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         int ret;
  91         struct btrfs_key key;
  92         struct btrfs_path *path;
  93
  94         path = btrfs_alloc_path();
  95         if (!path)
  96                 return -ENOMEM;
  97
  98         key.objectid = start;
  99         key.offset = len;
 100         key.type = BTRFS_EXTENT_ITEM_KEY;
 101         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 102         btrfs_free_path(path);
 103         return ret;
 104 }
 105
 106 /*
 107  * helper function to lookup reference count and flags of a tree block.
 108  *
 109  * the head node for delayed ref is used to store the sum of all the
 110  * reference count modifications queued up in the rbtree. the head
 111  * node may also store the extent flags to set. This way you can check
 112  * to see what the reference count and extent flags would be if all of
 113  * the delayed refs are not processed.
 114  */
 115 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 116                              struct btrfs_fs_info *fs_info, u64 bytenr,
 117                              u64 offset, int metadata, u64 *refs, u64 *flags)
 118 {
 119         struct btrfs_delayed_ref_head *head;
 120         struct btrfs_delayed_ref_root *delayed_refs;
 121         struct btrfs_path *path;
 122         struct btrfs_extent_item *ei;
 123         struct extent_buffer *leaf;
 124         struct btrfs_key key;
 125         u32 item_size;
 126         u64 num_refs;
 127         u64 extent_flags;
 128         int ret;
 129
 130         /*
 131          * If we don't have skinny metadata, don't bother doing anything
 132          * different
 133          */
 134         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 135                 offset = fs_info->nodesize;
 136                 metadata = 0;
 137         }
 138
 139         path = btrfs_alloc_path();
 140         if (!path)
 141                 return -ENOMEM;
 142
 143         if (!trans) {
 144                 path->skip_locking = 1;
 145                 path->search_commit_root = 1;
 146         }
 147
 148 search_again:
 149         key.objectid = bytenr;
 150         key.offset = offset;
 151         if (metadata)
 152                 key.type = BTRFS_METADATA_ITEM_KEY;
 153         else
 154                 key.type = BTRFS_EXTENT_ITEM_KEY;
 155
 156         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 157         if (ret < 0)
 158                 goto out_free;
 159
 160         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 161                 if (path->slots[0]) {
 162                         path->slots[0]--;
 163                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 164                                               path->slots[0]);
 165                         if (key.objectid == bytenr &&
 166                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 167                             key.offset == fs_info->nodesize)
 168                                 ret = 0;
 169                 }
 170         }
 171
 172         if (ret == 0) {
 173                 leaf = path->nodes[0];
 174                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 175                 if (item_size >= sizeof(*ei)) {
 176                         ei = btrfs_item_ptr(leaf, path->slots[0],
 177                                             struct btrfs_extent_item);
 178                         num_refs = btrfs_extent_refs(leaf, ei);
 179                         extent_flags = btrfs_extent_flags(leaf, ei);
 180                 } else {
 181                         ret = -EINVAL;
 182                         btrfs_print_v0_err(fs_info);
 183                         if (trans)
 184                                 btrfs_abort_transaction(trans, ret);
 185                         else
 186                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 187
 188                         goto out_free;
 189                 }
 190
 191                 BUG_ON(num_refs == 0);
 192         } else {
 193                 num_refs = 0;
 194                 extent_flags = 0;
 195                 ret = 0;
 196         }
 197
 198         if (!trans)
 199                 goto out;
 200
 201         delayed_refs = &trans->transaction->delayed_refs;
 202         spin_lock(&delayed_refs->lock);
 203         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 204         if (head) {
 205                 if (!mutex_trylock(&head->mutex)) {
 206                         refcount_inc(&head->refs);
 207                         spin_unlock(&delayed_refs->lock);
 208
 209                         btrfs_release_path(path);
 210
 211                         /*
 212                          * Mutex was contended, block until it's released and try
 213                          * again
 214                          */
 215                         mutex_lock(&head->mutex);
 216                         mutex_unlock(&head->mutex);
 217                         btrfs_put_delayed_ref_head(head);
 218                         goto search_again;
 219                 }
 220                 spin_lock(&head->lock);
 221                 if (head->extent_op && head->extent_op->update_flags)
 222                         extent_flags |= head->extent_op->flags_to_set;
 223                 else
 224                         BUG_ON(num_refs == 0);
 225
 226                 num_refs += head->ref_mod;
 227                 spin_unlock(&head->lock);
 228                 mutex_unlock(&head->mutex);
 229         }
 230         spin_unlock(&delayed_refs->lock);
 231 out:
 232         WARN_ON(num_refs == 0);
 233         if (refs)
 234                 *refs = num_refs;
 235         if (flags)
 236                 *flags = extent_flags;
 237 out_free:
 238         btrfs_free_path(path);
 239         return ret;
 240 }
 241
 242 /*
 243  * Back reference rules.  Back refs have three main goals:
 244  *
 245  * 1) differentiate between all holders of references to an extent so that
 246  *    when a reference is dropped we can make sure it was a valid reference
 247  *    before freeing the extent.
 248  *
 249  * 2) Provide enough information to quickly find the holders of an extent
 250  *    if we notice a given block is corrupted or bad.
 251  *
 252  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 253  *    maintenance.  This is actually the same as #2, but with a slightly
 254  *    different use case.
 255  *
 256  * There are two kinds of back refs. The implicit back refs is optimized
 257  * for pointers in non-shared tree blocks. For a given pointer in a block,
 258  * back refs of this kind provide information about the block's owner tree
 259  * and the pointer's key. These information allow us to find the block by
 260  * b-tree searching. The full back refs is for pointers in tree blocks not
 261  * referenced by their owner trees. The location of tree block is recorded
 262  * in the back refs. Actually the full back refs is generic, and can be
 263  * used in all cases the implicit back refs is used. The major shortcoming
 264  * of the full back refs is its overhead. Every time a tree block gets
 265  * COWed, we have to update back refs entry for all pointers in it.
 266  *
 267  * For a newly allocated tree block, we use implicit back refs for
 268  * pointers in it. This means most tree related operations only involve
 269  * implicit back refs. For a tree block created in old transaction, the
 270  * only way to drop a reference to it is COW it. So we can detect the
 271  * event that tree block loses its owner tree's reference and do the
 272  * back refs conversion.
 273  *
 274  * When a tree block is COWed through a tree, there are four cases:
 275  *
 276  * The reference count of the block is one and the tree is the block's
 277  * owner tree. Nothing to do in this case.
 278  *
 279  * The reference count of the block is one and the tree is not the
 280  * block's owner tree. In this case, full back refs is used for pointers
 281  * in the block. Remove these full back refs, add implicit back refs for
 282  * every pointers in the new block.
 283  *
 284  * The reference count of the block is greater than one and the tree is
 285  * the block's owner tree. In this case, implicit back refs is used for
 286  * pointers in the block. Add full back refs for every pointers in the
 287  * block, increase lower level extents' reference counts. The original
 288  * implicit back refs are entailed to the new block.
 289  *
 290  * The reference count of the block is greater than one and the tree is
 291  * not the block's owner tree. Add implicit back refs for every pointer in
 292  * the new block, increase lower level extents' reference count.
 293  *
 294  * Back Reference Key composing:
 295  *
 296  * The key objectid corresponds to the first byte in the extent,
 297  * The key type is used to differentiate between types of back refs.
 298  * There are different meanings of the key offset for different types
 299  * of back refs.
 300  *
 301  * File extents can be referenced by:
 302  *
 303  * - multiple snapshots, subvolumes, or different generations in one subvol
 304  * - different files inside a single subvolume
 305  * - different offsets inside a file (bookend extents in file.c)
 306  *
 307  * The extent ref structure for the implicit back refs has fields for:
 308  *
 309  * - Objectid of the subvolume root
 310  * - objectid of the file holding the reference
 311  * - original offset in the file
 312  * - how many bookend extents
 313  *
 314  * The key offset for the implicit back refs is hash of the first
 315  * three fields.
 316  *
 317  * The extent ref structure for the full back refs has field for:
 318  *
 319  * - number of pointers in the tree leaf
 320  *
 321  * The key offset for the implicit back refs is the first byte of
 322  * the tree leaf
 323  *
 324  * When a file extent is allocated, The implicit back refs is used.
 325  * the fields are filled in:
 326  *
 327  *     (root_key.objectid, inode objectid, offset in file, 1)
 328  *
 329  * When a file extent is removed file truncation, we find the
 330  * corresponding implicit back refs and check the following fields:
 331  *
 332  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 333  *
 334  * Btree extents can be referenced by:
 335  *
 336  * - Different subvolumes
 337  *
 338  * Both the implicit back refs and the full back refs for tree blocks
 339  * only consist of key. The key offset for the implicit back refs is
 340  * objectid of block's owner tree. The key offset for the full back refs
 341  * is the first byte of parent block.
 342  *
 343  * When implicit back refs is used, information about the lowest key and
 344  * level of the tree block are required. These information are stored in
 345  * tree block info structure.
 346  */
 347
 348 /*
 349  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 350  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 351  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 352  */
 353 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 354                                      struct btrfs_extent_inline_ref *iref,
 355                                      enum btrfs_inline_ref_type is_data)
 356 {
 357         int type = btrfs_extent_inline_ref_type(eb, iref);
 358         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 359
 360         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 361             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 362             type == BTRFS_SHARED_DATA_REF_KEY ||
 363             type == BTRFS_EXTENT_DATA_REF_KEY) {
 364                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 365                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 366                                 return type;
 367                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 368                                 ASSERT(eb->fs_info);
 369                                 /*
 370                                  * Every shared one has parent tree block,
 371                                  * which must be aligned to sector size.
 372                                  */
 373                                 if (offset &&
 374                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 375                                         return type;
 376                         }
 377                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 378                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 379                                 return type;
 380                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 381                                 ASSERT(eb->fs_info);
 382                                 /*
 383                                  * Every shared one has parent tree block,
 384                                  * which must be aligned to sector size.
 385                                  */
 386                                 if (offset &&
 387                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 388                                         return type;
 389                         }
 390                 } else {
 391                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 392                         return type;
 393                 }
 394         }
 395
 396         btrfs_print_leaf((struct extent_buffer *)eb);
 397         btrfs_err(eb->fs_info,
 398                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 399                   eb->start, (unsigned long)iref, type);
 400         WARN_ON(1);
 401
 402         return BTRFS_REF_TYPE_INVALID;
 403 }
 404
 405 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 406 {
 407         u32 high_crc = ~(u32)0;
 408         u32 low_crc = ~(u32)0;
 409         __le64 lenum;
 410
 411         lenum = cpu_to_le64(root_objectid);
 412         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 413         lenum = cpu_to_le64(owner);
 414         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 415         lenum = cpu_to_le64(offset);
 416         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 417
 418         return ((u64)high_crc << 31) ^ (u64)low_crc;
 419 }
 420
 421 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 422                                      struct btrfs_extent_data_ref *ref)
 423 {
 424         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 425                                     btrfs_extent_data_ref_objectid(leaf, ref),
 426                                     btrfs_extent_data_ref_offset(leaf, ref));
 427 }
 428
 429 static int match_extent_data_ref(struct extent_buffer *leaf,
 430                                  struct btrfs_extent_data_ref *ref,
 431                                  u64 root_objectid, u64 owner, u64 offset)
 432 {
 433         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 434             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 435             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 436                 return 0;
 437         return 1;
 438 }
 439
 440 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 441                                            struct btrfs_path *path,
 442                                            u64 bytenr, u64 parent,
 443                                            u64 root_objectid,
 444                                            u64 owner, u64 offset)
 445 {
 446         struct btrfs_root *root = trans->fs_info->extent_root;
 447         struct btrfs_key key;
 448         struct btrfs_extent_data_ref *ref;
 449         struct extent_buffer *leaf;
 450         u32 nritems;
 451         int ret;
 452         int recow;
 453         int err = -ENOENT;
 454
 455         key.objectid = bytenr;
 456         if (parent) {
 457                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 458                 key.offset = parent;
 459         } else {
 460                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 461                 key.offset = hash_extent_data_ref(root_objectid,
 462                                                   owner, offset);
 463         }
 464 again:
 465         recow = 0;
 466         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 467         if (ret < 0) {
 468                 err = ret;
 469                 goto fail;
 470         }
 471
 472         if (parent) {
 473                 if (!ret)
 474                         return 0;
 475                 goto fail;
 476         }
 477
 478         leaf = path->nodes[0];
 479         nritems = btrfs_header_nritems(leaf);
 480         while (1) {
 481                 if (path->slots[0] >= nritems) {
 482                         ret = btrfs_next_leaf(root, path);
 483                         if (ret < 0)
 484                                 err = ret;
 485                         if (ret)
 486                                 goto fail;
 487
 488                         leaf = path->nodes[0];
 489                         nritems = btrfs_header_nritems(leaf);
 490                         recow = 1;
 491                 }
 492
 493                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 494                 if (key.objectid != bytenr ||
 495                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 496                         goto fail;
 497
 498                 ref = btrfs_item_ptr(leaf, path->slots[0],
 499                                      struct btrfs_extent_data_ref);
 500
 501                 if (match_extent_data_ref(leaf, ref, root_objectid,
 502                                           owner, offset)) {
 503                         if (recow) {
 504                                 btrfs_release_path(path);
 505                                 goto again;
 506                         }
 507                         err = 0;
 508                         break;
 509                 }
 510                 path->slots[0]++;
 511         }
 512 fail:
 513         return err;
 514 }
 515
 516 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 517                                            struct btrfs_path *path,
 518                                            u64 bytenr, u64 parent,
 519                                            u64 root_objectid, u64 owner,
 520                                            u64 offset, int refs_to_add)
 521 {
 522         struct btrfs_root *root = trans->fs_info->extent_root;
 523         struct btrfs_key key;
 524         struct extent_buffer *leaf;
 525         u32 size;
 526         u32 num_refs;
 527         int ret;
 528
 529         key.objectid = bytenr;
 530         if (parent) {
 531                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 532                 key.offset = parent;
 533                 size = sizeof(struct btrfs_shared_data_ref);
 534         } else {
 535                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 536                 key.offset = hash_extent_data_ref(root_objectid,
 537                                                   owner, offset);
 538                 size = sizeof(struct btrfs_extent_data_ref);
 539         }
 540
 541         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 542         if (ret && ret != -EEXIST)
 543                 goto fail;
 544
 545         leaf = path->nodes[0];
 546         if (parent) {
 547                 struct btrfs_shared_data_ref *ref;
 548                 ref = btrfs_item_ptr(leaf, path->slots[0],
 549                                      struct btrfs_shared_data_ref);
 550                 if (ret == 0) {
 551                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 552                 } else {
 553                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 554                         num_refs += refs_to_add;
 555                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 556                 }
 557         } else {
 558                 struct btrfs_extent_data_ref *ref;
 559                 while (ret == -EEXIST) {
 560                         ref = btrfs_item_ptr(leaf, path->slots[0],
 561                                              struct btrfs_extent_data_ref);
 562                         if (match_extent_data_ref(leaf, ref, root_objectid,
 563                                                   owner, offset))
 564                                 break;
 565                         btrfs_release_path(path);
 566                         key.offset++;
 567                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 568                                                       size);
 569                         if (ret && ret != -EEXIST)
 570                                 goto fail;
 571
 572                         leaf = path->nodes[0];
 573                 }
 574                 ref = btrfs_item_ptr(leaf, path->slots[0],
 575                                      struct btrfs_extent_data_ref);
 576                 if (ret == 0) {
 577                         btrfs_set_extent_data_ref_root(leaf, ref,
 578                                                        root_objectid);
 579                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 580                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 581                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 582                 } else {
 583                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 584                         num_refs += refs_to_add;
 585                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 586                 }
 587         }
 588         btrfs_mark_buffer_dirty(leaf);
 589         ret = 0;
 590 fail:
 591         btrfs_release_path(path);
 592         return ret;
 593 }
 594
 595 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 596                                            struct btrfs_path *path,
 597                                            int refs_to_drop, int *last_ref)
 598 {
 599         struct btrfs_key key;
 600         struct btrfs_extent_data_ref *ref1 = NULL;
 601         struct btrfs_shared_data_ref *ref2 = NULL;
 602         struct extent_buffer *leaf;
 603         u32 num_refs = 0;
 604         int ret = 0;
 605
 606         leaf = path->nodes[0];
 607         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 608
 609         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 610                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 611                                       struct btrfs_extent_data_ref);
 612                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 613         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 614                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 615                                       struct btrfs_shared_data_ref);
 616                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 617         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 618                 btrfs_print_v0_err(trans->fs_info);
 619                 btrfs_abort_transaction(trans, -EINVAL);
 620                 return -EINVAL;
 621         } else {
 622                 BUG();
 623         }
 624
 625         BUG_ON(num_refs < refs_to_drop);
 626         num_refs -= refs_to_drop;
 627
 628         if (num_refs == 0) {
 629                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
 630                 *last_ref = 1;
 631         } else {
 632                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 633                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 634                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 635                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 636                 btrfs_mark_buffer_dirty(leaf);
 637         }
 638         return ret;
 639 }
 640
 641 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 642                                           struct btrfs_extent_inline_ref *iref)
 643 {
 644         struct btrfs_key key;
 645         struct extent_buffer *leaf;
 646         struct btrfs_extent_data_ref *ref1;
 647         struct btrfs_shared_data_ref *ref2;
 648         u32 num_refs = 0;
 649         int type;
 650
 651         leaf = path->nodes[0];
 652         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 653
 654         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 655         if (iref) {
 656                 /*
 657                  * If type is invalid, we should have bailed out earlier than
 658                  * this call.
 659                  */
 660                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 661                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 662                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 663                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 664                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 665                 } else {
 666                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 667                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 668                 }
 669         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 670                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 671                                       struct btrfs_extent_data_ref);
 672                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 673         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 674                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 675                                       struct btrfs_shared_data_ref);
 676                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 677         } else {
 678                 WARN_ON(1);
 679         }
 680         return num_refs;
 681 }
 682
 683 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 684                                           struct btrfs_path *path,
 685                                           u64 bytenr, u64 parent,
 686                                           u64 root_objectid)
 687 {
 688         struct btrfs_root *root = trans->fs_info->extent_root;
 689         struct btrfs_key key;
 690         int ret;
 691
 692         key.objectid = bytenr;
 693         if (parent) {
 694                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 695                 key.offset = parent;
 696         } else {
 697                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 698                 key.offset = root_objectid;
 699         }
 700
 701         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 702         if (ret > 0)
 703                 ret = -ENOENT;
 704         return ret;
 705 }
 706
 707 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 708                                           struct btrfs_path *path,
 709                                           u64 bytenr, u64 parent,
 710                                           u64 root_objectid)
 711 {
 712         struct btrfs_key key;
 713         int ret;
 714
 715         key.objectid = bytenr;
 716         if (parent) {
 717                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 718                 key.offset = parent;
 719         } else {
 720                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 721                 key.offset = root_objectid;
 722         }
 723
 724         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
 725                                       path, &key, 0);
 726         btrfs_release_path(path);
 727         return ret;
 728 }
 729
 730 static inline int extent_ref_type(u64 parent, u64 owner)
 731 {
 732         int type;
 733         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 734                 if (parent > 0)
 735                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 736                 else
 737                         type = BTRFS_TREE_BLOCK_REF_KEY;
 738         } else {
 739                 if (parent > 0)
 740                         type = BTRFS_SHARED_DATA_REF_KEY;
 741                 else
 742                         type = BTRFS_EXTENT_DATA_REF_KEY;
 743         }
 744         return type;
 745 }
 746
 747 static int find_next_key(struct btrfs_path *path, int level,
 748                          struct btrfs_key *key)
 749
 750 {
 751         for (; level < BTRFS_MAX_LEVEL; level++) {
 752                 if (!path->nodes[level])
 753                         break;
 754                 if (path->slots[level] + 1 >=
 755                     btrfs_header_nritems(path->nodes[level]))
 756                         continue;
 757                 if (level == 0)
 758                         btrfs_item_key_to_cpu(path->nodes[level], key,
 759                                               path->slots[level] + 1);
 760                 else
 761                         btrfs_node_key_to_cpu(path->nodes[level], key,
 762                                               path->slots[level] + 1);
 763                 return 0;
 764         }
 765         return 1;
 766 }
 767
 768 /*
 769  * look for inline back ref. if back ref is found, *ref_ret is set
 770  * to the address of inline back ref, and 0 is returned.
 771  *
 772  * if back ref isn't found, *ref_ret is set to the address where it
 773  * should be inserted, and -ENOENT is returned.
 774  *
 775  * if insert is true and there are too many inline back refs, the path
 776  * points to the extent item, and -EAGAIN is returned.
 777  *
 778  * NOTE: inline back refs are ordered in the same way that back ref
 779  *       items in the tree are ordered.
 780  */
 781 static noinline_for_stack
 782 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 783                                  struct btrfs_path *path,
 784                                  struct btrfs_extent_inline_ref **ref_ret,
 785                                  u64 bytenr, u64 num_bytes,
 786                                  u64 parent, u64 root_objectid,
 787                                  u64 owner, u64 offset, int insert)
 788 {
 789         struct btrfs_fs_info *fs_info = trans->fs_info;
 790         struct btrfs_root *root = fs_info->extent_root;
 791         struct btrfs_key key;
 792         struct extent_buffer *leaf;
 793         struct btrfs_extent_item *ei;
 794         struct btrfs_extent_inline_ref *iref;
 795         u64 flags;
 796         u64 item_size;
 797         unsigned long ptr;
 798         unsigned long end;
 799         int extra_size;
 800         int type;
 801         int want;
 802         int ret;
 803         int err = 0;
 804         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 805         int needed;
 806
 807         key.objectid = bytenr;
 808         key.type = BTRFS_EXTENT_ITEM_KEY;
 809         key.offset = num_bytes;
 810
 811         want = extent_ref_type(parent, owner);
 812         if (insert) {
 813                 extra_size = btrfs_extent_inline_ref_size(want);
 814                 path->search_for_extension = 1;
 815                 path->keep_locks = 1;
 816         } else
 817                 extra_size = -1;
 818
 819         /*
 820          * Owner is our level, so we can just add one to get the level for the
 821          * block we are interested in.
 822          */
 823         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 824                 key.type = BTRFS_METADATA_ITEM_KEY;
 825                 key.offset = owner;
 826         }
 827
 828 again:
 829         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 830         if (ret < 0) {
 831                 err = ret;
 832                 goto out;
 833         }
 834
 835         /*
 836          * We may be a newly converted file system which still has the old fat
 837          * extent entries for metadata, so try and see if we have one of those.
