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(NULL, 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                         struct btrfs_device *device = stripe->dev;
1344
1345                         if (!device->bdev) {
1346                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1347                                 continue;
1348                         }
1349
1350                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
1351                                 continue;
1352
1353                         ret = do_discard_extent(stripe, &bytes);
1354                         if (!ret) {
1355                                 discarded_bytes += bytes;
1356                         } else if (ret != -EOPNOTSUPP) {
1357                                 /*
1358                                  * Logic errors or -ENOMEM, or -EIO, but
1359                                  * unlikely to happen.
1360                                  *
1361                                  * And since there are two loops, explicitly
1362                                  * go to out to avoid confusion.
1363                                  */
1364                                 btrfs_put_bbio(bbio);
1365                                 goto out;
1366                         }
1367
1368                         /*
1369                          * Just in case we get back EOPNOTSUPP for some reason,
1370                          * just ignore the return value so we don't screw up
1371                          * people calling discard_extent.
1372                          */
1373                         ret = 0;
1374                 }
1375                 btrfs_put_bbio(bbio);
1376                 cur += num_bytes;
1377         }
1378 out:
1379         btrfs_bio_counter_dec(fs_info);
1380
1381         if (actual_bytes)
1382                 *actual_bytes = discarded_bytes;
1383
1384
1385         if (ret == -EOPNOTSUPP)
1386                 ret = 0;
1387         return ret;
1388 }
1389
1390 /* Can return -ENOMEM */
1391 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1392                          struct btrfs_ref *generic_ref)
1393 {
1394         struct btrfs_fs_info *fs_info = trans->fs_info;
1395         int ret;
1396
1397         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1398                generic_ref->action);
1399         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1400                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
1401
1402         if (generic_ref->type == BTRFS_REF_METADATA)
1403                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1404         else
1405                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1406
1407         btrfs_ref_tree_mod(fs_info, generic_ref);
1408
1409         return ret;
1410 }
1411
1412 /*
1413  * __btrfs_inc_extent_ref - insert backreference for a given extent
1414  *
1415  * The counterpart is in __btrfs_free_extent(), with examples and more details
1416  * how it works.
1417  *
1418  * @trans:          Handle of transaction
1419  *
1420  * @node:           The delayed ref node used to get the bytenr/length for
1421  *                  extent whose references are incremented.
1422  *
1423  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1424  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1425  *                  bytenr of the parent block. Since new extents are always
1426  *                  created with indirect references, this will only be the case
1427  *                  when relocating a shared extent. In that case, root_objectid
1428  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1429  *                  be 0
1430  *
1431  * @root_objectid:  The id of the root where this modification has originated,
1432  *                  this can be either one of the well-known metadata trees or
1433  *                  the subvolume id which references this extent.
1434  *
1435  * @owner:          For data extents it is the inode number of the owning file.
1436  *                  For metadata extents this parameter holds the level in the
1437  *                  tree of the extent.
1438  *
1439  * @offset:         For metadata extents the offset is ignored and is currently
1440  *                  always passed as 0. For data extents it is the fileoffset
1441  *                  this extent belongs to.
1442  *
1443  * @refs_to_add     Number of references to add
1444  *
1445  * @extent_op       Pointer to a structure, holding information necessary when
1446  *                  updating a tree block's flags
1447  *
1448  */
1449 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1450                                   struct btrfs_delayed_ref_node *node,
1451                                   u64 parent, u64 root_objectid,
1452                                   u64 owner, u64 offset, int refs_to_add,
1453                                   struct btrfs_delayed_extent_op *extent_op)
1454 {
1455         struct btrfs_path *path;
1456         struct extent_buffer *leaf;
1457         struct btrfs_extent_item *item;
1458         struct btrfs_key key;
1459         u64 bytenr = node->bytenr;
1460         u64 num_bytes = node->num_bytes;
1461         u64 refs;
1462         int ret;
1463
1464         path = btrfs_alloc_path();
1465         if (!path)
1466                 return -ENOMEM;
1467
1468         /* this will setup the path even if it fails to insert the back ref */
1469         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1470                                            parent, root_objectid, owner,
1471                                            offset, refs_to_add, extent_op);
1472         if ((ret < 0 && ret != -EAGAIN) || !ret)
1473                 goto out;
1474
1475         /*
1476          * Ok we had -EAGAIN which means we didn't have space to insert and
1477          * inline extent ref, so just update the reference count and add a
1478          * normal backref.
1479          */
1480         leaf = path->nodes[0];
1481         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1482         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1483         refs = btrfs_extent_refs(leaf, item);
1484         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1485         if (extent_op)
1486                 __run_delayed_extent_op(extent_op, leaf, item);
1487
1488         btrfs_mark_buffer_dirty(leaf);
1489         btrfs_release_path(path);
1490
1491         /* now insert the actual backref */
1492         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1493                 BUG_ON(refs_to_add != 1);
1494                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1495                                             root_objectid);
1496         } else {
1497                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1498                                              root_objectid, owner, offset,
1499                                              refs_to_add);
1500         }
1501         if (ret)
1502                 btrfs_abort_transaction(trans, ret);
1503 out:
1504         btrfs_free_path(path);
1505         return ret;
1506 }
1507
1508 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1509                                 struct btrfs_delayed_ref_node *node,
1510                                 struct btrfs_delayed_extent_op *extent_op,
1511                                 int insert_reserved)
1512 {
1513         int ret = 0;
1514         struct btrfs_delayed_data_ref *ref;
1515         struct btrfs_key ins;
1516         u64 parent = 0;
1517         u64 ref_root = 0;
1518         u64 flags = 0;
1519
1520         ins.objectid = node->bytenr;
1521         ins.offset = node->num_bytes;
1522         ins.type = BTRFS_EXTENT_ITEM_KEY;
1523
1524         ref = btrfs_delayed_node_to_data_ref(node);
1525         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1526
1527         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1528                 parent = ref->parent;
1529         ref_root = ref->root;
1530
1531         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1532                 if (extent_op)
1533                         flags |= extent_op->flags_to_set;
1534                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1535                                                  flags, ref->objectid,
1536                                                  ref->offset, &ins,
1537                                                  node->ref_mod);
1538         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1539                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1540                                              ref->objectid, ref->offset,
1541                                              node->ref_mod, extent_op);
1542         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1543                 ret = __btrfs_free_extent(trans, node, parent,
1544                                           ref_root, ref->objectid,
1545                                           ref->offset, node->ref_mod,
1546                                           extent_op);
1547         } else {
1548                 BUG();
1549         }
1550         return ret;
1551 }
1552
1553 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1554                                     struct extent_buffer *leaf,
1555                                     struct btrfs_extent_item *ei)
1556 {
1557         u64 flags = btrfs_extent_flags(leaf, ei);
1558         if (extent_op->update_flags) {
1559                 flags |= extent_op->flags_to_set;
1560                 btrfs_set_extent_flags(leaf, ei, flags);
1561         }
1562
1563         if (extent_op->update_key) {
1564                 struct btrfs_tree_block_info *bi;
1565                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1566                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1567                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1568         }
1569 }
1570
1571 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1572                                  struct btrfs_delayed_ref_head *head,
1573                                  struct btrfs_delayed_extent_op *extent_op)
1574 {
1575         struct btrfs_fs_info *fs_info = trans->fs_info;
1576         struct btrfs_key key;
1577         struct btrfs_path *path;
1578         struct btrfs_extent_item *ei;
1579         struct extent_buffer *leaf;
1580         u32 item_size;
1581         int ret;
1582         int err = 0;
1583         int metadata = !extent_op->is_data;
1584
1585         if (TRANS_ABORTED(trans))
1586                 return 0;
1587
1588         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1589                 metadata = 0;
1590
1591         path = btrfs_alloc_path();
1592         if (!path)
1593                 return -ENOMEM;
1594
1595         key.objectid = head->bytenr;
1596
1597         if (metadata) {
1598                 key.type = BTRFS_METADATA_ITEM_KEY;
1599                 key.offset = extent_op->level;
1600         } else {
1601                 key.type = BTRFS_EXTENT_ITEM_KEY;
1602                 key.offset = head->num_bytes;
1603         }
1604
1605 again:
1606         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
1607         if (ret < 0) {
1608                 err = ret;
1609                 goto out;
1610         }
1611         if (ret > 0) {
1612                 if (metadata) {
1613                         if (path->slots[0] > 0) {
1614                                 path->slots[0]--;
1615                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1616                                                       path->slots[0]);
1617                                 if (key.objectid == head->bytenr &&
1618                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1619                                     key.offset == head->num_bytes)
1620                                         ret = 0;
1621                         }
1622                         if (ret > 0) {
1623                                 btrfs_release_path(path);
1624                                 metadata = 0;
1625
1626                                 key.objectid = head->bytenr;
1627                                 key.offset = head->num_bytes;
1628                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1629                                 goto again;
1630                         }
1631                 } else {
1632                         err = -EIO;
1633                         goto out;
1634                 }
1635         }
1636
1637         leaf = path->nodes[0];
1638         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1639
1640         if (unlikely(item_size < sizeof(*ei))) {
1641                 err = -EINVAL;
1642                 btrfs_print_v0_err(fs_info);
1643                 btrfs_abort_transaction(trans, err);
1644                 goto out;
1645         }
1646
1647         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1648         __run_delayed_extent_op(extent_op, leaf, ei);
1649
1650         btrfs_mark_buffer_dirty(leaf);
1651 out:
1652         btrfs_free_path(path);
1653         return err;
1654 }
1655
1656 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1657                                 struct btrfs_delayed_ref_node *node,
1658                                 struct btrfs_delayed_extent_op *extent_op,
1659                                 int insert_reserved)
1660 {
1661         int ret = 0;
1662         struct btrfs_delayed_tree_ref *ref;
1663         u64 parent = 0;
1664         u64 ref_root = 0;
1665
1666         ref = btrfs_delayed_node_to_tree_ref(node);
1667         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1668
1669         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1670                 parent = ref->parent;
1671         ref_root = ref->root;
1672
1673         if (node->ref_mod != 1) {
1674                 btrfs_err(trans->fs_info,
1675         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1676                           node->bytenr, node->ref_mod, node->action, ref_root,
1677                           parent);
1678                 return -EIO;
1679         }
1680         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1681                 BUG_ON(!extent_op || !extent_op->update_flags);
1682                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1683         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1684                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1685                                              ref->level, 0, 1, extent_op);
1686         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1687                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1688                                           ref->level, 0, 1, extent_op);
1689         } else {
1690                 BUG();
1691         }
1692         return ret;
1693 }
1694
1695 /* helper function to actually process a single delayed ref entry */
1696 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1697                                struct btrfs_delayed_ref_node *node,
1698                                struct btrfs_delayed_extent_op *extent_op,
1699                                int insert_reserved)
1700 {
1701         int ret = 0;
1702
1703         if (TRANS_ABORTED(trans)) {
1704                 if (insert_reserved)
1705                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1706                 return 0;
1707         }
1708
1709         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1710             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1711                 ret = run_delayed_tree_ref(trans, node, extent_op,
1712                                            insert_reserved);
1713         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1714                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1715                 ret = run_delayed_data_ref(trans, node, extent_op,
1716                                            insert_reserved);
1717         else
1718                 BUG();
1719         if (ret && insert_reserved)
1720                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1721         return ret;
1722 }
1723
1724 static inline struct btrfs_delayed_ref_node *
1725 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1726 {
1727         struct btrfs_delayed_ref_node *ref;
1728
1729         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1730                 return NULL;
1731
1732         /*
1733          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1734          * This is to prevent a ref count from going down to zero, which deletes
1735          * the extent item from the extent tree, when there still are references
1736          * to add, which would fail because they would not find the extent item.
1737          */
1738         if (!list_empty(&head->ref_add_list))
1739                 return list_first_entry(&head->ref_add_list,
1740                                 struct btrfs_delayed_ref_node, add_list);
1741
1742         ref = rb_entry(rb_first_cached(&head->ref_tree),
1743                        struct btrfs_delayed_ref_node, ref_node);
1744         ASSERT(list_empty(&ref->add_list));
1745         return ref;
1746 }
1747
1748 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1749                                       struct btrfs_delayed_ref_head *head)
1750 {
1751         spin_lock(&delayed_refs->lock);
1752         head->processing = 0;
1753         delayed_refs->num_heads_ready++;
1754         spin_unlock(&delayed_refs->lock);
1755         btrfs_delayed_ref_unlock(head);
1756 }
1757
1758 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1759                                 struct btrfs_delayed_ref_head *head)
1760 {
1761         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1762
1763         if (!extent_op)
1764                 return NULL;
1765
1766         if (head->must_insert_reserved) {
1767                 head->extent_op = NULL;
1768                 btrfs_free_delayed_extent_op(extent_op);
1769                 return NULL;
1770         }
1771         return extent_op;
1772 }
1773
1774 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1775                                      struct btrfs_delayed_ref_head *head)
1776 {
1777         struct btrfs_delayed_extent_op *extent_op;
1778         int ret;
1779
1780         extent_op = cleanup_extent_op(head);
1781         if (!extent_op)
1782                 return 0;
1783         head->extent_op = NULL;
1784         spin_unlock(&head->lock);
1785         ret = run_delayed_extent_op(trans, head, extent_op);
1786         btrfs_free_delayed_extent_op(extent_op);
1787         return ret ? ret : 1;
1788 }
1789
1790 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1791                                   struct btrfs_delayed_ref_root *delayed_refs,
1792                                   struct btrfs_delayed_ref_head *head)
1793 {
1794         int nr_items = 1;       /* Dropping this ref head update. */
1795
1796         /*
1797          * We had csum deletions accounted for in our delayed refs rsv, we need
1798          * to drop the csum leaves for this update from our delayed_refs_rsv.
1799          */
1800         if (head->total_ref_mod < 0 && head->is_data) {
1801                 spin_lock(&delayed_refs->lock);
1802                 delayed_refs->pending_csums -= head->num_bytes;
1803                 spin_unlock(&delayed_refs->lock);
1804                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1805         }
1806
1807         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1808 }
1809
1810 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1811                             struct btrfs_delayed_ref_head *head)
1812 {
1813
1814         struct btrfs_fs_info *fs_info = trans->fs_info;
1815         struct btrfs_delayed_ref_root *delayed_refs;
1816         int ret;
1817
1818         delayed_refs = &trans->transaction->delayed_refs;
1819
1820         ret = run_and_cleanup_extent_op(trans, head);
1821         if (ret < 0) {
1822                 unselect_delayed_ref_head(delayed_refs, head);
1823                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1824                 return ret;
1825         } else if (ret) {
1826                 return ret;
1827         }
1828
1829         /*
1830          * Need to drop our head ref lock and re-acquire the delayed ref lock
1831          * and then re-check to make sure nobody got added.
1832          */
1833         spin_unlock(&head->lock);
1834         spin_lock(&delayed_refs->lock);
1835         spin_lock(&head->lock);
1836         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1837                 spin_unlock(&head->lock);
1838                 spin_unlock(&delayed_refs->lock);
1839                 return 1;
1840         }
1841         btrfs_delete_ref_head(delayed_refs, head);
1842         spin_unlock(&head->lock);
1843         spin_unlock(&delayed_refs->lock);
1844
1845         if (head->must_insert_reserved) {
1846                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1847                 if (head->is_data) {
1848                         ret = btrfs_del_csums(trans, fs_info->csum_root,
1849                                               head->bytenr, head->num_bytes);
1850                 }
1851         }
1852
1853         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1854
1855         trace_run_delayed_ref_head(fs_info, head, 0);
1856         btrfs_delayed_ref_unlock(head);
1857         btrfs_put_delayed_ref_head(head);
1858         return ret;
1859 }
1860
1861 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1862                                         struct btrfs_trans_handle *trans)
1863 {
1864         struct btrfs_delayed_ref_root *delayed_refs =
1865                 &trans->transaction->delayed_refs;
1866         struct btrfs_delayed_ref_head *head = NULL;
1867         int ret;
1868
1869         spin_lock(&delayed_refs->lock);
1870         head = btrfs_select_ref_head(delayed_refs);
1871         if (!head) {
1872                 spin_unlock(&delayed_refs->lock);
1873                 return head;
1874         }
1875
1876         /*
1877          * Grab the lock that says we are going to process all the refs for
1878          * this head
1879          */
1880         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1881         spin_unlock(&delayed_refs->lock);
1882
1883         /*
1884          * We may have dropped the spin lock to get the head mutex lock, and
1885          * that might have given someone else time to free the head.  If that's
1886          * true, it has been removed from our list and we can move on.
1887          */
1888         if (ret == -EAGAIN)
1889                 head = ERR_PTR(-EAGAIN);
1890
1891         return head;
1892 }
1893
1894 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1895                                     struct btrfs_delayed_ref_head *locked_ref,
1896                                     unsigned long *run_refs)
1897 {
1898         struct btrfs_fs_info *fs_info = trans->fs_info;
1899         struct btrfs_delayed_ref_root *delayed_refs;
1900         struct btrfs_delayed_extent_op *extent_op;
1901         struct btrfs_delayed_ref_node *ref;
1902         int must_insert_reserved = 0;
1903         int ret;
1904
1905         delayed_refs = &trans->transaction->delayed_refs;
1906
1907         lockdep_assert_held(&locked_ref->mutex);
1908         lockdep_assert_held(&locked_ref->lock);
1909
1910         while ((ref = select_delayed_ref(locked_ref))) {
1911                 if (ref->seq &&
1912                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1913                         spin_unlock(&locked_ref->lock);
1914                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1915                         return -EAGAIN;
1916                 }
1917
1918                 (*run_refs)++;
1919                 ref->in_tree = 0;
1920                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1921                 RB_CLEAR_NODE(&ref->ref_node);
1922                 if (!list_empty(&ref->add_list))
1923                         list_del(&ref->add_list);
1924                 /*
1925                  * When we play the delayed ref, also correct the ref_mod on
1926                  * head
1927                  */
1928                 switch (ref->action) {
1929                 case BTRFS_ADD_DELAYED_REF:
1930                 case BTRFS_ADD_DELAYED_EXTENT:
1931                         locked_ref->ref_mod -= ref->ref_mod;
1932                         break;
1933                 case BTRFS_DROP_DELAYED_REF:
1934                         locked_ref->ref_mod += ref->ref_mod;
1935                         break;
1936                 default:
1937                         WARN_ON(1);
1938                 }
1939                 atomic_dec(&delayed_refs->num_entries);
1940
1941                 /*
1942                  * Record the must_insert_reserved flag before we drop the
1943                  * spin lock.
