887  *
2888  * Keyed backref case:
2889  *
2890  * in extent tree we have:
2891  *
2892  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2893  *              refs 754 gen 6 flags DATA
2894  *      [...]
2895  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2896  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2897  *
2898  * This function get called with:
2899  *
2900  *    node->bytenr = 13631488
2901  *    node->num_bytes = 1048576
2902  *    root_objectid = FS_TREE
2903  *    owner_objectid = 866
2904  *    owner_offset = 0
2905  *    refs_to_drop = 1
2906  *
2907  * Then we should get some like:
2908  *
2909  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2910  *              refs 753 gen 6 flags DATA
2911  *
2912  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2913  */
2914 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2915                                struct btrfs_delayed_ref_node *node, u64 parent,
2916                                u64 root_objectid, u64 owner_objectid,
2917                                u64 owner_offset, int refs_to_drop,
2918                                struct btrfs_delayed_extent_op *extent_op)
2919 {
2920         struct btrfs_fs_info *info = trans->fs_info;
2921         struct btrfs_key key;
2922         struct btrfs_path *path;
2923         struct btrfs_root *extent_root;
2924         struct extent_buffer *leaf;
2925         struct btrfs_extent_item *ei;
2926         struct btrfs_extent_inline_ref *iref;
2927         int ret;
2928         int is_data;
2929         int extent_slot = 0;
2930         int found_extent = 0;
2931         int num_to_del = 1;
2932         u32 item_size;
2933         u64 refs;
2934         u64 bytenr = node->bytenr;
2935         u64 num_bytes = node->num_bytes;
2936         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2937
2938         extent_root = btrfs_extent_root(info, bytenr);
2939         ASSERT(extent_root);
2940
2941         path = btrfs_alloc_path();
2942         if (!path)
2943                 return -ENOMEM;
2944
2945         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2946
2947         if (!is_data && refs_to_drop != 1) {
2948                 btrfs_crit(info,
2949 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2950                            node->bytenr, refs_to_drop);
2951                 ret = -EINVAL;
2952                 btrfs_abort_transaction(trans, ret);
2953                 goto out;
2954         }
2955
2956         if (is_data)
2957                 skinny_metadata = false;
2958
2959         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2960                                     parent, root_objectid, owner_objectid,
2961                                     owner_offset);
2962         if (ret == 0) {
2963                 /*
2964                  * Either the inline backref or the SHARED_DATA_REF/
2965                  * SHARED_BLOCK_REF is found
2966                  *
2967                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2968                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2969                  */
2970                 extent_slot = path->slots[0];
2971                 while (extent_slot >= 0) {
2972                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2973                                               extent_slot);
2974                         if (key.objectid != bytenr)
2975                                 break;
2976                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2977                             key.offset == num_bytes) {
2978                                 found_extent = 1;
2979                                 break;
2980                         }
2981                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2982                             key.offset == owner_objectid) {
2983                                 found_extent = 1;
2984                                 break;
2985                         }
2986
2987                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
2988                         if (path->slots[0] - extent_slot > 5)
2989                                 break;
2990                         extent_slot--;
2991                 }
2992
2993                 if (!found_extent) {
2994                         if (iref) {
2995                                 btrfs_crit(info,
2996 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
2997                                 btrfs_abort_transaction(trans, -EUCLEAN);
2998                                 goto err_dump;
2999                         }
3000                         /* Must be SHARED_* item, remove the backref first */
3001                         ret = remove_extent_backref(trans, extent_root, path,
3002                                                     NULL, refs_to_drop, is_data);
3003                         if (ret) {
3004                                 btrfs_abort_transaction(trans, ret);
3005                                 goto out;
3006                         }
3007                         btrfs_release_path(path);
3008
3009                         /* Slow path to locate EXTENT/METADATA_ITEM */
3010                         key.objectid = bytenr;
3011                         key.type = BTRFS_EXTENT_ITEM_KEY;
3012                         key.offset = num_bytes;
3013
3014                         if (!is_data && skinny_metadata) {
3015                                 key.type = BTRFS_METADATA_ITEM_KEY;
3016                                 key.offset = owner_objectid;
3017                         }
3018
3019                         ret = btrfs_search_slot(trans, extent_root,
3020                                                 &key, path, -1, 1);
3021                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3022                                 /*
3023                                  * Couldn't find our skinny metadata item,
3024                                  * see if we have ye olde extent item.
3025                                  */
3026                                 path->slots[0]--;
3027                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3028                                                       path->slots[0]);
3029                                 if (key.objectid == bytenr &&
3030                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3031                                     key.offset == num_bytes)
3032                                         ret = 0;
3033                         }
3034
3035                         if (ret > 0 && skinny_metadata) {
3036                                 skinny_metadata = false;
3037                                 key.objectid = bytenr;
3038                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3039                                 key.offset = num_bytes;
3040                                 btrfs_release_path(path);
3041                                 ret = btrfs_search_slot(trans, extent_root,
3042                                                         &key, path, -1, 1);
3043                         }
3044
3045                         if (ret) {
3046                                 btrfs_err(info,
3047                                           "umm, got %d back from search, was looking for %llu",
3048                                           ret, bytenr);
3049                                 if (ret > 0)
3050                                         btrfs_print_leaf(path->nodes[0]);
3051                         }
3052                         if (ret < 0) {
3053                                 btrfs_abort_transaction(trans, ret);
3054                                 goto out;
3055                         }
3056                         extent_slot = path->slots[0];
3057                 }
3058         } else if (WARN_ON(ret == -ENOENT)) {
3059                 btrfs_print_leaf(path->nodes[0]);
3060                 btrfs_err(info,
3061                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3062                         bytenr, parent, root_objectid, owner_objectid,
3063                         owner_offset);
3064                 btrfs_abort_transaction(trans, ret);
3065                 goto out;
3066         } else {
3067                 btrfs_abort_transaction(trans, ret);
3068                 goto out;
3069         }
3070
3071         leaf = path->nodes[0];
3072         item_size = btrfs_item_size(leaf, extent_slot);
3073         if (unlikely(item_size < sizeof(*ei))) {
3074                 ret = -EINVAL;
3075                 btrfs_print_v0_err(info);
3076                 btrfs_abort_transaction(trans, ret);
3077                 goto out;
3078         }
3079         ei = btrfs_item_ptr(leaf, extent_slot,
3080                             struct btrfs_extent_item);
3081         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3082             key.type == BTRFS_EXTENT_ITEM_KEY) {
3083                 struct btrfs_tree_block_info *bi;
3084                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3085                         btrfs_crit(info,
3086 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3087                                    key.objectid, key.type, key.offset,
3088                                    owner_objectid, item_size,
3089                                    sizeof(*ei) + sizeof(*bi));
3090                         btrfs_abort_transaction(trans, -EUCLEAN);
3091                         goto err_dump;
3092                 }
3093                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3094                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3095         }
3096
3097         refs = btrfs_extent_refs(leaf, ei);
3098         if (refs < refs_to_drop) {
3099                 btrfs_crit(info,
3100                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3101                           refs_to_drop, refs, bytenr);
3102                 btrfs_abort_transaction(trans, -EUCLEAN);
3103                 goto err_dump;
3104         }
3105         refs -= refs_to_drop;
3106
3107         if (refs > 0) {
3108                 if (extent_op)
3109                         __run_delayed_extent_op(extent_op, leaf, ei);
3110                 /*
3111                  * In the case of inline back ref, reference count will
3112                  * be updated by remove_extent_backref
3113                  */
3114                 if (iref) {
3115                         if (!found_extent) {
3116                                 btrfs_crit(info,
3117 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3118                                 btrfs_abort_transaction(trans, -EUCLEAN);
3119                                 goto err_dump;
3120                         }
3121                 } else {
3122                         btrfs_set_extent_refs(leaf, ei, refs);
3123                         btrfs_mark_buffer_dirty(leaf);
3124                 }
3125                 if (found_extent) {
3126                         ret = remove_extent_backref(trans, extent_root, path,
3127                                                     iref, refs_to_drop, is_data);
3128                         if (ret) {
3129                                 btrfs_abort_transaction(trans, ret);
3130                                 goto out;
3131                         }
3132                 }
3133         } else {
3134                 /* In this branch refs == 1 */
3135                 if (found_extent) {
3136                         if (is_data && refs_to_drop !=
3137                             extent_data_ref_count(path, iref)) {
3138                                 btrfs_crit(info,
3139                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3140                                            extent_data_ref_count(path, iref),
3141                                            refs_to_drop);
3142                                 btrfs_abort_transaction(trans, -EUCLEAN);
3143                                 goto err_dump;
3144                         }
3145                         if (iref) {
3146                                 if (path->slots[0] != extent_slot) {
3147                                         btrfs_crit(info,
3148 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3149                                                    key.objectid, key.type,
3150                                                    key.offset);
3151                                         btrfs_abort_transaction(trans, -EUCLEAN);
3152                                         goto err_dump;
3153                                 }
3154                         } else {
3155                                 /*
3156                                  * No inline ref, we must be at SHARED_* item,
3157                                  * And it's single ref, it must be:
3158                                  * |    extent_slot       ||extent_slot + 1|
3159                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3160                                  */
3161                                 if (path->slots[0] != extent_slot + 1) {
3162                                         btrfs_crit(info,
3163         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3164                                         btrfs_abort_transaction(trans, -EUCLEAN);
3165                                         goto err_dump;
3166                                 }
3167                                 path->slots[0] = extent_slot;
3168                                 num_to_del = 2;
3169                         }
3170                 }
3171
3172                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3173                                       num_to_del);
3174                 if (ret) {
3175                         btrfs_abort_transaction(trans, ret);
3176                         goto out;
3177                 }
3178                 btrfs_release_path(path);
3179
3180                 ret = do_free_extent_accounting(trans, bytenr, num_bytes, is_data);
3181         }
3182         btrfs_release_path(path);
3183
3184 out:
3185         btrfs_free_path(path);
3186         return ret;
3187 err_dump:
3188         /*
3189          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3190          * dump for debug build.
3191          */
3192         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3193                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3194                            path->slots[0], extent_slot);
3195                 btrfs_print_leaf(path->nodes[0]);
3196         }
3197
3198         btrfs_free_path(path);
3199         return -EUCLEAN;
3200 }
3201
3202 /*
3203  * when we free an block, it is possible (and likely) that we free the last
3204  * delayed ref for that extent as well.  This searches the delayed ref tree for
3205  * a given extent, and if there are no other delayed refs to be processed, it
3206  * removes it from the tree.
3207  */
3208 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3209                                       u64 bytenr)
3210 {
3211         struct btrfs_delayed_ref_head *head;
3212         struct btrfs_delayed_ref_root *delayed_refs;
3213         int ret = 0;
3214
3215         delayed_refs = &trans->transaction->delayed_refs;
3216         spin_lock(&delayed_refs->lock);
3217         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3218         if (!head)
3219                 goto out_delayed_unlock;
3220
3221         spin_lock(&head->lock);
3222         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3223                 goto out;
3224
3225         if (cleanup_extent_op(head) != NULL)
3226                 goto out;
3227
3228         /*
3229          * waiting for the lock here would deadlock.  If someone else has it
3230          * locked they are already in the process of dropping it anyway
3231          */
3232         if (!mutex_trylock(&head->mutex))
3233                 goto out;
3234
3235         btrfs_delete_ref_head(delayed_refs, head);
3236         head->processing = 0;
3237
3238         spin_unlock(&head->lock);
3239         spin_unlock(&delayed_refs->lock);
3240
3241         BUG_ON(head->extent_op);
3242         if (head->must_insert_reserved)
3243                 ret = 1;
3244
3245         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3246         mutex_unlock(&head->mutex);
3247         btrfs_put_delayed_ref_head(head);
3248         return ret;
3249 out:
3250         spin_unlock(&head->lock);
3251
3252 out_delayed_unlock:
3253         spin_unlock(&delayed_refs->lock);
3254         return 0;
3255 }
3256
3257 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3258                            u64 root_id,
3259                            struct extent_buffer *buf,
3260                            u64 parent, int last_ref)
3261 {
3262         struct btrfs_fs_info *fs_info = trans->fs_info;
3263         struct btrfs_ref generic_ref = { 0 };
3264         int ret;
3265
3266         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3267                                buf->start, buf->len, parent);
3268         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3269                             root_id, 0, false);
3270
3271         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3272                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3273                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3274                 BUG_ON(ret); /* -ENOMEM */
3275         }
3276
3277         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3278                 struct btrfs_block_group *cache;
3279                 bool must_pin = false;
3280
3281                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3282                         ret = check_ref_cleanup(trans, buf->start);
3283                         if (!ret) {
3284                                 btrfs_redirty_list_add(trans->transaction, buf);
3285                                 goto out;
3286                         }
3287                 }
3288
3289                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3290
3291                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3292                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3293                         btrfs_put_block_group(cache);
3294                         goto out;
3295                 }
3296
3297                 /*
3298                  * If there are tree mod log users we may have recorded mod log
3299                  * operations for this node.  If we re-allocate this node we
3300                  * could replay operations on this node that happened when it
3301                  * existed in a completely different root.  For example if it
3302                  * was part of root A, then was reallocated to root B, and we
3303                  * are doing a btrfs_old_search_slot(root b), we could replay
3304                  * operations that happened when the block was part of root A,
3305                  * giving us an inconsistent view of the btree.
