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 #include "tree-mod-log.h"
  18
  19 /* Just an arbitrary number so we can be sure this happened */
  20 #define BACKREF_FOUND_SHARED 6
  21
  22 struct extent_inode_elem {
  23         u64 inum;
  24         u64 offset;
  25         struct extent_inode_elem *next;
  26 };
  27
  28 static int check_extent_in_eb(const struct btrfs_key *key,
  29                               const struct extent_buffer *eb,
  30                               const struct btrfs_file_extent_item *fi,
  31                               u64 extent_item_pos,
  32                               struct extent_inode_elem **eie,
  33                               bool ignore_offset)
  34 {
  35         u64 offset = 0;
  36         struct extent_inode_elem *e;
  37
  38         if (!ignore_offset &&
  39             !btrfs_file_extent_compression(eb, fi) &&
  40             !btrfs_file_extent_encryption(eb, fi) &&
  41             !btrfs_file_extent_other_encoding(eb, fi)) {
  42                 u64 data_offset;
  43                 u64 data_len;
  44
  45                 data_offset = btrfs_file_extent_offset(eb, fi);
  46                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  47
  48                 if (extent_item_pos < data_offset ||
  49                     extent_item_pos >= data_offset + data_len)
  50                         return 1;
  51                 offset = extent_item_pos - data_offset;
  52         }
  53
  54         e = kmalloc(sizeof(*e), GFP_NOFS);
  55         if (!e)
  56                 return -ENOMEM;
  57
  58         e->next = *eie;
  59         e->inum = key->objectid;
  60         e->offset = key->offset + offset;
  61         *eie = e;
  62
  63         return 0;
  64 }
  65
  66 static void free_inode_elem_list(struct extent_inode_elem *eie)
  67 {
  68         struct extent_inode_elem *eie_next;
  69
  70         for (; eie; eie = eie_next) {
  71                 eie_next = eie->next;
  72                 kfree(eie);
  73         }
  74 }
  75
  76 static int find_extent_in_eb(const struct extent_buffer *eb,
  77                              u64 wanted_disk_byte, u64 extent_item_pos,
  78                              struct extent_inode_elem **eie,
  79                              bool ignore_offset)
  80 {
  81         u64 disk_byte;
  82         struct btrfs_key key;
  83         struct btrfs_file_extent_item *fi;
  84         int slot;
  85         int nritems;
  86         int extent_type;
  87         int ret;
  88
  89         /*
  90          * from the shared data ref, we only have the leaf but we need
  91          * the key. thus, we must look into all items and see that we
  92          * find one (some) with a reference to our extent item.
  93          */
  94         nritems = btrfs_header_nritems(eb);
  95         for (slot = 0; slot < nritems; ++slot) {
  96                 btrfs_item_key_to_cpu(eb, &key, slot);
  97                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  98                         continue;
  99                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 100                 extent_type = btrfs_file_extent_type(eb, fi);
 101                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 102                         continue;
 103                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 104                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 105                 if (disk_byte != wanted_disk_byte)
 106                         continue;
 107
 108                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 109                 if (ret < 0)
 110                         return ret;
 111         }
 112
 113         return 0;
 114 }
 115
 116 struct preftree {
 117         struct rb_root_cached root;
 118         unsigned int count;
 119 };
 120
 121 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 122
 123 struct preftrees {
 124         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 125         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 126         struct preftree indirect_missing_keys;
 127 };
 128
 129 /*
 130  * Checks for a shared extent during backref search.
 131  *
 132  * The share_count tracks prelim_refs (direct and indirect) having a
 133  * ref->count >0:
 134  *  - incremented when a ref->count transitions to >0
 135  *  - decremented when a ref->count transitions to <1
 136  */
 137 struct share_check {
 138         u64 root_objectid;
 139         u64 inum;
 140         int share_count;
 141         bool have_delayed_delete_refs;
 142 };
 143
 144 static inline int extent_is_shared(struct share_check *sc)
 145 {
 146         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 147 }
 148
 149 static struct kmem_cache *btrfs_prelim_ref_cache;
 150
 151 int __init btrfs_prelim_ref_init(void)
 152 {
 153         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 154                                         sizeof(struct prelim_ref),
 155                                         0,
 156                                         SLAB_MEM_SPREAD,
 157                                         NULL);
 158         if (!btrfs_prelim_ref_cache)
 159                 return -ENOMEM;
 160         return 0;
 161 }
 162
 163 void __cold btrfs_prelim_ref_exit(void)
 164 {
 165         kmem_cache_destroy(btrfs_prelim_ref_cache);
 166 }
 167
 168 static void free_pref(struct prelim_ref *ref)
 169 {
 170         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 171 }
 172
 173 /*
 174  * Return 0 when both refs are for the same block (and can be merged).
 175  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 176  * indicates a 'higher' block.
 177  */
 178 static int prelim_ref_compare(struct prelim_ref *ref1,
 179                               struct prelim_ref *ref2)
 180 {
 181         if (ref1->level < ref2->level)
 182                 return -1;
 183         if (ref1->level > ref2->level)
 184                 return 1;
 185         if (ref1->root_id < ref2->root_id)
 186                 return -1;
 187         if (ref1->root_id > ref2->root_id)
 188                 return 1;
 189         if (ref1->key_for_search.type < ref2->key_for_search.type)
 190                 return -1;
 191         if (ref1->key_for_search.type > ref2->key_for_search.type)
 192                 return 1;
 193         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 194                 return -1;
 195         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 196                 return 1;
 197         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 198                 return -1;
 199         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 200                 return 1;
 201         if (ref1->parent < ref2->parent)
 202                 return -1;
 203         if (ref1->parent > ref2->parent)
 204                 return 1;
 205
 206         return 0;
 207 }
 208
 209 static void update_share_count(struct share_check *sc, int oldcount,
 210                                int newcount)
 211 {
 212         if ((!sc) || (oldcount == 0 && newcount < 1))
 213                 return;
 214
 215         if (oldcount > 0 && newcount < 1)
 216                 sc->share_count--;
 217         else if (oldcount < 1 && newcount > 0)
 218                 sc->share_count++;
 219 }
 220
 221 /*
 222  * Add @newref to the @root rbtree, merging identical refs.
 223  *
 224  * Callers should assume that newref has been freed after calling.
 225  */
 226 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 227                               struct preftree *preftree,
 228                               struct prelim_ref *newref,
 229                               struct share_check *sc)
 230 {
 231         struct rb_root_cached *root;
 232         struct rb_node **p;
 233         struct rb_node *parent = NULL;
 234         struct prelim_ref *ref;
 235         int result;
 236         bool leftmost = true;
 237
 238         root = &preftree->root;
 239         p = &root->rb_root.rb_node;
 240
 241         while (*p) {
 242                 parent = *p;
 243                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 244                 result = prelim_ref_compare(ref, newref);
 245                 if (result < 0) {
 246                         p = &(*p)->rb_left;
 247                 } else if (result > 0) {
 248                         p = &(*p)->rb_right;
 249                         leftmost = false;
 250                 } else {
 251                         /* Identical refs, merge them and free @newref */
 252                         struct extent_inode_elem *eie = ref->inode_list;
 253
 254                         while (eie && eie->next)
 255                                 eie = eie->next;
 256
 257                         if (!eie)
 258                                 ref->inode_list = newref->inode_list;
 259                         else
 260                                 eie->next = newref->inode_list;
 261                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 262                                                      preftree->count);
 263                         /*
 264                          * A delayed ref can have newref->count < 0.
 265                          * The ref->count is updated to follow any
 266                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 267                          */
 268                         update_share_count(sc, ref->count,
 269                                            ref->count + newref->count);
 270                         ref->count += newref->count;
 271                         free_pref(newref);
 272                         return;
 273                 }
 274         }
 275
 276         update_share_count(sc, 0, newref->count);
 277         preftree->count++;
 278         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 279         rb_link_node(&newref->rbnode, parent, p);
 280         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 281 }
 282
 283 /*
 284  * Release the entire tree.  We don't care about internal consistency so
 285  * just free everything and then reset the tree root.
 286  */
 287 static void prelim_release(struct preftree *preftree)
 288 {
 289         struct prelim_ref *ref, *next_ref;
 290
 291         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 292                                              &preftree->root.rb_root, rbnode)
 293                 free_pref(ref);
 294
 295         preftree->root = RB_ROOT_CACHED;
 296         preftree->count = 0;
 297 }
 298
 299 /*
 300  * the rules for all callers of this function are:
 301  * - obtaining the parent is the goal
 302  * - if you add a key, you must know that it is a correct key
 303  * - if you cannot add the parent or a correct key, then we will look into the
 304  *   block later to set a correct key
 305  *
 306  * delayed refs
 307  * ============
 308  *        backref type | shared | indirect | shared | indirect
 309  * information         |   tree |     tree |   data |     data
 310  * --------------------+--------+----------+--------+----------
 311  *      parent logical |    y   |     -    |    -   |     -
 312  *      key to resolve |    -   |     y    |    y   |     y
 313  *  tree block logical |    -   |     -    |    -   |     -
 314  *  root for resolving |    y   |     y    |    y   |     y
 315  *
 316  * - column 1:       we've the parent -> done
 317  * - column 2, 3, 4: we use the key to find the parent
 318  *
 319  * on disk refs (inline or keyed)
 320  * ==============================
 321  *        backref type | shared | indirect | shared | indirect
 322  * information         |   tree |     tree |   data |     data
 323  * --------------------+--------+----------+--------+----------
 324  *      parent logical |    y   |     -    |    y   |     -
 325  *      key to resolve |    -   |     -    |    -   |     y
 326  *  tree block logical |    y   |     y    |    y   |     y
 327  *  root for resolving |    -   |     y    |    y   |     y
 328  *
 329  * - column 1, 3: we've the parent -> done
 330  * - column 2:    we take the first key from the block to find the parent
 331  *                (see add_missing_keys)
 332  * - column 4:    we use the key to find the parent
 333  *
 334  * additional information that's available but not required to find the parent
 335  * block might help in merging entries to gain some speed.
