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 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
2494 {
2495         struct btrfs_block_group *block_group;
2496         int readonly = 0;
2497
2498         block_group = btrfs_lookup_block_group(fs_info, bytenr);
2499         if (!block_group || block_group->ro)
2500                 readonly = 1;
2501         if (block_group)
2502                 btrfs_put_block_group(block_group);
2503         return readonly;
2504 }
2505
2506 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2507 {
2508         struct btrfs_fs_info *fs_info = root->fs_info;
2509         u64 flags;
2510         u64 ret;
2511
2512         if (data)
2513                 flags = BTRFS_BLOCK_GROUP_DATA;
2514         else if (root == fs_info->chunk_root)
2515                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2516         else
2517                 flags = BTRFS_BLOCK_GROUP_METADATA;
2518
2519         ret = btrfs_get_alloc_profile(fs_info, flags);
2520         return ret;
2521 }
2522
2523 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
2524 {
2525         struct btrfs_block_group *cache;
2526         u64 bytenr;
2527
2528         spin_lock(&fs_info->block_group_cache_lock);
2529         bytenr = fs_info->first_logical_byte;
2530         spin_unlock(&fs_info->block_group_cache_lock);
2531
2532         if (bytenr < (u64)-1)
2533                 return bytenr;
2534
2535         cache = btrfs_lookup_first_block_group(fs_info, search_start);
2536         if (!cache)
2537                 return 0;
2538
2539         bytenr = cache->start;
2540         btrfs_put_block_group(cache);
2541
2542         return bytenr;
2543 }
2544
2545 static int pin_down_extent(struct btrfs_trans_handle *trans,
2546                            struct btrfs_block_group *cache,
2547                            u64 bytenr, u64 num_bytes, int reserved)
2548 {
2549         struct btrfs_fs_info *fs_info = cache->fs_info;
2550
2551         spin_lock(&cache->space_info->lock);
2552         spin_lock(&cache->lock);
2553         cache->pinned += num_bytes;
2554         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2555                                              num_bytes);
2556         if (reserved) {
2557                 cache->reserved -= num_bytes;
2558                 cache->space_info->bytes_reserved -= num_bytes;
2559         }
2560         spin_unlock(&cache->lock);
2561         spin_unlock(&cache->space_info->lock);
2562
2563         __btrfs_mod_total_bytes_pinned(cache->space_info, num_bytes);
2564         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2565                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2566         return 0;
2567 }
2568
2569 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2570                      u64 bytenr, u64 num_bytes, int reserved)
2571 {
2572         struct btrfs_block_group *cache;
2573
2574         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2575         BUG_ON(!cache); /* Logic error */
2576
2577         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2578
2579         btrfs_put_block_group(cache);
2580         return 0;
2581 }
2582
2583 /*
2584  * this function must be called within transaction
2585  */
2586 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2587                                     u64 bytenr, u64 num_bytes)
2588 {
2589         struct btrfs_block_group *cache;
2590         int ret;
2591
2592         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2593         if (!cache)
2594                 return -EINVAL;
2595
2596         /*
2597          * pull in the free space cache (if any) so that our pin
2598          * removes the free space from the cache.  We have load_only set
2599          * to one because the slow code to read in the free extents does check
2600          * the pinned extents.
2601          */
2602         btrfs_cache_block_group(cache, 1);
2603         /*
2604          * Make sure we wait until the cache is completely built in case it is
2605          * missing or is invalid and therefore needs to be rebuilt.
2606          */
2607         ret = btrfs_wait_block_group_cache_done(cache);
2608         if (ret)
2609                 goto out;
2610
2611         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2612
2613         /* remove us from the free space cache (if we're there at all) */
2614         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2615 out:
2616         btrfs_put_block_group(cache);
2617         return ret;
2618 }
2619
2620 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2621                                    u64 start, u64 num_bytes)
2622 {
2623         int ret;
2624         struct btrfs_block_group *block_group;
2625
2626         block_group = btrfs_lookup_block_group(fs_info, start);
2627         if (!block_group)
2628                 return -EINVAL;
2629
2630         btrfs_cache_block_group(block_group, 1);
2631         /*
2632          * Make sure we wait until the cache is completely built in case it is
2633          * missing or is invalid and therefore needs to be rebuilt.
2634          */
2635         ret = btrfs_wait_block_group_cache_done(block_group);
2636         if (ret)
2637                 goto out;
2638
2639         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2640 out:
2641         btrfs_put_block_group(block_group);
2642         return ret;
2643 }
2644
2645 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2646 {
2647         struct btrfs_fs_info *fs_info = eb->fs_info;
2648         struct btrfs_file_extent_item *item;
2649         struct btrfs_key key;
2650         int found_type;
2651         int i;
2652         int ret = 0;
2653
2654         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2655                 return 0;
2656
2657         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2658                 btrfs_item_key_to_cpu(eb, &key, i);
2659                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2660                         continue;
2661                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2662                 found_type = btrfs_file_extent_type(eb, item);
2663                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2664                         continue;
2665                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2666                         continue;
2667                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2668                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2669                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2670                 if (ret)
2671                         break;
2672         }
2673
2674         return ret;
2675 }
2676
2677 static void
2678 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2679 {
2680         atomic_inc(&bg->reservations);
2681 }
2682
2683 /*
2684  * Returns the free cluster for the given space info and sets empty_cluster to
2685  * what it should be based on the mount options.
2686  */
2687 static struct btrfs_free_cluster *
2688 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2689                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2690 {
2691         struct btrfs_free_cluster *ret = NULL;
2692
2693         *empty_cluster = 0;
2694         if (btrfs_mixed_space_info(space_info))
2695                 return ret;
2696
2697         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2698                 ret = &fs_info->meta_alloc_cluster;
2699                 if (btrfs_test_opt(fs_info, SSD))
2700                         *empty_cluster = SZ_2M;
2701                 else
2702                         *empty_cluster = SZ_64K;
2703         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2704                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2705                 *empty_cluster = SZ_2M;
2706                 ret = &fs_info->data_alloc_cluster;
2707         }
2708
2709         return ret;
2710 }
2711
2712 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2713                               u64 start, u64 end,
2714                               const bool return_free_space)
2715 {
2716         struct btrfs_block_group *cache = NULL;
2717         struct btrfs_space_info *space_info;
2718         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2719         struct btrfs_free_cluster *cluster = NULL;
2720         u64 len;
2721         u64 total_unpinned = 0;
2722         u64 empty_cluster = 0;
2723         bool readonly;
2724
2725         while (start <= end) {
2726                 readonly = false;
2727                 if (!cache ||
2728                     start >= cache->start + cache->length) {
2729                         if (cache)
2730                                 btrfs_put_block_group(cache);
2731                         total_unpinned = 0;
2732                         cache = btrfs_lookup_block_group(fs_info, start);
2733                         BUG_ON(!cache); /* Logic error */
2734
2735                         cluster = fetch_cluster_info(fs_info,
2736                                                      cache->space_info,
2737                                                      &empty_cluster);
2738                         empty_cluster <<= 1;
2739                 }
2740
2741                 len = cache->start + cache->length - start;
2742                 len = min(len, end + 1 - start);
2743
2744                 down_read(&fs_info->commit_root_sem);
2745                 if (start < cache->last_byte_to_unpin && return_free_space) {
2746                         u64 add_len = min(len, cache->last_byte_to_unpin - start);
2747
2748                         btrfs_add_free_space(cache, start, add_len);
2749                 }
2750                 up_read(&fs_info->commit_root_sem);
2751
2752                 start += len;
2753                 total_unpinned += len;
2754                 space_info = cache->space_info;
2755
2756                 /*
2757                  * If this space cluster has been marked as fragmented and we've
2758                  * unpinned enough in this block group to potentially allow a
2759                  * cluster to be created inside of it go ahead and clear the
2760                  * fragmented check.
2761                  */
2762                 if (cluster && cluster->fragmented &&
2763                     total_unpinned > empty_cluster) {
2764                         spin_lock(&cluster->lock);
2765                         cluster->fragmented = 0;
2766                         spin_unlock(&cluster->lock);
2767                 }
2768
2769                 spin_lock(&space_info->lock);
2770                 spin_lock(&cache->lock);
2771                 cache->pinned -= len;
2772                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2773                 space_info->max_extent_size = 0;
2774                 __btrfs_mod_total_bytes_pinned(space_info, -len);
2775                 if (cache->ro) {
2776                         space_info->bytes_readonly += len;
2777                         readonly = true;
2778                 } else if (btrfs_is_zoned(fs_info)) {
2779                         /* Need reset before reusing in a zoned block group */
2780                         space_info->bytes_zone_unusable += len;
2781                         readonly = true;
2782                 }
2783                 spin_unlock(&cache->lock);
2784                 if (!readonly && return_free_space &&
2785                     global_rsv->space_info == space_info) {
2786                         u64 to_add = len;
2787
2788                         spin_lock(&global_rsv->lock);
2789                         if (!global_rsv->full) {
2790                                 to_add = min(len, global_rsv->size -
2791                                              global_rsv->reserved);
2792                                 global_rsv->reserved += to_add;
2793                                 btrfs_space_info_update_bytes_may_use(fs_info,
2794                                                 space_info, to_add);
2795                                 if (global_rsv->reserved >= global_rsv->size)
2796                                         global_rsv->full = 1;
2797                                 len -= to_add;
2798                         }
2799                         spin_unlock(&global_rsv->lock);
2800                 }
2801                 /* Add to any tickets we may have */
2802                 if (!readonly && return_free_space && len)
2803                         btrfs_try_granting_tickets(fs_info, space_info);
2804                 spin_unlock(&space_info->lock);
2805         }
2806
2807         if (cache)
2808                 btrfs_put_block_group(cache);
2809         return 0;
2810 }
2811
2812 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2813 {
2814         struct btrfs_fs_info *fs_info = trans->fs_info;
2815         struct btrfs_block_group *block_group, *tmp;
2816         struct list_head *deleted_bgs;
2817         struct extent_io_tree *unpin;
2818         u64 start;
2819         u64 end;
2820         int ret;
2821
2822         unpin = &trans->transaction->pinned_extents;
2823
2824         while (!TRANS_ABORTED(trans)) {
2825                 struct extent_state *cached_state = NULL;
2826
2827                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2828                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2829                                             EXTENT_DIRTY, &cached_state);
2830                 if (ret) {
2831                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2832                         break;
2833                 }
2834
2835                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2836                         ret = btrfs_discard_extent(fs_info, start,
2837                                                    end + 1 - start, NULL);
2838
2839                 clear_extent_dirty(unpin, start, end, &cached_state);
2840                 unpin_extent_range(fs_info, start, end, true);
2841                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2842                 free_extent_state(cached_state);
2843                 cond_resched();
2844         }
2845
2846         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2847                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2848                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2849         }
2850
2851         /*
2852          * Transaction is finished.  We don't need the lock anymore.  We
2853          * do need to clean up the block groups in case of a transaction
2854          * abort.
2855          */
2856         deleted_bgs = &trans->transaction->deleted_bgs;
2857         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2858                 u64 trimmed = 0;
2859
2860                 ret = -EROFS;
2861                 if (!TRANS_ABORTED(trans))
2862                         ret = btrfs_discard_extent(fs_info,
2863                                                    block_group->start,
2864                                                    block_group->length,
2865                                                    &trimmed);
2866
2867                 list_del_init(&block_group->bg_list);
2868                 btrfs_unfreeze_block_group(block_group);
2869                 btrfs_put_block_group(block_group);
2870
2871                 if (ret) {
2872                         const char *errstr = btrfs_decode_error(ret);
2873                         btrfs_warn(fs_info,
2874                            "discard failed while removing blockgroup: errno=%d %s",
2875                                    ret, errstr);
2876                 }
2877         }
2878
2879         return 0;
2880 }
2881
2882 /*
2883  * Drop one or more refs of @node.
2884  *
2885  * 1. Locate the extent refs.
2886  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2887  *    Locate it, then reduce the refs number or remove the ref line completely.
2888  *
2889  * 2. Update the refs count in EXTENT/METADATA_ITEM
2890  *
2891  * Inline backref case:
2892  *
2893  * in extent tree we have:
2894  *
2895  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2896  *              refs 2 gen 6 flags DATA
2897  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2898  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2899  *
2900  * This function gets called with:
2901  *
2902  *    node->bytenr = 13631488
2903  *    node->num_bytes = 1048576
2904  *    root_objectid = FS_TREE
2905  *    owner_objectid = 257
2906  *    owner_offset = 0
2907  *    refs_to_drop = 1
2908  *
2909  * Then we should get some like:
2910  *
2911  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2912  *              refs 1 gen 6 flags DATA
2913  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2914  *
2915  * Keyed backref case:
2916  *
2917  * in extent tree we have:
2918  *
2919  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2920  *              refs 754 gen 6 flags DATA
2921  *      [...]