 838          */
 839         if (ret > 0 && skinny_metadata) {
 840                 skinny_metadata = false;
 841                 if (path->slots[0]) {
 842                         path->slots[0]--;
 843                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 844                                               path->slots[0]);
 845                         if (key.objectid == bytenr &&
 846                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 847                             key.offset == num_bytes)
 848                                 ret = 0;
 849                 }
 850                 if (ret) {
 851                         key.objectid = bytenr;
 852                         key.type = BTRFS_EXTENT_ITEM_KEY;
 853                         key.offset = num_bytes;
 854                         btrfs_release_path(path);
 855                         goto again;
 856                 }
 857         }
 858
 859         if (ret && !insert) {
 860                 err = -ENOENT;
 861                 goto out;
 862         } else if (WARN_ON(ret)) {
 863                 err = -EIO;
 864                 goto out;
 865         }
 866
 867         leaf = path->nodes[0];
 868         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 869         if (unlikely(item_size < sizeof(*ei))) {
 870                 err = -EINVAL;
 871                 btrfs_print_v0_err(fs_info);
 872                 btrfs_abort_transaction(trans, err);
 873                 goto out;
 874         }
 875
 876         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 877         flags = btrfs_extent_flags(leaf, ei);
 878
 879         ptr = (unsigned long)(ei + 1);
 880         end = (unsigned long)ei + item_size;
 881
 882         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 883                 ptr += sizeof(struct btrfs_tree_block_info);
 884                 BUG_ON(ptr > end);
 885         }
 886
 887         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 888                 needed = BTRFS_REF_TYPE_DATA;
 889         else
 890                 needed = BTRFS_REF_TYPE_BLOCK;
 891
 892         err = -ENOENT;
 893         while (1) {
 894                 if (ptr >= end) {
 895                         WARN_ON(ptr > end);
 896                         break;
 897                 }
 898                 iref = (struct btrfs_extent_inline_ref *)ptr;
 899                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 900                 if (type == BTRFS_REF_TYPE_INVALID) {
 901                         err = -EUCLEAN;
 902                         goto out;
 903                 }
 904
 905                 if (want < type)
 906                         break;
 907                 if (want > type) {
 908                         ptr += btrfs_extent_inline_ref_size(type);
 909                         continue;
 910                 }
 911
 912                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 913                         struct btrfs_extent_data_ref *dref;
 914                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 915                         if (match_extent_data_ref(leaf, dref, root_objectid,
 916                                                   owner, offset)) {
 917                                 err = 0;
 918                                 break;
 919                         }
 920                         if (hash_extent_data_ref_item(leaf, dref) <
 921                             hash_extent_data_ref(root_objectid, owner, offset))
 922                                 break;
 923                 } else {
 924                         u64 ref_offset;
 925                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 926                         if (parent > 0) {
 927                                 if (parent == ref_offset) {
 928                                         err = 0;
 929                                         break;
 930                                 }
 931                                 if (ref_offset < parent)
 932                                         break;
 933                         } else {
 934                                 if (root_objectid == ref_offset) {
 935                                         err = 0;
 936                                         break;
 937                                 }
 938                                 if (ref_offset < root_objectid)
 939                                         break;
 940                         }
 941                 }
 942                 ptr += btrfs_extent_inline_ref_size(type);
 943         }
 944         if (err == -ENOENT && insert) {
 945                 if (item_size + extra_size >=
 946                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 947                         err = -EAGAIN;
 948                         goto out;
 949                 }
 950                 /*
 951                  * To add new inline back ref, we have to make sure
 952                  * there is no corresponding back ref item.
 953                  * For simplicity, we just do not add new inline back
 954                  * ref if there is any kind of item for this block
 955                  */
 956                 if (find_next_key(path, 0, &key) == 0 &&
 957                     key.objectid == bytenr &&
 958                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 959                         err = -EAGAIN;
 960                         goto out;
 961                 }
 962         }
 963         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 964 out:
 965         if (insert) {
 966                 path->keep_locks = 0;
 967                 path->search_for_extension = 0;
 968                 btrfs_unlock_up_safe(path, 1);
 969         }
 970         return err;
 971 }
 972
 973 /*
 974  * helper to add new inline back ref
 975  */
 976 static noinline_for_stack
 977 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
 978                                  struct btrfs_path *path,
 979                                  struct btrfs_extent_inline_ref *iref,
 980                                  u64 parent, u64 root_objectid,
 981                                  u64 owner, u64 offset, int refs_to_add,
 982                                  struct btrfs_delayed_extent_op *extent_op)
 983 {
 984         struct extent_buffer *leaf;
 985         struct btrfs_extent_item *ei;
 986         unsigned long ptr;
 987         unsigned long end;
 988         unsigned long item_offset;
 989         u64 refs;
 990         int size;
 991         int type;
 992
 993         leaf = path->nodes[0];
 994         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 995         item_offset = (unsigned long)iref - (unsigned long)ei;
 996
 997         type = extent_ref_type(parent, owner);
 998         size = btrfs_extent_inline_ref_size(type);
 999
1000         btrfs_extend_item(path, size);
1001
1002         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1003         refs = btrfs_extent_refs(leaf, ei);
1004         refs += refs_to_add;
1005         btrfs_set_extent_refs(leaf, ei, refs);
1006         if (extent_op)
1007                 __run_delayed_extent_op(extent_op, leaf, ei);
1008
1009         ptr = (unsigned long)ei + item_offset;
1010         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1011         if (ptr < end - size)
1012                 memmove_extent_buffer(leaf, ptr + size, ptr,
1013                                       end - size - ptr);
1014
1015         iref = (struct btrfs_extent_inline_ref *)ptr;
1016         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1017         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1018                 struct btrfs_extent_data_ref *dref;
1019                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1020                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1021                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1022                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1023                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1024         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1025                 struct btrfs_shared_data_ref *sref;
1026                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1027                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1028                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1029         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1030                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1031         } else {
1032                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1033         }
1034         btrfs_mark_buffer_dirty(leaf);
1035 }
1036
1037 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1038                                  struct btrfs_path *path,
1039                                  struct btrfs_extent_inline_ref **ref_ret,
1040                                  u64 bytenr, u64 num_bytes, u64 parent,
1041                                  u64 root_objectid, u64 owner, u64 offset)
1042 {
1043         int ret;
1044
1045         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1046                                            num_bytes, parent, root_objectid,
1047                                            owner, offset, 0);
1048         if (ret != -ENOENT)
1049                 return ret;
1050
1051         btrfs_release_path(path);
1052         *ref_ret = NULL;
1053
1054         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1055                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1056                                             root_objectid);
1057         } else {
1058                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1059                                              root_objectid, owner, offset);
1060         }
1061         return ret;
1062 }
1063
1064 /*
1065  * helper to update/remove inline back ref
1066  */
1067 static noinline_for_stack
1068 void update_inline_extent_backref(struct btrfs_path *path,
1069                                   struct btrfs_extent_inline_ref *iref,
1070                                   int refs_to_mod,
1071                                   struct btrfs_delayed_extent_op *extent_op,
1072                                   int *last_ref)
1073 {
1074         struct extent_buffer *leaf = path->nodes[0];
1075         struct btrfs_extent_item *ei;
1076         struct btrfs_extent_data_ref *dref = NULL;
1077         struct btrfs_shared_data_ref *sref = NULL;
1078         unsigned long ptr;
1079         unsigned long end;
1080         u32 item_size;
1081         int size;
1082         int type;
1083         u64 refs;
1084
1085         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1086         refs = btrfs_extent_refs(leaf, ei);
1087         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1088         refs += refs_to_mod;
1089         btrfs_set_extent_refs(leaf, ei, refs);
1090         if (extent_op)
1091                 __run_delayed_extent_op(extent_op, leaf, ei);
1092
1093         /*
1094          * If type is invalid, we should have bailed out after
1095          * lookup_inline_extent_backref().
1096          */
1097         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1098         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1099
1100         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1101                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1102                 refs = btrfs_extent_data_ref_count(leaf, dref);
1103         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1104                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1105                 refs = btrfs_shared_data_ref_count(leaf, sref);
1106         } else {
1107                 refs = 1;
1108                 BUG_ON(refs_to_mod != -1);
1109         }
1110
1111         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1112         refs += refs_to_mod;
1113
1114         if (refs > 0) {
1115                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1116                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1117                 else
1118                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1119         } else {
1120                 *last_ref = 1;
1121                 size =  btrfs_extent_inline_ref_size(type);
1122                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1123                 ptr = (unsigned long)iref;
1124                 end = (unsigned long)ei + item_size;
1125                 if (ptr + size < end)
1126                         memmove_extent_buffer(leaf, ptr, ptr + size,
1127                                               end - ptr - size);
1128                 item_size -= size;
1129                 btrfs_truncate_item(path, item_size, 1);
1130         }
1131         btrfs_mark_buffer_dirty(leaf);
1132 }
1133
1134 static noinline_for_stack
1135 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1136                                  struct btrfs_path *path,
1137                                  u64 bytenr, u64 num_bytes, u64 parent,
1138                                  u64 root_objectid, u64 owner,
1139                                  u64 offset, int refs_to_add,
1140                                  struct btrfs_delayed_extent_op *extent_op)
1141 {
1142         struct btrfs_extent_inline_ref *iref;
1143         int ret;
1144
1145         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1146                                            num_bytes, parent, root_objectid,
1147                                            owner, offset, 1);
1148         if (ret == 0) {
1149                 /*
1150                  * We're adding refs to a tree block we already own, this
1151                  * should not happen at all.
1152                  */
1153                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1154                         btrfs_crit(trans->fs_info,
1155 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu",
1156                                    bytenr, num_bytes, root_objectid);
1157                         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
1158                                 WARN_ON(1);
1159                                 btrfs_crit(trans->fs_info,
1160                         "path->slots[0]=%d path->nodes[0]:", path->slots[0]);
1161                                 btrfs_print_leaf(path->nodes[0]);
1162                         }
1163                         return -EUCLEAN;
1164                 }
1165                 update_inline_extent_backref(path, iref, refs_to_add,
1166                                              extent_op, NULL);
1167         } else if (ret == -ENOENT) {
1168                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1169                                             root_objectid, owner, offset,
1170                                             refs_to_add, extent_op);
1171                 ret = 0;
1172         }
1173         return ret;
1174 }
1175
1176 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1177                                  struct btrfs_path *path,
1178                                  struct btrfs_extent_inline_ref *iref,
1179                                  int refs_to_drop, int is_data, int *last_ref)
1180 {
1181         int ret = 0;
1182
1183         BUG_ON(!is_data && refs_to_drop != 1);
1184         if (iref) {
1185                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1186                                              last_ref);
1187         } else if (is_data) {
1188                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1189                                              last_ref);
1190         } else {
1191                 *last_ref = 1;
1192                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1193         }
1194         return ret;
1195 }
1196
1197 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1198                                u64 *discarded_bytes)
1199 {
1200         int j, ret = 0;
1201         u64 bytes_left, end;
1202         u64 aligned_start = ALIGN(start, 1 << 9);
1203
1204         if (WARN_ON(start != aligned_start)) {
1205                 len -= aligned_start - start;
1206                 len = round_down(len, 1 << 9);
1207                 start = aligned_start;
1208         }
1209
1210         *discarded_bytes = 0;
1211
1212         if (!len)
1213                 return 0;
1214
1215         end = start + len;
1216         bytes_left = len;
1217
1218         /* Skip any superblocks on this device. */
1219         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1220                 u64 sb_start = btrfs_sb_offset(j);
1221                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1222                 u64 size = sb_start - start;
1223
1224                 if (!in_range(sb_start, start, bytes_left) &&
1225                     !in_range(sb_end, start, bytes_left) &&
1226                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1227                         continue;
1228
1229                 /*
1230                  * Superblock spans beginning of range.  Adjust start and
1231                  * try again.
1232                  */
1233                 if (sb_start <= start) {
1234                         start += sb_end - start;
1235                         if (start > end) {
1236                                 bytes_left = 0;
1237                                 break;
1238                         }
1239                         bytes_left = end - start;
1240                         continue;
1241                 }
1242
1243                 if (size) {
1244                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1245                                                    GFP_NOFS, 0);
1246                         if (!ret)
1247                                 *discarded_bytes += size;
1248                         else if (ret != -EOPNOTSUPP)
1249                                 return ret;
1250                 }
1251
1252                 start = sb_end;
1253                 if (start > end) {
1254                         bytes_left = 0;
1255                         break;
1256                 }
1257                 bytes_left = end - start;
1258         }
1259
1260         if (bytes_left) {
1261                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1262                                            GFP_NOFS, 0);
1263                 if (!ret)
1264                         *discarded_bytes += bytes_left;
1265         }
1266         return ret;
1267 }
1268
1269 static int do_discard_extent(struct btrfs_bio_stripe *stripe, u64 *bytes)
1270 {
1271         struct btrfs_device *dev = stripe->dev;
1272         struct btrfs_fs_info *fs_info = dev->fs_info;
1273         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1274         u64 phys = stripe->physical;
1275         u64 len = stripe->length;
1276         u64 discarded = 0;
1277         int ret = 0;
1278
1279         /* Zone reset on a zoned filesystem */
1280         if (btrfs_can_zone_reset(dev, phys, len)) {
1281                 u64 src_disc;
1282
1283                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1284                 if (ret)
1285                         goto out;
1286
1287                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1288                     dev != dev_replace->srcdev)
1289                         goto out;
1290
1291                 src_disc = discarded;
1292
1293                 /* Send to replace target as well */
1294                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1295                                               &discarded);
1296                 discarded += src_disc;
1297         } else if (blk_queue_discard(bdev_get_queue(stripe->dev->bdev))) {
1298                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1299         } else {
1300                 ret = 0;
1301                 *bytes = 0;
1302         }
1303
1304 out:
1305         *bytes = discarded;
1306         return ret;
1307 }
1308
1309 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1310                          u64 num_bytes, u64 *actual_bytes)
1311 {
1312         int ret = 0;
1313         u64 discarded_bytes = 0;
1314         u64 end = bytenr + num_bytes;
1315         u64 cur = bytenr;
1316         struct btrfs_bio *bbio = NULL;
1317
1318
1319         /*
1320          * Avoid races with device replace and make sure our bbio has devices
1321          * associated to its stripes that don't go away while we are discarding.
1322          */
1323         btrfs_bio_counter_inc_blocked(fs_info);
1324         while (cur < end) {
1325                 struct btrfs_bio_stripe *stripe;
1326                 int i;
1327
1328                 num_bytes = end - cur;
1329                 /* Tell the block device(s) that the sectors can be discarded */
1330                 ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, cur,
1331                                       &num_bytes, &bbio, 0);
1332                 /*
1333                  * Error can be -ENOMEM, -ENOENT (no such chunk mapping) or
1334                  * -EOPNOTSUPP. For any such error, @num_bytes is not updated,
1335                  * thus we can't continue anyway.
1336                  */
1337                 if (ret < 0)
1338                         goto out;
1339
1340                 stripe = bbio->stripes;
1341                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1342                         u64 bytes;
1343
1344                         if (!stripe->dev->bdev) {
1345                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1346                                 continue;
1347                         }
1348
1349                         ret = do_discard_extent(stripe, &bytes);
1350                         if (!ret) {
1351                                 discarded_bytes += bytes;
1352                         } else if (ret != -EOPNOTSUPP) {
1353                                 /*
1354                                  * Logic errors or -ENOMEM, or -EIO, but
1355                                  * unlikely to happen.
1356                                  *
1357                                  * And since there are two loops, explicitly
1358                                  * go to out to avoid confusion.
1359                                  */
1360                                 btrfs_put_bbio(bbio);
1361                                 goto out;
1362                         }
1363
1364                         /*
1365                          * Just in case we get back EOPNOTSUPP for some reason,
1366                          * just ignore the return value so we don't screw up
1367                          * people calling discard_extent.
1368                          */
1369                         ret = 0;
1370                 }
1371                 btrfs_put_bbio(bbio);
1372                 cur += num_bytes;
1373         }
1374 out:
1375         btrfs_bio_counter_dec(fs_info);
1376
1377         if (actual_bytes)
1378                 *actual_bytes = discarded_bytes;
1379
1380
1381         if (ret == -EOPNOTSUPP)
1382                 ret = 0;
1383         return ret;
1384 }
1385
1386 /* Can return -ENOMEM */
1387 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1388                          struct btrfs_ref *generic_ref)
1389 {
1390         struct btrfs_fs_info *fs_info = trans->fs_info;
1391         int ret;
1392
1393         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1394                generic_ref->action);
1395         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1396                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
1397
1398         if (generic_ref->type == BTRFS_REF_METADATA)
1399                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1400         else
1401                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1402
1403         btrfs_ref_tree_mod(fs_info, generic_ref);
1404
1405         return ret;
1406 }
1407
1408 /*
1409  * __btrfs_inc_extent_ref - insert backreference for a given extent
1410  *
1411  * The counterpart is in __btrfs_free_extent(), with examples and more details
1412  * how it works.
1413  *
1414  * @trans:          Handle of transaction
1415  *
1416  * @node:           The delayed ref node used to get the bytenr/length for
1417  *                  extent whose references are incremented.
1418  *
1419  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1420  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1421  *                  bytenr of the parent block. Since new extents are always
1422  *                  created with indirect references, this will only be the case
1423  *                  when relocating a shared extent. In that case, root_objectid
1424  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
1425  *                  be 0
1426  *
1427  * @root_objectid:  The id of the root where this modification has originated,
1428  *                  this can be either one of the well-known metadata trees or
1429  *                  the subvolume id which references this extent.
1430  *
1431  * @owner:          For data extents it is the inode number of the owning file.
1432  *                  For metadata extents this parameter holds the level in the
1433  *                  tree of the extent.
1434  *
1435  * @offset:         For metadata extents the offset is ignored and is currently
1436  *                  always passed as 0. For data extents it is the fileoffset
1437  *                  this extent belongs to.
1438  *
1439  * @refs_to_add     Number of references to add
1440  *
1441  * @extent_op       Pointer to a structure, holding information necessary when
1442  *                  updating a tree block's flags
1443  *
1444  */
1445 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1446                                   struct btrfs_delayed_ref_node *node,
1447                                   u64 parent, u64 root_objectid,
1448                                   u64 owner, u64 offset, int refs_to_add,
1449                                   struct btrfs_delayed_extent_op *extent_op)
1450 {
1451         struct btrfs_path *path;
1452         struct extent_buffer *leaf;
1453         struct btrfs_extent_item *item;
1454         struct btrfs_key key;
1455         u64 bytenr = node->bytenr;
1456         u64 num_bytes = node->num_bytes;
1457         u64 refs;
1458         int ret;
1459
1460         path = btrfs_alloc_path();
1461         if (!path)
1462                 return -ENOMEM;
1463
1464         /* this will setup the path even if it fails to insert the back ref */
1465         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1466                                            parent, root_objectid, owner,
1467                                            offset, refs_to_add, extent_op);
1468         if ((ret < 0 && ret != -EAGAIN) || !ret)
1469                 goto out;
1470
1471         /*
1472          * Ok we had -EAGAIN which means we didn't have space to insert and
1473          * inline extent ref, so just update the reference count and add a
1474          * normal backref.
1475          */
1476         leaf = path->nodes[0];
1477         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1478         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1479         refs = btrfs_extent_refs(leaf, item);
1480         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1481         if (extent_op)
1482                 __run_delayed_extent_op(extent_op, leaf, item);
1483
1484         btrfs_mark_buffer_dirty(leaf);
1485         btrfs_release_path(path);
1486
1487         /* now insert the actual backref */
1488         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1489                 BUG_ON(refs_to_add != 1);
1490                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1491                                             root_objectid);
1492         } else {
1493                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1494                                              root_objectid, owner, offset,
1495                                              refs_to_add);
1496         }
1497         if (ret)
1498                 btrfs_abort_transaction(trans, ret);
1499 out:
1500         btrfs_free_path(path);
1501         return ret;
1502 }
1503
1504 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1505                                 struct btrfs_delayed_ref_node *node,
1506                                 struct btrfs_delayed_extent_op *extent_op,
1507                                 int insert_reserved)
1508 {
1509         int ret = 0;
1510         struct btrfs_delayed_data_ref *ref;
1511         struct btrfs_key ins;
1512         u64 parent = 0;
1513         u64 ref_root = 0;
1514         u64 flags = 0;
1515
1516         ins.objectid = node->bytenr;
1517         ins.offset = node->num_bytes;
1518         ins.type = BTRFS_EXTENT_ITEM_KEY;
1519
1520         ref = btrfs_delayed_node_to_data_ref(node);
1521         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1522
1523         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1524                 parent = ref->parent;
1525         ref_root = ref->root;
1526
1527         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1528                 if (extent_op)
1529                         flags |= extent_op->flags_to_set;
1530                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1531                                                  flags, ref->objectid,
1532                                                  ref->offset, &ins,
1533                                                  node->ref_mod);
1534         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1535                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1536                                              ref->objectid, ref->offset,
1537                                              node->ref_mod, extent_op);
1538         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1539                 ret = __btrfs_free_extent(trans, node, parent,
1540                                           ref_root, ref->objectid,
1541                                           ref->offset, node->ref_mod,
1542                                           extent_op);
1543         } else {
1544                 BUG();
1545         }
1546         return ret;
1547 }
1548
1549 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1550                                     struct extent_buffer *leaf,
1551                                     struct btrfs_extent_item *ei)
1552 {
1553         u64 flags = btrfs_extent_flags(leaf, ei);
1554         if (extent_op->update_flags) {
1555                 flags |= extent_op->flags_to_set;
1556                 btrfs_set_extent_flags(leaf, ei, flags);
1557         }
1558
1559         if (extent_op->update_key) {
1560                 struct btrfs_tree_block_info *bi;
1561                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1562                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1563                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1564         }
1565 }
1566
1567 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1568                                  struct btrfs_delayed_ref_head *head,
1569                                  struct btrfs_delayed_extent_op *extent_op)
1570 {
1571         struct btrfs_fs_info *fs_info = trans->fs_info;
1572         struct btrfs_key key;
1573         struct btrfs_path *path;
1574         struct btrfs_extent_item *ei;
1575         struct extent_buffer *leaf;
1576         u32 item_size;
1577         int ret;
1578         int err = 0;
1579         int metadata = !extent_op->is_data;
1580
1581         if (TRANS_ABORTED(trans))
1582                 return 0;
1583
1584         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1585                 metadata = 0;
1586
1587         path = btrfs_alloc_path();
1588         if (!path)
1589                 return -ENOMEM;
1590
1591         key.objectid = head->bytenr;
1592
1593         if (metadata) {
1594                 key.type = BTRFS_METADATA_ITEM_KEY;
1595                 key.offset = extent_op->level;
1596         } else {
1597                 key.type = BTRFS_EXTENT_ITEM_KEY;
1598                 key.offset = head->num_bytes;
1599         }
1600
1601 again:
1602         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
1603         if (ret < 0) {
1604                 err = ret;
1605                 goto out;
1606         }
1607         if (ret > 0) {
1608                 if (metadata) {
1609                         if (path->slots[0] > 0) {
1610                                 path->slots[0]--;
1611                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1612                                                       path->slots[0]);
1613                                 if (key.objectid == head->bytenr &&
1614                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1615                                     key.offset == head->num_bytes)
1616                                         ret = 0;
1617                         }
1618                         if (ret > 0) {
1619                                 btrfs_release_path(path);
1620                                 metadata = 0;
1621
1622                                 key.objectid = head->bytenr;
1623                                 key.offset = head->num_bytes;
1624                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1625                                 goto again;
1626                         }
1627                 } else {
1628                         err = -EIO;
1629                         goto out;
1630                 }
1631         }
1632
1633         leaf = path->nodes[0];
1634         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1635
1636         if (unlikely(item_size < sizeof(*ei))) {
1637                 err = -EINVAL;
1638                 btrfs_print_v0_err(fs_info);
1639                 btrfs_abort_transaction(trans, err);
1640                 goto out;
1641         }
1642
1643         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1644         __run_delayed_extent_op(extent_op, leaf, ei);
1645
1646         btrfs_mark_buffer_dirty(leaf);
1647 out:
1648         btrfs_free_path(path);
1649         return err;
1650 }
1651
1652 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1653                                 struct btrfs_delayed_ref_node *node,
1654                                 struct btrfs_delayed_extent_op *extent_op,
1655                                 int insert_reserved)
1656 {
1657         int ret = 0;
1658         struct btrfs_delayed_tree_ref *ref;
1659         u64 parent = 0;
1660         u64 ref_root = 0;
1661
1662         ref = btrfs_delayed_node_to_tree_ref(node);
1663         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1664
1665         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1666                 parent = ref->parent;
1667         ref_root = ref->root;
1668
1669         if (node->ref_mod != 1) {
1670                 btrfs_err(trans->fs_info,
1671         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1672                           node->bytenr, node->ref_mod, node->action, ref_root,
1673                           parent);
1674                 return -EIO;
1675         }
1676         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1677                 BUG_ON(!extent_op || !extent_op->update_flags);
1678                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1679         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1680                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1681                                              ref->level, 0, 1, extent_op);
1682         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1683                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1684                                           ref->level, 0, 1, extent_op);
1685         } else {
1686                 BUG();
1687         }
1688         return ret;
1689 }
1690
1691 /* helper function to actually process a single delayed ref entry */
1692 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1693                                struct btrfs_delayed_ref_node *node,
1694                                struct btrfs_delayed_extent_op *extent_op,
1695                                int insert_reserved)
1696 {
1697         int ret = 0;
1698
1699         if (TRANS_ABORTED(trans)) {
1700                 if (insert_reserved)
1701                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1702                 return 0;
1703         }
1704
1705         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1706             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1707                 ret = run_delayed_tree_ref(trans, node, extent_op,
1708                                            insert_reserved);
1709         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1710                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1711                 ret = run_delayed_data_ref(trans, node, extent_op,
1712                                            insert_reserved);
1713         else
1714                 BUG();
1715         if (ret && insert_reserved)
1716                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1717         return ret;
1718 }
1719
1720 static inline struct btrfs_delayed_ref_node *
1721 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1722 {
1723         struct btrfs_delayed_ref_node *ref;
1724
1725         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1726                 return NULL;
1727
1728         /*
1729          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1730          * This is to prevent a ref count from going down to zero, which deletes
1731          * the extent item from the extent tree, when there still are references
1732          * to add, which would fail because they would not find the extent item.