1944                  */
1945                 must_insert_reserved = locked_ref->must_insert_reserved;
1946                 locked_ref->must_insert_reserved = 0;
1947
1948                 extent_op = locked_ref->extent_op;
1949                 locked_ref->extent_op = NULL;
1950                 spin_unlock(&locked_ref->lock);
1951
1952                 ret = run_one_delayed_ref(trans, ref, extent_op,
1953                                           must_insert_reserved);
1954
1955                 btrfs_free_delayed_extent_op(extent_op);
1956                 if (ret) {
1957                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1958                         btrfs_put_delayed_ref(ref);
1959                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
1960                                     ret);
1961                         return ret;
1962                 }
1963
1964                 btrfs_put_delayed_ref(ref);
1965                 cond_resched();
1966
1967                 spin_lock(&locked_ref->lock);
1968                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1969         }
1970
1971         return 0;
1972 }
1973
1974 /*
1975  * Returns 0 on success or if called with an already aborted transaction.
1976  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1977  */
1978 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1979                                              unsigned long nr)
1980 {
1981         struct btrfs_fs_info *fs_info = trans->fs_info;
1982         struct btrfs_delayed_ref_root *delayed_refs;
1983         struct btrfs_delayed_ref_head *locked_ref = NULL;
1984         ktime_t start = ktime_get();
1985         int ret;
1986         unsigned long count = 0;
1987         unsigned long actual_count = 0;
1988
1989         delayed_refs = &trans->transaction->delayed_refs;
1990         do {
1991                 if (!locked_ref) {
1992                         locked_ref = btrfs_obtain_ref_head(trans);
1993                         if (IS_ERR_OR_NULL(locked_ref)) {
1994                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
1995                                         continue;
1996                                 } else {
1997                                         break;
1998                                 }
1999                         }
2000                         count++;
2001                 }
2002                 /*
2003                  * We need to try and merge add/drops of the same ref since we
2004                  * can run into issues with relocate dropping the implicit ref
2005                  * and then it being added back again before the drop can
2006                  * finish.  If we merged anything we need to re-loop so we can
2007                  * get a good ref.
2008                  * Or we can get node references of the same type that weren't
2009                  * merged when created due to bumps in the tree mod seq, and
2010                  * we need to merge them to prevent adding an inline extent
2011                  * backref before dropping it (triggering a BUG_ON at
2012                  * insert_inline_extent_backref()).
2013                  */
2014                 spin_lock(&locked_ref->lock);
2015                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2016
2017                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2018                                                       &actual_count);
2019                 if (ret < 0 && ret != -EAGAIN) {
2020                         /*
2021                          * Error, btrfs_run_delayed_refs_for_head already
2022                          * unlocked everything so just bail out
2023                          */
2024                         return ret;
2025                 } else if (!ret) {
2026                         /*
2027                          * Success, perform the usual cleanup of a processed
2028                          * head
2029                          */
2030                         ret = cleanup_ref_head(trans, locked_ref);
2031                         if (ret > 0 ) {
2032                                 /* We dropped our lock, we need to loop. */
2033                                 ret = 0;
2034                                 continue;
2035                         } else if (ret) {
2036                                 return ret;
2037                         }
2038                 }
2039
2040                 /*
2041                  * Either success case or btrfs_run_delayed_refs_for_head
2042                  * returned -EAGAIN, meaning we need to select another head
2043                  */
2044
2045                 locked_ref = NULL;
2046                 cond_resched();
2047         } while ((nr != -1 && count < nr) || locked_ref);
2048
2049         /*
2050          * We don't want to include ref heads since we can have empty ref heads
2051          * and those will drastically skew our runtime down since we just do
2052          * accounting, no actual extent tree updates.
2053          */
2054         if (actual_count > 0) {
2055                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2056                 u64 avg;
2057
2058                 /*
2059                  * We weigh the current average higher than our current runtime
2060                  * to avoid large swings in the average.
2061                  */
2062                 spin_lock(&delayed_refs->lock);
2063                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2064                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2065                 spin_unlock(&delayed_refs->lock);
2066         }
2067         return 0;
2068 }
2069
2070 #ifdef SCRAMBLE_DELAYED_REFS
2071 /*
2072  * Normally delayed refs get processed in ascending bytenr order. This
2073  * correlates in most cases to the order added. To expose dependencies on this
2074  * order, we start to process the tree in the middle instead of the beginning
2075  */
2076 static u64 find_middle(struct rb_root *root)
2077 {
2078         struct rb_node *n = root->rb_node;
2079         struct btrfs_delayed_ref_node *entry;
2080         int alt = 1;
2081         u64 middle;
2082         u64 first = 0, last = 0;
2083
2084         n = rb_first(root);
2085         if (n) {
2086                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2087                 first = entry->bytenr;
2088         }
2089         n = rb_last(root);
2090         if (n) {
2091                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2092                 last = entry->bytenr;
2093         }
2094         n = root->rb_node;
2095
2096         while (n) {
2097                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2098                 WARN_ON(!entry->in_tree);
2099
2100                 middle = entry->bytenr;
2101
2102                 if (alt)
2103                         n = n->rb_left;
2104                 else
2105                         n = n->rb_right;
2106
2107                 alt = 1 - alt;
2108         }
2109         return middle;
2110 }
2111 #endif
2112
2113 /*
2114  * this starts processing the delayed reference count updates and
2115  * extent insertions we have queued up so far.  count can be
2116  * 0, which means to process everything in the tree at the start
2117  * of the run (but not newly added entries), or it can be some target
2118  * number you'd like to process.
2119  *
2120  * Returns 0 on success or if called with an aborted transaction
2121  * Returns <0 on error and aborts the transaction
2122  */
2123 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2124                            unsigned long count)
2125 {
2126         struct btrfs_fs_info *fs_info = trans->fs_info;
2127         struct rb_node *node;
2128         struct btrfs_delayed_ref_root *delayed_refs;
2129         struct btrfs_delayed_ref_head *head;
2130         int ret;
2131         int run_all = count == (unsigned long)-1;
2132
2133         /* We'll clean this up in btrfs_cleanup_transaction */
2134         if (TRANS_ABORTED(trans))
2135                 return 0;
2136
2137         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2138                 return 0;
2139
2140         delayed_refs = &trans->transaction->delayed_refs;
2141         if (count == 0)
2142                 count = delayed_refs->num_heads_ready;
2143
2144 again:
2145 #ifdef SCRAMBLE_DELAYED_REFS
2146         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2147 #endif
2148         ret = __btrfs_run_delayed_refs(trans, count);
2149         if (ret < 0) {
2150                 btrfs_abort_transaction(trans, ret);
2151                 return ret;
2152         }
2153
2154         if (run_all) {
2155                 btrfs_create_pending_block_groups(trans);
2156
2157                 spin_lock(&delayed_refs->lock);
2158                 node = rb_first_cached(&delayed_refs->href_root);
2159                 if (!node) {
2160                         spin_unlock(&delayed_refs->lock);
2161                         goto out;
2162                 }
2163                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2164                                 href_node);
2165                 refcount_inc(&head->refs);
2166                 spin_unlock(&delayed_refs->lock);
2167
2168                 /* Mutex was contended, block until it's released and retry. */
2169                 mutex_lock(&head->mutex);
2170                 mutex_unlock(&head->mutex);
2171
2172                 btrfs_put_delayed_ref_head(head);
2173                 cond_resched();
2174                 goto again;
2175         }
2176 out:
2177         return 0;
2178 }
2179
2180 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2181                                 struct extent_buffer *eb, u64 flags,
2182                                 int level, int is_data)
2183 {
2184         struct btrfs_delayed_extent_op *extent_op;
2185         int ret;
2186
2187         extent_op = btrfs_alloc_delayed_extent_op();
2188         if (!extent_op)
2189                 return -ENOMEM;
2190
2191         extent_op->flags_to_set = flags;
2192         extent_op->update_flags = true;
2193         extent_op->update_key = false;
2194         extent_op->is_data = is_data ? true : false;
2195         extent_op->level = level;
2196
2197         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2198         if (ret)
2199                 btrfs_free_delayed_extent_op(extent_op);
2200         return ret;
2201 }
2202
2203 static noinline int check_delayed_ref(struct btrfs_root *root,
2204                                       struct btrfs_path *path,
2205                                       u64 objectid, u64 offset, u64 bytenr)
2206 {
2207         struct btrfs_delayed_ref_head *head;
2208         struct btrfs_delayed_ref_node *ref;
2209         struct btrfs_delayed_data_ref *data_ref;
2210         struct btrfs_delayed_ref_root *delayed_refs;
2211         struct btrfs_transaction *cur_trans;
2212         struct rb_node *node;
2213         int ret = 0;
2214
2215         spin_lock(&root->fs_info->trans_lock);
2216         cur_trans = root->fs_info->running_transaction;
2217         if (cur_trans)
2218                 refcount_inc(&cur_trans->use_count);
2219         spin_unlock(&root->fs_info->trans_lock);
2220         if (!cur_trans)
2221                 return 0;
2222
2223         delayed_refs = &cur_trans->delayed_refs;
2224         spin_lock(&delayed_refs->lock);
2225         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2226         if (!head) {
2227                 spin_unlock(&delayed_refs->lock);
2228                 btrfs_put_transaction(cur_trans);
2229                 return 0;
2230         }
2231
2232         if (!mutex_trylock(&head->mutex)) {
2233                 refcount_inc(&head->refs);
2234                 spin_unlock(&delayed_refs->lock);
2235
2236                 btrfs_release_path(path);
2237
2238                 /*
2239                  * Mutex was contended, block until it's released and let
2240                  * caller try again
2241                  */
2242                 mutex_lock(&head->mutex);
2243                 mutex_unlock(&head->mutex);
2244                 btrfs_put_delayed_ref_head(head);
2245                 btrfs_put_transaction(cur_trans);
2246                 return -EAGAIN;
2247         }
2248         spin_unlock(&delayed_refs->lock);
2249
2250         spin_lock(&head->lock);
2251         /*
2252          * XXX: We should replace this with a proper search function in the
2253          * future.
2254          */
2255         for (node = rb_first_cached(&head->ref_tree); node;
2256              node = rb_next(node)) {
2257                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2258                 /* If it's a shared ref we know a cross reference exists */
2259                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2260                         ret = 1;
2261                         break;
2262                 }
2263
2264                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2265
2266                 /*
2267                  * If our ref doesn't match the one we're currently looking at
2268                  * then we have a cross reference.
2269                  */
2270                 if (data_ref->root != root->root_key.objectid ||
2271                     data_ref->objectid != objectid ||
2272                     data_ref->offset != offset) {
2273                         ret = 1;
2274                         break;
2275                 }
2276         }
2277         spin_unlock(&head->lock);
2278         mutex_unlock(&head->mutex);
2279         btrfs_put_transaction(cur_trans);
2280         return ret;
2281 }
2282
2283 static noinline int check_committed_ref(struct btrfs_root *root,
2284                                         struct btrfs_path *path,
2285                                         u64 objectid, u64 offset, u64 bytenr,
2286                                         bool strict)
2287 {
2288         struct btrfs_fs_info *fs_info = root->fs_info;
2289         struct btrfs_root *extent_root = fs_info->extent_root;
2290         struct extent_buffer *leaf;
2291         struct btrfs_extent_data_ref *ref;
2292         struct btrfs_extent_inline_ref *iref;
2293         struct btrfs_extent_item *ei;
2294         struct btrfs_key key;
2295         u32 item_size;
2296         int type;
2297         int ret;
2298
2299         key.objectid = bytenr;
2300         key.offset = (u64)-1;
2301         key.type = BTRFS_EXTENT_ITEM_KEY;
2302
2303         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2304         if (ret < 0)
2305                 goto out;
2306         BUG_ON(ret == 0); /* Corruption */
2307
2308         ret = -ENOENT;
2309         if (path->slots[0] == 0)
2310                 goto out;
2311
2312         path->slots[0]--;
2313         leaf = path->nodes[0];
2314         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2315
2316         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2317                 goto out;
2318
2319         ret = 1;
2320         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2321         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2322
2323         /* If extent item has more than 1 inline ref then it's shared */
2324         if (item_size != sizeof(*ei) +
2325             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2326                 goto out;
2327
2328         /*
2329          * If extent created before last snapshot => it's shared unless the
2330          * snapshot has been deleted. Use the heuristic if strict is false.
2331          */
2332         if (!strict &&
2333             (btrfs_extent_generation(leaf, ei) <=
2334              btrfs_root_last_snapshot(&root->root_item)))
2335                 goto out;
2336
2337         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2338
2339         /* If this extent has SHARED_DATA_REF then it's shared */
2340         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2341         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2342                 goto out;
2343
2344         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2345         if (btrfs_extent_refs(leaf, ei) !=
2346             btrfs_extent_data_ref_count(leaf, ref) ||
2347             btrfs_extent_data_ref_root(leaf, ref) !=
2348             root->root_key.objectid ||
2349             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2350             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2351                 goto out;
2352
2353         ret = 0;
2354 out:
2355         return ret;
2356 }
2357
2358 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2359                           u64 bytenr, bool strict)
2360 {
2361         struct btrfs_path *path;
2362         int ret;
2363
2364         path = btrfs_alloc_path();
2365         if (!path)
2366                 return -ENOMEM;
2367
2368         do {
2369                 ret = check_committed_ref(root, path, objectid,
2370                                           offset, bytenr, strict);
2371                 if (ret && ret != -ENOENT)
2372                         goto out;
2373
2374                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2375         } while (ret == -EAGAIN);
2376
2377 out:
2378         btrfs_free_path(path);
2379         if (btrfs_is_data_reloc_root(root))
2380                 WARN_ON(ret > 0);
2381         return ret;
2382 }
2383
2384 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2385                            struct btrfs_root *root,
2386                            struct extent_buffer *buf,
2387                            int full_backref, int inc)
2388 {
2389         struct btrfs_fs_info *fs_info = root->fs_info;
2390         u64 bytenr;
2391         u64 num_bytes;
2392         u64 parent;
2393         u64 ref_root;
2394         u32 nritems;
2395         struct btrfs_key key;
2396         struct btrfs_file_extent_item *fi;
2397         struct btrfs_ref generic_ref = { 0 };
2398         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2399         int i;
2400         int action;
2401         int level;
2402         int ret = 0;
2403
2404         if (btrfs_is_testing(fs_info))
2405                 return 0;
2406
2407         ref_root = btrfs_header_owner(buf);
2408         nritems = btrfs_header_nritems(buf);
2409         level = btrfs_header_level(buf);
2410
2411         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2412                 return 0;
2413
2414         if (full_backref)
2415                 parent = buf->start;
2416         else
2417                 parent = 0;
2418         if (inc)
2419                 action = BTRFS_ADD_DELAYED_REF;
2420         else
2421                 action = BTRFS_DROP_DELAYED_REF;
2422
2423         for (i = 0; i < nritems; i++) {
2424                 if (level == 0) {
2425                         btrfs_item_key_to_cpu(buf, &key, i);
2426                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2427                                 continue;
2428                         fi = btrfs_item_ptr(buf, i,
2429                                             struct btrfs_file_extent_item);
2430                         if (btrfs_file_extent_type(buf, fi) ==
2431                             BTRFS_FILE_EXTENT_INLINE)
2432                                 continue;
2433                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2434                         if (bytenr == 0)
2435                                 continue;
2436
2437                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2438                         key.offset -= btrfs_file_extent_offset(buf, fi);
2439                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2440                                                num_bytes, parent);
2441                         generic_ref.real_root = root->root_key.objectid;
2442                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2443                                             key.offset);
2444                         generic_ref.skip_qgroup = for_reloc;
2445                         if (inc)
2446                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2447                         else
2448                                 ret = btrfs_free_extent(trans, &generic_ref);
2449                         if (ret)
2450                                 goto fail;
2451                 } else {
2452                         bytenr = btrfs_node_blockptr(buf, i);
2453                         num_bytes = fs_info->nodesize;
2454                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2455                                                num_bytes, parent);
2456                         generic_ref.real_root = root->root_key.objectid;
2457                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
2458                         generic_ref.skip_qgroup = for_reloc;
2459                         if (inc)
2460                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2461                         else
2462                                 ret = btrfs_free_extent(trans, &generic_ref);
2463                         if (ret)
2464                                 goto fail;
2465                 }
2466         }
2467         return 0;
2468 fail:
2469         return ret;
2470 }
2471
2472 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2473                   struct extent_buffer *buf, int full_backref)
2474 {
2475         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2476 }
2477
2478 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2479                   struct extent_buffer *buf, int full_backref)
2480 {
2481         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2482 }
2483
2484 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2485 {
2486         struct btrfs_fs_info *fs_info = root->fs_info;
2487         u64 flags;
2488         u64 ret;
2489
2490         if (data)
2491                 flags = BTRFS_BLOCK_GROUP_DATA;
2492         else if (root == fs_info->chunk_root)
2493                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2494         else
2495                 flags = BTRFS_BLOCK_GROUP_METADATA;
2496
2497         ret = btrfs_get_alloc_profile(fs_info, flags);
2498         return ret;
2499 }
2500
2501 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
2502 {
2503         struct btrfs_block_group *cache;
2504         u64 bytenr;
2505
2506         spin_lock(&fs_info->block_group_cache_lock);
2507         bytenr = fs_info->first_logical_byte;
2508         spin_unlock(&fs_info->block_group_cache_lock);
2509
2510         if (bytenr < (u64)-1)
2511                 return bytenr;
2512
2513         cache = btrfs_lookup_first_block_group(fs_info, search_start);
2514         if (!cache)
2515                 return 0;
2516
2517         bytenr = cache->start;
2518         btrfs_put_block_group(cache);
2519
2520         return bytenr;
2521 }
2522
2523 static int pin_down_extent(struct btrfs_trans_handle *trans,
2524                            struct btrfs_block_group *cache,
2525                            u64 bytenr, u64 num_bytes, int reserved)
2526 {
2527         struct btrfs_fs_info *fs_info = cache->fs_info;
2528
2529         spin_lock(&cache->space_info->lock);
2530         spin_lock(&cache->lock);
2531         cache->pinned += num_bytes;
2532         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2533                                              num_bytes);
2534         if (reserved) {
2535                 cache->reserved -= num_bytes;
2536                 cache->space_info->bytes_reserved -= num_bytes;
2537         }
2538         spin_unlock(&cache->lock);
2539         spin_unlock(&cache->space_info->lock);
2540
2541         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2542                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2543         return 0;
2544 }
2545
2546 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2547                      u64 bytenr, u64 num_bytes, int reserved)
2548 {
2549         struct btrfs_block_group *cache;
2550
2551         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2552         BUG_ON(!cache); /* Logic error */
2553
2554         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2555
2556         btrfs_put_block_group(cache);
2557         return 0;
2558 }
2559
2560 /*
2561  * this function must be called within transaction
2562  */
2563 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2564                                     u64 bytenr, u64 num_bytes)
2565 {
2566         struct btrfs_block_group *cache;
2567         int ret;
2568
2569         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2570         if (!cache)
2571                 return -EINVAL;
2572
2573         /*
2574          * pull in the free space cache (if any) so that our pin
2575          * removes the free space from the cache.  We have load_only set
2576          * to one because the slow code to read in the free extents does check
2577          * the pinned extents.