2922  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2923  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2924  *
2925  * This function get called with:
2926  *
2927  *    node->bytenr = 13631488
2928  *    node->num_bytes = 1048576
2929  *    root_objectid = FS_TREE
2930  *    owner_objectid = 866
2931  *    owner_offset = 0
2932  *    refs_to_drop = 1
2933  *
2934  * Then we should get some like:
2935  *
2936  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2937  *              refs 753 gen 6 flags DATA
2938  *
2939  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2940  */
2941 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2942                                struct btrfs_delayed_ref_node *node, u64 parent,
2943                                u64 root_objectid, u64 owner_objectid,
2944                                u64 owner_offset, int refs_to_drop,
2945                                struct btrfs_delayed_extent_op *extent_op)
2946 {
2947         struct btrfs_fs_info *info = trans->fs_info;
2948         struct btrfs_key key;
2949         struct btrfs_path *path;
2950         struct btrfs_root *extent_root = info->extent_root;
2951         struct extent_buffer *leaf;
2952         struct btrfs_extent_item *ei;
2953         struct btrfs_extent_inline_ref *iref;
2954         int ret;
2955         int is_data;
2956         int extent_slot = 0;
2957         int found_extent = 0;
2958         int num_to_del = 1;
2959         u32 item_size;
2960         u64 refs;
2961         u64 bytenr = node->bytenr;
2962         u64 num_bytes = node->num_bytes;
2963         int last_ref = 0;
2964         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2965
2966         path = btrfs_alloc_path();
2967         if (!path)
2968                 return -ENOMEM;
2969
2970         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2971
2972         if (!is_data && refs_to_drop != 1) {
2973                 btrfs_crit(info,
2974 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2975                            node->bytenr, refs_to_drop);
2976                 ret = -EINVAL;
2977                 btrfs_abort_transaction(trans, ret);
2978                 goto out;
2979         }
2980
2981         if (is_data)
2982                 skinny_metadata = false;
2983
2984         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2985                                     parent, root_objectid, owner_objectid,
2986                                     owner_offset);
2987         if (ret == 0) {
2988                 /*
2989                  * Either the inline backref or the SHARED_DATA_REF/
2990                  * SHARED_BLOCK_REF is found
2991                  *
2992                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2993                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2994                  */
2995                 extent_slot = path->slots[0];
2996                 while (extent_slot >= 0) {
2997                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2998                                               extent_slot);
2999                         if (key.objectid != bytenr)
3000                                 break;
3001                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3002                             key.offset == num_bytes) {
3003                                 found_extent = 1;
3004                                 break;
3005                         }
3006                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3007                             key.offset == owner_objectid) {
3008                                 found_extent = 1;
3009                                 break;
3010                         }
3011
3012                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3013                         if (path->slots[0] - extent_slot > 5)
3014                                 break;
3015                         extent_slot--;
3016                 }
3017
3018                 if (!found_extent) {
3019                         if (iref) {
3020                                 btrfs_crit(info,
3021 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
3022                                 btrfs_abort_transaction(trans, -EUCLEAN);
3023                                 goto err_dump;
3024                         }
3025                         /* Must be SHARED_* item, remove the backref first */
3026                         ret = remove_extent_backref(trans, path, NULL,
3027                                                     refs_to_drop,
3028                                                     is_data, &last_ref);
3029                         if (ret) {
3030                                 btrfs_abort_transaction(trans, ret);
3031                                 goto out;
3032                         }
3033                         btrfs_release_path(path);
3034
3035                         /* Slow path to locate EXTENT/METADATA_ITEM */
3036                         key.objectid = bytenr;
3037                         key.type = BTRFS_EXTENT_ITEM_KEY;
3038                         key.offset = num_bytes;
3039
3040                         if (!is_data && skinny_metadata) {
3041                                 key.type = BTRFS_METADATA_ITEM_KEY;
3042                                 key.offset = owner_objectid;
3043                         }
3044
3045                         ret = btrfs_search_slot(trans, extent_root,
3046                                                 &key, path, -1, 1);
3047                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3048                                 /*
3049                                  * Couldn't find our skinny metadata item,
3050                                  * see if we have ye olde extent item.
3051                                  */
3052                                 path->slots[0]--;
3053                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3054                                                       path->slots[0]);
3055                                 if (key.objectid == bytenr &&
3056                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3057                                     key.offset == num_bytes)
3058                                         ret = 0;
3059                         }
3060
3061                         if (ret > 0 && skinny_metadata) {
3062                                 skinny_metadata = false;
3063                                 key.objectid = bytenr;
3064                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3065                                 key.offset = num_bytes;
3066                                 btrfs_release_path(path);
3067                                 ret = btrfs_search_slot(trans, extent_root,
3068                                                         &key, path, -1, 1);
3069                         }
3070
3071                         if (ret) {
3072                                 btrfs_err(info,
3073                                           "umm, got %d back from search, was looking for %llu",
3074                                           ret, bytenr);
3075                                 if (ret > 0)
3076                                         btrfs_print_leaf(path->nodes[0]);
3077                         }
3078                         if (ret < 0) {
3079                                 btrfs_abort_transaction(trans, ret);
3080                                 goto out;
3081                         }
3082                         extent_slot = path->slots[0];
3083                 }
3084         } else if (WARN_ON(ret == -ENOENT)) {
3085                 btrfs_print_leaf(path->nodes[0]);
3086                 btrfs_err(info,
3087                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3088                         bytenr, parent, root_objectid, owner_objectid,
3089                         owner_offset);
3090                 btrfs_abort_transaction(trans, ret);
3091                 goto out;
3092         } else {
3093                 btrfs_abort_transaction(trans, ret);
3094                 goto out;
3095         }
3096
3097         leaf = path->nodes[0];
3098         item_size = btrfs_item_size_nr(leaf, extent_slot);
3099         if (unlikely(item_size < sizeof(*ei))) {
3100                 ret = -EINVAL;
3101                 btrfs_print_v0_err(info);
3102                 btrfs_abort_transaction(trans, ret);
3103                 goto out;
3104         }
3105         ei = btrfs_item_ptr(leaf, extent_slot,
3106                             struct btrfs_extent_item);
3107         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3108             key.type == BTRFS_EXTENT_ITEM_KEY) {
3109                 struct btrfs_tree_block_info *bi;
3110                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3111                         btrfs_crit(info,
3112 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3113                                    key.objectid, key.type, key.offset,
3114                                    owner_objectid, item_size,
3115                                    sizeof(*ei) + sizeof(*bi));
3116                         btrfs_abort_transaction(trans, -EUCLEAN);
3117                         goto err_dump;
3118                 }
3119                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3120                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3121         }
3122
3123         refs = btrfs_extent_refs(leaf, ei);
3124         if (refs < refs_to_drop) {
3125                 btrfs_crit(info,
3126                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3127                           refs_to_drop, refs, bytenr);
3128                 btrfs_abort_transaction(trans, -EUCLEAN);
3129                 goto err_dump;
3130         }
3131         refs -= refs_to_drop;
3132
3133         if (refs > 0) {
3134                 if (extent_op)
3135                         __run_delayed_extent_op(extent_op, leaf, ei);
3136                 /*
3137                  * In the case of inline back ref, reference count will
3138                  * be updated by remove_extent_backref
3139                  */
3140                 if (iref) {
3141                         if (!found_extent) {
3142                                 btrfs_crit(info,
3143 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3144                                 btrfs_abort_transaction(trans, -EUCLEAN);
3145                                 goto err_dump;
3146                         }
3147                 } else {
3148                         btrfs_set_extent_refs(leaf, ei, refs);
3149                         btrfs_mark_buffer_dirty(leaf);
3150                 }
3151                 if (found_extent) {
3152                         ret = remove_extent_backref(trans, path, iref,
3153                                                     refs_to_drop, is_data,
3154                                                     &last_ref);
3155                         if (ret) {
3156                                 btrfs_abort_transaction(trans, ret);
3157                                 goto out;
3158                         }
3159                 }
3160         } else {
3161                 /* In this branch refs == 1 */
3162                 if (found_extent) {
3163                         if (is_data && refs_to_drop !=
3164                             extent_data_ref_count(path, iref)) {
3165                                 btrfs_crit(info,
3166                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3167                                            extent_data_ref_count(path, iref),
3168                                            refs_to_drop);
3169                                 btrfs_abort_transaction(trans, -EUCLEAN);
3170                                 goto err_dump;
3171                         }
3172                         if (iref) {
3173                                 if (path->slots[0] != extent_slot) {
3174                                         btrfs_crit(info,
3175 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3176                                                    key.objectid, key.type,
3177                                                    key.offset);
3178                                         btrfs_abort_transaction(trans, -EUCLEAN);
3179                                         goto err_dump;
3180                                 }
3181                         } else {
3182                                 /*
3183                                  * No inline ref, we must be at SHARED_* item,
3184                                  * And it's single ref, it must be:
3185                                  * |    extent_slot       ||extent_slot + 1|
3186                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3187                                  */
3188                                 if (path->slots[0] != extent_slot + 1) {
3189                                         btrfs_crit(info,
3190         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3191                                         btrfs_abort_transaction(trans, -EUCLEAN);
3192                                         goto err_dump;
3193                                 }
3194                                 path->slots[0] = extent_slot;
3195                                 num_to_del = 2;
3196                         }
3197                 }
3198
3199                 last_ref = 1;
3200                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3201                                       num_to_del);
3202                 if (ret) {
3203                         btrfs_abort_transaction(trans, ret);
3204                         goto out;
3205                 }
3206                 btrfs_release_path(path);
3207
3208                 if (is_data) {
3209                         ret = btrfs_del_csums(trans, info->csum_root, bytenr,
3210                                               num_bytes);
3211                         if (ret) {
3212                                 btrfs_abort_transaction(trans, ret);
3213                                 goto out;
3214                         }
3215                 }
3216
3217                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3218                 if (ret) {
3219                         btrfs_abort_transaction(trans, ret);
3220                         goto out;
3221                 }
3222
3223                 ret = btrfs_update_block_group(trans, bytenr, num_bytes, 0);
3224                 if (ret) {
3225                         btrfs_abort_transaction(trans, ret);
3226                         goto out;
3227                 }
3228         }
3229         btrfs_release_path(path);
3230
3231 out:
3232         btrfs_free_path(path);
3233         return ret;
3234 err_dump:
3235         /*
3236          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3237          * dump for debug build.
3238          */
3239         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3240                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3241                            path->slots[0], extent_slot);
3242                 btrfs_print_leaf(path->nodes[0]);
3243         }
3244
3245         btrfs_free_path(path);
3246         return -EUCLEAN;
3247 }
3248
3249 /*
3250  * when we free an block, it is possible (and likely) that we free the last
3251  * delayed ref for that extent as well.  This searches the delayed ref tree for
3252  * a given extent, and if there are no other delayed refs to be processed, it
3253  * removes it from the tree.
3254  */
3255 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3256                                       u64 bytenr)
3257 {
3258         struct btrfs_delayed_ref_head *head;
3259         struct btrfs_delayed_ref_root *delayed_refs;
3260         int ret = 0;
3261
3262         delayed_refs = &trans->transaction->delayed_refs;
3263         spin_lock(&delayed_refs->lock);
3264         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3265         if (!head)
3266                 goto out_delayed_unlock;
3267
3268         spin_lock(&head->lock);
3269         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3270                 goto out;
3271
3272         if (cleanup_extent_op(head) != NULL)
3273                 goto out;
3274
3275         /*
3276          * waiting for the lock here would deadlock.  If someone else has it
3277          * locked they are already in the process of dropping it anyway
3278          */
3279         if (!mutex_trylock(&head->mutex))
3280                 goto out;
3281
3282         btrfs_delete_ref_head(delayed_refs, head);
3283         head->processing = 0;
3284
3285         spin_unlock(&head->lock);
3286         spin_unlock(&delayed_refs->lock);
3287
3288         BUG_ON(head->extent_op);
3289         if (head->must_insert_reserved)
3290                 ret = 1;
3291
3292         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3293         mutex_unlock(&head->mutex);
3294         btrfs_put_delayed_ref_head(head);
3295         return ret;
3296 out:
3297         spin_unlock(&head->lock);
3298
3299 out_delayed_unlock:
3300         spin_unlock(&delayed_refs->lock);
3301         return 0;
3302 }
3303
3304 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3305                            struct btrfs_root *root,
3306                            struct extent_buffer *buf,
3307                            u64 parent, int last_ref)
3308 {
3309         struct btrfs_fs_info *fs_info = root->fs_info;
3310         struct btrfs_ref generic_ref = { 0 };
3311         int ret;
3312
3313         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3314                                buf->start, buf->len, parent);
3315         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3316                             root->root_key.objectid);
3317
3318         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3319                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3320                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3321                 BUG_ON(ret); /* -ENOMEM */
3322         }
3323
3324         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3325                 struct btrfs_block_group *cache;
3326
3327                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3328                         ret = check_ref_cleanup(trans, buf->start);
3329                         if (!ret) {
3330                                 btrfs_redirty_list_add(trans->transaction, buf);
3331                                 goto out;
3332                         }
3333                 }
3334
3335                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3336
3337                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3338                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3339                         btrfs_put_block_group(cache);
3340                         goto out;
3341                 }
3342
3343                 if (btrfs_is_zoned(fs_info)) {
3344                         btrfs_redirty_list_add(trans->transaction, buf);
3345                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3346                         btrfs_put_block_group(cache);
3347                         goto out;
3348                 }
3349
3350                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3351
3352                 btrfs_add_free_space(cache, buf->start, buf->len);
3353                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3354                 btrfs_put_block_group(cache);
3355                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3356         }
3357 out:
3358         if (last_ref) {
3359                 /*
3360                  * Deleting the buffer, clear the corrupt flag since it doesn't
3361                  * matter anymore.
3362                  */
3363                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3364         }
3365 }
3366
3367 /* Can return -ENOMEM */
3368 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3369 {
3370         struct btrfs_fs_info *fs_info = trans->fs_info;
3371         int ret;
3372
3373         if (btrfs_is_testing(fs_info))
3374                 return 0;
3375
3376         /*
3377          * tree log blocks never actually go into the extent allocation
3378          * tree, just update pinning info and exit early.