 336  */
 337 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 338                           struct preftree *preftree, u64 root_id,
 339                           const struct btrfs_key *key, int level, u64 parent,
 340                           u64 wanted_disk_byte, int count,
 341                           struct share_check *sc, gfp_t gfp_mask)
 342 {
 343         struct prelim_ref *ref;
 344
 345         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 346                 return 0;
 347
 348         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 349         if (!ref)
 350                 return -ENOMEM;
 351
 352         ref->root_id = root_id;
 353         if (key)
 354                 ref->key_for_search = *key;
 355         else
 356                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 357
 358         ref->inode_list = NULL;
 359         ref->level = level;
 360         ref->count = count;
 361         ref->parent = parent;
 362         ref->wanted_disk_byte = wanted_disk_byte;
 363         prelim_ref_insert(fs_info, preftree, ref, sc);
 364         return extent_is_shared(sc);
 365 }
 366
 367 /* direct refs use root == 0, key == NULL */
 368 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 369                           struct preftrees *preftrees, int level, u64 parent,
 370                           u64 wanted_disk_byte, int count,
 371                           struct share_check *sc, gfp_t gfp_mask)
 372 {
 373         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 374                               parent, wanted_disk_byte, count, sc, gfp_mask);
 375 }
 376
 377 /* indirect refs use parent == 0 */
 378 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 379                             struct preftrees *preftrees, u64 root_id,
 380                             const struct btrfs_key *key, int level,
 381                             u64 wanted_disk_byte, int count,
 382                             struct share_check *sc, gfp_t gfp_mask)
 383 {
 384         struct preftree *tree = &preftrees->indirect;
 385
 386         if (!key)
 387                 tree = &preftrees->indirect_missing_keys;
 388         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 389                               wanted_disk_byte, count, sc, gfp_mask);
 390 }
 391
 392 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 393 {
 394         struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 395         struct rb_node *parent = NULL;
 396         struct prelim_ref *ref = NULL;
 397         struct prelim_ref target = {};
 398         int result;
 399
 400         target.parent = bytenr;
 401
 402         while (*p) {
 403                 parent = *p;
 404                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 405                 result = prelim_ref_compare(ref, &target);
 406
 407                 if (result < 0)
 408                         p = &(*p)->rb_left;
 409                 else if (result > 0)
 410                         p = &(*p)->rb_right;
 411                 else
 412                         return 1;
 413         }
 414         return 0;
 415 }
 416
 417 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 418                            struct ulist *parents,
 419                            struct preftrees *preftrees, struct prelim_ref *ref,
 420                            int level, u64 time_seq, const u64 *extent_item_pos,
 421                            bool ignore_offset)
 422 {
 423         int ret = 0;
 424         int slot;
 425         struct extent_buffer *eb;
 426         struct btrfs_key key;
 427         struct btrfs_key *key_for_search = &ref->key_for_search;
 428         struct btrfs_file_extent_item *fi;
 429         struct extent_inode_elem *eie = NULL, *old = NULL;
 430         u64 disk_byte;
 431         u64 wanted_disk_byte = ref->wanted_disk_byte;
 432         u64 count = 0;
 433         u64 data_offset;
 434
 435         if (level != 0) {
 436                 eb = path->nodes[level];
 437                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 438                 if (ret < 0)
 439                         return ret;
 440                 return 0;
 441         }
 442
 443         /*
 444          * 1. We normally enter this function with the path already pointing to
 445          *    the first item to check. But sometimes, we may enter it with
 446          *    slot == nritems.
 447          * 2. We are searching for normal backref but bytenr of this leaf
 448          *    matches shared data backref
 449          * 3. The leaf owner is not equal to the root we are searching
 450          *
 451          * For these cases, go to the next leaf before we continue.
 452          */
 453         eb = path->nodes[0];
 454         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 455             is_shared_data_backref(preftrees, eb->start) ||
 456             ref->root_id != btrfs_header_owner(eb)) {
 457                 if (time_seq == BTRFS_SEQ_LAST)
 458                         ret = btrfs_next_leaf(root, path);
 459                 else
 460                         ret = btrfs_next_old_leaf(root, path, time_seq);
 461         }
 462
 463         while (!ret && count < ref->count) {
 464                 eb = path->nodes[0];
 465                 slot = path->slots[0];
 466
 467                 btrfs_item_key_to_cpu(eb, &key, slot);
 468
 469                 if (key.objectid != key_for_search->objectid ||
 470                     key.type != BTRFS_EXTENT_DATA_KEY)
 471                         break;
 472
 473                 /*
 474                  * We are searching for normal backref but bytenr of this leaf
 475                  * matches shared data backref, OR
 476                  * the leaf owner is not equal to the root we are searching for
 477                  */
 478                 if (slot == 0 &&
 479                     (is_shared_data_backref(preftrees, eb->start) ||
 480                      ref->root_id != btrfs_header_owner(eb))) {
 481                         if (time_seq == BTRFS_SEQ_LAST)
 482                                 ret = btrfs_next_leaf(root, path);
 483                         else
 484                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 485                         continue;
 486                 }
 487                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 488                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 489                 data_offset = btrfs_file_extent_offset(eb, fi);
 490
 491                 if (disk_byte == wanted_disk_byte) {
 492                         eie = NULL;
 493                         old = NULL;
 494                         if (ref->key_for_search.offset == key.offset - data_offset)
 495                                 count++;
 496                         else
 497                                 goto next;
 498                         if (extent_item_pos) {
 499                                 ret = check_extent_in_eb(&key, eb, fi,
 500                                                 *extent_item_pos,
 501                                                 &eie, ignore_offset);
 502                                 if (ret < 0)
 503                                         break;
 504                         }
 505                         if (ret > 0)
 506                                 goto next;
 507                         ret = ulist_add_merge_ptr(parents, eb->start,
 508                                                   eie, (void **)&old, GFP_NOFS);
 509                         if (ret < 0)
 510                                 break;
 511                         if (!ret && extent_item_pos) {
 512                                 while (old->next)
 513                                         old = old->next;
 514                                 old->next = eie;
 515                         }
 516                         eie = NULL;
 517                 }
 518 next:
 519                 if (time_seq == BTRFS_SEQ_LAST)
 520                         ret = btrfs_next_item(root, path);
 521                 else
 522                         ret = btrfs_next_old_item(root, path, time_seq);
 523         }
 524
 525         if (ret > 0)
 526                 ret = 0;
 527         else if (ret < 0)
 528                 free_inode_elem_list(eie);
 529         return ret;
 530 }
 531
 532 /*
 533  * resolve an indirect backref in the form (root_id, key, level)
 534  * to a logical address
 535  */
 536 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 537                                 struct btrfs_path *path, u64 time_seq,
 538                                 struct preftrees *preftrees,
 539                                 struct prelim_ref *ref, struct ulist *parents,
 540                                 const u64 *extent_item_pos, bool ignore_offset)
 541 {
 542         struct btrfs_root *root;
 543         struct extent_buffer *eb;
 544         int ret = 0;
 545         int root_level;
 546         int level = ref->level;
 547         struct btrfs_key search_key = ref->key_for_search;
 548
 549         /*
 550          * If we're search_commit_root we could possibly be holding locks on
 551          * other tree nodes.  This happens when qgroups does backref walks when
 552          * adding new delayed refs.  To deal with this we need to look in cache
 553          * for the root, and if we don't find it then we need to search the
 554          * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
 555          * here.
 556          */
 557         if (path->search_commit_root)
 558                 root = btrfs_get_fs_root_commit_root(fs_info, path, ref->root_id);
 559         else
 560                 root = btrfs_get_fs_root(fs_info, ref->root_id, false);
 561         if (IS_ERR(root)) {
 562                 ret = PTR_ERR(root);
 563                 goto out_free;
 564         }
 565
 566         if (!path->search_commit_root &&
 567             test_bit(BTRFS_ROOT_DELETING, &root->state)) {
 568                 ret = -ENOENT;
 569                 goto out;
 570         }
 571
 572         if (btrfs_is_testing(fs_info)) {
 573                 ret = -ENOENT;
 574                 goto out;
 575         }
 576
 577         if (path->search_commit_root)
 578                 root_level = btrfs_header_level(root->commit_root);
 579         else if (time_seq == BTRFS_SEQ_LAST)
 580                 root_level = btrfs_header_level(root->node);
 581         else
 582                 root_level = btrfs_old_root_level(root, time_seq);
 583
 584         if (root_level + 1 == level)
 585                 goto out;
 586
 587         /*
 588          * We can often find data backrefs with an offset that is too large
 589          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 590          * subtracting a file's offset with the data offset of its
 591          * corresponding extent data item. This can happen for example in the
 592          * clone ioctl.
 593          *
 594          * So if we detect such case we set the search key's offset to zero to
 595          * make sure we will find the matching file extent item at
 596          * add_all_parents(), otherwise we will miss it because the offset
 597          * taken form the backref is much larger then the offset of the file
 598          * extent item. This can make us scan a very large number of file
 599          * extent items, but at least it will not make us miss any.
 600          *
 601          * This is an ugly workaround for a behaviour that should have never
 602          * existed, but it does and a fix for the clone ioctl would touch a lot
 603          * of places, cause backwards incompatibility and would not fix the
 604          * problem for extents cloned with older kernels.
 605          */
 606         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 607             search_key.offset >= LLONG_MAX)
 608                 search_key.offset = 0;
 609         path->lowest_level = level;
 610         if (time_seq == BTRFS_SEQ_LAST)
 611                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 612         else
 613                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 614
 615         btrfs_debug(fs_info,
 616                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 617                  ref->root_id, level, ref->count, ret,
 618                  ref->key_for_search.objectid, ref->key_for_search.type,
 619                  ref->key_for_search.offset);
 620         if (ret < 0)
 621                 goto out;
 622
 623         eb = path->nodes[level];
 624         while (!eb) {
 625                 if (WARN_ON(!level)) {
 626                         ret = 1;
 627                         goto out;
 628                 }
 629                 level--;
 630                 eb = path->nodes[level];
 631         }
 632
 633         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 634                               time_seq, extent_item_pos, ignore_offset);
 635 out:
 636         btrfs_put_root(root);
 637 out_free:
 638         path->lowest_level = 0;
 639         btrfs_release_path(path);
 640         return ret;
 641 }
 642
 643 static struct extent_inode_elem *
 644 unode_aux_to_inode_list(struct ulist_node *node)
 645 {
 646         if (!node)
 647                 return NULL;
 648         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 649 }
 650
 651 /*
 652  * We maintain three separate rbtrees: one for direct refs, one for
 653  * indirect refs which have a key, and one for indirect refs which do not
 654  * have a key. Each tree does merge on insertion.
 655  *
 656  * Once all of the references are located, we iterate over the tree of
 657  * indirect refs with missing keys. An appropriate key is located and
 658  * the ref is moved onto the tree for indirect refs. After all missing
 659  * keys are thus located, we iterate over the indirect ref tree, resolve
 660  * each reference, and then insert the resolved reference onto the
 661  * direct tree (merging there too).
 662  *
 663  * New backrefs (i.e., for parent nodes) are added to the appropriate
 664  * rbtree as they are encountered. The new backrefs are subsequently
 665  * resolved as above.