2578          */
2579         btrfs_cache_block_group(cache, 1);
2580         /*
2581          * Make sure we wait until the cache is completely built in case it is
2582          * missing or is invalid and therefore needs to be rebuilt.
2583          */
2584         ret = btrfs_wait_block_group_cache_done(cache);
2585         if (ret)
2586                 goto out;
2587
2588         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2589
2590         /* remove us from the free space cache (if we're there at all) */
2591         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2592 out:
2593         btrfs_put_block_group(cache);
2594         return ret;
2595 }
2596
2597 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2598                                    u64 start, u64 num_bytes)
2599 {
2600         int ret;
2601         struct btrfs_block_group *block_group;
2602
2603         block_group = btrfs_lookup_block_group(fs_info, start);
2604         if (!block_group)
2605                 return -EINVAL;
2606
2607         btrfs_cache_block_group(block_group, 1);
2608         /*
2609          * Make sure we wait until the cache is completely built in case it is
2610          * missing or is invalid and therefore needs to be rebuilt.
2611          */
2612         ret = btrfs_wait_block_group_cache_done(block_group);
2613         if (ret)
2614                 goto out;
2615
2616         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2617 out:
2618         btrfs_put_block_group(block_group);
2619         return ret;
2620 }
2621
2622 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2623 {
2624         struct btrfs_fs_info *fs_info = eb->fs_info;
2625         struct btrfs_file_extent_item *item;
2626         struct btrfs_key key;
2627         int found_type;
2628         int i;
2629         int ret = 0;
2630
2631         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2632                 return 0;
2633
2634         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2635                 btrfs_item_key_to_cpu(eb, &key, i);
2636                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2637                         continue;
2638                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2639                 found_type = btrfs_file_extent_type(eb, item);
2640                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2641                         continue;
2642                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2643                         continue;
2644                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2645                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2646                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2647                 if (ret)
2648                         break;
2649         }
2650
2651         return ret;
2652 }
2653
2654 static void
2655 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2656 {
2657         atomic_inc(&bg->reservations);
2658 }
2659
2660 /*
2661  * Returns the free cluster for the given space info and sets empty_cluster to
2662  * what it should be based on the mount options.
2663  */
2664 static struct btrfs_free_cluster *
2665 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2666                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2667 {
2668         struct btrfs_free_cluster *ret = NULL;
2669
2670         *empty_cluster = 0;
2671         if (btrfs_mixed_space_info(space_info))
2672                 return ret;
2673
2674         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2675                 ret = &fs_info->meta_alloc_cluster;
2676                 if (btrfs_test_opt(fs_info, SSD))
2677                         *empty_cluster = SZ_2M;
2678                 else
2679                         *empty_cluster = SZ_64K;
2680         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2681                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2682                 *empty_cluster = SZ_2M;
2683                 ret = &fs_info->data_alloc_cluster;
2684         }
2685
2686         return ret;
2687 }
2688
2689 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2690                               u64 start, u64 end,
2691                               const bool return_free_space)
2692 {
2693         struct btrfs_block_group *cache = NULL;
2694         struct btrfs_space_info *space_info;
2695         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2696         struct btrfs_free_cluster *cluster = NULL;
2697         u64 len;
2698         u64 total_unpinned = 0;
2699         u64 empty_cluster = 0;
2700         bool readonly;
2701
2702         while (start <= end) {
2703                 readonly = false;
2704                 if (!cache ||
2705                     start >= cache->start + cache->length) {
2706                         if (cache)
2707                                 btrfs_put_block_group(cache);
2708                         total_unpinned = 0;
2709                         cache = btrfs_lookup_block_group(fs_info, start);
2710                         BUG_ON(!cache); /* Logic error */
2711
2712                         cluster = fetch_cluster_info(fs_info,
2713                                                      cache->space_info,
2714                                                      &empty_cluster);
2715                         empty_cluster <<= 1;
2716                 }
2717
2718                 len = cache->start + cache->length - start;
2719                 len = min(len, end + 1 - start);
2720
2721                 down_read(&fs_info->commit_root_sem);
2722                 if (start < cache->last_byte_to_unpin && return_free_space) {
2723                         u64 add_len = min(len, cache->last_byte_to_unpin - start);
2724
2725                         btrfs_add_free_space(cache, start, add_len);
2726                 }
2727                 up_read(&fs_info->commit_root_sem);
2728
2729                 start += len;
2730                 total_unpinned += len;
2731                 space_info = cache->space_info;
2732
2733                 /*
2734                  * If this space cluster has been marked as fragmented and we've
2735                  * unpinned enough in this block group to potentially allow a
2736                  * cluster to be created inside of it go ahead and clear the
2737                  * fragmented check.
2738                  */
2739                 if (cluster && cluster->fragmented &&
2740                     total_unpinned > empty_cluster) {
2741                         spin_lock(&cluster->lock);
2742                         cluster->fragmented = 0;
2743                         spin_unlock(&cluster->lock);
2744                 }
2745
2746                 spin_lock(&space_info->lock);
2747                 spin_lock(&cache->lock);
2748                 cache->pinned -= len;
2749                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2750                 space_info->max_extent_size = 0;
2751                 if (cache->ro) {
2752                         space_info->bytes_readonly += len;
2753                         readonly = true;
2754                 } else if (btrfs_is_zoned(fs_info)) {
2755                         /* Need reset before reusing in a zoned block group */
2756                         space_info->bytes_zone_unusable += len;
2757                         readonly = true;
2758                 }
2759                 spin_unlock(&cache->lock);
2760                 if (!readonly && return_free_space &&
2761                     global_rsv->space_info == space_info) {
2762                         u64 to_add = len;
2763
2764                         spin_lock(&global_rsv->lock);
2765                         if (!global_rsv->full) {
2766                                 to_add = min(len, global_rsv->size -
2767                                              global_rsv->reserved);
2768                                 global_rsv->reserved += to_add;
2769                                 btrfs_space_info_update_bytes_may_use(fs_info,
2770                                                 space_info, to_add);
2771                                 if (global_rsv->reserved >= global_rsv->size)
2772                                         global_rsv->full = 1;
2773                                 len -= to_add;
2774                         }
2775                         spin_unlock(&global_rsv->lock);
2776                 }
2777                 /* Add to any tickets we may have */
2778                 if (!readonly && return_free_space && len)
2779                         btrfs_try_granting_tickets(fs_info, space_info);
2780                 spin_unlock(&space_info->lock);
2781         }
2782
2783         if (cache)
2784                 btrfs_put_block_group(cache);
2785         return 0;
2786 }
2787
2788 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2789 {
2790         struct btrfs_fs_info *fs_info = trans->fs_info;
2791         struct btrfs_block_group *block_group, *tmp;
2792         struct list_head *deleted_bgs;
2793         struct extent_io_tree *unpin;
2794         u64 start;
2795         u64 end;
2796         int ret;
2797
2798         unpin = &trans->transaction->pinned_extents;
2799
2800         while (!TRANS_ABORTED(trans)) {
2801                 struct extent_state *cached_state = NULL;
2802
2803                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2804                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2805                                             EXTENT_DIRTY, &cached_state);
2806                 if (ret) {
2807                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2808                         break;
2809                 }
2810
2811                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2812                         ret = btrfs_discard_extent(fs_info, start,
2813                                                    end + 1 - start, NULL);
2814
2815                 clear_extent_dirty(unpin, start, end, &cached_state);
2816                 unpin_extent_range(fs_info, start, end, true);
2817                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2818                 free_extent_state(cached_state);
2819                 cond_resched();
2820         }
2821
2822         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2823                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2824                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2825         }
2826
2827         /*
2828          * Transaction is finished.  We don't need the lock anymore.  We
2829          * do need to clean up the block groups in case of a transaction
2830          * abort.
2831          */
2832         deleted_bgs = &trans->transaction->deleted_bgs;
2833         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2834                 u64 trimmed = 0;
2835
2836                 ret = -EROFS;
2837                 if (!TRANS_ABORTED(trans))
2838                         ret = btrfs_discard_extent(fs_info,
2839                                                    block_group->start,
2840                                                    block_group->length,
2841                                                    &trimmed);
2842
2843                 list_del_init(&block_group->bg_list);
2844                 btrfs_unfreeze_block_group(block_group);
2845                 btrfs_put_block_group(block_group);
2846
2847                 if (ret) {
2848                         const char *errstr = btrfs_decode_error(ret);
2849                         btrfs_warn(fs_info,
2850                            "discard failed while removing blockgroup: errno=%d %s",
2851                                    ret, errstr);
2852                 }
2853         }
2854
2855         return 0;
2856 }
2857
2858 /*
2859  * Drop one or more refs of @node.
2860  *
2861  * 1. Locate the extent refs.
2862  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2863  *    Locate it, then reduce the refs number or remove the ref line completely.
2864  *
2865  * 2. Update the refs count in EXTENT/METADATA_ITEM
2866  *
2867  * Inline backref case:
2868  *
2869  * in extent tree we have:
2870  *
2871  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2872  *              refs 2 gen 6 flags DATA
2873  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2874  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2875  *
2876  * This function gets called with:
2877  *
2878  *    node->bytenr = 13631488
2879  *    node->num_bytes = 1048576
2880  *    root_objectid = FS_TREE
2881  *    owner_objectid = 257
2882  *    owner_offset = 0
2883  *    refs_to_drop = 1
2884  *
2885  * Then we should get some like:
2886  *
2887  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2888  *              refs 1 gen 6 flags DATA
2889  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2890  *
2891  * Keyed backref case:
2892  *
2893  * in extent tree we have:
2894  *
2895  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2896  *              refs 754 gen 6 flags DATA
2897  *      [...]
2898  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2899  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2900  *
2901  * This function get called with:
2902  *
2903  *    node->bytenr = 13631488
2904  *    node->num_bytes = 1048576
2905  *    root_objectid = FS_TREE
2906  *    owner_objectid = 866
2907  *    owner_offset = 0
2908  *    refs_to_drop = 1
2909  *
2910  * Then we should get some like:
2911  *
2912  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2913  *              refs 753 gen 6 flags DATA
2914  *
2915  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2916  */
2917 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2918                                struct btrfs_delayed_ref_node *node, u64 parent,
2919                                u64 root_objectid, u64 owner_objectid,
2920                                u64 owner_offset, int refs_to_drop,
2921                                struct btrfs_delayed_extent_op *extent_op)
2922 {
2923         struct btrfs_fs_info *info = trans->fs_info;
2924         struct btrfs_key key;
2925         struct btrfs_path *path;
2926         struct btrfs_root *extent_root = info->extent_root;
2927         struct extent_buffer *leaf;
2928         struct btrfs_extent_item *ei;
2929         struct btrfs_extent_inline_ref *iref;
2930         int ret;
2931         int is_data;
2932         int extent_slot = 0;
2933         int found_extent = 0;
2934         int num_to_del = 1;
2935         u32 item_size;
2936         u64 refs;
2937         u64 bytenr = node->bytenr;
2938         u64 num_bytes = node->num_bytes;
2939         int last_ref = 0;
2940         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2941
2942         path = btrfs_alloc_path();
2943         if (!path)
2944                 return -ENOMEM;
2945
2946         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2947
2948         if (!is_data && refs_to_drop != 1) {
2949                 btrfs_crit(info,
2950 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2951                            node->bytenr, refs_to_drop);
2952                 ret = -EINVAL;
2953                 btrfs_abort_transaction(trans, ret);
2954                 goto out;
2955         }
2956
2957         if (is_data)
2958                 skinny_metadata = false;
2959
2960         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2961                                     parent, root_objectid, owner_objectid,
2962                                     owner_offset);
2963         if (ret == 0) {
2964                 /*
2965                  * Either the inline backref or the SHARED_DATA_REF/
2966                  * SHARED_BLOCK_REF is found
2967                  *
2968                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2969                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2970                  */
2971                 extent_slot = path->slots[0];
2972                 while (extent_slot >= 0) {
2973                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2974                                               extent_slot);
2975                         if (key.objectid != bytenr)
2976                                 break;
2977                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2978                             key.offset == num_bytes) {
2979                                 found_extent = 1;
2980                                 break;
2981                         }
2982                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2983                             key.offset == owner_objectid) {
2984                                 found_extent = 1;
2985                                 break;
2986                         }
2987
2988                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
2989                         if (path->slots[0] - extent_slot > 5)
2990                                 break;
2991                         extent_slot--;
2992                 }
2993
2994                 if (!found_extent) {
2995                         if (iref) {
2996                                 btrfs_crit(info,
2997 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
2998                                 btrfs_abort_transaction(trans, -EUCLEAN);
2999                                 goto err_dump;
3000                         }
3001                         /* Must be SHARED_* item, remove the backref first */
3002                         ret = remove_extent_backref(trans, path, NULL,
3003                                                     refs_to_drop,
3004                                                     is_data, &last_ref);
3005                         if (ret) {
3006                                 btrfs_abort_transaction(trans, ret);
3007                                 goto out;
3008                         }
3009                         btrfs_release_path(path);
3010
3011                         /* Slow path to locate EXTENT/METADATA_ITEM */
3012                         key.objectid = bytenr;
3013                         key.type = BTRFS_EXTENT_ITEM_KEY;
3014                         key.offset = num_bytes;
3015
3016                         if (!is_data && skinny_metadata) {
3017                                 key.type = BTRFS_METADATA_ITEM_KEY;
3018                                 key.offset = owner_objectid;
3019                         }
3020
3021                         ret = btrfs_search_slot(trans, extent_root,
3022                                                 &key, path, -1, 1);
3023                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3024                                 /*
3025                                  * Couldn't find our skinny metadata item,
3026                                  * see if we have ye olde extent item.