3379          */
3380         if ((ref->type == BTRFS_REF_METADATA &&
3381              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3382             (ref->type == BTRFS_REF_DATA &&
3383              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3384                 /* unlocks the pinned mutex */
3385                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3386                 ret = 0;
3387         } else if (ref->type == BTRFS_REF_METADATA) {
3388                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3389         } else {
3390                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3391         }
3392
3393         if (!((ref->type == BTRFS_REF_METADATA &&
3394                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3395               (ref->type == BTRFS_REF_DATA &&
3396                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3397                 btrfs_ref_tree_mod(fs_info, ref);
3398
3399         return ret;
3400 }
3401
3402 enum btrfs_loop_type {
3403         LOOP_CACHING_NOWAIT,
3404         LOOP_CACHING_WAIT,
3405         LOOP_ALLOC_CHUNK,
3406         LOOP_NO_EMPTY_SIZE,
3407 };
3408
3409 static inline void
3410 btrfs_lock_block_group(struct btrfs_block_group *cache,
3411                        int delalloc)
3412 {
3413         if (delalloc)
3414                 down_read(&cache->data_rwsem);
3415 }
3416
3417 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3418                        int delalloc)
3419 {
3420         btrfs_get_block_group(cache);
3421         if (delalloc)
3422                 down_read(&cache->data_rwsem);
3423 }
3424
3425 static struct btrfs_block_group *btrfs_lock_cluster(
3426                    struct btrfs_block_group *block_group,
3427                    struct btrfs_free_cluster *cluster,
3428                    int delalloc)
3429         __acquires(&cluster->refill_lock)
3430 {
3431         struct btrfs_block_group *used_bg = NULL;
3432
3433         spin_lock(&cluster->refill_lock);
3434         while (1) {
3435                 used_bg = cluster->block_group;
3436                 if (!used_bg)
3437                         return NULL;
3438
3439                 if (used_bg == block_group)
3440                         return used_bg;
3441
3442                 btrfs_get_block_group(used_bg);
3443
3444                 if (!delalloc)
3445                         return used_bg;
3446
3447                 if (down_read_trylock(&used_bg->data_rwsem))
3448                         return used_bg;
3449
3450                 spin_unlock(&cluster->refill_lock);
3451
3452                 /* We should only have one-level nested. */
3453                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3454
3455                 spin_lock(&cluster->refill_lock);
3456                 if (used_bg == cluster->block_group)
3457                         return used_bg;
3458
3459                 up_read(&used_bg->data_rwsem);
3460                 btrfs_put_block_group(used_bg);
3461         }
3462 }
3463
3464 static inline void
3465 btrfs_release_block_group(struct btrfs_block_group *cache,
3466                          int delalloc)
3467 {
3468         if (delalloc)
3469                 up_read(&cache->data_rwsem);
3470         btrfs_put_block_group(cache);
3471 }
3472
3473 enum btrfs_extent_allocation_policy {
3474         BTRFS_EXTENT_ALLOC_CLUSTERED,
3475         BTRFS_EXTENT_ALLOC_ZONED,
3476 };
3477
3478 /*
3479  * Structure used internally for find_free_extent() function.  Wraps needed
3480  * parameters.
3481  */
3482 struct find_free_extent_ctl {
3483         /* Basic allocation info */
3484         u64 num_bytes;
3485         u64 empty_size;
3486         u64 flags;
3487         int delalloc;
3488
3489         /* Where to start the search inside the bg */
3490         u64 search_start;
3491
3492         /* For clustered allocation */
3493         u64 empty_cluster;
3494         struct btrfs_free_cluster *last_ptr;
3495         bool use_cluster;
3496
3497         bool have_caching_bg;
3498         bool orig_have_caching_bg;
3499
3500         /* Allocation is called for tree-log */
3501         bool for_treelog;
3502
3503         /* RAID index, converted from flags */
3504         int index;
3505
3506         /*
3507          * Current loop number, check find_free_extent_update_loop() for details
3508          */
3509         int loop;
3510
3511         /*
3512          * Whether we're refilling a cluster, if true we need to re-search
3513          * current block group but don't try to refill the cluster again.
3514          */
3515         bool retry_clustered;
3516
3517         /*
3518          * Whether we're updating free space cache, if true we need to re-search
3519          * current block group but don't try updating free space cache again.
3520          */
3521         bool retry_unclustered;
3522
3523         /* If current block group is cached */
3524         int cached;
3525
3526         /* Max contiguous hole found */
3527         u64 max_extent_size;
3528
3529         /* Total free space from free space cache, not always contiguous */
3530         u64 total_free_space;
3531
3532         /* Found result */
3533         u64 found_offset;
3534
3535         /* Hint where to start looking for an empty space */
3536         u64 hint_byte;
3537
3538         /* Allocation policy */
3539         enum btrfs_extent_allocation_policy policy;
3540 };
3541
3542
3543 /*
3544  * Helper function for find_free_extent().
3545  *
3546  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3547  * Return -EAGAIN to inform caller that we need to re-search this block group
3548  * Return >0 to inform caller that we find nothing
3549  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3550  */
3551 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3552                                       struct find_free_extent_ctl *ffe_ctl,
3553                                       struct btrfs_block_group **cluster_bg_ret)
3554 {
3555         struct btrfs_block_group *cluster_bg;
3556         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3557         u64 aligned_cluster;
3558         u64 offset;
3559         int ret;
3560
3561         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3562         if (!cluster_bg)
3563                 goto refill_cluster;
3564         if (cluster_bg != bg && (cluster_bg->ro ||
3565             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3566                 goto release_cluster;
3567
3568         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3569                         ffe_ctl->num_bytes, cluster_bg->start,
3570                         &ffe_ctl->max_extent_size);
3571         if (offset) {
3572                 /* We have a block, we're done */
3573                 spin_unlock(&last_ptr->refill_lock);
3574                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3575                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3576                 *cluster_bg_ret = cluster_bg;
3577                 ffe_ctl->found_offset = offset;
3578                 return 0;
3579         }
3580         WARN_ON(last_ptr->block_group != cluster_bg);
3581
3582 release_cluster:
3583         /*
3584          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3585          * lets just skip it and let the allocator find whatever block it can
3586          * find. If we reach this point, we will have tried the cluster
3587          * allocator plenty of times and not have found anything, so we are
3588          * likely way too fragmented for the clustering stuff to find anything.
3589          *
3590          * However, if the cluster is taken from the current block group,
3591          * release the cluster first, so that we stand a better chance of
3592          * succeeding in the unclustered allocation.
3593          */
3594         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3595                 spin_unlock(&last_ptr->refill_lock);
3596                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3597                 return -ENOENT;
3598         }
3599
3600         /* This cluster didn't work out, free it and start over */
3601         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3602
3603         if (cluster_bg != bg)
3604                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3605
3606 refill_cluster:
3607         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3608                 spin_unlock(&last_ptr->refill_lock);
3609                 return -ENOENT;
3610         }
3611
3612         aligned_cluster = max_t(u64,
3613                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3614                         bg->full_stripe_len);
3615         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3616                         ffe_ctl->num_bytes, aligned_cluster);
3617         if (ret == 0) {
3618                 /* Now pull our allocation out of this cluster */
3619                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3620                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3621                                 &ffe_ctl->max_extent_size);
3622                 if (offset) {
3623                         /* We found one, proceed */
3624                         spin_unlock(&last_ptr->refill_lock);
3625                         trace_btrfs_reserve_extent_cluster(bg,
3626                                         ffe_ctl->search_start,
3627                                         ffe_ctl->num_bytes);
3628                         ffe_ctl->found_offset = offset;
3629                         return 0;
3630                 }
3631         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3632                    !ffe_ctl->retry_clustered) {
3633                 spin_unlock(&last_ptr->refill_lock);
3634
3635                 ffe_ctl->retry_clustered = true;
3636                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3637                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3638                 return -EAGAIN;
3639         }
3640         /*
3641          * At this point we either didn't find a cluster or we weren't able to
3642          * allocate a block from our cluster.  Free the cluster we've been
3643          * trying to use, and go to the next block group.
3644          */
3645         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3646         spin_unlock(&last_ptr->refill_lock);
3647         return 1;
3648 }
3649
3650 /*
3651  * Return >0 to inform caller that we find nothing
3652  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3653  * Return -EAGAIN to inform caller that we need to re-search this block group
3654  */
3655 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3656                                         struct find_free_extent_ctl *ffe_ctl)
3657 {
3658         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3659         u64 offset;
3660
3661         /*
3662          * We are doing an unclustered allocation, set the fragmented flag so
3663          * we don't bother trying to setup a cluster again until we get more
3664          * space.
3665          */
3666         if (unlikely(last_ptr)) {
3667                 spin_lock(&last_ptr->lock);
3668                 last_ptr->fragmented = 1;
3669                 spin_unlock(&last_ptr->lock);
3670         }
3671         if (ffe_ctl->cached) {
3672                 struct btrfs_free_space_ctl *free_space_ctl;
3673
3674                 free_space_ctl = bg->free_space_ctl;
3675                 spin_lock(&free_space_ctl->tree_lock);
3676                 if (free_space_ctl->free_space <
3677                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3678                     ffe_ctl->empty_size) {
3679                         ffe_ctl->total_free_space = max_t(u64,
3680                                         ffe_ctl->total_free_space,
3681                                         free_space_ctl->free_space);
3682                         spin_unlock(&free_space_ctl->tree_lock);
3683                         return 1;
3684                 }
3685                 spin_unlock(&free_space_ctl->tree_lock);
3686         }
3687
3688         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3689                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3690                         &ffe_ctl->max_extent_size);
3691
3692         /*
3693          * If we didn't find a chunk, and we haven't failed on this block group
3694          * before, and this block group is in the middle of caching and we are
3695          * ok with waiting, then go ahead and wait for progress to be made, and
3696          * set @retry_unclustered to true.
3697          *
3698          * If @retry_unclustered is true then we've already waited on this
3699          * block group once and should move on to the next block group.
3700          */
3701         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3702             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3703                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3704                                                       ffe_ctl->empty_size);
3705                 ffe_ctl->retry_unclustered = true;
3706                 return -EAGAIN;
3707         } else if (!offset) {
3708                 return 1;
3709         }
3710         ffe_ctl->found_offset = offset;
3711         return 0;
3712 }
3713
3714 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3715                                    struct find_free_extent_ctl *ffe_ctl,
3716                                    struct btrfs_block_group **bg_ret)
3717 {
3718         int ret;
3719
3720         /* We want to try and use the cluster allocator, so lets look there */
3721         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3722                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3723                 if (ret >= 0 || ret == -EAGAIN)
3724                         return ret;
3725                 /* ret == -ENOENT case falls through */
3726         }
3727
3728         return find_free_extent_unclustered(block_group, ffe_ctl);
3729 }
3730
3731 /*
3732  * Tree-log block group locking
3733  * ============================
3734  *
3735  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3736  * indicates the starting address of a block group, which is reserved only
3737  * for tree-log metadata.
3738  *
3739  * Lock nesting
3740  * ============
3741  *
3742  * space_info::lock
3743  *   block_group::lock
3744  *     fs_info::treelog_bg_lock
3745  */
3746
3747 /*
3748  * Simple allocator for sequential-only block group. It only allows sequential
3749  * allocation. No need to play with trees. This function also reserves the
3750  * bytes as in btrfs_add_reserved_bytes.
3751  */
3752 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3753                                struct find_free_extent_ctl *ffe_ctl,
3754                                struct btrfs_block_group **bg_ret)
3755 {
3756         struct btrfs_fs_info *fs_info = block_group->fs_info;
3757         struct btrfs_space_info *space_info = block_group->space_info;
3758         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3759         u64 start = block_group->start;
3760         u64 num_bytes = ffe_ctl->num_bytes;
3761         u64 avail;
3762         u64 bytenr = block_group->start;
3763         u64 log_bytenr;
3764         int ret = 0;
3765         bool skip;
3766
3767         ASSERT(btrfs_is_zoned(block_group->fs_info));
3768
3769         /*
3770          * Do not allow non-tree-log blocks in the dedicated tree-log block
3771          * group, and vice versa.
3772          */
3773         spin_lock(&fs_info->treelog_bg_lock);
3774         log_bytenr = fs_info->treelog_bg;
3775         skip = log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3776                               (!ffe_ctl->for_treelog && bytenr == log_bytenr));
3777         spin_unlock(&fs_info->treelog_bg_lock);
3778         if (skip)
3779                 return 1;
3780
3781         spin_lock(&space_info->lock);
3782         spin_lock(&block_group->lock);
3783         spin_lock(&fs_info->treelog_bg_lock);
3784
3785         ASSERT(!ffe_ctl->for_treelog ||
3786                block_group->start == fs_info->treelog_bg ||
3787                fs_info->treelog_bg == 0);
3788
3789         if (block_group->ro) {
3790                 ret = 1;
3791                 goto out;
3792         }
3793
3794         /*
3795          * Do not allow currently using block group to be tree-log dedicated
3796          * block group.
3797          */
3798         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3799             (block_group->used || block_group->reserved)) {
3800                 ret = 1;
3801                 goto out;
3802         }
3803
3804         avail = block_group->length - block_group->alloc_offset;
3805         if (avail < num_bytes) {
3806                 if (ffe_ctl->max_extent_size < avail) {
3807                         /*
3808                          * With sequential allocator, free space is always
3809                          * contiguous
3810                          */
3811                         ffe_ctl->max_extent_size = avail;
3812                         ffe_ctl->total_free_space = avail;
3813                 }
3814                 ret = 1;
3815                 goto out;
3816         }
3817
3818         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3819                 fs_info->treelog_bg = block_group->start;
3820
3821         ffe_ctl->found_offset = start + block_group->alloc_offset;
3822         block_group->alloc_offset += num_bytes;
3823         spin_lock(&ctl->tree_lock);
3824         ctl->free_space -= num_bytes;
3825         spin_unlock(&ctl->tree_lock);
3826
3827         /*
3828          * We do not check if found_offset is aligned to stripesize. The
3829          * address is anyway rewritten when using zone append writing.