3306                  *
3307                  * We are safe from races here because at this point no other
3308                  * node or root points to this extent buffer, so if after this
3309                  * check a new tree mod log user joins we will not have an
3310                  * existing log of operations on this node that we have to
3311                  * contend with.
3312                  */
3313                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3314                         must_pin = true;
3315
3316                 if (must_pin || btrfs_is_zoned(fs_info)) {
3317                         btrfs_redirty_list_add(trans->transaction, buf);
3318                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3319                         btrfs_put_block_group(cache);
3320                         goto out;
3321                 }
3322
3323                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3324
3325                 btrfs_add_free_space(cache, buf->start, buf->len);
3326                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3327                 btrfs_put_block_group(cache);
3328                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3329         }
3330 out:
3331         if (last_ref) {
3332                 /*
3333                  * Deleting the buffer, clear the corrupt flag since it doesn't
3334                  * matter anymore.
3335                  */
3336                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3337         }
3338 }
3339
3340 /* Can return -ENOMEM */
3341 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3342 {
3343         struct btrfs_fs_info *fs_info = trans->fs_info;
3344         int ret;
3345
3346         if (btrfs_is_testing(fs_info))
3347                 return 0;
3348
3349         /*
3350          * tree log blocks never actually go into the extent allocation
3351          * tree, just update pinning info and exit early.
3352          */
3353         if ((ref->type == BTRFS_REF_METADATA &&
3354              ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3355             (ref->type == BTRFS_REF_DATA &&
3356              ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)) {
3357                 /* unlocks the pinned mutex */
3358                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3359                 ret = 0;
3360         } else if (ref->type == BTRFS_REF_METADATA) {
3361                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3362         } else {
3363                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3364         }
3365
3366         if (!((ref->type == BTRFS_REF_METADATA &&
3367                ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3368               (ref->type == BTRFS_REF_DATA &&
3369                ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)))
3370                 btrfs_ref_tree_mod(fs_info, ref);
3371
3372         return ret;
3373 }
3374
3375 enum btrfs_loop_type {
3376         LOOP_CACHING_NOWAIT,
3377         LOOP_CACHING_WAIT,
3378         LOOP_ALLOC_CHUNK,
3379         LOOP_NO_EMPTY_SIZE,
3380 };
3381
3382 static inline void
3383 btrfs_lock_block_group(struct btrfs_block_group *cache,
3384                        int delalloc)
3385 {
3386         if (delalloc)
3387                 down_read(&cache->data_rwsem);
3388 }
3389
3390 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3391                        int delalloc)
3392 {
3393         btrfs_get_block_group(cache);
3394         if (delalloc)
3395                 down_read(&cache->data_rwsem);
3396 }
3397
3398 static struct btrfs_block_group *btrfs_lock_cluster(
3399                    struct btrfs_block_group *block_group,
3400                    struct btrfs_free_cluster *cluster,
3401                    int delalloc)
3402         __acquires(&cluster->refill_lock)
3403 {
3404         struct btrfs_block_group *used_bg = NULL;
3405
3406         spin_lock(&cluster->refill_lock);
3407         while (1) {
3408                 used_bg = cluster->block_group;
3409                 if (!used_bg)
3410                         return NULL;
3411
3412                 if (used_bg == block_group)
3413                         return used_bg;
3414
3415                 btrfs_get_block_group(used_bg);
3416
3417                 if (!delalloc)
3418                         return used_bg;
3419
3420                 if (down_read_trylock(&used_bg->data_rwsem))
3421                         return used_bg;
3422
3423                 spin_unlock(&cluster->refill_lock);
3424
3425                 /* We should only have one-level nested. */
3426                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3427
3428                 spin_lock(&cluster->refill_lock);
3429                 if (used_bg == cluster->block_group)
3430                         return used_bg;
3431
3432                 up_read(&used_bg->data_rwsem);
3433                 btrfs_put_block_group(used_bg);
3434         }
3435 }
3436
3437 static inline void
3438 btrfs_release_block_group(struct btrfs_block_group *cache,
3439                          int delalloc)
3440 {
3441         if (delalloc)
3442                 up_read(&cache->data_rwsem);
3443         btrfs_put_block_group(cache);
3444 }
3445
3446 enum btrfs_extent_allocation_policy {
3447         BTRFS_EXTENT_ALLOC_CLUSTERED,
3448         BTRFS_EXTENT_ALLOC_ZONED,
3449 };
3450
3451 /*
3452  * Structure used internally for find_free_extent() function.  Wraps needed
3453  * parameters.
3454  */
3455 struct find_free_extent_ctl {
3456         /* Basic allocation info */
3457         u64 ram_bytes;
3458         u64 num_bytes;
3459         u64 min_alloc_size;
3460         u64 empty_size;
3461         u64 flags;
3462         int delalloc;
3463
3464         /* Where to start the search inside the bg */
3465         u64 search_start;
3466
3467         /* For clustered allocation */
3468         u64 empty_cluster;
3469         struct btrfs_free_cluster *last_ptr;
3470         bool use_cluster;
3471
3472         bool have_caching_bg;
3473         bool orig_have_caching_bg;
3474
3475         /* Allocation is called for tree-log */
3476         bool for_treelog;
3477
3478         /* Allocation is called for data relocation */
3479         bool for_data_reloc;
3480
3481         /* RAID index, converted from flags */
3482         int index;
3483
3484         /*
3485          * Current loop number, check find_free_extent_update_loop() for details
3486          */
3487         int loop;
3488
3489         /*
3490          * Whether we're refilling a cluster, if true we need to re-search
3491          * current block group but don't try to refill the cluster again.
3492          */
3493         bool retry_clustered;
3494
3495         /*
3496          * Whether we're updating free space cache, if true we need to re-search
3497          * current block group but don't try updating free space cache again.
3498          */
3499         bool retry_unclustered;
3500
3501         /* If current block group is cached */
3502         int cached;
3503
3504         /* Max contiguous hole found */
3505         u64 max_extent_size;
3506
3507         /* Total free space from free space cache, not always contiguous */
3508         u64 total_free_space;
3509
3510         /* Found result */
3511         u64 found_offset;
3512
3513         /* Hint where to start looking for an empty space */
3514         u64 hint_byte;
3515
3516         /* Allocation policy */
3517         enum btrfs_extent_allocation_policy policy;
3518 };
3519
3520
3521 /*
3522  * Helper function for find_free_extent().
3523  *
3524  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3525  * Return -EAGAIN to inform caller that we need to re-search this block group
3526  * Return >0 to inform caller that we find nothing
3527  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3528  */
3529 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3530                                       struct find_free_extent_ctl *ffe_ctl,
3531                                       struct btrfs_block_group **cluster_bg_ret)
3532 {
3533         struct btrfs_block_group *cluster_bg;
3534         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3535         u64 aligned_cluster;
3536         u64 offset;
3537         int ret;
3538
3539         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3540         if (!cluster_bg)
3541                 goto refill_cluster;
3542         if (cluster_bg != bg && (cluster_bg->ro ||
3543             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3544                 goto release_cluster;
3545
3546         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3547                         ffe_ctl->num_bytes, cluster_bg->start,
3548                         &ffe_ctl->max_extent_size);
3549         if (offset) {
3550                 /* We have a block, we're done */
3551                 spin_unlock(&last_ptr->refill_lock);
3552                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3553                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3554                 *cluster_bg_ret = cluster_bg;
3555                 ffe_ctl->found_offset = offset;
3556                 return 0;
3557         }
3558         WARN_ON(last_ptr->block_group != cluster_bg);
3559
3560 release_cluster:
3561         /*
3562          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3563          * lets just skip it and let the allocator find whatever block it can
3564          * find. If we reach this point, we will have tried the cluster
3565          * allocator plenty of times and not have found anything, so we are
3566          * likely way too fragmented for the clustering stuff to find anything.
3567          *
3568          * However, if the cluster is taken from the current block group,
3569          * release the cluster first, so that we stand a better chance of
3570          * succeeding in the unclustered allocation.
3571          */
3572         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3573                 spin_unlock(&last_ptr->refill_lock);
3574                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3575                 return -ENOENT;
3576         }
3577
3578         /* This cluster didn't work out, free it and start over */
3579         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3580
3581         if (cluster_bg != bg)
3582                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3583
3584 refill_cluster:
3585         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3586                 spin_unlock(&last_ptr->refill_lock);
3587                 return -ENOENT;
3588         }
3589
3590         aligned_cluster = max_t(u64,
3591                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3592                         bg->full_stripe_len);
3593         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3594                         ffe_ctl->num_bytes, aligned_cluster);
3595         if (ret == 0) {
3596                 /* Now pull our allocation out of this cluster */
3597                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3598                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3599                                 &ffe_ctl->max_extent_size);
3600                 if (offset) {
3601                         /* We found one, proceed */
3602                         spin_unlock(&last_ptr->refill_lock);
3603                         trace_btrfs_reserve_extent_cluster(bg,
3604                                         ffe_ctl->search_start,
3605                                         ffe_ctl->num_bytes);
3606                         ffe_ctl->found_offset = offset;
3607                         return 0;
3608                 }
3609         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3610                    !ffe_ctl->retry_clustered) {
3611                 spin_unlock(&last_ptr->refill_lock);
3612
3613                 ffe_ctl->retry_clustered = true;
3614                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3615                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3616                 return -EAGAIN;
3617         }
3618         /*
3619          * At this point we either didn't find a cluster or we weren't able to
3620          * allocate a block from our cluster.  Free the cluster we've been
3621          * trying to use, and go to the next block group.
3622          */
3623         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3624         spin_unlock(&last_ptr->refill_lock);
3625         return 1;
3626 }
3627
3628 /*
3629  * Return >0 to inform caller that we find nothing
3630  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3631  * Return -EAGAIN to inform caller that we need to re-search this block group
3632  */
3633 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3634                                         struct find_free_extent_ctl *ffe_ctl)
3635 {
3636         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3637         u64 offset;
3638
3639         /*
3640          * We are doing an unclustered allocation, set the fragmented flag so
3641          * we don't bother trying to setup a cluster again until we get more
3642          * space.
3643          */
3644         if (unlikely(last_ptr)) {
3645                 spin_lock(&last_ptr->lock);
3646                 last_ptr->fragmented = 1;
3647                 spin_unlock(&last_ptr->lock);
3648         }
3649         if (ffe_ctl->cached) {
3650                 struct btrfs_free_space_ctl *free_space_ctl;
3651
3652                 free_space_ctl = bg->free_space_ctl;
3653                 spin_lock(&free_space_ctl->tree_lock);
3654                 if (free_space_ctl->free_space <
3655                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3656                     ffe_ctl->empty_size) {
3657                         ffe_ctl->total_free_space = max_t(u64,
3658                                         ffe_ctl->total_free_space,
3659                                         free_space_ctl->free_space);
3660                         spin_unlock(&free_space_ctl->tree_lock);
3661                         return 1;
3662                 }
3663                 spin_unlock(&free_space_ctl->tree_lock);
3664         }
3665
3666         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3667                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3668                         &ffe_ctl->max_extent_size);
3669
3670         /*
3671          * If we didn't find a chunk, and we haven't failed on this block group
3672          * before, and this block group is in the middle of caching and we are
3673          * ok with waiting, then go ahead and wait for progress to be made, and
3674          * set @retry_unclustered to true.
3675          *
3676          * If @retry_unclustered is true then we've already waited on this
3677          * block group once and should move on to the next block group.
3678          */
3679         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3680             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3681                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3682                                                       ffe_ctl->empty_size);
3683                 ffe_ctl->retry_unclustered = true;
3684                 return -EAGAIN;
3685         } else if (!offset) {
3686                 return 1;
3687         }
3688         ffe_ctl->found_offset = offset;
3689         return 0;
3690 }
3691
3692 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3693                                    struct find_free_extent_ctl *ffe_ctl,
3694                                    struct btrfs_block_group **bg_ret)
3695 {
3696         int ret;
3697
3698         /* We want to try and use the cluster allocator, so lets look there */
3699         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3700                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3701                 if (ret >= 0 || ret == -EAGAIN)
3702                         return ret;
3703                 /* ret == -ENOENT case falls through */
3704         }
3705
3706         return find_free_extent_unclustered(block_group, ffe_ctl);
3707 }
3708
3709 /*
3710  * Tree-log block group locking
3711  * ============================
3712  *
3713  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3714  * indicates the starting address of a block group, which is reserved only
3715  * for tree-log metadata.
3716  *
3717  * Lock nesting
3718  * ============
3719  *
3720  * space_info::lock
3721  *   block_group::lock
3722  *     fs_info::treelog_bg_lock
3723  */
3724
3725 /*
3726  * Simple allocator for sequential-only block group. It only allows sequential
3727  * allocation. No need to play with trees. This function also reserves the
3728  * bytes as in btrfs_add_reserved_bytes.