3027                                  */
3028                                 path->slots[0]--;
3029                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3030                                                       path->slots[0]);
3031                                 if (key.objectid == bytenr &&
3032                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3033                                     key.offset == num_bytes)
3034                                         ret = 0;
3035                         }
3036
3037                         if (ret > 0 && skinny_metadata) {
3038                                 skinny_metadata = false;
3039                                 key.objectid = bytenr;
3040                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3041                                 key.offset = num_bytes;
3042                                 btrfs_release_path(path);
3043                                 ret = btrfs_search_slot(trans, extent_root,
3044                                                         &key, path, -1, 1);
3045                         }
3046
3047                         if (ret) {
3048                                 btrfs_err(info,
3049                                           "umm, got %d back from search, was looking for %llu",
3050                                           ret, bytenr);
3051                                 if (ret > 0)
3052                                         btrfs_print_leaf(path->nodes[0]);
3053                         }
3054                         if (ret < 0) {
3055                                 btrfs_abort_transaction(trans, ret);
3056                                 goto out;
3057                         }
3058                         extent_slot = path->slots[0];
3059                 }
3060         } else if (WARN_ON(ret == -ENOENT)) {
3061                 btrfs_print_leaf(path->nodes[0]);
3062                 btrfs_err(info,
3063                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3064                         bytenr, parent, root_objectid, owner_objectid,
3065                         owner_offset);
3066                 btrfs_abort_transaction(trans, ret);
3067                 goto out;
3068         } else {
3069                 btrfs_abort_transaction(trans, ret);
3070                 goto out;
3071         }
3072
3073         leaf = path->nodes[0];
3074         item_size = btrfs_item_size_nr(leaf, extent_slot);
3075         if (unlikely(item_size < sizeof(*ei))) {
3076                 ret = -EINVAL;
3077                 btrfs_print_v0_err(info);
3078                 btrfs_abort_transaction(trans, ret);
3079                 goto out;
3080         }
3081         ei = btrfs_item_ptr(leaf, extent_slot,
3082                             struct btrfs_extent_item);
3083         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3084             key.type == BTRFS_EXTENT_ITEM_KEY) {
3085                 struct btrfs_tree_block_info *bi;
3086                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3087                         btrfs_crit(info,
3088 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3089                                    key.objectid, key.type, key.offset,
3090                                    owner_objectid, item_size,
3091                                    sizeof(*ei) + sizeof(*bi));
3092                         btrfs_abort_transaction(trans, -EUCLEAN);
3093                         goto err_dump;
3094                 }
3095                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3096                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3097         }
3098
3099         refs = btrfs_extent_refs(leaf, ei);
3100         if (refs < refs_to_drop) {
3101                 btrfs_crit(info,
3102                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3103                           refs_to_drop, refs, bytenr);
3104                 btrfs_abort_transaction(trans, -EUCLEAN);
3105                 goto err_dump;
3106         }
3107         refs -= refs_to_drop;
3108
3109         if (refs > 0) {
3110                 if (extent_op)
3111                         __run_delayed_extent_op(extent_op, leaf, ei);
3112                 /*
3113                  * In the case of inline back ref, reference count will
3114                  * be updated by remove_extent_backref
3115                  */
3116                 if (iref) {
3117                         if (!found_extent) {
3118                                 btrfs_crit(info,
3119 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3120                                 btrfs_abort_transaction(trans, -EUCLEAN);
3121                                 goto err_dump;
3122                         }
3123                 } else {
3124                         btrfs_set_extent_refs(leaf, ei, refs);
3125                         btrfs_mark_buffer_dirty(leaf);
3126                 }
3127                 if (found_extent) {
3128                         ret = remove_extent_backref(trans, path, iref,
3129                                                     refs_to_drop, is_data,
3130                                                     &last_ref);
3131                         if (ret) {
3132                                 btrfs_abort_transaction(trans, ret);
3133                                 goto out;
3134                         }
3135                 }
3136         } else {
3137                 /* In this branch refs == 1 */
3138                 if (found_extent) {
3139                         if (is_data && refs_to_drop !=
3140                             extent_data_ref_count(path, iref)) {
3141                                 btrfs_crit(info,
3142                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3143                                            extent_data_ref_count(path, iref),
3144                                            refs_to_drop);
3145                                 btrfs_abort_transaction(trans, -EUCLEAN);
3146                                 goto err_dump;
3147                         }
3148                         if (iref) {
3149                                 if (path->slots[0] != extent_slot) {
3150                                         btrfs_crit(info,
3151 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3152                                                    key.objectid, key.type,
3153                                                    key.offset);
3154                                         btrfs_abort_transaction(trans, -EUCLEAN);
3155                                         goto err_dump;
3156                                 }
3157                         } else {
3158                                 /*
3159                                  * No inline ref, we must be at SHARED_* item,
3160                                  * And it's single ref, it must be:
3161                                  * |    extent_slot       ||extent_slot + 1|
3162                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3163                                  */
3164                                 if (path->slots[0] != extent_slot + 1) {
3165                                         btrfs_crit(info,
3166         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3167                                         btrfs_abort_transaction(trans, -EUCLEAN);
3168                                         goto err_dump;
3169                                 }
3170                                 path->slots[0] = extent_slot;
3171                                 num_to_del = 2;
3172                         }
3173                 }
3174
3175                 last_ref = 1;
3176                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3177                                       num_to_del);
3178                 if (ret) {
3179                         btrfs_abort_transaction(trans, ret);
3180                         goto out;
3181                 }
3182                 btrfs_release_path(path);
3183
3184                 if (is_data) {
3185                         ret = btrfs_del_csums(trans, info->csum_root, bytenr,
3186                                               num_bytes);
3187                         if (ret) {
3188                                 btrfs_abort_transaction(trans, ret);
3189                                 goto out;
3190                         }
3191                 }
3192
3193                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3194                 if (ret) {
3195                         btrfs_abort_transaction(trans, ret);
3196                         goto out;
3197                 }
3198
3199                 ret = btrfs_update_block_group(trans, bytenr, num_bytes, 0);
3200                 if (ret) {
3201                         btrfs_abort_transaction(trans, ret);
3202                         goto out;
3203                 }
3204         }
3205         btrfs_release_path(path);
3206
3207 out:
3208         btrfs_free_path(path);
3209         return ret;
3210 err_dump:
3211         /*
3212          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3213          * dump for debug build.
3214          */
3215         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3216                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3217                            path->slots[0], extent_slot);
3218                 btrfs_print_leaf(path->nodes[0]);
3219         }
3220
3221         btrfs_free_path(path);
3222         return -EUCLEAN;
3223 }
3224
3225 /*
3226  * when we free an block, it is possible (and likely) that we free the last
3227  * delayed ref for that extent as well.  This searches the delayed ref tree for
3228  * a given extent, and if there are no other delayed refs to be processed, it
3229  * removes it from the tree.
3230  */
3231 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3232                                       u64 bytenr)
3233 {
3234         struct btrfs_delayed_ref_head *head;
3235         struct btrfs_delayed_ref_root *delayed_refs;
3236         int ret = 0;
3237
3238         delayed_refs = &trans->transaction->delayed_refs;
3239         spin_lock(&delayed_refs->lock);
3240         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3241         if (!head)
3242                 goto out_delayed_unlock;
3243
3244         spin_lock(&head->lock);
3245         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3246                 goto out;
3247
3248         if (cleanup_extent_op(head) != NULL)
3249                 goto out;
3250
3251         /*
3252          * waiting for the lock here would deadlock.  If someone else has it
3253          * locked they are already in the process of dropping it anyway
3254          */
3255         if (!mutex_trylock(&head->mutex))
3256                 goto out;
3257
3258         btrfs_delete_ref_head(delayed_refs, head);
3259         head->processing = 0;
3260
3261         spin_unlock(&head->lock);
3262         spin_unlock(&delayed_refs->lock);
3263
3264         BUG_ON(head->extent_op);
3265         if (head->must_insert_reserved)
3266                 ret = 1;
3267
3268         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3269         mutex_unlock(&head->mutex);
3270         btrfs_put_delayed_ref_head(head);
3271         return ret;
3272 out:
3273         spin_unlock(&head->lock);
3274
3275 out_delayed_unlock:
3276         spin_unlock(&delayed_refs->lock);
3277         return 0;
3278 }
3279
3280 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3281                            struct btrfs_root *root,
3282                            struct extent_buffer *buf,
3283                            u64 parent, int last_ref)
3284 {
3285         struct btrfs_fs_info *fs_info = root->fs_info;
3286         struct btrfs_ref generic_ref = { 0 };
3287         int ret;
3288
3289         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3290                                buf->start, buf->len, parent);
3291         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3292                             root->root_key.objectid);
3293
3294         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3295                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3296                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3297                 BUG_ON(ret); /* -ENOMEM */
3298         }
3299
3300         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3301                 struct btrfs_block_group *cache;
3302                 bool must_pin = false;
3303
3304                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3305                         ret = check_ref_cleanup(trans, buf->start);
3306                         if (!ret) {
3307                                 btrfs_redirty_list_add(trans->transaction, buf);
3308                                 goto out;
3309                         }
3310                 }
3311
3312                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3313
3314                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3315                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3316                         btrfs_put_block_group(cache);
3317                         goto out;
3318                 }
3319
3320                 /*
3321                  * If this is a leaf and there are tree mod log users, we may
3322                  * have recorded mod log operations that point to this leaf.
3323                  * So we must make sure no one reuses this leaf's extent before
3324                  * mod log operations are applied to a node, otherwise after
3325                  * rewinding a node using the mod log operations we get an
3326                  * inconsistent btree, as the leaf's extent may now be used as
3327                  * a node or leaf for another different btree.
3328                  * We are safe from races here because at this point no other
3329                  * node or root points to this extent buffer, so if after this
3330                  * check a new tree mod log user joins, it will not be able to
3331                  * find a node pointing to this leaf and record operations that
3332                  * point to this leaf.
3333                  */
3334                 if (btrfs_header_level(buf) == 0 &&
3335                     test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3336                         must_pin = true;
3337
3338                 if (must_pin || btrfs_is_zoned(fs_info)) {
3339                         btrfs_redirty_list_add(trans->transaction, buf);
3340                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3341                         btrfs_put_block_group(cache);
3342                         goto out;
3343                 }
3344
3345                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3346
3347                 btrfs_add_free_space(cache, buf->start, buf->len);
3348                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3349                 btrfs_put_block_group(cache);
3350                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3351         }
3352 out:
3353         if (last_ref) {
3354                 /*
3355                  * Deleting the buffer, clear the corrupt flag since it doesn't
3356                  * matter anymore.
3357                  */
3358                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3359         }
3360 }
3361
3362 /* Can return -ENOMEM */
3363 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3364 {
3365         struct btrfs_fs_info *fs_info = trans->fs_info;
3366         int ret;
3367
3368         if (btrfs_is_testing(fs_info))
3369                 return 0;
3370
3371         /*
3372          * tree log blocks never actually go into the extent allocation
3373          * tree, just update pinning info and exit early.
3374          */
3375         if ((ref->type == BTRFS_REF_METADATA &&
3376              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3377             (ref->type == BTRFS_REF_DATA &&
3378              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3379                 /* unlocks the pinned mutex */
3380                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3381                 ret = 0;
3382         } else if (ref->type == BTRFS_REF_METADATA) {
3383                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3384         } else {
3385                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3386         }
3387
3388         if (!((ref->type == BTRFS_REF_METADATA &&
3389                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3390               (ref->type == BTRFS_REF_DATA &&
3391                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3392                 btrfs_ref_tree_mod(fs_info, ref);
3393
3394         return ret;
3395 }
3396
3397 enum btrfs_loop_type {
3398         LOOP_CACHING_NOWAIT,
3399         LOOP_CACHING_WAIT,
3400         LOOP_ALLOC_CHUNK,
3401         LOOP_NO_EMPTY_SIZE,
3402 };
3403
3404 static inline void
3405 btrfs_lock_block_group(struct btrfs_block_group *cache,
3406                        int delalloc)
3407 {
3408         if (delalloc)
3409                 down_read(&cache->data_rwsem);
3410 }
3411
3412 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3413                        int delalloc)
3414 {
3415         btrfs_get_block_group(cache);
3416         if (delalloc)
3417                 down_read(&cache->data_rwsem);
3418 }
3419
3420 static struct btrfs_block_group *btrfs_lock_cluster(
3421                    struct btrfs_block_group *block_group,
3422                    struct btrfs_free_cluster *cluster,
3423                    int delalloc)
3424         __acquires(&cluster->refill_lock)
3425 {
3426         struct btrfs_block_group *used_bg = NULL;
3427
3428         spin_lock(&cluster->refill_lock);
3429         while (1) {
3430                 used_bg = cluster->block_group;
3431                 if (!used_bg)
3432                         return NULL;
3433
3434                 if (used_bg == block_group)
3435                         return used_bg;
3436
3437                 btrfs_get_block_group(used_bg);
3438
3439                 if (!delalloc)
3440                         return used_bg;
3441
3442                 if (down_read_trylock(&used_bg->data_rwsem))
3443                         return used_bg;
3444
3445                 spin_unlock(&cluster->refill_lock);
3446
3447                 /* We should only have one-level nested. */
3448                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3449
3450                 spin_lock(&cluster->refill_lock);
3451                 if (used_bg == cluster->block_group)
3452                         return used_bg;
3453
3454                 up_read(&used_bg->data_rwsem);
3455                 btrfs_put_block_group(used_bg);
3456         }
3457 }
3458
3459 static inline void
3460 btrfs_release_block_group(struct btrfs_block_group *cache,
3461                          int delalloc)
3462 {
3463         if (delalloc)
3464                 up_read(&cache->data_rwsem);
3465         btrfs_put_block_group(cache);
3466 }
3467
3468 enum btrfs_extent_allocation_policy {
3469         BTRFS_EXTENT_ALLOC_CLUSTERED,
3470         BTRFS_EXTENT_ALLOC_ZONED,
3471 };
3472
3473 /*
3474  * Structure used internally for find_free_extent() function.  Wraps needed
3475  * parameters.
3476  */
3477 struct find_free_extent_ctl {
3478         /* Basic allocation info */
3479         u64 ram_bytes;
3480         u64 num_bytes;
3481         u64 min_alloc_size;
3482         u64 empty_size;
3483         u64 flags;
3484         int delalloc;
3485
3486         /* Where to start the search inside the bg */
3487         u64 search_start;
3488
3489         /* For clustered allocation */
3490         u64 empty_cluster;
3491         struct btrfs_free_cluster *last_ptr;
3492         bool use_cluster;
3493
3494         bool have_caching_bg;
3495         bool orig_have_caching_bg;
3496
3497         /* Allocation is called for tree-log */
3498         bool for_treelog;
3499
3500         /* Allocation is called for data relocation */
3501         bool for_data_reloc;
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         u64 data_reloc_bytenr;
3765         int ret = 0;
3766         bool skip;
3767
3768         ASSERT(btrfs_is_zoned(block_group->fs_info));
3769
3770         /*
3771          * Do not allow non-tree-log blocks in the dedicated tree-log block
3772          * group, and vice versa.
3773          */
3774         spin_lock(&fs_info->treelog_bg_lock);
3775         log_bytenr = fs_info->treelog_bg;
3776         skip = log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3777                               (!ffe_ctl->for_treelog && bytenr == log_bytenr));
3778         spin_unlock(&fs_info->treelog_bg_lock);
3779         if (skip)
3780                 return 1;
3781
3782         /*
3783          * Do not allow non-relocation blocks in the dedicated relocation block
3784          * group, and vice versa.
3785          */
3786         spin_lock(&fs_info->relocation_bg_lock);
3787         data_reloc_bytenr = fs_info->data_reloc_bg;
3788         if (data_reloc_bytenr &&
3789             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3790              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3791                 skip = true;
3792         spin_unlock(&fs_info->relocation_bg_lock);
3793         if (skip)
3794                 return 1;
3795         /* Check RO and no space case before trying to activate it */
3796         spin_lock(&block_group->lock);
3797         if (block_group->ro ||
3798             block_group->alloc_offset == block_group->zone_capacity) {
3799                 spin_unlock(&block_group->lock);
3800                 return 1;
3801         }
3802         spin_unlock(&block_group->lock);
3803
3804         if (!btrfs_zone_activate(block_group))
3805                 return 1;
3806
3807         spin_lock(&space_info->lock);
3808         spin_lock(&block_group->lock);
3809         spin_lock(&fs_info->treelog_bg_lock);
3810         spin_lock(&fs_info->relocation_bg_lock);
3811
3812         ASSERT(!ffe_ctl->for_treelog ||
3813                block_group->start == fs_info->treelog_bg ||
3814                fs_info->treelog_bg == 0);
3815         ASSERT(!ffe_ctl->for_data_reloc ||
3816                block_group->start == fs_info->data_reloc_bg ||
3817                fs_info->data_reloc_bg == 0);
3818
3819         if (block_group->ro) {
3820                 ret = 1;
3821                 goto out;
3822         }
3823
3824         /*
3825          * Do not allow currently using block group to be tree-log dedicated
3826          * block group.
3827          */
3828         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3829             (block_group->used || block_group->reserved)) {
3830                 ret = 1;
3831                 goto out;
3832         }
3833
3834         /*
3835          * Do not allow currently used block group to be the data relocation
3836          * dedicated block group.
3837          */
3838         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3839             (block_group->used || block_group->reserved)) {
3840                 ret = 1;
3841                 goto out;
3842         }
3843
3844         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3845         avail = block_group->zone_capacity - block_group->alloc_offset;
3846         if (avail < num_bytes) {
3847                 if (ffe_ctl->max_extent_size < avail) {
3848                         /*
3849                          * With sequential allocator, free space is always
3850                          * contiguous
3851                          */
3852                         ffe_ctl->max_extent_size = avail;
3853                         ffe_ctl->total_free_space = avail;
3854                 }
3855                 ret = 1;
3856                 goto out;
3857         }
3858
3859         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3860                 fs_info->treelog_bg = block_group->start;
3861
3862         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg)
3863                 fs_info->data_reloc_bg = block_group->start;
3864
3865         ffe_ctl->found_offset = start + block_group->alloc_offset;
3866         block_group->alloc_offset += num_bytes;
3867         spin_lock(&ctl->tree_lock);
3868         ctl->free_space -= num_bytes;
3869         spin_unlock(&ctl->tree_lock);
3870
3871         /*
3872          * We do not check if found_offset is aligned to stripesize. The
3873          * address is anyway rewritten when using zone append writing.
