fs/btrfs/backref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/rbtree.h>
   8 #include <trace/events/btrfs.h>
   9 #include "ctree.h"
  10 #include "disk-io.h"
  11 #include "backref.h"
  12 #include "ulist.h"
  13 #include "transaction.h"
  14 #include "delayed-ref.h"
  15 #include "locking.h"
  16 #include "misc.h"
  17
  18 /* Just an arbitrary number so we can be sure this happened */
  19 #define BACKREF_FOUND_SHARED 6
  20
  21 struct extent_inode_elem {
  22         u64 inum;
  23         u64 offset;
  24         struct extent_inode_elem *next;
  25 };
  26
  27 static int check_extent_in_eb(const struct btrfs_key *key,
  28                               const struct extent_buffer *eb,
  29                               const struct btrfs_file_extent_item *fi,
  30                               u64 extent_item_pos,
  31                               struct extent_inode_elem **eie,
  32                               bool ignore_offset)
  33 {
  34         u64 offset = 0;
  35         struct extent_inode_elem *e;
  36
  37         if (!ignore_offset &&
  38             !btrfs_file_extent_compression(eb, fi) &&
  39             !btrfs_file_extent_encryption(eb, fi) &&
  40             !btrfs_file_extent_other_encoding(eb, fi)) {
  41                 u64 data_offset;
  42                 u64 data_len;
  43
  44                 data_offset = btrfs_file_extent_offset(eb, fi);
  45                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  46
  47                 if (extent_item_pos < data_offset ||
  48                     extent_item_pos >= data_offset + data_len)
  49                         return 1;
  50                 offset = extent_item_pos - data_offset;
  51         }
  52
  53         e = kmalloc(sizeof(*e), GFP_NOFS);
  54         if (!e)
  55                 return -ENOMEM;
  56
  57         e->next = *eie;
  58         e->inum = key->objectid;
  59         e->offset = key->offset + offset;
  60         *eie = e;
  61
  62         return 0;
  63 }
  64
  65 static void free_inode_elem_list(struct extent_inode_elem *eie)
  66 {
  67         struct extent_inode_elem *eie_next;
  68
  69         for (; eie; eie = eie_next) {
  70                 eie_next = eie->next;
  71                 kfree(eie);
  72         }
  73 }
  74
  75 static int find_extent_in_eb(const struct extent_buffer *eb,
  76                              u64 wanted_disk_byte, u64 extent_item_pos,
  77                              struct extent_inode_elem **eie,
  78                              bool ignore_offset)
  79 {
  80         u64 disk_byte;
  81         struct btrfs_key key;
  82         struct btrfs_file_extent_item *fi;
  83         int slot;
  84         int nritems;
  85         int extent_type;
  86         int ret;
  87
  88         /*
  89          * from the shared data ref, we only have the leaf but we need
  90          * the key. thus, we must look into all items and see that we
  91          * find one (some) with a reference to our extent item.
  92          */
  93         nritems = btrfs_header_nritems(eb);
  94         for (slot = 0; slot < nritems; ++slot) {
  95                 btrfs_item_key_to_cpu(eb, &key, slot);
  96                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  97                         continue;
  98                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
  99                 extent_type = btrfs_file_extent_type(eb, fi);
 100                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 101                         continue;
 102                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 103                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 104                 if (disk_byte != wanted_disk_byte)
 105                         continue;
 106
 107                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 108                 if (ret < 0)
 109                         return ret;
 110         }
 111
 112         return 0;
 113 }
 114
 115 struct preftree {
 116         struct rb_root_cached root;
 117         unsigned int count;
 118 };
 119
 120 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 121
 122 struct preftrees {
 123         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 124         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 125         struct preftree indirect_missing_keys;
 126 };
 127
 128 /*
 129  * Checks for a shared extent during backref search.
 130  *
 131  * The share_count tracks prelim_refs (direct and indirect) having a
 132  * ref->count >0:
 133  *  - incremented when a ref->count transitions to >0
 134  *  - decremented when a ref->count transitions to <1
 135  */
 136 struct share_check {
 137         u64 root_objectid;
 138         u64 inum;
 139         int share_count;
 140 };
 141
 142 static inline int extent_is_shared(struct share_check *sc)
 143 {
 144         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 145 }
 146
 147 static struct kmem_cache *btrfs_prelim_ref_cache;
 148
 149 int __init btrfs_prelim_ref_init(void)
 150 {
 151         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 152                                         sizeof(struct prelim_ref),
 153                                         0,
 154                                         SLAB_MEM_SPREAD,
 155                                         NULL);
 156         if (!btrfs_prelim_ref_cache)
 157                 return -ENOMEM;
 158         return 0;
 159 }
 160
 161 void __cold btrfs_prelim_ref_exit(void)
 162 {
 163         kmem_cache_destroy(btrfs_prelim_ref_cache);
 164 }
 165
 166 static void free_pref(struct prelim_ref *ref)
 167 {
 168         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 169 }
 170
 171 /*
 172  * Return 0 when both refs are for the same block (and can be merged).
 173  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 174  * indicates a 'higher' block.
 175  */
 176 static int prelim_ref_compare(struct prelim_ref *ref1,
 177                               struct prelim_ref *ref2)
 178 {
 179         if (ref1->level < ref2->level)
 180                 return -1;
 181         if (ref1->level > ref2->level)
 182                 return 1;
 183         if (ref1->root_id < ref2->root_id)
 184                 return -1;
 185         if (ref1->root_id > ref2->root_id)
 186                 return 1;
 187         if (ref1->key_for_search.type < ref2->key_for_search.type)
 188                 return -1;
 189         if (ref1->key_for_search.type > ref2->key_for_search.type)
 190                 return 1;
 191         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 192                 return -1;
 193         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 194                 return 1;
 195         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 196                 return -1;
 197         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 198                 return 1;
 199         if (ref1->parent < ref2->parent)
 200                 return -1;
 201         if (ref1->parent > ref2->parent)
 202                 return 1;
 203
 204         return 0;
 205 }
 206
 207 static void update_share_count(struct share_check *sc, int oldcount,
 208                                int newcount)
 209 {
 210         if ((!sc) || (oldcount == 0 && newcount < 1))
 211                 return;
 212
 213         if (oldcount > 0 && newcount < 1)
 214                 sc->share_count--;
 215         else if (oldcount < 1 && newcount > 0)
 216                 sc->share_count++;
 217 }
 218
 219 /*
 220  * Add @newref to the @root rbtree, merging identical refs.
 221  *
 222  * Callers should assume that newref has been freed after calling.
 223  */
 224 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 225                               struct preftree *preftree,
 226                               struct prelim_ref *newref,
 227                               struct share_check *sc)
 228 {
 229         struct rb_root_cached *root;
 230         struct rb_node **p;
 231         struct rb_node *parent = NULL;
 232         struct prelim_ref *ref;
 233         int result;
 234         bool leftmost = true;
 235
 236         root = &preftree->root;
 237         p = &root->rb_root.rb_node;
 238
 239         while (*p) {
 240                 parent = *p;
 241                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 242                 result = prelim_ref_compare(ref, newref);
 243                 if (result < 0) {
 244                         p = &(*p)->rb_left;
 245                 } else if (result > 0) {
 246                         p = &(*p)->rb_right;
 247                         leftmost = false;
 248                 } else {
 249                         /* Identical refs, merge them and free @newref */
 250                         struct extent_inode_elem *eie = ref->inode_list;
 251
 252                         while (eie && eie->next)
 253                                 eie = eie->next;
 254
 255                         if (!eie)
 256                                 ref->inode_list = newref->inode_list;
 257                         else
 258                                 eie->next = newref->inode_list;
 259                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 260                                                      preftree->count);
 261                         /*
 262                          * A delayed ref can have newref->count < 0.
 263                          * The ref->count is updated to follow any
 264                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 265                          */
 266                         update_share_count(sc, ref->count,
 267                                            ref->count + newref->count);
 268                         ref->count += newref->count;
 269                         free_pref(newref);
 270                         return;
 271                 }
 272         }
 273
 274         update_share_count(sc, 0, newref->count);
 275         preftree->count++;
 276         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 277         rb_link_node(&newref->rbnode, parent, p);
 278         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 279 }
 280
 281 /*
 282  * Release the entire tree.  We don't care about internal consistency so
 283  * just free everything and then reset the tree root.
 284  */
 285 static void prelim_release(struct preftree *preftree)
 286 {
 287         struct prelim_ref *ref, *next_ref;
 288
 289         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 290                                              &preftree->root.rb_root, rbnode)
 291                 free_pref(ref);
 292
 293         preftree->root = RB_ROOT_CACHED;
 294         preftree->count = 0;
 295 }
 296
 297 /*
 298  * the rules for all callers of this function are:
 299  * - obtaining the parent is the goal
 300  * - if you add a key, you must know that it is a correct key
 301  * - if you cannot add the parent or a correct key, then we will look into the
 302  *   block later to set a correct key
 303  *
 304  * delayed refs
 305  * ============
 306  *        backref type | shared | indirect | shared | indirect
 307  * information         |   tree |     tree |   data |     data
 308  * --------------------+--------+----------+--------+----------
 309  *      parent logical |    y   |     -    |    -   |     -
 310  *      key to resolve |    -   |     y    |    y   |     y
 311  *  tree block logical |    -   |     -    |    -   |     -
 312  *  root for resolving |    y   |     y    |    y   |     y
 313  *
 314  * - column 1:       we've the parent -> done
 315  * - column 2, 3, 4: we use the key to find the parent
 316  *
 317  * on disk refs (inline or keyed)
 318  * ==============================
 319  *        backref type | shared | indirect | shared | indirect
 320  * information         |   tree |     tree |   data |     data
 321  * --------------------+--------+----------+--------+----------
 322  *      parent logical |    y   |     -    |    y   |     -
 323  *      key to resolve |    -   |     -    |    -   |     y
 324  *  tree block logical |    y   |     y    |    y   |     y
 325  *  root for resolving |    -   |     y    |    y   |     y
 326  *
 327  * - column 1, 3: we've the parent -> done
 328  * - column 2:    we take the first key from the block to find the parent
 329  *                (see add_missing_keys)
 330  * - column 4:    we use the key to find the parent
 331  *
 332  * additional information that's available but not required to find the parent
 333  * block might help in merging entries to gain some speed.
 334  */
 335 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 336                           struct preftree *preftree, u64 root_id,
 337                           const struct btrfs_key *key, int level, u64 parent,
 338                           u64 wanted_disk_byte, int count,
 339                           struct share_check *sc, gfp_t gfp_mask)
 340 {
 341         struct prelim_ref *ref;
 342
 343         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 344                 return 0;
 345
 346         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 347         if (!ref)
 348                 return -ENOMEM;
 349
 350         ref->root_id = root_id;
 351         if (key)
 352                 ref->key_for_search = *key;
 353         else
 354                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 355
 356         ref->inode_list = NULL;
 357         ref->level = level;
 358         ref->count = count;
 359         ref->parent = parent;
 360         ref->wanted_disk_byte = wanted_disk_byte;
 361         prelim_ref_insert(fs_info, preftree, ref, sc);
 362         return extent_is_shared(sc);
 363 }
 364
 365 /* direct refs use root == 0, key == NULL */
 366 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 367                           struct preftrees *preftrees, int level, u64 parent,
 368                           u64 wanted_disk_byte, int count,
 369                           struct share_check *sc, gfp_t gfp_mask)
 370 {
 371         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 372                               parent, wanted_disk_byte, count, sc, gfp_mask);
 373 }
 374
 375 /* indirect refs use parent == 0 */
 376 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 377                             struct preftrees *preftrees, u64 root_id,
 378                             const struct btrfs_key *key, int level,
 379                             u64 wanted_disk_byte, int count,
 380                             struct share_check *sc, gfp_t gfp_mask)
 381 {
 382         struct preftree *tree = &preftrees->indirect;
 383
 384         if (!key)
 385                 tree = &preftrees->indirect_missing_keys;
 386         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 387                               wanted_disk_byte, count, sc, gfp_mask);
 388 }
 389
 390 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 391 {
 392         struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 393         struct rb_node *parent = NULL;
 394         struct prelim_ref *ref = NULL;
 395         struct prelim_ref target = {};
 396         int result;
 397
 398         target.parent = bytenr;
 399
 400         while (*p) {
 401                 parent = *p;
 402                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 403                 result = prelim_ref_compare(ref, &target);
 404
 405                 if (result < 0)
 406                         p = &(*p)->rb_left;
 407                 else if (result > 0)
 408                         p = &(*p)->rb_right;
 409                 else
 410                         return 1;
 411         }
 412         return 0;
 413 }
 414
 415 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 416                            struct ulist *parents,
 417                            struct preftrees *preftrees, struct prelim_ref *ref,
 418                            int level, u64 time_seq, const u64 *extent_item_pos,
 419                            bool ignore_offset)
 420 {
 421         int ret = 0;
 422         int slot;
 423         struct extent_buffer *eb;
 424         struct btrfs_key key;
 425         struct btrfs_key *key_for_search = &ref->key_for_search;
 426         struct btrfs_file_extent_item *fi;
 427         struct extent_inode_elem *eie = NULL, *old = NULL;
 428         u64 disk_byte;
 429         u64 wanted_disk_byte = ref->wanted_disk_byte;
 430         u64 count = 0;
 431         u64 data_offset;
 432
 433         if (level != 0) {
 434                 eb = path->nodes[level];
 435                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 436                 if (ret < 0)
 437                         return ret;
 438                 return 0;
 439         }
 440
 441         /*
 442          * 1. We normally enter this function with the path already pointing to
 443          *    the first item to check. But sometimes, we may enter it with
 444          *    slot == nritems.