3830          */
3831
3832         ffe_ctl->search_start = ffe_ctl->found_offset;
3833
3834 out:
3835         if (ret && ffe_ctl->for_treelog)
3836                 fs_info->treelog_bg = 0;
3837         spin_unlock(&fs_info->treelog_bg_lock);
3838         spin_unlock(&block_group->lock);
3839         spin_unlock(&space_info->lock);
3840         return ret;
3841 }
3842
3843 static int do_allocation(struct btrfs_block_group *block_group,
3844                          struct find_free_extent_ctl *ffe_ctl,
3845                          struct btrfs_block_group **bg_ret)
3846 {
3847         switch (ffe_ctl->policy) {
3848         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3849                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3850         case BTRFS_EXTENT_ALLOC_ZONED:
3851                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3852         default:
3853                 BUG();
3854         }
3855 }
3856
3857 static void release_block_group(struct btrfs_block_group *block_group,
3858                                 struct find_free_extent_ctl *ffe_ctl,
3859                                 int delalloc)
3860 {
3861         switch (ffe_ctl->policy) {
3862         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3863                 ffe_ctl->retry_clustered = false;
3864                 ffe_ctl->retry_unclustered = false;
3865                 break;
3866         case BTRFS_EXTENT_ALLOC_ZONED:
3867                 /* Nothing to do */
3868                 break;
3869         default:
3870                 BUG();
3871         }
3872
3873         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3874                ffe_ctl->index);
3875         btrfs_release_block_group(block_group, delalloc);
3876 }
3877
3878 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3879                                    struct btrfs_key *ins)
3880 {
3881         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3882
3883         if (!ffe_ctl->use_cluster && last_ptr) {
3884                 spin_lock(&last_ptr->lock);
3885                 last_ptr->window_start = ins->objectid;
3886                 spin_unlock(&last_ptr->lock);
3887         }
3888 }
3889
3890 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3891                          struct btrfs_key *ins)
3892 {
3893         switch (ffe_ctl->policy) {
3894         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3895                 found_extent_clustered(ffe_ctl, ins);
3896                 break;
3897         case BTRFS_EXTENT_ALLOC_ZONED:
3898                 /* Nothing to do */
3899                 break;
3900         default:
3901                 BUG();
3902         }
3903 }
3904
3905 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
3906 {
3907         switch (ffe_ctl->policy) {
3908         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3909                 /*
3910                  * If we can't allocate a new chunk we've already looped through
3911                  * at least once, move on to the NO_EMPTY_SIZE case.
3912                  */
3913                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
3914                 return 0;
3915         case BTRFS_EXTENT_ALLOC_ZONED:
3916                 /* Give up here */
3917                 return -ENOSPC;
3918         default:
3919                 BUG();
3920         }
3921 }
3922
3923 /*
3924  * Return >0 means caller needs to re-search for free extent
3925  * Return 0 means we have the needed free extent.
3926  * Return <0 means we failed to locate any free extent.
3927  */
3928 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3929                                         struct btrfs_key *ins,
3930                                         struct find_free_extent_ctl *ffe_ctl,
3931                                         bool full_search)
3932 {
3933         struct btrfs_root *root = fs_info->extent_root;
3934         int ret;
3935
3936         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3937             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3938                 ffe_ctl->orig_have_caching_bg = true;
3939
3940         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
3941             ffe_ctl->have_caching_bg)
3942                 return 1;
3943
3944         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
3945                 return 1;
3946
3947         if (ins->objectid) {
3948                 found_extent(ffe_ctl, ins);
3949                 return 0;
3950         }
3951
3952         /*
3953          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
3954          *                      caching kthreads as we move along
3955          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
3956          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
3957          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
3958          *                     again
3959          */
3960         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
3961                 ffe_ctl->index = 0;
3962                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
3963                         /*
3964                          * We want to skip the LOOP_CACHING_WAIT step if we
3965                          * don't have any uncached bgs and we've already done a
3966                          * full search through.
3967                          */
3968                         if (ffe_ctl->orig_have_caching_bg || !full_search)
3969                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
3970                         else
3971                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
3972                 } else {
3973                         ffe_ctl->loop++;
3974                 }
3975
3976                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
3977                         struct btrfs_trans_handle *trans;
3978                         int exist = 0;
3979
3980                         trans = current->journal_info;
3981                         if (trans)
3982                                 exist = 1;
3983                         else
3984                                 trans = btrfs_join_transaction(root);
3985
3986                         if (IS_ERR(trans)) {
3987                                 ret = PTR_ERR(trans);
3988                                 return ret;
3989                         }
3990
3991                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
3992                                                 CHUNK_ALLOC_FORCE);
3993
3994                         /* Do not bail out on ENOSPC since we can do more. */
3995                         if (ret == -ENOSPC)
3996                                 ret = chunk_allocation_failed(ffe_ctl);
3997                         else if (ret < 0)
3998                                 btrfs_abort_transaction(trans, ret);
3999                         else
4000                                 ret = 0;
4001                         if (!exist)
4002                                 btrfs_end_transaction(trans);
4003                         if (ret)
4004                                 return ret;
4005                 }
4006
4007                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4008                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4009                                 return -ENOSPC;
4010
4011                         /*
4012                          * Don't loop again if we already have no empty_size and
4013                          * no empty_cluster.
4014                          */
4015                         if (ffe_ctl->empty_size == 0 &&
4016                             ffe_ctl->empty_cluster == 0)
4017                                 return -ENOSPC;
4018                         ffe_ctl->empty_size = 0;
4019                         ffe_ctl->empty_cluster = 0;
4020                 }
4021                 return 1;
4022         }
4023         return -ENOSPC;
4024 }
4025
4026 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4027                                         struct find_free_extent_ctl *ffe_ctl,
4028                                         struct btrfs_space_info *space_info,
4029                                         struct btrfs_key *ins)
4030 {
4031         /*
4032          * If our free space is heavily fragmented we may not be able to make
4033          * big contiguous allocations, so instead of doing the expensive search
4034          * for free space, simply return ENOSPC with our max_extent_size so we
4035          * can go ahead and search for a more manageable chunk.
4036          *
4037          * If our max_extent_size is large enough for our allocation simply
4038          * disable clustering since we will likely not be able to find enough
4039          * space to create a cluster and induce latency trying.
4040          */
4041         if (space_info->max_extent_size) {
4042                 spin_lock(&space_info->lock);
4043                 if (space_info->max_extent_size &&
4044                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4045                         ins->offset = space_info->max_extent_size;
4046                         spin_unlock(&space_info->lock);
4047                         return -ENOSPC;
4048                 } else if (space_info->max_extent_size) {
4049                         ffe_ctl->use_cluster = false;
4050                 }
4051                 spin_unlock(&space_info->lock);
4052         }
4053
4054         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4055                                                &ffe_ctl->empty_cluster);
4056         if (ffe_ctl->last_ptr) {
4057                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4058
4059                 spin_lock(&last_ptr->lock);
4060                 if (last_ptr->block_group)
4061                         ffe_ctl->hint_byte = last_ptr->window_start;
4062                 if (last_ptr->fragmented) {
4063                         /*
4064                          * We still set window_start so we can keep track of the
4065                          * last place we found an allocation to try and save
4066                          * some time.
4067                          */
4068                         ffe_ctl->hint_byte = last_ptr->window_start;
4069                         ffe_ctl->use_cluster = false;
4070                 }
4071                 spin_unlock(&last_ptr->lock);
4072         }
4073
4074         return 0;
4075 }
4076
4077 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4078                               struct find_free_extent_ctl *ffe_ctl,
4079                               struct btrfs_space_info *space_info,
4080                               struct btrfs_key *ins)
4081 {
4082         switch (ffe_ctl->policy) {
4083         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4084                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4085                                                     space_info, ins);
4086         case BTRFS_EXTENT_ALLOC_ZONED:
4087                 if (ffe_ctl->for_treelog) {
4088                         spin_lock(&fs_info->treelog_bg_lock);
4089                         if (fs_info->treelog_bg)
4090                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4091                         spin_unlock(&fs_info->treelog_bg_lock);
4092                 }
4093                 return 0;
4094         default:
4095                 BUG();
4096         }
4097 }
4098
4099 /*
4100  * walks the btree of allocated extents and find a hole of a given size.
4101  * The key ins is changed to record the hole:
4102  * ins->objectid == start position
4103  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4104  * ins->offset == the size of the hole.
4105  * Any available blocks before search_start are skipped.
4106  *
4107  * If there is no suitable free space, we will record the max size of
4108  * the free space extent currently.
4109  *
4110  * The overall logic and call chain:
4111  *
4112  * find_free_extent()
4113  * |- Iterate through all block groups
4114  * |  |- Get a valid block group
4115  * |  |- Try to do clustered allocation in that block group
4116  * |  |- Try to do unclustered allocation in that block group
4117  * |  |- Check if the result is valid
4118  * |  |  |- If valid, then exit
4119  * |  |- Jump to next block group
4120  * |
4121  * |- Push harder to find free extents
4122  *    |- If not found, re-iterate all block groups
4123  */
4124 static noinline int find_free_extent(struct btrfs_root *root,
4125                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
4126                                 u64 hint_byte_orig, struct btrfs_key *ins,
4127                                 u64 flags, int delalloc)
4128 {
4129         struct btrfs_fs_info *fs_info = root->fs_info;
4130         int ret = 0;
4131         int cache_block_group_error = 0;
4132         struct btrfs_block_group *block_group = NULL;
4133         struct find_free_extent_ctl ffe_ctl = {0};
4134         struct btrfs_space_info *space_info;
4135         bool full_search = false;
4136         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4137
4138         WARN_ON(num_bytes < fs_info->sectorsize);
4139
4140         ffe_ctl.num_bytes = num_bytes;
4141         ffe_ctl.empty_size = empty_size;
4142         ffe_ctl.flags = flags;
4143         ffe_ctl.search_start = 0;
4144         ffe_ctl.delalloc = delalloc;
4145         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
4146         ffe_ctl.have_caching_bg = false;
4147         ffe_ctl.orig_have_caching_bg = false;
4148         ffe_ctl.found_offset = 0;
4149         ffe_ctl.hint_byte = hint_byte_orig;
4150         ffe_ctl.for_treelog = for_treelog;
4151         ffe_ctl.policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4152
4153         /* For clustered allocation */
4154         ffe_ctl.retry_clustered = false;
4155         ffe_ctl.retry_unclustered = false;
4156         ffe_ctl.last_ptr = NULL;
4157         ffe_ctl.use_cluster = true;
4158
4159         if (btrfs_is_zoned(fs_info))
4160                 ffe_ctl.policy = BTRFS_EXTENT_ALLOC_ZONED;
4161
4162         ins->type = BTRFS_EXTENT_ITEM_KEY;
4163         ins->objectid = 0;
4164         ins->offset = 0;
4165
4166         trace_find_free_extent(root, num_bytes, empty_size, flags);
4167
4168         space_info = btrfs_find_space_info(fs_info, flags);
4169         if (!space_info) {
4170                 btrfs_err(fs_info, "No space info for %llu", flags);
4171                 return -ENOSPC;
4172         }
4173
4174         ret = prepare_allocation(fs_info, &ffe_ctl, space_info, ins);
4175         if (ret < 0)
4176                 return ret;
4177
4178         ffe_ctl.search_start = max(ffe_ctl.search_start,
4179                                    first_logical_byte(fs_info, 0));
4180         ffe_ctl.search_start = max(ffe_ctl.search_start, ffe_ctl.hint_byte);
4181         if (ffe_ctl.search_start == ffe_ctl.hint_byte) {
4182                 block_group = btrfs_lookup_block_group(fs_info,
4183                                                        ffe_ctl.search_start);
4184                 /*
4185                  * we don't want to use the block group if it doesn't match our
4186                  * allocation bits, or if its not cached.
4187                  *
4188                  * However if we are re-searching with an ideal block group
4189                  * picked out then we don't care that the block group is cached.
4190                  */
4191                 if (block_group && block_group_bits(block_group, flags) &&
4192                     block_group->cached != BTRFS_CACHE_NO) {
4193                         down_read(&space_info->groups_sem);
4194                         if (list_empty(&block_group->list) ||
4195                             block_group->ro) {
4196                                 /*
4197                                  * someone is removing this block group,
4198                                  * we can't jump into the have_block_group
4199                                  * target because our list pointers are not
4200                                  * valid
4201                                  */
4202                                 btrfs_put_block_group(block_group);
4203                                 up_read(&space_info->groups_sem);
4204                         } else {
4205                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
4206                                                 block_group->flags);
4207                                 btrfs_lock_block_group(block_group, delalloc);
4208                                 goto have_block_group;
4209                         }
4210                 } else if (block_group) {
4211                         btrfs_put_block_group(block_group);
4212                 }
4213         }
4214 search:
4215         ffe_ctl.have_caching_bg = false;
4216         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
4217             ffe_ctl.index == 0)
4218                 full_search = true;
4219         down_read(&space_info->groups_sem);
4220         list_for_each_entry(block_group,
4221                             &space_info->block_groups[ffe_ctl.index], list) {
4222                 struct btrfs_block_group *bg_ret;
4223
4224                 /* If the block group is read-only, we can skip it entirely. */
4225                 if (unlikely(block_group->ro)) {
4226                         if (for_treelog)
4227                                 btrfs_clear_treelog_bg(block_group);
4228                         continue;
4229                 }
4230
4231                 btrfs_grab_block_group(block_group, delalloc);
4232                 ffe_ctl.search_start = block_group->start;
4233
4234                 /*
4235                  * this can happen if we end up cycling through all the
4236                  * raid types, but we want to make sure we only allocate
4237                  * for the proper type.
4238                  */
4239                 if (!block_group_bits(block_group, flags)) {
4240                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4241                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4242                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4243                                 BTRFS_BLOCK_GROUP_RAID10;
4244
4245                         /*
4246                          * if they asked for extra copies and this block group
4247                          * doesn't provide them, bail.  This does allow us to
4248                          * fill raid0 from raid1.
4249                          */
4250                         if ((flags & extra) && !(block_group->flags & extra))
4251                                 goto loop;
4252
4253                         /*
4254                          * This block group has different flags than we want.
4255                          * It's possible that we have MIXED_GROUP flag but no
4256                          * block group is mixed.  Just skip such block group.
4257                          */
4258                         btrfs_release_block_group(block_group, delalloc);
4259                         continue;
4260                 }
4261
4262 have_block_group:
4263                 ffe_ctl.cached = btrfs_block_group_done(block_group);
4264                 if (unlikely(!ffe_ctl.cached)) {
4265                         ffe_ctl.have_caching_bg = true;
4266                         ret = btrfs_cache_block_group(block_group, 0);
4267
4268                         /*
4269                          * If we get ENOMEM here or something else we want to
4270                          * try other block groups, because it may not be fatal.