3874          */
3875
3876         ffe_ctl->search_start = ffe_ctl->found_offset;
3877
3878 out:
3879         if (ret && ffe_ctl->for_treelog)
3880                 fs_info->treelog_bg = 0;
3881         if (ret && ffe_ctl->for_data_reloc)
3882                 fs_info->data_reloc_bg = 0;
3883         spin_unlock(&fs_info->relocation_bg_lock);
3884         spin_unlock(&fs_info->treelog_bg_lock);
3885         spin_unlock(&block_group->lock);
3886         spin_unlock(&space_info->lock);
3887         return ret;
3888 }
3889
3890 static int do_allocation(struct btrfs_block_group *block_group,
3891                          struct find_free_extent_ctl *ffe_ctl,
3892                          struct btrfs_block_group **bg_ret)
3893 {
3894         switch (ffe_ctl->policy) {
3895         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3896                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3897         case BTRFS_EXTENT_ALLOC_ZONED:
3898                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3899         default:
3900                 BUG();
3901         }
3902 }
3903
3904 static void release_block_group(struct btrfs_block_group *block_group,
3905                                 struct find_free_extent_ctl *ffe_ctl,
3906                                 int delalloc)
3907 {
3908         switch (ffe_ctl->policy) {
3909         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3910                 ffe_ctl->retry_clustered = false;
3911                 ffe_ctl->retry_unclustered = false;
3912                 break;
3913         case BTRFS_EXTENT_ALLOC_ZONED:
3914                 /* Nothing to do */
3915                 break;
3916         default:
3917                 BUG();
3918         }
3919
3920         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3921                ffe_ctl->index);
3922         btrfs_release_block_group(block_group, delalloc);
3923 }
3924
3925 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3926                                    struct btrfs_key *ins)
3927 {
3928         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3929
3930         if (!ffe_ctl->use_cluster && last_ptr) {
3931                 spin_lock(&last_ptr->lock);
3932                 last_ptr->window_start = ins->objectid;
3933                 spin_unlock(&last_ptr->lock);
3934         }
3935 }
3936
3937 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3938                          struct btrfs_key *ins)
3939 {
3940         switch (ffe_ctl->policy) {
3941         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3942                 found_extent_clustered(ffe_ctl, ins);
3943                 break;
3944         case BTRFS_EXTENT_ALLOC_ZONED:
3945                 /* Nothing to do */
3946                 break;
3947         default:
3948                 BUG();
3949         }
3950 }
3951
3952 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
3953 {
3954         switch (ffe_ctl->policy) {
3955         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3956                 /*
3957                  * If we can't allocate a new chunk we've already looped through
3958                  * at least once, move on to the NO_EMPTY_SIZE case.
3959                  */
3960                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
3961                 return 0;
3962         case BTRFS_EXTENT_ALLOC_ZONED:
3963                 /* Give up here */
3964                 return -ENOSPC;
3965         default:
3966                 BUG();
3967         }
3968 }
3969
3970 /*
3971  * Return >0 means caller needs to re-search for free extent
3972  * Return 0 means we have the needed free extent.
3973  * Return <0 means we failed to locate any free extent.
3974  */
3975 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3976                                         struct btrfs_key *ins,
3977                                         struct find_free_extent_ctl *ffe_ctl,
3978                                         bool full_search)
3979 {
3980         struct btrfs_root *root = fs_info->extent_root;
3981         int ret;
3982
3983         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3984             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3985                 ffe_ctl->orig_have_caching_bg = true;
3986
3987         if (ins->objectid) {
3988                 found_extent(ffe_ctl, ins);
3989                 return 0;
3990         }
3991
3992         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size &&
3993             !btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->index)) {
3994                 /*
3995                  * If we have enough free space left in an already active block
3996                  * group and we can't activate any other zone now, retry the
3997                  * active ones with a smaller allocation size.  Returning early
3998                  * from here will tell btrfs_reserve_extent() to haven the
3999                  * size.
4000                  */
4001                 return -ENOSPC;
4002         }
4003
4004         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4005                 return 1;
4006
4007         ffe_ctl->index++;
4008         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4009                 return 1;
4010
4011         /*
4012          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
4013          *                      caching kthreads as we move along
4014          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
4015          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
4016          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
4017          *                     again
4018          */
4019         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4020                 ffe_ctl->index = 0;
4021                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
4022                         /*
4023                          * We want to skip the LOOP_CACHING_WAIT step if we
4024                          * don't have any uncached bgs and we've already done a
4025                          * full search through.
4026                          */
4027                         if (ffe_ctl->orig_have_caching_bg || !full_search)
4028                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
4029                         else
4030                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
4031                 } else {
4032                         ffe_ctl->loop++;
4033                 }
4034
4035                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4036                         struct btrfs_trans_handle *trans;
4037                         int exist = 0;
4038
4039                         trans = current->journal_info;
4040                         if (trans)
4041                                 exist = 1;
4042                         else
4043                                 trans = btrfs_join_transaction(root);
4044
4045                         if (IS_ERR(trans)) {
4046                                 ret = PTR_ERR(trans);
4047                                 return ret;
4048                         }
4049
4050                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4051                                                 CHUNK_ALLOC_FORCE);
4052
4053                         /* Do not bail out on ENOSPC since we can do more. */
4054                         if (ret == -ENOSPC)
4055                                 ret = chunk_allocation_failed(ffe_ctl);
4056                         else if (ret < 0)
4057                                 btrfs_abort_transaction(trans, ret);
4058                         else
4059                                 ret = 0;
4060                         if (!exist)
4061                                 btrfs_end_transaction(trans);
4062                         if (ret)
4063                                 return ret;
4064                 }
4065
4066                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4067                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4068                                 return -ENOSPC;
4069
4070                         /*
4071                          * Don't loop again if we already have no empty_size and
4072                          * no empty_cluster.
4073                          */
4074                         if (ffe_ctl->empty_size == 0 &&
4075                             ffe_ctl->empty_cluster == 0)
4076                                 return -ENOSPC;
4077                         ffe_ctl->empty_size = 0;
4078                         ffe_ctl->empty_cluster = 0;
4079                 }
4080                 return 1;
4081         }
4082         return -ENOSPC;
4083 }
4084
4085 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4086                                         struct find_free_extent_ctl *ffe_ctl,
4087                                         struct btrfs_space_info *space_info,
4088                                         struct btrfs_key *ins)
4089 {
4090         /*
4091          * If our free space is heavily fragmented we may not be able to make
4092          * big contiguous allocations, so instead of doing the expensive search
4093          * for free space, simply return ENOSPC with our max_extent_size so we
4094          * can go ahead and search for a more manageable chunk.
4095          *
4096          * If our max_extent_size is large enough for our allocation simply
4097          * disable clustering since we will likely not be able to find enough
4098          * space to create a cluster and induce latency trying.
4099          */
4100         if (space_info->max_extent_size) {
4101                 spin_lock(&space_info->lock);
4102                 if (space_info->max_extent_size &&
4103                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4104                         ins->offset = space_info->max_extent_size;
4105                         spin_unlock(&space_info->lock);
4106                         return -ENOSPC;
4107                 } else if (space_info->max_extent_size) {
4108                         ffe_ctl->use_cluster = false;
4109                 }
4110                 spin_unlock(&space_info->lock);
4111         }
4112
4113         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4114                                                &ffe_ctl->empty_cluster);
4115         if (ffe_ctl->last_ptr) {
4116                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4117
4118                 spin_lock(&last_ptr->lock);
4119                 if (last_ptr->block_group)
4120                         ffe_ctl->hint_byte = last_ptr->window_start;
4121                 if (last_ptr->fragmented) {
4122                         /*
4123                          * We still set window_start so we can keep track of the
4124                          * last place we found an allocation to try and save
4125                          * some time.
4126                          */
4127                         ffe_ctl->hint_byte = last_ptr->window_start;
4128                         ffe_ctl->use_cluster = false;
4129                 }
4130                 spin_unlock(&last_ptr->lock);
4131         }
4132
4133         return 0;
4134 }
4135
4136 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4137                               struct find_free_extent_ctl *ffe_ctl,
4138                               struct btrfs_space_info *space_info,
4139                               struct btrfs_key *ins)
4140 {
4141         switch (ffe_ctl->policy) {
4142         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4143                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4144                                                     space_info, ins);
4145         case BTRFS_EXTENT_ALLOC_ZONED:
4146                 if (ffe_ctl->for_treelog) {
4147                         spin_lock(&fs_info->treelog_bg_lock);
4148                         if (fs_info->treelog_bg)
4149                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4150                         spin_unlock(&fs_info->treelog_bg_lock);
4151                 }
4152                 if (ffe_ctl->for_data_reloc) {
4153                         spin_lock(&fs_info->relocation_bg_lock);
4154                         if (fs_info->data_reloc_bg)
4155                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4156                         spin_unlock(&fs_info->relocation_bg_lock);
4157                 }
4158                 return 0;
4159         default:
4160                 BUG();
4161         }
4162 }
4163
4164 /*
4165  * walks the btree of allocated extents and find a hole of a given size.
4166  * The key ins is changed to record the hole:
4167  * ins->objectid == start position
4168  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4169  * ins->offset == the size of the hole.
4170  * Any available blocks before search_start are skipped.
4171  *
4172  * If there is no suitable free space, we will record the max size of
4173  * the free space extent currently.
4174  *
4175  * The overall logic and call chain:
4176  *
4177  * find_free_extent()
4178  * |- Iterate through all block groups
4179  * |  |- Get a valid block group
4180  * |  |- Try to do clustered allocation in that block group
4181  * |  |- Try to do unclustered allocation in that block group
4182  * |  |- Check if the result is valid
4183  * |  |  |- If valid, then exit
4184  * |  |- Jump to next block group
4185  * |
4186  * |- Push harder to find free extents
4187  *    |- If not found, re-iterate all block groups
4188  */
4189 static noinline int find_free_extent(struct btrfs_root *root,
4190                                      struct btrfs_key *ins,
4191                                      struct find_free_extent_ctl *ffe_ctl)
4192 {
4193         struct btrfs_fs_info *fs_info = root->fs_info;
4194         int ret = 0;
4195         int cache_block_group_error = 0;
4196         struct btrfs_block_group *block_group = NULL;
4197         struct btrfs_space_info *space_info;
4198         bool full_search = false;
4199
4200         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4201
4202         ffe_ctl->search_start = 0;
4203         /* For clustered allocation */
4204         ffe_ctl->empty_cluster = 0;
4205         ffe_ctl->last_ptr = NULL;
4206         ffe_ctl->use_cluster = true;
4207         ffe_ctl->have_caching_bg = false;
4208         ffe_ctl->orig_have_caching_bg = false;
4209         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4210         ffe_ctl->loop = 0;
4211         /* For clustered allocation */
4212         ffe_ctl->retry_clustered = false;
4213         ffe_ctl->retry_unclustered = false;
4214         ffe_ctl->cached = 0;
4215         ffe_ctl->max_extent_size = 0;
4216         ffe_ctl->total_free_space = 0;
4217         ffe_ctl->found_offset = 0;
4218         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4219
4220         if (btrfs_is_zoned(fs_info))
4221                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4222
4223         ins->type = BTRFS_EXTENT_ITEM_KEY;
4224         ins->objectid = 0;
4225         ins->offset = 0;
4226
4227         trace_find_free_extent(root, ffe_ctl->num_bytes, ffe_ctl->empty_size,
4228                                ffe_ctl->flags);
4229
4230         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4231         if (!space_info) {
4232                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4233                 return -ENOSPC;
4234         }
4235
4236         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4237         if (ret < 0)
4238                 return ret;
4239
4240         ffe_ctl->search_start = max(ffe_ctl->search_start,
4241                                     first_logical_byte(fs_info, 0));
4242         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4243         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4244                 block_group = btrfs_lookup_block_group(fs_info,
4245                                                        ffe_ctl->search_start);
4246                 /*
4247                  * we don't want to use the block group if it doesn't match our
4248                  * allocation bits, or if its not cached.
4249                  *
4250                  * However if we are re-searching with an ideal block group
4251                  * picked out then we don't care that the block group is cached.
4252                  */
4253                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4254                     block_group->cached != BTRFS_CACHE_NO) {
4255                         down_read(&space_info->groups_sem);
4256                         if (list_empty(&block_group->list) ||
4257                             block_group->ro) {
4258                                 /*
4259                                  * someone is removing this block group,
4260                                  * we can't jump into the have_block_group
4261                                  * target because our list pointers are not
4262                                  * valid
4263                                  */
4264                                 btrfs_put_block_group(block_group);
4265                                 up_read(&space_info->groups_sem);
4266                         } else {
4267                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4268                                                         block_group->flags);
4269                                 btrfs_lock_block_group(block_group,
4270                                                        ffe_ctl->delalloc);
4271                                 goto have_block_group;
4272                         }
4273                 } else if (block_group) {
4274                         btrfs_put_block_group(block_group);
4275                 }
4276         }
4277 search:
4278         ffe_ctl->have_caching_bg = false;
4279         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4280             ffe_ctl->index == 0)
4281                 full_search = true;
4282         down_read(&space_info->groups_sem);
4283         list_for_each_entry(block_group,
4284                             &space_info->block_groups[ffe_ctl->index], list) {
4285                 struct btrfs_block_group *bg_ret;
4286
4287                 /* If the block group is read-only, we can skip it entirely. */
4288                 if (unlikely(block_group->ro)) {
4289                         if (ffe_ctl->for_treelog)
4290                                 btrfs_clear_treelog_bg(block_group);
4291                         if (ffe_ctl->for_data_reloc)
4292                                 btrfs_clear_data_reloc_bg(block_group);
4293                         continue;
4294                 }
4295
4296                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4297                 ffe_ctl->search_start = block_group->start;
4298
4299                 /*
4300                  * this can happen if we end up cycling through all the
4301                  * raid types, but we want to make sure we only allocate
4302                  * for the proper type.
4303                  */
4304                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4305                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4306                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4307                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4308                                 BTRFS_BLOCK_GROUP_RAID10;
4309
4310                         /*
4311                          * if they asked for extra copies and this block group
4312                          * doesn't provide them, bail.  This does allow us to
4313                          * fill raid0 from raid1.
4314                          */
4315                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4316                                 goto loop;
4317
4318                         /*
4319                          * This block group has different flags than we want.
4320                          * It's possible that we have MIXED_GROUP flag but no
4321                          * block group is mixed.  Just skip such block group.
4322                          */
4323                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4324                         continue;
4325                 }
4326
4327 have_block_group:
4328                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4329                 if (unlikely(!ffe_ctl->cached)) {
4330                         ffe_ctl->have_caching_bg = true;
4331                         ret = btrfs_cache_block_group(block_group, 0);
4332
4333                         /*
4334                          * If we get ENOMEM here or something else we want to
4335                          * try other block groups, because it may not be fatal.
4336                          * However if we can't find anything else we need to
4337                          * save our return here so that we return the actual
4338                          * error that caused problems, not ENOSPC.
4339                          */
4340                         if (ret < 0) {
4341                                 if (!cache_block_group_error)
4342                                         cache_block_group_error = ret;
4343                                 ret = 0;
4344                                 goto loop;
4345                         }
4346                         ret = 0;
4347                 }
4348
4349                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4350                         goto loop;
4351
4352                 bg_ret = NULL;
4353                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4354                 if (ret == 0) {
4355                         if (bg_ret && bg_ret != block_group) {
4356                                 btrfs_release_block_group(block_group,
4357                                                           ffe_ctl->delalloc);
4358                                 block_group = bg_ret;
4359                         }
4360                 } else if (ret == -EAGAIN) {
4361                         goto have_block_group;
4362                 } else if (ret > 0) {
4363                         goto loop;
4364                 }
4365
4366                 /* Checks */
4367                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4368                                                  fs_info->stripesize);
4369
4370                 /* move on to the next group */
4371                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4372                     block_group->start + block_group->length) {
4373                         btrfs_add_free_space_unused(block_group,
4374                                             ffe_ctl->found_offset,
4375                                             ffe_ctl->num_bytes);
4376                         goto loop;
4377                 }
4378
4379                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4380                         btrfs_add_free_space_unused(block_group,
4381                                         ffe_ctl->found_offset,
4382                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4383
4384                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4385                                                ffe_ctl->num_bytes,
4386                                                ffe_ctl->delalloc);
4387                 if (ret == -EAGAIN) {
4388                         btrfs_add_free_space_unused(block_group,
4389                                         ffe_ctl->found_offset,
4390                                         ffe_ctl->num_bytes);
4391                         goto loop;
4392                 }
4393                 btrfs_inc_block_group_reservations(block_group);
4394
4395                 /* we are all good, lets return */
4396                 ins->objectid = ffe_ctl->search_start;
4397                 ins->offset = ffe_ctl->num_bytes;
4398
4399                 trace_btrfs_reserve_extent(block_group, ffe_ctl->search_start,
4400                                            ffe_ctl->num_bytes);
4401                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4402                 break;
4403 loop:
4404                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4405                 cond_resched();
4406         }
4407         up_read(&space_info->groups_sem);
4408
4409         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4410         if (ret > 0)
4411                 goto search;
4412
4413         if (ret == -ENOSPC && !cache_block_group_error) {
4414                 /*
4415                  * Use ffe_ctl->total_free_space as fallback if we can't find
4416                  * any contiguous hole.
4417                  */
4418                 if (!ffe_ctl->max_extent_size)
4419                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4420                 spin_lock(&space_info->lock);
4421                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4422                 spin_unlock(&space_info->lock);
4423                 ins->offset = ffe_ctl->max_extent_size;
4424         } else if (ret == -ENOSPC) {
4425                 ret = cache_block_group_error;
4426         }
4427         return ret;
4428 }
4429
4430 /*
4431  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4432  *                        hole that is at least as big as @num_bytes.
4433  *
4434  * @root           -    The root that will contain this extent
4435  *
4436  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4437  *                      is used for accounting purposes. This value differs
4438  *                      from @num_bytes only in the case of compressed extents.
4439  *
4440  * @num_bytes      -    Number of bytes to allocate on-disk.