1733          */
1734         if (!list_empty(&head->ref_add_list))
1735                 return list_first_entry(&head->ref_add_list,
1736                                 struct btrfs_delayed_ref_node, add_list);
1737
1738         ref = rb_entry(rb_first_cached(&head->ref_tree),
1739                        struct btrfs_delayed_ref_node, ref_node);
1740         ASSERT(list_empty(&ref->add_list));
1741         return ref;
1742 }
1743
1744 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1745                                       struct btrfs_delayed_ref_head *head)
1746 {
1747         spin_lock(&delayed_refs->lock);
1748         head->processing = 0;
1749         delayed_refs->num_heads_ready++;
1750         spin_unlock(&delayed_refs->lock);
1751         btrfs_delayed_ref_unlock(head);
1752 }
1753
1754 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1755                                 struct btrfs_delayed_ref_head *head)
1756 {
1757         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1758
1759         if (!extent_op)
1760                 return NULL;
1761
1762         if (head->must_insert_reserved) {
1763                 head->extent_op = NULL;
1764                 btrfs_free_delayed_extent_op(extent_op);
1765                 return NULL;
1766         }
1767         return extent_op;
1768 }
1769
1770 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1771                                      struct btrfs_delayed_ref_head *head)
1772 {
1773         struct btrfs_delayed_extent_op *extent_op;
1774         int ret;
1775
1776         extent_op = cleanup_extent_op(head);
1777         if (!extent_op)
1778                 return 0;
1779         head->extent_op = NULL;
1780         spin_unlock(&head->lock);
1781         ret = run_delayed_extent_op(trans, head, extent_op);
1782         btrfs_free_delayed_extent_op(extent_op);
1783         return ret ? ret : 1;
1784 }
1785
1786 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1787                                   struct btrfs_delayed_ref_root *delayed_refs,
1788                                   struct btrfs_delayed_ref_head *head)
1789 {
1790         int nr_items = 1;       /* Dropping this ref head update. */
1791
1792         /*
1793          * We had csum deletions accounted for in our delayed refs rsv, we need
1794          * to drop the csum leaves for this update from our delayed_refs_rsv.
1795          */
1796         if (head->total_ref_mod < 0 && head->is_data) {
1797                 spin_lock(&delayed_refs->lock);
1798                 delayed_refs->pending_csums -= head->num_bytes;
1799                 spin_unlock(&delayed_refs->lock);
1800                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1801         }
1802
1803         /*
1804          * We were dropping refs, or had a new ref and dropped it, and thus must
1805          * adjust down our total_bytes_pinned, the space may or may not have
1806          * been pinned and so is accounted for properly in the pinned space by
1807          * now.
1808          */
1809         if (head->total_ref_mod < 0 ||
1810             (head->total_ref_mod == 0 && head->must_insert_reserved)) {
1811                 u64 flags = btrfs_ref_head_to_space_flags(head);
1812
1813                 btrfs_mod_total_bytes_pinned(fs_info, flags, -head->num_bytes);
1814         }
1815
1816         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1817 }
1818
1819 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1820                             struct btrfs_delayed_ref_head *head)
1821 {
1822
1823         struct btrfs_fs_info *fs_info = trans->fs_info;
1824         struct btrfs_delayed_ref_root *delayed_refs;
1825         int ret;
1826
1827         delayed_refs = &trans->transaction->delayed_refs;
1828
1829         ret = run_and_cleanup_extent_op(trans, head);
1830         if (ret < 0) {
1831                 unselect_delayed_ref_head(delayed_refs, head);
1832                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1833                 return ret;
1834         } else if (ret) {
1835                 return ret;
1836         }
1837
1838         /*
1839          * Need to drop our head ref lock and re-acquire the delayed ref lock
1840          * and then re-check to make sure nobody got added.
1841          */
1842         spin_unlock(&head->lock);
1843         spin_lock(&delayed_refs->lock);
1844         spin_lock(&head->lock);
1845         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1846                 spin_unlock(&head->lock);
1847                 spin_unlock(&delayed_refs->lock);
1848                 return 1;
1849         }
1850         btrfs_delete_ref_head(delayed_refs, head);
1851         spin_unlock(&head->lock);
1852         spin_unlock(&delayed_refs->lock);
1853
1854         if (head->must_insert_reserved) {
1855                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1856                 if (head->is_data) {
1857                         ret = btrfs_del_csums(trans, fs_info->csum_root,
1858                                               head->bytenr, head->num_bytes);
1859                 }
1860         }
1861
1862         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1863
1864         trace_run_delayed_ref_head(fs_info, head, 0);
1865         btrfs_delayed_ref_unlock(head);
1866         btrfs_put_delayed_ref_head(head);
1867         return 0;
1868 }
1869
1870 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1871                                         struct btrfs_trans_handle *trans)
1872 {
1873         struct btrfs_delayed_ref_root *delayed_refs =
1874                 &trans->transaction->delayed_refs;
1875         struct btrfs_delayed_ref_head *head = NULL;
1876         int ret;
1877
1878         spin_lock(&delayed_refs->lock);
1879         head = btrfs_select_ref_head(delayed_refs);
1880         if (!head) {
1881                 spin_unlock(&delayed_refs->lock);
1882                 return head;
1883         }
1884
1885         /*
1886          * Grab the lock that says we are going to process all the refs for
1887          * this head
1888          */
1889         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1890         spin_unlock(&delayed_refs->lock);
1891
1892         /*
1893          * We may have dropped the spin lock to get the head mutex lock, and
1894          * that might have given someone else time to free the head.  If that's
1895          * true, it has been removed from our list and we can move on.
1896          */
1897         if (ret == -EAGAIN)
1898                 head = ERR_PTR(-EAGAIN);
1899
1900         return head;
1901 }
1902
1903 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1904                                     struct btrfs_delayed_ref_head *locked_ref,
1905                                     unsigned long *run_refs)
1906 {
1907         struct btrfs_fs_info *fs_info = trans->fs_info;
1908         struct btrfs_delayed_ref_root *delayed_refs;
1909         struct btrfs_delayed_extent_op *extent_op;
1910         struct btrfs_delayed_ref_node *ref;
1911         int must_insert_reserved = 0;
1912         int ret;
1913
1914         delayed_refs = &trans->transaction->delayed_refs;
1915
1916         lockdep_assert_held(&locked_ref->mutex);
1917         lockdep_assert_held(&locked_ref->lock);
1918
1919         while ((ref = select_delayed_ref(locked_ref))) {
1920                 if (ref->seq &&
1921                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1922                         spin_unlock(&locked_ref->lock);
1923                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1924                         return -EAGAIN;
1925                 }
1926
1927                 (*run_refs)++;
1928                 ref->in_tree = 0;
1929                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1930                 RB_CLEAR_NODE(&ref->ref_node);
1931                 if (!list_empty(&ref->add_list))
1932                         list_del(&ref->add_list);
1933                 /*
1934                  * When we play the delayed ref, also correct the ref_mod on
1935                  * head
1936                  */
1937                 switch (ref->action) {
1938                 case BTRFS_ADD_DELAYED_REF:
1939                 case BTRFS_ADD_DELAYED_EXTENT:
1940                         locked_ref->ref_mod -= ref->ref_mod;
1941                         break;
1942                 case BTRFS_DROP_DELAYED_REF:
1943                         locked_ref->ref_mod += ref->ref_mod;
1944                         break;
1945                 default:
1946                         WARN_ON(1);
1947                 }
1948                 atomic_dec(&delayed_refs->num_entries);
1949
1950                 /*
1951                  * Record the must_insert_reserved flag before we drop the
1952                  * spin lock.
1953                  */
1954                 must_insert_reserved = locked_ref->must_insert_reserved;
1955                 locked_ref->must_insert_reserved = 0;
1956
1957                 extent_op = locked_ref->extent_op;
1958                 locked_ref->extent_op = NULL;
1959                 spin_unlock(&locked_ref->lock);
1960
1961                 ret = run_one_delayed_ref(trans, ref, extent_op,
1962                                           must_insert_reserved);
1963
1964                 btrfs_free_delayed_extent_op(extent_op);
1965                 if (ret) {
1966                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1967                         btrfs_put_delayed_ref(ref);
1968                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
1969                                     ret);
1970                         return ret;
1971                 }
1972
1973                 btrfs_put_delayed_ref(ref);
1974                 cond_resched();
1975
1976                 spin_lock(&locked_ref->lock);
1977                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1978         }
1979
1980         return 0;
1981 }
1982
1983 /*
1984  * Returns 0 on success or if called with an already aborted transaction.
1985  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1986  */
1987 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1988                                              unsigned long nr)
1989 {
1990         struct btrfs_fs_info *fs_info = trans->fs_info;
1991         struct btrfs_delayed_ref_root *delayed_refs;
1992         struct btrfs_delayed_ref_head *locked_ref = NULL;
1993         ktime_t start = ktime_get();
1994         int ret;
1995         unsigned long count = 0;
1996         unsigned long actual_count = 0;
1997
1998         delayed_refs = &trans->transaction->delayed_refs;
1999         do {
2000                 if (!locked_ref) {
2001                         locked_ref = btrfs_obtain_ref_head(trans);
2002                         if (IS_ERR_OR_NULL(locked_ref)) {
2003                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2004                                         continue;
2005                                 } else {
2006                                         break;
2007                                 }
2008                         }
2009                         count++;
2010                 }
2011                 /*
2012                  * We need to try and merge add/drops of the same ref since we
2013                  * can run into issues with relocate dropping the implicit ref
2014                  * and then it being added back again before the drop can
2015                  * finish.  If we merged anything we need to re-loop so we can
2016                  * get a good ref.
2017                  * Or we can get node references of the same type that weren't
2018                  * merged when created due to bumps in the tree mod seq, and
2019                  * we need to merge them to prevent adding an inline extent
2020                  * backref before dropping it (triggering a BUG_ON at
2021                  * insert_inline_extent_backref()).
2022                  */
2023                 spin_lock(&locked_ref->lock);
2024                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2025
2026                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2027                                                       &actual_count);
2028                 if (ret < 0 && ret != -EAGAIN) {
2029                         /*
2030                          * Error, btrfs_run_delayed_refs_for_head already
2031                          * unlocked everything so just bail out
2032                          */
2033                         return ret;
2034                 } else if (!ret) {
2035                         /*
2036                          * Success, perform the usual cleanup of a processed
2037                          * head
2038                          */
2039                         ret = cleanup_ref_head(trans, locked_ref);
2040                         if (ret > 0 ) {
2041                                 /* We dropped our lock, we need to loop. */
2042                                 ret = 0;
2043                                 continue;
2044                         } else if (ret) {
2045                                 return ret;
2046                         }
2047                 }
2048
2049                 /*
2050                  * Either success case or btrfs_run_delayed_refs_for_head
2051                  * returned -EAGAIN, meaning we need to select another head
2052                  */
2053
2054                 locked_ref = NULL;
2055                 cond_resched();
2056         } while ((nr != -1 && count < nr) || locked_ref);
2057
2058         /*
2059          * We don't want to include ref heads since we can have empty ref heads
2060          * and those will drastically skew our runtime down since we just do
2061          * accounting, no actual extent tree updates.
2062          */
2063         if (actual_count > 0) {
2064                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2065                 u64 avg;
2066
2067                 /*
2068                  * We weigh the current average higher than our current runtime
2069                  * to avoid large swings in the average.
2070                  */
2071                 spin_lock(&delayed_refs->lock);
2072                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2073                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2074                 spin_unlock(&delayed_refs->lock);
2075         }
2076         return 0;
2077 }
2078
2079 #ifdef SCRAMBLE_DELAYED_REFS
2080 /*
2081  * Normally delayed refs get processed in ascending bytenr order. This
2082  * correlates in most cases to the order added. To expose dependencies on this
2083  * order, we start to process the tree in the middle instead of the beginning
2084  */
2085 static u64 find_middle(struct rb_root *root)
2086 {
2087         struct rb_node *n = root->rb_node;
2088         struct btrfs_delayed_ref_node *entry;
2089         int alt = 1;
2090         u64 middle;
2091         u64 first = 0, last = 0;
2092
2093         n = rb_first(root);
2094         if (n) {
2095                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2096                 first = entry->bytenr;
2097         }
2098         n = rb_last(root);
2099         if (n) {
2100                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2101                 last = entry->bytenr;
2102         }
2103         n = root->rb_node;
2104
2105         while (n) {
2106                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2107                 WARN_ON(!entry->in_tree);
2108
2109                 middle = entry->bytenr;
2110
2111                 if (alt)
2112                         n = n->rb_left;
2113                 else
2114                         n = n->rb_right;
2115
2116                 alt = 1 - alt;
2117         }
2118         return middle;
2119 }
2120 #endif
2121
2122 /*
2123  * this starts processing the delayed reference count updates and
2124  * extent insertions we have queued up so far.  count can be
2125  * 0, which means to process everything in the tree at the start
2126  * of the run (but not newly added entries), or it can be some target
2127  * number you'd like to process.
2128  *
2129  * Returns 0 on success or if called with an aborted transaction
2130  * Returns <0 on error and aborts the transaction
2131  */
2132 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2133                            unsigned long count)
2134 {
2135         struct btrfs_fs_info *fs_info = trans->fs_info;
2136         struct rb_node *node;
2137         struct btrfs_delayed_ref_root *delayed_refs;
2138         struct btrfs_delayed_ref_head *head;
2139         int ret;
2140         int run_all = count == (unsigned long)-1;
2141
2142         /* We'll clean this up in btrfs_cleanup_transaction */
2143         if (TRANS_ABORTED(trans))
2144                 return 0;
2145
2146         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2147                 return 0;
2148
2149         delayed_refs = &trans->transaction->delayed_refs;
2150         if (count == 0)
2151                 count = delayed_refs->num_heads_ready;
2152
2153 again:
2154 #ifdef SCRAMBLE_DELAYED_REFS
2155         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2156 #endif
2157         ret = __btrfs_run_delayed_refs(trans, count);
2158         if (ret < 0) {
2159                 btrfs_abort_transaction(trans, ret);
2160                 return ret;
2161         }
2162
2163         if (run_all) {
2164                 btrfs_create_pending_block_groups(trans);
2165
2166                 spin_lock(&delayed_refs->lock);
2167                 node = rb_first_cached(&delayed_refs->href_root);
2168                 if (!node) {
2169                         spin_unlock(&delayed_refs->lock);
2170                         goto out;
2171                 }
2172                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2173                                 href_node);
2174                 refcount_inc(&head->refs);
2175                 spin_unlock(&delayed_refs->lock);
2176
2177                 /* Mutex was contended, block until it's released and retry. */
2178                 mutex_lock(&head->mutex);
2179                 mutex_unlock(&head->mutex);
2180
2181                 btrfs_put_delayed_ref_head(head);
2182                 cond_resched();
2183                 goto again;
2184         }
2185 out:
2186         return 0;
2187 }
2188
2189 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2190                                 struct extent_buffer *eb, u64 flags,
2191                                 int level, int is_data)
2192 {
2193         struct btrfs_delayed_extent_op *extent_op;
2194         int ret;
2195
2196         extent_op = btrfs_alloc_delayed_extent_op();
2197         if (!extent_op)
2198                 return -ENOMEM;
2199
2200         extent_op->flags_to_set = flags;
2201         extent_op->update_flags = true;
2202         extent_op->update_key = false;
2203         extent_op->is_data = is_data ? true : false;
2204         extent_op->level = level;
2205
2206         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2207         if (ret)
2208                 btrfs_free_delayed_extent_op(extent_op);
2209         return ret;
2210 }
2211
2212 static noinline int check_delayed_ref(struct btrfs_root *root,
2213                                       struct btrfs_path *path,
2214                                       u64 objectid, u64 offset, u64 bytenr)
2215 {
2216         struct btrfs_delayed_ref_head *head;
2217         struct btrfs_delayed_ref_node *ref;
2218         struct btrfs_delayed_data_ref *data_ref;
2219         struct btrfs_delayed_ref_root *delayed_refs;
2220         struct btrfs_transaction *cur_trans;
2221         struct rb_node *node;
2222         int ret = 0;
2223
2224         spin_lock(&root->fs_info->trans_lock);
2225         cur_trans = root->fs_info->running_transaction;
2226         if (cur_trans)
2227                 refcount_inc(&cur_trans->use_count);
2228         spin_unlock(&root->fs_info->trans_lock);
2229         if (!cur_trans)
2230                 return 0;
2231
2232         delayed_refs = &cur_trans->delayed_refs;
2233         spin_lock(&delayed_refs->lock);
2234         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2235         if (!head) {
2236                 spin_unlock(&delayed_refs->lock);
2237                 btrfs_put_transaction(cur_trans);
2238                 return 0;
2239         }
2240
2241         if (!mutex_trylock(&head->mutex)) {
2242                 refcount_inc(&head->refs);
2243                 spin_unlock(&delayed_refs->lock);
2244
2245                 btrfs_release_path(path);
2246
2247                 /*
2248                  * Mutex was contended, block until it's released and let
2249                  * caller try again
2250                  */
2251                 mutex_lock(&head->mutex);
2252                 mutex_unlock(&head->mutex);
2253                 btrfs_put_delayed_ref_head(head);
2254                 btrfs_put_transaction(cur_trans);
2255                 return -EAGAIN;
2256         }
2257         spin_unlock(&delayed_refs->lock);
2258
2259         spin_lock(&head->lock);
2260         /*
2261          * XXX: We should replace this with a proper search function in the
2262          * future.
2263          */
2264         for (node = rb_first_cached(&head->ref_tree); node;
2265              node = rb_next(node)) {
2266                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2267                 /* If it's a shared ref we know a cross reference exists */
2268                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2269                         ret = 1;
2270                         break;
2271                 }
2272
2273                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2274
2275                 /*
2276                  * If our ref doesn't match the one we're currently looking at
2277                  * then we have a cross reference.
2278                  */
2279                 if (data_ref->root != root->root_key.objectid ||
2280                     data_ref->objectid != objectid ||
2281                     data_ref->offset != offset) {
2282                         ret = 1;
2283                         break;
2284                 }
2285         }
2286         spin_unlock(&head->lock);
2287         mutex_unlock(&head->mutex);
2288         btrfs_put_transaction(cur_trans);
2289         return ret;
2290 }
2291
2292 static noinline int check_committed_ref(struct btrfs_root *root,
2293                                         struct btrfs_path *path,
2294                                         u64 objectid, u64 offset, u64 bytenr,
2295                                         bool strict)
2296 {
2297         struct btrfs_fs_info *fs_info = root->fs_info;
2298         struct btrfs_root *extent_root = fs_info->extent_root;
2299         struct extent_buffer *leaf;
2300         struct btrfs_extent_data_ref *ref;
2301         struct btrfs_extent_inline_ref *iref;
2302         struct btrfs_extent_item *ei;
2303         struct btrfs_key key;
2304         u32 item_size;
2305         int type;
2306         int ret;
2307
2308         key.objectid = bytenr;
2309         key.offset = (u64)-1;
2310         key.type = BTRFS_EXTENT_ITEM_KEY;
2311
2312         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2313         if (ret < 0)
2314                 goto out;
2315         BUG_ON(ret == 0); /* Corruption */
2316
2317         ret = -ENOENT;
2318         if (path->slots[0] == 0)
2319                 goto out;
2320
2321         path->slots[0]--;
2322         leaf = path->nodes[0];
2323         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2324
2325         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2326                 goto out;
2327
2328         ret = 1;
2329         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2330         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2331
2332         /* If extent item has more than 1 inline ref then it's shared */
2333         if (item_size != sizeof(*ei) +
2334             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2335                 goto out;
2336
2337         /*
2338          * If extent created before last snapshot => it's shared unless the
2339          * snapshot has been deleted. Use the heuristic if strict is false.