4271                          * However if we can't find anything else we need to
4272                          * save our return here so that we return the actual
4273                          * error that caused problems, not ENOSPC.
4274                          */
4275                         if (ret < 0) {
4276                                 if (!cache_block_group_error)
4277                                         cache_block_group_error = ret;
4278                                 ret = 0;
4279                                 goto loop;
4280                         }
4281                         ret = 0;
4282                 }
4283
4284                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4285                         goto loop;
4286
4287                 bg_ret = NULL;
4288                 ret = do_allocation(block_group, &ffe_ctl, &bg_ret);
4289                 if (ret == 0) {
4290                         if (bg_ret && bg_ret != block_group) {
4291                                 btrfs_release_block_group(block_group, delalloc);
4292                                 block_group = bg_ret;
4293                         }
4294                 } else if (ret == -EAGAIN) {
4295                         goto have_block_group;
4296                 } else if (ret > 0) {
4297                         goto loop;
4298                 }
4299
4300                 /* Checks */
4301                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
4302                                              fs_info->stripesize);
4303
4304                 /* move on to the next group */
4305                 if (ffe_ctl.search_start + num_bytes >
4306                     block_group->start + block_group->length) {
4307                         btrfs_add_free_space_unused(block_group,
4308                                             ffe_ctl.found_offset, num_bytes);
4309                         goto loop;
4310                 }
4311
4312                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
4313                         btrfs_add_free_space_unused(block_group,
4314                                         ffe_ctl.found_offset,
4315                                         ffe_ctl.search_start - ffe_ctl.found_offset);
4316
4317                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
4318                                 num_bytes, delalloc);
4319                 if (ret == -EAGAIN) {
4320                         btrfs_add_free_space_unused(block_group,
4321                                         ffe_ctl.found_offset, num_bytes);
4322                         goto loop;
4323                 }
4324                 btrfs_inc_block_group_reservations(block_group);
4325
4326                 /* we are all good, lets return */
4327                 ins->objectid = ffe_ctl.search_start;
4328                 ins->offset = num_bytes;
4329
4330                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
4331                                            num_bytes);
4332                 btrfs_release_block_group(block_group, delalloc);
4333                 break;
4334 loop:
4335                 release_block_group(block_group, &ffe_ctl, delalloc);
4336                 cond_resched();
4337         }
4338         up_read(&space_info->groups_sem);
4339
4340         ret = find_free_extent_update_loop(fs_info, ins, &ffe_ctl, full_search);
4341         if (ret > 0)
4342                 goto search;
4343
4344         if (ret == -ENOSPC && !cache_block_group_error) {
4345                 /*
4346                  * Use ffe_ctl->total_free_space as fallback if we can't find
4347                  * any contiguous hole.
4348                  */
4349                 if (!ffe_ctl.max_extent_size)
4350                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
4351                 spin_lock(&space_info->lock);
4352                 space_info->max_extent_size = ffe_ctl.max_extent_size;
4353                 spin_unlock(&space_info->lock);
4354                 ins->offset = ffe_ctl.max_extent_size;
4355         } else if (ret == -ENOSPC) {
4356                 ret = cache_block_group_error;
4357         }
4358         return ret;
4359 }
4360
4361 /*
4362  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4363  *                        hole that is at least as big as @num_bytes.
4364  *
4365  * @root           -    The root that will contain this extent
4366  *
4367  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4368  *                      is used for accounting purposes. This value differs
4369  *                      from @num_bytes only in the case of compressed extents.
4370  *
4371  * @num_bytes      -    Number of bytes to allocate on-disk.
4372  *
4373  * @min_alloc_size -    Indicates the minimum amount of space that the
4374  *                      allocator should try to satisfy. In some cases
4375  *                      @num_bytes may be larger than what is required and if
4376  *                      the filesystem is fragmented then allocation fails.
4377  *                      However, the presence of @min_alloc_size gives a
4378  *                      chance to try and satisfy the smaller allocation.
4379  *
4380  * @empty_size     -    A hint that you plan on doing more COW. This is the
4381  *                      size in bytes the allocator should try to find free
4382  *                      next to the block it returns.  This is just a hint and
4383  *                      may be ignored by the allocator.
4384  *
4385  * @hint_byte      -    Hint to the allocator to start searching above the byte
4386  *                      address passed. It might be ignored.
4387  *
4388  * @ins            -    This key is modified to record the found hole. It will
4389  *                      have the following values:
4390  *                      ins->objectid == start position
4391  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4392  *                      ins->offset == the size of the hole.
4393  *
4394  * @is_data        -    Boolean flag indicating whether an extent is
4395  *                      allocated for data (true) or metadata (false)
4396  *
4397  * @delalloc       -    Boolean flag indicating whether this allocation is for
4398  *                      delalloc or not. If 'true' data_rwsem of block groups
4399  *                      is going to be acquired.
4400  *
4401  *
4402  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4403  * case -ENOSPC is returned then @ins->offset will contain the size of the
4404  * largest available hole the allocator managed to find.
4405  */
4406 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4407                          u64 num_bytes, u64 min_alloc_size,
4408                          u64 empty_size, u64 hint_byte,
4409                          struct btrfs_key *ins, int is_data, int delalloc)
4410 {
4411         struct btrfs_fs_info *fs_info = root->fs_info;
4412         bool final_tried = num_bytes == min_alloc_size;
4413         u64 flags;
4414         int ret;
4415         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4416
4417         flags = get_alloc_profile_by_root(root, is_data);
4418 again:
4419         WARN_ON(num_bytes < fs_info->sectorsize);
4420         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
4421                                hint_byte, ins, flags, delalloc);
4422         if (!ret && !is_data) {
4423                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4424         } else if (ret == -ENOSPC) {
4425                 if (!final_tried && ins->offset) {
4426                         num_bytes = min(num_bytes >> 1, ins->offset);
4427                         num_bytes = round_down(num_bytes,
4428                                                fs_info->sectorsize);
4429                         num_bytes = max(num_bytes, min_alloc_size);
4430                         ram_bytes = num_bytes;
4431                         if (num_bytes == min_alloc_size)
4432                                 final_tried = true;
4433                         goto again;
4434                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4435                         struct btrfs_space_info *sinfo;
4436
4437                         sinfo = btrfs_find_space_info(fs_info, flags);
4438                         btrfs_err(fs_info,
4439                         "allocation failed flags %llu, wanted %llu tree-log %d",
4440                                   flags, num_bytes, for_treelog);
4441                         if (sinfo)
4442                                 btrfs_dump_space_info(fs_info, sinfo,
4443                                                       num_bytes, 1);
4444                 }
4445         }
4446
4447         return ret;
4448 }
4449
4450 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4451                                u64 start, u64 len, int delalloc)
4452 {
4453         struct btrfs_block_group *cache;
4454
4455         cache = btrfs_lookup_block_group(fs_info, start);
4456         if (!cache) {
4457                 btrfs_err(fs_info, "Unable to find block group for %llu",
4458                           start);
4459                 return -ENOSPC;
4460         }
4461
4462         btrfs_add_free_space(cache, start, len);
4463         btrfs_free_reserved_bytes(cache, len, delalloc);
4464         trace_btrfs_reserved_extent_free(fs_info, start, len);
4465
4466         btrfs_put_block_group(cache);
4467         return 0;
4468 }
4469
4470 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4471                               u64 len)
4472 {
4473         struct btrfs_block_group *cache;
4474         int ret = 0;
4475
4476         cache = btrfs_lookup_block_group(trans->fs_info, start);
4477         if (!cache) {
4478                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4479                           start);
4480                 return -ENOSPC;
4481         }
4482
4483         ret = pin_down_extent(trans, cache, start, len, 1);
4484         btrfs_put_block_group(cache);
4485         return ret;
4486 }
4487
4488 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4489                                       u64 parent, u64 root_objectid,
4490                                       u64 flags, u64 owner, u64 offset,
4491                                       struct btrfs_key *ins, int ref_mod)
4492 {
4493         struct btrfs_fs_info *fs_info = trans->fs_info;
4494         int ret;
4495         struct btrfs_extent_item *extent_item;
4496         struct btrfs_extent_inline_ref *iref;
4497         struct btrfs_path *path;
4498         struct extent_buffer *leaf;
4499         int type;
4500         u32 size;
4501
4502         if (parent > 0)
4503                 type = BTRFS_SHARED_DATA_REF_KEY;
4504         else
4505                 type = BTRFS_EXTENT_DATA_REF_KEY;
4506
4507         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4508
4509         path = btrfs_alloc_path();
4510         if (!path)
4511                 return -ENOMEM;
4512
4513         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4514                                       ins, size);
4515         if (ret) {
4516                 btrfs_free_path(path);
4517                 return ret;
4518         }
4519
4520         leaf = path->nodes[0];
4521         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4522                                      struct btrfs_extent_item);
4523         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4524         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4525         btrfs_set_extent_flags(leaf, extent_item,
4526                                flags | BTRFS_EXTENT_FLAG_DATA);
4527
4528         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4529         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4530         if (parent > 0) {
4531                 struct btrfs_shared_data_ref *ref;
4532                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4533                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4534                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4535         } else {
4536                 struct btrfs_extent_data_ref *ref;
4537                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4538                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4539                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4540                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4541                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4542         }
4543
4544         btrfs_mark_buffer_dirty(path->nodes[0]);
4545         btrfs_free_path(path);
4546
4547         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
4548         if (ret)
4549                 return ret;
4550
4551         ret = btrfs_update_block_group(trans, ins->objectid, ins->offset, 1);
4552         if (ret) { /* -ENOENT, logic error */
4553                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4554                         ins->objectid, ins->offset);
4555                 BUG();
4556         }
4557         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
4558         return ret;
4559 }
4560
4561 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4562                                      struct btrfs_delayed_ref_node *node,
4563                                      struct btrfs_delayed_extent_op *extent_op)
4564 {
4565         struct btrfs_fs_info *fs_info = trans->fs_info;
4566         int ret;
4567         struct btrfs_extent_item *extent_item;
4568         struct btrfs_key extent_key;
4569         struct btrfs_tree_block_info *block_info;
4570         struct btrfs_extent_inline_ref *iref;
4571         struct btrfs_path *path;
4572         struct extent_buffer *leaf;
4573         struct btrfs_delayed_tree_ref *ref;
4574         u32 size = sizeof(*extent_item) + sizeof(*iref);
4575         u64 num_bytes;
4576         u64 flags = extent_op->flags_to_set;
4577         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4578
4579         ref = btrfs_delayed_node_to_tree_ref(node);
4580
4581         extent_key.objectid = node->bytenr;
4582         if (skinny_metadata) {
4583                 extent_key.offset = ref->level;
4584                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4585                 num_bytes = fs_info->nodesize;
4586         } else {
4587                 extent_key.offset = node->num_bytes;
4588                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4589                 size += sizeof(*block_info);
4590                 num_bytes = node->num_bytes;
4591         }
4592
4593         path = btrfs_alloc_path();
4594         if (!path)
4595                 return -ENOMEM;
4596
4597         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4598                                       &extent_key, size);
4599         if (ret) {
4600                 btrfs_free_path(path);
4601                 return ret;
4602         }
4603
4604         leaf = path->nodes[0];
4605         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4606                                      struct btrfs_extent_item);
4607         btrfs_set_extent_refs(leaf, extent_item, 1);
4608         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4609         btrfs_set_extent_flags(leaf, extent_item,
4610                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4611
4612         if (skinny_metadata) {
4613                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4614         } else {
4615                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4616                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4617                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4618                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4619         }
4620
4621         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4622                 btrfs_set_extent_inline_ref_type(leaf, iref,
4623                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4624                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4625         } else {
4626                 btrfs_set_extent_inline_ref_type(leaf, iref,
4627                                                  BTRFS_TREE_BLOCK_REF_KEY);
4628                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4629         }
4630
4631         btrfs_mark_buffer_dirty(leaf);
4632         btrfs_free_path(path);
4633
4634         ret = remove_from_free_space_tree(trans, extent_key.objectid,
4635                                           num_bytes);
4636         if (ret)
4637                 return ret;
4638
4639         ret = btrfs_update_block_group(trans, extent_key.objectid,
4640                                        fs_info->nodesize, 1);
4641         if (ret) { /* -ENOENT, logic error */
4642                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4643                         extent_key.objectid, extent_key.offset);
4644                 BUG();
4645         }
4646
4647         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
4648                                           fs_info->nodesize);
4649         return ret;
4650 }
4651
4652 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4653                                      struct btrfs_root *root, u64 owner,
4654                                      u64 offset, u64 ram_bytes,
4655                                      struct btrfs_key *ins)
4656 {
4657         struct btrfs_ref generic_ref = { 0 };
4658
4659         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4660
4661         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4662                                ins->objectid, ins->offset, 0);
4663         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
4664         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4665
4666         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4667 }
4668
4669 /*
4670  * this is used by the tree logging recovery code.  It records that
4671  * an extent has been allocated and makes sure to clear the free
4672  * space cache bits as well
4673  */
4674 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4675                                    u64 root_objectid, u64 owner, u64 offset,
4676                                    struct btrfs_key *ins)
4677 {
4678         struct btrfs_fs_info *fs_info = trans->fs_info;
4679         int ret;
4680         struct btrfs_block_group *block_group;
4681         struct btrfs_space_info *space_info;
4682
4683         /*
4684          * Mixed block groups will exclude before processing the log so we only
4685          * need to do the exclude dance if this fs isn't mixed.
4686          */
4687         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4688                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4689                                               ins->offset);
4690                 if (ret)
4691                         return ret;
4692         }
4693
4694         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4695         if (!block_group)
4696                 return -EINVAL;
4697
4698         space_info = block_group->space_info;
4699         spin_lock(&space_info->lock);
4700         spin_lock(&block_group->lock);
4701         space_info->bytes_reserved += ins->offset;
4702         block_group->reserved += ins->offset;
4703         spin_unlock(&block_group->lock);
4704         spin_unlock(&space_info->lock);
4705
4706         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4707                                          offset, ins, 1);
4708         if (ret)
4709                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4710         btrfs_put_block_group(block_group);
4711         return ret;
4712 }
4713
4714 static struct extent_buffer *
4715 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4716                       u64 bytenr, int level, u64 owner,
4717                       enum btrfs_lock_nesting nest)
4718 {
4719         struct btrfs_fs_info *fs_info = root->fs_info;
4720         struct extent_buffer *buf;
4721
4722         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4723         if (IS_ERR(buf))
4724                 return buf;
4725
4726         /*
4727          * Extra safety check in case the extent tree is corrupted and extent
4728          * allocator chooses to use a tree block which is already used and
4729          * locked.