 445          * 2. We are searching for normal backref but bytenr of this leaf
 446          *    matches shared data backref
 447          * 3. The leaf owner is not equal to the root we are searching
 448          *
 449          * For these cases, go to the next leaf before we continue.
 450          */
 451         eb = path->nodes[0];
 452         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 453             is_shared_data_backref(preftrees, eb->start) ||
 454             ref->root_id != btrfs_header_owner(eb)) {
 455                 if (time_seq == SEQ_LAST)
 456                         ret = btrfs_next_leaf(root, path);
 457                 else
 458                         ret = btrfs_next_old_leaf(root, path, time_seq);
 459         }
 460
 461         while (!ret && count < ref->count) {
 462                 eb = path->nodes[0];
 463                 slot = path->slots[0];
 464
 465                 btrfs_item_key_to_cpu(eb, &key, slot);
 466
 467                 if (key.objectid != key_for_search->objectid ||
 468                     key.type != BTRFS_EXTENT_DATA_KEY)
 469                         break;
 470
 471                 /*
 472                  * We are searching for normal backref but bytenr of this leaf
 473                  * matches shared data backref, OR
 474                  * the leaf owner is not equal to the root we are searching for
 475                  */
 476                 if (slot == 0 &&
 477                     (is_shared_data_backref(preftrees, eb->start) ||
 478                      ref->root_id != btrfs_header_owner(eb))) {
 479                         if (time_seq == SEQ_LAST)
 480                                 ret = btrfs_next_leaf(root, path);
 481                         else
 482                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 483                         continue;
 484                 }
 485                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 486                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 487                 data_offset = btrfs_file_extent_offset(eb, fi);
 488
 489                 if (disk_byte == wanted_disk_byte) {
 490                         eie = NULL;
 491                         old = NULL;
 492                         if (ref->key_for_search.offset == key.offset - data_offset)
 493                                 count++;
 494                         else
 495                                 goto next;
 496                         if (extent_item_pos) {
 497                                 ret = check_extent_in_eb(&key, eb, fi,
 498                                                 *extent_item_pos,
 499                                                 &eie, ignore_offset);
 500                                 if (ret < 0)
 501                                         break;
 502                         }
 503                         if (ret > 0)
 504                                 goto next;
 505                         ret = ulist_add_merge_ptr(parents, eb->start,
 506                                                   eie, (void **)&old, GFP_NOFS);
 507                         if (ret < 0)
 508                                 break;
 509                         if (!ret && extent_item_pos) {
 510                                 while (old->next)
 511                                         old = old->next;
 512                                 old->next = eie;
 513                         }
 514                         eie = NULL;
 515                 }
 516 next:
 517                 if (time_seq == SEQ_LAST)
 518                         ret = btrfs_next_item(root, path);
 519                 else
 520                         ret = btrfs_next_old_item(root, path, time_seq);
 521         }
 522
 523         if (ret > 0)
 524                 ret = 0;
 525         else if (ret < 0)
 526                 free_inode_elem_list(eie);
 527         return ret;
 528 }
 529
 530 /*
 531  * resolve an indirect backref in the form (root_id, key, level)
 532  * to a logical address
 533  */
 534 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 535                                 struct btrfs_path *path, u64 time_seq,
 536                                 struct preftrees *preftrees,
 537                                 struct prelim_ref *ref, struct ulist *parents,
 538                                 const u64 *extent_item_pos, bool ignore_offset)
 539 {
 540         struct btrfs_root *root;
 541         struct extent_buffer *eb;
 542         int ret = 0;
 543         int root_level;
 544         int level = ref->level;
 545         struct btrfs_key search_key = ref->key_for_search;
 546
 547         root = btrfs_get_fs_root(fs_info, ref->root_id, false);
 548         if (IS_ERR(root)) {
 549                 ret = PTR_ERR(root);
 550                 goto out_free;
 551         }
 552
 553         if (!path->search_commit_root &&
 554             test_bit(BTRFS_ROOT_DELETING, &root->state)) {
 555                 ret = -ENOENT;
 556                 goto out;
 557         }
 558
 559         if (btrfs_is_testing(fs_info)) {
 560                 ret = -ENOENT;
 561                 goto out;
 562         }
 563
 564         if (path->search_commit_root)
 565                 root_level = btrfs_header_level(root->commit_root);
 566         else if (time_seq == SEQ_LAST)
 567                 root_level = btrfs_header_level(root->node);
 568         else
 569                 root_level = btrfs_old_root_level(root, time_seq);
 570
 571         if (root_level + 1 == level)
 572                 goto out;
 573
 574         /*
 575          * We can often find data backrefs with an offset that is too large
 576          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 577          * subtracting a file's offset with the data offset of its
 578          * corresponding extent data item. This can happen for example in the
 579          * clone ioctl.
 580          *
 581          * So if we detect such case we set the search key's offset to zero to
 582          * make sure we will find the matching file extent item at
 583          * add_all_parents(), otherwise we will miss it because the offset
 584          * taken form the backref is much larger then the offset of the file
 585          * extent item. This can make us scan a very large number of file
 586          * extent items, but at least it will not make us miss any.
 587          *
 588          * This is an ugly workaround for a behaviour that should have never
 589          * existed, but it does and a fix for the clone ioctl would touch a lot
 590          * of places, cause backwards incompatibility and would not fix the
 591          * problem for extents cloned with older kernels.
 592          */
 593         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 594             search_key.offset >= LLONG_MAX)
 595                 search_key.offset = 0;
 596         path->lowest_level = level;
 597         if (time_seq == SEQ_LAST)
 598                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 599         else
 600                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 601
 602         btrfs_debug(fs_info,
 603                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 604                  ref->root_id, level, ref->count, ret,
 605                  ref->key_for_search.objectid, ref->key_for_search.type,
 606                  ref->key_for_search.offset);
 607         if (ret < 0)
 608                 goto out;
 609
 610         eb = path->nodes[level];
 611         while (!eb) {
 612                 if (WARN_ON(!level)) {
 613                         ret = 1;
 614                         goto out;
 615                 }
 616                 level--;
 617                 eb = path->nodes[level];
 618         }
 619
 620         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 621                               time_seq, extent_item_pos, ignore_offset);
 622 out:
 623         btrfs_put_root(root);
 624 out_free:
 625         path->lowest_level = 0;
 626         btrfs_release_path(path);
 627         return ret;
 628 }
 629
 630 static struct extent_inode_elem *
 631 unode_aux_to_inode_list(struct ulist_node *node)
 632 {
 633         if (!node)
 634                 return NULL;
 635         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 636 }
 637
 638 /*
 639  * We maintain three separate rbtrees: one for direct refs, one for
 640  * indirect refs which have a key, and one for indirect refs which do not
 641  * have a key. Each tree does merge on insertion.
 642  *
 643  * Once all of the references are located, we iterate over the tree of
 644  * indirect refs with missing keys. An appropriate key is located and
 645  * the ref is moved onto the tree for indirect refs. After all missing
 646  * keys are thus located, we iterate over the indirect ref tree, resolve
 647  * each reference, and then insert the resolved reference onto the
 648  * direct tree (merging there too).
 649  *
 650  * New backrefs (i.e., for parent nodes) are added to the appropriate
 651  * rbtree as they are encountered. The new backrefs are subsequently
 652  * resolved as above.
 653  */
 654 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 655                                  struct btrfs_path *path, u64 time_seq,
 656                                  struct preftrees *preftrees,
 657                                  const u64 *extent_item_pos,
 658                                  struct share_check *sc, bool ignore_offset)
 659 {
 660         int err;
 661         int ret = 0;
 662         struct ulist *parents;
 663         struct ulist_node *node;
 664         struct ulist_iterator uiter;
 665         struct rb_node *rnode;
 666
 667         parents = ulist_alloc(GFP_NOFS);
 668         if (!parents)
 669                 return -ENOMEM;
 670
 671         /*
 672          * We could trade memory usage for performance here by iterating
 673          * the tree, allocating new refs for each insertion, and then
 674          * freeing the entire indirect tree when we're done.  In some test
 675          * cases, the tree can grow quite large (~200k objects).
 676          */
 677         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 678                 struct prelim_ref *ref;
 679
 680                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 681                 if (WARN(ref->parent,
 682                          "BUG: direct ref found in indirect tree")) {
 683                         ret = -EINVAL;
 684                         goto out;
 685                 }
 686
 687                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 688                 preftrees->indirect.count--;
 689
 690                 if (ref->count == 0) {
 691                         free_pref(ref);
 692                         continue;
 693                 }
 694
 695                 if (sc && sc->root_objectid &&
 696                     ref->root_id != sc->root_objectid) {
 697                         free_pref(ref);
 698                         ret = BACKREF_FOUND_SHARED;
 699                         goto out;
 700                 }
 701                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 702                                            ref, parents, extent_item_pos,
 703                                            ignore_offset);
 704                 /*
 705                  * we can only tolerate ENOENT,otherwise,we should catch error
 706                  * and return directly.
 707                  */
 708                 if (err == -ENOENT) {
 709                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 710                                           NULL);
 711                         continue;
 712                 } else if (err) {
 713                         free_pref(ref);
 714                         ret = err;
 715                         goto out;
 716                 }
 717
 718                 /* we put the first parent into the ref at hand */
 719                 ULIST_ITER_INIT(&uiter);
 720                 node = ulist_next(parents, &uiter);
 721                 ref->parent = node ? node->val : 0;
 722                 ref->inode_list = unode_aux_to_inode_list(node);
 723
 724                 /* Add a prelim_ref(s) for any other parent(s). */
 725                 while ((node = ulist_next(parents, &uiter))) {
 726                         struct prelim_ref *new_ref;
 727
 728                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 729                                                    GFP_NOFS);
 730                         if (!new_ref) {
 731                                 free_pref(ref);
 732                                 ret = -ENOMEM;
 733                                 goto out;
 734                         }
 735                         memcpy(new_ref, ref, sizeof(*ref));
 736                         new_ref->parent = node->val;
 737                         new_ref->inode_list = unode_aux_to_inode_list(node);
 738                         prelim_ref_insert(fs_info, &preftrees->direct,
 739                                           new_ref, NULL);
 740                 }
 741
 742                 /*
 743                  * Now it's a direct ref, put it in the direct tree. We must
 744                  * do this last because the ref could be merged/freed here.
 745                  */
 746                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 747
 748                 ulist_reinit(parents);
 749                 cond_resched();
 750         }
 751 out:
 752         ulist_free(parents);
 753         return ret;
 754 }
 755
 756 /*
 757  * read tree blocks and add keys where required.