2340          */
2341         if (!strict &&
2342             (btrfs_extent_generation(leaf, ei) <=
2343              btrfs_root_last_snapshot(&root->root_item)))
2344                 goto out;
2345
2346         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2347
2348         /* If this extent has SHARED_DATA_REF then it's shared */
2349         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2350         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2351                 goto out;
2352
2353         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2354         if (btrfs_extent_refs(leaf, ei) !=
2355             btrfs_extent_data_ref_count(leaf, ref) ||
2356             btrfs_extent_data_ref_root(leaf, ref) !=
2357             root->root_key.objectid ||
2358             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2359             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2360                 goto out;
2361
2362         ret = 0;
2363 out:
2364         return ret;
2365 }
2366
2367 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2368                           u64 bytenr, bool strict)
2369 {
2370         struct btrfs_path *path;
2371         int ret;
2372
2373         path = btrfs_alloc_path();
2374         if (!path)
2375                 return -ENOMEM;
2376
2377         do {
2378                 ret = check_committed_ref(root, path, objectid,
2379                                           offset, bytenr, strict);
2380                 if (ret && ret != -ENOENT)
2381                         goto out;
2382
2383                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2384         } while (ret == -EAGAIN);
2385
2386 out:
2387         btrfs_free_path(path);
2388         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2389                 WARN_ON(ret > 0);
2390         return ret;
2391 }
2392
2393 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2394                            struct btrfs_root *root,
2395                            struct extent_buffer *buf,
2396                            int full_backref, int inc)
2397 {
2398         struct btrfs_fs_info *fs_info = root->fs_info;
2399         u64 bytenr;
2400         u64 num_bytes;
2401         u64 parent;
2402         u64 ref_root;
2403         u32 nritems;
2404         struct btrfs_key key;
2405         struct btrfs_file_extent_item *fi;
2406         struct btrfs_ref generic_ref = { 0 };
2407         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2408         int i;
2409         int action;
2410         int level;
2411         int ret = 0;
2412
2413         if (btrfs_is_testing(fs_info))
2414                 return 0;
2415
2416         ref_root = btrfs_header_owner(buf);
2417         nritems = btrfs_header_nritems(buf);
2418         level = btrfs_header_level(buf);
2419
2420         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2421                 return 0;
2422
2423         if (full_backref)
2424                 parent = buf->start;
2425         else
2426                 parent = 0;
2427         if (inc)
2428                 action = BTRFS_ADD_DELAYED_REF;
2429         else
2430                 action = BTRFS_DROP_DELAYED_REF;
2431
2432         for (i = 0; i < nritems; i++) {
2433                 if (level == 0) {
2434                         btrfs_item_key_to_cpu(buf, &key, i);
2435                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2436                                 continue;
2437                         fi = btrfs_item_ptr(buf, i,
2438                                             struct btrfs_file_extent_item);
2439                         if (btrfs_file_extent_type(buf, fi) ==
2440                             BTRFS_FILE_EXTENT_INLINE)
2441                                 continue;
2442                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2443                         if (bytenr == 0)
2444                                 continue;
2445
2446                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2447                         key.offset -= btrfs_file_extent_offset(buf, fi);
2448                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2449                                                num_bytes, parent);
2450                         generic_ref.real_root = root->root_key.objectid;
2451                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2452                                             key.offset);
2453                         generic_ref.skip_qgroup = for_reloc;
2454                         if (inc)
2455                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2456                         else
2457                                 ret = btrfs_free_extent(trans, &generic_ref);
2458                         if (ret)
2459                                 goto fail;
2460                 } else {
2461                         bytenr = btrfs_node_blockptr(buf, i);
2462                         num_bytes = fs_info->nodesize;
2463                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2464                                                num_bytes, parent);
2465                         generic_ref.real_root = root->root_key.objectid;
2466                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
2467                         generic_ref.skip_qgroup = for_reloc;
2468                         if (inc)
2469                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2470                         else
2471                                 ret = btrfs_free_extent(trans, &generic_ref);
2472                         if (ret)
2473                                 goto fail;
2474                 }
2475         }
2476         return 0;
2477 fail:
2478         return ret;
2479 }
2480
2481 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2482                   struct extent_buffer *buf, int full_backref)
2483 {
2484         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2485 }
2486
2487 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2488                   struct extent_buffer *buf, int full_backref)
2489 {
2490         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2491 }
2492
2493 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2494 {
2495         struct btrfs_fs_info *fs_info = root->fs_info;
2496         u64 flags;
2497         u64 ret;
2498
2499         if (data)
2500                 flags = BTRFS_BLOCK_GROUP_DATA;
2501         else if (root == fs_info->chunk_root)
2502                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2503         else
2504                 flags = BTRFS_BLOCK_GROUP_METADATA;
2505
2506         ret = btrfs_get_alloc_profile(fs_info, flags);
2507         return ret;
2508 }
2509
2510 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
2511 {
2512         struct btrfs_block_group *cache;
2513         u64 bytenr;
2514
2515         spin_lock(&fs_info->block_group_cache_lock);
2516         bytenr = fs_info->first_logical_byte;
2517         spin_unlock(&fs_info->block_group_cache_lock);
2518
2519         if (bytenr < (u64)-1)
2520                 return bytenr;
2521
2522         cache = btrfs_lookup_first_block_group(fs_info, search_start);
2523         if (!cache)
2524                 return 0;
2525
2526         bytenr = cache->start;
2527         btrfs_put_block_group(cache);
2528
2529         return bytenr;
2530 }
2531
2532 static int pin_down_extent(struct btrfs_trans_handle *trans,
2533                            struct btrfs_block_group *cache,
2534                            u64 bytenr, u64 num_bytes, int reserved)
2535 {
2536         struct btrfs_fs_info *fs_info = cache->fs_info;
2537
2538         spin_lock(&cache->space_info->lock);
2539         spin_lock(&cache->lock);
2540         cache->pinned += num_bytes;
2541         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2542                                              num_bytes);
2543         if (reserved) {
2544                 cache->reserved -= num_bytes;
2545                 cache->space_info->bytes_reserved -= num_bytes;
2546         }
2547         spin_unlock(&cache->lock);
2548         spin_unlock(&cache->space_info->lock);
2549
2550         __btrfs_mod_total_bytes_pinned(cache->space_info, num_bytes);
2551         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2552                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2553         return 0;
2554 }
2555
2556 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2557                      u64 bytenr, u64 num_bytes, int reserved)
2558 {
2559         struct btrfs_block_group *cache;
2560
2561         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2562         BUG_ON(!cache); /* Logic error */
2563
2564         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2565
2566         btrfs_put_block_group(cache);
2567         return 0;
2568 }
2569
2570 /*
2571  * this function must be called within transaction
2572  */
2573 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2574                                     u64 bytenr, u64 num_bytes)
2575 {
2576         struct btrfs_block_group *cache;
2577         int ret;
2578
2579         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2580         if (!cache)
2581                 return -EINVAL;
2582
2583         /*
2584          * pull in the free space cache (if any) so that our pin
2585          * removes the free space from the cache.  We have load_only set
2586          * to one because the slow code to read in the free extents does check
2587          * the pinned extents.
2588          */
2589         btrfs_cache_block_group(cache, 1);
2590         /*
2591          * Make sure we wait until the cache is completely built in case it is
2592          * missing or is invalid and therefore needs to be rebuilt.
2593          */
2594         ret = btrfs_wait_block_group_cache_done(cache);
2595         if (ret)
2596                 goto out;
2597
2598         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2599
2600         /* remove us from the free space cache (if we're there at all) */
2601         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2602 out:
2603         btrfs_put_block_group(cache);
2604         return ret;
2605 }
2606
2607 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2608                                    u64 start, u64 num_bytes)
2609 {
2610         int ret;
2611         struct btrfs_block_group *block_group;
2612
2613         block_group = btrfs_lookup_block_group(fs_info, start);
2614         if (!block_group)
2615                 return -EINVAL;
2616
2617         btrfs_cache_block_group(block_group, 1);
2618         /*
2619          * Make sure we wait until the cache is completely built in case it is
2620          * missing or is invalid and therefore needs to be rebuilt.
2621          */
2622         ret = btrfs_wait_block_group_cache_done(block_group);
2623         if (ret)
2624                 goto out;
2625
2626         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2627 out:
2628         btrfs_put_block_group(block_group);
2629         return ret;
2630 }
2631
2632 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2633 {
2634         struct btrfs_fs_info *fs_info = eb->fs_info;
2635         struct btrfs_file_extent_item *item;
2636         struct btrfs_key key;
2637         int found_type;
2638         int i;
2639         int ret = 0;
2640
2641         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2642                 return 0;
2643
2644         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2645                 btrfs_item_key_to_cpu(eb, &key, i);
2646                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2647                         continue;
2648                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2649                 found_type = btrfs_file_extent_type(eb, item);
2650                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2651                         continue;
2652                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2653                         continue;
2654                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2655                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2656                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2657                 if (ret)
2658                         break;
2659         }
2660
2661         return ret;
2662 }
2663
2664 static void
2665 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2666 {
2667         atomic_inc(&bg->reservations);
2668 }
2669
2670 /*
2671  * Returns the free cluster for the given space info and sets empty_cluster to
2672  * what it should be based on the mount options.
2673  */
2674 static struct btrfs_free_cluster *
2675 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2676                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2677 {
2678         struct btrfs_free_cluster *ret = NULL;
2679
2680         *empty_cluster = 0;
2681         if (btrfs_mixed_space_info(space_info))
2682                 return ret;
2683
2684         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2685                 ret = &fs_info->meta_alloc_cluster;
2686                 if (btrfs_test_opt(fs_info, SSD))
2687                         *empty_cluster = SZ_2M;
2688                 else
2689                         *empty_cluster = SZ_64K;
2690         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2691                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2692                 *empty_cluster = SZ_2M;
2693                 ret = &fs_info->data_alloc_cluster;
2694         }
2695
2696         return ret;
2697 }
2698
2699 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2700                               u64 start, u64 end,
2701                               const bool return_free_space)
2702 {
2703         struct btrfs_block_group *cache = NULL;
2704         struct btrfs_space_info *space_info;
2705         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2706         struct btrfs_free_cluster *cluster = NULL;
2707         u64 len;
2708         u64 total_unpinned = 0;
2709         u64 empty_cluster = 0;
2710         bool readonly;
2711
2712         while (start <= end) {
2713                 readonly = false;
2714                 if (!cache ||
2715                     start >= cache->start + cache->length) {
2716                         if (cache)
2717                                 btrfs_put_block_group(cache);
2718                         total_unpinned = 0;
2719                         cache = btrfs_lookup_block_group(fs_info, start);
2720                         BUG_ON(!cache); /* Logic error */
2721
2722                         cluster = fetch_cluster_info(fs_info,
2723                                                      cache->space_info,
2724                                                      &empty_cluster);
2725                         empty_cluster <<= 1;
2726                 }
2727
2728                 len = cache->start + cache->length - start;
2729                 len = min(len, end + 1 - start);
2730
2731                 down_read(&fs_info->commit_root_sem);
2732                 if (start < cache->last_byte_to_unpin && return_free_space) {
2733                         u64 add_len = min(len, cache->last_byte_to_unpin - start);
2734
2735                         btrfs_add_free_space(cache, start, add_len);
2736                 }
2737                 up_read(&fs_info->commit_root_sem);
2738
2739                 start += len;
2740                 total_unpinned += len;
2741                 space_info = cache->space_info;
2742
2743                 /*
2744                  * If this space cluster has been marked as fragmented and we've
2745                  * unpinned enough in this block group to potentially allow a
2746                  * cluster to be created inside of it go ahead and clear the
2747                  * fragmented check.
2748                  */
2749                 if (cluster && cluster->fragmented &&
2750                     total_unpinned > empty_cluster) {
2751                         spin_lock(&cluster->lock);
2752                         cluster->fragmented = 0;
2753                         spin_unlock(&cluster->lock);
2754                 }
2755
2756                 spin_lock(&space_info->lock);
2757                 spin_lock(&cache->lock);
2758                 cache->pinned -= len;
2759                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2760                 space_info->max_extent_size = 0;
2761                 __btrfs_mod_total_bytes_pinned(space_info, -len);
2762                 if (cache->ro) {
2763                         space_info->bytes_readonly += len;
2764                         readonly = true;
2765                 } else if (btrfs_is_zoned(fs_info)) {
2766                         /* Need reset before reusing in a zoned block group */
2767                         space_info->bytes_zone_unusable += len;
2768                         readonly = true;
2769                 }
2770                 spin_unlock(&cache->lock);
2771                 if (!readonly && return_free_space &&
2772                     global_rsv->space_info == space_info) {
2773                         u64 to_add = len;
2774
2775                         spin_lock(&global_rsv->lock);
2776                         if (!global_rsv->full) {
2777                                 to_add = min(len, global_rsv->size -
2778                                              global_rsv->reserved);
2779                                 global_rsv->reserved += to_add;
2780                                 btrfs_space_info_update_bytes_may_use(fs_info,
2781                                                 space_info, to_add);
2782                                 if (global_rsv->reserved >= global_rsv->size)
2783                                         global_rsv->full = 1;
2784                                 len -= to_add;
2785                         }
2786                         spin_unlock(&global_rsv->lock);
2787                 }
2788                 /* Add to any tickets we may have */
2789                 if (!readonly && return_free_space && len)
2790                         btrfs_try_granting_tickets(fs_info, space_info);
2791                 spin_unlock(&space_info->lock);
2792         }
2793
2794         if (cache)
2795                 btrfs_put_block_group(cache);
2796         return 0;
2797 }
2798
2799 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2800 {
2801         struct btrfs_fs_info *fs_info = trans->fs_info;
2802         struct btrfs_block_group *block_group, *tmp;
2803         struct list_head *deleted_bgs;
2804         struct extent_io_tree *unpin;
2805         u64 start;
2806         u64 end;
2807         int ret;
2808
2809         unpin = &trans->transaction->pinned_extents;
2810
2811         while (!TRANS_ABORTED(trans)) {
2812                 struct extent_state *cached_state = NULL;
2813
2814                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2815                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2816                                             EXTENT_DIRTY, &cached_state);
2817                 if (ret) {
2818                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2819                         break;
2820                 }
2821
2822                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2823                         ret = btrfs_discard_extent(fs_info, start,
2824                                                    end + 1 - start, NULL);
2825
2826                 clear_extent_dirty(unpin, start, end, &cached_state);
2827                 unpin_extent_range(fs_info, start, end, true);
2828                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2829                 free_extent_state(cached_state);
2830                 cond_resched();
2831         }
2832
2833         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2834                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2835                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2836         }
2837
2838         /*
2839          * Transaction is finished.  We don't need the lock anymore.  We
2840          * do need to clean up the block groups in case of a transaction
2841          * abort.
2842          */
2843         deleted_bgs = &trans->transaction->deleted_bgs;
2844         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2845                 u64 trimmed = 0;
2846
2847                 ret = -EROFS;
2848                 if (!TRANS_ABORTED(trans))
2849                         ret = btrfs_discard_extent(fs_info,
2850                                                    block_group->start,
2851                                                    block_group->length,
2852                                                    &trimmed);
2853
2854                 list_del_init(&block_group->bg_list);
2855                 btrfs_unfreeze_block_group(block_group);
2856                 btrfs_put_block_group(block_group);
2857
2858                 if (ret) {
2859                         const char *errstr = btrfs_decode_error(ret);
2860                         btrfs_warn(fs_info,
2861                            "discard failed while removing blockgroup: errno=%d %s",
2862                                    ret, errstr);
2863                 }
2864         }
2865
2866         return 0;
2867 }
2868
2869 /*
2870  * Drop one or more refs of @node.
2871  *
2872  * 1. Locate the extent refs.
2873  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2874  *    Locate it, then reduce the refs number or remove the ref line completely.
2875  *
2876  * 2. Update the refs count in EXTENT/METADATA_ITEM
2877  *
2878  * Inline backref case:
2879  *
2880  * in extent tree we have:
2881  *
2882  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2883  *              refs 2 gen 6 flags DATA
2884  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2885  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2886  *
2887  * This function gets called with:
2888  *
2889  *    node->bytenr = 13631488
2890  *    node->num_bytes = 1048576
2891  *    root_objectid = FS_TREE
2892  *    owner_objectid = 257
2893  *    owner_offset = 0
2894  *    refs_to_drop = 1
2895  *
2896  * Then we should get some like:
2897  *
2898  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2899  *              refs 1 gen 6 flags DATA
2900  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2901  *
2902  * Keyed backref case:
2903  *
2904  * in extent tree we have:
2905  *
2906  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2907  *              refs 754 gen 6 flags DATA
2908  *      [...]
2909  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2910  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2911  *
2912  * This function get called with:
2913  *
2914  *    node->bytenr = 13631488
2915  *    node->num_bytes = 1048576
2916  *    root_objectid = FS_TREE
2917  *    owner_objectid = 866
2918  *    owner_offset = 0
2919  *    refs_to_drop = 1
2920  *
2921  * Then we should get some like:
2922  *
2923  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2924  *              refs 753 gen 6 flags DATA
2925  *
2926  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2927  */
2928 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2929                                struct btrfs_delayed_ref_node *node, u64 parent,
2930                                u64 root_objectid, u64 owner_objectid,
2931                                u64 owner_offset, int refs_to_drop,
2932                                struct btrfs_delayed_extent_op *extent_op)
2933 {
2934         struct btrfs_fs_info *info = trans->fs_info;
2935         struct btrfs_key key;
2936         struct btrfs_path *path;
2937         struct btrfs_root *extent_root = info->extent_root;
2938         struct extent_buffer *leaf;
2939         struct btrfs_extent_item *ei;
2940         struct btrfs_extent_inline_ref *iref;
2941         int ret;
2942         int is_data;
2943         int extent_slot = 0;
2944         int found_extent = 0;
2945         int num_to_del = 1;
2946         u32 item_size;
2947         u64 refs;
2948         u64 bytenr = node->bytenr;
2949         u64 num_bytes = node->num_bytes;
2950         int last_ref = 0;
2951         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2952
2953         path = btrfs_alloc_path();
2954         if (!path)
2955                 return -ENOMEM;
2956
2957         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2958
2959         if (!is_data && refs_to_drop != 1) {
2960                 btrfs_crit(info,
2961 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2962                            node->bytenr, refs_to_drop);
2963                 ret = -EINVAL;
2964                 btrfs_abort_transaction(trans, ret);
2965                 goto out;
2966         }
2967
2968         if (is_data)
2969                 skinny_metadata = false;
2970
2971         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2972                                     parent, root_objectid, owner_objectid,
2973                                     owner_offset);
2974         if (ret == 0) {
2975                 /*
2976                  * Either the inline backref or the SHARED_DATA_REF/
2977                  * SHARED_BLOCK_REF is found
2978                  *
2979                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2980                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2981                  */
2982                 extent_slot = path->slots[0];
2983                 while (extent_slot >= 0) {
2984                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2985                                               extent_slot);
2986                         if (key.objectid != bytenr)
2987                                 break;
2988                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2989                             key.offset == num_bytes) {
2990                                 found_extent = 1;
2991                                 break;
2992                         }
2993                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2994                             key.offset == owner_objectid) {
2995                                 found_extent = 1;
2996                                 break;
2997                         }
2998
2999                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3000                         if (path->slots[0] - extent_slot > 5)
3001                                 break;
3002                         extent_slot--;
3003                 }
3004
3005                 if (!found_extent) {
3006                         if (iref) {
3007                                 btrfs_crit(info,
3008 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
3009                                 btrfs_abort_transaction(trans, -EUCLEAN);
3010                                 goto err_dump;
3011                         }
3012                         /* Must be SHARED_* item, remove the backref first */
3013                         ret = remove_extent_backref(trans, path, NULL,
3014                                                     refs_to_drop,
3015                                                     is_data, &last_ref);
3016                         if (ret) {
3017                                 btrfs_abort_transaction(trans, ret);
3018                                 goto out;
3019                         }
3020                         btrfs_release_path(path);
3021
3022                         /* Slow path to locate EXTENT/METADATA_ITEM */
3023                         key.objectid = bytenr;
3024                         key.type = BTRFS_EXTENT_ITEM_KEY;
3025                         key.offset = num_bytes;
3026
3027                         if (!is_data && skinny_metadata) {
3028                                 key.type = BTRFS_METADATA_ITEM_KEY;
3029                                 key.offset = owner_objectid;
3030                         }
3031
3032                         ret = btrfs_search_slot(trans, extent_root,
3033                                                 &key, path, -1, 1);
3034                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3035                                 /*
3036                                  * Couldn't find our skinny metadata item,
3037                                  * see if we have ye olde extent item.