4730          */
4731         if (buf->lock_owner == current->pid) {
4732                 btrfs_err_rl(fs_info,
4733 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4734                         buf->start, btrfs_header_owner(buf), current->pid);
4735                 free_extent_buffer(buf);
4736                 return ERR_PTR(-EUCLEAN);
4737         }
4738
4739         /*
4740          * This needs to stay, because we could allocate a freed block from an
4741          * old tree into a new tree, so we need to make sure this new block is
4742          * set to the appropriate level and owner.
4743          */
4744         btrfs_set_buffer_lockdep_class(owner, buf, level);
4745         __btrfs_tree_lock(buf, nest);
4746         btrfs_clean_tree_block(buf);
4747         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4748         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4749
4750         set_extent_buffer_uptodate(buf);
4751
4752         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4753         btrfs_set_header_level(buf, level);
4754         btrfs_set_header_bytenr(buf, buf->start);
4755         btrfs_set_header_generation(buf, trans->transid);
4756         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4757         btrfs_set_header_owner(buf, owner);
4758         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4759         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4760         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4761                 buf->log_index = root->log_transid % 2;
4762                 /*
4763                  * we allow two log transactions at a time, use different
4764                  * EXTENT bit to differentiate dirty pages.
4765                  */
4766                 if (buf->log_index == 0)
4767                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4768                                         buf->start + buf->len - 1, GFP_NOFS);
4769                 else
4770                         set_extent_new(&root->dirty_log_pages, buf->start,
4771                                         buf->start + buf->len - 1);
4772         } else {
4773                 buf->log_index = -1;
4774                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4775                          buf->start + buf->len - 1, GFP_NOFS);
4776         }
4777         trans->dirty = true;
4778         /* this returns a buffer locked for blocking */
4779         return buf;
4780 }
4781
4782 /*
4783  * finds a free extent and does all the dirty work required for allocation
4784  * returns the tree buffer or an ERR_PTR on error.
4785  */
4786 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4787                                              struct btrfs_root *root,
4788                                              u64 parent, u64 root_objectid,
4789                                              const struct btrfs_disk_key *key,
4790                                              int level, u64 hint,
4791                                              u64 empty_size,
4792                                              enum btrfs_lock_nesting nest)
4793 {
4794         struct btrfs_fs_info *fs_info = root->fs_info;
4795         struct btrfs_key ins;
4796         struct btrfs_block_rsv *block_rsv;
4797         struct extent_buffer *buf;
4798         struct btrfs_delayed_extent_op *extent_op;
4799         struct btrfs_ref generic_ref = { 0 };
4800         u64 flags = 0;
4801         int ret;
4802         u32 blocksize = fs_info->nodesize;
4803         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4804
4805 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4806         if (btrfs_is_testing(fs_info)) {
4807                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4808                                             level, root_objectid, nest);
4809                 if (!IS_ERR(buf))
4810                         root->alloc_bytenr += blocksize;
4811                 return buf;
4812         }
4813 #endif
4814
4815         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4816         if (IS_ERR(block_rsv))
4817                 return ERR_CAST(block_rsv);
4818
4819         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4820                                    empty_size, hint, &ins, 0, 0);
4821         if (ret)
4822                 goto out_unuse;
4823
4824         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4825                                     root_objectid, nest);
4826         if (IS_ERR(buf)) {
4827                 ret = PTR_ERR(buf);
4828                 goto out_free_reserved;
4829         }
4830
4831         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4832                 if (parent == 0)
4833                         parent = ins.objectid;
4834                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4835         } else
4836                 BUG_ON(parent > 0);
4837
4838         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4839                 extent_op = btrfs_alloc_delayed_extent_op();
4840                 if (!extent_op) {
4841                         ret = -ENOMEM;
4842                         goto out_free_buf;
4843                 }
4844                 if (key)
4845                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4846                 else
4847                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4848                 extent_op->flags_to_set = flags;
4849                 extent_op->update_key = skinny_metadata ? false : true;
4850                 extent_op->update_flags = true;
4851                 extent_op->is_data = false;
4852                 extent_op->level = level;
4853
4854                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4855                                        ins.objectid, ins.offset, parent);
4856                 generic_ref.real_root = root->root_key.objectid;
4857                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
4858                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4859                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
4860                 if (ret)
4861                         goto out_free_delayed;
4862         }
4863         return buf;
4864
4865 out_free_delayed:
4866         btrfs_free_delayed_extent_op(extent_op);
4867 out_free_buf:
4868         free_extent_buffer(buf);
4869 out_free_reserved:
4870         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4871 out_unuse:
4872         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4873         return ERR_PTR(ret);
4874 }
4875
4876 struct walk_control {
4877         u64 refs[BTRFS_MAX_LEVEL];
4878         u64 flags[BTRFS_MAX_LEVEL];
4879         struct btrfs_key update_progress;
4880         struct btrfs_key drop_progress;
4881         int drop_level;
4882         int stage;
4883         int level;
4884         int shared_level;
4885         int update_ref;
4886         int keep_locks;
4887         int reada_slot;
4888         int reada_count;
4889         int restarted;
4890 };
4891
4892 #define DROP_REFERENCE  1
4893 #define UPDATE_BACKREF  2
4894
4895 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
4896                                      struct btrfs_root *root,
4897                                      struct walk_control *wc,
4898                                      struct btrfs_path *path)
4899 {
4900         struct btrfs_fs_info *fs_info = root->fs_info;
4901         u64 bytenr;
4902         u64 generation;
4903         u64 refs;
4904         u64 flags;
4905         u32 nritems;
4906         struct btrfs_key key;
4907         struct extent_buffer *eb;
4908         int ret;
4909         int slot;
4910         int nread = 0;
4911
4912         if (path->slots[wc->level] < wc->reada_slot) {
4913                 wc->reada_count = wc->reada_count * 2 / 3;
4914                 wc->reada_count = max(wc->reada_count, 2);
4915         } else {
4916                 wc->reada_count = wc->reada_count * 3 / 2;
4917                 wc->reada_count = min_t(int, wc->reada_count,
4918                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
4919         }
4920
4921         eb = path->nodes[wc->level];
4922         nritems = btrfs_header_nritems(eb);
4923
4924         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
4925                 if (nread >= wc->reada_count)
4926                         break;
4927
4928                 cond_resched();
4929                 bytenr = btrfs_node_blockptr(eb, slot);
4930                 generation = btrfs_node_ptr_generation(eb, slot);
4931
4932                 if (slot == path->slots[wc->level])
4933                         goto reada;
4934
4935                 if (wc->stage == UPDATE_BACKREF &&
4936                     generation <= root->root_key.offset)
4937                         continue;
4938
4939                 /* We don't lock the tree block, it's OK to be racy here */
4940                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
4941                                                wc->level - 1, 1, &refs,
4942                                                &flags);
4943                 /* We don't care about errors in readahead. */
4944                 if (ret < 0)
4945                         continue;
4946                 BUG_ON(refs == 0);
4947
4948                 if (wc->stage == DROP_REFERENCE) {
4949                         if (refs == 1)
4950                                 goto reada;
4951
4952                         if (wc->level == 1 &&
4953                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4954                                 continue;
4955                         if (!wc->update_ref ||
4956                             generation <= root->root_key.offset)
4957                                 continue;
4958                         btrfs_node_key_to_cpu(eb, &key, slot);
4959                         ret = btrfs_comp_cpu_keys(&key,
4960                                                   &wc->update_progress);
4961                         if (ret < 0)
4962                                 continue;
4963                 } else {
4964                         if (wc->level == 1 &&
4965                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4966                                 continue;
4967                 }
4968 reada:
4969                 btrfs_readahead_node_child(eb, slot);
4970                 nread++;
4971         }
4972         wc->reada_slot = slot;
4973 }
4974
4975 /*
4976  * helper to process tree block while walking down the tree.
4977  *
4978  * when wc->stage == UPDATE_BACKREF, this function updates
4979  * back refs for pointers in the block.
4980  *
4981  * NOTE: return value 1 means we should stop walking down.
4982  */
4983 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
4984                                    struct btrfs_root *root,
4985                                    struct btrfs_path *path,
4986                                    struct walk_control *wc, int lookup_info)
4987 {
4988         struct btrfs_fs_info *fs_info = root->fs_info;
4989         int level = wc->level;
4990         struct extent_buffer *eb = path->nodes[level];
4991         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
4992         int ret;
4993
4994         if (wc->stage == UPDATE_BACKREF &&
4995             btrfs_header_owner(eb) != root->root_key.objectid)
4996                 return 1;
4997
4998         /*
4999          * when reference count of tree block is 1, it won't increase
5000          * again. once full backref flag is set, we never clear it.
5001          */
5002         if (lookup_info &&
5003             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5004              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5005                 BUG_ON(!path->locks[level]);
5006                 ret = btrfs_lookup_extent_info(trans, fs_info,
5007                                                eb->start, level, 1,
5008                                                &wc->refs[level],
5009                                                &wc->flags[level]);
5010                 BUG_ON(ret == -ENOMEM);
5011                 if (ret)
5012                         return ret;
5013                 BUG_ON(wc->refs[level] == 0);
5014         }
5015
5016         if (wc->stage == DROP_REFERENCE) {
5017                 if (wc->refs[level] > 1)
5018                         return 1;
5019
5020                 if (path->locks[level] && !wc->keep_locks) {
5021                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5022                         path->locks[level] = 0;
5023                 }
5024                 return 0;
5025         }
5026
5027         /* wc->stage == UPDATE_BACKREF */
5028         if (!(wc->flags[level] & flag)) {
5029                 BUG_ON(!path->locks[level]);
5030                 ret = btrfs_inc_ref(trans, root, eb, 1);
5031                 BUG_ON(ret); /* -ENOMEM */
5032                 ret = btrfs_dec_ref(trans, root, eb, 0);
5033                 BUG_ON(ret); /* -ENOMEM */
5034                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5035                                                   btrfs_header_level(eb), 0);
5036                 BUG_ON(ret); /* -ENOMEM */
5037                 wc->flags[level] |= flag;
5038         }
5039
5040         /*
5041          * the block is shared by multiple trees, so it's not good to
5042          * keep the tree lock
5043          */
5044         if (path->locks[level] && level > 0) {
5045                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5046                 path->locks[level] = 0;
5047         }
5048         return 0;
5049 }
5050
5051 /*
5052  * This is used to verify a ref exists for this root to deal with a bug where we
5053  * would have a drop_progress key that hadn't been updated properly.
5054  */
5055 static int check_ref_exists(struct btrfs_trans_handle *trans,
5056                             struct btrfs_root *root, u64 bytenr, u64 parent,
5057                             int level)
5058 {
5059         struct btrfs_path *path;
5060         struct btrfs_extent_inline_ref *iref;
5061         int ret;
5062
5063         path = btrfs_alloc_path();
5064         if (!path)
5065                 return -ENOMEM;
5066
5067         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5068                                     root->fs_info->nodesize, parent,
5069                                     root->root_key.objectid, level, 0);
5070         btrfs_free_path(path);
5071         if (ret == -ENOENT)
5072                 return 0;
5073         if (ret < 0)
5074                 return ret;
5075         return 1;
5076 }
5077
5078 /*
5079  * helper to process tree block pointer.
5080  *
5081  * when wc->stage == DROP_REFERENCE, this function checks
5082  * reference count of the block pointed to. if the block
5083  * is shared and we need update back refs for the subtree
5084  * rooted at the block, this function changes wc->stage to
5085  * UPDATE_BACKREF. if the block is shared and there is no
5086  * need to update back, this function drops the reference
5087  * to the block.
5088  *
5089  * NOTE: return value 1 means we should stop walking down.