3729  */
3730 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3731                                struct find_free_extent_ctl *ffe_ctl,
3732                                struct btrfs_block_group **bg_ret)
3733 {
3734         struct btrfs_fs_info *fs_info = block_group->fs_info;
3735         struct btrfs_space_info *space_info = block_group->space_info;
3736         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3737         u64 start = block_group->start;
3738         u64 num_bytes = ffe_ctl->num_bytes;
3739         u64 avail;
3740         u64 bytenr = block_group->start;
3741         u64 log_bytenr;
3742         u64 data_reloc_bytenr;
3743         int ret = 0;
3744         bool skip = false;
3745
3746         ASSERT(btrfs_is_zoned(block_group->fs_info));
3747
3748         /*
3749          * Do not allow non-tree-log blocks in the dedicated tree-log block
3750          * group, and vice versa.
3751          */
3752         spin_lock(&fs_info->treelog_bg_lock);
3753         log_bytenr = fs_info->treelog_bg;
3754         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3755                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3756                 skip = true;
3757         spin_unlock(&fs_info->treelog_bg_lock);
3758         if (skip)
3759                 return 1;
3760
3761         /*
3762          * Do not allow non-relocation blocks in the dedicated relocation block
3763          * group, and vice versa.
3764          */
3765         spin_lock(&fs_info->relocation_bg_lock);
3766         data_reloc_bytenr = fs_info->data_reloc_bg;
3767         if (data_reloc_bytenr &&
3768             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3769              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3770                 skip = true;
3771         spin_unlock(&fs_info->relocation_bg_lock);
3772         if (skip)
3773                 return 1;
3774
3775         /* Check RO and no space case before trying to activate it */
3776         spin_lock(&block_group->lock);
3777         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3778                 ret = 1;
3779                 /*
3780                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3781                  * Return the error after taking the locks.
3782                  */
3783         }
3784         spin_unlock(&block_group->lock);
3785
3786         if (!ret && !btrfs_zone_activate(block_group)) {
3787                 ret = 1;
3788                 /*
3789                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3790                  * Return the error after taking the locks.
3791                  */
3792         }
3793
3794         spin_lock(&space_info->lock);
3795         spin_lock(&block_group->lock);
3796         spin_lock(&fs_info->treelog_bg_lock);
3797         spin_lock(&fs_info->relocation_bg_lock);
3798
3799         if (ret)
3800                 goto out;
3801
3802         ASSERT(!ffe_ctl->for_treelog ||
3803                block_group->start == fs_info->treelog_bg ||
3804                fs_info->treelog_bg == 0);
3805         ASSERT(!ffe_ctl->for_data_reloc ||
3806                block_group->start == fs_info->data_reloc_bg ||
3807                fs_info->data_reloc_bg == 0);
3808
3809         if (block_group->ro ||
3810             test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
3811                 ret = 1;
3812                 goto out;
3813         }
3814
3815         /*
3816          * Do not allow currently using block group to be tree-log dedicated
3817          * block group.
3818          */
3819         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3820             (block_group->used || block_group->reserved)) {
3821                 ret = 1;
3822                 goto out;
3823         }
3824
3825         /*
3826          * Do not allow currently used block group to be the data relocation
3827          * dedicated block group.
3828          */
3829         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3830             (block_group->used || block_group->reserved)) {
3831                 ret = 1;
3832                 goto out;
3833         }
3834
3835         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3836         avail = block_group->zone_capacity - block_group->alloc_offset;
3837         if (avail < num_bytes) {
3838                 if (ffe_ctl->max_extent_size < avail) {
3839                         /*
3840                          * With sequential allocator, free space is always
3841                          * contiguous
3842                          */
3843                         ffe_ctl->max_extent_size = avail;
3844                         ffe_ctl->total_free_space = avail;
3845                 }
3846                 ret = 1;
3847                 goto out;
3848         }
3849
3850         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3851                 fs_info->treelog_bg = block_group->start;
3852
3853         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg)
3854                 fs_info->data_reloc_bg = block_group->start;
3855
3856         ffe_ctl->found_offset = start + block_group->alloc_offset;
3857         block_group->alloc_offset += num_bytes;
3858         spin_lock(&ctl->tree_lock);
3859         ctl->free_space -= num_bytes;
3860         spin_unlock(&ctl->tree_lock);
3861
3862         /*
3863          * We do not check if found_offset is aligned to stripesize. The
3864          * address is anyway rewritten when using zone append writing.
3865          */
3866
3867         ffe_ctl->search_start = ffe_ctl->found_offset;
3868
3869 out:
3870         if (ret && ffe_ctl->for_treelog)
3871                 fs_info->treelog_bg = 0;
3872         if (ret && ffe_ctl->for_data_reloc &&
3873             fs_info->data_reloc_bg == block_group->start) {
3874                 /*
3875                  * Do not allow further allocations from this block group.
3876                  * Compared to increasing the ->ro, setting the
3877                  * ->zoned_data_reloc_ongoing flag still allows nocow
3878                  *  writers to come in. See btrfs_inc_nocow_writers().
3879                  *
3880                  * We need to disable an allocation to avoid an allocation of
3881                  * regular (non-relocation data) extent. With mix of relocation
3882                  * extents and regular extents, we can dispatch WRITE commands
3883                  * (for relocation extents) and ZONE APPEND commands (for
3884                  * regular extents) at the same time to the same zone, which
3885                  * easily break the write pointer.
3886                  */
3887                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3888                 fs_info->data_reloc_bg = 0;
3889         }
3890         spin_unlock(&fs_info->relocation_bg_lock);
3891         spin_unlock(&fs_info->treelog_bg_lock);
3892         spin_unlock(&block_group->lock);
3893         spin_unlock(&space_info->lock);
3894         return ret;
3895 }
3896
3897 static int do_allocation(struct btrfs_block_group *block_group,
3898                          struct find_free_extent_ctl *ffe_ctl,
3899                          struct btrfs_block_group **bg_ret)
3900 {
3901         switch (ffe_ctl->policy) {
3902         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3903                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3904         case BTRFS_EXTENT_ALLOC_ZONED:
3905                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3906         default:
3907                 BUG();
3908         }
3909 }
3910
3911 static void release_block_group(struct btrfs_block_group *block_group,
3912                                 struct find_free_extent_ctl *ffe_ctl,
3913                                 int delalloc)
3914 {
3915         switch (ffe_ctl->policy) {
3916         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3917                 ffe_ctl->retry_clustered = false;
3918                 ffe_ctl->retry_unclustered = false;
3919                 break;
3920         case BTRFS_EXTENT_ALLOC_ZONED:
3921                 /* Nothing to do */
3922                 break;
3923         default:
3924                 BUG();
3925         }
3926
3927         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3928                ffe_ctl->index);
3929         btrfs_release_block_group(block_group, delalloc);
3930 }
3931
3932 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3933                                    struct btrfs_key *ins)
3934 {
3935         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3936
3937         if (!ffe_ctl->use_cluster && last_ptr) {
3938                 spin_lock(&last_ptr->lock);
3939                 last_ptr->window_start = ins->objectid;
3940                 spin_unlock(&last_ptr->lock);
3941         }
3942 }
3943
3944 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3945                          struct btrfs_key *ins)
3946 {
3947         switch (ffe_ctl->policy) {
3948         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3949                 found_extent_clustered(ffe_ctl, ins);
3950                 break;
3951         case BTRFS_EXTENT_ALLOC_ZONED:
3952                 /* Nothing to do */
3953                 break;
3954         default:
3955                 BUG();
3956         }
3957 }
3958
3959 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
3960                                     struct find_free_extent_ctl *ffe_ctl)
3961 {
3962         /* If we can activate new zone, just allocate a chunk and use it */
3963         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
3964                 return 0;
3965
3966         /*
3967          * We already reached the max active zones. Try to finish one block
3968          * group to make a room for a new block group. This is only possible
3969          * for a data block group because btrfs_zone_finish() may need to wait
3970          * for a running transaction which can cause a deadlock for metadata
3971          * allocation.
3972          */
3973         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
3974                 int ret = btrfs_zone_finish_one_bg(fs_info);
3975
3976                 if (ret == 1)
3977                         return 0;
3978                 else if (ret < 0)
3979                         return ret;
3980         }
3981
3982         /*
3983          * If we have enough free space left in an already active block group
3984          * and we can't activate any other zone now, do not allow allocating a
3985          * new chunk and let find_free_extent() retry with a smaller size.
3986          */
3987         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
3988                 return -ENOSPC;
3989
3990         /*
3991          * Even min_alloc_size is not left in any block groups. Since we cannot
3992          * activate a new block group, allocating it may not help. Let's tell a
3993          * caller to try again and hope it progress something by writing some
3994          * parts of the region. That is only possible for data block groups,
3995          * where a part of the region can be written.
3996          */
3997         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
3998                 return -EAGAIN;
3999
4000         /*
4001          * We cannot activate a new block group and no enough space left in any
4002          * block groups. So, allocating a new block group may not help. But,
4003          * there is nothing to do anyway, so let's go with it.
4004          */
4005         return 0;
4006 }
4007
4008 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4009                               struct find_free_extent_ctl *ffe_ctl)
4010 {
4011         switch (ffe_ctl->policy) {
4012         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4013                 return 0;
4014         case BTRFS_EXTENT_ALLOC_ZONED:
4015                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4016         default:
4017                 BUG();
4018         }
4019 }
4020
4021 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
4022 {
4023         switch (ffe_ctl->policy) {
4024         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4025                 /*
4026                  * If we can't allocate a new chunk we've already looped through
4027                  * at least once, move on to the NO_EMPTY_SIZE case.
4028                  */
4029                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
4030                 return 0;
4031         case BTRFS_EXTENT_ALLOC_ZONED:
4032                 /* Give up here */
4033                 return -ENOSPC;
4034         default:
4035                 BUG();
4036         }
4037 }
4038
4039 /*
4040  * Return >0 means caller needs to re-search for free extent
4041  * Return 0 means we have the needed free extent.
4042  * Return <0 means we failed to locate any free extent.
4043  */
4044 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4045                                         struct btrfs_key *ins,
4046                                         struct find_free_extent_ctl *ffe_ctl,
4047                                         bool full_search)
4048 {
4049         struct btrfs_root *root = fs_info->chunk_root;
4050         int ret;
4051
4052         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4053             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4054                 ffe_ctl->orig_have_caching_bg = true;
4055
4056         if (ins->objectid) {
4057                 found_extent(ffe_ctl, ins);
4058                 return 0;
4059         }
4060
4061         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4062                 return 1;
4063
4064         ffe_ctl->index++;
4065         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4066                 return 1;
4067
4068         /*
4069          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
4070          *                      caching kthreads as we move along
4071          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
4072          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
4073          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
4074          *                     again
4075          */
4076         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4077                 ffe_ctl->index = 0;
4078                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
4079                         /*
4080                          * We want to skip the LOOP_CACHING_WAIT step if we
4081                          * don't have any uncached bgs and we've already done a
4082                          * full search through.
4083                          */
4084                         if (ffe_ctl->orig_have_caching_bg || !full_search)
4085                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
4086                         else
4087                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
4088                 } else {
4089                         ffe_ctl->loop++;
4090                 }
4091
4092                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4093                         struct btrfs_trans_handle *trans;
4094                         int exist = 0;
4095
4096                         /*Check if allocation policy allows to create a new chunk */
4097                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4098                         if (ret)
4099                                 return ret;
4100
4101                         trans = current->journal_info;
4102                         if (trans)
4103                                 exist = 1;
4104                         else
4105                                 trans = btrfs_join_transaction(root);
4106
4107                         if (IS_ERR(trans)) {
4108                                 ret = PTR_ERR(trans);
4109                                 return ret;
4110                         }
4111
4112                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4113                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4114
4115                         /* Do not bail out on ENOSPC since we can do more. */
4116                         if (ret == -ENOSPC)
4117                                 ret = chunk_allocation_failed(ffe_ctl);
4118                         else if (ret < 0)
4119                                 btrfs_abort_transaction(trans, ret);
4120                         else
4121                                 ret = 0;
4122                         if (!exist)
4123                                 btrfs_end_transaction(trans);
4124                         if (ret)
4125                                 return ret;
4126                 }
4127
4128                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4129                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4130                                 return -ENOSPC;
4131
4132                         /*
4133                          * Don't loop again if we already have no empty_size and
4134                          * no empty_cluster.
4135                          */
4136                         if (ffe_ctl->empty_size == 0 &&
4137                             ffe_ctl->empty_cluster == 0)
4138                                 return -ENOSPC;
4139                         ffe_ctl->empty_size = 0;
4140                         ffe_ctl->empty_cluster = 0;
4141                 }
4142                 return 1;
4143         }
4144         return -ENOSPC;
4145 }
4146
4147 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4148                                         struct find_free_extent_ctl *ffe_ctl,
4149                                         struct btrfs_space_info *space_info,
4150                                         struct btrfs_key *ins)
4151 {
4152         /*
4153          * If our free space is heavily fragmented we may not be able to make
4154          * big contiguous allocations, so instead of doing the expensive search
4155          * for free space, simply return ENOSPC with our max_extent_size so we
4156          * can go ahead and search for a more manageable chunk.