3038                                  */
3039                                 path->slots[0]--;
3040                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3041                                                       path->slots[0]);
3042                                 if (key.objectid == bytenr &&
3043                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3044                                     key.offset == num_bytes)
3045                                         ret = 0;
3046                         }
3047
3048                         if (ret > 0 && skinny_metadata) {
3049                                 skinny_metadata = false;
3050                                 key.objectid = bytenr;
3051                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3052                                 key.offset = num_bytes;
3053                                 btrfs_release_path(path);
3054                                 ret = btrfs_search_slot(trans, extent_root,
3055                                                         &key, path, -1, 1);
3056                         }
3057
3058                         if (ret) {
3059                                 btrfs_err(info,
3060                                           "umm, got %d back from search, was looking for %llu",
3061                                           ret, bytenr);
3062                                 if (ret > 0)
3063                                         btrfs_print_leaf(path->nodes[0]);
3064                         }
3065                         if (ret < 0) {
3066                                 btrfs_abort_transaction(trans, ret);
3067                                 goto out;
3068                         }
3069                         extent_slot = path->slots[0];
3070                 }
3071         } else if (WARN_ON(ret == -ENOENT)) {
3072                 btrfs_print_leaf(path->nodes[0]);
3073                 btrfs_err(info,
3074                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3075                         bytenr, parent, root_objectid, owner_objectid,
3076                         owner_offset);
3077                 btrfs_abort_transaction(trans, ret);
3078                 goto out;
3079         } else {
3080                 btrfs_abort_transaction(trans, ret);
3081                 goto out;
3082         }
3083
3084         leaf = path->nodes[0];
3085         item_size = btrfs_item_size_nr(leaf, extent_slot);
3086         if (unlikely(item_size < sizeof(*ei))) {
3087                 ret = -EINVAL;
3088                 btrfs_print_v0_err(info);
3089                 btrfs_abort_transaction(trans, ret);
3090                 goto out;
3091         }
3092         ei = btrfs_item_ptr(leaf, extent_slot,
3093                             struct btrfs_extent_item);
3094         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3095             key.type == BTRFS_EXTENT_ITEM_KEY) {
3096                 struct btrfs_tree_block_info *bi;
3097                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3098                         btrfs_crit(info,
3099 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3100                                    key.objectid, key.type, key.offset,
3101                                    owner_objectid, item_size,
3102                                    sizeof(*ei) + sizeof(*bi));
3103                         btrfs_abort_transaction(trans, -EUCLEAN);
3104                         goto err_dump;
3105                 }
3106                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3107                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3108         }
3109
3110         refs = btrfs_extent_refs(leaf, ei);
3111         if (refs < refs_to_drop) {
3112                 btrfs_crit(info,
3113                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3114                           refs_to_drop, refs, bytenr);
3115                 btrfs_abort_transaction(trans, -EUCLEAN);
3116                 goto err_dump;
3117         }
3118         refs -= refs_to_drop;
3119
3120         if (refs > 0) {
3121                 if (extent_op)
3122                         __run_delayed_extent_op(extent_op, leaf, ei);
3123                 /*
3124                  * In the case of inline back ref, reference count will
3125                  * be updated by remove_extent_backref
3126                  */
3127                 if (iref) {
3128                         if (!found_extent) {
3129                                 btrfs_crit(info,
3130 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3131                                 btrfs_abort_transaction(trans, -EUCLEAN);
3132                                 goto err_dump;
3133                         }
3134                 } else {
3135                         btrfs_set_extent_refs(leaf, ei, refs);
3136                         btrfs_mark_buffer_dirty(leaf);
3137                 }
3138                 if (found_extent) {
3139                         ret = remove_extent_backref(trans, path, iref,
3140                                                     refs_to_drop, is_data,
3141                                                     &last_ref);
3142                         if (ret) {
3143                                 btrfs_abort_transaction(trans, ret);
3144                                 goto out;
3145                         }
3146                 }
3147         } else {
3148                 /* In this branch refs == 1 */
3149                 if (found_extent) {
3150                         if (is_data && refs_to_drop !=
3151                             extent_data_ref_count(path, iref)) {
3152                                 btrfs_crit(info,
3153                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3154                                            extent_data_ref_count(path, iref),
3155                                            refs_to_drop);
3156                                 btrfs_abort_transaction(trans, -EUCLEAN);
3157                                 goto err_dump;
3158                         }
3159                         if (iref) {
3160                                 if (path->slots[0] != extent_slot) {
3161                                         btrfs_crit(info,
3162 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3163                                                    key.objectid, key.type,
3164                                                    key.offset);
3165                                         btrfs_abort_transaction(trans, -EUCLEAN);
3166                                         goto err_dump;
3167                                 }
3168                         } else {
3169                                 /*
3170                                  * No inline ref, we must be at SHARED_* item,
3171                                  * And it's single ref, it must be:
3172                                  * |    extent_slot       ||extent_slot + 1|
3173                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3174                                  */
3175                                 if (path->slots[0] != extent_slot + 1) {
3176                                         btrfs_crit(info,
3177         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3178                                         btrfs_abort_transaction(trans, -EUCLEAN);
3179                                         goto err_dump;
3180                                 }
3181                                 path->slots[0] = extent_slot;
3182                                 num_to_del = 2;
3183                         }
3184                 }
3185
3186                 last_ref = 1;
3187                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3188                                       num_to_del);
3189                 if (ret) {
3190                         btrfs_abort_transaction(trans, ret);
3191                         goto out;
3192                 }
3193                 btrfs_release_path(path);
3194
3195                 if (is_data) {
3196                         ret = btrfs_del_csums(trans, info->csum_root, bytenr,
3197                                               num_bytes);
3198                         if (ret) {
3199                                 btrfs_abort_transaction(trans, ret);
3200                                 goto out;
3201                         }
3202                 }
3203
3204                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3205                 if (ret) {
3206                         btrfs_abort_transaction(trans, ret);
3207                         goto out;
3208                 }
3209
3210                 ret = btrfs_update_block_group(trans, bytenr, num_bytes, 0);
3211                 if (ret) {
3212                         btrfs_abort_transaction(trans, ret);
3213                         goto out;
3214                 }
3215         }
3216         btrfs_release_path(path);
3217
3218 out:
3219         btrfs_free_path(path);
3220         return ret;
3221 err_dump:
3222         /*
3223          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3224          * dump for debug build.
3225          */
3226         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3227                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3228                            path->slots[0], extent_slot);
3229                 btrfs_print_leaf(path->nodes[0]);
3230         }
3231
3232         btrfs_free_path(path);
3233         return -EUCLEAN;
3234 }
3235
3236 /*
3237  * when we free an block, it is possible (and likely) that we free the last
3238  * delayed ref for that extent as well.  This searches the delayed ref tree for
3239  * a given extent, and if there are no other delayed refs to be processed, it
3240  * removes it from the tree.
3241  */
3242 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3243                                       u64 bytenr)
3244 {
3245         struct btrfs_delayed_ref_head *head;
3246         struct btrfs_delayed_ref_root *delayed_refs;
3247         int ret = 0;
3248
3249         delayed_refs = &trans->transaction->delayed_refs;
3250         spin_lock(&delayed_refs->lock);
3251         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3252         if (!head)
3253                 goto out_delayed_unlock;
3254
3255         spin_lock(&head->lock);
3256         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3257                 goto out;
3258
3259         if (cleanup_extent_op(head) != NULL)
3260                 goto out;
3261
3262         /*
3263          * waiting for the lock here would deadlock.  If someone else has it
3264          * locked they are already in the process of dropping it anyway
3265          */
3266         if (!mutex_trylock(&head->mutex))
3267                 goto out;
3268
3269         btrfs_delete_ref_head(delayed_refs, head);
3270         head->processing = 0;
3271
3272         spin_unlock(&head->lock);
3273         spin_unlock(&delayed_refs->lock);
3274
3275         BUG_ON(head->extent_op);
3276         if (head->must_insert_reserved)
3277                 ret = 1;
3278
3279         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3280         mutex_unlock(&head->mutex);
3281         btrfs_put_delayed_ref_head(head);
3282         return ret;
3283 out:
3284         spin_unlock(&head->lock);
3285
3286 out_delayed_unlock:
3287         spin_unlock(&delayed_refs->lock);
3288         return 0;
3289 }
3290
3291 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3292                            struct btrfs_root *root,
3293                            struct extent_buffer *buf,
3294                            u64 parent, int last_ref)
3295 {
3296         struct btrfs_fs_info *fs_info = root->fs_info;
3297         struct btrfs_ref generic_ref = { 0 };
3298         int ret;
3299
3300         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3301                                buf->start, buf->len, parent);
3302         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3303                             root->root_key.objectid);
3304
3305         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3306                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3307                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3308                 BUG_ON(ret); /* -ENOMEM */
3309         }
3310
3311         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3312                 struct btrfs_block_group *cache;
3313                 bool must_pin = false;
3314
3315                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3316                         ret = check_ref_cleanup(trans, buf->start);
3317                         if (!ret) {
3318                                 btrfs_redirty_list_add(trans->transaction, buf);
3319                                 goto out;
3320                         }
3321                 }
3322
3323                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3324
3325                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3326                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3327                         btrfs_put_block_group(cache);
3328                         goto out;
3329                 }
3330
3331                 /*
3332                  * If this is a leaf and there are tree mod log users, we may
3333                  * have recorded mod log operations that point to this leaf.
3334                  * So we must make sure no one reuses this leaf's extent before
3335                  * mod log operations are applied to a node, otherwise after
3336                  * rewinding a node using the mod log operations we get an
3337                  * inconsistent btree, as the leaf's extent may now be used as
3338                  * a node or leaf for another different btree.
3339                  * We are safe from races here because at this point no other
3340                  * node or root points to this extent buffer, so if after this
3341                  * check a new tree mod log user joins, it will not be able to
3342                  * find a node pointing to this leaf and record operations that
3343                  * point to this leaf.
3344                  */
3345                 if (btrfs_header_level(buf) == 0 &&
3346                     test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3347                         must_pin = true;
3348
3349                 if (must_pin || btrfs_is_zoned(fs_info)) {
3350                         btrfs_redirty_list_add(trans->transaction, buf);
3351                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3352                         btrfs_put_block_group(cache);
3353                         goto out;
3354                 }
3355
3356                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3357
3358                 btrfs_add_free_space(cache, buf->start, buf->len);
3359                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3360                 btrfs_put_block_group(cache);
3361                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3362         }
3363 out:
3364         if (last_ref) {
3365                 /*
3366                  * Deleting the buffer, clear the corrupt flag since it doesn't
3367                  * matter anymore.
3368                  */
3369                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3370         }
3371 }
3372
3373 /* Can return -ENOMEM */
3374 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3375 {
3376         struct btrfs_fs_info *fs_info = trans->fs_info;
3377         int ret;
3378
3379         if (btrfs_is_testing(fs_info))
3380                 return 0;
3381
3382         /*
3383          * tree log blocks never actually go into the extent allocation
3384          * tree, just update pinning info and exit early.
3385          */
3386         if ((ref->type == BTRFS_REF_METADATA &&
3387              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3388             (ref->type == BTRFS_REF_DATA &&
3389              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3390                 /* unlocks the pinned mutex */
3391                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3392                 ret = 0;
3393         } else if (ref->type == BTRFS_REF_METADATA) {
3394                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3395         } else {
3396                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3397         }
3398
3399         if (!((ref->type == BTRFS_REF_METADATA &&
3400                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3401               (ref->type == BTRFS_REF_DATA &&
3402                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3403                 btrfs_ref_tree_mod(fs_info, ref);
3404
3405         return ret;
3406 }
3407
3408 enum btrfs_loop_type {
3409         LOOP_CACHING_NOWAIT,
3410         LOOP_CACHING_WAIT,
3411         LOOP_ALLOC_CHUNK,
3412         LOOP_NO_EMPTY_SIZE,
3413 };
3414
3415 static inline void
3416 btrfs_lock_block_group(struct btrfs_block_group *cache,
3417                        int delalloc)
3418 {
3419         if (delalloc)
3420                 down_read(&cache->data_rwsem);
3421 }
3422
3423 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3424                        int delalloc)
3425 {
3426         btrfs_get_block_group(cache);
3427         if (delalloc)
3428                 down_read(&cache->data_rwsem);
3429 }
3430
3431 static struct btrfs_block_group *btrfs_lock_cluster(
3432                    struct btrfs_block_group *block_group,
3433                    struct btrfs_free_cluster *cluster,
3434                    int delalloc)
3435         __acquires(&cluster->refill_lock)
3436 {
3437         struct btrfs_block_group *used_bg = NULL;
3438
3439         spin_lock(&cluster->refill_lock);
3440         while (1) {
3441                 used_bg = cluster->block_group;
3442                 if (!used_bg)
3443                         return NULL;
3444
3445                 if (used_bg == block_group)
3446                         return used_bg;
3447
3448                 btrfs_get_block_group(used_bg);
3449
3450                 if (!delalloc)
3451                         return used_bg;
3452
3453                 if (down_read_trylock(&used_bg->data_rwsem))
3454                         return used_bg;
3455
3456                 spin_unlock(&cluster->refill_lock);
3457
3458                 /* We should only have one-level nested. */
3459                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3460
3461                 spin_lock(&cluster->refill_lock);
3462                 if (used_bg == cluster->block_group)
3463                         return used_bg;
3464
3465                 up_read(&used_bg->data_rwsem);
3466                 btrfs_put_block_group(used_bg);
3467         }
3468 }
3469
3470 static inline void
3471 btrfs_release_block_group(struct btrfs_block_group *cache,
3472                          int delalloc)
3473 {
3474         if (delalloc)
3475                 up_read(&cache->data_rwsem);
3476         btrfs_put_block_group(cache);
3477 }
3478
3479 enum btrfs_extent_allocation_policy {
3480         BTRFS_EXTENT_ALLOC_CLUSTERED,
3481         BTRFS_EXTENT_ALLOC_ZONED,
3482 };
3483
3484 /*
3485  * Structure used internally for find_free_extent() function.  Wraps needed
3486  * parameters.
3487  */
3488 struct find_free_extent_ctl {
3489         /* Basic allocation info */
3490         u64 num_bytes;
3491         u64 empty_size;
3492         u64 flags;
3493         int delalloc;
3494
3495         /* Where to start the search inside the bg */
3496         u64 search_start;
3497
3498         /* For clustered allocation */
3499         u64 empty_cluster;
3500         struct btrfs_free_cluster *last_ptr;
3501         bool use_cluster;
3502
3503         bool have_caching_bg;
3504         bool orig_have_caching_bg;
3505
3506         /* Allocation is called for tree-log */
3507         bool for_treelog;
3508
3509         /* RAID index, converted from flags */
3510         int index;
3511
3512         /*
3513          * Current loop number, check find_free_extent_update_loop() for details
3514          */
3515         int loop;
3516
3517         /*
3518          * Whether we're refilling a cluster, if true we need to re-search
3519          * current block group but don't try to refill the cluster again.
3520          */
3521         bool retry_clustered;
3522
3523         /*
3524          * Whether we're updating free space cache, if true we need to re-search
3525          * current block group but don't try updating free space cache again.
3526          */
3527         bool retry_unclustered;
3528
3529         /* If current block group is cached */
3530         int cached;
3531
3532         /* Max contiguous hole found */
3533         u64 max_extent_size;
3534
3535         /* Total free space from free space cache, not always contiguous */
3536         u64 total_free_space;
3537
3538         /* Found result */
3539         u64 found_offset;
3540
3541         /* Hint where to start looking for an empty space */
3542         u64 hint_byte;
3543
3544         /* Allocation policy */
3545         enum btrfs_extent_allocation_policy policy;
3546 };
3547
3548
3549 /*
3550  * Helper function for find_free_extent().
3551  *
3552  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3553  * Return -EAGAIN to inform caller that we need to re-search this block group
3554  * Return >0 to inform caller that we find nothing
3555  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3556  */
3557 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3558                                       struct find_free_extent_ctl *ffe_ctl,
3559                                       struct btrfs_block_group **cluster_bg_ret)
3560 {
3561         struct btrfs_block_group *cluster_bg;
3562         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3563         u64 aligned_cluster;
3564         u64 offset;
3565         int ret;
3566
3567         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3568         if (!cluster_bg)
3569                 goto refill_cluster;
3570         if (cluster_bg != bg && (cluster_bg->ro ||
3571             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3572                 goto release_cluster;
3573
3574         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3575                         ffe_ctl->num_bytes, cluster_bg->start,
3576                         &ffe_ctl->max_extent_size);
3577         if (offset) {
3578                 /* We have a block, we're done */
3579                 spin_unlock(&last_ptr->refill_lock);
3580                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3581                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3582                 *cluster_bg_ret = cluster_bg;
3583                 ffe_ctl->found_offset = offset;
3584                 return 0;
3585         }
3586         WARN_ON(last_ptr->block_group != cluster_bg);
3587
3588 release_cluster:
3589         /*
3590          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3591          * lets just skip it and let the allocator find whatever block it can
3592          * find. If we reach this point, we will have tried the cluster
3593          * allocator plenty of times and not have found anything, so we are
3594          * likely way too fragmented for the clustering stuff to find anything.
3595          *
3596          * However, if the cluster is taken from the current block group,
3597          * release the cluster first, so that we stand a better chance of
3598          * succeeding in the unclustered allocation.
3599          */
3600         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3601                 spin_unlock(&last_ptr->refill_lock);
3602                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3603                 return -ENOENT;
3604         }
3605
3606         /* This cluster didn't work out, free it and start over */
3607         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3608
3609         if (cluster_bg != bg)
3610                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3611
3612 refill_cluster:
3613         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3614                 spin_unlock(&last_ptr->refill_lock);
3615                 return -ENOENT;
3616         }
3617
3618         aligned_cluster = max_t(u64,
3619                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3620                         bg->full_stripe_len);
3621         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3622                         ffe_ctl->num_bytes, aligned_cluster);
3623         if (ret == 0) {
3624                 /* Now pull our allocation out of this cluster */
3625                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3626                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3627                                 &ffe_ctl->max_extent_size);
3628                 if (offset) {
3629                         /* We found one, proceed */
3630                         spin_unlock(&last_ptr->refill_lock);
3631                         trace_btrfs_reserve_extent_cluster(bg,
3632                                         ffe_ctl->search_start,
3633                                         ffe_ctl->num_bytes);
3634                         ffe_ctl->found_offset = offset;
3635                         return 0;
3636                 }
3637         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3638                    !ffe_ctl->retry_clustered) {
3639                 spin_unlock(&last_ptr->refill_lock);
3640
3641                 ffe_ctl->retry_clustered = true;
3642                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3643                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3644                 return -EAGAIN;
3645         }
3646         /*
3647          * At this point we either didn't find a cluster or we weren't able to
3648          * allocate a block from our cluster.  Free the cluster we've been
3649          * trying to use, and go to the next block group.
3650          */
3651         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3652         spin_unlock(&last_ptr->refill_lock);
3653         return 1;
3654 }
3655
3656 /*
3657  * Return >0 to inform caller that we find nothing
3658  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3659  * Return -EAGAIN to inform caller that we need to re-search this block group
3660  */
3661 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3662                                         struct find_free_extent_ctl *ffe_ctl)
3663 {
3664         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3665         u64 offset;
3666
3667         /*
3668          * We are doing an unclustered allocation, set the fragmented flag so
3669          * we don't bother trying to setup a cluster again until we get more
3670          * space.
3671          */
3672         if (unlikely(last_ptr)) {
3673                 spin_lock(&last_ptr->lock);
3674                 last_ptr->fragmented = 1;
3675                 spin_unlock(&last_ptr->lock);
3676         }
3677         if (ffe_ctl->cached) {
3678                 struct btrfs_free_space_ctl *free_space_ctl;
3679
3680                 free_space_ctl = bg->free_space_ctl;
3681                 spin_lock(&free_space_ctl->tree_lock);
3682                 if (free_space_ctl->free_space <
3683                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3684                     ffe_ctl->empty_size) {
3685                         ffe_ctl->total_free_space = max_t(u64,
3686                                         ffe_ctl->total_free_space,
3687                                         free_space_ctl->free_space);
3688                         spin_unlock(&free_space_ctl->tree_lock);
3689                         return 1;
3690                 }
3691                 spin_unlock(&free_space_ctl->tree_lock);
3692         }
3693
3694         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3695                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3696                         &ffe_ctl->max_extent_size);
3697
3698         /*
3699          * If we didn't find a chunk, and we haven't failed on this block group
3700          * before, and this block group is in the middle of caching and we are
3701          * ok with waiting, then go ahead and wait for progress to be made, and
3702          * set @retry_unclustered to true.
3703          *
3704          * If @retry_unclustered is true then we've already waited on this
3705          * block group once and should move on to the next block group.
3706          */
3707         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3708             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3709                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3710                                                       ffe_ctl->empty_size);
3711                 ffe_ctl->retry_unclustered = true;
3712                 return -EAGAIN;
3713         } else if (!offset) {
3714                 return 1;
3715         }
3716         ffe_ctl->found_offset = offset;
3717         return 0;
3718 }
3719
3720 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3721                                    struct find_free_extent_ctl *ffe_ctl,
3722                                    struct btrfs_block_group **bg_ret)
3723 {
3724         int ret;
3725
3726         /* We want to try and use the cluster allocator, so lets look there */
3727         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3728                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3729                 if (ret >= 0 || ret == -EAGAIN)
3730                         return ret;
3731                 /* ret == -ENOENT case falls through */
3732         }
3733
3734         return find_free_extent_unclustered(block_group, ffe_ctl);
3735 }
3736
3737 /*
3738  * Tree-log block group locking
3739  * ============================
3740  *
3741  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3742  * indicates the starting address of a block group, which is reserved only
3743  * for tree-log metadata.
3744  *
3745  * Lock nesting
3746  * ============
3747  *
3748  * space_info::lock
3749  *   block_group::lock
3750  *     fs_info::treelog_bg_lock
3751  */
3752
3753 /*
3754  * Simple allocator for sequential-only block group. It only allows sequential
3755  * allocation. No need to play with trees. This function also reserves the
3756  * bytes as in btrfs_add_reserved_bytes.
3757  */
3758 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3759                                struct find_free_extent_ctl *ffe_ctl,
3760                                struct btrfs_block_group **bg_ret)
3761 {
3762         struct btrfs_fs_info *fs_info = block_group->fs_info;
3763         struct btrfs_space_info *space_info = block_group->space_info;
3764         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3765         u64 start = block_group->start;
3766         u64 num_bytes = ffe_ctl->num_bytes;
3767         u64 avail;
3768         u64 bytenr = block_group->start;
3769         u64 log_bytenr;
3770         int ret = 0;
3771         bool skip;
3772
3773         ASSERT(btrfs_is_zoned(block_group->fs_info));
3774
3775         /*
3776          * Do not allow non-tree-log blocks in the dedicated tree-log block
3777          * group, and vice versa.
3778          */
3779         spin_lock(&fs_info->treelog_bg_lock);
3780         log_bytenr = fs_info->treelog_bg;
3781         skip = log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3782                               (!ffe_ctl->for_treelog && bytenr == log_bytenr));
3783         spin_unlock(&fs_info->treelog_bg_lock);
3784         if (skip)
3785                 return 1;
3786
3787         spin_lock(&space_info->lock);
3788         spin_lock(&block_group->lock);
3789         spin_lock(&fs_info->treelog_bg_lock);
3790
3791         ASSERT(!ffe_ctl->for_treelog ||
3792                block_group->start == fs_info->treelog_bg ||
3793                fs_info->treelog_bg == 0);
3794
3795         if (block_group->ro) {
3796                 ret = 1;
3797                 goto out;
3798         }
3799
3800         /*
3801          * Do not allow currently using block group to be tree-log dedicated
3802          * block group.
3803          */
3804         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3805             (block_group->used || block_group->reserved)) {
3806                 ret = 1;
3807                 goto out;
3808         }
3809
3810         avail = block_group->length - block_group->alloc_offset;
3811         if (avail < num_bytes) {
3812                 if (ffe_ctl->max_extent_size < avail) {
3813                         /*
3814                          * With sequential allocator, free space is always
3815                          * contiguous
3816                          */
3817                         ffe_ctl->max_extent_size = avail;
3818                         ffe_ctl->total_free_space = avail;
3819                 }
3820                 ret = 1;
3821                 goto out;
3822         }
3823
3824         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3825                 fs_info->treelog_bg = block_group->start;
3826
3827         ffe_ctl->found_offset = start + block_group->alloc_offset;
3828         block_group->alloc_offset += num_bytes;
3829         spin_lock(&ctl->tree_lock);
3830         ctl->free_space -= num_bytes;
3831         spin_unlock(&ctl->tree_lock);
3832
3833         /*
3834          * We do not check if found_offset is aligned to stripesize. The
3835          * address is anyway rewritten when using zone append writing.