5090  */
5091 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5092                                  struct btrfs_root *root,
5093                                  struct btrfs_path *path,
5094                                  struct walk_control *wc, int *lookup_info)
5095 {
5096         struct btrfs_fs_info *fs_info = root->fs_info;
5097         u64 bytenr;
5098         u64 generation;
5099         u64 parent;
5100         struct btrfs_key key;
5101         struct btrfs_key first_key;
5102         struct btrfs_ref ref = { 0 };
5103         struct extent_buffer *next;
5104         int level = wc->level;
5105         int reada = 0;
5106         int ret = 0;
5107         bool need_account = false;
5108
5109         generation = btrfs_node_ptr_generation(path->nodes[level],
5110                                                path->slots[level]);
5111         /*
5112          * if the lower level block was created before the snapshot
5113          * was created, we know there is no need to update back refs
5114          * for the subtree
5115          */
5116         if (wc->stage == UPDATE_BACKREF &&
5117             generation <= root->root_key.offset) {
5118                 *lookup_info = 1;
5119                 return 1;
5120         }
5121
5122         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5123         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
5124                               path->slots[level]);
5125
5126         next = find_extent_buffer(fs_info, bytenr);
5127         if (!next) {
5128                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5129                                 root->root_key.objectid, level - 1);
5130                 if (IS_ERR(next))
5131                         return PTR_ERR(next);
5132                 reada = 1;
5133         }
5134         btrfs_tree_lock(next);
5135
5136         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5137                                        &wc->refs[level - 1],
5138                                        &wc->flags[level - 1]);
5139         if (ret < 0)
5140                 goto out_unlock;
5141
5142         if (unlikely(wc->refs[level - 1] == 0)) {
5143                 btrfs_err(fs_info, "Missing references.");
5144                 ret = -EIO;
5145                 goto out_unlock;
5146         }
5147         *lookup_info = 0;
5148
5149         if (wc->stage == DROP_REFERENCE) {
5150                 if (wc->refs[level - 1] > 1) {
5151                         need_account = true;
5152                         if (level == 1 &&
5153                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5154                                 goto skip;
5155
5156                         if (!wc->update_ref ||
5157                             generation <= root->root_key.offset)
5158                                 goto skip;
5159
5160                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5161                                               path->slots[level]);
5162                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5163                         if (ret < 0)
5164                                 goto skip;
5165
5166                         wc->stage = UPDATE_BACKREF;
5167                         wc->shared_level = level - 1;
5168                 }
5169         } else {
5170                 if (level == 1 &&
5171                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5172                         goto skip;
5173         }
5174
5175         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5176                 btrfs_tree_unlock(next);
5177                 free_extent_buffer(next);
5178                 next = NULL;
5179                 *lookup_info = 1;
5180         }
5181
5182         if (!next) {
5183                 if (reada && level == 1)
5184                         reada_walk_down(trans, root, wc, path);
5185                 next = read_tree_block(fs_info, bytenr, root->root_key.objectid,
5186                                        generation, level - 1, &first_key);
5187                 if (IS_ERR(next)) {
5188                         return PTR_ERR(next);
5189                 } else if (!extent_buffer_uptodate(next)) {
5190                         free_extent_buffer(next);
5191                         return -EIO;
5192                 }
5193                 btrfs_tree_lock(next);
5194         }
5195
5196         level--;
5197         ASSERT(level == btrfs_header_level(next));
5198         if (level != btrfs_header_level(next)) {
5199                 btrfs_err(root->fs_info, "mismatched level");
5200                 ret = -EIO;
5201                 goto out_unlock;
5202         }
5203         path->nodes[level] = next;
5204         path->slots[level] = 0;
5205         path->locks[level] = BTRFS_WRITE_LOCK;
5206         wc->level = level;
5207         if (wc->level == 1)
5208                 wc->reada_slot = 0;
5209         return 0;
5210 skip:
5211         wc->refs[level - 1] = 0;
5212         wc->flags[level - 1] = 0;
5213         if (wc->stage == DROP_REFERENCE) {
5214                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5215                         parent = path->nodes[level]->start;
5216                 } else {
5217                         ASSERT(root->root_key.objectid ==
5218                                btrfs_header_owner(path->nodes[level]));
5219                         if (root->root_key.objectid !=
5220                             btrfs_header_owner(path->nodes[level])) {
5221                                 btrfs_err(root->fs_info,
5222                                                 "mismatched block owner");
5223                                 ret = -EIO;
5224                                 goto out_unlock;
5225                         }
5226                         parent = 0;
5227                 }
5228
5229                 /*
5230                  * If we had a drop_progress we need to verify the refs are set
5231                  * as expected.  If we find our ref then we know that from here
5232                  * on out everything should be correct, and we can clear the
5233                  * ->restarted flag.
5234                  */
5235                 if (wc->restarted) {
5236                         ret = check_ref_exists(trans, root, bytenr, parent,
5237                                                level - 1);
5238                         if (ret < 0)
5239                                 goto out_unlock;
5240                         if (ret == 0)
5241                                 goto no_delete;
5242                         ret = 0;
5243                         wc->restarted = 0;
5244                 }
5245
5246                 /*
5247                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5248                  * already accounted them at merge time (replace_path),
5249                  * thus we could skip expensive subtree trace here.
5250                  */
5251                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5252                     need_account) {
5253                         ret = btrfs_qgroup_trace_subtree(trans, next,
5254                                                          generation, level - 1);
5255                         if (ret) {
5256                                 btrfs_err_rl(fs_info,
5257                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5258                                              ret);
5259                         }
5260                 }
5261
5262                 /*
5263                  * We need to update the next key in our walk control so we can
5264                  * update the drop_progress key accordingly.  We don't care if
5265                  * find_next_key doesn't find a key because that means we're at
5266                  * the end and are going to clean up now.
5267                  */
5268                 wc->drop_level = level;
5269                 find_next_key(path, level, &wc->drop_progress);
5270
5271                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5272                                        fs_info->nodesize, parent);
5273                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
5274                 ret = btrfs_free_extent(trans, &ref);
5275                 if (ret)
5276                         goto out_unlock;
5277         }
5278 no_delete:
5279         *lookup_info = 1;
5280         ret = 1;
5281
5282 out_unlock:
5283         btrfs_tree_unlock(next);
5284         free_extent_buffer(next);
5285
5286         return ret;
5287 }
5288
5289 /*
5290  * helper to process tree block while walking up the tree.
5291  *
5292  * when wc->stage == DROP_REFERENCE, this function drops
5293  * reference count on the block.
5294  *
5295  * when wc->stage == UPDATE_BACKREF, this function changes
5296  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5297  * to UPDATE_BACKREF previously while processing the block.
5298  *
5299  * NOTE: return value 1 means we should stop walking up.
5300  */
5301 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5302                                  struct btrfs_root *root,
5303                                  struct btrfs_path *path,
5304                                  struct walk_control *wc)
5305 {
5306         struct btrfs_fs_info *fs_info = root->fs_info;
5307         int ret;
5308         int level = wc->level;
5309         struct extent_buffer *eb = path->nodes[level];
5310         u64 parent = 0;
5311
5312         if (wc->stage == UPDATE_BACKREF) {
5313                 BUG_ON(wc->shared_level < level);
5314                 if (level < wc->shared_level)
5315                         goto out;
5316
5317                 ret = find_next_key(path, level + 1, &wc->update_progress);
5318                 if (ret > 0)
5319                         wc->update_ref = 0;
5320
5321                 wc->stage = DROP_REFERENCE;
5322                 wc->shared_level = -1;
5323                 path->slots[level] = 0;
5324
5325                 /*
5326                  * check reference count again if the block isn't locked.
5327                  * we should start walking down the tree again if reference
5328                  * count is one.
5329                  */
5330                 if (!path->locks[level]) {
5331                         BUG_ON(level == 0);
5332                         btrfs_tree_lock(eb);
5333                         path->locks[level] = BTRFS_WRITE_LOCK;
5334
5335                         ret = btrfs_lookup_extent_info(trans, fs_info,
5336                                                        eb->start, level, 1,
5337                                                        &wc->refs[level],
5338                                                        &wc->flags[level]);
5339                         if (ret < 0) {
5340                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5341                                 path->locks[level] = 0;
5342                                 return ret;
5343                         }
5344                         BUG_ON(wc->refs[level] == 0);
5345                         if (wc->refs[level] == 1) {
5346                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5347                                 path->locks[level] = 0;
5348                                 return 1;
5349                         }
5350                 }
5351         }
5352
5353         /* wc->stage == DROP_REFERENCE */
5354         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5355
5356         if (wc->refs[level] == 1) {
5357                 if (level == 0) {
5358                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5359                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5360                         else
5361                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5362                         BUG_ON(ret); /* -ENOMEM */
5363                         if (is_fstree(root->root_key.objectid)) {
5364                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5365                                 if (ret) {
5366                                         btrfs_err_rl(fs_info,
5367         "error %d accounting leaf items, quota is out of sync, rescan required",
5368                                              ret);
5369                                 }
5370                         }
5371                 }
5372                 /* make block locked assertion in btrfs_clean_tree_block happy */
5373                 if (!path->locks[level] &&
5374                     btrfs_header_generation(eb) == trans->transid) {
5375                         btrfs_tree_lock(eb);
5376                         path->locks[level] = BTRFS_WRITE_LOCK;
5377                 }
5378                 btrfs_clean_tree_block(eb);
5379         }
5380
5381         if (eb == root->node) {
5382                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5383                         parent = eb->start;
5384                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5385                         goto owner_mismatch;
5386         } else {
5387                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5388                         parent = path->nodes[level + 1]->start;
5389                 else if (root->root_key.objectid !=
5390                          btrfs_header_owner(path->nodes[level + 1]))
5391                         goto owner_mismatch;
5392         }
5393
5394         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
5395 out:
5396         wc->refs[level] = 0;
5397         wc->flags[level] = 0;
5398         return 0;
5399
5400 owner_mismatch:
5401         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5402                      btrfs_header_owner(eb), root->root_key.objectid);
5403         return -EUCLEAN;
5404 }
5405
5406 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5407                                    struct btrfs_root *root,
5408                                    struct btrfs_path *path,
5409                                    struct walk_control *wc)
5410 {
5411         int level = wc->level;
5412         int lookup_info = 1;
5413         int ret;
5414
5415         while (level >= 0) {
5416                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5417                 if (ret > 0)
5418                         break;
5419
5420                 if (level == 0)
5421                         break;
5422
5423                 if (path->slots[level] >=
5424                     btrfs_header_nritems(path->nodes[level]))
5425                         break;
5426
5427                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5428                 if (ret > 0) {
5429                         path->slots[level]++;
5430                         continue;
5431                 } else if (ret < 0)
5432                         return ret;
5433                 level = wc->level;
5434         }
5435         return 0;
5436 }
5437
5438 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5439                                  struct btrfs_root *root,
5440                                  struct btrfs_path *path,
5441                                  struct walk_control *wc, int max_level)
5442 {
5443         int level = wc->level;
5444         int ret;
5445
5446         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5447         while (level < max_level && path->nodes[level]) {
5448                 wc->level = level;
5449                 if (path->slots[level] + 1 <
5450                     btrfs_header_nritems(path->nodes[level])) {
5451                         path->slots[level]++;
5452                         return 0;
5453                 } else {
5454                         ret = walk_up_proc(trans, root, path, wc);
5455                         if (ret > 0)
5456                                 return 0;
5457                         if (ret < 0)
5458                                 return ret;
5459
5460                         if (path->locks[level]) {
5461                                 btrfs_tree_unlock_rw(path->nodes[level],
5462                                                      path->locks[level]);
5463                                 path->locks[level] = 0;
5464                         }
5465                         free_extent_buffer(path->nodes[level]);
5466                         path->nodes[level] = NULL;
5467                         level++;
5468                 }
5469         }
5470         return 1;
5471 }
5472
5473 /*
5474  * drop a subvolume tree.
5475  *
5476  * this function traverses the tree freeing any blocks that only
5477  * referenced by the tree.
5478  *
5479  * when a shared tree block is found. this function decreases its
5480  * reference count by one. if update_ref is true, this function
5481  * also make sure backrefs for the shared block and all lower level
5482  * blocks are properly updated.
5483  *
5484  * If called with for_reloc == 0, may exit early with -EAGAIN
5485  */
5486 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5487 {
5488         struct btrfs_fs_info *fs_info = root->fs_info;
5489         struct btrfs_path *path;
5490         struct btrfs_trans_handle *trans;
5491         struct btrfs_root *tree_root = fs_info->tree_root;
5492         struct btrfs_root_item *root_item = &root->root_item;
5493         struct walk_control *wc;
5494         struct btrfs_key key;
5495         int err = 0;
5496         int ret;
5497         int level;
5498         bool root_dropped = false;
5499
5500         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5501
5502         path = btrfs_alloc_path();
5503         if (!path) {
5504                 err = -ENOMEM;
5505                 goto out;
5506         }
5507
5508         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5509         if (!wc) {
5510                 btrfs_free_path(path);
5511                 err = -ENOMEM;
5512                 goto out;
5513         }
5514
5515         /*
5516          * Use join to avoid potential EINTR from transaction start. See
5517          * wait_reserve_ticket and the whole reservation callchain.
5518          */
5519         if (for_reloc)
5520                 trans = btrfs_join_transaction(tree_root);
5521         else
5522                 trans = btrfs_start_transaction(tree_root, 0);
5523         if (IS_ERR(trans)) {
5524                 err = PTR_ERR(trans);
5525                 goto out_free;
5526         }
5527
5528         err = btrfs_run_delayed_items(trans);
5529         if (err)
5530                 goto out_end_trans;
5531
5532         /*
5533          * This will help us catch people modifying the fs tree while we're
5534          * dropping it.  It is unsafe to mess with the fs tree while it's being
5535          * dropped as we unlock the root node and parent nodes as we walk down
5536          * the tree, assuming nothing will change.  If something does change
5537          * then we'll have stale information and drop references to blocks we've
5538          * already dropped.