4157          *
4158          * If our max_extent_size is large enough for our allocation simply
4159          * disable clustering since we will likely not be able to find enough
4160          * space to create a cluster and induce latency trying.
4161          */
4162         if (space_info->max_extent_size) {
4163                 spin_lock(&space_info->lock);
4164                 if (space_info->max_extent_size &&
4165                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4166                         ins->offset = space_info->max_extent_size;
4167                         spin_unlock(&space_info->lock);
4168                         return -ENOSPC;
4169                 } else if (space_info->max_extent_size) {
4170                         ffe_ctl->use_cluster = false;
4171                 }
4172                 spin_unlock(&space_info->lock);
4173         }
4174
4175         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4176                                                &ffe_ctl->empty_cluster);
4177         if (ffe_ctl->last_ptr) {
4178                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4179
4180                 spin_lock(&last_ptr->lock);
4181                 if (last_ptr->block_group)
4182                         ffe_ctl->hint_byte = last_ptr->window_start;
4183                 if (last_ptr->fragmented) {
4184                         /*
4185                          * We still set window_start so we can keep track of the
4186                          * last place we found an allocation to try and save
4187                          * some time.
4188                          */
4189                         ffe_ctl->hint_byte = last_ptr->window_start;
4190                         ffe_ctl->use_cluster = false;
4191                 }
4192                 spin_unlock(&last_ptr->lock);
4193         }
4194
4195         return 0;
4196 }
4197
4198 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4199                               struct find_free_extent_ctl *ffe_ctl,
4200                               struct btrfs_space_info *space_info,
4201                               struct btrfs_key *ins)
4202 {
4203         switch (ffe_ctl->policy) {
4204         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4205                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4206                                                     space_info, ins);
4207         case BTRFS_EXTENT_ALLOC_ZONED:
4208                 if (ffe_ctl->for_treelog) {
4209                         spin_lock(&fs_info->treelog_bg_lock);
4210                         if (fs_info->treelog_bg)
4211                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4212                         spin_unlock(&fs_info->treelog_bg_lock);
4213                 }
4214                 if (ffe_ctl->for_data_reloc) {
4215                         spin_lock(&fs_info->relocation_bg_lock);
4216                         if (fs_info->data_reloc_bg)
4217                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4218                         spin_unlock(&fs_info->relocation_bg_lock);
4219                 }
4220                 return 0;
4221         default:
4222                 BUG();
4223         }
4224 }
4225
4226 /*
4227  * walks the btree of allocated extents and find a hole of a given size.
4228  * The key ins is changed to record the hole:
4229  * ins->objectid == start position
4230  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4231  * ins->offset == the size of the hole.
4232  * Any available blocks before search_start are skipped.
4233  *
4234  * If there is no suitable free space, we will record the max size of
4235  * the free space extent currently.
4236  *
4237  * The overall logic and call chain:
4238  *
4239  * find_free_extent()
4240  * |- Iterate through all block groups
4241  * |  |- Get a valid block group
4242  * |  |- Try to do clustered allocation in that block group
4243  * |  |- Try to do unclustered allocation in that block group
4244  * |  |- Check if the result is valid
4245  * |  |  |- If valid, then exit
4246  * |  |- Jump to next block group
4247  * |
4248  * |- Push harder to find free extents
4249  *    |- If not found, re-iterate all block groups
4250  */
4251 static noinline int find_free_extent(struct btrfs_root *root,
4252                                      struct btrfs_key *ins,
4253                                      struct find_free_extent_ctl *ffe_ctl)
4254 {
4255         struct btrfs_fs_info *fs_info = root->fs_info;
4256         int ret = 0;
4257         int cache_block_group_error = 0;
4258         struct btrfs_block_group *block_group = NULL;
4259         struct btrfs_space_info *space_info;
4260         bool full_search = false;
4261
4262         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4263
4264         ffe_ctl->search_start = 0;
4265         /* For clustered allocation */
4266         ffe_ctl->empty_cluster = 0;
4267         ffe_ctl->last_ptr = NULL;
4268         ffe_ctl->use_cluster = true;
4269         ffe_ctl->have_caching_bg = false;
4270         ffe_ctl->orig_have_caching_bg = false;
4271         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4272         ffe_ctl->loop = 0;
4273         /* For clustered allocation */
4274         ffe_ctl->retry_clustered = false;
4275         ffe_ctl->retry_unclustered = false;
4276         ffe_ctl->cached = 0;
4277         ffe_ctl->max_extent_size = 0;
4278         ffe_ctl->total_free_space = 0;
4279         ffe_ctl->found_offset = 0;
4280         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4281
4282         if (btrfs_is_zoned(fs_info))
4283                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4284
4285         ins->type = BTRFS_EXTENT_ITEM_KEY;
4286         ins->objectid = 0;
4287         ins->offset = 0;
4288
4289         trace_find_free_extent(root, ffe_ctl->num_bytes, ffe_ctl->empty_size,
4290                                ffe_ctl->flags);
4291
4292         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4293         if (!space_info) {
4294                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4295                 return -ENOSPC;
4296         }
4297
4298         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4299         if (ret < 0)
4300                 return ret;
4301
4302         ffe_ctl->search_start = max(ffe_ctl->search_start,
4303                                     first_logical_byte(fs_info));
4304         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4305         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4306                 block_group = btrfs_lookup_block_group(fs_info,
4307                                                        ffe_ctl->search_start);
4308                 /*
4309                  * we don't want to use the block group if it doesn't match our
4310                  * allocation bits, or if its not cached.
4311                  *
4312                  * However if we are re-searching with an ideal block group
4313                  * picked out then we don't care that the block group is cached.
4314                  */
4315                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4316                     block_group->cached != BTRFS_CACHE_NO) {
4317                         down_read(&space_info->groups_sem);
4318                         if (list_empty(&block_group->list) ||
4319                             block_group->ro) {
4320                                 /*
4321                                  * someone is removing this block group,
4322                                  * we can't jump into the have_block_group
4323                                  * target because our list pointers are not
4324                                  * valid
4325                                  */
4326                                 btrfs_put_block_group(block_group);
4327                                 up_read(&space_info->groups_sem);
4328                         } else {
4329                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4330                                                         block_group->flags);
4331                                 btrfs_lock_block_group(block_group,
4332                                                        ffe_ctl->delalloc);
4333                                 goto have_block_group;
4334                         }
4335                 } else if (block_group) {
4336                         btrfs_put_block_group(block_group);
4337                 }
4338         }
4339 search:
4340         ffe_ctl->have_caching_bg = false;
4341         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4342             ffe_ctl->index == 0)
4343                 full_search = true;
4344         down_read(&space_info->groups_sem);
4345         list_for_each_entry(block_group,
4346                             &space_info->block_groups[ffe_ctl->index], list) {
4347                 struct btrfs_block_group *bg_ret;
4348
4349                 /* If the block group is read-only, we can skip it entirely. */
4350                 if (unlikely(block_group->ro)) {
4351                         if (ffe_ctl->for_treelog)
4352                                 btrfs_clear_treelog_bg(block_group);
4353                         if (ffe_ctl->for_data_reloc)
4354                                 btrfs_clear_data_reloc_bg(block_group);
4355                         continue;
4356                 }
4357
4358                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4359                 ffe_ctl->search_start = block_group->start;
4360
4361                 /*
4362                  * this can happen if we end up cycling through all the
4363                  * raid types, but we want to make sure we only allocate
4364                  * for the proper type.
4365                  */
4366                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4367                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4368                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4369                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4370                                 BTRFS_BLOCK_GROUP_RAID10;
4371
4372                         /*
4373                          * if they asked for extra copies and this block group
4374                          * doesn't provide them, bail.  This does allow us to
4375                          * fill raid0 from raid1.
4376                          */
4377                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4378                                 goto loop;
4379
4380                         /*
4381                          * This block group has different flags than we want.
4382                          * It's possible that we have MIXED_GROUP flag but no
4383                          * block group is mixed.  Just skip such block group.
4384                          */
4385                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4386                         continue;
4387                 }
4388
4389 have_block_group:
4390                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4391                 if (unlikely(!ffe_ctl->cached)) {
4392                         ffe_ctl->have_caching_bg = true;
4393                         ret = btrfs_cache_block_group(block_group, false);
4394
4395                         /*
4396                          * If we get ENOMEM here or something else we want to
4397                          * try other block groups, because it may not be fatal.
4398                          * However if we can't find anything else we need to
4399                          * save our return here so that we return the actual
4400                          * error that caused problems, not ENOSPC.
4401                          */
4402                         if (ret < 0) {
4403                                 if (!cache_block_group_error)
4404                                         cache_block_group_error = ret;
4405                                 ret = 0;
4406                                 goto loop;
4407                         }
4408                         ret = 0;
4409                 }
4410
4411                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4412                         goto loop;
4413
4414                 bg_ret = NULL;
4415                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4416                 if (ret == 0) {
4417                         if (bg_ret && bg_ret != block_group) {
4418                                 btrfs_release_block_group(block_group,
4419                                                           ffe_ctl->delalloc);
4420                                 block_group = bg_ret;
4421                         }
4422                 } else if (ret == -EAGAIN) {
4423                         goto have_block_group;
4424                 } else if (ret > 0) {
4425                         goto loop;
4426                 }
4427
4428                 /* Checks */
4429                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4430                                                  fs_info->stripesize);
4431
4432                 /* move on to the next group */
4433                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4434                     block_group->start + block_group->length) {
4435                         btrfs_add_free_space_unused(block_group,
4436                                             ffe_ctl->found_offset,
4437                                             ffe_ctl->num_bytes);
4438                         goto loop;
4439                 }
4440
4441                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4442                         btrfs_add_free_space_unused(block_group,
4443                                         ffe_ctl->found_offset,
4444                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4445
4446                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4447                                                ffe_ctl->num_bytes,
4448                                                ffe_ctl->delalloc);
4449                 if (ret == -EAGAIN) {
4450                         btrfs_add_free_space_unused(block_group,
4451                                         ffe_ctl->found_offset,
4452                                         ffe_ctl->num_bytes);
4453                         goto loop;
4454                 }
4455                 btrfs_inc_block_group_reservations(block_group);
4456
4457                 /* we are all good, lets return */
4458                 ins->objectid = ffe_ctl->search_start;
4459                 ins->offset = ffe_ctl->num_bytes;
4460
4461                 trace_btrfs_reserve_extent(block_group, ffe_ctl->search_start,
4462                                            ffe_ctl->num_bytes);
4463                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4464                 break;
4465 loop:
4466                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4467                 cond_resched();
4468         }
4469         up_read(&space_info->groups_sem);
4470
4471         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4472         if (ret > 0)
4473                 goto search;
4474
4475         if (ret == -ENOSPC && !cache_block_group_error) {
4476                 /*
4477                  * Use ffe_ctl->total_free_space as fallback if we can't find
4478                  * any contiguous hole.
4479                  */
4480                 if (!ffe_ctl->max_extent_size)
4481                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4482                 spin_lock(&space_info->lock);
4483                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4484                 spin_unlock(&space_info->lock);
4485                 ins->offset = ffe_ctl->max_extent_size;
4486         } else if (ret == -ENOSPC) {
4487                 ret = cache_block_group_error;
4488         }
4489         return ret;
4490 }
4491
4492 /*
4493  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4494  *                        hole that is at least as big as @num_bytes.
4495  *
4496  * @root           -    The root that will contain this extent
4497  *
4498  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4499  *                      is used for accounting purposes. This value differs
4500  *                      from @num_bytes only in the case of compressed extents.
4501  *
4502  * @num_bytes      -    Number of bytes to allocate on-disk.
4503  *
4504  * @min_alloc_size -    Indicates the minimum amount of space that the
4505  *                      allocator should try to satisfy. In some cases
4506  *                      @num_bytes may be larger than what is required and if
4507  *                      the filesystem is fragmented then allocation fails.
4508  *                      However, the presence of @min_alloc_size gives a
4509  *                      chance to try and satisfy the smaller allocation.
4510  *
4511  * @empty_size     -    A hint that you plan on doing more COW. This is the
4512  *                      size in bytes the allocator should try to find free
4513  *                      next to the block it returns.  This is just a hint and
4514  *                      may be ignored by the allocator.
4515  *
4516  * @hint_byte      -    Hint to the allocator to start searching above the byte
4517  *                      address passed. It might be ignored.
4518  *
4519  * @ins            -    This key is modified to record the found hole. It will
4520  *                      have the following values:
4521  *                      ins->objectid == start position
4522  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4523  *                      ins->offset == the size of the hole.
4524  *
4525  * @is_data        -    Boolean flag indicating whether an extent is
4526  *                      allocated for data (true) or metadata (false)
4527  *
4528  * @delalloc       -    Boolean flag indicating whether this allocation is for
4529  *                      delalloc or not. If 'true' data_rwsem of block groups
4530  *                      is going to be acquired.
4531  *
4532  *
4533  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4534  * case -ENOSPC is returned then @ins->offset will contain the size of the
4535  * largest available hole the allocator managed to find.