 666  */
 667 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 668                                  struct btrfs_path *path, u64 time_seq,
 669                                  struct preftrees *preftrees,
 670                                  const u64 *extent_item_pos,
 671                                  struct share_check *sc, bool ignore_offset)
 672 {
 673         int err;
 674         int ret = 0;
 675         struct ulist *parents;
 676         struct ulist_node *node;
 677         struct ulist_iterator uiter;
 678         struct rb_node *rnode;
 679
 680         parents = ulist_alloc(GFP_NOFS);
 681         if (!parents)
 682                 return -ENOMEM;
 683
 684         /*
 685          * We could trade memory usage for performance here by iterating
 686          * the tree, allocating new refs for each insertion, and then
 687          * freeing the entire indirect tree when we're done.  In some test
 688          * cases, the tree can grow quite large (~200k objects).
 689          */
 690         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 691                 struct prelim_ref *ref;
 692
 693                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 694                 if (WARN(ref->parent,
 695                          "BUG: direct ref found in indirect tree")) {
 696                         ret = -EINVAL;
 697                         goto out;
 698                 }
 699
 700                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 701                 preftrees->indirect.count--;
 702
 703                 if (ref->count == 0) {
 704                         free_pref(ref);
 705                         continue;
 706                 }
 707
 708                 if (sc && sc->root_objectid &&
 709                     ref->root_id != sc->root_objectid) {
 710                         free_pref(ref);
 711                         ret = BACKREF_FOUND_SHARED;
 712                         goto out;
 713                 }
 714                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 715                                            ref, parents, extent_item_pos,
 716                                            ignore_offset);
 717                 /*
 718                  * we can only tolerate ENOENT,otherwise,we should catch error
 719                  * and return directly.
 720                  */
 721                 if (err == -ENOENT) {
 722                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 723                                           NULL);
 724                         continue;
 725                 } else if (err) {
 726                         free_pref(ref);
 727                         ret = err;
 728                         goto out;
 729                 }
 730
 731                 /* we put the first parent into the ref at hand */
 732                 ULIST_ITER_INIT(&uiter);
 733                 node = ulist_next(parents, &uiter);
 734                 ref->parent = node ? node->val : 0;
 735                 ref->inode_list = unode_aux_to_inode_list(node);
 736
 737                 /* Add a prelim_ref(s) for any other parent(s). */
 738                 while ((node = ulist_next(parents, &uiter))) {
 739                         struct prelim_ref *new_ref;
 740
 741                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 742                                                    GFP_NOFS);
 743                         if (!new_ref) {
 744                                 free_pref(ref);
 745                                 ret = -ENOMEM;
 746                                 goto out;
 747                         }
 748                         memcpy(new_ref, ref, sizeof(*ref));
 749                         new_ref->parent = node->val;
 750                         new_ref->inode_list = unode_aux_to_inode_list(node);
 751                         prelim_ref_insert(fs_info, &preftrees->direct,
 752                                           new_ref, NULL);
 753                 }
 754
 755                 /*
 756                  * Now it's a direct ref, put it in the direct tree. We must
 757                  * do this last because the ref could be merged/freed here.
 758                  */
 759                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 760
 761                 ulist_reinit(parents);
 762                 cond_resched();
 763         }
 764 out:
 765         ulist_free(parents);
 766         return ret;
 767 }
 768
 769 /*
 770  * read tree blocks and add keys where required.
 771  */
 772 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 773                             struct preftrees *preftrees, bool lock)
 774 {
 775         struct prelim_ref *ref;
 776         struct extent_buffer *eb;
 777         struct preftree *tree = &preftrees->indirect_missing_keys;
 778         struct rb_node *node;
 779
 780         while ((node = rb_first_cached(&tree->root))) {
 781                 ref = rb_entry(node, struct prelim_ref, rbnode);
 782                 rb_erase_cached(node, &tree->root);
 783
 784                 BUG_ON(ref->parent);    /* should not be a direct ref */
 785                 BUG_ON(ref->key_for_search.type);
 786                 BUG_ON(!ref->wanted_disk_byte);
 787
 788                 eb = read_tree_block(fs_info, ref->wanted_disk_byte,
 789                                      ref->root_id, 0, ref->level - 1, NULL);
 790                 if (IS_ERR(eb)) {
 791                         free_pref(ref);
 792                         return PTR_ERR(eb);
 793                 } else if (!extent_buffer_uptodate(eb)) {
 794                         free_pref(ref);
 795                         free_extent_buffer(eb);
 796                         return -EIO;
 797                 }
 798                 if (lock)
 799                         btrfs_tree_read_lock(eb);
 800                 if (btrfs_header_level(eb) == 0)
 801                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 802                 else
 803                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 804                 if (lock)
 805                         btrfs_tree_read_unlock(eb);
 806                 free_extent_buffer(eb);
 807                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 808                 cond_resched();
 809         }
 810         return 0;
 811 }
 812
 813 /*
 814  * add all currently queued delayed refs from this head whose seq nr is
 815  * smaller or equal that seq to the list
 816  */
 817 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 818                             struct btrfs_delayed_ref_head *head, u64 seq,
 819                             struct preftrees *preftrees, struct share_check *sc)
 820 {
 821         struct btrfs_delayed_ref_node *node;
 822         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 823         struct btrfs_key key;
 824         struct btrfs_key tmp_op_key;
 825         struct rb_node *n;
 826         int count;
 827         int ret = 0;
 828
 829         if (extent_op && extent_op->update_key)
 830                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 831
 832         spin_lock(&head->lock);
 833         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 834                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 835                                 ref_node);
 836                 if (node->seq > seq)
 837                         continue;
 838
 839                 switch (node->action) {
 840                 case BTRFS_ADD_DELAYED_EXTENT:
 841                 case BTRFS_UPDATE_DELAYED_HEAD:
 842                         WARN_ON(1);
 843                         continue;
 844                 case BTRFS_ADD_DELAYED_REF:
 845                         count = node->ref_mod;
 846                         break;
 847                 case BTRFS_DROP_DELAYED_REF:
 848                         count = node->ref_mod * -1;
 849                         break;
 850                 default:
 851                         BUG();
 852                 }
 853                 switch (node->type) {
 854                 case BTRFS_TREE_BLOCK_REF_KEY: {
 855                         /* NORMAL INDIRECT METADATA backref */
 856                         struct btrfs_delayed_tree_ref *ref;
 857
 858                         ref = btrfs_delayed_node_to_tree_ref(node);
 859                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 860                                                &tmp_op_key, ref->level + 1,
 861                                                node->bytenr, count, sc,
 862                                                GFP_ATOMIC);
 863                         break;
 864                 }
 865                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 866                         /* SHARED DIRECT METADATA backref */
 867                         struct btrfs_delayed_tree_ref *ref;
 868
 869                         ref = btrfs_delayed_node_to_tree_ref(node);
 870
 871                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 872                                              ref->parent, node->bytenr, count,
 873                                              sc, GFP_ATOMIC);
 874                         break;
 875                 }
 876                 case BTRFS_EXTENT_DATA_REF_KEY: {
 877                         /* NORMAL INDIRECT DATA backref */
 878                         struct btrfs_delayed_data_ref *ref;
 879                         ref = btrfs_delayed_node_to_data_ref(node);
 880
 881                         key.objectid = ref->objectid;
 882                         key.type = BTRFS_EXTENT_DATA_KEY;
 883                         key.offset = ref->offset;
 884
 885                         /*
 886                          * If we have a share check context and a reference for
 887                          * another inode, we can't exit immediately. This is
 888                          * because even if this is a BTRFS_ADD_DELAYED_REF
 889                          * reference we may find next a BTRFS_DROP_DELAYED_REF
 890                          * which cancels out this ADD reference.
 891                          *
 892                          * If this is a DROP reference and there was no previous
 893                          * ADD reference, then we need to signal that when we
 894                          * process references from the extent tree (through
 895                          * add_inline_refs() and add_keyed_refs()), we should
 896                          * not exit early if we find a reference for another
 897                          * inode, because one of the delayed DROP references
 898                          * may cancel that reference in the extent tree.
 899                          */
 900                         if (sc && count < 0)
 901                                 sc->have_delayed_delete_refs = true;
 902
 903                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 904                                                &key, 0, node->bytenr, count, sc,
 905                                                GFP_ATOMIC);
 906                         break;
 907                 }
 908                 case BTRFS_SHARED_DATA_REF_KEY: {
 909                         /* SHARED DIRECT FULL backref */
 910                         struct btrfs_delayed_data_ref *ref;
 911
 912                         ref = btrfs_delayed_node_to_data_ref(node);
 913
 914                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 915                                              node->bytenr, count, sc,
 916                                              GFP_ATOMIC);
 917                         break;
 918                 }
 919                 default:
 920                         WARN_ON(1);
 921                 }
 922                 /*
 923                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 924                  * refs have been checked.