5539          */
5540         set_bit(BTRFS_ROOT_DELETING, &root->state);
5541         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5542                 level = btrfs_header_level(root->node);
5543                 path->nodes[level] = btrfs_lock_root_node(root);
5544                 path->slots[level] = 0;
5545                 path->locks[level] = BTRFS_WRITE_LOCK;
5546                 memset(&wc->update_progress, 0,
5547                        sizeof(wc->update_progress));
5548         } else {
5549                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5550                 memcpy(&wc->update_progress, &key,
5551                        sizeof(wc->update_progress));
5552
5553                 level = btrfs_root_drop_level(root_item);
5554                 BUG_ON(level == 0);
5555                 path->lowest_level = level;
5556                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5557                 path->lowest_level = 0;
5558                 if (ret < 0) {
5559                         err = ret;
5560                         goto out_end_trans;
5561                 }
5562                 WARN_ON(ret > 0);
5563
5564                 /*
5565                  * unlock our path, this is safe because only this
5566                  * function is allowed to delete this snapshot
5567                  */
5568                 btrfs_unlock_up_safe(path, 0);
5569
5570                 level = btrfs_header_level(root->node);
5571                 while (1) {
5572                         btrfs_tree_lock(path->nodes[level]);
5573                         path->locks[level] = BTRFS_WRITE_LOCK;
5574
5575                         ret = btrfs_lookup_extent_info(trans, fs_info,
5576                                                 path->nodes[level]->start,
5577                                                 level, 1, &wc->refs[level],
5578                                                 &wc->flags[level]);
5579                         if (ret < 0) {
5580                                 err = ret;
5581                                 goto out_end_trans;
5582                         }
5583                         BUG_ON(wc->refs[level] == 0);
5584
5585                         if (level == btrfs_root_drop_level(root_item))
5586                                 break;
5587
5588                         btrfs_tree_unlock(path->nodes[level]);
5589                         path->locks[level] = 0;
5590                         WARN_ON(wc->refs[level] != 1);
5591                         level--;
5592                 }
5593         }
5594
5595         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5596         wc->level = level;
5597         wc->shared_level = -1;
5598         wc->stage = DROP_REFERENCE;
5599         wc->update_ref = update_ref;
5600         wc->keep_locks = 0;
5601         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5602
5603         while (1) {
5604
5605                 ret = walk_down_tree(trans, root, path, wc);
5606                 if (ret < 0) {
5607                         err = ret;
5608                         break;
5609                 }
5610
5611                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5612                 if (ret < 0) {
5613                         err = ret;
5614                         break;
5615                 }
5616
5617                 if (ret > 0) {
5618                         BUG_ON(wc->stage != DROP_REFERENCE);
5619                         break;
5620                 }
5621
5622                 if (wc->stage == DROP_REFERENCE) {
5623                         wc->drop_level = wc->level;
5624                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5625                                               &wc->drop_progress,
5626                                               path->slots[wc->drop_level]);
5627                 }
5628                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5629                                       &wc->drop_progress);
5630                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5631
5632                 BUG_ON(wc->level == 0);
5633                 if (btrfs_should_end_transaction(trans) ||
5634                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5635                         ret = btrfs_update_root(trans, tree_root,
5636                                                 &root->root_key,
5637                                                 root_item);
5638                         if (ret) {
5639                                 btrfs_abort_transaction(trans, ret);
5640                                 err = ret;
5641                                 goto out_end_trans;
5642                         }
5643
5644                         btrfs_end_transaction_throttle(trans);
5645                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5646                                 btrfs_debug(fs_info,
5647                                             "drop snapshot early exit");
5648                                 err = -EAGAIN;
5649                                 goto out_free;
5650                         }
5651
5652                        /*
5653                         * Use join to avoid potential EINTR from transaction
5654                         * start. See wait_reserve_ticket and the whole
5655                         * reservation callchain.
5656                         */
5657                         if (for_reloc)
5658                                 trans = btrfs_join_transaction(tree_root);
5659                         else
5660                                 trans = btrfs_start_transaction(tree_root, 0);
5661                         if (IS_ERR(trans)) {
5662                                 err = PTR_ERR(trans);
5663                                 goto out_free;
5664                         }
5665                 }
5666         }
5667         btrfs_release_path(path);
5668         if (err)
5669                 goto out_end_trans;
5670
5671         ret = btrfs_del_root(trans, &root->root_key);
5672         if (ret) {
5673                 btrfs_abort_transaction(trans, ret);
5674                 err = ret;
5675                 goto out_end_trans;
5676         }
5677
5678         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5679                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5680                                       NULL, NULL);
5681                 if (ret < 0) {
5682                         btrfs_abort_transaction(trans, ret);
5683                         err = ret;
5684                         goto out_end_trans;
5685                 } else if (ret > 0) {
5686                         /* if we fail to delete the orphan item this time
5687                          * around, it'll get picked up the next time.
5688                          *
5689                          * The most common failure here is just -ENOENT.
5690                          */
5691                         btrfs_del_orphan_item(trans, tree_root,
5692                                               root->root_key.objectid);
5693                 }
5694         }
5695
5696         /*
5697          * This subvolume is going to be completely dropped, and won't be
5698          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5699          * commit transaction time.  So free it here manually.
5700          */
5701         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5702         btrfs_qgroup_free_meta_all_pertrans(root);
5703
5704         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5705                 btrfs_add_dropped_root(trans, root);
5706         else
5707                 btrfs_put_root(root);
5708         root_dropped = true;
5709 out_end_trans:
5710         btrfs_end_transaction_throttle(trans);
5711 out_free:
5712         kfree(wc);
5713         btrfs_free_path(path);
5714 out:
5715         /*
5716          * So if we need to stop dropping the snapshot for whatever reason we
5717          * need to make sure to add it back to the dead root list so that we
5718          * keep trying to do the work later.  This also cleans up roots if we
5719          * don't have it in the radix (like when we recover after a power fail
5720          * or unmount) so we don't leak memory.
5721          */
5722         if (!for_reloc && !root_dropped)
5723                 btrfs_add_dead_root(root);
5724         return err;
5725 }
5726
5727 /*
5728  * drop subtree rooted at tree block 'node'.
5729  *
5730  * NOTE: this function will unlock and release tree block 'node'
5731  * only used by relocation code
5732  */
5733 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5734                         struct btrfs_root *root,
5735                         struct extent_buffer *node,
5736                         struct extent_buffer *parent)
5737 {
5738         struct btrfs_fs_info *fs_info = root->fs_info;
5739         struct btrfs_path *path;
5740         struct walk_control *wc;
5741         int level;
5742         int parent_level;
5743         int ret = 0;
5744         int wret;
5745
5746         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5747
5748         path = btrfs_alloc_path();
5749         if (!path)
5750                 return -ENOMEM;
5751
5752         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5753         if (!wc) {
5754                 btrfs_free_path(path);
5755                 return -ENOMEM;
5756         }
5757
5758         btrfs_assert_tree_locked(parent);
5759         parent_level = btrfs_header_level(parent);
5760         atomic_inc(&parent->refs);
5761         path->nodes[parent_level] = parent;
5762         path->slots[parent_level] = btrfs_header_nritems(parent);
5763
5764         btrfs_assert_tree_locked(node);
5765         level = btrfs_header_level(node);
5766         path->nodes[level] = node;
5767         path->slots[level] = 0;
5768         path->locks[level] = BTRFS_WRITE_LOCK;
5769
5770         wc->refs[parent_level] = 1;
5771         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5772         wc->level = level;
5773         wc->shared_level = -1;
5774         wc->stage = DROP_REFERENCE;
5775         wc->update_ref = 0;
5776         wc->keep_locks = 1;
5777         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5778
5779         while (1) {
5780                 wret = walk_down_tree(trans, root, path, wc);
5781                 if (wret < 0) {
5782                         ret = wret;
5783                         break;
5784                 }
5785
5786                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5787                 if (wret < 0)
5788                         ret = wret;
5789                 if (wret != 0)
5790                         break;
5791         }
5792
5793         kfree(wc);
5794         btrfs_free_path(path);
5795         return ret;
5796 }
5797
5798 /*
5799  * helper to account the unused space of all the readonly block group in the
5800  * space_info. takes mirrors into account.
5801  */
5802 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
5803 {
5804         struct btrfs_block_group *block_group;
5805         u64 free_bytes = 0;
5806         int factor;
5807
5808         /* It's df, we don't care if it's racy */
5809         if (list_empty(&sinfo->ro_bgs))
5810                 return 0;
5811
5812         spin_lock(&sinfo->lock);
5813         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
5814                 spin_lock(&block_group->lock);
5815
5816                 if (!block_group->ro) {
5817                         spin_unlock(&block_group->lock);
5818                         continue;
5819                 }
5820
5821                 factor = btrfs_bg_type_to_factor(block_group->flags);
5822                 free_bytes += (block_group->length -
5823                                block_group->used) * factor;
5824
5825                 spin_unlock(&block_group->lock);
5826         }
5827         spin_unlock(&sinfo->lock);
5828
5829         return free_bytes;
5830 }
5831
5832 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5833                                    u64 start, u64 end)
5834 {
5835         return unpin_extent_range(fs_info, start, end, false);
5836 }
5837
5838 /*
5839  * It used to be that old block groups would be left around forever.
5840  * Iterating over them would be enough to trim unused space.  Since we
5841  * now automatically remove them, we also need to iterate over unallocated
5842  * space.
5843  *
5844  * We don't want a transaction for this since the discard may take a
5845  * substantial amount of time.  We don't require that a transaction be
5846  * running, but we do need to take a running transaction into account
5847  * to ensure that we're not discarding chunks that were released or
5848  * allocated in the current transaction.
5849  *
5850  * Holding the chunks lock will prevent other threads from allocating
5851  * or releasing chunks, but it won't prevent a running transaction
5852  * from committing and releasing the memory that the pending chunks
5853  * list head uses.  For that, we need to take a reference to the
5854  * transaction and hold the commit root sem.  We only need to hold
5855  * it while performing the free space search since we have already
5856  * held back allocations.
5857  */
5858 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5859 {
5860         u64 start = SZ_1M, len = 0, end = 0;
5861         int ret;
5862
5863         *trimmed = 0;
5864
5865         /* Discard not supported = nothing to do. */
5866         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
5867                 return 0;
5868
5869         /* Not writable = nothing to do. */
5870         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5871                 return 0;
5872
5873         /* No free space = nothing to do. */
5874         if (device->total_bytes <= device->bytes_used)
5875                 return 0;
5876
5877         ret = 0;
5878
5879         while (1) {
5880                 struct btrfs_fs_info *fs_info = device->fs_info;
5881                 u64 bytes;
5882
5883                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5884                 if (ret)
5885                         break;
5886
5887                 find_first_clear_extent_bit(&device->alloc_state, start,
5888                                             &start, &end,
5889                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5890
5891                 /* Check if there are any CHUNK_* bits left */
5892                 if (start > device->total_bytes) {
5893                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
5894                         btrfs_warn_in_rcu(fs_info,
5895 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
5896                                           start, end - start + 1,
5897                                           rcu_str_deref(device->name),
5898                                           device->total_bytes);
5899                         mutex_unlock(&fs_info->chunk_mutex);
5900                         ret = 0;
5901                         break;
5902                 }
5903
5904                 /* Ensure we skip the reserved area in the first 1M */
5905                 start = max_t(u64, start, SZ_1M);
5906
5907                 /*
5908                  * If find_first_clear_extent_bit find a range that spans the
5909                  * end of the device it will set end to -1, in this case it's up
5910                  * to the caller to trim the value to the size of the device.
5911                  */
5912                 end = min(end, device->total_bytes - 1);
5913
5914                 len = end - start + 1;
5915
5916                 /* We didn't find any extents */
5917                 if (!len) {
5918                         mutex_unlock(&fs_info->chunk_mutex);
5919                         ret = 0;
5920                         break;
5921                 }
5922
5923                 ret = btrfs_issue_discard(device->bdev, start, len,
5924                                           &bytes);
5925                 if (!ret)
5926                         set_extent_bits(&device->alloc_state, start,
5927                                         start + bytes - 1,
5928                                         CHUNK_TRIMMED);
5929                 mutex_unlock(&fs_info->chunk_mutex);
5930
5931                 if (ret)
5932                         break;
5933
5934                 start += len;
5935                 *trimmed += bytes;
5936
5937                 if (fatal_signal_pending(current)) {
5938                         ret = -ERESTARTSYS;
5939                         break;
5940                 }
5941
5942                 cond_resched();
5943         }
5944
5945         return ret;
5946 }
5947
5948 /*
5949  * Trim the whole filesystem by:
5950  * 1) trimming the free space in each block group
5951  * 2) trimming the unallocated space on each device
5952  *
5953  * This will also continue trimming even if a block group or device encounters
5954  * an error.  The return value will be the last error, or 0 if nothing bad
5955  * happens.
5956  */
5957 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
5958 {
5959         struct btrfs_block_group *cache = NULL;
5960         struct btrfs_device *device;
5961         struct list_head *devices;
5962         u64 group_trimmed;
5963         u64 range_end = U64_MAX;
5964         u64 start;
5965         u64 end;
5966         u64 trimmed = 0;
5967         u64 bg_failed = 0;
5968         u64 dev_failed = 0;
5969         int bg_ret = 0;
5970         int dev_ret = 0;
5971         int ret = 0;
5972
5973         /*
5974          * Check range overflow if range->len is set.
5975          * The default range->len is U64_MAX.
5976          */
5977         if (range->len != U64_MAX &&
5978             check_add_overflow(range->start, range->len, &range_end))
5979                 return -EINVAL;
5980
5981         cache = btrfs_lookup_first_block_group(fs_info, range->start);
5982         for (; cache; cache = btrfs_next_block_group(cache)) {
5983                 if (cache->start >= range_end) {
5984                         btrfs_put_block_group(cache);
5985                         break;
5986                 }
5987
5988                 start = max(range->start, cache->start);
5989                 end = min(range_end, cache->start + cache->length);
5990
5991                 if (end - start >= range->minlen) {
5992                         if (!btrfs_block_group_done(cache)) {
5993                                 ret = btrfs_cache_block_group(cache, 0);
5994                                 if (ret) {
5995                                         bg_failed++;
5996                                         bg_ret = ret;
5997                                         continue;
5998                                 }
5999                                 ret = btrfs_wait_block_group_cache_done(cache);
6000                                 if (ret) {
6001                                         bg_failed++;
6002                                         bg_ret = ret;
6003                                         continue;
6004                                 }
6005                         }
6006                         ret = btrfs_trim_block_group(cache,
6007                                                      &group_trimmed,
6008                                                      start,
6009                                                      end,
6010                                                      range->minlen);
6011
6012                         trimmed += group_trimmed;
6013                         if (ret) {
6014                                 bg_failed++;
6015                                 bg_ret = ret;
6016                                 continue;
6017                         }
6018                 }
6019         }
6020
6021         if (bg_failed)
6022                 btrfs_warn(fs_info,
6023                         "failed to trim %llu block group(s), last error %d",
6024                         bg_failed, bg_ret);
6025         mutex_lock(&fs_info->fs_devices->device_list_mutex);
6026         devices = &fs_info->fs_devices->devices;
6027         list_for_each_entry(device, devices, dev_list) {
6028                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6029                 if (ret) {
6030                         dev_failed++;
6031                         dev_ret = ret;
6032                         break;
6033                 }
6034
6035                 trimmed += group_trimmed;
6036         }
6037         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
6038
6039         if (dev_failed)
6040                 btrfs_warn(fs_info,
6041                         "failed to trim %llu device(s), last error %d",
6042                         dev_failed, dev_ret);
6043         range->len = trimmed;
6044         if (bg_ret)
6045                 return bg_ret;
6046         return dev_ret;
6047 }