3836          */
3837
3838         ffe_ctl->search_start = ffe_ctl->found_offset;
3839
3840 out:
3841         if (ret && ffe_ctl->for_treelog)
3842                 fs_info->treelog_bg = 0;
3843         spin_unlock(&fs_info->treelog_bg_lock);
3844         spin_unlock(&block_group->lock);
3845         spin_unlock(&space_info->lock);
3846         return ret;
3847 }
3848
3849 static int do_allocation(struct btrfs_block_group *block_group,
3850                          struct find_free_extent_ctl *ffe_ctl,
3851                          struct btrfs_block_group **bg_ret)
3852 {
3853         switch (ffe_ctl->policy) {
3854         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3855                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3856         case BTRFS_EXTENT_ALLOC_ZONED:
3857                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3858         default:
3859                 BUG();
3860         }
3861 }
3862
3863 static void release_block_group(struct btrfs_block_group *block_group,
3864                                 struct find_free_extent_ctl *ffe_ctl,
3865                                 int delalloc)
3866 {
3867         switch (ffe_ctl->policy) {
3868         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3869                 ffe_ctl->retry_clustered = false;
3870                 ffe_ctl->retry_unclustered = false;
3871                 break;
3872         case BTRFS_EXTENT_ALLOC_ZONED:
3873                 /* Nothing to do */
3874                 break;
3875         default:
3876                 BUG();
3877         }
3878
3879         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3880                ffe_ctl->index);
3881         btrfs_release_block_group(block_group, delalloc);
3882 }
3883
3884 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3885                                    struct btrfs_key *ins)
3886 {
3887         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3888
3889         if (!ffe_ctl->use_cluster && last_ptr) {
3890                 spin_lock(&last_ptr->lock);
3891                 last_ptr->window_start = ins->objectid;
3892                 spin_unlock(&last_ptr->lock);
3893         }
3894 }
3895
3896 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3897                          struct btrfs_key *ins)
3898 {
3899         switch (ffe_ctl->policy) {
3900         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3901                 found_extent_clustered(ffe_ctl, ins);
3902                 break;
3903         case BTRFS_EXTENT_ALLOC_ZONED:
3904                 /* Nothing to do */
3905                 break;
3906         default:
3907                 BUG();
3908         }
3909 }
3910
3911 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
3912 {
3913         switch (ffe_ctl->policy) {
3914         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3915                 /*
3916                  * If we can't allocate a new chunk we've already looped through
3917                  * at least once, move on to the NO_EMPTY_SIZE case.
3918                  */
3919                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
3920                 return 0;
3921         case BTRFS_EXTENT_ALLOC_ZONED:
3922                 /* Give up here */
3923                 return -ENOSPC;
3924         default:
3925                 BUG();
3926         }
3927 }
3928
3929 /*
3930  * Return >0 means caller needs to re-search for free extent
3931  * Return 0 means we have the needed free extent.
3932  * Return <0 means we failed to locate any free extent.
3933  */
3934 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3935                                         struct btrfs_key *ins,
3936                                         struct find_free_extent_ctl *ffe_ctl,
3937                                         bool full_search)
3938 {
3939         struct btrfs_root *root = fs_info->extent_root;
3940         int ret;
3941
3942         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3943             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3944                 ffe_ctl->orig_have_caching_bg = true;
3945
3946         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
3947             ffe_ctl->have_caching_bg)
3948                 return 1;
3949
3950         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
3951                 return 1;
3952
3953         if (ins->objectid) {
3954                 found_extent(ffe_ctl, ins);
3955                 return 0;
3956         }
3957
3958         /*
3959          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
3960          *                      caching kthreads as we move along
3961          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
3962          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
3963          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
3964          *                     again
3965          */
3966         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
3967                 ffe_ctl->index = 0;
3968                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
3969                         /*
3970                          * We want to skip the LOOP_CACHING_WAIT step if we
3971                          * don't have any uncached bgs and we've already done a
3972                          * full search through.
3973                          */
3974                         if (ffe_ctl->orig_have_caching_bg || !full_search)
3975                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
3976                         else
3977                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
3978                 } else {
3979                         ffe_ctl->loop++;
3980                 }
3981
3982                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
3983                         struct btrfs_trans_handle *trans;
3984                         int exist = 0;
3985
3986                         trans = current->journal_info;
3987                         if (trans)
3988                                 exist = 1;
3989                         else
3990                                 trans = btrfs_join_transaction(root);
3991
3992                         if (IS_ERR(trans)) {
3993                                 ret = PTR_ERR(trans);
3994                                 return ret;
3995                         }
3996
3997                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
3998                                                 CHUNK_ALLOC_FORCE);
3999
4000                         /* Do not bail out on ENOSPC since we can do more. */
4001                         if (ret == -ENOSPC)
4002                                 ret = chunk_allocation_failed(ffe_ctl);
4003                         else if (ret < 0)
4004                                 btrfs_abort_transaction(trans, ret);
4005                         else
4006                                 ret = 0;
4007                         if (!exist)
4008                                 btrfs_end_transaction(trans);
4009                         if (ret)
4010                                 return ret;
4011                 }
4012
4013                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4014                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4015                                 return -ENOSPC;
4016
4017                         /*
4018                          * Don't loop again if we already have no empty_size and
4019                          * no empty_cluster.
4020                          */
4021                         if (ffe_ctl->empty_size == 0 &&
4022                             ffe_ctl->empty_cluster == 0)
4023                                 return -ENOSPC;
4024                         ffe_ctl->empty_size = 0;
4025                         ffe_ctl->empty_cluster = 0;
4026                 }
4027                 return 1;
4028         }
4029         return -ENOSPC;
4030 }
4031
4032 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4033                                         struct find_free_extent_ctl *ffe_ctl,
4034                                         struct btrfs_space_info *space_info,
4035                                         struct btrfs_key *ins)
4036 {
4037         /*
4038          * If our free space is heavily fragmented we may not be able to make
4039          * big contiguous allocations, so instead of doing the expensive search
4040          * for free space, simply return ENOSPC with our max_extent_size so we
4041          * can go ahead and search for a more manageable chunk.
4042          *
4043          * If our max_extent_size is large enough for our allocation simply
4044          * disable clustering since we will likely not be able to find enough
4045          * space to create a cluster and induce latency trying.
4046          */
4047         if (space_info->max_extent_size) {
4048                 spin_lock(&space_info->lock);
4049                 if (space_info->max_extent_size &&
4050                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4051                         ins->offset = space_info->max_extent_size;
4052                         spin_unlock(&space_info->lock);
4053                         return -ENOSPC;
4054                 } else if (space_info->max_extent_size) {
4055                         ffe_ctl->use_cluster = false;
4056                 }
4057                 spin_unlock(&space_info->lock);
4058         }
4059
4060         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4061                                                &ffe_ctl->empty_cluster);
4062         if (ffe_ctl->last_ptr) {
4063                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4064
4065                 spin_lock(&last_ptr->lock);
4066                 if (last_ptr->block_group)
4067                         ffe_ctl->hint_byte = last_ptr->window_start;
4068                 if (last_ptr->fragmented) {
4069                         /*
4070                          * We still set window_start so we can keep track of the
4071                          * last place we found an allocation to try and save
4072                          * some time.
4073                          */
4074                         ffe_ctl->hint_byte = last_ptr->window_start;
4075                         ffe_ctl->use_cluster = false;
4076                 }
4077                 spin_unlock(&last_ptr->lock);
4078         }
4079
4080         return 0;
4081 }
4082
4083 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4084                               struct find_free_extent_ctl *ffe_ctl,
4085                               struct btrfs_space_info *space_info,
4086                               struct btrfs_key *ins)
4087 {
4088         switch (ffe_ctl->policy) {
4089         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4090                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4091                                                     space_info, ins);
4092         case BTRFS_EXTENT_ALLOC_ZONED:
4093                 if (ffe_ctl->for_treelog) {
4094                         spin_lock(&fs_info->treelog_bg_lock);
4095                         if (fs_info->treelog_bg)
4096                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4097                         spin_unlock(&fs_info->treelog_bg_lock);
4098                 }
4099                 return 0;
4100         default:
4101                 BUG();
4102         }
4103 }
4104
4105 /*
4106  * walks the btree of allocated extents and find a hole of a given size.
4107  * The key ins is changed to record the hole:
4108  * ins->objectid == start position
4109  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4110  * ins->offset == the size of the hole.
4111  * Any available blocks before search_start are skipped.
4112  *
4113  * If there is no suitable free space, we will record the max size of
4114  * the free space extent currently.
4115  *
4116  * The overall logic and call chain:
4117  *
4118  * find_free_extent()
4119  * |- Iterate through all block groups
4120  * |  |- Get a valid block group
4121  * |  |- Try to do clustered allocation in that block group
4122  * |  |- Try to do unclustered allocation in that block group
4123  * |  |- Check if the result is valid
4124  * |  |  |- If valid, then exit
4125  * |  |- Jump to next block group
4126  * |
4127  * |- Push harder to find free extents
4128  *    |- If not found, re-iterate all block groups
4129  */
4130 static noinline int find_free_extent(struct btrfs_root *root,
4131                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
4132                                 u64 hint_byte_orig, struct btrfs_key *ins,
4133                                 u64 flags, int delalloc)
4134 {
4135         struct btrfs_fs_info *fs_info = root->fs_info;
4136         int ret = 0;
4137         int cache_block_group_error = 0;
4138         struct btrfs_block_group *block_group = NULL;
4139         struct find_free_extent_ctl ffe_ctl = {0};
4140         struct btrfs_space_info *space_info;
4141         bool full_search = false;
4142         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4143
4144         WARN_ON(num_bytes < fs_info->sectorsize);
4145
4146         ffe_ctl.num_bytes = num_bytes;
4147         ffe_ctl.empty_size = empty_size;
4148         ffe_ctl.flags = flags;
4149         ffe_ctl.search_start = 0;
4150         ffe_ctl.delalloc = delalloc;
4151         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
4152         ffe_ctl.have_caching_bg = false;
4153         ffe_ctl.orig_have_caching_bg = false;
4154         ffe_ctl.found_offset = 0;
4155         ffe_ctl.hint_byte = hint_byte_orig;
4156         ffe_ctl.for_treelog = for_treelog;
4157         ffe_ctl.policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4158
4159         /* For clustered allocation */
4160         ffe_ctl.retry_clustered = false;
4161         ffe_ctl.retry_unclustered = false;
4162         ffe_ctl.last_ptr = NULL;
4163         ffe_ctl.use_cluster = true;
4164
4165         if (btrfs_is_zoned(fs_info))
4166                 ffe_ctl.policy = BTRFS_EXTENT_ALLOC_ZONED;
4167
4168         ins->type = BTRFS_EXTENT_ITEM_KEY;
4169         ins->objectid = 0;
4170         ins->offset = 0;
4171
4172         trace_find_free_extent(root, num_bytes, empty_size, flags);
4173
4174         space_info = btrfs_find_space_info(fs_info, flags);
4175         if (!space_info) {
4176                 btrfs_err(fs_info, "No space info for %llu", flags);
4177                 return -ENOSPC;
4178         }
4179
4180         ret = prepare_allocation(fs_info, &ffe_ctl, space_info, ins);
4181         if (ret < 0)
4182                 return ret;
4183
4184         ffe_ctl.search_start = max(ffe_ctl.search_start,
4185                                    first_logical_byte(fs_info, 0));
4186         ffe_ctl.search_start = max(ffe_ctl.search_start, ffe_ctl.hint_byte);
4187         if (ffe_ctl.search_start == ffe_ctl.hint_byte) {
4188                 block_group = btrfs_lookup_block_group(fs_info,
4189                                                        ffe_ctl.search_start);
4190                 /*
4191                  * we don't want to use the block group if it doesn't match our
4192                  * allocation bits, or if its not cached.
4193                  *
4194                  * However if we are re-searching with an ideal block group
4195                  * picked out then we don't care that the block group is cached.
4196                  */
4197                 if (block_group && block_group_bits(block_group, flags) &&
4198                     block_group->cached != BTRFS_CACHE_NO) {
4199                         down_read(&space_info->groups_sem);
4200                         if (list_empty(&block_group->list) ||
4201                             block_group->ro) {
4202                                 /*
4203                                  * someone is removing this block group,
4204                                  * we can't jump into the have_block_group
4205                                  * target because our list pointers are not
4206                                  * valid
4207                                  */
4208                                 btrfs_put_block_group(block_group);
4209                                 up_read(&space_info->groups_sem);
4210                         } else {
4211                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
4212                                                 block_group->flags);
4213                                 btrfs_lock_block_group(block_group, delalloc);
4214                                 goto have_block_group;
4215                         }
4216                 } else if (block_group) {
4217                         btrfs_put_block_group(block_group);
4218                 }
4219         }
4220 search:
4221         ffe_ctl.have_caching_bg = false;
4222         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
4223             ffe_ctl.index == 0)
4224                 full_search = true;
4225         down_read(&space_info->groups_sem);
4226         list_for_each_entry(block_group,
4227                             &space_info->block_groups[ffe_ctl.index], list) {
4228                 struct btrfs_block_group *bg_ret;
4229
4230                 /* If the block group is read-only, we can skip it entirely. */
4231                 if (unlikely(block_group->ro)) {
4232                         if (for_treelog)
4233                                 btrfs_clear_treelog_bg(block_group);
4234                         continue;
4235                 }
4236
4237                 btrfs_grab_block_group(block_group, delalloc);
4238                 ffe_ctl.search_start = block_group->start;
4239
4240                 /*
4241                  * this can happen if we end up cycling through all the
4242                  * raid types, but we want to make sure we only allocate
4243                  * for the proper type.
4244                  */
4245                 if (!block_group_bits(block_group, flags)) {
4246                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4247                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4248                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4249                                 BTRFS_BLOCK_GROUP_RAID10;
4250
4251                         /*
4252                          * if they asked for extra copies and this block group
4253                          * doesn't provide them, bail.  This does allow us to
4254                          * fill raid0 from raid1.
4255                          */
4256                         if ((flags & extra) && !(block_group->flags & extra))
4257                                 goto loop;
4258
4259                         /*
4260                          * This block group has different flags than we want.
4261                          * It's possible that we have MIXED_GROUP flag but no
4262                          * block group is mixed.  Just skip such block group.
4263                          */
4264                         btrfs_release_block_group(block_group, delalloc);
4265                         continue;
4266                 }
4267
4268 have_block_group:
4269                 ffe_ctl.cached = btrfs_block_group_done(block_group);
4270                 if (unlikely(!ffe_ctl.cached)) {
4271                         ffe_ctl.have_caching_bg = true;
4272                         ret = btrfs_cache_block_group(block_group, 0);
4273
4274                         /*
4275                          * If we get ENOMEM here or something else we want to
4276                          * try other block groups, because it may not be fatal.
4277                          * However if we can't find anything else we need to
4278                          * save our return here so that we return the actual
4279                          * error that caused problems, not ENOSPC.
4280                          */
4281                         if (ret < 0) {
4282                                 if (!cache_block_group_error)
4283                                         cache_block_group_error = ret;
4284                                 ret = 0;
4285                                 goto loop;
4286                         }
4287                         ret = 0;
4288                 }
4289
4290                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4291                         goto loop;
4292
4293                 bg_ret = NULL;
4294                 ret = do_allocation(block_group, &ffe_ctl, &bg_ret);
4295                 if (ret == 0) {
4296                         if (bg_ret && bg_ret != block_group) {
4297                                 btrfs_release_block_group(block_group, delalloc);
4298                                 block_group = bg_ret;
4299                         }
4300                 } else if (ret == -EAGAIN) {
4301                         goto have_block_group;
4302                 } else if (ret > 0) {
4303                         goto loop;
4304                 }
4305
4306                 /* Checks */
4307                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
4308                                              fs_info->stripesize);
4309
4310                 /* move on to the next group */
4311                 if (ffe_ctl.search_start + num_bytes >
4312                     block_group->start + block_group->length) {
4313                         btrfs_add_free_space_unused(block_group,
4314                                             ffe_ctl.found_offset, num_bytes);
4315                         goto loop;
4316                 }
4317
4318                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
4319                         btrfs_add_free_space_unused(block_group,
4320                                         ffe_ctl.found_offset,
4321                                         ffe_ctl.search_start - ffe_ctl.found_offset);
4322
4323                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
4324                                 num_bytes, delalloc);
4325                 if (ret == -EAGAIN) {
4326                         btrfs_add_free_space_unused(block_group,
4327                                         ffe_ctl.found_offset, num_bytes);
4328                         goto loop;
4329                 }
4330                 btrfs_inc_block_group_reservations(block_group);
4331
4332                 /* we are all good, lets return */
4333                 ins->objectid = ffe_ctl.search_start;
4334                 ins->offset = num_bytes;
4335
4336                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
4337                                            num_bytes);
4338                 btrfs_release_block_group(block_group, delalloc);
4339                 break;
4340 loop:
4341                 release_block_group(block_group, &ffe_ctl, delalloc);
4342                 cond_resched();
4343         }
4344         up_read(&space_info->groups_sem);
4345
4346         ret = find_free_extent_update_loop(fs_info, ins, &ffe_ctl, full_search);
4347         if (ret > 0)
4348                 goto search;
4349
4350         if (ret == -ENOSPC && !cache_block_group_error) {
4351                 /*
4352                  * Use ffe_ctl->total_free_space as fallback if we can't find
4353                  * any contiguous hole.
4354                  */
4355                 if (!ffe_ctl.max_extent_size)
4356                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
4357                 spin_lock(&space_info->lock);
4358                 space_info->max_extent_size = ffe_ctl.max_extent_size;
4359                 spin_unlock(&space_info->lock);
4360                 ins->offset = ffe_ctl.max_extent_size;
4361         } else if (ret == -ENOSPC) {
4362                 ret = cache_block_group_error;
4363         }
4364         return ret;
4365 }
4366
4367 /*
4368  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4369  *                        hole that is at least as big as @num_bytes.
4370  *
4371  * @root           -    The root that will contain this extent
4372  *
4373  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4374  *                      is used for accounting purposes. This value differs
4375  *                      from @num_bytes only in the case of compressed extents.
4376  *
4377  * @num_bytes      -    Number of bytes to allocate on-disk.
4378  *
4379  * @min_alloc_size -    Indicates the minimum amount of space that the
4380  *                      allocator should try to satisfy. In some cases
4381  *                      @num_bytes may be larger than what is required and if
4382  *                      the filesystem is fragmented then allocation fails.
4383  *                      However, the presence of @min_alloc_size gives a
4384  *                      chance to try and satisfy the smaller allocation.
4385  *
4386  * @empty_size     -    A hint that you plan on doing more COW. This is the
4387  *                      size in bytes the allocator should try to find free
4388  *                      next to the block it returns.  This is just a hint and
4389  *                      may be ignored by the allocator.
4390  *
4391  * @hint_byte      -    Hint to the allocator to start searching above the byte
4392  *                      address passed. It might be ignored.
4393  *
4394  * @ins            -    This key is modified to record the found hole. It will
4395  *                      have the following values:
4396  *                      ins->objectid == start position
4397  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4398  *                      ins->offset == the size of the hole.
4399  *
4400  * @is_data        -    Boolean flag indicating whether an extent is
4401  *                      allocated for data (true) or metadata (false)
4402  *
4403  * @delalloc       -    Boolean flag indicating whether this allocation is for
4404  *                      delalloc or not. If 'true' data_rwsem of block groups
4405  *                      is going to be acquired.
4406  *
4407  *
4408  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4409  * case -ENOSPC is returned then @ins->offset will contain the size of the
4410  * largest available hole the allocator managed to find.
4411  */
4412 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4413                          u64 num_bytes, u64 min_alloc_size,
4414                          u64 empty_size, u64 hint_byte,
4415                          struct btrfs_key *ins, int is_data, int delalloc)
4416 {
4417         struct btrfs_fs_info *fs_info = root->fs_info;
4418         bool final_tried = num_bytes == min_alloc_size;
4419         u64 flags;
4420         int ret;
4421         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4422
4423         flags = get_alloc_profile_by_root(root, is_data);
4424 again:
4425         WARN_ON(num_bytes < fs_info->sectorsize);
4426         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
4427                                hint_byte, ins, flags, delalloc);
4428         if (!ret && !is_data) {
4429                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4430         } else if (ret == -ENOSPC) {
4431                 if (!final_tried && ins->offset) {
4432                         num_bytes = min(num_bytes >> 1, ins->offset);
4433                         num_bytes = round_down(num_bytes,
4434                                                fs_info->sectorsize);
4435                         num_bytes = max(num_bytes, min_alloc_size);
4436                         ram_bytes = num_bytes;
4437                         if (num_bytes == min_alloc_size)
4438                                 final_tried = true;
4439                         goto again;
4440                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4441                         struct btrfs_space_info *sinfo;
4442
4443                         sinfo = btrfs_find_space_info(fs_info, flags);
4444                         btrfs_err(fs_info,
4445                         "allocation failed flags %llu, wanted %llu tree-log %d",
4446                                   flags, num_bytes, for_treelog);
4447                         if (sinfo)
4448                                 btrfs_dump_space_info(fs_info, sinfo,
4449                                                       num_bytes, 1);
4450                 }
4451         }
4452
4453         return ret;
4454 }
4455
4456 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4457                                u64 start, u64 len, int delalloc)
4458 {
4459         struct btrfs_block_group *cache;
4460
4461         cache = btrfs_lookup_block_group(fs_info, start);
4462         if (!cache) {
4463                 btrfs_err(fs_info, "Unable to find block group for %llu",
4464                           start);
4465                 return -ENOSPC;
4466         }
4467
4468         btrfs_add_free_space(cache, start, len);
4469         btrfs_free_reserved_bytes(cache, len, delalloc);
4470         trace_btrfs_reserved_extent_free(fs_info, start, len);
4471
4472         btrfs_put_block_group(cache);
4473         return 0;
4474 }
4475
4476 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4477                               u64 len)
4478 {
4479         struct btrfs_block_group *cache;
4480         int ret = 0;
4481
4482         cache = btrfs_lookup_block_group(trans->fs_info, start);
4483         if (!cache) {
4484                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4485                           start);
4486                 return -ENOSPC;
4487         }
4488
4489         ret = pin_down_extent(trans, cache, start, len, 1);
4490         btrfs_put_block_group(cache);
4491         return ret;
4492 }
4493
4494 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4495                                       u64 parent, u64 root_objectid,
4496                                       u64 flags, u64 owner, u64 offset,
4497                                       struct btrfs_key *ins, int ref_mod)
4498 {
4499         struct btrfs_fs_info *fs_info = trans->fs_info;
4500         int ret;
4501         struct btrfs_extent_item *extent_item;
4502         struct btrfs_extent_inline_ref *iref;
4503         struct btrfs_path *path;
4504         struct extent_buffer *leaf;
4505         int type;
4506         u32 size;
4507
4508         if (parent > 0)
4509                 type = BTRFS_SHARED_DATA_REF_KEY;
4510         else
4511                 type = BTRFS_EXTENT_DATA_REF_KEY;
4512
4513         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4514
4515         path = btrfs_alloc_path();
4516         if (!path)
4517                 return -ENOMEM;
4518
4519         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4520                                       ins, size);
4521         if (ret) {
4522                 btrfs_free_path(path);
4523                 return ret;
4524         }
4525
4526         leaf = path->nodes[0];
4527         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4528                                      struct btrfs_extent_item);
4529         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4530         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4531         btrfs_set_extent_flags(leaf, extent_item,
4532                                flags | BTRFS_EXTENT_FLAG_DATA);
4533
4534         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4535         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4536         if (parent > 0) {
4537                 struct btrfs_shared_data_ref *ref;
4538                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4539                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4540                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4541         } else {
4542                 struct btrfs_extent_data_ref *ref;
4543                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4544                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4545                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4546                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4547                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4548         }
4549
4550         btrfs_mark_buffer_dirty(path->nodes[0]);
4551         btrfs_free_path(path);
4552
4553         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
4554         if (ret)
4555                 return ret;
4556
4557         ret = btrfs_update_block_group(trans, ins->objectid, ins->offset, 1);
4558         if (ret) { /* -ENOENT, logic error */
4559                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4560                         ins->objectid, ins->offset);
4561                 BUG();
4562         }
4563         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
4564         return ret;
4565 }
4566
4567 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4568                                      struct btrfs_delayed_ref_node *node,
4569                                      struct btrfs_delayed_extent_op *extent_op)
4570 {
4571         struct btrfs_fs_info *fs_info = trans->fs_info;
4572         int ret;
4573         struct btrfs_extent_item *extent_item;
4574         struct btrfs_key extent_key;
4575         struct btrfs_tree_block_info *block_info;
4576         struct btrfs_extent_inline_ref *iref;
4577         struct btrfs_path *path;
4578         struct extent_buffer *leaf;
4579         struct btrfs_delayed_tree_ref *ref;
4580         u32 size = sizeof(*extent_item) + sizeof(*iref);
4581         u64 num_bytes;
4582         u64 flags = extent_op->flags_to_set;
4583         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4584
4585         ref = btrfs_delayed_node_to_tree_ref(node);
4586
4587         extent_key.objectid = node->bytenr;
4588         if (skinny_metadata) {
4589                 extent_key.offset = ref->level;
4590                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4591                 num_bytes = fs_info->nodesize;
4592         } else {
4593                 extent_key.offset = node->num_bytes;
4594                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4595                 size += sizeof(*block_info);
4596                 num_bytes = node->num_bytes;
4597         }
4598
4599         path = btrfs_alloc_path();
4600         if (!path)
4601                 return -ENOMEM;
4602
4603         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4604                                       &extent_key, size);
4605         if (ret) {
4606                 btrfs_free_path(path);
4607                 return ret;
4608         }
4609
4610         leaf = path->nodes[0];
4611         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4612                                      struct btrfs_extent_item);
4613         btrfs_set_extent_refs(leaf, extent_item, 1);
4614         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4615         btrfs_set_extent_flags(leaf, extent_item,
4616                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4617
4618         if (skinny_metadata) {
4619                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4620         } else {
4621                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4622                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4623                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4624                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4625         }
4626
4627         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4628                 btrfs_set_extent_inline_ref_type(leaf, iref,
4629                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4630                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4631         } else {
4632                 btrfs_set_extent_inline_ref_type(leaf, iref,
4633                                                  BTRFS_TREE_BLOCK_REF_KEY);
4634                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4635         }
4636
4637         btrfs_mark_buffer_dirty(leaf);
4638         btrfs_free_path(path);
4639
4640         ret = remove_from_free_space_tree(trans, extent_key.objectid,
4641                                           num_bytes);
4642         if (ret)
4643                 return ret;
4644
4645         ret = btrfs_update_block_group(trans, extent_key.objectid,
4646                                        fs_info->nodesize, 1);
4647         if (ret) { /* -ENOENT, logic error */
4648                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4649                         extent_key.objectid, extent_key.offset);
4650                 BUG();
4651         }
4652
4653         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
4654                                           fs_info->nodesize);
4655         return ret;
4656 }
4657
4658 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4659                                      struct btrfs_root *root, u64 owner,
4660                                      u64 offset, u64 ram_bytes,
4661                                      struct btrfs_key *ins)
4662 {
4663         struct btrfs_ref generic_ref = { 0 };
4664
4665         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4666
4667         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4668                                ins->objectid, ins->offset, 0);
4669         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
4670         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4671
4672         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4673 }
4674
4675 /*
4676  * this is used by the tree logging recovery code.  It records that
4677  * an extent has been allocated and makes sure to clear the free
4678  * space cache bits as well
4679  */
4680 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4681                                    u64 root_objectid, u64 owner, u64 offset,
4682                                    struct btrfs_key *ins)
4683 {
4684         struct btrfs_fs_info *fs_info = trans->fs_info;
4685         int ret;
4686         struct btrfs_block_group *block_group;
4687         struct btrfs_space_info *space_info;
4688
4689         /*
4690          * Mixed block groups will exclude before processing the log so we only
4691          * need to do the exclude dance if this fs isn't mixed.