 925                  */
 926                 if (ret && (ret != BACKREF_FOUND_SHARED))
 927                         break;
 928         }
 929         if (!ret)
 930                 ret = extent_is_shared(sc);
 931
 932         spin_unlock(&head->lock);
 933         return ret;
 934 }
 935
 936 /*
 937  * add all inline backrefs for bytenr to the list
 938  *
 939  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 940  */
 941 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 942                            struct btrfs_path *path, u64 bytenr,
 943                            int *info_level, struct preftrees *preftrees,
 944                            struct share_check *sc)
 945 {
 946         int ret = 0;
 947         int slot;
 948         struct extent_buffer *leaf;
 949         struct btrfs_key key;
 950         struct btrfs_key found_key;
 951         unsigned long ptr;
 952         unsigned long end;
 953         struct btrfs_extent_item *ei;
 954         u64 flags;
 955         u64 item_size;
 956
 957         /*
 958          * enumerate all inline refs
 959          */
 960         leaf = path->nodes[0];
 961         slot = path->slots[0];
 962
 963         item_size = btrfs_item_size_nr(leaf, slot);
 964         BUG_ON(item_size < sizeof(*ei));
 965
 966         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 967         flags = btrfs_extent_flags(leaf, ei);
 968         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 969
 970         ptr = (unsigned long)(ei + 1);
 971         end = (unsigned long)ei + item_size;
 972
 973         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 974             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 975                 struct btrfs_tree_block_info *info;
 976
 977                 info = (struct btrfs_tree_block_info *)ptr;
 978                 *info_level = btrfs_tree_block_level(leaf, info);
 979                 ptr += sizeof(struct btrfs_tree_block_info);
 980                 BUG_ON(ptr > end);
 981         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 982                 *info_level = found_key.offset;
 983         } else {
 984                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 985         }
 986
 987         while (ptr < end) {
 988                 struct btrfs_extent_inline_ref *iref;
 989                 u64 offset;
 990                 int type;
 991
 992                 iref = (struct btrfs_extent_inline_ref *)ptr;
 993                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 994                                                         BTRFS_REF_TYPE_ANY);
 995                 if (type == BTRFS_REF_TYPE_INVALID)
 996                         return -EUCLEAN;
 997
 998                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 999
1000                 switch (type) {
1001                 case BTRFS_SHARED_BLOCK_REF_KEY:
1002                         ret = add_direct_ref(fs_info, preftrees,
1003                                              *info_level + 1, offset,
1004                                              bytenr, 1, NULL, GFP_NOFS);
1005                         break;
1006                 case BTRFS_SHARED_DATA_REF_KEY: {
1007                         struct btrfs_shared_data_ref *sdref;
1008                         int count;
1009
1010                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1011                         count = btrfs_shared_data_ref_count(leaf, sdref);
1012
1013                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
1014                                              bytenr, count, sc, GFP_NOFS);
1015                         break;
1016                 }
1017                 case BTRFS_TREE_BLOCK_REF_KEY:
1018                         ret = add_indirect_ref(fs_info, preftrees, offset,
1019                                                NULL, *info_level + 1,
1020                                                bytenr, 1, NULL, GFP_NOFS);
1021                         break;
1022                 case BTRFS_EXTENT_DATA_REF_KEY: {
1023                         struct btrfs_extent_data_ref *dref;
1024                         int count;
1025                         u64 root;
1026
1027                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1028                         count = btrfs_extent_data_ref_count(leaf, dref);
1029                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1030                                                                       dref);
1031                         key.type = BTRFS_EXTENT_DATA_KEY;
1032                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1033
1034                         if (sc && sc->inum && key.objectid != sc->inum &&
1035                             !sc->have_delayed_delete_refs) {
1036                                 ret = BACKREF_FOUND_SHARED;
1037                                 break;
1038                         }
1039
1040                         root = btrfs_extent_data_ref_root(leaf, dref);
1041
1042                         ret = add_indirect_ref(fs_info, preftrees, root,
1043                                                &key, 0, bytenr, count,
1044                                                sc, GFP_NOFS);
1045
1046                         break;
1047                 }
1048                 default:
1049                         WARN_ON(1);
1050                 }
1051                 if (ret)
1052                         return ret;
1053                 ptr += btrfs_extent_inline_ref_size(type);
1054         }
1055
1056         return 0;
1057 }
1058
1059 /*
1060  * add all non-inline backrefs for bytenr to the list
1061  *
1062  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1063  */
1064 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1065                           struct btrfs_path *path, u64 bytenr,
1066                           int info_level, struct preftrees *preftrees,
1067                           struct share_check *sc)
1068 {
1069         struct btrfs_root *extent_root = fs_info->extent_root;
1070         int ret;
1071         int slot;
1072         struct extent_buffer *leaf;
1073         struct btrfs_key key;
1074
1075         while (1) {
1076                 ret = btrfs_next_item(extent_root, path);
1077                 if (ret < 0)
1078                         break;
1079                 if (ret) {
1080                         ret = 0;
1081                         break;
1082                 }
1083
1084                 slot = path->slots[0];
1085                 leaf = path->nodes[0];
1086                 btrfs_item_key_to_cpu(leaf, &key, slot);
1087
1088                 if (key.objectid != bytenr)
1089                         break;
1090                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1091                         continue;
1092                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1093                         break;
1094
1095                 switch (key.type) {
1096                 case BTRFS_SHARED_BLOCK_REF_KEY:
1097                         /* SHARED DIRECT METADATA backref */
1098                         ret = add_direct_ref(fs_info, preftrees,
1099                                              info_level + 1, key.offset,
1100                                              bytenr, 1, NULL, GFP_NOFS);
1101                         break;
1102                 case BTRFS_SHARED_DATA_REF_KEY: {
1103                         /* SHARED DIRECT FULL backref */
1104                         struct btrfs_shared_data_ref *sdref;
1105                         int count;
1106
1107                         sdref = btrfs_item_ptr(leaf, slot,
1108                                               struct btrfs_shared_data_ref);
1109                         count = btrfs_shared_data_ref_count(leaf, sdref);
1110                         ret = add_direct_ref(fs_info, preftrees, 0,
1111                                              key.offset, bytenr, count,
1112                                              sc, GFP_NOFS);
1113                         break;
1114                 }
1115                 case BTRFS_TREE_BLOCK_REF_KEY:
1116                         /* NORMAL INDIRECT METADATA backref */
1117                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1118                                                NULL, info_level + 1, bytenr,
1119                                                1, NULL, GFP_NOFS);
1120                         break;
1121                 case BTRFS_EXTENT_DATA_REF_KEY: {
1122                         /* NORMAL INDIRECT DATA backref */
1123                         struct btrfs_extent_data_ref *dref;
1124                         int count;
1125                         u64 root;
1126
1127                         dref = btrfs_item_ptr(leaf, slot,
1128                                               struct btrfs_extent_data_ref);
1129                         count = btrfs_extent_data_ref_count(leaf, dref);
1130                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1131                                                                       dref);
1132                         key.type = BTRFS_EXTENT_DATA_KEY;
1133                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1134
1135                         if (sc && sc->inum && key.objectid != sc->inum &&
1136                             !sc->have_delayed_delete_refs) {
1137                                 ret = BACKREF_FOUND_SHARED;
1138                                 break;
1139                         }
1140
1141                         root = btrfs_extent_data_ref_root(leaf, dref);
1142                         ret = add_indirect_ref(fs_info, preftrees, root,
1143                                                &key, 0, bytenr, count,
1144                                                sc, GFP_NOFS);
1145                         break;
1146                 }
1147                 default:
1148                         WARN_ON(1);
1149                 }
1150                 if (ret)
1151                         return ret;
1152
1153         }
1154
1155         return ret;
1156 }
1157
1158 /*
1159  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1160  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1161  * indirect refs to their parent bytenr.
1162  * When roots are found, they're added to the roots list
1163  *
1164  * If time_seq is set to BTRFS_SEQ_LAST, it will not search delayed_refs, and
1165  * behave much like trans == NULL case, the difference only lies in it will not
1166  * commit root.
1167  * The special case is for qgroup to search roots in commit_transaction().
1168  *
1169  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1170  * shared extent is detected.
1171  *
1172  * Otherwise this returns 0 for success and <0 for an error.
1173  *
1174  * If ignore_offset is set to false, only extent refs whose offsets match
1175  * extent_item_pos are returned.  If true, every extent ref is returned
1176  * and extent_item_pos is ignored.
1177  *
1178  * FIXME some caching might speed things up
1179  */
1180 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1181                              struct btrfs_fs_info *fs_info, u64 bytenr,
1182                              u64 time_seq, struct ulist *refs,
1183                              struct ulist *roots, const u64 *extent_item_pos,
1184                              struct share_check *sc, bool ignore_offset)
1185 {
1186         struct btrfs_key key;
1187         struct btrfs_path *path;
1188         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1189         struct btrfs_delayed_ref_head *head;
1190         int info_level = 0;
1191         int ret;
1192         struct prelim_ref *ref;
1193         struct rb_node *node;
1194         struct extent_inode_elem *eie = NULL;
1195         struct preftrees preftrees = {
1196                 .direct = PREFTREE_INIT,
1197                 .indirect = PREFTREE_INIT,
1198                 .indirect_missing_keys = PREFTREE_INIT
1199         };
1200
1201         key.objectid = bytenr;
1202         key.offset = (u64)-1;
1203         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1204                 key.type = BTRFS_METADATA_ITEM_KEY;
1205         else
1206                 key.type = BTRFS_EXTENT_ITEM_KEY;
1207
1208         path = btrfs_alloc_path();
1209         if (!path)
1210                 return -ENOMEM;
1211         if (!trans) {
1212                 path->search_commit_root = 1;
1213                 path->skip_locking = 1;
1214         }
1215
1216         if (time_seq == BTRFS_SEQ_LAST)
1217                 path->skip_locking = 1;
1218
1219         /*
1220          * grab both a lock on the path and a lock on the delayed ref head.
1221          * We need both to get a consistent picture of how the refs look
1222          * at a specified point in time
1223          */
1224 again:
1225         head = NULL;
1226
1227         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1228         if (ret < 0)
1229                 goto out;
1230         if (ret == 0) {
1231                 /* This shouldn't happen, indicates a bug or fs corruption. */
1232                 ASSERT(ret != 0);
1233                 ret = -EUCLEAN;
1234                 goto out;
1235         }
1236
1237 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1238         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1239             time_seq != BTRFS_SEQ_LAST) {
1240 #else
1241         if (trans && time_seq != BTRFS_SEQ_LAST) {
1242 #endif
1243                 /*
1244                  * look if there are updates for this ref queued and lock the
1245                  * head
1246                  */
1247                 delayed_refs = &trans->transaction->delayed_refs;
1248                 spin_lock(&delayed_refs->lock);
1249                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1250                 if (head) {
1251                         if (!mutex_trylock(&head->mutex)) {
1252                                 refcount_inc(&head->refs);
1253                                 spin_unlock(&delayed_refs->lock);
1254
1255                                 btrfs_release_path(path);
1256
1257                                 /*
1258                                  * Mutex was contended, block until it's
1259                                  * released and try again
1260                                  */
1261                                 mutex_lock(&head->mutex);
1262                                 mutex_unlock(&head->mutex);
1263                                 btrfs_put_delayed_ref_head(head);
1264                                 goto again;
1265                         }
1266                         spin_unlock(&delayed_refs->lock);
1267                         ret = add_delayed_refs(fs_info, head, time_seq,
1268                                                &preftrees, sc);
1269                         mutex_unlock(&head->mutex);
1270                         if (ret)
1271                                 goto out;
1272                 } else {
1273                         spin_unlock(&delayed_refs->lock);
1274                 }
1275         }
1276
1277         if (path->slots[0]) {
1278                 struct extent_buffer *leaf;
1279                 int slot;
1280
1281                 path->slots[0]--;
1282                 leaf = path->nodes[0];
1283                 slot = path->slots[0];
1284                 btrfs_item_key_to_cpu(leaf, &key, slot);
1285                 if (key.objectid == bytenr &&
1286                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1287                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1288                         ret = add_inline_refs(fs_info, path, bytenr,
1289                                               &info_level, &preftrees, sc);
1290                         if (ret)
1291                                 goto out;
1292                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1293                                              &preftrees, sc);
1294                         if (ret)
1295                                 goto out;
1296                 }
1297         }
1298
1299         btrfs_release_path(path);
1300
1301         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1302         if (ret)
1303                 goto out;
1304
1305         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1306
1307         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1308                                     extent_item_pos, sc, ignore_offset);
1309         if (ret)
1310                 goto out;
1311
1312         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1313
1314         /*
1315          * This walks the tree of merged and resolved refs. Tree blocks are
1316          * read in as needed. Unique entries are added to the ulist, and
1317          * the list of found roots is updated.
1318          *
1319          * We release the entire tree in one go before returning.