4536  */
4537 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4538                          u64 num_bytes, u64 min_alloc_size,
4539                          u64 empty_size, u64 hint_byte,
4540                          struct btrfs_key *ins, int is_data, int delalloc)
4541 {
4542         struct btrfs_fs_info *fs_info = root->fs_info;
4543         struct find_free_extent_ctl ffe_ctl = {};
4544         bool final_tried = num_bytes == min_alloc_size;
4545         u64 flags;
4546         int ret;
4547         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4548         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4549
4550         flags = get_alloc_profile_by_root(root, is_data);
4551 again:
4552         WARN_ON(num_bytes < fs_info->sectorsize);
4553
4554         ffe_ctl.ram_bytes = ram_bytes;
4555         ffe_ctl.num_bytes = num_bytes;
4556         ffe_ctl.min_alloc_size = min_alloc_size;
4557         ffe_ctl.empty_size = empty_size;
4558         ffe_ctl.flags = flags;
4559         ffe_ctl.delalloc = delalloc;
4560         ffe_ctl.hint_byte = hint_byte;
4561         ffe_ctl.for_treelog = for_treelog;
4562         ffe_ctl.for_data_reloc = for_data_reloc;
4563
4564         ret = find_free_extent(root, ins, &ffe_ctl);
4565         if (!ret && !is_data) {
4566                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4567         } else if (ret == -ENOSPC) {
4568                 if (!final_tried && ins->offset) {
4569                         num_bytes = min(num_bytes >> 1, ins->offset);
4570                         num_bytes = round_down(num_bytes,
4571                                                fs_info->sectorsize);
4572                         num_bytes = max(num_bytes, min_alloc_size);
4573                         ram_bytes = num_bytes;
4574                         if (num_bytes == min_alloc_size)
4575                                 final_tried = true;
4576                         goto again;
4577                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4578                         struct btrfs_space_info *sinfo;
4579
4580                         sinfo = btrfs_find_space_info(fs_info, flags);
4581                         btrfs_err(fs_info,
4582         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4583                                   flags, num_bytes, for_treelog, for_data_reloc);
4584                         if (sinfo)
4585                                 btrfs_dump_space_info(fs_info, sinfo,
4586                                                       num_bytes, 1);
4587                 }
4588         }
4589
4590         return ret;
4591 }
4592
4593 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4594                                u64 start, u64 len, int delalloc)
4595 {
4596         struct btrfs_block_group *cache;
4597
4598         cache = btrfs_lookup_block_group(fs_info, start);
4599         if (!cache) {
4600                 btrfs_err(fs_info, "Unable to find block group for %llu",
4601                           start);
4602                 return -ENOSPC;
4603         }
4604
4605         btrfs_add_free_space(cache, start, len);
4606         btrfs_free_reserved_bytes(cache, len, delalloc);
4607         trace_btrfs_reserved_extent_free(fs_info, start, len);
4608
4609         btrfs_put_block_group(cache);
4610         return 0;
4611 }
4612
4613 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4614                               u64 len)
4615 {
4616         struct btrfs_block_group *cache;
4617         int ret = 0;
4618
4619         cache = btrfs_lookup_block_group(trans->fs_info, start);
4620         if (!cache) {
4621                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4622                           start);
4623                 return -ENOSPC;
4624         }
4625
4626         ret = pin_down_extent(trans, cache, start, len, 1);
4627         btrfs_put_block_group(cache);
4628         return ret;
4629 }
4630
4631 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4632                                  u64 num_bytes)
4633 {
4634         struct btrfs_fs_info *fs_info = trans->fs_info;
4635         int ret;
4636
4637         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4638         if (ret)
4639                 return ret;
4640
4641         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4642         if (ret) {
4643                 ASSERT(!ret);
4644                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4645                           bytenr, num_bytes);
4646                 return ret;
4647         }
4648
4649         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4650         return 0;
4651 }
4652
4653 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4654                                       u64 parent, u64 root_objectid,
4655                                       u64 flags, u64 owner, u64 offset,
4656                                       struct btrfs_key *ins, int ref_mod)
4657 {
4658         struct btrfs_fs_info *fs_info = trans->fs_info;
4659         struct btrfs_root *extent_root;
4660         int ret;
4661         struct btrfs_extent_item *extent_item;
4662         struct btrfs_extent_inline_ref *iref;
4663         struct btrfs_path *path;
4664         struct extent_buffer *leaf;
4665         int type;
4666         u32 size;
4667
4668         if (parent > 0)
4669                 type = BTRFS_SHARED_DATA_REF_KEY;
4670         else
4671                 type = BTRFS_EXTENT_DATA_REF_KEY;
4672
4673         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4674
4675         path = btrfs_alloc_path();
4676         if (!path)
4677                 return -ENOMEM;
4678
4679         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4680         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4681         if (ret) {
4682                 btrfs_free_path(path);
4683                 return ret;
4684         }
4685
4686         leaf = path->nodes[0];
4687         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4688                                      struct btrfs_extent_item);
4689         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4690         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4691         btrfs_set_extent_flags(leaf, extent_item,
4692                                flags | BTRFS_EXTENT_FLAG_DATA);
4693
4694         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4695         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4696         if (parent > 0) {
4697                 struct btrfs_shared_data_ref *ref;
4698                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4699                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4700                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4701         } else {
4702                 struct btrfs_extent_data_ref *ref;
4703                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4704                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4705                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4706                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4707                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4708         }
4709
4710         btrfs_mark_buffer_dirty(path->nodes[0]);
4711         btrfs_free_path(path);
4712
4713         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4714 }
4715
4716 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4717                                      struct btrfs_delayed_ref_node *node,
4718                                      struct btrfs_delayed_extent_op *extent_op)
4719 {
4720         struct btrfs_fs_info *fs_info = trans->fs_info;
4721         struct btrfs_root *extent_root;
4722         int ret;
4723         struct btrfs_extent_item *extent_item;
4724         struct btrfs_key extent_key;
4725         struct btrfs_tree_block_info *block_info;
4726         struct btrfs_extent_inline_ref *iref;
4727         struct btrfs_path *path;
4728         struct extent_buffer *leaf;
4729         struct btrfs_delayed_tree_ref *ref;
4730         u32 size = sizeof(*extent_item) + sizeof(*iref);
4731         u64 flags = extent_op->flags_to_set;
4732         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4733
4734         ref = btrfs_delayed_node_to_tree_ref(node);
4735
4736         extent_key.objectid = node->bytenr;
4737         if (skinny_metadata) {
4738                 extent_key.offset = ref->level;
4739                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4740         } else {
4741                 extent_key.offset = node->num_bytes;
4742                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4743                 size += sizeof(*block_info);
4744         }
4745
4746         path = btrfs_alloc_path();
4747         if (!path)
4748                 return -ENOMEM;
4749
4750         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4751         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4752                                       size);
4753         if (ret) {
4754                 btrfs_free_path(path);
4755                 return ret;
4756         }
4757
4758         leaf = path->nodes[0];
4759         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4760                                      struct btrfs_extent_item);
4761         btrfs_set_extent_refs(leaf, extent_item, 1);
4762         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4763         btrfs_set_extent_flags(leaf, extent_item,
4764                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4765
4766         if (skinny_metadata) {
4767                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4768         } else {
4769                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4770                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4771                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4772                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4773         }
4774
4775         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4776                 btrfs_set_extent_inline_ref_type(leaf, iref,
4777                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4778                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4779         } else {
4780                 btrfs_set_extent_inline_ref_type(leaf, iref,
4781                                                  BTRFS_TREE_BLOCK_REF_KEY);
4782                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4783         }
4784
4785         btrfs_mark_buffer_dirty(leaf);
4786         btrfs_free_path(path);
4787
4788         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4789 }
4790
4791 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4792                                      struct btrfs_root *root, u64 owner,
4793                                      u64 offset, u64 ram_bytes,
4794                                      struct btrfs_key *ins)
4795 {
4796         struct btrfs_ref generic_ref = { 0 };
4797
4798         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4799
4800         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4801                                ins->objectid, ins->offset, 0);
4802         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner,
4803                             offset, 0, false);
4804         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4805
4806         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4807 }
4808
4809 /*
4810  * this is used by the tree logging recovery code.  It records that
4811  * an extent has been allocated and makes sure to clear the free
4812  * space cache bits as well
4813  */
4814 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4815                                    u64 root_objectid, u64 owner, u64 offset,
4816                                    struct btrfs_key *ins)
4817 {
4818         struct btrfs_fs_info *fs_info = trans->fs_info;
4819         int ret;
4820         struct btrfs_block_group *block_group;
4821         struct btrfs_space_info *space_info;
4822
4823         /*
4824          * Mixed block groups will exclude before processing the log so we only
4825          * need to do the exclude dance if this fs isn't mixed.
4826          */
4827         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4828                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4829                                               ins->offset);
4830                 if (ret)
4831                         return ret;
4832         }
4833
4834         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4835         if (!block_group)
4836                 return -EINVAL;
4837
4838         space_info = block_group->space_info;
4839         spin_lock(&space_info->lock);
4840         spin_lock(&block_group->lock);
4841         space_info->bytes_reserved += ins->offset;
4842         block_group->reserved += ins->offset;
4843         spin_unlock(&block_group->lock);
4844         spin_unlock(&space_info->lock);
4845
4846         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4847                                          offset, ins, 1);
4848         if (ret)
4849                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4850         btrfs_put_block_group(block_group);
4851         return ret;
4852 }
4853
4854 static struct extent_buffer *
4855 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4856                       u64 bytenr, int level, u64 owner,
4857                       enum btrfs_lock_nesting nest)
4858 {
4859         struct btrfs_fs_info *fs_info = root->fs_info;
4860         struct extent_buffer *buf;
4861         u64 lockdep_owner = owner;
4862
4863         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4864         if (IS_ERR(buf))
4865                 return buf;
4866
4867         /*
4868          * Extra safety check in case the extent tree is corrupted and extent
4869          * allocator chooses to use a tree block which is already used and
4870          * locked.
4871          */
4872         if (buf->lock_owner == current->pid) {
4873                 btrfs_err_rl(fs_info,
4874 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4875                         buf->start, btrfs_header_owner(buf), current->pid);
4876                 free_extent_buffer(buf);
4877                 return ERR_PTR(-EUCLEAN);
4878         }
4879
4880         /*
4881          * The reloc trees are just snapshots, so we need them to appear to be
4882          * just like any other fs tree WRT lockdep.
4883          *
4884          * The exception however is in replace_path() in relocation, where we
4885          * hold the lock on the original fs root and then search for the reloc
4886          * root.  At that point we need to make sure any reloc root buffers are
4887          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
4888          * lockdep happy.
4889          */
4890         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
4891             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
4892                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4893
4894         /* btrfs_clean_tree_block() accesses generation field. */
4895         btrfs_set_header_generation(buf, trans->transid);
4896
4897         /*
4898          * This needs to stay, because we could allocate a freed block from an
4899          * old tree into a new tree, so we need to make sure this new block is
4900          * set to the appropriate level and owner.
4901          */
4902         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
4903
4904         __btrfs_tree_lock(buf, nest);
4905         btrfs_clean_tree_block(buf);
4906         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4907         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4908
4909         set_extent_buffer_uptodate(buf);
4910
4911         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4912         btrfs_set_header_level(buf, level);
4913         btrfs_set_header_bytenr(buf, buf->start);
4914         btrfs_set_header_generation(buf, trans->transid);
4915         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4916         btrfs_set_header_owner(buf, owner);
4917         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4918         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4919         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4920                 buf->log_index = root->log_transid % 2;
4921                 /*
4922                  * we allow two log transactions at a time, use different
4923                  * EXTENT bit to differentiate dirty pages.
4924                  */
4925                 if (buf->log_index == 0)
4926                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4927                                         buf->start + buf->len - 1, GFP_NOFS);
4928                 else
4929                         set_extent_new(&root->dirty_log_pages, buf->start,
4930                                         buf->start + buf->len - 1);
4931         } else {
4932                 buf->log_index = -1;
4933                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4934                          buf->start + buf->len - 1, GFP_NOFS);
4935         }
4936         /* this returns a buffer locked for blocking */
4937         return buf;
4938 }
4939
4940 /*
4941  * finds a free extent and does all the dirty work required for allocation
4942  * returns the tree buffer or an ERR_PTR on error.