4441  *
4442  * @min_alloc_size -    Indicates the minimum amount of space that the
4443  *                      allocator should try to satisfy. In some cases
4444  *                      @num_bytes may be larger than what is required and if
4445  *                      the filesystem is fragmented then allocation fails.
4446  *                      However, the presence of @min_alloc_size gives a
4447  *                      chance to try and satisfy the smaller allocation.
4448  *
4449  * @empty_size     -    A hint that you plan on doing more COW. This is the
4450  *                      size in bytes the allocator should try to find free
4451  *                      next to the block it returns.  This is just a hint and
4452  *                      may be ignored by the allocator.
4453  *
4454  * @hint_byte      -    Hint to the allocator to start searching above the byte
4455  *                      address passed. It might be ignored.
4456  *
4457  * @ins            -    This key is modified to record the found hole. It will
4458  *                      have the following values:
4459  *                      ins->objectid == start position
4460  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4461  *                      ins->offset == the size of the hole.
4462  *
4463  * @is_data        -    Boolean flag indicating whether an extent is
4464  *                      allocated for data (true) or metadata (false)
4465  *
4466  * @delalloc       -    Boolean flag indicating whether this allocation is for
4467  *                      delalloc or not. If 'true' data_rwsem of block groups
4468  *                      is going to be acquired.
4469  *
4470  *
4471  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4472  * case -ENOSPC is returned then @ins->offset will contain the size of the
4473  * largest available hole the allocator managed to find.
4474  */
4475 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4476                          u64 num_bytes, u64 min_alloc_size,
4477                          u64 empty_size, u64 hint_byte,
4478                          struct btrfs_key *ins, int is_data, int delalloc)
4479 {
4480         struct btrfs_fs_info *fs_info = root->fs_info;
4481         struct find_free_extent_ctl ffe_ctl = {};
4482         bool final_tried = num_bytes == min_alloc_size;
4483         u64 flags;
4484         int ret;
4485         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4486         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4487
4488         flags = get_alloc_profile_by_root(root, is_data);
4489 again:
4490         WARN_ON(num_bytes < fs_info->sectorsize);
4491
4492         ffe_ctl.ram_bytes = ram_bytes;
4493         ffe_ctl.num_bytes = num_bytes;
4494         ffe_ctl.min_alloc_size = min_alloc_size;
4495         ffe_ctl.empty_size = empty_size;
4496         ffe_ctl.flags = flags;
4497         ffe_ctl.delalloc = delalloc;
4498         ffe_ctl.hint_byte = hint_byte;
4499         ffe_ctl.for_treelog = for_treelog;
4500         ffe_ctl.for_data_reloc = for_data_reloc;
4501
4502         ret = find_free_extent(root, ins, &ffe_ctl);
4503         if (!ret && !is_data) {
4504                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4505         } else if (ret == -ENOSPC) {
4506                 if (!final_tried && ins->offset) {
4507                         num_bytes = min(num_bytes >> 1, ins->offset);
4508                         num_bytes = round_down(num_bytes,
4509                                                fs_info->sectorsize);
4510                         num_bytes = max(num_bytes, min_alloc_size);
4511                         ram_bytes = num_bytes;
4512                         if (num_bytes == min_alloc_size)
4513                                 final_tried = true;
4514                         goto again;
4515                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4516                         struct btrfs_space_info *sinfo;
4517
4518                         sinfo = btrfs_find_space_info(fs_info, flags);
4519                         btrfs_err(fs_info,
4520         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4521                                   flags, num_bytes, for_treelog, for_data_reloc);
4522                         if (sinfo)
4523                                 btrfs_dump_space_info(fs_info, sinfo,
4524                                                       num_bytes, 1);
4525                 }
4526         }
4527
4528         return ret;
4529 }
4530
4531 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4532                                u64 start, u64 len, int delalloc)
4533 {
4534         struct btrfs_block_group *cache;
4535
4536         cache = btrfs_lookup_block_group(fs_info, start);
4537         if (!cache) {
4538                 btrfs_err(fs_info, "Unable to find block group for %llu",
4539                           start);
4540                 return -ENOSPC;
4541         }
4542
4543         btrfs_add_free_space(cache, start, len);
4544         btrfs_free_reserved_bytes(cache, len, delalloc);
4545         trace_btrfs_reserved_extent_free(fs_info, start, len);
4546
4547         btrfs_put_block_group(cache);
4548         return 0;
4549 }
4550
4551 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4552                               u64 len)
4553 {
4554         struct btrfs_block_group *cache;
4555         int ret = 0;
4556
4557         cache = btrfs_lookup_block_group(trans->fs_info, start);
4558         if (!cache) {
4559                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4560                           start);
4561                 return -ENOSPC;
4562         }
4563
4564         ret = pin_down_extent(trans, cache, start, len, 1);
4565         btrfs_put_block_group(cache);
4566         return ret;
4567 }
4568
4569 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4570                                       u64 parent, u64 root_objectid,
4571                                       u64 flags, u64 owner, u64 offset,
4572                                       struct btrfs_key *ins, int ref_mod)
4573 {
4574         struct btrfs_fs_info *fs_info = trans->fs_info;
4575         int ret;
4576         struct btrfs_extent_item *extent_item;
4577         struct btrfs_extent_inline_ref *iref;
4578         struct btrfs_path *path;
4579         struct extent_buffer *leaf;
4580         int type;
4581         u32 size;
4582
4583         if (parent > 0)
4584                 type = BTRFS_SHARED_DATA_REF_KEY;
4585         else
4586                 type = BTRFS_EXTENT_DATA_REF_KEY;
4587
4588         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4589
4590         path = btrfs_alloc_path();
4591         if (!path)
4592                 return -ENOMEM;
4593
4594         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4595                                       ins, size);
4596         if (ret) {
4597                 btrfs_free_path(path);
4598                 return ret;
4599         }
4600
4601         leaf = path->nodes[0];
4602         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4603                                      struct btrfs_extent_item);
4604         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4605         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4606         btrfs_set_extent_flags(leaf, extent_item,
4607                                flags | BTRFS_EXTENT_FLAG_DATA);
4608
4609         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4610         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4611         if (parent > 0) {
4612                 struct btrfs_shared_data_ref *ref;
4613                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4614                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4615                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4616         } else {
4617                 struct btrfs_extent_data_ref *ref;
4618                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4619                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4620                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4621                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4622                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4623         }
4624
4625         btrfs_mark_buffer_dirty(path->nodes[0]);
4626         btrfs_free_path(path);
4627
4628         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
4629         if (ret)
4630                 return ret;
4631
4632         ret = btrfs_update_block_group(trans, ins->objectid, ins->offset, 1);
4633         if (ret) { /* -ENOENT, logic error */
4634                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4635                         ins->objectid, ins->offset);
4636                 BUG();
4637         }
4638         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
4639         return ret;
4640 }
4641
4642 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4643                                      struct btrfs_delayed_ref_node *node,
4644                                      struct btrfs_delayed_extent_op *extent_op)
4645 {
4646         struct btrfs_fs_info *fs_info = trans->fs_info;
4647         int ret;
4648         struct btrfs_extent_item *extent_item;
4649         struct btrfs_key extent_key;
4650         struct btrfs_tree_block_info *block_info;
4651         struct btrfs_extent_inline_ref *iref;
4652         struct btrfs_path *path;
4653         struct extent_buffer *leaf;
4654         struct btrfs_delayed_tree_ref *ref;
4655         u32 size = sizeof(*extent_item) + sizeof(*iref);
4656         u64 num_bytes;
4657         u64 flags = extent_op->flags_to_set;
4658         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4659
4660         ref = btrfs_delayed_node_to_tree_ref(node);
4661
4662         extent_key.objectid = node->bytenr;
4663         if (skinny_metadata) {
4664                 extent_key.offset = ref->level;
4665                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4666                 num_bytes = fs_info->nodesize;
4667         } else {
4668                 extent_key.offset = node->num_bytes;
4669                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4670                 size += sizeof(*block_info);
4671                 num_bytes = node->num_bytes;
4672         }
4673
4674         path = btrfs_alloc_path();
4675         if (!path)
4676                 return -ENOMEM;
4677
4678         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4679                                       &extent_key, size);
4680         if (ret) {
4681                 btrfs_free_path(path);
4682                 return ret;
4683         }
4684
4685         leaf = path->nodes[0];
4686         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4687                                      struct btrfs_extent_item);
4688         btrfs_set_extent_refs(leaf, extent_item, 1);
4689         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4690         btrfs_set_extent_flags(leaf, extent_item,
4691                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4692
4693         if (skinny_metadata) {
4694                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4695         } else {
4696                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4697                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4698                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4699                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4700         }
4701
4702         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4703                 btrfs_set_extent_inline_ref_type(leaf, iref,
4704                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4705                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4706         } else {
4707                 btrfs_set_extent_inline_ref_type(leaf, iref,
4708                                                  BTRFS_TREE_BLOCK_REF_KEY);
4709                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4710         }
4711
4712         btrfs_mark_buffer_dirty(leaf);
4713         btrfs_free_path(path);
4714
4715         ret = remove_from_free_space_tree(trans, extent_key.objectid,
4716                                           num_bytes);
4717         if (ret)
4718                 return ret;
4719
4720         ret = btrfs_update_block_group(trans, extent_key.objectid,
4721                                        fs_info->nodesize, 1);
4722         if (ret) { /* -ENOENT, logic error */
4723                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4724                         extent_key.objectid, extent_key.offset);
4725                 BUG();
4726         }
4727
4728         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
4729                                           fs_info->nodesize);
4730         return ret;
4731 }
4732
4733 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4734                                      struct btrfs_root *root, u64 owner,
4735                                      u64 offset, u64 ram_bytes,
4736                                      struct btrfs_key *ins)
4737 {
4738         struct btrfs_ref generic_ref = { 0 };
4739
4740         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4741
4742         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4743                                ins->objectid, ins->offset, 0);
4744         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
4745         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4746
4747         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4748 }
4749
4750 /*
4751  * this is used by the tree logging recovery code.  It records that
4752  * an extent has been allocated and makes sure to clear the free
4753  * space cache bits as well
4754  */
4755 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4756                                    u64 root_objectid, u64 owner, u64 offset,
4757                                    struct btrfs_key *ins)
4758 {
4759         struct btrfs_fs_info *fs_info = trans->fs_info;
4760         int ret;
4761         struct btrfs_block_group *block_group;
4762         struct btrfs_space_info *space_info;
4763
4764         /*
4765          * Mixed block groups will exclude before processing the log so we only
4766          * need to do the exclude dance if this fs isn't mixed.
4767          */
4768         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4769                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4770                                               ins->offset);
4771                 if (ret)
4772                         return ret;
4773         }
4774
4775         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4776         if (!block_group)
4777                 return -EINVAL;
4778
4779         space_info = block_group->space_info;
4780         spin_lock(&space_info->lock);
4781         spin_lock(&block_group->lock);
4782         space_info->bytes_reserved += ins->offset;
4783         block_group->reserved += ins->offset;
4784         spin_unlock(&block_group->lock);
4785         spin_unlock(&space_info->lock);
4786
4787         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4788                                          offset, ins, 1);
4789         if (ret)
4790                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4791         btrfs_put_block_group(block_group);
4792         return ret;
4793 }
4794
4795 static struct extent_buffer *
4796 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4797                       u64 bytenr, int level, u64 owner,
4798                       enum btrfs_lock_nesting nest)
4799 {
4800         struct btrfs_fs_info *fs_info = root->fs_info;
4801         struct extent_buffer *buf;
4802
4803         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4804         if (IS_ERR(buf))
4805                 return buf;
4806
4807         /*
4808          * Extra safety check in case the extent tree is corrupted and extent
4809          * allocator chooses to use a tree block which is already used and
4810          * locked.
4811          */
4812         if (buf->lock_owner == current->pid) {
4813                 btrfs_err_rl(fs_info,
4814 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4815                         buf->start, btrfs_header_owner(buf), current->pid);
4816                 free_extent_buffer(buf);
4817                 return ERR_PTR(-EUCLEAN);
4818         }
4819
4820         /*
4821          * This needs to stay, because we could allocate a freed block from an
4822          * old tree into a new tree, so we need to make sure this new block is
4823          * set to the appropriate level and owner.
4824          */
4825         btrfs_set_buffer_lockdep_class(owner, buf, level);
4826         __btrfs_tree_lock(buf, nest);
4827         btrfs_clean_tree_block(buf);
4828         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4829         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4830
4831         set_extent_buffer_uptodate(buf);
4832
4833         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4834         btrfs_set_header_level(buf, level);
4835         btrfs_set_header_bytenr(buf, buf->start);
4836         btrfs_set_header_generation(buf, trans->transid);
4837         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4838         btrfs_set_header_owner(buf, owner);
4839         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4840         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4841         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4842                 buf->log_index = root->log_transid % 2;
4843                 /*
4844                  * we allow two log transactions at a time, use different
4845                  * EXTENT bit to differentiate dirty pages.
4846                  */
4847                 if (buf->log_index == 0)
4848                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4849                                         buf->start + buf->len - 1, GFP_NOFS);
4850                 else
4851                         set_extent_new(&root->dirty_log_pages, buf->start,
4852                                         buf->start + buf->len - 1);
4853         } else {
4854                 buf->log_index = -1;
4855                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4856                          buf->start + buf->len - 1, GFP_NOFS);
4857         }
4858         /* this returns a buffer locked for blocking */
4859         return buf;
4860 }
4861
4862 /*
4863  * finds a free extent and does all the dirty work required for allocation
4864  * returns the tree buffer or an ERR_PTR on error.
4865  */
4866 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4867                                              struct btrfs_root *root,
4868                                              u64 parent, u64 root_objectid,
4869                                              const struct btrfs_disk_key *key,
4870                                              int level, u64 hint,
4871                                              u64 empty_size,
4872                                              enum btrfs_lock_nesting nest)
4873 {
4874         struct btrfs_fs_info *fs_info = root->fs_info;
4875         struct btrfs_key ins;
4876         struct btrfs_block_rsv *block_rsv;
4877         struct extent_buffer *buf;
4878         struct btrfs_delayed_extent_op *extent_op;
4879         struct btrfs_ref generic_ref = { 0 };
4880         u64 flags = 0;
4881         int ret;
4882         u32 blocksize = fs_info->nodesize;
4883         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4884
4885 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4886         if (btrfs_is_testing(fs_info)) {
4887                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4888                                             level, root_objectid, nest);
4889                 if (!IS_ERR(buf))
4890                         root->alloc_bytenr += blocksize;
4891                 return buf;
4892         }
4893 #endif
4894
4895         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4896         if (IS_ERR(block_rsv))
4897                 return ERR_CAST(block_rsv);
4898
4899         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4900                                    empty_size, hint, &ins, 0, 0);
4901         if (ret)
4902                 goto out_unuse;
4903
4904         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4905                                     root_objectid, nest);
4906         if (IS_ERR(buf)) {
4907                 ret = PTR_ERR(buf);
4908                 goto out_free_reserved;
4909         }
4910
4911         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4912                 if (parent == 0)
4913                         parent = ins.objectid;
4914                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4915         } else
4916                 BUG_ON(parent > 0);
4917
4918         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4919                 extent_op = btrfs_alloc_delayed_extent_op();
4920                 if (!extent_op) {
4921                         ret = -ENOMEM;
4922                         goto out_free_buf;
4923                 }
4924                 if (key)
4925                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4926                 else
4927                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4928                 extent_op->flags_to_set = flags;
4929                 extent_op->update_key = skinny_metadata ? false : true;
4930                 extent_op->update_flags = true;
4931                 extent_op->is_data = false;
4932                 extent_op->level = level;
4933
4934                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4935                                        ins.objectid, ins.offset, parent);
4936                 generic_ref.real_root = root->root_key.objectid;
4937                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
4938                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4939                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
4940                 if (ret)
4941                         goto out_free_delayed;
4942         }
4943         return buf;
4944
4945 out_free_delayed:
4946         btrfs_free_delayed_extent_op(extent_op);
4947 out_free_buf:
4948         btrfs_tree_unlock(buf);
4949         free_extent_buffer(buf);
4950 out_free_reserved:
4951         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4952 out_unuse:
4953         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4954         return ERR_PTR(ret);
4955 }
4956
4957 struct walk_control {
4958         u64 refs[BTRFS_MAX_LEVEL];
4959         u64 flags[BTRFS_MAX_LEVEL];
4960         struct btrfs_key update_progress;
4961         struct btrfs_key drop_progress;
4962         int drop_level;
4963         int stage;
4964         int level;
4965         int shared_level;
4966         int update_ref;
4967         int keep_locks;
4968         int reada_slot;
4969         int reada_count;
4970         int restarted;
4971 };
4972
4973 #define DROP_REFERENCE  1
4974 #define UPDATE_BACKREF  2
4975
4976 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
4977                                      struct btrfs_root *root,
4978                                      struct walk_control *wc,
4979                                      struct btrfs_path *path)
4980 {
4981         struct btrfs_fs_info *fs_info = root->fs_info;
4982         u64 bytenr;
4983         u64 generation;
4984         u64 refs;
4985         u64 flags;
4986         u32 nritems;
4987         struct btrfs_key key;
4988         struct extent_buffer *eb;
4989         int ret;
4990         int slot;
4991         int nread = 0;
4992
4993         if (path->slots[wc->level] < wc->reada_slot) {
4994                 wc->reada_count = wc->reada_count * 2 / 3;
4995                 wc->reada_count = max(wc->reada_count, 2);
4996         } else {
4997                 wc->reada_count = wc->reada_count * 3 / 2;
4998                 wc->reada_count = min_t(int, wc->reada_count,
4999                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5000         }
5001
5002         eb = path->nodes[wc->level];
5003         nritems = btrfs_header_nritems(eb);
5004
5005         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5006                 if (nread >= wc->reada_count)
5007                         break;
5008
5009                 cond_resched();
5010                 bytenr = btrfs_node_blockptr(eb, slot);
5011                 generation = btrfs_node_ptr_generation(eb, slot);
5012
5013                 if (slot == path->slots[wc->level])
5014                         goto reada;
5015
5016                 if (wc->stage == UPDATE_BACKREF &&
5017                     generation <= root->root_key.offset)
5018                         continue;
5019
5020                 /* We don't lock the tree block, it's OK to be racy here */
5021                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5022                                                wc->level - 1, 1, &refs,
5023                                                &flags);
5024                 /* We don't care about errors in readahead. */
5025                 if (ret < 0)
5026                         continue;
5027                 BUG_ON(refs == 0);
5028
5029                 if (wc->stage == DROP_REFERENCE) {
5030                         if (refs == 1)
5031                                 goto reada;
5032
5033                         if (wc->level == 1 &&
5034                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5035                                 continue;
5036                         if (!wc->update_ref ||
5037                             generation <= root->root_key.offset)
5038                                 continue;
5039                         btrfs_node_key_to_cpu(eb, &key, slot);
5040                         ret = btrfs_comp_cpu_keys(&key,
5041                                                   &wc->update_progress);
5042                         if (ret < 0)
5043                                 continue;
5044                 } else {
5045                         if (wc->level == 1 &&
5046                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5047                                 continue;
5048                 }
5049 reada:
5050                 btrfs_readahead_node_child(eb, slot);
5051                 nread++;
5052         }
5053         wc->reada_slot = slot;
5054 }
5055
5056 /*
5057  * helper to process tree block while walking down the tree.