 758  */
 759 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 760                             struct preftrees *preftrees, bool lock)
 761 {
 762         struct prelim_ref *ref;
 763         struct extent_buffer *eb;
 764         struct preftree *tree = &preftrees->indirect_missing_keys;
 765         struct rb_node *node;
 766
 767         while ((node = rb_first_cached(&tree->root))) {
 768                 ref = rb_entry(node, struct prelim_ref, rbnode);
 769                 rb_erase_cached(node, &tree->root);
 770
 771                 BUG_ON(ref->parent);    /* should not be a direct ref */
 772                 BUG_ON(ref->key_for_search.type);
 773                 BUG_ON(!ref->wanted_disk_byte);
 774
 775                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 776                                      ref->level - 1, NULL);
 777                 if (IS_ERR(eb)) {
 778                         free_pref(ref);
 779                         return PTR_ERR(eb);
 780                 } else if (!extent_buffer_uptodate(eb)) {
 781                         free_pref(ref);
 782                         free_extent_buffer(eb);
 783                         return -EIO;
 784                 }
 785                 if (lock)
 786                         btrfs_tree_read_lock(eb);
 787                 if (btrfs_header_level(eb) == 0)
 788                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 789                 else
 790                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 791                 if (lock)
 792                         btrfs_tree_read_unlock(eb);
 793                 free_extent_buffer(eb);
 794                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 795                 cond_resched();
 796         }
 797         return 0;
 798 }
 799
 800 /*
 801  * add all currently queued delayed refs from this head whose seq nr is
 802  * smaller or equal that seq to the list
 803  */
 804 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 805                             struct btrfs_delayed_ref_head *head, u64 seq,
 806                             struct preftrees *preftrees, struct share_check *sc)
 807 {
 808         struct btrfs_delayed_ref_node *node;
 809         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 810         struct btrfs_key key;
 811         struct btrfs_key tmp_op_key;
 812         struct rb_node *n;
 813         int count;
 814         int ret = 0;
 815
 816         if (extent_op && extent_op->update_key)
 817                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 818
 819         spin_lock(&head->lock);
 820         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 821                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 822                                 ref_node);
 823                 if (node->seq > seq)
 824                         continue;
 825
 826                 switch (node->action) {
 827                 case BTRFS_ADD_DELAYED_EXTENT:
 828                 case BTRFS_UPDATE_DELAYED_HEAD:
 829                         WARN_ON(1);
 830                         continue;
 831                 case BTRFS_ADD_DELAYED_REF:
 832                         count = node->ref_mod;
 833                         break;
 834                 case BTRFS_DROP_DELAYED_REF:
 835                         count = node->ref_mod * -1;
 836                         break;
 837                 default:
 838                         BUG();
 839                 }
 840                 switch (node->type) {
 841                 case BTRFS_TREE_BLOCK_REF_KEY: {
 842                         /* NORMAL INDIRECT METADATA backref */
 843                         struct btrfs_delayed_tree_ref *ref;
 844
 845                         ref = btrfs_delayed_node_to_tree_ref(node);
 846                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 847                                                &tmp_op_key, ref->level + 1,
 848                                                node->bytenr, count, sc,
 849                                                GFP_ATOMIC);
 850                         break;
 851                 }
 852                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 853                         /* SHARED DIRECT METADATA backref */
 854                         struct btrfs_delayed_tree_ref *ref;
 855
 856                         ref = btrfs_delayed_node_to_tree_ref(node);
 857
 858                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 859                                              ref->parent, node->bytenr, count,
 860                                              sc, GFP_ATOMIC);
 861                         break;
 862                 }
 863                 case BTRFS_EXTENT_DATA_REF_KEY: {
 864                         /* NORMAL INDIRECT DATA backref */
 865                         struct btrfs_delayed_data_ref *ref;
 866                         ref = btrfs_delayed_node_to_data_ref(node);
 867
 868                         key.objectid = ref->objectid;
 869                         key.type = BTRFS_EXTENT_DATA_KEY;
 870                         key.offset = ref->offset;
 871
 872                         /*
 873                          * Found a inum that doesn't match our known inum, we
 874                          * know it's shared.
 875                          */
 876                         if (sc && sc->inum && ref->objectid != sc->inum) {
 877                                 ret = BACKREF_FOUND_SHARED;
 878                                 goto out;
 879                         }
 880
 881                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 882                                                &key, 0, node->bytenr, count, sc,
 883                                                GFP_ATOMIC);
 884                         break;
 885                 }
 886                 case BTRFS_SHARED_DATA_REF_KEY: {
 887                         /* SHARED DIRECT FULL backref */
 888                         struct btrfs_delayed_data_ref *ref;
 889
 890                         ref = btrfs_delayed_node_to_data_ref(node);
 891
 892                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 893                                              node->bytenr, count, sc,
 894                                              GFP_ATOMIC);
 895                         break;
 896                 }
 897                 default:
 898                         WARN_ON(1);
 899                 }
 900                 /*
 901                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 902                  * refs have been checked.
 903                  */
 904                 if (ret && (ret != BACKREF_FOUND_SHARED))
 905                         break;
 906         }
 907         if (!ret)
 908                 ret = extent_is_shared(sc);
 909 out:
 910         spin_unlock(&head->lock);
 911         return ret;
 912 }
 913
 914 /*
 915  * add all inline backrefs for bytenr to the list
 916  *
 917  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 918  */
 919 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 920                            struct btrfs_path *path, u64 bytenr,
 921                            int *info_level, struct preftrees *preftrees,
 922                            struct share_check *sc)
 923 {
 924         int ret = 0;
 925         int slot;
 926         struct extent_buffer *leaf;
 927         struct btrfs_key key;
 928         struct btrfs_key found_key;
 929         unsigned long ptr;
 930         unsigned long end;
 931         struct btrfs_extent_item *ei;
 932         u64 flags;
 933         u64 item_size;
 934
 935         /*
 936          * enumerate all inline refs
 937          */
 938         leaf = path->nodes[0];
 939         slot = path->slots[0];
 940
 941         item_size = btrfs_item_size_nr(leaf, slot);
 942         BUG_ON(item_size < sizeof(*ei));
 943
 944         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 945         flags = btrfs_extent_flags(leaf, ei);
 946         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 947
 948         ptr = (unsigned long)(ei + 1);
 949         end = (unsigned long)ei + item_size;
 950
 951         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 952             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 953                 struct btrfs_tree_block_info *info;
 954
 955                 info = (struct btrfs_tree_block_info *)ptr;
 956                 *info_level = btrfs_tree_block_level(leaf, info);
 957                 ptr += sizeof(struct btrfs_tree_block_info);
 958                 BUG_ON(ptr > end);
 959         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 960                 *info_level = found_key.offset;
 961         } else {
 962                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 963         }
 964
 965         while (ptr < end) {
 966                 struct btrfs_extent_inline_ref *iref;
 967                 u64 offset;
 968                 int type;
 969
 970                 iref = (struct btrfs_extent_inline_ref *)ptr;
 971                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 972                                                         BTRFS_REF_TYPE_ANY);
 973                 if (type == BTRFS_REF_TYPE_INVALID)
 974                         return -EUCLEAN;
 975
 976                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 977
 978                 switch (type) {
 979                 case BTRFS_SHARED_BLOCK_REF_KEY:
 980                         ret = add_direct_ref(fs_info, preftrees,
 981                                              *info_level + 1, offset,
 982                                              bytenr, 1, NULL, GFP_NOFS);
 983                         break;
 984                 case BTRFS_SHARED_DATA_REF_KEY: {
 985                         struct btrfs_shared_data_ref *sdref;
 986                         int count;
 987
 988                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 989                         count = btrfs_shared_data_ref_count(leaf, sdref);
 990
 991                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 992                                              bytenr, count, sc, GFP_NOFS);
 993                         break;
 994                 }
 995                 case BTRFS_TREE_BLOCK_REF_KEY:
 996                         ret = add_indirect_ref(fs_info, preftrees, offset,
 997                                                NULL, *info_level + 1,
 998                                                bytenr, 1, NULL, GFP_NOFS);
 999                         break;
1000                 case BTRFS_EXTENT_DATA_REF_KEY: {
1001                         struct btrfs_extent_data_ref *dref;
1002                         int count;
1003                         u64 root;
1004
1005                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1006                         count = btrfs_extent_data_ref_count(leaf, dref);
1007                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1008                                                                       dref);
1009                         key.type = BTRFS_EXTENT_DATA_KEY;
1010                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1011
1012                         if (sc && sc->inum && key.objectid != sc->inum) {
1013                                 ret = BACKREF_FOUND_SHARED;
1014                                 break;
1015                         }
1016
1017                         root = btrfs_extent_data_ref_root(leaf, dref);
1018
1019                         ret = add_indirect_ref(fs_info, preftrees, root,
1020                                                &key, 0, bytenr, count,
1021                                                sc, GFP_NOFS);
1022                         break;
1023                 }
1024                 default:
1025                         WARN_ON(1);
1026                 }
1027                 if (ret)
1028                         return ret;
1029                 ptr += btrfs_extent_inline_ref_size(type);
1030         }
1031
1032         return 0;
1033 }
1034
1035 /*
1036  * add all non-inline backrefs for bytenr to the list
1037  *
1038  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1039  */
1040 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1041                           struct btrfs_path *path, u64 bytenr,
1042                           int info_level, struct preftrees *preftrees,
1043                           struct share_check *sc)
1044 {
1045         struct btrfs_root *extent_root = fs_info->extent_root;
1046         int ret;
1047         int slot;
1048         struct extent_buffer *leaf;
1049         struct btrfs_key key;
1050
1051         while (1) {
1052                 ret = btrfs_next_item(extent_root, path);
1053                 if (ret < 0)
1054                         break;
1055                 if (ret) {
1056                         ret = 0;
1057                         break;
1058                 }
1059
1060                 slot = path->slots[0];
1061                 leaf = path->nodes[0];
1062                 btrfs_item_key_to_cpu(leaf, &key, slot);
1063
1064                 if (key.objectid != bytenr)
1065                         break;
1066                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1067                         continue;
1068                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1069                         break;
1070
1071                 switch (key.type) {
1072                 case BTRFS_SHARED_BLOCK_REF_KEY:
1073                         /* SHARED DIRECT METADATA backref */
1074                         ret = add_direct_ref(fs_info, preftrees,
1075                                              info_level + 1, key.offset,
1076                                              bytenr, 1, NULL, GFP_NOFS);
1077                         break;
1078                 case BTRFS_SHARED_DATA_REF_KEY: {
1079                         /* SHARED DIRECT FULL backref */
1080                         struct btrfs_shared_data_ref *sdref;
1081                         int count;
1082
1083                         sdref = btrfs_item_ptr(leaf, slot,
1084                                               struct btrfs_shared_data_ref);
1085                         count = btrfs_shared_data_ref_count(leaf, sdref);
1086                         ret = add_direct_ref(fs_info, preftrees, 0,
1087                                              key.offset, bytenr, count,
1088                                              sc, GFP_NOFS);
1089                         break;
1090                 }
1091                 case BTRFS_TREE_BLOCK_REF_KEY:
1092                         /* NORMAL INDIRECT METADATA backref */
1093                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1094                                                NULL, info_level + 1, bytenr,
1095                                                1, NULL, GFP_NOFS);
1096                         break;
1097                 case BTRFS_EXTENT_DATA_REF_KEY: {
1098                         /* NORMAL INDIRECT DATA backref */
1099                         struct btrfs_extent_data_ref *dref;
1100                         int count;
1101                         u64 root;
1102
1103                         dref = btrfs_item_ptr(leaf, slot,
1104                                               struct btrfs_extent_data_ref);
1105                         count = btrfs_extent_data_ref_count(leaf, dref);
1106                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1107                                                                       dref);
1108                         key.type = BTRFS_EXTENT_DATA_KEY;
1109                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1110
1111                         if (sc && sc->inum && key.objectid != sc->inum) {
1112                                 ret = BACKREF_FOUND_SHARED;
1113                                 break;
1114                         }
1115
1116                         root = btrfs_extent_data_ref_root(leaf, dref);
1117                         ret = add_indirect_ref(fs_info, preftrees, root,
1118                                                &key, 0, bytenr, count,
1119                                                sc, GFP_NOFS);
1120                         break;
1121                 }
1122                 default:
1123                         WARN_ON(1);
1124                 }
1125                 if (ret)
1126                         return ret;
1127
1128         }
1129
1130         return ret;
1131 }
1132
1133 /*
1134  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1135  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1136  * indirect refs to their parent bytenr.
1137  * When roots are found, they're added to the roots list
1138  *
1139  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1140  * much like trans == NULL case, the difference only lies in it will not
1141  * commit root.
1142  * The special case is for qgroup to search roots in commit_transaction().
1143  *
1144  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1145  * shared extent is detected.
1146  *
1147  * Otherwise this returns 0 for success and <0 for an error.
1148  *
1149  * If ignore_offset is set to false, only extent refs whose offsets match
1150  * extent_item_pos are returned.  If true, every extent ref is returned
1151  * and extent_item_pos is ignored.