4943  */
4944 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4945                                              struct btrfs_root *root,
4946                                              u64 parent, u64 root_objectid,
4947                                              const struct btrfs_disk_key *key,
4948                                              int level, u64 hint,
4949                                              u64 empty_size,
4950                                              enum btrfs_lock_nesting nest)
4951 {
4952         struct btrfs_fs_info *fs_info = root->fs_info;
4953         struct btrfs_key ins;
4954         struct btrfs_block_rsv *block_rsv;
4955         struct extent_buffer *buf;
4956         struct btrfs_delayed_extent_op *extent_op;
4957         struct btrfs_ref generic_ref = { 0 };
4958         u64 flags = 0;
4959         int ret;
4960         u32 blocksize = fs_info->nodesize;
4961         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4962
4963 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4964         if (btrfs_is_testing(fs_info)) {
4965                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4966                                             level, root_objectid, nest);
4967                 if (!IS_ERR(buf))
4968                         root->alloc_bytenr += blocksize;
4969                 return buf;
4970         }
4971 #endif
4972
4973         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4974         if (IS_ERR(block_rsv))
4975                 return ERR_CAST(block_rsv);
4976
4977         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4978                                    empty_size, hint, &ins, 0, 0);
4979         if (ret)
4980                 goto out_unuse;
4981
4982         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4983                                     root_objectid, nest);
4984         if (IS_ERR(buf)) {
4985                 ret = PTR_ERR(buf);
4986                 goto out_free_reserved;
4987         }
4988
4989         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4990                 if (parent == 0)
4991                         parent = ins.objectid;
4992                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4993         } else
4994                 BUG_ON(parent > 0);
4995
4996         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4997                 extent_op = btrfs_alloc_delayed_extent_op();
4998                 if (!extent_op) {
4999                         ret = -ENOMEM;
5000                         goto out_free_buf;
5001                 }
5002                 if (key)
5003                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5004                 else
5005                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5006                 extent_op->flags_to_set = flags;
5007                 extent_op->update_key = skinny_metadata ? false : true;
5008                 extent_op->update_flags = true;
5009                 extent_op->level = level;
5010
5011                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
5012                                        ins.objectid, ins.offset, parent);
5013                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
5014                                     root->root_key.objectid, false);
5015                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5016                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5017                 if (ret)
5018                         goto out_free_delayed;
5019         }
5020         return buf;
5021
5022 out_free_delayed:
5023         btrfs_free_delayed_extent_op(extent_op);
5024 out_free_buf:
5025         btrfs_tree_unlock(buf);
5026         free_extent_buffer(buf);
5027 out_free_reserved:
5028         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5029 out_unuse:
5030         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5031         return ERR_PTR(ret);
5032 }
5033
5034 struct walk_control {
5035         u64 refs[BTRFS_MAX_LEVEL];
5036         u64 flags[BTRFS_MAX_LEVEL];
5037         struct btrfs_key update_progress;
5038         struct btrfs_key drop_progress;
5039         int drop_level;
5040         int stage;
5041         int level;
5042         int shared_level;
5043         int update_ref;
5044         int keep_locks;
5045         int reada_slot;
5046         int reada_count;
5047         int restarted;
5048 };
5049
5050 #define DROP_REFERENCE  1
5051 #define UPDATE_BACKREF  2
5052
5053 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5054                                      struct btrfs_root *root,
5055                                      struct walk_control *wc,
5056                                      struct btrfs_path *path)
5057 {
5058         struct btrfs_fs_info *fs_info = root->fs_info;
5059         u64 bytenr;
5060         u64 generation;
5061         u64 refs;
5062         u64 flags;
5063         u32 nritems;
5064         struct btrfs_key key;
5065         struct extent_buffer *eb;
5066         int ret;
5067         int slot;
5068         int nread = 0;
5069
5070         if (path->slots[wc->level] < wc->reada_slot) {
5071                 wc->reada_count = wc->reada_count * 2 / 3;
5072                 wc->reada_count = max(wc->reada_count, 2);
5073         } else {
5074                 wc->reada_count = wc->reada_count * 3 / 2;
5075                 wc->reada_count = min_t(int, wc->reada_count,
5076                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5077         }
5078
5079         eb = path->nodes[wc->level];
5080         nritems = btrfs_header_nritems(eb);
5081
5082         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5083                 if (nread >= wc->reada_count)
5084                         break;
5085
5086                 cond_resched();
5087                 bytenr = btrfs_node_blockptr(eb, slot);
5088                 generation = btrfs_node_ptr_generation(eb, slot);
5089
5090                 if (slot == path->slots[wc->level])
5091                         goto reada;
5092
5093                 if (wc->stage == UPDATE_BACKREF &&
5094                     generation <= root->root_key.offset)
5095                         continue;
5096
5097                 /* We don't lock the tree block, it's OK to be racy here */
5098                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5099                                                wc->level - 1, 1, &refs,
5100                                                &flags);
5101                 /* We don't care about errors in readahead. */
5102                 if (ret < 0)
5103                         continue;
5104                 BUG_ON(refs == 0);
5105
5106                 if (wc->stage == DROP_REFERENCE) {
5107                         if (refs == 1)
5108                                 goto reada;
5109
5110                         if (wc->level == 1 &&
5111                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5112                                 continue;
5113                         if (!wc->update_ref ||
5114                             generation <= root->root_key.offset)
5115                                 continue;
5116                         btrfs_node_key_to_cpu(eb, &key, slot);
5117                         ret = btrfs_comp_cpu_keys(&key,
5118                                                   &wc->update_progress);
5119                         if (ret < 0)
5120                                 continue;
5121                 } else {
5122                         if (wc->level == 1 &&
5123                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5124                                 continue;
5125                 }
5126 reada:
5127                 btrfs_readahead_node_child(eb, slot);
5128                 nread++;
5129         }
5130         wc->reada_slot = slot;
5131 }
5132
5133 /*
5134  * helper to process tree block while walking down the tree.
5135  *
5136  * when wc->stage == UPDATE_BACKREF, this function updates
5137  * back refs for pointers in the block.
5138  *
5139  * NOTE: return value 1 means we should stop walking down.
5140  */
5141 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5142                                    struct btrfs_root *root,
5143                                    struct btrfs_path *path,
5144                                    struct walk_control *wc, int lookup_info)
5145 {
5146         struct btrfs_fs_info *fs_info = root->fs_info;
5147         int level = wc->level;
5148         struct extent_buffer *eb = path->nodes[level];
5149         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5150         int ret;
5151
5152         if (wc->stage == UPDATE_BACKREF &&
5153             btrfs_header_owner(eb) != root->root_key.objectid)
5154                 return 1;
5155
5156         /*
5157          * when reference count of tree block is 1, it won't increase
5158          * again. once full backref flag is set, we never clear it.
5159          */
5160         if (lookup_info &&
5161             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5162              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5163                 BUG_ON(!path->locks[level]);
5164                 ret = btrfs_lookup_extent_info(trans, fs_info,
5165                                                eb->start, level, 1,
5166                                                &wc->refs[level],
5167                                                &wc->flags[level]);
5168                 BUG_ON(ret == -ENOMEM);
5169                 if (ret)
5170                         return ret;
5171                 BUG_ON(wc->refs[level] == 0);
5172         }
5173
5174         if (wc->stage == DROP_REFERENCE) {
5175                 if (wc->refs[level] > 1)
5176                         return 1;
5177
5178                 if (path->locks[level] && !wc->keep_locks) {
5179                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5180                         path->locks[level] = 0;
5181                 }
5182                 return 0;
5183         }
5184
5185         /* wc->stage == UPDATE_BACKREF */
5186         if (!(wc->flags[level] & flag)) {
5187                 BUG_ON(!path->locks[level]);
5188                 ret = btrfs_inc_ref(trans, root, eb, 1);
5189                 BUG_ON(ret); /* -ENOMEM */
5190                 ret = btrfs_dec_ref(trans, root, eb, 0);
5191                 BUG_ON(ret); /* -ENOMEM */
5192                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5193                                                   btrfs_header_level(eb));
5194                 BUG_ON(ret); /* -ENOMEM */
5195                 wc->flags[level] |= flag;
5196         }
5197
5198         /*
5199          * the block is shared by multiple trees, so it's not good to
5200          * keep the tree lock
5201          */
5202         if (path->locks[level] && level > 0) {
5203                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5204                 path->locks[level] = 0;
5205         }
5206         return 0;
5207 }
5208
5209 /*
5210  * This is used to verify a ref exists for this root to deal with a bug where we
5211  * would have a drop_progress key that hadn't been updated properly.
5212  */
5213 static int check_ref_exists(struct btrfs_trans_handle *trans,
5214                             struct btrfs_root *root, u64 bytenr, u64 parent,
5215                             int level)
5216 {
5217         struct btrfs_path *path;
5218         struct btrfs_extent_inline_ref *iref;
5219         int ret;
5220
5221         path = btrfs_alloc_path();
5222         if (!path)
5223                 return -ENOMEM;
5224
5225         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5226                                     root->fs_info->nodesize, parent,
5227                                     root->root_key.objectid, level, 0);
5228         btrfs_free_path(path);
5229         if (ret == -ENOENT)
5230                 return 0;
5231         if (ret < 0)
5232                 return ret;
5233         return 1;
5234 }
5235
5236 /*
5237  * helper to process tree block pointer.
5238  *
5239  * when wc->stage == DROP_REFERENCE, this function checks
5240  * reference count of the block pointed to. if the block
5241  * is shared and we need update back refs for the subtree
5242  * rooted at the block, this function changes wc->stage to
5243  * UPDATE_BACKREF. if the block is shared and there is no
5244  * need to update back, this function drops the reference
5245  * to the block.
5246  *
5247  * NOTE: return value 1 means we should stop walking down.
5248  */
5249 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5250                                  struct btrfs_root *root,
5251                                  struct btrfs_path *path,
5252                                  struct walk_control *wc, int *lookup_info)
5253 {
5254         struct btrfs_fs_info *fs_info = root->fs_info;
5255         u64 bytenr;
5256         u64 generation;
5257         u64 parent;
5258         struct btrfs_key key;
5259         struct btrfs_key first_key;
5260         struct btrfs_ref ref = { 0 };
5261         struct extent_buffer *next;
5262         int level = wc->level;
5263         int reada = 0;
5264         int ret = 0;
5265         bool need_account = false;
5266
5267         generation = btrfs_node_ptr_generation(path->nodes[level],
5268                                                path->slots[level]);
5269         /*
5270          * if the lower level block was created before the snapshot
5271          * was created, we know there is no need to update back refs
5272          * for the subtree
5273          */
5274         if (wc->stage == UPDATE_BACKREF &&
5275             generation <= root->root_key.offset) {
5276                 *lookup_info = 1;
5277                 return 1;
5278         }
5279
5280         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5281         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
5282                               path->slots[level]);
5283
5284         next = find_extent_buffer(fs_info, bytenr);
5285         if (!next) {
5286                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5287                                 root->root_key.objectid, level - 1);
5288                 if (IS_ERR(next))
5289                         return PTR_ERR(next);
5290                 reada = 1;
5291         }
5292         btrfs_tree_lock(next);
5293
5294         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5295                                        &wc->refs[level - 1],
5296                                        &wc->flags[level - 1]);
5297         if (ret < 0)
5298                 goto out_unlock;
5299
5300         if (unlikely(wc->refs[level - 1] == 0)) {
5301                 btrfs_err(fs_info, "Missing references.");
5302                 ret = -EIO;
5303                 goto out_unlock;
5304         }
5305         *lookup_info = 0;
5306
5307         if (wc->stage == DROP_REFERENCE) {
5308                 if (wc->refs[level - 1] > 1) {
5309                         need_account = true;
5310                         if (level == 1 &&
5311                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5312                                 goto skip;
5313
5314                         if (!wc->update_ref ||
5315                             generation <= root->root_key.offset)
5316                                 goto skip;
5317
5318                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5319                                               path->slots[level]);
5320                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5321                         if (ret < 0)
5322                                 goto skip;
5323
5324                         wc->stage = UPDATE_BACKREF;
5325                         wc->shared_level = level - 1;
5326                 }
5327         } else {
5328                 if (level == 1 &&
5329                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5330                         goto skip;
5331         }
5332
5333         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5334                 btrfs_tree_unlock(next);
5335                 free_extent_buffer(next);
5336                 next = NULL;
5337                 *lookup_info = 1;
5338         }
5339
5340         if (!next) {
5341                 if (reada && level == 1)
5342                         reada_walk_down(trans, root, wc, path);
5343                 next = read_tree_block(fs_info, bytenr, root->root_key.objectid,
5344                                        generation, level - 1, &first_key);
5345                 if (IS_ERR(next)) {
5346                         return PTR_ERR(next);
5347                 } else if (!extent_buffer_uptodate(next)) {
5348                         free_extent_buffer(next);
5349                         return -EIO;
5350                 }
5351                 btrfs_tree_lock(next);
5352         }
5353
5354         level--;
5355         ASSERT(level == btrfs_header_level(next));
5356         if (level != btrfs_header_level(next)) {
5357                 btrfs_err(root->fs_info, "mismatched level");
5358                 ret = -EIO;
5359                 goto out_unlock;
5360         }
5361         path->nodes[level] = next;
5362         path->slots[level] = 0;
5363         path->locks[level] = BTRFS_WRITE_LOCK;
5364         wc->level = level;
5365         if (wc->level == 1)
5366                 wc->reada_slot = 0;
5367         return 0;
5368 skip:
5369         wc->refs[level - 1] = 0;
5370         wc->flags[level - 1] = 0;
5371         if (wc->stage == DROP_REFERENCE) {
5372                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5373                         parent = path->nodes[level]->start;
5374                 } else {
5375                         ASSERT(root->root_key.objectid ==
5376                                btrfs_header_owner(path->nodes[level]));
5377                         if (root->root_key.objectid !=
5378                             btrfs_header_owner(path->nodes[level])) {
5379                                 btrfs_err(root->fs_info,
5380                                                 "mismatched block owner");
5381                                 ret = -EIO;
5382                                 goto out_unlock;
5383                         }
5384                         parent = 0;
5385                 }
5386
5387                 /*
5388                  * If we had a drop_progress we need to verify the refs are set
5389                  * as expected.  If we find our ref then we know that from here
5390                  * on out everything should be correct, and we can clear the
5391                  * ->restarted flag.