5058  *
5059  * when wc->stage == UPDATE_BACKREF, this function updates
5060  * back refs for pointers in the block.
5061  *
5062  * NOTE: return value 1 means we should stop walking down.
5063  */
5064 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5065                                    struct btrfs_root *root,
5066                                    struct btrfs_path *path,
5067                                    struct walk_control *wc, int lookup_info)
5068 {
5069         struct btrfs_fs_info *fs_info = root->fs_info;
5070         int level = wc->level;
5071         struct extent_buffer *eb = path->nodes[level];
5072         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5073         int ret;
5074
5075         if (wc->stage == UPDATE_BACKREF &&
5076             btrfs_header_owner(eb) != root->root_key.objectid)
5077                 return 1;
5078
5079         /*
5080          * when reference count of tree block is 1, it won't increase
5081          * again. once full backref flag is set, we never clear it.
5082          */
5083         if (lookup_info &&
5084             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5085              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5086                 BUG_ON(!path->locks[level]);
5087                 ret = btrfs_lookup_extent_info(trans, fs_info,
5088                                                eb->start, level, 1,
5089                                                &wc->refs[level],
5090                                                &wc->flags[level]);
5091                 BUG_ON(ret == -ENOMEM);
5092                 if (ret)
5093                         return ret;
5094                 BUG_ON(wc->refs[level] == 0);
5095         }
5096
5097         if (wc->stage == DROP_REFERENCE) {
5098                 if (wc->refs[level] > 1)
5099                         return 1;
5100
5101                 if (path->locks[level] && !wc->keep_locks) {
5102                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5103                         path->locks[level] = 0;
5104                 }
5105                 return 0;
5106         }
5107
5108         /* wc->stage == UPDATE_BACKREF */
5109         if (!(wc->flags[level] & flag)) {
5110                 BUG_ON(!path->locks[level]);
5111                 ret = btrfs_inc_ref(trans, root, eb, 1);
5112                 BUG_ON(ret); /* -ENOMEM */
5113                 ret = btrfs_dec_ref(trans, root, eb, 0);
5114                 BUG_ON(ret); /* -ENOMEM */
5115                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5116                                                   btrfs_header_level(eb), 0);
5117                 BUG_ON(ret); /* -ENOMEM */
5118                 wc->flags[level] |= flag;
5119         }
5120
5121         /*
5122          * the block is shared by multiple trees, so it's not good to
5123          * keep the tree lock
5124          */
5125         if (path->locks[level] && level > 0) {
5126                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5127                 path->locks[level] = 0;
5128         }
5129         return 0;
5130 }
5131
5132 /*
5133  * This is used to verify a ref exists for this root to deal with a bug where we
5134  * would have a drop_progress key that hadn't been updated properly.
5135  */
5136 static int check_ref_exists(struct btrfs_trans_handle *trans,
5137                             struct btrfs_root *root, u64 bytenr, u64 parent,
5138                             int level)
5139 {
5140         struct btrfs_path *path;
5141         struct btrfs_extent_inline_ref *iref;
5142         int ret;
5143
5144         path = btrfs_alloc_path();
5145         if (!path)
5146                 return -ENOMEM;
5147
5148         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5149                                     root->fs_info->nodesize, parent,
5150                                     root->root_key.objectid, level, 0);
5151         btrfs_free_path(path);
5152         if (ret == -ENOENT)
5153                 return 0;
5154         if (ret < 0)
5155                 return ret;
5156         return 1;
5157 }
5158
5159 /*
5160  * helper to process tree block pointer.
5161  *
5162  * when wc->stage == DROP_REFERENCE, this function checks
5163  * reference count of the block pointed to. if the block
5164  * is shared and we need update back refs for the subtree
5165  * rooted at the block, this function changes wc->stage to
5166  * UPDATE_BACKREF. if the block is shared and there is no
5167  * need to update back, this function drops the reference
5168  * to the block.
5169  *
5170  * NOTE: return value 1 means we should stop walking down.
5171  */
5172 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5173                                  struct btrfs_root *root,
5174                                  struct btrfs_path *path,
5175                                  struct walk_control *wc, int *lookup_info)
5176 {
5177         struct btrfs_fs_info *fs_info = root->fs_info;
5178         u64 bytenr;
5179         u64 generation;
5180         u64 parent;
5181         struct btrfs_key key;
5182         struct btrfs_key first_key;
5183         struct btrfs_ref ref = { 0 };
5184         struct extent_buffer *next;
5185         int level = wc->level;
5186         int reada = 0;
5187         int ret = 0;
5188         bool need_account = false;
5189
5190         generation = btrfs_node_ptr_generation(path->nodes[level],
5191                                                path->slots[level]);
5192         /*
5193          * if the lower level block was created before the snapshot
5194          * was created, we know there is no need to update back refs
5195          * for the subtree
5196          */
5197         if (wc->stage == UPDATE_BACKREF &&
5198             generation <= root->root_key.offset) {
5199                 *lookup_info = 1;
5200                 return 1;
5201         }
5202
5203         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5204         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
5205                               path->slots[level]);
5206
5207         next = find_extent_buffer(fs_info, bytenr);
5208         if (!next) {
5209                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5210                                 root->root_key.objectid, level - 1);
5211                 if (IS_ERR(next))
5212                         return PTR_ERR(next);
5213                 reada = 1;
5214         }
5215         btrfs_tree_lock(next);
5216
5217         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5218                                        &wc->refs[level - 1],
5219                                        &wc->flags[level - 1]);
5220         if (ret < 0)
5221                 goto out_unlock;
5222
5223         if (unlikely(wc->refs[level - 1] == 0)) {
5224                 btrfs_err(fs_info, "Missing references.");
5225                 ret = -EIO;
5226                 goto out_unlock;
5227         }
5228         *lookup_info = 0;
5229
5230         if (wc->stage == DROP_REFERENCE) {
5231                 if (wc->refs[level - 1] > 1) {
5232                         need_account = true;
5233                         if (level == 1 &&
5234                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5235                                 goto skip;
5236
5237                         if (!wc->update_ref ||
5238                             generation <= root->root_key.offset)
5239                                 goto skip;
5240
5241                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5242                                               path->slots[level]);
5243                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5244                         if (ret < 0)
5245                                 goto skip;
5246
5247                         wc->stage = UPDATE_BACKREF;
5248                         wc->shared_level = level - 1;
5249                 }
5250         } else {
5251                 if (level == 1 &&
5252                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5253                         goto skip;
5254         }
5255
5256         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5257                 btrfs_tree_unlock(next);
5258                 free_extent_buffer(next);
5259                 next = NULL;
5260                 *lookup_info = 1;
5261         }
5262
5263         if (!next) {
5264                 if (reada && level == 1)
5265                         reada_walk_down(trans, root, wc, path);
5266                 next = read_tree_block(fs_info, bytenr, root->root_key.objectid,
5267                                        generation, level - 1, &first_key);
5268                 if (IS_ERR(next)) {
5269                         return PTR_ERR(next);
5270                 } else if (!extent_buffer_uptodate(next)) {
5271                         free_extent_buffer(next);
5272                         return -EIO;
5273                 }
5274                 btrfs_tree_lock(next);
5275         }
5276
5277         level--;
5278         ASSERT(level == btrfs_header_level(next));
5279         if (level != btrfs_header_level(next)) {
5280                 btrfs_err(root->fs_info, "mismatched level");
5281                 ret = -EIO;
5282                 goto out_unlock;
5283         }
5284         path->nodes[level] = next;
5285         path->slots[level] = 0;
5286         path->locks[level] = BTRFS_WRITE_LOCK;
5287         wc->level = level;
5288         if (wc->level == 1)
5289                 wc->reada_slot = 0;
5290         return 0;
5291 skip:
5292         wc->refs[level - 1] = 0;
5293         wc->flags[level - 1] = 0;
5294         if (wc->stage == DROP_REFERENCE) {
5295                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5296                         parent = path->nodes[level]->start;
5297                 } else {
5298                         ASSERT(root->root_key.objectid ==
5299                                btrfs_header_owner(path->nodes[level]));
5300                         if (root->root_key.objectid !=
5301                             btrfs_header_owner(path->nodes[level])) {
5302                                 btrfs_err(root->fs_info,
5303                                                 "mismatched block owner");
5304                                 ret = -EIO;
5305                                 goto out_unlock;
5306                         }
5307                         parent = 0;
5308                 }
5309
5310                 /*
5311                  * If we had a drop_progress we need to verify the refs are set
5312                  * as expected.  If we find our ref then we know that from here
5313                  * on out everything should be correct, and we can clear the
5314                  * ->restarted flag.
5315                  */
5316                 if (wc->restarted) {
5317                         ret = check_ref_exists(trans, root, bytenr, parent,
5318                                                level - 1);
5319                         if (ret < 0)
5320                                 goto out_unlock;
5321                         if (ret == 0)
5322                                 goto no_delete;
5323                         ret = 0;
5324                         wc->restarted = 0;
5325                 }
5326
5327                 /*
5328                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5329                  * already accounted them at merge time (replace_path),
5330                  * thus we could skip expensive subtree trace here.
5331                  */
5332                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5333                     need_account) {
5334                         ret = btrfs_qgroup_trace_subtree(trans, next,
5335                                                          generation, level - 1);
5336                         if (ret) {
5337                                 btrfs_err_rl(fs_info,
5338                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5339                                              ret);
5340                         }
5341                 }
5342
5343                 /*
5344                  * We need to update the next key in our walk control so we can
5345                  * update the drop_progress key accordingly.  We don't care if
5346                  * find_next_key doesn't find a key because that means we're at
5347                  * the end and are going to clean up now.
5348                  */
5349                 wc->drop_level = level;
5350                 find_next_key(path, level, &wc->drop_progress);
5351
5352                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5353                                        fs_info->nodesize, parent);
5354                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
5355                 ret = btrfs_free_extent(trans, &ref);
5356                 if (ret)
5357                         goto out_unlock;
5358         }
5359 no_delete:
5360         *lookup_info = 1;
5361         ret = 1;
5362
5363 out_unlock:
5364         btrfs_tree_unlock(next);
5365         free_extent_buffer(next);
5366
5367         return ret;
5368 }
5369
5370 /*
5371  * helper to process tree block while walking up the tree.
5372  *
5373  * when wc->stage == DROP_REFERENCE, this function drops
5374  * reference count on the block.
5375  *
5376  * when wc->stage == UPDATE_BACKREF, this function changes
5377  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5378  * to UPDATE_BACKREF previously while processing the block.
5379  *
5380  * NOTE: return value 1 means we should stop walking up.
5381  */
5382 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5383                                  struct btrfs_root *root,
5384                                  struct btrfs_path *path,
5385                                  struct walk_control *wc)
5386 {
5387         struct btrfs_fs_info *fs_info = root->fs_info;
5388         int ret;
5389         int level = wc->level;
5390         struct extent_buffer *eb = path->nodes[level];
5391         u64 parent = 0;
5392
5393         if (wc->stage == UPDATE_BACKREF) {
5394                 BUG_ON(wc->shared_level < level);
5395                 if (level < wc->shared_level)
5396                         goto out;
5397
5398                 ret = find_next_key(path, level + 1, &wc->update_progress);
5399                 if (ret > 0)
5400                         wc->update_ref = 0;
5401
5402                 wc->stage = DROP_REFERENCE;
5403                 wc->shared_level = -1;
5404                 path->slots[level] = 0;
5405
5406                 /*
5407                  * check reference count again if the block isn't locked.
5408                  * we should start walking down the tree again if reference
5409                  * count is one.
5410                  */
5411                 if (!path->locks[level]) {
5412                         BUG_ON(level == 0);
5413                         btrfs_tree_lock(eb);
5414                         path->locks[level] = BTRFS_WRITE_LOCK;
5415
5416                         ret = btrfs_lookup_extent_info(trans, fs_info,
5417                                                        eb->start, level, 1,
5418                                                        &wc->refs[level],
5419                                                        &wc->flags[level]);
5420                         if (ret < 0) {
5421                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5422                                 path->locks[level] = 0;
5423                                 return ret;
5424                         }
5425                         BUG_ON(wc->refs[level] == 0);
5426                         if (wc->refs[level] == 1) {
5427                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5428                                 path->locks[level] = 0;
5429                                 return 1;
5430                         }
5431                 }
5432         }
5433
5434         /* wc->stage == DROP_REFERENCE */
5435         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5436
5437         if (wc->refs[level] == 1) {
5438                 if (level == 0) {
5439                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5440                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5441                         else
5442                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5443                         BUG_ON(ret); /* -ENOMEM */
5444                         if (is_fstree(root->root_key.objectid)) {
5445                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5446                                 if (ret) {
5447                                         btrfs_err_rl(fs_info,
5448         "error %d accounting leaf items, quota is out of sync, rescan required",
5449                                              ret);
5450                                 }
5451                         }
5452                 }
5453                 /* make block locked assertion in btrfs_clean_tree_block happy */
5454                 if (!path->locks[level] &&
5455                     btrfs_header_generation(eb) == trans->transid) {
5456                         btrfs_tree_lock(eb);
5457                         path->locks[level] = BTRFS_WRITE_LOCK;
5458                 }
5459                 btrfs_clean_tree_block(eb);
5460         }
5461
5462         if (eb == root->node) {
5463                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5464                         parent = eb->start;
5465                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5466                         goto owner_mismatch;
5467         } else {
5468                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5469                         parent = path->nodes[level + 1]->start;
5470                 else if (root->root_key.objectid !=
5471                          btrfs_header_owner(path->nodes[level + 1]))
5472                         goto owner_mismatch;
5473         }
5474
5475         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
5476 out:
5477         wc->refs[level] = 0;
5478         wc->flags[level] = 0;
5479         return 0;
5480
5481 owner_mismatch:
5482         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5483                      btrfs_header_owner(eb), root->root_key.objectid);
5484         return -EUCLEAN;
5485 }
5486
5487 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5488                                    struct btrfs_root *root,
5489                                    struct btrfs_path *path,
5490                                    struct walk_control *wc)
5491 {
5492         int level = wc->level;
5493         int lookup_info = 1;
5494         int ret;
5495
5496         while (level >= 0) {
5497                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5498                 if (ret > 0)
5499                         break;
5500
5501                 if (level == 0)
5502                         break;
5503
5504                 if (path->slots[level] >=
5505                     btrfs_header_nritems(path->nodes[level]))
5506                         break;
5507
5508                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5509                 if (ret > 0) {
5510                         path->slots[level]++;
5511                         continue;
5512                 } else if (ret < 0)
5513                         return ret;
5514                 level = wc->level;
5515         }
5516         return 0;
5517 }
5518
5519 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5520                                  struct btrfs_root *root,
5521                                  struct btrfs_path *path,
5522                                  struct walk_control *wc, int max_level)
5523 {
5524         int level = wc->level;
5525         int ret;
5526
5527         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5528         while (level < max_level && path->nodes[level]) {
5529                 wc->level = level;
5530                 if (path->slots[level] + 1 <
5531                     btrfs_header_nritems(path->nodes[level])) {
5532                         path->slots[level]++;
5533                         return 0;
5534                 } else {
5535                         ret = walk_up_proc(trans, root, path, wc);
5536                         if (ret > 0)
5537                                 return 0;
5538                         if (ret < 0)
5539                                 return ret;
5540
5541                         if (path->locks[level]) {
5542                                 btrfs_tree_unlock_rw(path->nodes[level],
5543                                                      path->locks[level]);
5544                                 path->locks[level] = 0;
5545                         }
5546                         free_extent_buffer(path->nodes[level]);
5547                         path->nodes[level] = NULL;
5548                         level++;
5549                 }
5550         }
5551         return 1;
5552 }
5553
5554 /*
5555  * drop a subvolume tree.
5556  *
5557  * this function traverses the tree freeing any blocks that only
5558  * referenced by the tree.