1320          */
1321         node = rb_first_cached(&preftrees.direct.root);
1322         while (node) {
1323                 ref = rb_entry(node, struct prelim_ref, rbnode);
1324                 node = rb_next(&ref->rbnode);
1325                 /*
1326                  * ref->count < 0 can happen here if there are delayed
1327                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1328                  * prelim_ref_insert() relies on this when merging
1329                  * identical refs to keep the overall count correct.
1330                  * prelim_ref_insert() will merge only those refs
1331                  * which compare identically.  Any refs having
1332                  * e.g. different offsets would not be merged,
1333                  * and would retain their original ref->count < 0.
1334                  */
1335                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1336                         if (sc && sc->root_objectid &&
1337                             ref->root_id != sc->root_objectid) {
1338                                 ret = BACKREF_FOUND_SHARED;
1339                                 goto out;
1340                         }
1341
1342                         /* no parent == root of tree */
1343                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1344                         if (ret < 0)
1345                                 goto out;
1346                 }
1347                 if (ref->count && ref->parent) {
1348                         if (extent_item_pos && !ref->inode_list &&
1349                             ref->level == 0) {
1350                                 struct extent_buffer *eb;
1351
1352                                 eb = read_tree_block(fs_info, ref->parent, 0,
1353                                                      0, ref->level, NULL);
1354                                 if (IS_ERR(eb)) {
1355                                         ret = PTR_ERR(eb);
1356                                         goto out;
1357                                 } else if (!extent_buffer_uptodate(eb)) {
1358                                         free_extent_buffer(eb);
1359                                         ret = -EIO;
1360                                         goto out;
1361                                 }
1362
1363                                 if (!path->skip_locking)
1364                                         btrfs_tree_read_lock(eb);
1365                                 ret = find_extent_in_eb(eb, bytenr,
1366                                                         *extent_item_pos, &eie, ignore_offset);
1367                                 if (!path->skip_locking)
1368                                         btrfs_tree_read_unlock(eb);
1369                                 free_extent_buffer(eb);
1370                                 if (ret < 0)
1371                                         goto out;
1372                                 ref->inode_list = eie;
1373                         }
1374                         ret = ulist_add_merge_ptr(refs, ref->parent,
1375                                                   ref->inode_list,
1376                                                   (void **)&eie, GFP_NOFS);
1377                         if (ret < 0)
1378                                 goto out;
1379                         if (!ret && extent_item_pos) {
1380                                 /*
1381                                  * We've recorded that parent, so we must extend
1382                                  * its inode list here.
1383                                  *
1384                                  * However if there was corruption we may not
1385                                  * have found an eie, return an error in this
1386                                  * case.
1387                                  */
1388                                 ASSERT(eie);
1389                                 if (!eie) {
1390                                         ret = -EUCLEAN;
1391                                         goto out;
1392                                 }
1393                                 while (eie->next)
1394                                         eie = eie->next;
1395                                 eie->next = ref->inode_list;
1396                         }
1397                         eie = NULL;
1398                 }
1399                 cond_resched();
1400         }
1401
1402 out:
1403         btrfs_free_path(path);
1404
1405         prelim_release(&preftrees.direct);
1406         prelim_release(&preftrees.indirect);
1407         prelim_release(&preftrees.indirect_missing_keys);
1408
1409         if (ret < 0)
1410                 free_inode_elem_list(eie);
1411         return ret;
1412 }
1413
1414 static void free_leaf_list(struct ulist *blocks)
1415 {
1416         struct ulist_node *node = NULL;
1417         struct extent_inode_elem *eie;
1418         struct ulist_iterator uiter;
1419
1420         ULIST_ITER_INIT(&uiter);
1421         while ((node = ulist_next(blocks, &uiter))) {
1422                 if (!node->aux)
1423                         continue;
1424                 eie = unode_aux_to_inode_list(node);
1425                 free_inode_elem_list(eie);
1426                 node->aux = 0;
1427         }
1428
1429         ulist_free(blocks);
1430 }
1431
1432 /*
1433  * Finds all leafs with a reference to the specified combination of bytenr and
1434  * offset. key_list_head will point to a list of corresponding keys (caller must
1435  * free each list element). The leafs will be stored in the leafs ulist, which
1436  * must be freed with ulist_free.
1437  *
1438  * returns 0 on success, <0 on error
1439  */
1440 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1441                          struct btrfs_fs_info *fs_info, u64 bytenr,
1442                          u64 time_seq, struct ulist **leafs,
1443                          const u64 *extent_item_pos, bool ignore_offset)
1444 {
1445         int ret;
1446
1447         *leafs = ulist_alloc(GFP_NOFS);
1448         if (!*leafs)
1449                 return -ENOMEM;
1450
1451         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1452                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1453         if (ret < 0 && ret != -ENOENT) {
1454                 free_leaf_list(*leafs);
1455                 return ret;
1456         }
1457
1458         return 0;
1459 }
1460
1461 /*
1462  * walk all backrefs for a given extent to find all roots that reference this
1463  * extent. Walking a backref means finding all extents that reference this
1464  * extent and in turn walk the backrefs of those, too. Naturally this is a
1465  * recursive process, but here it is implemented in an iterative fashion: We
1466  * find all referencing extents for the extent in question and put them on a
1467  * list. In turn, we find all referencing extents for those, further appending
1468  * to the list. The way we iterate the list allows adding more elements after
1469  * the current while iterating. The process stops when we reach the end of the
1470  * list. Found roots are added to the roots list.
1471  *
1472  * returns 0 on success, < 0 on error.
1473  */
1474 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1475                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1476                                      u64 time_seq, struct ulist **roots,
1477                                      bool ignore_offset)
1478 {
1479         struct ulist *tmp;
1480         struct ulist_node *node = NULL;
1481         struct ulist_iterator uiter;
1482         int ret;
1483
1484         tmp = ulist_alloc(GFP_NOFS);
1485         if (!tmp)
1486                 return -ENOMEM;
1487         *roots = ulist_alloc(GFP_NOFS);
1488         if (!*roots) {
1489                 ulist_free(tmp);
1490                 return -ENOMEM;
1491         }
1492
1493         ULIST_ITER_INIT(&uiter);
1494         while (1) {
1495                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1496                                         tmp, *roots, NULL, NULL, ignore_offset);
1497                 if (ret < 0 && ret != -ENOENT) {
1498                         ulist_free(tmp);
1499                         ulist_free(*roots);
1500                         *roots = NULL;
1501                         return ret;
1502                 }
1503                 node = ulist_next(tmp, &uiter);
1504                 if (!node)
1505                         break;
1506                 bytenr = node->val;
1507                 cond_resched();
1508         }
1509
1510         ulist_free(tmp);
1511         return 0;
1512 }
1513
1514 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1515                          struct btrfs_fs_info *fs_info, u64 bytenr,
1516                          u64 time_seq, struct ulist **roots,
1517                          bool skip_commit_root_sem)
1518 {
1519         int ret;
1520
1521         if (!trans && !skip_commit_root_sem)
1522                 down_read(&fs_info->commit_root_sem);
1523         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1524                                         time_seq, roots, false);
1525         if (!trans && !skip_commit_root_sem)
1526                 up_read(&fs_info->commit_root_sem);
1527         return ret;
1528 }
1529
1530 /**
1531  * Check if an extent is shared or not
1532  *
1533  * @root:   root inode belongs to
1534  * @inum:   inode number of the inode whose extent we are checking
1535  * @bytenr: logical bytenr of the extent we are checking
1536  * @roots:  list of roots this extent is shared among
1537  * @tmp:    temporary list used for iteration
1538  *
1539  * btrfs_check_shared uses the backref walking code but will short
1540  * circuit as soon as it finds a root or inode that doesn't match the
1541  * one passed in. This provides a significant performance benefit for
1542  * callers (such as fiemap) which want to know whether the extent is
1543  * shared but do not need a ref count.
1544  *
1545  * This attempts to attach to the running transaction in order to account for
1546  * delayed refs, but continues on even when no running transaction exists.
1547  *
1548  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1549  */
1550 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1551                 struct ulist *roots, struct ulist *tmp)
1552 {
1553         struct btrfs_fs_info *fs_info = root->fs_info;
1554         struct btrfs_trans_handle *trans;
1555         struct ulist_iterator uiter;
1556         struct ulist_node *node;
1557         struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1558         int ret = 0;
1559         struct share_check shared = {
1560                 .root_objectid = root->root_key.objectid,
1561                 .inum = inum,
1562                 .share_count = 0,
1563                 .have_delayed_delete_refs = false,
1564         };
1565
1566         ulist_init(roots);
1567         ulist_init(tmp);
1568
1569         trans = btrfs_join_transaction_nostart(root);
1570         if (IS_ERR(trans)) {
1571                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1572                         ret = PTR_ERR(trans);
1573                         goto out;
1574                 }
1575                 trans = NULL;
1576                 down_read(&fs_info->commit_root_sem);
1577         } else {
1578                 btrfs_get_tree_mod_seq(fs_info, &elem);
1579         }
1580
1581         ULIST_ITER_INIT(&uiter);
1582         while (1) {
1583                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1584                                         roots, NULL, &shared, false);
1585                 if (ret == BACKREF_FOUND_SHARED) {
1586                         /* this is the only condition under which we return 1 */
1587                         ret = 1;
1588                         break;
1589                 }
1590                 if (ret < 0 && ret != -ENOENT)
1591                         break;
1592                 ret = 0;
1593                 node = ulist_next(tmp, &uiter);
1594                 if (!node)
1595                         break;
1596                 bytenr = node->val;
1597                 shared.share_count = 0;
1598                 shared.have_delayed_delete_refs = false;
1599                 cond_resched();
1600         }
1601
1602         if (trans) {
1603                 btrfs_put_tree_mod_seq(fs_info, &elem);
1604                 btrfs_end_transaction(trans);
1605         } else {
1606                 up_read(&fs_info->commit_root_sem);
1607         }
1608 out:
1609         ulist_release(roots);
1610         ulist_release(tmp);
1611         return ret;
1612 }
1613
1614 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1615                           u64 start_off, struct btrfs_path *path,
1616                           struct btrfs_inode_extref **ret_extref,
1617                           u64 *found_off)
1618 {
1619         int ret, slot;
1620         struct btrfs_key key;
1621         struct btrfs_key found_key;
1622         struct btrfs_inode_extref *extref;
1623         const struct extent_buffer *leaf;
1624         unsigned long ptr;
1625
1626         key.objectid = inode_objectid;
1627         key.type = BTRFS_INODE_EXTREF_KEY;
1628         key.offset = start_off;
1629
1630         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1631         if (ret < 0)
1632                 return ret;
1633
1634         while (1) {
1635                 leaf = path->nodes[0];
1636                 slot = path->slots[0];
1637                 if (slot >= btrfs_header_nritems(leaf)) {
1638                         /*
1639                          * If the item at offset is not found,
1640                          * btrfs_search_slot will point us to the slot
1641                          * where it should be inserted. In our case
1642                          * that will be the slot directly before the
1643                          * next INODE_REF_KEY_V2 item. In the case
1644                          * that we're pointing to the last slot in a
1645                          * leaf, we must move one leaf over.