5392                  */
5393                 if (wc->restarted) {
5394                         ret = check_ref_exists(trans, root, bytenr, parent,
5395                                                level - 1);
5396                         if (ret < 0)
5397                                 goto out_unlock;
5398                         if (ret == 0)
5399                                 goto no_delete;
5400                         ret = 0;
5401                         wc->restarted = 0;
5402                 }
5403
5404                 /*
5405                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5406                  * already accounted them at merge time (replace_path),
5407                  * thus we could skip expensive subtree trace here.
5408                  */
5409                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5410                     need_account) {
5411                         ret = btrfs_qgroup_trace_subtree(trans, next,
5412                                                          generation, level - 1);
5413                         if (ret) {
5414                                 btrfs_err_rl(fs_info,
5415                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5416                                              ret);
5417                         }
5418                 }
5419
5420                 /*
5421                  * We need to update the next key in our walk control so we can
5422                  * update the drop_progress key accordingly.  We don't care if
5423                  * find_next_key doesn't find a key because that means we're at
5424                  * the end and are going to clean up now.
5425                  */
5426                 wc->drop_level = level;
5427                 find_next_key(path, level, &wc->drop_progress);
5428
5429                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5430                                        fs_info->nodesize, parent);
5431                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5432                                     0, false);
5433                 ret = btrfs_free_extent(trans, &ref);
5434                 if (ret)
5435                         goto out_unlock;
5436         }
5437 no_delete:
5438         *lookup_info = 1;
5439         ret = 1;
5440
5441 out_unlock:
5442         btrfs_tree_unlock(next);
5443         free_extent_buffer(next);
5444
5445         return ret;
5446 }
5447
5448 /*
5449  * helper to process tree block while walking up the tree.
5450  *
5451  * when wc->stage == DROP_REFERENCE, this function drops
5452  * reference count on the block.
5453  *
5454  * when wc->stage == UPDATE_BACKREF, this function changes
5455  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5456  * to UPDATE_BACKREF previously while processing the block.
5457  *
5458  * NOTE: return value 1 means we should stop walking up.
5459  */
5460 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5461                                  struct btrfs_root *root,
5462                                  struct btrfs_path *path,
5463                                  struct walk_control *wc)
5464 {
5465         struct btrfs_fs_info *fs_info = root->fs_info;
5466         int ret;
5467         int level = wc->level;
5468         struct extent_buffer *eb = path->nodes[level];
5469         u64 parent = 0;
5470
5471         if (wc->stage == UPDATE_BACKREF) {
5472                 BUG_ON(wc->shared_level < level);
5473                 if (level < wc->shared_level)
5474                         goto out;
5475
5476                 ret = find_next_key(path, level + 1, &wc->update_progress);
5477                 if (ret > 0)
5478                         wc->update_ref = 0;
5479
5480                 wc->stage = DROP_REFERENCE;
5481                 wc->shared_level = -1;
5482                 path->slots[level] = 0;
5483
5484                 /*
5485                  * check reference count again if the block isn't locked.
5486                  * we should start walking down the tree again if reference
5487                  * count is one.
5488                  */
5489                 if (!path->locks[level]) {
5490                         BUG_ON(level == 0);
5491                         btrfs_tree_lock(eb);
5492                         path->locks[level] = BTRFS_WRITE_LOCK;
5493
5494                         ret = btrfs_lookup_extent_info(trans, fs_info,
5495                                                        eb->start, level, 1,
5496                                                        &wc->refs[level],
5497                                                        &wc->flags[level]);
5498                         if (ret < 0) {
5499                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5500                                 path->locks[level] = 0;
5501                                 return ret;
5502                         }
5503                         BUG_ON(wc->refs[level] == 0);
5504                         if (wc->refs[level] == 1) {
5505                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5506                                 path->locks[level] = 0;
5507                                 return 1;
5508                         }
5509                 }
5510         }
5511
5512         /* wc->stage == DROP_REFERENCE */
5513         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5514
5515         if (wc->refs[level] == 1) {
5516                 if (level == 0) {
5517                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5518                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5519                         else
5520                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5521                         BUG_ON(ret); /* -ENOMEM */
5522                         if (is_fstree(root->root_key.objectid)) {
5523                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5524                                 if (ret) {
5525                                         btrfs_err_rl(fs_info,
5526         "error %d accounting leaf items, quota is out of sync, rescan required",
5527                                              ret);
5528                                 }
5529                         }
5530                 }
5531                 /* make block locked assertion in btrfs_clean_tree_block happy */
5532                 if (!path->locks[level] &&
5533                     btrfs_header_generation(eb) == trans->transid) {
5534                         btrfs_tree_lock(eb);
5535                         path->locks[level] = BTRFS_WRITE_LOCK;
5536                 }
5537                 btrfs_clean_tree_block(eb);
5538         }
5539
5540         if (eb == root->node) {
5541                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5542                         parent = eb->start;
5543                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5544                         goto owner_mismatch;
5545         } else {
5546                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5547                         parent = path->nodes[level + 1]->start;
5548                 else if (root->root_key.objectid !=
5549                          btrfs_header_owner(path->nodes[level + 1]))
5550                         goto owner_mismatch;
5551         }
5552
5553         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5554                               wc->refs[level] == 1);
5555 out:
5556         wc->refs[level] = 0;
5557         wc->flags[level] = 0;
5558         return 0;
5559
5560 owner_mismatch:
5561         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5562                      btrfs_header_owner(eb), root->root_key.objectid);
5563         return -EUCLEAN;
5564 }
5565
5566 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5567                                    struct btrfs_root *root,
5568                                    struct btrfs_path *path,
5569                                    struct walk_control *wc)
5570 {
5571         int level = wc->level;
5572         int lookup_info = 1;
5573         int ret;
5574
5575         while (level >= 0) {
5576                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5577                 if (ret > 0)
5578                         break;
5579
5580                 if (level == 0)
5581                         break;
5582
5583                 if (path->slots[level] >=
5584                     btrfs_header_nritems(path->nodes[level]))
5585                         break;
5586
5587                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5588                 if (ret > 0) {
5589                         path->slots[level]++;
5590                         continue;
5591                 } else if (ret < 0)
5592                         return ret;
5593                 level = wc->level;
5594         }
5595         return 0;
5596 }
5597
5598 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5599                                  struct btrfs_root *root,
5600                                  struct btrfs_path *path,
5601                                  struct walk_control *wc, int max_level)
5602 {
5603         int level = wc->level;
5604         int ret;
5605
5606         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5607         while (level < max_level && path->nodes[level]) {
5608                 wc->level = level;
5609                 if (path->slots[level] + 1 <
5610                     btrfs_header_nritems(path->nodes[level])) {
5611                         path->slots[level]++;
5612                         return 0;
5613                 } else {
5614                         ret = walk_up_proc(trans, root, path, wc);
5615                         if (ret > 0)
5616                                 return 0;
5617                         if (ret < 0)
5618                                 return ret;
5619
5620                         if (path->locks[level]) {
5621                                 btrfs_tree_unlock_rw(path->nodes[level],
5622                                                      path->locks[level]);
5623                                 path->locks[level] = 0;
5624                         }
5625                         free_extent_buffer(path->nodes[level]);
5626                         path->nodes[level] = NULL;
5627                         level++;
5628                 }
5629         }
5630         return 1;
5631 }
5632
5633 /*
5634  * drop a subvolume tree.
5635  *
5636  * this function traverses the tree freeing any blocks that only
5637  * referenced by the tree.
5638  *
5639  * when a shared tree block is found. this function decreases its
5640  * reference count by one. if update_ref is true, this function
5641  * also make sure backrefs for the shared block and all lower level
5642  * blocks are properly updated.
5643  *
5644  * If called with for_reloc == 0, may exit early with -EAGAIN
5645  */
5646 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5647 {
5648         const bool is_reloc_root = (root->root_key.objectid ==
5649                                     BTRFS_TREE_RELOC_OBJECTID);
5650         struct btrfs_fs_info *fs_info = root->fs_info;
5651         struct btrfs_path *path;
5652         struct btrfs_trans_handle *trans;
5653         struct btrfs_root *tree_root = fs_info->tree_root;
5654         struct btrfs_root_item *root_item = &root->root_item;
5655         struct walk_control *wc;
5656         struct btrfs_key key;
5657         int err = 0;
5658         int ret;
5659         int level;
5660         bool root_dropped = false;
5661         bool unfinished_drop = false;
5662
5663         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5664
5665         path = btrfs_alloc_path();
5666         if (!path) {
5667                 err = -ENOMEM;
5668                 goto out;
5669         }
5670
5671         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5672         if (!wc) {
5673                 btrfs_free_path(path);
5674                 err = -ENOMEM;
5675                 goto out;
5676         }
5677
5678         /*
5679          * Use join to avoid potential EINTR from transaction start. See
5680          * wait_reserve_ticket and the whole reservation callchain.
5681          */
5682         if (for_reloc)
5683                 trans = btrfs_join_transaction(tree_root);
5684         else
5685                 trans = btrfs_start_transaction(tree_root, 0);
5686         if (IS_ERR(trans)) {
5687                 err = PTR_ERR(trans);
5688                 goto out_free;
5689         }
5690
5691         err = btrfs_run_delayed_items(trans);
5692         if (err)
5693                 goto out_end_trans;
5694
5695         /*
5696          * This will help us catch people modifying the fs tree while we're
5697          * dropping it.  It is unsafe to mess with the fs tree while it's being
5698          * dropped as we unlock the root node and parent nodes as we walk down
5699          * the tree, assuming nothing will change.  If something does change
5700          * then we'll have stale information and drop references to blocks we've
5701          * already dropped.
5702          */
5703         set_bit(BTRFS_ROOT_DELETING, &root->state);
5704         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5705
5706         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5707                 level = btrfs_header_level(root->node);
5708                 path->nodes[level] = btrfs_lock_root_node(root);
5709                 path->slots[level] = 0;
5710                 path->locks[level] = BTRFS_WRITE_LOCK;
5711                 memset(&wc->update_progress, 0,
5712                        sizeof(wc->update_progress));
5713         } else {
5714                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5715                 memcpy(&wc->update_progress, &key,
5716                        sizeof(wc->update_progress));
5717
5718                 level = btrfs_root_drop_level(root_item);
5719                 BUG_ON(level == 0);
5720                 path->lowest_level = level;
5721                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5722                 path->lowest_level = 0;
5723                 if (ret < 0) {
5724                         err = ret;
5725                         goto out_end_trans;
5726                 }
5727                 WARN_ON(ret > 0);
5728
5729                 /*
5730                  * unlock our path, this is safe because only this
5731                  * function is allowed to delete this snapshot
5732                  */
5733                 btrfs_unlock_up_safe(path, 0);
5734
5735                 level = btrfs_header_level(root->node);
5736                 while (1) {
5737                         btrfs_tree_lock(path->nodes[level]);
5738                         path->locks[level] = BTRFS_WRITE_LOCK;
5739
5740                         ret = btrfs_lookup_extent_info(trans, fs_info,
5741                                                 path->nodes[level]->start,
5742                                                 level, 1, &wc->refs[level],
5743                                                 &wc->flags[level]);
5744                         if (ret < 0) {
5745                                 err = ret;
5746                                 goto out_end_trans;
5747                         }
5748                         BUG_ON(wc->refs[level] == 0);
5749
5750                         if (level == btrfs_root_drop_level(root_item))
5751                                 break;
5752
5753                         btrfs_tree_unlock(path->nodes[level]);
5754                         path->locks[level] = 0;
5755                         WARN_ON(wc->refs[level] != 1);
5756                         level--;
5757                 }
5758         }
5759
5760         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5761         wc->level = level;
5762         wc->shared_level = -1;
5763         wc->stage = DROP_REFERENCE;
5764         wc->update_ref = update_ref;
5765         wc->keep_locks = 0;
5766         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5767
5768         while (1) {
5769
5770                 ret = walk_down_tree(trans, root, path, wc);
5771                 if (ret < 0) {
5772                         err = ret;
5773                         break;
5774                 }
5775
5776                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5777                 if (ret < 0) {
5778                         err = ret;
5779                         break;
5780                 }
5781
5782                 if (ret > 0) {
5783                         BUG_ON(wc->stage != DROP_REFERENCE);
5784                         break;
5785                 }
5786
5787                 if (wc->stage == DROP_REFERENCE) {
5788                         wc->drop_level = wc->level;
5789                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5790                                               &wc->drop_progress,
5791                                               path->slots[wc->drop_level]);
5792                 }
5793                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5794                                       &wc->drop_progress);
5795                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5796
5797                 BUG_ON(wc->level == 0);
5798                 if (btrfs_should_end_transaction(trans) ||
5799                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5800                         ret = btrfs_update_root(trans, tree_root,
5801                                                 &root->root_key,
5802                                                 root_item);
5803                         if (ret) {
5804                                 btrfs_abort_transaction(trans, ret);
5805                                 err = ret;
5806                                 goto out_end_trans;
5807                         }
5808
5809                         if (!is_reloc_root)
5810                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5811
5812                         btrfs_end_transaction_throttle(trans);
5813                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5814                                 btrfs_debug(fs_info,
5815                                             "drop snapshot early exit");
5816                                 err = -EAGAIN;
5817                                 goto out_free;
5818                         }
5819
5820                        /*
5821                         * Use join to avoid potential EINTR from transaction
5822                         * start. See wait_reserve_ticket and the whole
5823                         * reservation callchain.