1152  *
1153  * FIXME some caching might speed things up
1154  */
1155 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1156                              struct btrfs_fs_info *fs_info, u64 bytenr,
1157                              u64 time_seq, struct ulist *refs,
1158                              struct ulist *roots, const u64 *extent_item_pos,
1159                              struct share_check *sc, bool ignore_offset)
1160 {
1161         struct btrfs_key key;
1162         struct btrfs_path *path;
1163         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1164         struct btrfs_delayed_ref_head *head;
1165         int info_level = 0;
1166         int ret;
1167         struct prelim_ref *ref;
1168         struct rb_node *node;
1169         struct extent_inode_elem *eie = NULL;
1170         struct preftrees preftrees = {
1171                 .direct = PREFTREE_INIT,
1172                 .indirect = PREFTREE_INIT,
1173                 .indirect_missing_keys = PREFTREE_INIT
1174         };
1175
1176         key.objectid = bytenr;
1177         key.offset = (u64)-1;
1178         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1179                 key.type = BTRFS_METADATA_ITEM_KEY;
1180         else
1181                 key.type = BTRFS_EXTENT_ITEM_KEY;
1182
1183         path = btrfs_alloc_path();
1184         if (!path)
1185                 return -ENOMEM;
1186         if (!trans) {
1187                 path->search_commit_root = 1;
1188                 path->skip_locking = 1;
1189         }
1190
1191         if (time_seq == SEQ_LAST)
1192                 path->skip_locking = 1;
1193
1194         /*
1195          * grab both a lock on the path and a lock on the delayed ref head.
1196          * We need both to get a consistent picture of how the refs look
1197          * at a specified point in time
1198          */
1199 again:
1200         head = NULL;
1201
1202         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1203         if (ret < 0)
1204                 goto out;
1205         BUG_ON(ret == 0);
1206
1207 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1208         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1209             time_seq != SEQ_LAST) {
1210 #else
1211         if (trans && time_seq != SEQ_LAST) {
1212 #endif
1213                 /*
1214                  * look if there are updates for this ref queued and lock the
1215                  * head
1216                  */
1217                 delayed_refs = &trans->transaction->delayed_refs;
1218                 spin_lock(&delayed_refs->lock);
1219                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1220                 if (head) {
1221                         if (!mutex_trylock(&head->mutex)) {
1222                                 refcount_inc(&head->refs);
1223                                 spin_unlock(&delayed_refs->lock);
1224
1225                                 btrfs_release_path(path);
1226
1227                                 /*
1228                                  * Mutex was contended, block until it's
1229                                  * released and try again
1230                                  */
1231                                 mutex_lock(&head->mutex);
1232                                 mutex_unlock(&head->mutex);
1233                                 btrfs_put_delayed_ref_head(head);
1234                                 goto again;
1235                         }
1236                         spin_unlock(&delayed_refs->lock);
1237                         ret = add_delayed_refs(fs_info, head, time_seq,
1238                                                &preftrees, sc);
1239                         mutex_unlock(&head->mutex);
1240                         if (ret)
1241                                 goto out;
1242                 } else {
1243                         spin_unlock(&delayed_refs->lock);
1244                 }
1245         }
1246
1247         if (path->slots[0]) {
1248                 struct extent_buffer *leaf;
1249                 int slot;
1250
1251                 path->slots[0]--;
1252                 leaf = path->nodes[0];
1253                 slot = path->slots[0];
1254                 btrfs_item_key_to_cpu(leaf, &key, slot);
1255                 if (key.objectid == bytenr &&
1256                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1257                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1258                         ret = add_inline_refs(fs_info, path, bytenr,
1259                                               &info_level, &preftrees, sc);
1260                         if (ret)
1261                                 goto out;
1262                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1263                                              &preftrees, sc);
1264                         if (ret)
1265                                 goto out;
1266                 }
1267         }
1268
1269         btrfs_release_path(path);
1270
1271         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1272         if (ret)
1273                 goto out;
1274
1275         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1276
1277         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1278                                     extent_item_pos, sc, ignore_offset);
1279         if (ret)
1280                 goto out;
1281
1282         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1283
1284         /*
1285          * This walks the tree of merged and resolved refs. Tree blocks are
1286          * read in as needed. Unique entries are added to the ulist, and
1287          * the list of found roots is updated.
1288          *
1289          * We release the entire tree in one go before returning.
1290          */
1291         node = rb_first_cached(&preftrees.direct.root);
1292         while (node) {
1293                 ref = rb_entry(node, struct prelim_ref, rbnode);
1294                 node = rb_next(&ref->rbnode);
1295                 /*
1296                  * ref->count < 0 can happen here if there are delayed
1297                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1298                  * prelim_ref_insert() relies on this when merging
1299                  * identical refs to keep the overall count correct.
1300                  * prelim_ref_insert() will merge only those refs
1301                  * which compare identically.  Any refs having
1302                  * e.g. different offsets would not be merged,
1303                  * and would retain their original ref->count < 0.
1304                  */
1305                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1306                         if (sc && sc->root_objectid &&
1307                             ref->root_id != sc->root_objectid) {
1308                                 ret = BACKREF_FOUND_SHARED;
1309                                 goto out;
1310                         }
1311
1312                         /* no parent == root of tree */
1313                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1314                         if (ret < 0)
1315                                 goto out;
1316                 }
1317                 if (ref->count && ref->parent) {
1318                         if (extent_item_pos && !ref->inode_list &&
1319                             ref->level == 0) {
1320                                 struct extent_buffer *eb;
1321
1322                                 eb = read_tree_block(fs_info, ref->parent, 0,
1323                                                      ref->level, NULL);
1324                                 if (IS_ERR(eb)) {
1325                                         ret = PTR_ERR(eb);
1326                                         goto out;
1327                                 } else if (!extent_buffer_uptodate(eb)) {
1328                                         free_extent_buffer(eb);
1329                                         ret = -EIO;
1330                                         goto out;
1331                                 }
1332
1333                                 if (!path->skip_locking) {
1334                                         btrfs_tree_read_lock(eb);
1335                                         btrfs_set_lock_blocking_read(eb);
1336                                 }
1337                                 ret = find_extent_in_eb(eb, bytenr,
1338                                                         *extent_item_pos, &eie, ignore_offset);
1339                                 if (!path->skip_locking)
1340                                         btrfs_tree_read_unlock_blocking(eb);
1341                                 free_extent_buffer(eb);
1342                                 if (ret < 0)
1343                                         goto out;
1344                                 ref->inode_list = eie;
1345                         }
1346                         ret = ulist_add_merge_ptr(refs, ref->parent,
1347                                                   ref->inode_list,
1348                                                   (void **)&eie, GFP_NOFS);
1349                         if (ret < 0)
1350                                 goto out;
1351                         if (!ret && extent_item_pos) {
1352                                 /*
1353                                  * we've recorded that parent, so we must extend
1354                                  * its inode list here
1355                                  */
1356                                 BUG_ON(!eie);
1357                                 while (eie->next)
1358                                         eie = eie->next;
1359                                 eie->next = ref->inode_list;
1360                         }
1361                         eie = NULL;
1362                 }
1363                 cond_resched();
1364         }
1365
1366 out:
1367         btrfs_free_path(path);
1368
1369         prelim_release(&preftrees.direct);
1370         prelim_release(&preftrees.indirect);
1371         prelim_release(&preftrees.indirect_missing_keys);
1372
1373         if (ret < 0)
1374                 free_inode_elem_list(eie);
1375         return ret;
1376 }
1377
1378 static void free_leaf_list(struct ulist *blocks)
1379 {
1380         struct ulist_node *node = NULL;
1381         struct extent_inode_elem *eie;
1382         struct ulist_iterator uiter;
1383
1384         ULIST_ITER_INIT(&uiter);
1385         while ((node = ulist_next(blocks, &uiter))) {
1386                 if (!node->aux)
1387                         continue;
1388                 eie = unode_aux_to_inode_list(node);
1389                 free_inode_elem_list(eie);
1390                 node->aux = 0;
1391         }
1392
1393         ulist_free(blocks);
1394 }
1395
1396 /*
1397  * Finds all leafs with a reference to the specified combination of bytenr and
1398  * offset. key_list_head will point to a list of corresponding keys (caller must
1399  * free each list element). The leafs will be stored in the leafs ulist, which
1400  * must be freed with ulist_free.
1401  *
1402  * returns 0 on success, <0 on error
1403  */
1404 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1405                          struct btrfs_fs_info *fs_info, u64 bytenr,
1406                          u64 time_seq, struct ulist **leafs,
1407                          const u64 *extent_item_pos, bool ignore_offset)
1408 {
1409         int ret;
1410
1411         *leafs = ulist_alloc(GFP_NOFS);
1412         if (!*leafs)
1413                 return -ENOMEM;
1414
1415         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1416                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1417         if (ret < 0 && ret != -ENOENT) {
1418                 free_leaf_list(*leafs);
1419                 return ret;
1420         }
1421
1422         return 0;
1423 }
1424
1425 /*
1426  * walk all backrefs for a given extent to find all roots that reference this
1427  * extent. Walking a backref means finding all extents that reference this
1428  * extent and in turn walk the backrefs of those, too. Naturally this is a
1429  * recursive process, but here it is implemented in an iterative fashion: We
1430  * find all referencing extents for the extent in question and put them on a
1431  * list. In turn, we find all referencing extents for those, further appending
1432  * to the list. The way we iterate the list allows adding more elements after
1433  * the current while iterating. The process stops when we reach the end of the
1434  * list. Found roots are added to the roots list.
1435  *
1436  * returns 0 on success, < 0 on error.
1437  */
1438 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1439                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1440                                      u64 time_seq, struct ulist **roots,
1441                                      bool ignore_offset)
1442 {
1443         struct ulist *tmp;
1444         struct ulist_node *node = NULL;
1445         struct ulist_iterator uiter;
1446         int ret;
1447
1448         tmp = ulist_alloc(GFP_NOFS);
1449         if (!tmp)
1450                 return -ENOMEM;
1451         *roots = ulist_alloc(GFP_NOFS);
1452         if (!*roots) {
1453                 ulist_free(tmp);
1454                 return -ENOMEM;
1455         }
1456
1457         ULIST_ITER_INIT(&uiter);
1458         while (1) {
1459                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1460                                         tmp, *roots, NULL, NULL, ignore_offset);
1461                 if (ret < 0 && ret != -ENOENT) {
1462                         ulist_free(tmp);
1463                         ulist_free(*roots);
1464                         return ret;
1465                 }
1466                 node = ulist_next(tmp, &uiter);
1467                 if (!node)
1468                         break;
1469                 bytenr = node->val;
1470                 cond_resched();
1471         }
1472
1473         ulist_free(tmp);
1474         return 0;
1475 }
1476
1477 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1478                          struct btrfs_fs_info *fs_info, u64 bytenr,
1479                          u64 time_seq, struct ulist **roots,
1480                          bool ignore_offset)
1481 {
1482         int ret;
1483
1484         if (!trans)
1485                 down_read(&fs_info->commit_root_sem);
1486         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1487                                         time_seq, roots, ignore_offset);
1488         if (!trans)
1489                 up_read(&fs_info->commit_root_sem);
1490         return ret;
1491 }
1492
1493 /**
1494  * btrfs_check_shared - tell us whether an extent is shared
1495  *
1496  * btrfs_check_shared uses the backref walking code but will short
1497  * circuit as soon as it finds a root or inode that doesn't match the
1498  * one passed in. This provides a significant performance benefit for
1499  * callers (such as fiemap) which want to know whether the extent is
1500  * shared but do not need a ref count.
1501  *
1502  * This attempts to attach to the running transaction in order to account for
1503  * delayed refs, but continues on even when no running transaction exists.