1646                          */
1647                         ret = btrfs_next_leaf(root, path);
1648                         if (ret) {
1649                                 if (ret >= 1)
1650                                         ret = -ENOENT;
1651                                 break;
1652                         }
1653                         continue;
1654                 }
1655
1656                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1657
1658                 /*
1659                  * Check that we're still looking at an extended ref key for
1660                  * this particular objectid. If we have different
1661                  * objectid or type then there are no more to be found
1662                  * in the tree and we can exit.
1663                  */
1664                 ret = -ENOENT;
1665                 if (found_key.objectid != inode_objectid)
1666                         break;
1667                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1668                         break;
1669
1670                 ret = 0;
1671                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1672                 extref = (struct btrfs_inode_extref *)ptr;
1673                 *ret_extref = extref;
1674                 if (found_off)
1675                         *found_off = found_key.offset;
1676                 break;
1677         }
1678
1679         return ret;
1680 }
1681
1682 /*
1683  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1684  * Elements of the path are separated by '/' and the path is guaranteed to be
1685  * 0-terminated. the path is only given within the current file system.
1686  * Therefore, it never starts with a '/'. the caller is responsible to provide
1687  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1688  * the start point of the resulting string is returned. this pointer is within
1689  * dest, normally.
1690  * in case the path buffer would overflow, the pointer is decremented further
1691  * as if output was written to the buffer, though no more output is actually
1692  * generated. that way, the caller can determine how much space would be
1693  * required for the path to fit into the buffer. in that case, the returned
1694  * value will be smaller than dest. callers must check this!
1695  */
1696 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1697                         u32 name_len, unsigned long name_off,
1698                         struct extent_buffer *eb_in, u64 parent,
1699                         char *dest, u32 size)
1700 {
1701         int slot;
1702         u64 next_inum;
1703         int ret;
1704         s64 bytes_left = ((s64)size) - 1;
1705         struct extent_buffer *eb = eb_in;
1706         struct btrfs_key found_key;
1707         struct btrfs_inode_ref *iref;
1708
1709         if (bytes_left >= 0)
1710                 dest[bytes_left] = '\0';
1711
1712         while (1) {
1713                 bytes_left -= name_len;
1714                 if (bytes_left >= 0)
1715                         read_extent_buffer(eb, dest + bytes_left,
1716                                            name_off, name_len);
1717                 if (eb != eb_in) {
1718                         if (!path->skip_locking)
1719                                 btrfs_tree_read_unlock(eb);
1720                         free_extent_buffer(eb);
1721                 }
1722                 ret = btrfs_find_item(fs_root, path, parent, 0,
1723                                 BTRFS_INODE_REF_KEY, &found_key);
1724                 if (ret > 0)
1725                         ret = -ENOENT;
1726                 if (ret)
1727                         break;
1728
1729                 next_inum = found_key.offset;
1730
1731                 /* regular exit ahead */
1732                 if (parent == next_inum)
1733                         break;
1734
1735                 slot = path->slots[0];
1736                 eb = path->nodes[0];
1737                 /* make sure we can use eb after releasing the path */
1738                 if (eb != eb_in) {
1739                         path->nodes[0] = NULL;
1740                         path->locks[0] = 0;
1741                 }
1742                 btrfs_release_path(path);
1743                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1744
1745                 name_len = btrfs_inode_ref_name_len(eb, iref);
1746                 name_off = (unsigned long)(iref + 1);
1747
1748                 parent = next_inum;
1749                 --bytes_left;
1750                 if (bytes_left >= 0)
1751                         dest[bytes_left] = '/';
1752         }
1753
1754         btrfs_release_path(path);
1755
1756         if (ret)
1757                 return ERR_PTR(ret);
1758
1759         return dest + bytes_left;
1760 }
1761
1762 /*
1763  * this makes the path point to (logical EXTENT_ITEM *)
1764  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1765  * tree blocks and <0 on error.
1766  */
1767 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1768                         struct btrfs_path *path, struct btrfs_key *found_key,
1769                         u64 *flags_ret)
1770 {
1771         int ret;
1772         u64 flags;
1773         u64 size = 0;
1774         u32 item_size;
1775         const struct extent_buffer *eb;
1776         struct btrfs_extent_item *ei;
1777         struct btrfs_key key;
1778
1779         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1780                 key.type = BTRFS_METADATA_ITEM_KEY;
1781         else
1782                 key.type = BTRFS_EXTENT_ITEM_KEY;
1783         key.objectid = logical;
1784         key.offset = (u64)-1;
1785
1786         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1787         if (ret < 0)
1788                 return ret;
1789
1790         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1791         if (ret) {
1792                 if (ret > 0)
1793                         ret = -ENOENT;
1794                 return ret;
1795         }
1796         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1797         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1798                 size = fs_info->nodesize;
1799         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1800                 size = found_key->offset;
1801
1802         if (found_key->objectid > logical ||
1803             found_key->objectid + size <= logical) {
1804                 btrfs_debug(fs_info,
1805                         "logical %llu is not within any extent", logical);
1806                 return -ENOENT;
1807         }
1808
1809         eb = path->nodes[0];
1810         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1811         BUG_ON(item_size < sizeof(*ei));
1812
1813         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1814         flags = btrfs_extent_flags(eb, ei);
1815
1816         btrfs_debug(fs_info,
1817                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1818                  logical, logical - found_key->objectid, found_key->objectid,
1819                  found_key->offset, flags, item_size);
1820
1821         WARN_ON(!flags_ret);
1822         if (flags_ret) {
1823                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1824                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1825                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1826                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1827                 else
1828                         BUG();
1829                 return 0;
1830         }
1831
1832         return -EIO;
1833 }
1834
1835 /*
1836  * helper function to iterate extent inline refs. ptr must point to a 0 value
1837  * for the first call and may be modified. it is used to track state.
1838  * if more refs exist, 0 is returned and the next call to
1839  * get_extent_inline_ref must pass the modified ptr parameter to get the
1840  * next ref. after the last ref was processed, 1 is returned.
1841  * returns <0 on error
1842  */
1843 static int get_extent_inline_ref(unsigned long *ptr,
1844                                  const struct extent_buffer *eb,
1845                                  const struct btrfs_key *key,
1846                                  const struct btrfs_extent_item *ei,
1847                                  u32 item_size,
1848                                  struct btrfs_extent_inline_ref **out_eiref,
1849                                  int *out_type)
1850 {
1851         unsigned long end;
1852         u64 flags;
1853         struct btrfs_tree_block_info *info;
1854
1855         if (!*ptr) {
1856                 /* first call */
1857                 flags = btrfs_extent_flags(eb, ei);
1858                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1859                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1860                                 /* a skinny metadata extent */
1861                                 *out_eiref =
1862                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1863                         } else {
1864                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1865                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1866                                 *out_eiref =
1867                                    (struct btrfs_extent_inline_ref *)(info + 1);
1868                         }
1869                 } else {
1870                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1871                 }
1872                 *ptr = (unsigned long)*out_eiref;
1873                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1874                         return -ENOENT;
1875         }
1876
1877         end = (unsigned long)ei + item_size;
1878         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1879         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1880                                                      BTRFS_REF_TYPE_ANY);
1881         if (*out_type == BTRFS_REF_TYPE_INVALID)
1882                 return -EUCLEAN;
1883
1884         *ptr += btrfs_extent_inline_ref_size(*out_type);
1885         WARN_ON(*ptr > end);
1886         if (*ptr == end)
1887                 return 1; /* last */
1888
1889         return 0;
1890 }
1891
1892 /*
1893  * reads the tree block backref for an extent. tree level and root are returned
1894  * through out_level and out_root. ptr must point to a 0 value for the first
1895  * call and may be modified (see get_extent_inline_ref comment).
1896  * returns 0 if data was provided, 1 if there was no more data to provide or
1897  * <0 on error.
1898  */
1899 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1900                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1901                             u32 item_size, u64 *out_root, u8 *out_level)
1902 {
1903         int ret;
1904         int type;
1905         struct btrfs_extent_inline_ref *eiref;
1906
1907         if (*ptr == (unsigned long)-1)
1908                 return 1;
1909
1910         while (1) {
1911                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1912                                               &eiref, &type);
1913                 if (ret < 0)
1914                         return ret;
1915
1916                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1917                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1918                         break;
1919
1920                 if (ret == 1)
1921                         return 1;
1922         }
1923
1924         /* we can treat both ref types equally here */
1925         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1926
1927         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1928                 struct btrfs_tree_block_info *info;
1929
1930                 info = (struct btrfs_tree_block_info *)(ei + 1);
1931                 *out_level = btrfs_tree_block_level(eb, info);
1932         } else {
1933                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1934                 *out_level = (u8)key->offset;
1935         }
1936
1937         if (ret == 1)
1938                 *ptr = (unsigned long)-1;
1939
1940         return 0;
1941 }
1942
1943 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1944                              struct extent_inode_elem *inode_list,
1945                              u64 root, u64 extent_item_objectid,
1946                              iterate_extent_inodes_t *iterate, void *ctx)
1947 {
1948         struct extent_inode_elem *eie;
1949         int ret = 0;
1950
1951         for (eie = inode_list; eie; eie = eie->next) {
1952                 btrfs_debug(fs_info,
1953                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1954                             extent_item_objectid, eie->inum,
1955                             eie->offset, root);
1956                 ret = iterate(eie->inum, eie->offset, root, ctx);
1957                 if (ret) {
1958                         btrfs_debug(fs_info,
1959                                     "stopping iteration for %llu due to ret=%d",
1960                                     extent_item_objectid, ret);
1961                         break;
1962                 }
1963         }
1964
1965         return ret;
1966 }
1967
1968 /*
1969  * calls iterate() for every inode that references the extent identified by
1970  * the given parameters.
1971  * when the iterator function returns a non-zero value, iteration stops.