5824                         */
5825                         if (for_reloc)
5826                                 trans = btrfs_join_transaction(tree_root);
5827                         else
5828                                 trans = btrfs_start_transaction(tree_root, 0);
5829                         if (IS_ERR(trans)) {
5830                                 err = PTR_ERR(trans);
5831                                 goto out_free;
5832                         }
5833                 }
5834         }
5835         btrfs_release_path(path);
5836         if (err)
5837                 goto out_end_trans;
5838
5839         ret = btrfs_del_root(trans, &root->root_key);
5840         if (ret) {
5841                 btrfs_abort_transaction(trans, ret);
5842                 err = ret;
5843                 goto out_end_trans;
5844         }
5845
5846         if (!is_reloc_root) {
5847                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5848                                       NULL, NULL);
5849                 if (ret < 0) {
5850                         btrfs_abort_transaction(trans, ret);
5851                         err = ret;
5852                         goto out_end_trans;
5853                 } else if (ret > 0) {
5854                         /* if we fail to delete the orphan item this time
5855                          * around, it'll get picked up the next time.
5856                          *
5857                          * The most common failure here is just -ENOENT.
5858                          */
5859                         btrfs_del_orphan_item(trans, tree_root,
5860                                               root->root_key.objectid);
5861                 }
5862         }
5863
5864         /*
5865          * This subvolume is going to be completely dropped, and won't be
5866          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5867          * commit transaction time.  So free it here manually.
5868          */
5869         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5870         btrfs_qgroup_free_meta_all_pertrans(root);
5871
5872         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5873                 btrfs_add_dropped_root(trans, root);
5874         else
5875                 btrfs_put_root(root);
5876         root_dropped = true;
5877 out_end_trans:
5878         if (!is_reloc_root)
5879                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5880
5881         btrfs_end_transaction_throttle(trans);
5882 out_free:
5883         kfree(wc);
5884         btrfs_free_path(path);
5885 out:
5886         /*
5887          * We were an unfinished drop root, check to see if there are any
5888          * pending, and if not clear and wake up any waiters.
5889          */
5890         if (!err && unfinished_drop)
5891                 btrfs_maybe_wake_unfinished_drop(fs_info);
5892
5893         /*
5894          * So if we need to stop dropping the snapshot for whatever reason we
5895          * need to make sure to add it back to the dead root list so that we
5896          * keep trying to do the work later.  This also cleans up roots if we
5897          * don't have it in the radix (like when we recover after a power fail
5898          * or unmount) so we don't leak memory.
5899          */
5900         if (!for_reloc && !root_dropped)
5901                 btrfs_add_dead_root(root);
5902         return err;
5903 }
5904
5905 /*
5906  * drop subtree rooted at tree block 'node'.
5907  *
5908  * NOTE: this function will unlock and release tree block 'node'
5909  * only used by relocation code
5910  */
5911 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5912                         struct btrfs_root *root,
5913                         struct extent_buffer *node,
5914                         struct extent_buffer *parent)
5915 {
5916         struct btrfs_fs_info *fs_info = root->fs_info;
5917         struct btrfs_path *path;
5918         struct walk_control *wc;
5919         int level;
5920         int parent_level;
5921         int ret = 0;
5922         int wret;
5923
5924         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5925
5926         path = btrfs_alloc_path();
5927         if (!path)
5928                 return -ENOMEM;
5929
5930         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5931         if (!wc) {
5932                 btrfs_free_path(path);
5933                 return -ENOMEM;
5934         }
5935
5936         btrfs_assert_tree_write_locked(parent);
5937         parent_level = btrfs_header_level(parent);
5938         atomic_inc(&parent->refs);
5939         path->nodes[parent_level] = parent;
5940         path->slots[parent_level] = btrfs_header_nritems(parent);
5941
5942         btrfs_assert_tree_write_locked(node);
5943         level = btrfs_header_level(node);
5944         path->nodes[level] = node;
5945         path->slots[level] = 0;
5946         path->locks[level] = BTRFS_WRITE_LOCK;
5947
5948         wc->refs[parent_level] = 1;
5949         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5950         wc->level = level;
5951         wc->shared_level = -1;
5952         wc->stage = DROP_REFERENCE;
5953         wc->update_ref = 0;
5954         wc->keep_locks = 1;
5955         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5956
5957         while (1) {
5958                 wret = walk_down_tree(trans, root, path, wc);
5959                 if (wret < 0) {
5960                         ret = wret;
5961                         break;
5962                 }
5963
5964                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5965                 if (wret < 0)
5966                         ret = wret;
5967                 if (wret != 0)
5968                         break;
5969         }
5970
5971         kfree(wc);
5972         btrfs_free_path(path);
5973         return ret;
5974 }
5975
5976 /*
5977  * helper to account the unused space of all the readonly block group in the
5978  * space_info. takes mirrors into account.
5979  */
5980 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
5981 {
5982         struct btrfs_block_group *block_group;
5983         u64 free_bytes = 0;
5984         int factor;
5985
5986         /* It's df, we don't care if it's racy */
5987         if (list_empty(&sinfo->ro_bgs))
5988                 return 0;
5989
5990         spin_lock(&sinfo->lock);
5991         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
5992                 spin_lock(&block_group->lock);
5993
5994                 if (!block_group->ro) {
5995                         spin_unlock(&block_group->lock);
5996                         continue;
5997                 }
5998
5999                 factor = btrfs_bg_type_to_factor(block_group->flags);
6000                 free_bytes += (block_group->length -
6001                                block_group->used) * factor;
6002
6003                 spin_unlock(&block_group->lock);
6004         }
6005         spin_unlock(&sinfo->lock);
6006
6007         return free_bytes;
6008 }
6009
6010 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
6011                                    u64 start, u64 end)
6012 {
6013         return unpin_extent_range(fs_info, start, end, false);
6014 }
6015
6016 /*
6017  * It used to be that old block groups would be left around forever.
6018  * Iterating over them would be enough to trim unused space.  Since we
6019  * now automatically remove them, we also need to iterate over unallocated
6020  * space.
6021  *
6022  * We don't want a transaction for this since the discard may take a
6023  * substantial amount of time.  We don't require that a transaction be
6024  * running, but we do need to take a running transaction into account
6025  * to ensure that we're not discarding chunks that were released or
6026  * allocated in the current transaction.
6027  *
6028  * Holding the chunks lock will prevent other threads from allocating
6029  * or releasing chunks, but it won't prevent a running transaction
6030  * from committing and releasing the memory that the pending chunks
6031  * list head uses.  For that, we need to take a reference to the
6032  * transaction and hold the commit root sem.  We only need to hold
6033  * it while performing the free space search since we have already
6034  * held back allocations.
6035  */
6036 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6037 {
6038         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6039         int ret;
6040
6041         *trimmed = 0;
6042
6043         /* Discard not supported = nothing to do. */
6044         if (!bdev_max_discard_sectors(device->bdev))
6045                 return 0;
6046
6047         /* Not writable = nothing to do. */
6048         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6049                 return 0;
6050
6051         /* No free space = nothing to do. */
6052         if (device->total_bytes <= device->bytes_used)
6053                 return 0;
6054
6055         ret = 0;
6056
6057         while (1) {
6058                 struct btrfs_fs_info *fs_info = device->fs_info;
6059                 u64 bytes;
6060
6061                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6062                 if (ret)
6063                         break;
6064
6065                 find_first_clear_extent_bit(&device->alloc_state, start,
6066                                             &start, &end,
6067                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6068
6069                 /* Check if there are any CHUNK_* bits left */
6070                 if (start > device->total_bytes) {
6071                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6072                         btrfs_warn_in_rcu(fs_info,
6073 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6074                                           start, end - start + 1,
6075                                           rcu_str_deref(device->name),
6076                                           device->total_bytes);
6077                         mutex_unlock(&fs_info->chunk_mutex);
6078                         ret = 0;
6079                         break;
6080                 }
6081
6082                 /* Ensure we skip the reserved space on each device. */
6083                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6084
6085                 /*
6086                  * If find_first_clear_extent_bit find a range that spans the
6087                  * end of the device it will set end to -1, in this case it's up
6088                  * to the caller to trim the value to the size of the device.
6089                  */
6090                 end = min(end, device->total_bytes - 1);
6091
6092                 len = end - start + 1;
6093
6094                 /* We didn't find any extents */
6095                 if (!len) {
6096                         mutex_unlock(&fs_info->chunk_mutex);
6097                         ret = 0;
6098                         break;
6099                 }
6100
6101                 ret = btrfs_issue_discard(device->bdev, start, len,
6102                                           &bytes);
6103                 if (!ret)
6104                         set_extent_bits(&device->alloc_state, start,
6105                                         start + bytes - 1,
6106                                         CHUNK_TRIMMED);
6107                 mutex_unlock(&fs_info->chunk_mutex);
6108
6109                 if (ret)
6110                         break;
6111
6112                 start += len;
6113                 *trimmed += bytes;
6114
6115                 if (fatal_signal_pending(current)) {
6116                         ret = -ERESTARTSYS;
6117                         break;
6118                 }
6119
6120                 cond_resched();
6121         }
6122
6123         return ret;
6124 }
6125
6126 /*
6127  * Trim the whole filesystem by:
6128  * 1) trimming the free space in each block group
6129  * 2) trimming the unallocated space on each device
6130  *
6131  * This will also continue trimming even if a block group or device encounters
6132  * an error.  The return value will be the last error, or 0 if nothing bad
6133  * happens.
6134  */
6135 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6136 {
6137         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6138         struct btrfs_block_group *cache = NULL;
6139         struct btrfs_device *device;
6140         u64 group_trimmed;
6141         u64 range_end = U64_MAX;
6142         u64 start;
6143         u64 end;
6144         u64 trimmed = 0;
6145         u64 bg_failed = 0;
6146         u64 dev_failed = 0;
6147         int bg_ret = 0;
6148         int dev_ret = 0;
6149         int ret = 0;
6150
6151         if (range->start == U64_MAX)
6152                 return -EINVAL;
6153
6154         /*
6155          * Check range overflow if range->len is set.
6156          * The default range->len is U64_MAX.
6157          */
6158         if (range->len != U64_MAX &&
6159             check_add_overflow(range->start, range->len, &range_end))
6160                 return -EINVAL;
6161
6162         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6163         for (; cache; cache = btrfs_next_block_group(cache)) {
6164                 if (cache->start >= range_end) {
6165                         btrfs_put_block_group(cache);
6166                         break;
6167                 }
6168
6169                 start = max(range->start, cache->start);
6170                 end = min(range_end, cache->start + cache->length);
6171
6172                 if (end - start >= range->minlen) {
6173                         if (!btrfs_block_group_done(cache)) {
6174                                 ret = btrfs_cache_block_group(cache, true);
6175                                 if (ret) {
6176                                         bg_failed++;
6177                                         bg_ret = ret;
6178                                         continue;
6179                                 }
6180                         }
6181                         ret = btrfs_trim_block_group(cache,
6182                                                      &group_trimmed,
6183                                                      start,
6184                                                      end,
6185                                                      range->minlen);
6186
6187                         trimmed += group_trimmed;
6188                         if (ret) {
6189                                 bg_failed++;
6190                                 bg_ret = ret;
6191                                 continue;
6192                         }
6193                 }
6194         }
6195
6196         if (bg_failed)
6197                 btrfs_warn(fs_info,
6198                         "failed to trim %llu block group(s), last error %d",
6199                         bg_failed, bg_ret);
6200
6201         mutex_lock(&fs_devices->device_list_mutex);
6202         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6203                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6204                         continue;
6205
6206                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6207                 if (ret) {
6208                         dev_failed++;
6209                         dev_ret = ret;
6210                         break;
6211                 }
6212
6213                 trimmed += group_trimmed;
6214         }
6215         mutex_unlock(&fs_devices->device_list_mutex);
6216
6217         if (dev_failed)
6218                 btrfs_warn(fs_info,
6219                         "failed to trim %llu device(s), last error %d",
6220                         dev_failed, dev_ret);
6221         range->len = trimmed;
6222         if (bg_ret)
6223                 return bg_ret;
6224         return dev_ret;
6225 }