4692          */
4693         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4694                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4695                                               ins->offset);
4696                 if (ret)
4697                         return ret;
4698         }
4699
4700         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4701         if (!block_group)
4702                 return -EINVAL;
4703
4704         space_info = block_group->space_info;
4705         spin_lock(&space_info->lock);
4706         spin_lock(&block_group->lock);
4707         space_info->bytes_reserved += ins->offset;
4708         block_group->reserved += ins->offset;
4709         spin_unlock(&block_group->lock);
4710         spin_unlock(&space_info->lock);
4711
4712         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4713                                          offset, ins, 1);
4714         if (ret)
4715                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4716         btrfs_put_block_group(block_group);
4717         return ret;
4718 }
4719
4720 static struct extent_buffer *
4721 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4722                       u64 bytenr, int level, u64 owner,
4723                       enum btrfs_lock_nesting nest)
4724 {
4725         struct btrfs_fs_info *fs_info = root->fs_info;
4726         struct extent_buffer *buf;
4727
4728         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4729         if (IS_ERR(buf))
4730                 return buf;
4731
4732         /*
4733          * Extra safety check in case the extent tree is corrupted and extent
4734          * allocator chooses to use a tree block which is already used and
4735          * locked.
4736          */
4737         if (buf->lock_owner == current->pid) {
4738                 btrfs_err_rl(fs_info,
4739 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4740                         buf->start, btrfs_header_owner(buf), current->pid);
4741                 free_extent_buffer(buf);
4742                 return ERR_PTR(-EUCLEAN);
4743         }
4744
4745         /*
4746          * This needs to stay, because we could allocate a freed block from an
4747          * old tree into a new tree, so we need to make sure this new block is
4748          * set to the appropriate level and owner.
4749          */
4750         btrfs_set_buffer_lockdep_class(owner, buf, level);
4751         __btrfs_tree_lock(buf, nest);
4752         btrfs_clean_tree_block(buf);
4753         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4754         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4755
4756         set_extent_buffer_uptodate(buf);
4757
4758         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4759         btrfs_set_header_level(buf, level);
4760         btrfs_set_header_bytenr(buf, buf->start);
4761         btrfs_set_header_generation(buf, trans->transid);
4762         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4763         btrfs_set_header_owner(buf, owner);
4764         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4765         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4766         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4767                 buf->log_index = root->log_transid % 2;
4768                 /*
4769                  * we allow two log transactions at a time, use different
4770                  * EXTENT bit to differentiate dirty pages.
4771                  */
4772                 if (buf->log_index == 0)
4773                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4774                                         buf->start + buf->len - 1, GFP_NOFS);
4775                 else
4776                         set_extent_new(&root->dirty_log_pages, buf->start,
4777                                         buf->start + buf->len - 1);
4778         } else {
4779                 buf->log_index = -1;
4780                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4781                          buf->start + buf->len - 1, GFP_NOFS);
4782         }
4783         trans->dirty = true;
4784         /* this returns a buffer locked for blocking */
4785         return buf;
4786 }
4787
4788 /*
4789  * finds a free extent and does all the dirty work required for allocation
4790  * returns the tree buffer or an ERR_PTR on error.
4791  */
4792 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4793                                              struct btrfs_root *root,
4794                                              u64 parent, u64 root_objectid,
4795                                              const struct btrfs_disk_key *key,
4796                                              int level, u64 hint,
4797                                              u64 empty_size,
4798                                              enum btrfs_lock_nesting nest)
4799 {
4800         struct btrfs_fs_info *fs_info = root->fs_info;
4801         struct btrfs_key ins;
4802         struct btrfs_block_rsv *block_rsv;
4803         struct extent_buffer *buf;
4804         struct btrfs_delayed_extent_op *extent_op;
4805         struct btrfs_ref generic_ref = { 0 };
4806         u64 flags = 0;
4807         int ret;
4808         u32 blocksize = fs_info->nodesize;
4809         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4810
4811 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4812         if (btrfs_is_testing(fs_info)) {
4813                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4814                                             level, root_objectid, nest);
4815                 if (!IS_ERR(buf))
4816                         root->alloc_bytenr += blocksize;
4817                 return buf;
4818         }
4819 #endif
4820
4821         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4822         if (IS_ERR(block_rsv))
4823                 return ERR_CAST(block_rsv);
4824
4825         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4826                                    empty_size, hint, &ins, 0, 0);
4827         if (ret)
4828                 goto out_unuse;
4829
4830         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4831                                     root_objectid, nest);
4832         if (IS_ERR(buf)) {
4833                 ret = PTR_ERR(buf);
4834                 goto out_free_reserved;
4835         }
4836
4837         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4838                 if (parent == 0)
4839                         parent = ins.objectid;
4840                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4841         } else
4842                 BUG_ON(parent > 0);
4843
4844         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4845                 extent_op = btrfs_alloc_delayed_extent_op();
4846                 if (!extent_op) {
4847                         ret = -ENOMEM;
4848                         goto out_free_buf;
4849                 }
4850                 if (key)
4851                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4852                 else
4853                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4854                 extent_op->flags_to_set = flags;
4855                 extent_op->update_key = skinny_metadata ? false : true;
4856                 extent_op->update_flags = true;
4857                 extent_op->is_data = false;
4858                 extent_op->level = level;
4859
4860                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4861                                        ins.objectid, ins.offset, parent);
4862                 generic_ref.real_root = root->root_key.objectid;
4863                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
4864                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4865                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
4866                 if (ret)
4867                         goto out_free_delayed;
4868         }
4869         return buf;
4870
4871 out_free_delayed:
4872         btrfs_free_delayed_extent_op(extent_op);
4873 out_free_buf:
4874         free_extent_buffer(buf);
4875 out_free_reserved:
4876         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4877 out_unuse:
4878         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4879         return ERR_PTR(ret);
4880 }
4881
4882 struct walk_control {
4883         u64 refs[BTRFS_MAX_LEVEL];
4884         u64 flags[BTRFS_MAX_LEVEL];
4885         struct btrfs_key update_progress;
4886         struct btrfs_key drop_progress;
4887         int drop_level;
4888         int stage;
4889         int level;
4890         int shared_level;
4891         int update_ref;
4892         int keep_locks;
4893         int reada_slot;
4894         int reada_count;
4895         int restarted;
4896 };
4897
4898 #define DROP_REFERENCE  1
4899 #define UPDATE_BACKREF  2
4900
4901 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
4902                                      struct btrfs_root *root,
4903                                      struct walk_control *wc,
4904                                      struct btrfs_path *path)
4905 {
4906         struct btrfs_fs_info *fs_info = root->fs_info;
4907         u64 bytenr;
4908         u64 generation;
4909         u64 refs;
4910         u64 flags;
4911         u32 nritems;
4912         struct btrfs_key key;
4913         struct extent_buffer *eb;
4914         int ret;
4915         int slot;
4916         int nread = 0;
4917
4918         if (path->slots[wc->level] < wc->reada_slot) {
4919                 wc->reada_count = wc->reada_count * 2 / 3;
4920                 wc->reada_count = max(wc->reada_count, 2);
4921         } else {
4922                 wc->reada_count = wc->reada_count * 3 / 2;
4923                 wc->reada_count = min_t(int, wc->reada_count,
4924                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
4925         }
4926
4927         eb = path->nodes[wc->level];
4928         nritems = btrfs_header_nritems(eb);
4929
4930         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
4931                 if (nread >= wc->reada_count)
4932                         break;
4933
4934                 cond_resched();
4935                 bytenr = btrfs_node_blockptr(eb, slot);
4936                 generation = btrfs_node_ptr_generation(eb, slot);
4937
4938                 if (slot == path->slots[wc->level])
4939                         goto reada;
4940
4941                 if (wc->stage == UPDATE_BACKREF &&
4942                     generation <= root->root_key.offset)
4943                         continue;
4944
4945                 /* We don't lock the tree block, it's OK to be racy here */
4946                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
4947                                                wc->level - 1, 1, &refs,
4948                                                &flags);
4949                 /* We don't care about errors in readahead. */
4950                 if (ret < 0)
4951                         continue;
4952                 BUG_ON(refs == 0);
4953
4954                 if (wc->stage == DROP_REFERENCE) {
4955                         if (refs == 1)
4956                                 goto reada;
4957
4958                         if (wc->level == 1 &&
4959                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4960                                 continue;
4961                         if (!wc->update_ref ||
4962                             generation <= root->root_key.offset)
4963                                 continue;
4964                         btrfs_node_key_to_cpu(eb, &key, slot);
4965                         ret = btrfs_comp_cpu_keys(&key,
4966                                                   &wc->update_progress);
4967                         if (ret < 0)
4968                                 continue;
4969                 } else {
4970                         if (wc->level == 1 &&
4971                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4972                                 continue;
4973                 }
4974 reada:
4975                 btrfs_readahead_node_child(eb, slot);
4976                 nread++;
4977         }
4978         wc->reada_slot = slot;
4979 }
4980
4981 /*
4982  * helper to process tree block while walking down the tree.
4983  *
4984  * when wc->stage == UPDATE_BACKREF, this function updates
4985  * back refs for pointers in the block.
4986  *
4987  * NOTE: return value 1 means we should stop walking down.
4988  */
4989 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
4990                                    struct btrfs_root *root,
4991                                    struct btrfs_path *path,
4992                                    struct walk_control *wc, int lookup_info)
4993 {
4994         struct btrfs_fs_info *fs_info = root->fs_info;
4995         int level = wc->level;
4996         struct extent_buffer *eb = path->nodes[level];
4997         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
4998         int ret;
4999
5000         if (wc->stage == UPDATE_BACKREF &&
5001             btrfs_header_owner(eb) != root->root_key.objectid)
5002                 return 1;
5003
5004         /*
5005          * when reference count of tree block is 1, it won't increase
5006          * again. once full backref flag is set, we never clear it.
5007          */
5008         if (lookup_info &&
5009             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5010              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5011                 BUG_ON(!path->locks[level]);
5012                 ret = btrfs_lookup_extent_info(trans, fs_info,
5013                                                eb->start, level, 1,
5014                                                &wc->refs[level],
5015                                                &wc->flags[level]);
5016                 BUG_ON(ret == -ENOMEM);
5017                 if (ret)
5018                         return ret;
5019                 BUG_ON(wc->refs[level] == 0);
5020         }
5021
5022         if (wc->stage == DROP_REFERENCE) {
5023                 if (wc->refs[level] > 1)
5024                         return 1;
5025
5026                 if (path->locks[level] && !wc->keep_locks) {
5027                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5028                         path->locks[level] = 0;
5029                 }
5030                 return 0;
5031         }
5032
5033         /* wc->stage == UPDATE_BACKREF */
5034         if (!(wc->flags[level] & flag)) {
5035                 BUG_ON(!path->locks[level]);
5036                 ret = btrfs_inc_ref(trans, root, eb, 1);
5037                 BUG_ON(ret); /* -ENOMEM */
5038                 ret = btrfs_dec_ref(trans, root, eb, 0);
5039                 BUG_ON(ret); /* -ENOMEM */
5040                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5041                                                   btrfs_header_level(eb), 0);
5042                 BUG_ON(ret); /* -ENOMEM */
5043                 wc->flags[level] |= flag;
5044         }
5045
5046         /*
5047          * the block is shared by multiple trees, so it's not good to
5048          * keep the tree lock
5049          */
5050         if (path->locks[level] && level > 0) {
5051                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5052                 path->locks[level] = 0;
5053         }
5054         return 0;
5055 }
5056
5057 /*
5058  * This is used to verify a ref exists for this root to deal with a bug where we
5059  * would have a drop_progress key that hadn't been updated properly.
5060  */
5061 static int check_ref_exists(struct btrfs_trans_handle *trans,
5062                             struct btrfs_root *root, u64 bytenr, u64 parent,
5063                             int level)
5064 {
5065         struct btrfs_path *path;
5066         struct btrfs_extent_inline_ref *iref;
5067         int ret;
5068
5069         path = btrfs_alloc_path();
5070         if (!path)
5071                 return -ENOMEM;
5072
5073         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5074                                     root->fs_info->nodesize, parent,
5075                                     root->root_key.objectid, level, 0);
5076         btrfs_free_path(path);
5077         if (ret == -ENOENT)
5078                 return 0;
5079         if (ret < 0)
5080                 return ret;
5081         return 1;
5082 }
5083
5084 /*
5085  * helper to process tree block pointer.
5086  *
5087  * when wc->stage == DROP_REFERENCE, this function checks
5088  * reference count of the block pointed to. if the block
5089  * is shared and we need update back refs for the subtree
5090  * rooted at the block, this function changes wc->stage to
5091  * UPDATE_BACKREF. if the block is shared and there is no
5092  * need to update back, this function drops the reference
5093  * to the block.
5094  *
5095  * NOTE: return value 1 means we should stop walking down.
5096  */
5097 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5098                                  struct btrfs_root *root,
5099                                  struct btrfs_path *path,
5100                                  struct walk_control *wc, int *lookup_info)
5101 {
5102         struct btrfs_fs_info *fs_info = root->fs_info;
5103         u64 bytenr;
5104         u64 generation;
5105         u64 parent;
5106         struct btrfs_key key;
5107         struct btrfs_key first_key;
5108         struct btrfs_ref ref = { 0 };
5109         struct extent_buffer *next;
5110         int level = wc->level;
5111         int reada = 0;
5112         int ret = 0;
5113         bool need_account = false;
5114
5115         generation = btrfs_node_ptr_generation(path->nodes[level],
5116                                                path->slots[level]);
5117         /*
5118          * if the lower level block was created before the snapshot
5119          * was created, we know there is no need to update back refs
5120          * for the subtree
5121          */
5122         if (wc->stage == UPDATE_BACKREF &&
5123             generation <= root->root_key.offset) {
5124                 *lookup_info = 1;
5125                 return 1;
5126         }
5127
5128         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5129         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
5130                               path->slots[level]);
5131
5132         next = find_extent_buffer(fs_info, bytenr);
5133         if (!next) {
5134                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5135                                 root->root_key.objectid, level - 1);
5136                 if (IS_ERR(next))
5137                         return PTR_ERR(next);
5138                 reada = 1;
5139         }
5140         btrfs_tree_lock(next);
5141
5142         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5143                                        &wc->refs[level - 1],
5144                                        &wc->flags[level - 1]);
5145         if (ret < 0)
5146                 goto out_unlock;
5147
5148         if (unlikely(wc->refs[level - 1] == 0)) {
5149                 btrfs_err(fs_info, "Missing references.");
5150                 ret = -EIO;
5151                 goto out_unlock;
5152         }
5153         *lookup_info = 0;
5154
5155         if (wc->stage == DROP_REFERENCE) {
5156                 if (wc->refs[level - 1] > 1) {
5157                         need_account = true;
5158                         if (level == 1 &&
5159                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5160                                 goto skip;
5161
5162                         if (!wc->update_ref ||
5163                             generation <= root->root_key.offset)
5164                                 goto skip;
5165
5166                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5167                                               path->slots[level]);
5168                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5169                         if (ret < 0)
5170                                 goto skip;
5171
5172                         wc->stage = UPDATE_BACKREF;
5173                         wc->shared_level = level - 1;
5174                 }
5175         } else {
5176                 if (level == 1 &&
5177                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5178                         goto skip;
5179         }
5180
5181         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5182                 btrfs_tree_unlock(next);
5183                 free_extent_buffer(next);
5184                 next = NULL;
5185                 *lookup_info = 1;
5186         }
5187
5188         if (!next) {
5189                 if (reada && level == 1)
5190                         reada_walk_down(trans, root, wc, path);
5191                 next = read_tree_block(fs_info, bytenr, root->root_key.objectid,
5192                                        generation, level - 1, &first_key);
5193                 if (IS_ERR(next)) {
5194                         return PTR_ERR(next);
5195                 } else if (!extent_buffer_uptodate(next)) {
5196                         free_extent_buffer(next);
5197                         return -EIO;
5198                 }
5199                 btrfs_tree_lock(next);
5200         }
5201
5202         level--;
5203         ASSERT(level == btrfs_header_level(next));
5204         if (level != btrfs_header_level(next)) {
5205                 btrfs_err(root->fs_info, "mismatched level");
5206                 ret = -EIO;
5207                 goto out_unlock;
5208         }
5209         path->nodes[level] = next;
5210         path->slots[level] = 0;
5211         path->locks[level] = BTRFS_WRITE_LOCK;
5212         wc->level = level;
5213         if (wc->level == 1)
5214                 wc->reada_slot = 0;
5215         return 0;
5216 skip:
5217         wc->refs[level - 1] = 0;
5218         wc->flags[level - 1] = 0;
5219         if (wc->stage == DROP_REFERENCE) {
5220                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5221                         parent = path->nodes[level]->start;
5222                 } else {
5223                         ASSERT(root->root_key.objectid ==
5224                                btrfs_header_owner(path->nodes[level]));
5225                         if (root->root_key.objectid !=
5226                             btrfs_header_owner(path->nodes[level])) {
5227                                 btrfs_err(root->fs_info,
5228                                                 "mismatched block owner");
5229                                 ret = -EIO;
5230                                 goto out_unlock;
5231                         }
5232                         parent = 0;
5233                 }
5234
5235                 /*
5236                  * If we had a drop_progress we need to verify the refs are set
5237                  * as expected.  If we find our ref then we know that from here
5238                  * on out everything should be correct, and we can clear the
5239                  * ->restarted flag.
5240                  */
5241                 if (wc->restarted) {
5242                         ret = check_ref_exists(trans, root, bytenr, parent,
5243                                                level - 1);
5244                         if (ret < 0)
5245                                 goto out_unlock;
5246                         if (ret == 0)
5247                                 goto no_delete;
5248                         ret = 0;
5249                         wc->restarted = 0;
5250                 }
5251
5252                 /*
5253                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5254                  * already accounted them at merge time (replace_path),
5255                  * thus we could skip expensive subtree trace here.
5256                  */
5257                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5258                     need_account) {
5259                         ret = btrfs_qgroup_trace_subtree(trans, next,
5260                                                          generation, level - 1);
5261                         if (ret) {
5262                                 btrfs_err_rl(fs_info,
5263                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5264                                              ret);
5265                         }
5266                 }
5267
5268                 /*
5269                  * We need to update the next key in our walk control so we can
5270                  * update the drop_progress key accordingly.  We don't care if
5271                  * find_next_key doesn't find a key because that means we're at
5272                  * the end and are going to clean up now.
5273                  */
5274                 wc->drop_level = level;
5275                 find_next_key(path, level, &wc->drop_progress);
5276
5277                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5278                                        fs_info->nodesize, parent);
5279                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
5280                 ret = btrfs_free_extent(trans, &ref);
5281                 if (ret)
5282                         goto out_unlock;
5283         }
5284 no_delete:
5285         *lookup_info = 1;
5286         ret = 1;
5287
5288 out_unlock:
5289         btrfs_tree_unlock(next);
5290         free_extent_buffer(next);
5291
5292         return ret;
5293 }
5294
5295 /*
5296  * helper to process tree block while walking up the tree.
5297  *
5298  * when wc->stage == DROP_REFERENCE, this function drops
5299  * reference count on the block.
5300  *
5301  * when wc->stage == UPDATE_BACKREF, this function changes
5302  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5303  * to UPDATE_BACKREF previously while processing the block.