1504  *
1505  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1506  */
1507 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1508                 struct ulist *roots, struct ulist *tmp)
1509 {
1510         struct btrfs_fs_info *fs_info = root->fs_info;
1511         struct btrfs_trans_handle *trans;
1512         struct ulist_iterator uiter;
1513         struct ulist_node *node;
1514         struct seq_list elem = SEQ_LIST_INIT(elem);
1515         int ret = 0;
1516         struct share_check shared = {
1517                 .root_objectid = root->root_key.objectid,
1518                 .inum = inum,
1519                 .share_count = 0,
1520         };
1521
1522         ulist_init(roots);
1523         ulist_init(tmp);
1524
1525         trans = btrfs_join_transaction_nostart(root);
1526         if (IS_ERR(trans)) {
1527                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1528                         ret = PTR_ERR(trans);
1529                         goto out;
1530                 }
1531                 trans = NULL;
1532                 down_read(&fs_info->commit_root_sem);
1533         } else {
1534                 btrfs_get_tree_mod_seq(fs_info, &elem);
1535         }
1536
1537         ULIST_ITER_INIT(&uiter);
1538         while (1) {
1539                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1540                                         roots, NULL, &shared, false);
1541                 if (ret == BACKREF_FOUND_SHARED) {
1542                         /* this is the only condition under which we return 1 */
1543                         ret = 1;
1544                         break;
1545                 }
1546                 if (ret < 0 && ret != -ENOENT)
1547                         break;
1548                 ret = 0;
1549                 node = ulist_next(tmp, &uiter);
1550                 if (!node)
1551                         break;
1552                 bytenr = node->val;
1553                 shared.share_count = 0;
1554                 cond_resched();
1555         }
1556
1557         if (trans) {
1558                 btrfs_put_tree_mod_seq(fs_info, &elem);
1559                 btrfs_end_transaction(trans);
1560         } else {
1561                 up_read(&fs_info->commit_root_sem);
1562         }
1563 out:
1564         ulist_release(roots);
1565         ulist_release(tmp);
1566         return ret;
1567 }
1568
1569 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1570                           u64 start_off, struct btrfs_path *path,
1571                           struct btrfs_inode_extref **ret_extref,
1572                           u64 *found_off)
1573 {
1574         int ret, slot;
1575         struct btrfs_key key;
1576         struct btrfs_key found_key;
1577         struct btrfs_inode_extref *extref;
1578         const struct extent_buffer *leaf;
1579         unsigned long ptr;
1580
1581         key.objectid = inode_objectid;
1582         key.type = BTRFS_INODE_EXTREF_KEY;
1583         key.offset = start_off;
1584
1585         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1586         if (ret < 0)
1587                 return ret;
1588
1589         while (1) {
1590                 leaf = path->nodes[0];
1591                 slot = path->slots[0];
1592                 if (slot >= btrfs_header_nritems(leaf)) {
1593                         /*
1594                          * If the item at offset is not found,
1595                          * btrfs_search_slot will point us to the slot
1596                          * where it should be inserted. In our case
1597                          * that will be the slot directly before the
1598                          * next INODE_REF_KEY_V2 item. In the case
1599                          * that we're pointing to the last slot in a
1600                          * leaf, we must move one leaf over.
1601                          */
1602                         ret = btrfs_next_leaf(root, path);
1603                         if (ret) {
1604                                 if (ret >= 1)
1605                                         ret = -ENOENT;
1606                                 break;
1607                         }
1608                         continue;
1609                 }
1610
1611                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1612
1613                 /*
1614                  * Check that we're still looking at an extended ref key for
1615                  * this particular objectid. If we have different
1616                  * objectid or type then there are no more to be found
1617                  * in the tree and we can exit.
1618                  */
1619                 ret = -ENOENT;
1620                 if (found_key.objectid != inode_objectid)
1621                         break;
1622                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1623                         break;
1624
1625                 ret = 0;
1626                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1627                 extref = (struct btrfs_inode_extref *)ptr;
1628                 *ret_extref = extref;
1629                 if (found_off)
1630                         *found_off = found_key.offset;
1631                 break;
1632         }
1633
1634         return ret;
1635 }
1636
1637 /*
1638  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1639  * Elements of the path are separated by '/' and the path is guaranteed to be
1640  * 0-terminated. the path is only given within the current file system.
1641  * Therefore, it never starts with a '/'. the caller is responsible to provide
1642  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1643  * the start point of the resulting string is returned. this pointer is within
1644  * dest, normally.
1645  * in case the path buffer would overflow, the pointer is decremented further
1646  * as if output was written to the buffer, though no more output is actually
1647  * generated. that way, the caller can determine how much space would be
1648  * required for the path to fit into the buffer. in that case, the returned
1649  * value will be smaller than dest. callers must check this!
1650  */
1651 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1652                         u32 name_len, unsigned long name_off,
1653                         struct extent_buffer *eb_in, u64 parent,
1654                         char *dest, u32 size)
1655 {
1656         int slot;
1657         u64 next_inum;
1658         int ret;
1659         s64 bytes_left = ((s64)size) - 1;
1660         struct extent_buffer *eb = eb_in;
1661         struct btrfs_key found_key;
1662         int leave_spinning = path->leave_spinning;
1663         struct btrfs_inode_ref *iref;
1664
1665         if (bytes_left >= 0)
1666                 dest[bytes_left] = '\0';
1667
1668         path->leave_spinning = 1;
1669         while (1) {
1670                 bytes_left -= name_len;
1671                 if (bytes_left >= 0)
1672                         read_extent_buffer(eb, dest + bytes_left,
1673                                            name_off, name_len);
1674                 if (eb != eb_in) {
1675                         if (!path->skip_locking)
1676                                 btrfs_tree_read_unlock_blocking(eb);
1677                         free_extent_buffer(eb);
1678                 }
1679                 ret = btrfs_find_item(fs_root, path, parent, 0,
1680                                 BTRFS_INODE_REF_KEY, &found_key);
1681                 if (ret > 0)
1682                         ret = -ENOENT;
1683                 if (ret)
1684                         break;
1685
1686                 next_inum = found_key.offset;
1687
1688                 /* regular exit ahead */
1689                 if (parent == next_inum)
1690                         break;
1691
1692                 slot = path->slots[0];
1693                 eb = path->nodes[0];
1694                 /* make sure we can use eb after releasing the path */
1695                 if (eb != eb_in) {
1696                         if (!path->skip_locking)
1697                                 btrfs_set_lock_blocking_read(eb);
1698                         path->nodes[0] = NULL;
1699                         path->locks[0] = 0;
1700                 }
1701                 btrfs_release_path(path);
1702                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1703
1704                 name_len = btrfs_inode_ref_name_len(eb, iref);
1705                 name_off = (unsigned long)(iref + 1);
1706
1707                 parent = next_inum;
1708                 --bytes_left;
1709                 if (bytes_left >= 0)
1710                         dest[bytes_left] = '/';
1711         }
1712
1713         btrfs_release_path(path);
1714         path->leave_spinning = leave_spinning;
1715
1716         if (ret)
1717                 return ERR_PTR(ret);
1718
1719         return dest + bytes_left;
1720 }
1721
1722 /*
1723  * this makes the path point to (logical EXTENT_ITEM *)
1724  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1725  * tree blocks and <0 on error.
1726  */
1727 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1728                         struct btrfs_path *path, struct btrfs_key *found_key,
1729                         u64 *flags_ret)
1730 {
1731         int ret;
1732         u64 flags;
1733         u64 size = 0;
1734         u32 item_size;
1735         const struct extent_buffer *eb;
1736         struct btrfs_extent_item *ei;
1737         struct btrfs_key key;
1738
1739         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1740                 key.type = BTRFS_METADATA_ITEM_KEY;
1741         else
1742                 key.type = BTRFS_EXTENT_ITEM_KEY;
1743         key.objectid = logical;
1744         key.offset = (u64)-1;
1745
1746         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1747         if (ret < 0)
1748                 return ret;
1749
1750         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1751         if (ret) {
1752                 if (ret > 0)
1753                         ret = -ENOENT;
1754                 return ret;
1755         }
1756         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1757         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1758                 size = fs_info->nodesize;
1759         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1760                 size = found_key->offset;
1761
1762         if (found_key->objectid > logical ||
1763             found_key->objectid + size <= logical) {
1764                 btrfs_debug(fs_info,
1765                         "logical %llu is not within any extent", logical);
1766                 return -ENOENT;
1767         }
1768
1769         eb = path->nodes[0];
1770         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1771         BUG_ON(item_size < sizeof(*ei));
1772
1773         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1774         flags = btrfs_extent_flags(eb, ei);
1775
1776         btrfs_debug(fs_info,
1777                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1778                  logical, logical - found_key->objectid, found_key->objectid,
1779                  found_key->offset, flags, item_size);
1780
1781         WARN_ON(!flags_ret);
1782         if (flags_ret) {
1783                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1784                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1785                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1786                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1787                 else
1788                         BUG();
1789                 return 0;
1790         }
1791
1792         return -EIO;
1793 }
1794
1795 /*
1796  * helper function to iterate extent inline refs. ptr must point to a 0 value
1797  * for the first call and may be modified. it is used to track state.
1798  * if more refs exist, 0 is returned and the next call to
1799  * get_extent_inline_ref must pass the modified ptr parameter to get the
1800  * next ref. after the last ref was processed, 1 is returned.
1801  * returns <0 on error
1802  */
1803 static int get_extent_inline_ref(unsigned long *ptr,
1804                                  const struct extent_buffer *eb,
1805                                  const struct btrfs_key *key,
1806                                  const struct btrfs_extent_item *ei,
1807                                  u32 item_size,
1808                                  struct btrfs_extent_inline_ref **out_eiref,
1809                                  int *out_type)
1810 {
1811         unsigned long end;
1812         u64 flags;
1813         struct btrfs_tree_block_info *info;
1814
1815         if (!*ptr) {
1816                 /* first call */
1817                 flags = btrfs_extent_flags(eb, ei);
1818                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1819                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1820                                 /* a skinny metadata extent */
1821                                 *out_eiref =
1822                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1823                         } else {
1824                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1825                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1826                                 *out_eiref =
1827                                    (struct btrfs_extent_inline_ref *)(info + 1);
1828                         }
1829                 } else {
1830                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1831                 }
1832                 *ptr = (unsigned long)*out_eiref;
1833                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1834                         return -ENOENT;
1835         }
1836
1837         end = (unsigned long)ei + item_size;
1838         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1839         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1840                                                      BTRFS_REF_TYPE_ANY);
1841         if (*out_type == BTRFS_REF_TYPE_INVALID)
1842                 return -EUCLEAN;
1843
1844         *ptr += btrfs_extent_inline_ref_size(*out_type);
1845         WARN_ON(*ptr > end);
1846         if (*ptr == end)
1847                 return 1; /* last */
1848
1849         return 0;
1850 }
1851
1852 /*
1853  * reads the tree block backref for an extent. tree level and root are returned
1854  * through out_level and out_root. ptr must point to a 0 value for the first
1855  * call and may be modified (see get_extent_inline_ref comment).
1856  * returns 0 if data was provided, 1 if there was no more data to provide or
1857  * <0 on error.
1858  */
1859 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1860                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1861                             u32 item_size, u64 *out_root, u8 *out_level)
1862 {
1863         int ret;
1864         int type;
1865         struct btrfs_extent_inline_ref *eiref;
1866
1867         if (*ptr == (unsigned long)-1)
1868                 return 1;
1869
1870         while (1) {
1871                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1872                                               &eiref, &type);
1873                 if (ret < 0)
1874                         return ret;
1875
1876                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1877                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1878                         break;
1879
1880                 if (ret == 1)
1881                         return 1;
1882         }
1883
1884         /* we can treat both ref types equally here */
1885         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1886
1887         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1888                 struct btrfs_tree_block_info *info;
1889
1890                 info = (struct btrfs_tree_block_info *)(ei + 1);
1891                 *out_level = btrfs_tree_block_level(eb, info);
1892         } else {
1893                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1894                 *out_level = (u8)key->offset;
1895         }
1896
1897         if (ret == 1)
1898                 *ptr = (unsigned long)-1;
1899
1900         return 0;
1901 }
1902
1903 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1904                              struct extent_inode_elem *inode_list,
1905                              u64 root, u64 extent_item_objectid,
1906                              iterate_extent_inodes_t *iterate, void *ctx)
1907 {
1908         struct extent_inode_elem *eie;
1909         int ret = 0;
1910
1911         for (eie = inode_list; eie; eie = eie->next) {
1912                 btrfs_debug(fs_info,
1913                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1914                             extent_item_objectid, eie->inum,
1915                             eie->offset, root);
1916                 ret = iterate(eie->inum, eie->offset, root, ctx);
1917                 if (ret) {
1918                         btrfs_debug(fs_info,
1919                                     "stopping iteration for %llu due to ret=%d",
1920                                     extent_item_objectid, ret);
1921                         break;
1922                 }
1923         }
1924
1925         return ret;
1926 }
1927
1928 /*
1929  * calls iterate() for every inode that references the extent identified by
1930  * the given parameters.
1931  * when the iterator function returns a non-zero value, iteration stops.