1972  */
1973 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1974                                 u64 extent_item_objectid, u64 extent_item_pos,
1975                                 int search_commit_root,
1976                                 iterate_extent_inodes_t *iterate, void *ctx,
1977                                 bool ignore_offset)
1978 {
1979         int ret;
1980         struct btrfs_trans_handle *trans = NULL;
1981         struct ulist *refs = NULL;
1982         struct ulist *roots = NULL;
1983         struct ulist_node *ref_node = NULL;
1984         struct ulist_node *root_node = NULL;
1985         struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
1986         struct ulist_iterator ref_uiter;
1987         struct ulist_iterator root_uiter;
1988
1989         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1990                         extent_item_objectid);
1991
1992         if (!search_commit_root) {
1993                 trans = btrfs_attach_transaction(fs_info->extent_root);
1994                 if (IS_ERR(trans)) {
1995                         if (PTR_ERR(trans) != -ENOENT &&
1996                             PTR_ERR(trans) != -EROFS)
1997                                 return PTR_ERR(trans);
1998                         trans = NULL;
1999                 }
2000         }
2001
2002         if (trans)
2003                 btrfs_get_tree_mod_seq(fs_info, &seq_elem);
2004         else
2005                 down_read(&fs_info->commit_root_sem);
2006
2007         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2008                                    seq_elem.seq, &refs,
2009                                    &extent_item_pos, ignore_offset);
2010         if (ret)
2011                 goto out;
2012
2013         ULIST_ITER_INIT(&ref_uiter);
2014         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2015                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
2016                                                 seq_elem.seq, &roots,
2017                                                 ignore_offset);
2018                 if (ret)
2019                         break;
2020                 ULIST_ITER_INIT(&root_uiter);
2021                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2022                         btrfs_debug(fs_info,
2023                                     "root %llu references leaf %llu, data list %#llx",
2024                                     root_node->val, ref_node->val,
2025                                     ref_node->aux);
2026                         ret = iterate_leaf_refs(fs_info,
2027                                                 (struct extent_inode_elem *)
2028                                                 (uintptr_t)ref_node->aux,
2029                                                 root_node->val,
2030                                                 extent_item_objectid,
2031                                                 iterate, ctx);
2032                 }
2033                 ulist_free(roots);
2034         }
2035
2036         free_leaf_list(refs);
2037 out:
2038         if (trans) {
2039                 btrfs_put_tree_mod_seq(fs_info, &seq_elem);
2040                 btrfs_end_transaction(trans);
2041         } else {
2042                 up_read(&fs_info->commit_root_sem);
2043         }
2044
2045         return ret;
2046 }
2047
2048 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2049                                 struct btrfs_path *path,
2050                                 iterate_extent_inodes_t *iterate, void *ctx,
2051                                 bool ignore_offset)
2052 {
2053         int ret;
2054         u64 extent_item_pos;
2055         u64 flags = 0;
2056         struct btrfs_key found_key;
2057         int search_commit_root = path->search_commit_root;
2058
2059         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2060         btrfs_release_path(path);
2061         if (ret < 0)
2062                 return ret;
2063         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2064                 return -EINVAL;
2065
2066         extent_item_pos = logical - found_key.objectid;
2067         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2068                                         extent_item_pos, search_commit_root,
2069                                         iterate, ctx, ignore_offset);
2070
2071         return ret;
2072 }
2073
2074 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2075                               struct extent_buffer *eb, void *ctx);
2076
2077 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2078                               struct btrfs_path *path,
2079                               iterate_irefs_t *iterate, void *ctx)
2080 {
2081         int ret = 0;
2082         int slot;
2083         u32 cur;
2084         u32 len;
2085         u32 name_len;
2086         u64 parent = 0;
2087         int found = 0;
2088         struct extent_buffer *eb;
2089         struct btrfs_item *item;
2090         struct btrfs_inode_ref *iref;
2091         struct btrfs_key found_key;
2092
2093         while (!ret) {
2094                 ret = btrfs_find_item(fs_root, path, inum,
2095                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2096                                 &found_key);
2097
2098                 if (ret < 0)
2099                         break;
2100                 if (ret) {
2101                         ret = found ? 0 : -ENOENT;
2102                         break;
2103                 }
2104                 ++found;
2105
2106                 parent = found_key.offset;
2107                 slot = path->slots[0];
2108                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2109                 if (!eb) {
2110                         ret = -ENOMEM;
2111                         break;
2112                 }
2113                 btrfs_release_path(path);
2114
2115                 item = btrfs_item_nr(slot);
2116                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2117
2118                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2119                         name_len = btrfs_inode_ref_name_len(eb, iref);
2120                         /* path must be released before calling iterate()! */
2121                         btrfs_debug(fs_root->fs_info,
2122                                 "following ref at offset %u for inode %llu in tree %llu",
2123                                 cur, found_key.objectid,
2124                                 fs_root->root_key.objectid);
2125                         ret = iterate(parent, name_len,
2126                                       (unsigned long)(iref + 1), eb, ctx);
2127                         if (ret)
2128                                 break;
2129                         len = sizeof(*iref) + name_len;
2130                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2131                 }
2132                 free_extent_buffer(eb);
2133         }
2134
2135         btrfs_release_path(path);
2136
2137         return ret;
2138 }
2139
2140 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2141                                  struct btrfs_path *path,
2142                                  iterate_irefs_t *iterate, void *ctx)
2143 {
2144         int ret;
2145         int slot;
2146         u64 offset = 0;
2147         u64 parent;
2148         int found = 0;
2149         struct extent_buffer *eb;
2150         struct btrfs_inode_extref *extref;
2151         u32 item_size;
2152         u32 cur_offset;
2153         unsigned long ptr;
2154
2155         while (1) {
2156                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2157                                             &offset);
2158                 if (ret < 0)
2159                         break;
2160                 if (ret) {
2161                         ret = found ? 0 : -ENOENT;
2162                         break;
2163                 }
2164                 ++found;
2165
2166                 slot = path->slots[0];
2167                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2168                 if (!eb) {
2169                         ret = -ENOMEM;
2170                         break;
2171                 }
2172                 btrfs_release_path(path);
2173
2174                 item_size = btrfs_item_size_nr(eb, slot);
2175                 ptr = btrfs_item_ptr_offset(eb, slot);
2176                 cur_offset = 0;
2177
2178                 while (cur_offset < item_size) {
2179                         u32 name_len;
2180
2181                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2182                         parent = btrfs_inode_extref_parent(eb, extref);
2183                         name_len = btrfs_inode_extref_name_len(eb, extref);
2184                         ret = iterate(parent, name_len,
2185                                       (unsigned long)&extref->name, eb, ctx);
2186                         if (ret)
2187                                 break;
2188
2189                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2190                         cur_offset += sizeof(*extref);
2191                 }
2192                 free_extent_buffer(eb);
2193
2194                 offset++;
2195         }
2196
2197         btrfs_release_path(path);
2198
2199         return ret;
2200 }
2201
2202 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2203                          struct btrfs_path *path, iterate_irefs_t *iterate,
2204                          void *ctx)
2205 {
2206         int ret;
2207         int found_refs = 0;
2208
2209         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2210         if (!ret)
2211                 ++found_refs;
2212         else if (ret != -ENOENT)
2213                 return ret;
2214
2215         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2216         if (ret == -ENOENT && found_refs)
2217                 return 0;
2218
2219         return ret;
2220 }
2221
2222 /*
2223  * returns 0 if the path could be dumped (probably truncated)
2224  * returns <0 in case of an error
2225  */
2226 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2227                          struct extent_buffer *eb, void *ctx)
2228 {
2229         struct inode_fs_paths *ipath = ctx;
2230         char *fspath;
2231         char *fspath_min;
2232         int i = ipath->fspath->elem_cnt;
2233         const int s_ptr = sizeof(char *);
2234         u32 bytes_left;
2235
2236         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2237                                         ipath->fspath->bytes_left - s_ptr : 0;
2238
2239         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2240         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2241                                    name_off, eb, inum, fspath_min, bytes_left);
2242         if (IS_ERR(fspath))
2243                 return PTR_ERR(fspath);
2244
2245         if (fspath > fspath_min) {
2246                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2247                 ++ipath->fspath->elem_cnt;
2248                 ipath->fspath->bytes_left = fspath - fspath_min;
2249         } else {
2250                 ++ipath->fspath->elem_missed;
2251                 ipath->fspath->bytes_missing += fspath_min - fspath;
2252                 ipath->fspath->bytes_left = 0;
2253         }
2254
2255         return 0;
2256 }
2257
2258 /*
2259  * this dumps all file system paths to the inode into the ipath struct, provided
2260  * is has been created large enough. each path is zero-terminated and accessed
2261  * from ipath->fspath->val[i].
2262  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2263  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2264  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2265  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2266  * have been needed to return all paths.
2267  */
2268 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2269 {
2270         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2271                              inode_to_path, ipath);
2272 }
2273
2274 struct btrfs_data_container *init_data_container(u32 total_bytes)
2275 {
2276         struct btrfs_data_container *data;
2277         size_t alloc_bytes;
2278
2279         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2280         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2281         if (!data)
2282                 return ERR_PTR(-ENOMEM);
2283
2284         if (total_bytes >= sizeof(*data)) {
2285                 data->bytes_left = total_bytes - sizeof(*data);
2286                 data->bytes_missing = 0;
2287         } else {
2288                 data->bytes_missing = sizeof(*data) - total_bytes;
2289                 data->bytes_left = 0;
2290         }
2291
2292         data->elem_cnt = 0;
2293         data->elem_missed = 0;
2294
2295         return data;
2296 }
2297
2298 /*
2299  * allocates space to return multiple file system paths for an inode.
2300  * total_bytes to allocate are passed, note that space usable for actual path
2301  * information will be total_bytes - sizeof(struct inode_fs_paths).
2302  * the returned pointer must be freed with free_ipath() in the end.
2303  */
2304 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2305                                         struct btrfs_path *path)
2306 {
2307         struct inode_fs_paths *ifp;
2308         struct btrfs_data_container *fspath;
2309
2310         fspath = init_data_container(total_bytes);
2311         if (IS_ERR(fspath))
2312                 return ERR_CAST(fspath);
2313
2314         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2315         if (!ifp) {
2316                 kvfree(fspath);
2317                 return ERR_PTR(-ENOMEM);
2318         }
2319
2320         ifp->btrfs_path = path;
2321         ifp->fspath = fspath;
2322         ifp->fs_root = fs_root;
2323
2324         return ifp;
2325 }
2326
2327 void free_ipath(struct inode_fs_paths *ipath)
2328 {
2329         if (!ipath)
2330                 return;
2331         kvfree(ipath->fspath);
2332         kfree(ipath);
2333 }
2334
2335 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2336                 struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2337 {
2338         struct btrfs_backref_iter *ret;
2339
2340         ret = kzalloc(sizeof(*ret), gfp_flag);
2341         if (!ret)
2342                 return NULL;
2343
2344         ret->path = btrfs_alloc_path();
2345         if (!ret->path) {
2346                 kfree(ret);
2347                 return NULL;
2348         }
2349
2350         /* Current backref iterator only supports iteration in commit root */
2351         ret->path->search_commit_root = 1;
2352         ret->path->skip_locking = 1;
2353         ret->fs_info = fs_info;
2354
2355         return ret;
2356 }
2357
2358 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2359 {
2360         struct btrfs_fs_info *fs_info = iter->fs_info;
2361         struct btrfs_path *path = iter->path;
2362         struct btrfs_extent_item *ei;
2363         struct btrfs_key key;
2364         int ret;
2365
2366         key.objectid = bytenr;
2367         key.type = BTRFS_METADATA_ITEM_KEY;
2368         key.offset = (u64)-1;
2369         iter->bytenr = bytenr;
2370
2371         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
2372         if (ret < 0)
2373                 return ret;
2374         if (ret == 0) {
2375                 ret = -EUCLEAN;
2376                 goto release;
2377         }
2378         if (path->slots[0] == 0) {
2379                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2380                 ret = -EUCLEAN;
2381                 goto release;
2382         }
2383         path->slots[0]--;
2384
2385         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2386         if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2387              key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2388                 ret = -ENOENT;
2389                 goto release;
2390         }
2391         memcpy(&iter->cur_key, &key, sizeof(key));
2392         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2393                                                     path->slots[0]);
2394         iter->end_ptr = (u32)(iter->item_ptr +
2395                         btrfs_item_size_nr(path->nodes[0], path->slots[0]));
2396         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2397                             struct btrfs_extent_item);
2398
2399         /*
2400          * Only support iteration on tree backref yet.