5559  *
5560  * when a shared tree block is found. this function decreases its
5561  * reference count by one. if update_ref is true, this function
5562  * also make sure backrefs for the shared block and all lower level
5563  * blocks are properly updated.
5564  *
5565  * If called with for_reloc == 0, may exit early with -EAGAIN
5566  */
5567 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5568 {
5569         struct btrfs_fs_info *fs_info = root->fs_info;
5570         struct btrfs_path *path;
5571         struct btrfs_trans_handle *trans;
5572         struct btrfs_root *tree_root = fs_info->tree_root;
5573         struct btrfs_root_item *root_item = &root->root_item;
5574         struct walk_control *wc;
5575         struct btrfs_key key;
5576         int err = 0;
5577         int ret;
5578         int level;
5579         bool root_dropped = false;
5580
5581         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5582
5583         path = btrfs_alloc_path();
5584         if (!path) {
5585                 err = -ENOMEM;
5586                 goto out;
5587         }
5588
5589         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5590         if (!wc) {
5591                 btrfs_free_path(path);
5592                 err = -ENOMEM;
5593                 goto out;
5594         }
5595
5596         /*
5597          * Use join to avoid potential EINTR from transaction start. See
5598          * wait_reserve_ticket and the whole reservation callchain.
5599          */
5600         if (for_reloc)
5601                 trans = btrfs_join_transaction(tree_root);
5602         else
5603                 trans = btrfs_start_transaction(tree_root, 0);
5604         if (IS_ERR(trans)) {
5605                 err = PTR_ERR(trans);
5606                 goto out_free;
5607         }
5608
5609         err = btrfs_run_delayed_items(trans);
5610         if (err)
5611                 goto out_end_trans;
5612
5613         /*
5614          * This will help us catch people modifying the fs tree while we're
5615          * dropping it.  It is unsafe to mess with the fs tree while it's being
5616          * dropped as we unlock the root node and parent nodes as we walk down
5617          * the tree, assuming nothing will change.  If something does change
5618          * then we'll have stale information and drop references to blocks we've
5619          * already dropped.
5620          */
5621         set_bit(BTRFS_ROOT_DELETING, &root->state);
5622         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5623                 level = btrfs_header_level(root->node);
5624                 path->nodes[level] = btrfs_lock_root_node(root);
5625                 path->slots[level] = 0;
5626                 path->locks[level] = BTRFS_WRITE_LOCK;
5627                 memset(&wc->update_progress, 0,
5628                        sizeof(wc->update_progress));
5629         } else {
5630                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5631                 memcpy(&wc->update_progress, &key,
5632                        sizeof(wc->update_progress));
5633
5634                 level = btrfs_root_drop_level(root_item);
5635                 BUG_ON(level == 0);
5636                 path->lowest_level = level;
5637                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5638                 path->lowest_level = 0;
5639                 if (ret < 0) {
5640                         err = ret;
5641                         goto out_end_trans;
5642                 }
5643                 WARN_ON(ret > 0);
5644
5645                 /*
5646                  * unlock our path, this is safe because only this
5647                  * function is allowed to delete this snapshot
5648                  */
5649                 btrfs_unlock_up_safe(path, 0);
5650
5651                 level = btrfs_header_level(root->node);
5652                 while (1) {
5653                         btrfs_tree_lock(path->nodes[level]);
5654                         path->locks[level] = BTRFS_WRITE_LOCK;
5655
5656                         ret = btrfs_lookup_extent_info(trans, fs_info,
5657                                                 path->nodes[level]->start,
5658                                                 level, 1, &wc->refs[level],
5659                                                 &wc->flags[level]);
5660                         if (ret < 0) {
5661                                 err = ret;
5662                                 goto out_end_trans;
5663                         }
5664                         BUG_ON(wc->refs[level] == 0);
5665
5666                         if (level == btrfs_root_drop_level(root_item))
5667                                 break;
5668
5669                         btrfs_tree_unlock(path->nodes[level]);
5670                         path->locks[level] = 0;
5671                         WARN_ON(wc->refs[level] != 1);
5672                         level--;
5673                 }
5674         }
5675
5676         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5677         wc->level = level;
5678         wc->shared_level = -1;
5679         wc->stage = DROP_REFERENCE;
5680         wc->update_ref = update_ref;
5681         wc->keep_locks = 0;
5682         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5683
5684         while (1) {
5685
5686                 ret = walk_down_tree(trans, root, path, wc);
5687                 if (ret < 0) {
5688                         err = ret;
5689                         break;
5690                 }
5691
5692                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5693                 if (ret < 0) {
5694                         err = ret;
5695                         break;
5696                 }
5697
5698                 if (ret > 0) {
5699                         BUG_ON(wc->stage != DROP_REFERENCE);
5700                         break;
5701                 }
5702
5703                 if (wc->stage == DROP_REFERENCE) {
5704                         wc->drop_level = wc->level;
5705                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5706                                               &wc->drop_progress,
5707                                               path->slots[wc->drop_level]);
5708                 }
5709                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5710                                       &wc->drop_progress);
5711                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5712
5713                 BUG_ON(wc->level == 0);
5714                 if (btrfs_should_end_transaction(trans) ||
5715                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5716                         ret = btrfs_update_root(trans, tree_root,
5717                                                 &root->root_key,
5718                                                 root_item);
5719                         if (ret) {
5720                                 btrfs_abort_transaction(trans, ret);
5721                                 err = ret;
5722                                 goto out_end_trans;
5723                         }
5724
5725                         btrfs_end_transaction_throttle(trans);
5726                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5727                                 btrfs_debug(fs_info,
5728                                             "drop snapshot early exit");
5729                                 err = -EAGAIN;
5730                                 goto out_free;
5731                         }
5732
5733                        /*
5734                         * Use join to avoid potential EINTR from transaction
5735                         * start. See wait_reserve_ticket and the whole
5736                         * reservation callchain.
5737                         */
5738                         if (for_reloc)
5739                                 trans = btrfs_join_transaction(tree_root);
5740                         else
5741                                 trans = btrfs_start_transaction(tree_root, 0);
5742                         if (IS_ERR(trans)) {
5743                                 err = PTR_ERR(trans);
5744                                 goto out_free;
5745                         }
5746                 }
5747         }
5748         btrfs_release_path(path);
5749         if (err)
5750                 goto out_end_trans;
5751
5752         ret = btrfs_del_root(trans, &root->root_key);
5753         if (ret) {
5754                 btrfs_abort_transaction(trans, ret);
5755                 err = ret;
5756                 goto out_end_trans;
5757         }
5758
5759         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5760                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5761                                       NULL, NULL);
5762                 if (ret < 0) {
5763                         btrfs_abort_transaction(trans, ret);
5764                         err = ret;
5765                         goto out_end_trans;
5766                 } else if (ret > 0) {
5767                         /* if we fail to delete the orphan item this time
5768                          * around, it'll get picked up the next time.
5769                          *
5770                          * The most common failure here is just -ENOENT.
5771                          */
5772                         btrfs_del_orphan_item(trans, tree_root,
5773                                               root->root_key.objectid);
5774                 }
5775         }
5776
5777         /*
5778          * This subvolume is going to be completely dropped, and won't be
5779          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5780          * commit transaction time.  So free it here manually.
5781          */
5782         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5783         btrfs_qgroup_free_meta_all_pertrans(root);
5784
5785         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5786                 btrfs_add_dropped_root(trans, root);
5787         else
5788                 btrfs_put_root(root);
5789         root_dropped = true;
5790 out_end_trans:
5791         btrfs_end_transaction_throttle(trans);
5792 out_free:
5793         kfree(wc);
5794         btrfs_free_path(path);
5795 out:
5796         /*
5797          * So if we need to stop dropping the snapshot for whatever reason we
5798          * need to make sure to add it back to the dead root list so that we
5799          * keep trying to do the work later.  This also cleans up roots if we
5800          * don't have it in the radix (like when we recover after a power fail
5801          * or unmount) so we don't leak memory.
5802          */
5803         if (!for_reloc && !root_dropped)
5804                 btrfs_add_dead_root(root);
5805         return err;
5806 }
5807
5808 /*
5809  * drop subtree rooted at tree block 'node'.
5810  *
5811  * NOTE: this function will unlock and release tree block 'node'
5812  * only used by relocation code
5813  */
5814 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5815                         struct btrfs_root *root,
5816                         struct extent_buffer *node,
5817                         struct extent_buffer *parent)
5818 {
5819         struct btrfs_fs_info *fs_info = root->fs_info;
5820         struct btrfs_path *path;
5821         struct walk_control *wc;
5822         int level;
5823         int parent_level;
5824         int ret = 0;
5825         int wret;
5826
5827         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5828
5829         path = btrfs_alloc_path();
5830         if (!path)
5831                 return -ENOMEM;
5832
5833         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5834         if (!wc) {
5835                 btrfs_free_path(path);
5836                 return -ENOMEM;
5837         }
5838
5839         btrfs_assert_tree_locked(parent);
5840         parent_level = btrfs_header_level(parent);
5841         atomic_inc(&parent->refs);
5842         path->nodes[parent_level] = parent;
5843         path->slots[parent_level] = btrfs_header_nritems(parent);
5844
5845         btrfs_assert_tree_locked(node);
5846         level = btrfs_header_level(node);
5847         path->nodes[level] = node;
5848         path->slots[level] = 0;
5849         path->locks[level] = BTRFS_WRITE_LOCK;
5850
5851         wc->refs[parent_level] = 1;
5852         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5853         wc->level = level;
5854         wc->shared_level = -1;
5855         wc->stage = DROP_REFERENCE;
5856         wc->update_ref = 0;
5857         wc->keep_locks = 1;
5858         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5859
5860         while (1) {
5861                 wret = walk_down_tree(trans, root, path, wc);
5862                 if (wret < 0) {
5863                         ret = wret;
5864                         break;
5865                 }
5866
5867                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5868                 if (wret < 0)
5869                         ret = wret;
5870                 if (wret != 0)
5871                         break;
5872         }
5873
5874         kfree(wc);
5875         btrfs_free_path(path);
5876         return ret;
5877 }
5878
5879 /*
5880  * helper to account the unused space of all the readonly block group in the
5881  * space_info. takes mirrors into account.
5882  */
5883 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
5884 {
5885         struct btrfs_block_group *block_group;
5886         u64 free_bytes = 0;
5887         int factor;
5888
5889         /* It's df, we don't care if it's racy */
5890         if (list_empty(&sinfo->ro_bgs))
5891                 return 0;
5892
5893         spin_lock(&sinfo->lock);
5894         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
5895                 spin_lock(&block_group->lock);
5896
5897                 if (!block_group->ro) {
5898                         spin_unlock(&block_group->lock);
5899                         continue;
5900                 }
5901
5902                 factor = btrfs_bg_type_to_factor(block_group->flags);
5903                 free_bytes += (block_group->length -
5904                                block_group->used) * factor;
5905
5906                 spin_unlock(&block_group->lock);
5907         }
5908         spin_unlock(&sinfo->lock);
5909
5910         return free_bytes;
5911 }
5912
5913 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5914                                    u64 start, u64 end)
5915 {
5916         return unpin_extent_range(fs_info, start, end, false);
5917 }
5918
5919 /*
5920  * It used to be that old block groups would be left around forever.
5921  * Iterating over them would be enough to trim unused space.  Since we
5922  * now automatically remove them, we also need to iterate over unallocated
5923  * space.
5924  *
5925  * We don't want a transaction for this since the discard may take a
5926  * substantial amount of time.  We don't require that a transaction be
5927  * running, but we do need to take a running transaction into account
5928  * to ensure that we're not discarding chunks that were released or
5929  * allocated in the current transaction.
5930  *
5931  * Holding the chunks lock will prevent other threads from allocating
5932  * or releasing chunks, but it won't prevent a running transaction
5933  * from committing and releasing the memory that the pending chunks
5934  * list head uses.  For that, we need to take a reference to the
5935  * transaction and hold the commit root sem.  We only need to hold
5936  * it while performing the free space search since we have already
5937  * held back allocations.
5938  */
5939 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5940 {
5941         u64 start = SZ_1M, len = 0, end = 0;
5942         int ret;
5943
5944         *trimmed = 0;
5945
5946         /* Discard not supported = nothing to do. */
5947         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
5948                 return 0;
5949
5950         /* Not writable = nothing to do. */
5951         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5952                 return 0;
5953
5954         /* No free space = nothing to do. */
5955         if (device->total_bytes <= device->bytes_used)
5956                 return 0;
5957
5958         ret = 0;
5959
5960         while (1) {
5961                 struct btrfs_fs_info *fs_info = device->fs_info;
5962                 u64 bytes;
5963
5964                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5965                 if (ret)
5966                         break;
5967
5968                 find_first_clear_extent_bit(&device->alloc_state, start,
5969                                             &start, &end,
5970                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5971
5972                 /* Check if there are any CHUNK_* bits left */
5973                 if (start > device->total_bytes) {
5974                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
5975                         btrfs_warn_in_rcu(fs_info,
5976 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
5977                                           start, end - start + 1,
5978                                           rcu_str_deref(device->name),
5979                                           device->total_bytes);
5980                         mutex_unlock(&fs_info->chunk_mutex);
5981                         ret = 0;
5982                         break;
5983                 }
5984
5985                 /* Ensure we skip the reserved area in the first 1M */
5986                 start = max_t(u64, start, SZ_1M);
5987
5988                 /*
5989                  * If find_first_clear_extent_bit find a range that spans the
5990                  * end of the device it will set end to -1, in this case it's up
5991                  * to the caller to trim the value to the size of the device.
5992                  */
5993                 end = min(end, device->total_bytes - 1);
5994
5995                 len = end - start + 1;
5996
5997                 /* We didn't find any extents */
5998                 if (!len) {
5999                         mutex_unlock(&fs_info->chunk_mutex);
6000                         ret = 0;
6001                         break;
6002                 }
6003
6004                 ret = btrfs_issue_discard(device->bdev, start, len,
6005                                           &bytes);
6006                 if (!ret)
6007                         set_extent_bits(&device->alloc_state, start,
6008                                         start + bytes - 1,
6009                                         CHUNK_TRIMMED);
6010                 mutex_unlock(&fs_info->chunk_mutex);
6011
6012                 if (ret)
6013                         break;
6014
6015                 start += len;
6016                 *trimmed += bytes;
6017
6018                 if (fatal_signal_pending(current)) {
6019                         ret = -ERESTARTSYS;
6020                         break;
6021                 }
6022
6023                 cond_resched();
6024         }
6025
6026         return ret;
6027 }
6028
6029 /*
6030  * Trim the whole filesystem by:
6031  * 1) trimming the free space in each block group
6032  * 2) trimming the unallocated space on each device
6033  *
6034  * This will also continue trimming even if a block group or device encounters
6035  * an error.  The return value will be the last error, or 0 if nothing bad
6036  * happens.
6037  */
6038 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6039 {
6040         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6041         struct btrfs_block_group *cache = NULL;
6042         struct btrfs_device *device;
6043         u64 group_trimmed;
6044         u64 range_end = U64_MAX;
6045         u64 start;
6046         u64 end;
6047         u64 trimmed = 0;
6048         u64 bg_failed = 0;
6049         u64 dev_failed = 0;
6050         int bg_ret = 0;
6051         int dev_ret = 0;
6052         int ret = 0;
6053
6054         /*
6055          * Check range overflow if range->len is set.
6056          * The default range->len is U64_MAX.
6057          */
6058         if (range->len != U64_MAX &&
6059             check_add_overflow(range->start, range->len, &range_end))
6060                 return -EINVAL;
6061
6062         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6063         for (; cache; cache = btrfs_next_block_group(cache)) {
6064                 if (cache->start >= range_end) {
6065                         btrfs_put_block_group(cache);
6066                         break;
6067                 }
6068
6069                 start = max(range->start, cache->start);
6070                 end = min(range_end, cache->start + cache->length);
6071
6072                 if (end - start >= range->minlen) {
6073                         if (!btrfs_block_group_done(cache)) {
6074                                 ret = btrfs_cache_block_group(cache, 0);
6075                                 if (ret) {
6076                                         bg_failed++;
6077                                         bg_ret = ret;
6078                                         continue;
6079                                 }
6080                                 ret = btrfs_wait_block_group_cache_done(cache);
6081                                 if (ret) {
6082                                         bg_failed++;
6083                                         bg_ret = ret;
6084                                         continue;
6085                                 }
6086                         }
6087                         ret = btrfs_trim_block_group(cache,
6088                                                      &group_trimmed,
6089                                                      start,
6090                                                      end,
6091                                                      range->minlen);
6092
6093                         trimmed += group_trimmed;
6094                         if (ret) {
6095                                 bg_failed++;
6096                                 bg_ret = ret;
6097                                 continue;
6098                         }
6099                 }
6100         }
6101
6102         if (bg_failed)
6103                 btrfs_warn(fs_info,
6104                         "failed to trim %llu block group(s), last error %d",
6105                         bg_failed, bg_ret);
6106
6107         mutex_lock(&fs_devices->device_list_mutex);
6108         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6109                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6110                         continue;
6111
6112                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6113                 if (ret) {
6114                         dev_failed++;
6115                         dev_ret = ret;
6116                         break;
6117                 }
6118
6119                 trimmed += group_trimmed;
6120         }
6121         mutex_unlock(&fs_devices->device_list_mutex);
6122
6123         if (dev_failed)
6124                 btrfs_warn(fs_info,
6125                         "failed to trim %llu device(s), last error %d",
6126                         dev_failed, dev_ret);
6127         range->len = trimmed;
6128         if (bg_ret)
6129                 return bg_ret;
6130         return dev_ret;
6131 }