1932  */
1933 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1934                                 u64 extent_item_objectid, u64 extent_item_pos,
1935                                 int search_commit_root,
1936                                 iterate_extent_inodes_t *iterate, void *ctx,
1937                                 bool ignore_offset)
1938 {
1939         int ret;
1940         struct btrfs_trans_handle *trans = NULL;
1941         struct ulist *refs = NULL;
1942         struct ulist *roots = NULL;
1943         struct ulist_node *ref_node = NULL;
1944         struct ulist_node *root_node = NULL;
1945         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1946         struct ulist_iterator ref_uiter;
1947         struct ulist_iterator root_uiter;
1948
1949         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1950                         extent_item_objectid);
1951
1952         if (!search_commit_root) {
1953                 trans = btrfs_attach_transaction(fs_info->extent_root);
1954                 if (IS_ERR(trans)) {
1955                         if (PTR_ERR(trans) != -ENOENT &&
1956                             PTR_ERR(trans) != -EROFS)
1957                                 return PTR_ERR(trans);
1958                         trans = NULL;
1959                 }
1960         }
1961
1962         if (trans)
1963                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1964         else
1965                 down_read(&fs_info->commit_root_sem);
1966
1967         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1968                                    tree_mod_seq_elem.seq, &refs,
1969                                    &extent_item_pos, ignore_offset);
1970         if (ret)
1971                 goto out;
1972
1973         ULIST_ITER_INIT(&ref_uiter);
1974         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1975                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1976                                                 tree_mod_seq_elem.seq, &roots,
1977                                                 ignore_offset);
1978                 if (ret)
1979                         break;
1980                 ULIST_ITER_INIT(&root_uiter);
1981                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1982                         btrfs_debug(fs_info,
1983                                     "root %llu references leaf %llu, data list %#llx",
1984                                     root_node->val, ref_node->val,
1985                                     ref_node->aux);
1986                         ret = iterate_leaf_refs(fs_info,
1987                                                 (struct extent_inode_elem *)
1988                                                 (uintptr_t)ref_node->aux,
1989                                                 root_node->val,
1990                                                 extent_item_objectid,
1991                                                 iterate, ctx);
1992                 }
1993                 ulist_free(roots);
1994         }
1995
1996         free_leaf_list(refs);
1997 out:
1998         if (trans) {
1999                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2000                 btrfs_end_transaction(trans);
2001         } else {
2002                 up_read(&fs_info->commit_root_sem);
2003         }
2004
2005         return ret;
2006 }
2007
2008 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2009                                 struct btrfs_path *path,
2010                                 iterate_extent_inodes_t *iterate, void *ctx,
2011                                 bool ignore_offset)
2012 {
2013         int ret;
2014         u64 extent_item_pos;
2015         u64 flags = 0;
2016         struct btrfs_key found_key;
2017         int search_commit_root = path->search_commit_root;
2018
2019         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2020         btrfs_release_path(path);
2021         if (ret < 0)
2022                 return ret;
2023         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2024                 return -EINVAL;
2025
2026         extent_item_pos = logical - found_key.objectid;
2027         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2028                                         extent_item_pos, search_commit_root,
2029                                         iterate, ctx, ignore_offset);
2030
2031         return ret;
2032 }
2033
2034 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2035                               struct extent_buffer *eb, void *ctx);
2036
2037 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2038                               struct btrfs_path *path,
2039                               iterate_irefs_t *iterate, void *ctx)
2040 {
2041         int ret = 0;
2042         int slot;
2043         u32 cur;
2044         u32 len;
2045         u32 name_len;
2046         u64 parent = 0;
2047         int found = 0;
2048         struct extent_buffer *eb;
2049         struct btrfs_item *item;
2050         struct btrfs_inode_ref *iref;
2051         struct btrfs_key found_key;
2052
2053         while (!ret) {
2054                 ret = btrfs_find_item(fs_root, path, inum,
2055                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2056                                 &found_key);
2057
2058                 if (ret < 0)
2059                         break;
2060                 if (ret) {
2061                         ret = found ? 0 : -ENOENT;
2062                         break;
2063                 }
2064                 ++found;
2065
2066                 parent = found_key.offset;
2067                 slot = path->slots[0];
2068                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2069                 if (!eb) {
2070                         ret = -ENOMEM;
2071                         break;
2072                 }
2073                 btrfs_release_path(path);
2074
2075                 item = btrfs_item_nr(slot);
2076                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2077
2078                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2079                         name_len = btrfs_inode_ref_name_len(eb, iref);
2080                         /* path must be released before calling iterate()! */
2081                         btrfs_debug(fs_root->fs_info,
2082                                 "following ref at offset %u for inode %llu in tree %llu",
2083                                 cur, found_key.objectid,
2084                                 fs_root->root_key.objectid);
2085                         ret = iterate(parent, name_len,
2086                                       (unsigned long)(iref + 1), eb, ctx);
2087                         if (ret)
2088                                 break;
2089                         len = sizeof(*iref) + name_len;
2090                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2091                 }
2092                 free_extent_buffer(eb);
2093         }
2094
2095         btrfs_release_path(path);
2096
2097         return ret;
2098 }
2099
2100 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2101                                  struct btrfs_path *path,
2102                                  iterate_irefs_t *iterate, void *ctx)
2103 {
2104         int ret;
2105         int slot;
2106         u64 offset = 0;
2107         u64 parent;
2108         int found = 0;
2109         struct extent_buffer *eb;
2110         struct btrfs_inode_extref *extref;
2111         u32 item_size;
2112         u32 cur_offset;
2113         unsigned long ptr;
2114
2115         while (1) {
2116                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2117                                             &offset);
2118                 if (ret < 0)
2119                         break;
2120                 if (ret) {
2121                         ret = found ? 0 : -ENOENT;
2122                         break;
2123                 }
2124                 ++found;
2125
2126                 slot = path->slots[0];
2127                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2128                 if (!eb) {
2129                         ret = -ENOMEM;
2130                         break;
2131                 }
2132                 btrfs_release_path(path);
2133
2134                 item_size = btrfs_item_size_nr(eb, slot);
2135                 ptr = btrfs_item_ptr_offset(eb, slot);
2136                 cur_offset = 0;
2137
2138                 while (cur_offset < item_size) {
2139                         u32 name_len;
2140
2141                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2142                         parent = btrfs_inode_extref_parent(eb, extref);
2143                         name_len = btrfs_inode_extref_name_len(eb, extref);
2144                         ret = iterate(parent, name_len,
2145                                       (unsigned long)&extref->name, eb, ctx);
2146                         if (ret)
2147                                 break;
2148
2149                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2150                         cur_offset += sizeof(*extref);
2151                 }
2152                 free_extent_buffer(eb);
2153
2154                 offset++;
2155         }
2156
2157         btrfs_release_path(path);
2158
2159         return ret;
2160 }
2161
2162 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2163                          struct btrfs_path *path, iterate_irefs_t *iterate,
2164                          void *ctx)
2165 {
2166         int ret;
2167         int found_refs = 0;
2168
2169         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2170         if (!ret)
2171                 ++found_refs;
2172         else if (ret != -ENOENT)
2173                 return ret;
2174
2175         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2176         if (ret == -ENOENT && found_refs)
2177                 return 0;
2178
2179         return ret;
2180 }
2181
2182 /*
2183  * returns 0 if the path could be dumped (probably truncated)
2184  * returns <0 in case of an error
2185  */
2186 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2187                          struct extent_buffer *eb, void *ctx)
2188 {
2189         struct inode_fs_paths *ipath = ctx;
2190         char *fspath;
2191         char *fspath_min;
2192         int i = ipath->fspath->elem_cnt;
2193         const int s_ptr = sizeof(char *);
2194         u32 bytes_left;
2195
2196         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2197                                         ipath->fspath->bytes_left - s_ptr : 0;
2198
2199         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2200         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2201                                    name_off, eb, inum, fspath_min, bytes_left);
2202         if (IS_ERR(fspath))
2203                 return PTR_ERR(fspath);
2204
2205         if (fspath > fspath_min) {
2206                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2207                 ++ipath->fspath->elem_cnt;
2208                 ipath->fspath->bytes_left = fspath - fspath_min;
2209         } else {
2210                 ++ipath->fspath->elem_missed;
2211                 ipath->fspath->bytes_missing += fspath_min - fspath;
2212                 ipath->fspath->bytes_left = 0;
2213         }
2214
2215         return 0;
2216 }
2217
2218 /*
2219  * this dumps all file system paths to the inode into the ipath struct, provided
2220  * is has been created large enough. each path is zero-terminated and accessed
2221  * from ipath->fspath->val[i].
2222  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2223  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2224  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2225  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2226  * have been needed to return all paths.
2227  */
2228 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2229 {
2230         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2231                              inode_to_path, ipath);
2232 }
2233
2234 struct btrfs_data_container *init_data_container(u32 total_bytes)
2235 {
2236         struct btrfs_data_container *data;
2237         size_t alloc_bytes;
2238
2239         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2240         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2241         if (!data)
2242                 return ERR_PTR(-ENOMEM);
2243
2244         if (total_bytes >= sizeof(*data)) {
2245                 data->bytes_left = total_bytes - sizeof(*data);
2246                 data->bytes_missing = 0;
2247         } else {
2248                 data->bytes_missing = sizeof(*data) - total_bytes;
2249                 data->bytes_left = 0;
2250         }
2251
2252         data->elem_cnt = 0;
2253         data->elem_missed = 0;
2254
2255         return data;
2256 }
2257
2258 /*
2259  * allocates space to return multiple file system paths for an inode.
2260  * total_bytes to allocate are passed, note that space usable for actual path
2261  * information will be total_bytes - sizeof(struct inode_fs_paths).
2262  * the returned pointer must be freed with free_ipath() in the end.
2263  */
2264 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2265                                         struct btrfs_path *path)
2266 {
2267         struct inode_fs_paths *ifp;
2268         struct btrfs_data_container *fspath;
2269
2270         fspath = init_data_container(total_bytes);
2271         if (IS_ERR(fspath))
2272                 return ERR_CAST(fspath);
2273
2274         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2275         if (!ifp) {
2276                 kvfree(fspath);
2277                 return ERR_PTR(-ENOMEM);
2278         }
2279
2280         ifp->btrfs_path = path;
2281         ifp->fspath = fspath;
2282         ifp->fs_root = fs_root;
2283
2284         return ifp;
2285 }
2286
2287 void free_ipath(struct inode_fs_paths *ipath)
2288 {
2289         if (!ipath)
2290                 return;
2291         kvfree(ipath->fspath);
2292         kfree(ipath);
2293 }
2294
2295 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2296                 struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2297 {
2298         struct btrfs_backref_iter *ret;
2299
2300         ret = kzalloc(sizeof(*ret), gfp_flag);
2301         if (!ret)
2302                 return NULL;
2303
2304         ret->path = btrfs_alloc_path();
2305         if (!ret) {
2306                 kfree(ret);
2307                 return NULL;
2308         }
2309
2310         /* Current backref iterator only supports iteration in commit root */
2311         ret->path->search_commit_root = 1;
2312         ret->path->skip_locking = 1;
2313         ret->fs_info = fs_info;
2314
2315         return ret;
2316 }
2317
2318 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2319 {
2320         struct btrfs_fs_info *fs_info = iter->fs_info;
2321         struct btrfs_path *path = iter->path;
2322         struct btrfs_extent_item *ei;
2323         struct btrfs_key key;
2324         int ret;
2325
2326         key.objectid = bytenr;
2327         key.type = BTRFS_METADATA_ITEM_KEY;
2328         key.offset = (u64)-1;
2329         iter->bytenr = bytenr;
2330
2331         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
2332         if (ret < 0)
2333                 return ret;
2334         if (ret == 0) {
2335                 ret = -EUCLEAN;
2336                 goto release;
2337         }
2338         if (path->slots[0] == 0) {
2339                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2340                 ret = -EUCLEAN;
2341                 goto release;
2342         }
2343         path->slots[0]--;
2344
2345         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2346         if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2347              key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2348                 ret = -ENOENT;
2349                 goto release;
2350         }
2351         memcpy(&iter->cur_key, &key, sizeof(key));
2352         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2353                                                     path->slots[0]);
2354         iter->end_ptr = (u32)(iter->item_ptr +
2355                         btrfs_item_size_nr(path->nodes[0], path->slots[0]));
2356         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2357                             struct btrfs_extent_item);
2358
2359         /*
2360          * Only support iteration on tree backref yet.
2361          *
2362          * This is an extra precaution for non skinny-metadata, where
2363          * EXTENT_ITEM is also used for tree blocks, that we can only use
2364          * extent flags to determine if it's a tree block.