5304  *
5305  * NOTE: return value 1 means we should stop walking up.
5306  */
5307 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5308                                  struct btrfs_root *root,
5309                                  struct btrfs_path *path,
5310                                  struct walk_control *wc)
5311 {
5312         struct btrfs_fs_info *fs_info = root->fs_info;
5313         int ret;
5314         int level = wc->level;
5315         struct extent_buffer *eb = path->nodes[level];
5316         u64 parent = 0;
5317
5318         if (wc->stage == UPDATE_BACKREF) {
5319                 BUG_ON(wc->shared_level < level);
5320                 if (level < wc->shared_level)
5321                         goto out;
5322
5323                 ret = find_next_key(path, level + 1, &wc->update_progress);
5324                 if (ret > 0)
5325                         wc->update_ref = 0;
5326
5327                 wc->stage = DROP_REFERENCE;
5328                 wc->shared_level = -1;
5329                 path->slots[level] = 0;
5330
5331                 /*
5332                  * check reference count again if the block isn't locked.
5333                  * we should start walking down the tree again if reference
5334                  * count is one.
5335                  */
5336                 if (!path->locks[level]) {
5337                         BUG_ON(level == 0);
5338                         btrfs_tree_lock(eb);
5339                         path->locks[level] = BTRFS_WRITE_LOCK;
5340
5341                         ret = btrfs_lookup_extent_info(trans, fs_info,
5342                                                        eb->start, level, 1,
5343                                                        &wc->refs[level],
5344                                                        &wc->flags[level]);
5345                         if (ret < 0) {
5346                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5347                                 path->locks[level] = 0;
5348                                 return ret;
5349                         }
5350                         BUG_ON(wc->refs[level] == 0);
5351                         if (wc->refs[level] == 1) {
5352                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5353                                 path->locks[level] = 0;
5354                                 return 1;
5355                         }
5356                 }
5357         }
5358
5359         /* wc->stage == DROP_REFERENCE */
5360         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5361
5362         if (wc->refs[level] == 1) {
5363                 if (level == 0) {
5364                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5365                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5366                         else
5367                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5368                         BUG_ON(ret); /* -ENOMEM */
5369                         if (is_fstree(root->root_key.objectid)) {
5370                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5371                                 if (ret) {
5372                                         btrfs_err_rl(fs_info,
5373         "error %d accounting leaf items, quota is out of sync, rescan required",
5374                                              ret);
5375                                 }
5376                         }
5377                 }
5378                 /* make block locked assertion in btrfs_clean_tree_block happy */
5379                 if (!path->locks[level] &&
5380                     btrfs_header_generation(eb) == trans->transid) {
5381                         btrfs_tree_lock(eb);
5382                         path->locks[level] = BTRFS_WRITE_LOCK;
5383                 }
5384                 btrfs_clean_tree_block(eb);
5385         }
5386
5387         if (eb == root->node) {
5388                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5389                         parent = eb->start;
5390                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5391                         goto owner_mismatch;
5392         } else {
5393                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5394                         parent = path->nodes[level + 1]->start;
5395                 else if (root->root_key.objectid !=
5396                          btrfs_header_owner(path->nodes[level + 1]))
5397                         goto owner_mismatch;
5398         }
5399
5400         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
5401 out:
5402         wc->refs[level] = 0;
5403         wc->flags[level] = 0;
5404         return 0;
5405
5406 owner_mismatch:
5407         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5408                      btrfs_header_owner(eb), root->root_key.objectid);
5409         return -EUCLEAN;
5410 }
5411
5412 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5413                                    struct btrfs_root *root,
5414                                    struct btrfs_path *path,
5415                                    struct walk_control *wc)
5416 {
5417         int level = wc->level;
5418         int lookup_info = 1;
5419         int ret;
5420
5421         while (level >= 0) {
5422                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5423                 if (ret > 0)
5424                         break;
5425
5426                 if (level == 0)
5427                         break;
5428
5429                 if (path->slots[level] >=
5430                     btrfs_header_nritems(path->nodes[level]))
5431                         break;
5432
5433                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5434                 if (ret > 0) {
5435                         path->slots[level]++;
5436                         continue;
5437                 } else if (ret < 0)
5438                         return ret;
5439                 level = wc->level;
5440         }
5441         return 0;
5442 }
5443
5444 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5445                                  struct btrfs_root *root,
5446                                  struct btrfs_path *path,
5447                                  struct walk_control *wc, int max_level)
5448 {
5449         int level = wc->level;
5450         int ret;
5451
5452         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5453         while (level < max_level && path->nodes[level]) {
5454                 wc->level = level;
5455                 if (path->slots[level] + 1 <
5456                     btrfs_header_nritems(path->nodes[level])) {
5457                         path->slots[level]++;
5458                         return 0;
5459                 } else {
5460                         ret = walk_up_proc(trans, root, path, wc);
5461                         if (ret > 0)
5462                                 return 0;
5463                         if (ret < 0)
5464                                 return ret;
5465
5466                         if (path->locks[level]) {
5467                                 btrfs_tree_unlock_rw(path->nodes[level],
5468                                                      path->locks[level]);
5469                                 path->locks[level] = 0;
5470                         }
5471                         free_extent_buffer(path->nodes[level]);
5472                         path->nodes[level] = NULL;
5473                         level++;
5474                 }
5475         }
5476         return 1;
5477 }
5478
5479 /*
5480  * drop a subvolume tree.
5481  *
5482  * this function traverses the tree freeing any blocks that only
5483  * referenced by the tree.
5484  *
5485  * when a shared tree block is found. this function decreases its
5486  * reference count by one. if update_ref is true, this function
5487  * also make sure backrefs for the shared block and all lower level
5488  * blocks are properly updated.
5489  *
5490  * If called with for_reloc == 0, may exit early with -EAGAIN
5491  */
5492 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5493 {
5494         struct btrfs_fs_info *fs_info = root->fs_info;
5495         struct btrfs_path *path;
5496         struct btrfs_trans_handle *trans;
5497         struct btrfs_root *tree_root = fs_info->tree_root;
5498         struct btrfs_root_item *root_item = &root->root_item;
5499         struct walk_control *wc;
5500         struct btrfs_key key;
5501         int err = 0;
5502         int ret;
5503         int level;
5504         bool root_dropped = false;
5505
5506         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5507
5508         path = btrfs_alloc_path();
5509         if (!path) {
5510                 err = -ENOMEM;
5511                 goto out;
5512         }
5513
5514         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5515         if (!wc) {
5516                 btrfs_free_path(path);
5517                 err = -ENOMEM;
5518                 goto out;
5519         }
5520
5521         /*
5522          * Use join to avoid potential EINTR from transaction start. See
5523          * wait_reserve_ticket and the whole reservation callchain.
5524          */
5525         if (for_reloc)
5526                 trans = btrfs_join_transaction(tree_root);
5527         else
5528                 trans = btrfs_start_transaction(tree_root, 0);
5529         if (IS_ERR(trans)) {
5530                 err = PTR_ERR(trans);
5531                 goto out_free;
5532         }
5533
5534         err = btrfs_run_delayed_items(trans);
5535         if (err)
5536                 goto out_end_trans;
5537
5538         /*
5539          * This will help us catch people modifying the fs tree while we're
5540          * dropping it.  It is unsafe to mess with the fs tree while it's being
5541          * dropped as we unlock the root node and parent nodes as we walk down
5542          * the tree, assuming nothing will change.  If something does change
5543          * then we'll have stale information and drop references to blocks we've
5544          * already dropped.
5545          */
5546         set_bit(BTRFS_ROOT_DELETING, &root->state);
5547         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5548                 level = btrfs_header_level(root->node);
5549                 path->nodes[level] = btrfs_lock_root_node(root);
5550                 path->slots[level] = 0;
5551                 path->locks[level] = BTRFS_WRITE_LOCK;
5552                 memset(&wc->update_progress, 0,
5553                        sizeof(wc->update_progress));
5554         } else {
5555                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5556                 memcpy(&wc->update_progress, &key,
5557                        sizeof(wc->update_progress));
5558
5559                 level = btrfs_root_drop_level(root_item);
5560                 BUG_ON(level == 0);
5561                 path->lowest_level = level;
5562                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5563                 path->lowest_level = 0;
5564                 if (ret < 0) {
5565                         err = ret;
5566                         goto out_end_trans;
5567                 }
5568                 WARN_ON(ret > 0);
5569
5570                 /*
5571                  * unlock our path, this is safe because only this
5572                  * function is allowed to delete this snapshot
5573                  */
5574                 btrfs_unlock_up_safe(path, 0);
5575
5576                 level = btrfs_header_level(root->node);
5577                 while (1) {
5578                         btrfs_tree_lock(path->nodes[level]);
5579                         path->locks[level] = BTRFS_WRITE_LOCK;
5580
5581                         ret = btrfs_lookup_extent_info(trans, fs_info,
5582                                                 path->nodes[level]->start,
5583                                                 level, 1, &wc->refs[level],
5584                                                 &wc->flags[level]);
5585                         if (ret < 0) {
5586                                 err = ret;
5587                                 goto out_end_trans;
5588                         }
5589                         BUG_ON(wc->refs[level] == 0);
5590
5591                         if (level == btrfs_root_drop_level(root_item))
5592                                 break;
5593
5594                         btrfs_tree_unlock(path->nodes[level]);
5595                         path->locks[level] = 0;
5596                         WARN_ON(wc->refs[level] != 1);
5597                         level--;
5598                 }
5599         }
5600
5601         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5602         wc->level = level;
5603         wc->shared_level = -1;
5604         wc->stage = DROP_REFERENCE;
5605         wc->update_ref = update_ref;
5606         wc->keep_locks = 0;
5607         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5608
5609         while (1) {
5610
5611                 ret = walk_down_tree(trans, root, path, wc);
5612                 if (ret < 0) {
5613                         err = ret;
5614                         break;
5615                 }
5616
5617                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5618                 if (ret < 0) {
5619                         err = ret;
5620                         break;
5621                 }
5622
5623                 if (ret > 0) {
5624                         BUG_ON(wc->stage != DROP_REFERENCE);
5625                         break;
5626                 }
5627
5628                 if (wc->stage == DROP_REFERENCE) {
5629                         wc->drop_level = wc->level;
5630                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5631                                               &wc->drop_progress,
5632                                               path->slots[wc->drop_level]);
5633                 }
5634                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5635                                       &wc->drop_progress);
5636                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5637
5638                 BUG_ON(wc->level == 0);
5639                 if (btrfs_should_end_transaction(trans) ||
5640                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5641                         ret = btrfs_update_root(trans, tree_root,
5642                                                 &root->root_key,
5643                                                 root_item);
5644                         if (ret) {
5645                                 btrfs_abort_transaction(trans, ret);
5646                                 err = ret;
5647                                 goto out_end_trans;
5648                         }
5649
5650                         btrfs_end_transaction_throttle(trans);
5651                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5652                                 btrfs_debug(fs_info,
5653                                             "drop snapshot early exit");
5654                                 err = -EAGAIN;
5655                                 goto out_free;
5656                         }
5657
5658                        /*
5659                         * Use join to avoid potential EINTR from transaction
5660                         * start. See wait_reserve_ticket and the whole
5661                         * reservation callchain.
5662                         */
5663                         if (for_reloc)
5664                                 trans = btrfs_join_transaction(tree_root);
5665                         else
5666                                 trans = btrfs_start_transaction(tree_root, 0);
5667                         if (IS_ERR(trans)) {
5668                                 err = PTR_ERR(trans);
5669                                 goto out_free;
5670                         }
5671                 }
5672         }
5673         btrfs_release_path(path);
5674         if (err)
5675                 goto out_end_trans;
5676
5677         ret = btrfs_del_root(trans, &root->root_key);
5678         if (ret) {
5679                 btrfs_abort_transaction(trans, ret);
5680                 err = ret;
5681                 goto out_end_trans;
5682         }
5683
5684         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5685                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5686                                       NULL, NULL);
5687                 if (ret < 0) {
5688                         btrfs_abort_transaction(trans, ret);
5689                         err = ret;
5690                         goto out_end_trans;
5691                 } else if (ret > 0) {
5692                         /* if we fail to delete the orphan item this time
5693                          * around, it'll get picked up the next time.
5694                          *
5695                          * The most common failure here is just -ENOENT.
5696                          */
5697                         btrfs_del_orphan_item(trans, tree_root,
5698                                               root->root_key.objectid);
5699                 }
5700         }
5701
5702         /*
5703          * This subvolume is going to be completely dropped, and won't be
5704          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5705          * commit transaction time.  So free it here manually.
5706          */
5707         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5708         btrfs_qgroup_free_meta_all_pertrans(root);
5709
5710         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5711                 btrfs_add_dropped_root(trans, root);
5712         else
5713                 btrfs_put_root(root);
5714         root_dropped = true;
5715 out_end_trans:
5716         btrfs_end_transaction_throttle(trans);
5717 out_free:
5718         kfree(wc);
5719         btrfs_free_path(path);
5720 out:
5721         /*
5722          * So if we need to stop dropping the snapshot for whatever reason we
5723          * need to make sure to add it back to the dead root list so that we
5724          * keep trying to do the work later.  This also cleans up roots if we
5725          * don't have it in the radix (like when we recover after a power fail
5726          * or unmount) so we don't leak memory.
5727          */
5728         if (!for_reloc && !root_dropped)
5729                 btrfs_add_dead_root(root);
5730         return err;
5731 }
5732
5733 /*
5734  * drop subtree rooted at tree block 'node'.
5735  *
5736  * NOTE: this function will unlock and release tree block 'node'
5737  * only used by relocation code
5738  */
5739 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5740                         struct btrfs_root *root,
5741                         struct extent_buffer *node,
5742                         struct extent_buffer *parent)
5743 {
5744         struct btrfs_fs_info *fs_info = root->fs_info;
5745         struct btrfs_path *path;
5746         struct walk_control *wc;
5747         int level;
5748         int parent_level;
5749         int ret = 0;
5750         int wret;
5751
5752         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5753
5754         path = btrfs_alloc_path();
5755         if (!path)
5756                 return -ENOMEM;
5757
5758         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5759         if (!wc) {
5760                 btrfs_free_path(path);
5761                 return -ENOMEM;
5762         }
5763
5764         btrfs_assert_tree_locked(parent);
5765         parent_level = btrfs_header_level(parent);
5766         atomic_inc(&parent->refs);
5767         path->nodes[parent_level] = parent;
5768         path->slots[parent_level] = btrfs_header_nritems(parent);
5769
5770         btrfs_assert_tree_locked(node);
5771         level = btrfs_header_level(node);
5772         path->nodes[level] = node;
5773         path->slots[level] = 0;
5774         path->locks[level] = BTRFS_WRITE_LOCK;
5775
5776         wc->refs[parent_level] = 1;
5777         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5778         wc->level = level;
5779         wc->shared_level = -1;
5780         wc->stage = DROP_REFERENCE;
5781         wc->update_ref = 0;
5782         wc->keep_locks = 1;
5783         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5784
5785         while (1) {
5786                 wret = walk_down_tree(trans, root, path, wc);
5787                 if (wret < 0) {
5788                         ret = wret;
5789                         break;
5790                 }
5791
5792                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5793                 if (wret < 0)
5794                         ret = wret;
5795                 if (wret != 0)
5796                         break;
5797         }
5798
5799         kfree(wc);
5800         btrfs_free_path(path);
5801         return ret;
5802 }
5803
5804 /*
5805  * helper to account the unused space of all the readonly block group in the
5806  * space_info. takes mirrors into account.
5807  */
5808 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
5809 {
5810         struct btrfs_block_group *block_group;
5811         u64 free_bytes = 0;
5812         int factor;
5813
5814         /* It's df, we don't care if it's racy */
5815         if (list_empty(&sinfo->ro_bgs))
5816                 return 0;
5817
5818         spin_lock(&sinfo->lock);
5819         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
5820                 spin_lock(&block_group->lock);
5821
5822                 if (!block_group->ro) {
5823                         spin_unlock(&block_group->lock);
5824                         continue;
5825                 }
5826
5827                 factor = btrfs_bg_type_to_factor(block_group->flags);
5828                 free_bytes += (block_group->length -
5829                                block_group->used) * factor;
5830
5831                 spin_unlock(&block_group->lock);
5832         }
5833         spin_unlock(&sinfo->lock);
5834
5835         return free_bytes;
5836 }
5837
5838 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5839                                    u64 start, u64 end)
5840 {
5841         return unpin_extent_range(fs_info, start, end, false);
5842 }
5843
5844 /*
5845  * It used to be that old block groups would be left around forever.
5846  * Iterating over them would be enough to trim unused space.  Since we
5847  * now automatically remove them, we also need to iterate over unallocated
5848  * space.
5849  *
5850  * We don't want a transaction for this since the discard may take a
5851  * substantial amount of time.  We don't require that a transaction be
5852  * running, but we do need to take a running transaction into account
5853  * to ensure that we're not discarding chunks that were released or
5854  * allocated in the current transaction.
5855  *
5856  * Holding the chunks lock will prevent other threads from allocating
5857  * or releasing chunks, but it won't prevent a running transaction
5858  * from committing and releasing the memory that the pending chunks
5859  * list head uses.  For that, we need to take a reference to the
5860  * transaction and hold the commit root sem.  We only need to hold
5861  * it while performing the free space search since we have already
5862  * held back allocations.
5863  */
5864 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5865 {
5866         u64 start = SZ_1M, len = 0, end = 0;
5867         int ret;
5868
5869         *trimmed = 0;
5870
5871         /* Discard not supported = nothing to do. */
5872         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
5873                 return 0;
5874
5875         /* Not writable = nothing to do. */
5876         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5877                 return 0;
5878
5879         /* No free space = nothing to do. */
5880         if (device->total_bytes <= device->bytes_used)
5881                 return 0;
5882
5883         ret = 0;
5884
5885         while (1) {
5886                 struct btrfs_fs_info *fs_info = device->fs_info;
5887                 u64 bytes;
5888
5889                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5890                 if (ret)
5891                         break;
5892
5893                 find_first_clear_extent_bit(&device->alloc_state, start,
5894                                             &start, &end,
5895                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5896
5897                 /* Check if there are any CHUNK_* bits left */
5898                 if (start > device->total_bytes) {
5899                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
5900                         btrfs_warn_in_rcu(fs_info,
5901 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
5902                                           start, end - start + 1,
5903                                           rcu_str_deref(device->name),
5904                                           device->total_bytes);
5905                         mutex_unlock(&fs_info->chunk_mutex);
5906                         ret = 0;
5907                         break;
5908                 }
5909
5910                 /* Ensure we skip the reserved area in the first 1M */
5911                 start = max_t(u64, start, SZ_1M);
5912
5913                 /*
5914                  * If find_first_clear_extent_bit find a range that spans the
5915                  * end of the device it will set end to -1, in this case it's up
5916                  * to the caller to trim the value to the size of the device.
5917                  */
5918                 end = min(end, device->total_bytes - 1);
5919
5920                 len = end - start + 1;
5921
5922                 /* We didn't find any extents */
5923                 if (!len) {
5924                         mutex_unlock(&fs_info->chunk_mutex);
5925                         ret = 0;
5926                         break;
5927                 }
5928
5929                 ret = btrfs_issue_discard(device->bdev, start, len,
5930                                           &bytes);
5931                 if (!ret)
5932                         set_extent_bits(&device->alloc_state, start,
5933                                         start + bytes - 1,
5934                                         CHUNK_TRIMMED);
5935                 mutex_unlock(&fs_info->chunk_mutex);
5936
5937                 if (ret)
5938                         break;
5939
5940                 start += len;
5941                 *trimmed += bytes;
5942
5943                 if (fatal_signal_pending(current)) {
5944                         ret = -ERESTARTSYS;
5945                         break;
5946                 }
5947
5948                 cond_resched();
5949         }
5950
5951         return ret;
5952 }
5953
5954 /*
5955  * Trim the whole filesystem by:
5956  * 1) trimming the free space in each block group
5957  * 2) trimming the unallocated space on each device
5958  *
5959  * This will also continue trimming even if a block group or device encounters
5960  * an error.  The return value will be the last error, or 0 if nothing bad
5961  * happens.
5962  */
5963 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
5964 {
5965         struct btrfs_block_group *cache = NULL;
5966         struct btrfs_device *device;
5967         struct list_head *devices;
5968         u64 group_trimmed;
5969         u64 range_end = U64_MAX;
5970         u64 start;
5971         u64 end;
5972         u64 trimmed = 0;
5973         u64 bg_failed = 0;
5974         u64 dev_failed = 0;
5975         int bg_ret = 0;
5976         int dev_ret = 0;
5977         int ret = 0;
5978
5979         /*
5980          * Check range overflow if range->len is set.
5981          * The default range->len is U64_MAX.
5982          */
5983         if (range->len != U64_MAX &&
5984             check_add_overflow(range->start, range->len, &range_end))
5985                 return -EINVAL;
5986
5987         cache = btrfs_lookup_first_block_group(fs_info, range->start);
5988         for (; cache; cache = btrfs_next_block_group(cache)) {
5989                 if (cache->start >= range_end) {
5990                         btrfs_put_block_group(cache);
5991                         break;
5992                 }
5993
5994                 start = max(range->start, cache->start);
5995                 end = min(range_end, cache->start + cache->length);
5996
5997                 if (end - start >= range->minlen) {
5998                         if (!btrfs_block_group_done(cache)) {
5999                                 ret = btrfs_cache_block_group(cache, 0);
6000                                 if (ret) {
6001                                         bg_failed++;
6002                                         bg_ret = ret;
6003                                         continue;
6004                                 }
6005                                 ret = btrfs_wait_block_group_cache_done(cache);
6006                                 if (ret) {
6007                                         bg_failed++;
6008                                         bg_ret = ret;
6009                                         continue;
6010                                 }
6011                         }
6012                         ret = btrfs_trim_block_group(cache,
6013                                                      &group_trimmed,
6014                                                      start,
6015                                                      end,
6016                                                      range->minlen);
6017
6018                         trimmed += group_trimmed;
6019                         if (ret) {
6020                                 bg_failed++;
6021                                 bg_ret = ret;
6022                                 continue;
6023                         }
6024                 }
6025         }
6026
6027         if (bg_failed)
6028                 btrfs_warn(fs_info,
6029                         "failed to trim %llu block group(s), last error %d",
6030                         bg_failed, bg_ret);
6031         mutex_lock(&fs_info->fs_devices->device_list_mutex);
6032         devices = &fs_info->fs_devices->devices;
6033         list_for_each_entry(device, devices, dev_list) {
6034                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6035                 if (ret) {
6036                         dev_failed++;
6037                         dev_ret = ret;
6038                         break;
6039                 }
6040
6041                 trimmed += group_trimmed;
6042         }
6043         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
6044
6045         if (dev_failed)
6046                 btrfs_warn(fs_info,
6047                         "failed to trim %llu device(s), last error %d",
6048                         dev_failed, dev_ret);
6049         range->len = trimmed;
6050         if (bg_ret)
6051                 return bg_ret;
6052         return dev_ret;
6053 }