2365          */
2366         if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2367                 ret = -ENOTSUPP;
2368                 goto release;
2369         }
2370         iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2371
2372         /* If there is no inline backref, go search for keyed backref */
2373         if (iter->cur_ptr >= iter->end_ptr) {
2374                 ret = btrfs_next_item(fs_info->extent_root, path);
2375
2376                 /* No inline nor keyed ref */
2377                 if (ret > 0) {
2378                         ret = -ENOENT;
2379                         goto release;
2380                 }
2381                 if (ret < 0)
2382                         goto release;
2383
2384                 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2385                                 path->slots[0]);
2386                 if (iter->cur_key.objectid != bytenr ||
2387                     (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2388                      iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2389                         ret = -ENOENT;
2390                         goto release;
2391                 }
2392                 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2393                                                            path->slots[0]);
2394                 iter->item_ptr = iter->cur_ptr;
2395                 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size_nr(
2396                                       path->nodes[0], path->slots[0]));
2397         }
2398
2399         return 0;
2400 release:
2401         btrfs_backref_iter_release(iter);
2402         return ret;
2403 }
2404
2405 /*
2406  * Go to the next backref item of current bytenr, can be either inlined or
2407  * keyed.
2408  *
2409  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2410  *
2411  * Return 0 if we get next backref without problem.
2412  * Return >0 if there is no extra backref for this bytenr.
2413  * Return <0 if there is something wrong happened.
2414  */
2415 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2416 {
2417         struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2418         struct btrfs_path *path = iter->path;
2419         struct btrfs_extent_inline_ref *iref;
2420         int ret;
2421         u32 size;
2422
2423         if (btrfs_backref_iter_is_inline_ref(iter)) {
2424                 /* We're still inside the inline refs */
2425                 ASSERT(iter->cur_ptr < iter->end_ptr);
2426
2427                 if (btrfs_backref_has_tree_block_info(iter)) {
2428                         /* First tree block info */
2429                         size = sizeof(struct btrfs_tree_block_info);
2430                 } else {
2431                         /* Use inline ref type to determine the size */
2432                         int type;
2433
2434                         iref = (struct btrfs_extent_inline_ref *)
2435                                 ((unsigned long)iter->cur_ptr);
2436                         type = btrfs_extent_inline_ref_type(eb, iref);
2437
2438                         size = btrfs_extent_inline_ref_size(type);
2439                 }
2440                 iter->cur_ptr += size;
2441                 if (iter->cur_ptr < iter->end_ptr)
2442                         return 0;
2443
2444                 /* All inline items iterated, fall through */
2445         }
2446
2447         /* We're at keyed items, there is no inline item, go to the next one */
2448         ret = btrfs_next_item(iter->fs_info->extent_root, iter->path);
2449         if (ret)
2450                 return ret;
2451
2452         btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2453         if (iter->cur_key.objectid != iter->bytenr ||
2454             (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2455              iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2456                 return 1;
2457         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2458                                         path->slots[0]);
2459         iter->cur_ptr = iter->item_ptr;
2460         iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size_nr(path->nodes[0],
2461                                                 path->slots[0]);
2462         return 0;
2463 }
2464
2465 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2466                               struct btrfs_backref_cache *cache, int is_reloc)
2467 {
2468         int i;
2469
2470         cache->rb_root = RB_ROOT;
2471         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2472                 INIT_LIST_HEAD(&cache->pending[i]);
2473         INIT_LIST_HEAD(&cache->changed);
2474         INIT_LIST_HEAD(&cache->detached);
2475         INIT_LIST_HEAD(&cache->leaves);
2476         INIT_LIST_HEAD(&cache->pending_edge);
2477         INIT_LIST_HEAD(&cache->useless_node);
2478         cache->fs_info = fs_info;
2479         cache->is_reloc = is_reloc;
2480 }
2481
2482 struct btrfs_backref_node *btrfs_backref_alloc_node(
2483                 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2484 {
2485         struct btrfs_backref_node *node;
2486
2487         ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2488         node = kzalloc(sizeof(*node), GFP_NOFS);
2489         if (!node)
2490                 return node;
2491
2492         INIT_LIST_HEAD(&node->list);
2493         INIT_LIST_HEAD(&node->upper);
2494         INIT_LIST_HEAD(&node->lower);
2495         RB_CLEAR_NODE(&node->rb_node);
2496         cache->nr_nodes++;
2497         node->level = level;
2498         node->bytenr = bytenr;
2499
2500         return node;
2501 }
2502
2503 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2504                 struct btrfs_backref_cache *cache)
2505 {
2506         struct btrfs_backref_edge *edge;
2507
2508         edge = kzalloc(sizeof(*edge), GFP_NOFS);
2509         if (edge)
2510                 cache->nr_edges++;
2511         return edge;
2512 }
2513
2514 /*
2515  * Drop the backref node from cache, also cleaning up all its
2516  * upper edges and any uncached nodes in the path.
2517  *
2518  * This cleanup happens bottom up, thus the node should either
2519  * be the lowest node in the cache or a detached node.
2520  */
2521 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2522                                 struct btrfs_backref_node *node)
2523 {
2524         struct btrfs_backref_node *upper;
2525         struct btrfs_backref_edge *edge;
2526
2527         if (!node)
2528                 return;
2529
2530         BUG_ON(!node->lowest && !node->detached);
2531         while (!list_empty(&node->upper)) {
2532                 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2533                                   list[LOWER]);
2534                 upper = edge->node[UPPER];
2535                 list_del(&edge->list[LOWER]);
2536                 list_del(&edge->list[UPPER]);
2537                 btrfs_backref_free_edge(cache, edge);
2538
2539                 if (RB_EMPTY_NODE(&upper->rb_node)) {
2540                         BUG_ON(!list_empty(&node->upper));
2541                         btrfs_backref_drop_node(cache, node);
2542                         node = upper;
2543                         node->lowest = 1;
2544                         continue;
2545                 }
2546                 /*
2547                  * Add the node to leaf node list if no other child block
2548                  * cached.
2549                  */
2550                 if (list_empty(&upper->lower)) {
2551                         list_add_tail(&upper->lower, &cache->leaves);
2552                         upper->lowest = 1;
2553                 }
2554         }
2555
2556         btrfs_backref_drop_node(cache, node);
2557 }
2558
2559 /*
2560  * Release all nodes/edges from current cache
2561  */
2562 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2563 {
2564         struct btrfs_backref_node *node;
2565         int i;
2566
2567         while (!list_empty(&cache->detached)) {
2568                 node = list_entry(cache->detached.next,
2569                                   struct btrfs_backref_node, list);
2570                 btrfs_backref_cleanup_node(cache, node);
2571         }
2572
2573         while (!list_empty(&cache->leaves)) {
2574                 node = list_entry(cache->leaves.next,
2575                                   struct btrfs_backref_node, lower);
2576                 btrfs_backref_cleanup_node(cache, node);
2577         }
2578
2579         cache->last_trans = 0;
2580
2581         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2582                 ASSERT(list_empty(&cache->pending[i]));
2583         ASSERT(list_empty(&cache->pending_edge));
2584         ASSERT(list_empty(&cache->useless_node));
2585         ASSERT(list_empty(&cache->changed));
2586         ASSERT(list_empty(&cache->detached));
2587         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2588         ASSERT(!cache->nr_nodes);
2589         ASSERT(!cache->nr_edges);
2590 }
2591
2592 /*
2593  * Handle direct tree backref
2594  *
2595  * Direct tree backref means, the backref item shows its parent bytenr
2596  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2597  *
2598  * @ref_key:    The converted backref key.
2599  *              For keyed backref, it's the item key.
2600  *              For inlined backref, objectid is the bytenr,
2601  *              type is btrfs_inline_ref_type, offset is
2602  *              btrfs_inline_ref_offset.
2603  */
2604 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2605                                       struct btrfs_key *ref_key,
2606                                       struct btrfs_backref_node *cur)
2607 {
2608         struct btrfs_backref_edge *edge;
2609         struct btrfs_backref_node *upper;
2610         struct rb_node *rb_node;
2611
2612         ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2613
2614         /* Only reloc root uses backref pointing to itself */
2615         if (ref_key->objectid == ref_key->offset) {
2616                 struct btrfs_root *root;
2617
2618                 cur->is_reloc_root = 1;
2619                 /* Only reloc backref cache cares about a specific root */
2620                 if (cache->is_reloc) {
2621                         root = find_reloc_root(cache->fs_info, cur->bytenr);
2622                         if (WARN_ON(!root))
2623                                 return -ENOENT;
2624                         cur->root = root;
2625                 } else {
2626                         /*
2627                          * For generic purpose backref cache, reloc root node
2628                          * is useless.
2629                          */
2630                         list_add(&cur->list, &cache->useless_node);
2631                 }
2632                 return 0;
2633         }
2634
2635         edge = btrfs_backref_alloc_edge(cache);
2636         if (!edge)
2637                 return -ENOMEM;
2638
2639         rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2640         if (!rb_node) {
2641                 /* Parent node not yet cached */
2642                 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2643                                            cur->level + 1);
2644                 if (!upper) {
2645                         btrfs_backref_free_edge(cache, edge);
2646                         return -ENOMEM;
2647                 }
2648
2649                 /*
2650                  *  Backrefs for the upper level block isn't cached, add the
2651                  *  block to pending list
2652                  */
2653                 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2654         } else {
2655                 /* Parent node already cached */
2656                 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2657                 ASSERT(upper->checked);
2658                 INIT_LIST_HEAD(&edge->list[UPPER]);
2659         }
2660         btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2661         return 0;
2662 }
2663
2664 /*
2665  * Handle indirect tree backref
2666  *
2667  * Indirect tree backref means, we only know which tree the node belongs to.
2668  * We still need to do a tree search to find out the parents. This is for
2669  * TREE_BLOCK_REF backref (keyed or inlined).
2670  *
2671  * @ref_key:    The same as @ref_key in  handle_direct_tree_backref()
2672  * @tree_key:   The first key of this tree block.
2673  * @path:       A clean (released) path, to avoid allocating path everytime
2674  *              the function get called.
2675  */
2676 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2677                                         struct btrfs_path *path,
2678                                         struct btrfs_key *ref_key,
2679                                         struct btrfs_key *tree_key,
2680                                         struct btrfs_backref_node *cur)
2681 {
2682         struct btrfs_fs_info *fs_info = cache->fs_info;
2683         struct btrfs_backref_node *upper;
2684         struct btrfs_backref_node *lower;
2685         struct btrfs_backref_edge *edge;
2686         struct extent_buffer *eb;
2687         struct btrfs_root *root;
2688         struct rb_node *rb_node;
2689         int level;
2690         bool need_check = true;
2691         int ret;
2692
2693         root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2694         if (IS_ERR(root))
2695                 return PTR_ERR(root);
2696         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2697                 cur->cowonly = 1;
2698
2699         if (btrfs_root_level(&root->root_item) == cur->level) {
2700                 /* Tree root */
2701                 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2702                 /*
2703                  * For reloc backref cache, we may ignore reloc root.  But for
2704                  * general purpose backref cache, we can't rely on
2705                  * btrfs_should_ignore_reloc_root() as it may conflict with
2706                  * current running relocation and lead to missing root.
2707                  *
2708                  * For general purpose backref cache, reloc root detection is
2709                  * completely relying on direct backref (key->offset is parent
2710                  * bytenr), thus only do such check for reloc cache.
2711                  */
2712                 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2713                         btrfs_put_root(root);
2714                         list_add(&cur->list, &cache->useless_node);
2715                 } else {
2716                         cur->root = root;
2717                 }
2718                 return 0;
2719         }
2720
2721         level = cur->level + 1;
2722
2723         /* Search the tree to find parent blocks referring to the block */
2724         path->search_commit_root = 1;
2725         path->skip_locking = 1;
2726         path->lowest_level = level;
2727         ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2728         path->lowest_level = 0;
2729         if (ret < 0) {
2730                 btrfs_put_root(root);
2731                 return ret;
2732         }
2733         if (ret > 0 && path->slots[level] > 0)
2734                 path->slots[level]--;
2735
2736         eb = path->nodes[level];
2737         if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2738                 btrfs_err(fs_info,
2739 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2740                           cur->bytenr, level - 1, root->root_key.objectid,
2741                           tree_key->objectid, tree_key->type, tree_key->offset);
2742                 btrfs_put_root(root);
2743                 ret = -ENOENT;
2744                 goto out;
2745         }
2746         lower = cur;
2747
2748         /* Add all nodes and edges in the path */
2749         for (; level < BTRFS_MAX_LEVEL; level++) {
2750                 if (!path->nodes[level]) {
2751                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
2752                                lower->bytenr);
2753                         /* Same as previous should_ignore_reloc_root() call */
2754                         if (btrfs_should_ignore_reloc_root(root) &&
2755                             cache->is_reloc) {
2756                                 btrfs_put_root(root);
2757                                 list_add(&lower->list, &cache->useless_node);
2758                         } else {
2759                                 lower->root = root;
2760                         }
2761                         break;
2762                 }
2763
2764                 edge = btrfs_backref_alloc_edge(cache);
2765                 if (!edge) {
2766                         btrfs_put_root(root);
2767                         ret = -ENOMEM;
2768                         goto out;
2769                 }
2770
2771                 eb = path->nodes[level];
2772                 rb_node = rb_simple_search(&cache->rb_root, eb->start);
2773                 if (!rb_node) {
2774                         upper = btrfs_backref_alloc_node(cache, eb->start,
2775                                                          lower->level + 1);
2776                         if (!upper) {
2777                                 btrfs_put_root(root);
2778                                 btrfs_backref_free_edge(cache, edge);
2779                                 ret = -ENOMEM;
2780                                 goto out;
2781                         }
2782                         upper->owner = btrfs_header_owner(eb);
2783                         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2784                                 upper->cowonly = 1;
2785
2786                         /*
2787                          * If we know the block isn't shared we can avoid
2788                          * checking its backrefs.
2789                          */
2790                         if (btrfs_block_can_be_shared(root, eb))
2791                                 upper->checked = 0;
2792                         else
2793                                 upper->checked = 1;
2794
2795                         /*
2796                          * Add the block to pending list if we need to check its
2797                          * backrefs, we only do this once while walking up a
2798                          * tree as we will catch anything else later on.
2799                          */
2800                         if (!upper->checked && need_check) {
2801                                 need_check = false;
2802                                 list_add_tail(&edge->list[UPPER],
2803                                               &cache->pending_edge);
2804                         } else {
2805                                 if (upper->checked)
2806                                         need_check = true;
2807                                 INIT_LIST_HEAD(&edge->list[UPPER]);
2808                         }
2809                 } else {
2810                         upper = rb_entry(rb_node, struct btrfs_backref_node,
2811                                          rb_node);
2812                         ASSERT(upper->checked);
2813                         INIT_LIST_HEAD(&edge->list[UPPER]);
2814                         if (!upper->owner)
2815                                 upper->owner = btrfs_header_owner(eb);
2816                 }
2817                 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2818
2819                 if (rb_node) {
2820                         btrfs_put_root(root);
2821                         break;
2822                 }
2823                 lower = upper;
2824                 upper = NULL;
2825         }
2826 out:
2827         btrfs_release_path(path);
2828         return ret;
2829 }
2830
2831 /*
2832  * Add backref node @cur into @cache.
2833  *
2834  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2835  *       links aren't yet bi-directional. Needs to finish such links.
2836  *       Use btrfs_backref_finish_upper_links() to finish such linkage.
2837  *
2838  * @path:       Released path for indirect tree backref lookup
2839  * @iter:       Released backref iter for extent tree search
2840  * @node_key:   The first key of the tree block
2841  */
2842 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2843                                 struct btrfs_path *path,
2844                                 struct btrfs_backref_iter *iter,
2845                                 struct btrfs_key *node_key,
2846                                 struct btrfs_backref_node *cur)
2847 {
2848         struct btrfs_fs_info *fs_info = cache->fs_info;
2849         struct btrfs_backref_edge *edge;
2850         struct btrfs_backref_node *exist;
2851         int ret;
2852
2853         ret = btrfs_backref_iter_start(iter, cur->bytenr);
2854         if (ret < 0)
2855                 return ret;
2856         /*
2857          * We skip the first btrfs_tree_block_info, as we don't use the key
2858          * stored in it, but fetch it from the tree block
2859          */
2860         if (btrfs_backref_has_tree_block_info(iter)) {
2861                 ret = btrfs_backref_iter_next(iter);
2862                 if (ret < 0)
2863                         goto out;
2864                 /* No extra backref? This means the tree block is corrupted */
2865                 if (ret > 0) {
2866                         ret = -EUCLEAN;
2867                         goto out;
2868                 }
2869         }
2870         WARN_ON(cur->checked);
2871         if (!list_empty(&cur->upper)) {
2872                 /*
2873                  * The backref was added previously when processing backref of
2874                  * type BTRFS_TREE_BLOCK_REF_KEY
2875                  */
2876                 ASSERT(list_is_singular(&cur->upper));
2877                 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2878                                   list[LOWER]);
2879                 ASSERT(list_empty(&edge->list[UPPER]));
2880                 exist = edge->node[UPPER];
2881                 /*
2882                  * Add the upper level block to pending list if we need check
2883                  * its backrefs
2884                  */
2885                 if (!exist->checked)
2886                         list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2887         } else {
2888                 exist = NULL;
2889         }
2890
2891         for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2892                 struct extent_buffer *eb;
2893                 struct btrfs_key key;
2894                 int type;
2895
2896                 cond_resched();
2897                 eb = btrfs_backref_get_eb(iter);
2898
2899                 key.objectid = iter->bytenr;
2900                 if (btrfs_backref_iter_is_inline_ref(iter)) {
2901                         struct btrfs_extent_inline_ref *iref;
2902
2903                         /* Update key for inline backref */
2904                         iref = (struct btrfs_extent_inline_ref *)
2905                                 ((unsigned long)iter->cur_ptr);
2906                         type = btrfs_get_extent_inline_ref_type(eb, iref,
2907                                                         BTRFS_REF_TYPE_BLOCK);
2908                         if (type == BTRFS_REF_TYPE_INVALID) {
2909                                 ret = -EUCLEAN;
2910                                 goto out;
2911                         }
2912                         key.type = type;
2913                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2914                 } else {
2915                         key.type = iter->cur_key.type;
2916                         key.offset = iter->cur_key.offset;
2917                 }
2918
2919                 /*
2920                  * Parent node found and matches current inline ref, no need to
2921                  * rebuild this node for this inline ref
2922                  */
2923                 if (exist &&
2924                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2925                       exist->owner == key.offset) ||
2926                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2927                       exist->bytenr == key.offset))) {
2928                         exist = NULL;
2929                         continue;
2930                 }
2931
2932                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
2933                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2934                         ret = handle_direct_tree_backref(cache, &key, cur);
2935                         if (ret < 0)
2936                                 goto out;
2937                         continue;
2938                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
2939                         ret = -EINVAL;
2940                         btrfs_print_v0_err(fs_info);
2941                         btrfs_handle_fs_error(fs_info, ret, NULL);
2942                         goto out;
2943                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
2944                         continue;
2945                 }
2946
2947                 /*
2948                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
2949                  * means the root objectid. We need to search the tree to get
2950                  * its parent bytenr.
2951                  */
2952                 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
2953                                                    cur);
2954                 if (ret < 0)
2955                         goto out;
2956         }
2957         ret = 0;
2958         cur->checked = 1;
2959         WARN_ON(exist);
2960 out:
2961         btrfs_backref_iter_release(iter);
2962         return ret;
2963 }
2964
2965 /*
2966  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
2967  */
2968 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
2969                                      struct btrfs_backref_node *start)
2970 {
2971         struct list_head *useless_node = &cache->useless_node;
2972         struct btrfs_backref_edge *edge;
2973         struct rb_node *rb_node;
2974         LIST_HEAD(pending_edge);
2975
2976         ASSERT(start->checked);
2977
2978         /* Insert this node to cache if it's not COW-only */
2979         if (!start->cowonly) {
2980                 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
2981                                            &start->rb_node);
2982                 if (rb_node)
2983                         btrfs_backref_panic(cache->fs_info, start->bytenr,
2984                                             -EEXIST);
2985                 list_add_tail(&start->lower, &cache->leaves);
2986         }
2987
2988         /*
2989          * Use breadth first search to iterate all related edges.
2990          *
2991          * The starting points are all the edges of this node
2992          */
2993         list_for_each_entry(edge, &start->upper, list[LOWER])
2994                 list_add_tail(&edge->list[UPPER], &pending_edge);
2995
2996         while (!list_empty(&pending_edge)) {
2997                 struct btrfs_backref_node *upper;
2998                 struct btrfs_backref_node *lower;
2999                 struct rb_node *rb_node;
3000
3001                 edge = list_first_entry(&pending_edge,
3002                                 struct btrfs_backref_edge, list[UPPER]);
3003                 list_del_init(&edge->list[UPPER]);
3004                 upper = edge->node[UPPER];
3005                 lower = edge->node[LOWER];
3006
3007                 /* Parent is detached, no need to keep any edges */
3008                 if (upper->detached) {
3009                         list_del(&edge->list[LOWER]);
3010                         btrfs_backref_free_edge(cache, edge);
3011
3012                         /* Lower node is orphan, queue for cleanup */
3013                         if (list_empty(&lower->upper))
3014                                 list_add(&lower->list, useless_node);
3015                         continue;
3016                 }
3017
3018                 /*
3019                  * All new nodes added in current build_backref_tree() haven't
3020                  * been linked to the cache rb tree.
3021                  * So if we have upper->rb_node populated, this means a cache
3022                  * hit. We only need to link the edge, as @upper and all its
3023                  * parents have already been linked.
3024                  */
3025                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3026                         if (upper->lowest) {
3027                                 list_del_init(&upper->lower);
3028                                 upper->lowest = 0;
3029                         }
3030
3031                         list_add_tail(&edge->list[UPPER], &upper->lower);
3032                         continue;
3033                 }
3034
3035                 /* Sanity check, we shouldn't have any unchecked nodes */
3036                 if (!upper->checked) {
3037                         ASSERT(0);
3038                         return -EUCLEAN;
3039                 }
3040
3041                 /* Sanity check, COW-only node has non-COW-only parent */
3042                 if (start->cowonly != upper->cowonly) {
3043                         ASSERT(0);
3044                         return -EUCLEAN;
3045                 }
3046
3047                 /* Only cache non-COW-only (subvolume trees) tree blocks */
3048                 if (!upper->cowonly) {
3049                         rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3050                                                    &upper->rb_node);
3051                         if (rb_node) {
3052                                 btrfs_backref_panic(cache->fs_info,
3053                                                 upper->bytenr, -EEXIST);
3054                                 return -EUCLEAN;
3055                         }
3056                 }
3057
3058                 list_add_tail(&edge->list[UPPER], &upper->lower);
3059
3060                 /*
3061                  * Also queue all the parent edges of this uncached node
3062                  * to finish the upper linkage
3063                  */
3064                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3065                         list_add_tail(&edge->list[UPPER], &pending_edge);
3066         }
3067         return 0;
3068 }
3069
3070 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3071                                  struct btrfs_backref_node *node)
3072 {
3073         struct btrfs_backref_node *lower;
3074         struct btrfs_backref_node *upper;
3075         struct btrfs_backref_edge *edge;
3076
3077         while (!list_empty(&cache->useless_node)) {
3078                 lower = list_first_entry(&cache->useless_node,
3079                                    struct btrfs_backref_node, list);
3080                 list_del_init(&lower->list);
3081         }
3082         while (!list_empty(&cache->pending_edge)) {
3083                 edge = list_first_entry(&cache->pending_edge,
3084                                 struct btrfs_backref_edge, list[UPPER]);
3085                 list_del(&edge->list[UPPER]);
3086                 list_del(&edge->list[LOWER]);
3087                 lower = edge->node[LOWER];
3088                 upper = edge->node[UPPER];
3089                 btrfs_backref_free_edge(cache, edge);
3090
3091                 /*
3092                  * Lower is no longer linked to any upper backref nodes and
3093                  * isn't in the cache, we can free it ourselves.
3094                  */
3095                 if (list_empty(&lower->upper) &&
3096                     RB_EMPTY_NODE(&lower->rb_node))
3097                         list_add(&lower->list, &cache->useless_node);
3098
3099                 if (!RB_EMPTY_NODE(&upper->rb_node))
3100                         continue;
3101
3102                 /* Add this guy's upper edges to the list to process */
3103                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3104                         list_add_tail(&edge->list[UPPER],
3105                                       &cache->pending_edge);
3106                 if (list_empty(&upper->upper))
3107                         list_add(&upper->list, &cache->useless_node);
3108         }
3109
3110         while (!list_empty(&cache->useless_node)) {
3111                 lower = list_first_entry(&cache->useless_node,
3112                                    struct btrfs_backref_node, list);
3113                 list_del_init(&lower->list);
3114                 if (lower == node)
3115                         node = NULL;
3116                 btrfs_backref_free_node(cache, lower);
3117         }
3118
3119         btrfs_backref_cleanup_node(cache, node);
3120         ASSERT(list_empty(&cache->useless_node) &&
3121                list_empty(&cache->pending_edge));
3122 }