2401          *
2402          * This is an extra precaution for non skinny-metadata, where
2403          * EXTENT_ITEM is also used for tree blocks, that we can only use
2404          * extent flags to determine if it's a tree block.
2405          */
2406         if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2407                 ret = -ENOTSUPP;
2408                 goto release;
2409         }
2410         iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2411
2412         /* If there is no inline backref, go search for keyed backref */
2413         if (iter->cur_ptr >= iter->end_ptr) {
2414                 ret = btrfs_next_item(fs_info->extent_root, path);
2415
2416                 /* No inline nor keyed ref */
2417                 if (ret > 0) {
2418                         ret = -ENOENT;
2419                         goto release;
2420                 }
2421                 if (ret < 0)
2422                         goto release;
2423
2424                 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2425                                 path->slots[0]);
2426                 if (iter->cur_key.objectid != bytenr ||
2427                     (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2428                      iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2429                         ret = -ENOENT;
2430                         goto release;
2431                 }
2432                 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2433                                                            path->slots[0]);
2434                 iter->item_ptr = iter->cur_ptr;
2435                 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size_nr(
2436                                       path->nodes[0], path->slots[0]));
2437         }
2438
2439         return 0;
2440 release:
2441         btrfs_backref_iter_release(iter);
2442         return ret;
2443 }
2444
2445 /*
2446  * Go to the next backref item of current bytenr, can be either inlined or
2447  * keyed.
2448  *
2449  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2450  *
2451  * Return 0 if we get next backref without problem.
2452  * Return >0 if there is no extra backref for this bytenr.
2453  * Return <0 if there is something wrong happened.
2454  */
2455 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2456 {
2457         struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2458         struct btrfs_path *path = iter->path;
2459         struct btrfs_extent_inline_ref *iref;
2460         int ret;
2461         u32 size;
2462
2463         if (btrfs_backref_iter_is_inline_ref(iter)) {
2464                 /* We're still inside the inline refs */
2465                 ASSERT(iter->cur_ptr < iter->end_ptr);
2466
2467                 if (btrfs_backref_has_tree_block_info(iter)) {
2468                         /* First tree block info */
2469                         size = sizeof(struct btrfs_tree_block_info);
2470                 } else {
2471                         /* Use inline ref type to determine the size */
2472                         int type;
2473
2474                         iref = (struct btrfs_extent_inline_ref *)
2475                                 ((unsigned long)iter->cur_ptr);
2476                         type = btrfs_extent_inline_ref_type(eb, iref);
2477
2478                         size = btrfs_extent_inline_ref_size(type);
2479                 }
2480                 iter->cur_ptr += size;
2481                 if (iter->cur_ptr < iter->end_ptr)
2482                         return 0;
2483
2484                 /* All inline items iterated, fall through */
2485         }
2486
2487         /* We're at keyed items, there is no inline item, go to the next one */
2488         ret = btrfs_next_item(iter->fs_info->extent_root, iter->path);
2489         if (ret)
2490                 return ret;
2491
2492         btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2493         if (iter->cur_key.objectid != iter->bytenr ||
2494             (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2495              iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2496                 return 1;
2497         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2498                                         path->slots[0]);
2499         iter->cur_ptr = iter->item_ptr;
2500         iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size_nr(path->nodes[0],
2501                                                 path->slots[0]);
2502         return 0;
2503 }
2504
2505 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2506                               struct btrfs_backref_cache *cache, int is_reloc)
2507 {
2508         int i;
2509
2510         cache->rb_root = RB_ROOT;
2511         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2512                 INIT_LIST_HEAD(&cache->pending[i]);
2513         INIT_LIST_HEAD(&cache->changed);
2514         INIT_LIST_HEAD(&cache->detached);
2515         INIT_LIST_HEAD(&cache->leaves);
2516         INIT_LIST_HEAD(&cache->pending_edge);
2517         INIT_LIST_HEAD(&cache->useless_node);
2518         cache->fs_info = fs_info;
2519         cache->is_reloc = is_reloc;
2520 }
2521
2522 struct btrfs_backref_node *btrfs_backref_alloc_node(
2523                 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2524 {
2525         struct btrfs_backref_node *node;
2526
2527         ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2528         node = kzalloc(sizeof(*node), GFP_NOFS);
2529         if (!node)
2530                 return node;
2531
2532         INIT_LIST_HEAD(&node->list);
2533         INIT_LIST_HEAD(&node->upper);
2534         INIT_LIST_HEAD(&node->lower);
2535         RB_CLEAR_NODE(&node->rb_node);
2536         cache->nr_nodes++;
2537         node->level = level;
2538         node->bytenr = bytenr;
2539
2540         return node;
2541 }
2542
2543 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2544                 struct btrfs_backref_cache *cache)
2545 {
2546         struct btrfs_backref_edge *edge;
2547
2548         edge = kzalloc(sizeof(*edge), GFP_NOFS);
2549         if (edge)
2550                 cache->nr_edges++;
2551         return edge;
2552 }
2553
2554 /*
2555  * Drop the backref node from cache, also cleaning up all its
2556  * upper edges and any uncached nodes in the path.
2557  *
2558  * This cleanup happens bottom up, thus the node should either
2559  * be the lowest node in the cache or a detached node.
2560  */
2561 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2562                                 struct btrfs_backref_node *node)
2563 {
2564         struct btrfs_backref_node *upper;
2565         struct btrfs_backref_edge *edge;
2566
2567         if (!node)
2568                 return;
2569
2570         BUG_ON(!node->lowest && !node->detached);
2571         while (!list_empty(&node->upper)) {
2572                 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2573                                   list[LOWER]);
2574                 upper = edge->node[UPPER];
2575                 list_del(&edge->list[LOWER]);
2576                 list_del(&edge->list[UPPER]);
2577                 btrfs_backref_free_edge(cache, edge);
2578
2579                 /*
2580                  * Add the node to leaf node list if no other child block
2581                  * cached.
2582                  */
2583                 if (list_empty(&upper->lower)) {
2584                         list_add_tail(&upper->lower, &cache->leaves);
2585                         upper->lowest = 1;
2586                 }
2587         }
2588
2589         btrfs_backref_drop_node(cache, node);
2590 }
2591
2592 /*
2593  * Release all nodes/edges from current cache
2594  */
2595 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2596 {
2597         struct btrfs_backref_node *node;
2598         int i;
2599
2600         while (!list_empty(&cache->detached)) {
2601                 node = list_entry(cache->detached.next,
2602                                   struct btrfs_backref_node, list);
2603                 btrfs_backref_cleanup_node(cache, node);
2604         }
2605
2606         while (!list_empty(&cache->leaves)) {
2607                 node = list_entry(cache->leaves.next,
2608                                   struct btrfs_backref_node, lower);
2609                 btrfs_backref_cleanup_node(cache, node);
2610         }
2611
2612         cache->last_trans = 0;
2613
2614         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2615                 ASSERT(list_empty(&cache->pending[i]));
2616         ASSERT(list_empty(&cache->pending_edge));
2617         ASSERT(list_empty(&cache->useless_node));
2618         ASSERT(list_empty(&cache->changed));
2619         ASSERT(list_empty(&cache->detached));
2620         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2621         ASSERT(!cache->nr_nodes);
2622         ASSERT(!cache->nr_edges);
2623 }
2624
2625 /*
2626  * Handle direct tree backref
2627  *
2628  * Direct tree backref means, the backref item shows its parent bytenr
2629  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2630  *
2631  * @ref_key:    The converted backref key.
2632  *              For keyed backref, it's the item key.
2633  *              For inlined backref, objectid is the bytenr,
2634  *              type is btrfs_inline_ref_type, offset is
2635  *              btrfs_inline_ref_offset.
2636  */
2637 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2638                                       struct btrfs_key *ref_key,
2639                                       struct btrfs_backref_node *cur)
2640 {
2641         struct btrfs_backref_edge *edge;
2642         struct btrfs_backref_node *upper;
2643         struct rb_node *rb_node;
2644
2645         ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2646
2647         /* Only reloc root uses backref pointing to itself */
2648         if (ref_key->objectid == ref_key->offset) {
2649                 struct btrfs_root *root;
2650
2651                 cur->is_reloc_root = 1;
2652                 /* Only reloc backref cache cares about a specific root */
2653                 if (cache->is_reloc) {
2654                         root = find_reloc_root(cache->fs_info, cur->bytenr);
2655                         if (!root)
2656                                 return -ENOENT;
2657                         cur->root = root;
2658                 } else {
2659                         /*
2660                          * For generic purpose backref cache, reloc root node
2661                          * is useless.
2662                          */
2663                         list_add(&cur->list, &cache->useless_node);
2664                 }
2665                 return 0;
2666         }
2667
2668         edge = btrfs_backref_alloc_edge(cache);
2669         if (!edge)
2670                 return -ENOMEM;
2671
2672         rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2673         if (!rb_node) {
2674                 /* Parent node not yet cached */
2675                 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2676                                            cur->level + 1);
2677                 if (!upper) {
2678                         btrfs_backref_free_edge(cache, edge);
2679                         return -ENOMEM;
2680                 }
2681
2682                 /*
2683                  *  Backrefs for the upper level block isn't cached, add the
2684                  *  block to pending list
2685                  */
2686                 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2687         } else {
2688                 /* Parent node already cached */
2689                 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2690                 ASSERT(upper->checked);
2691                 INIT_LIST_HEAD(&edge->list[UPPER]);
2692         }
2693         btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2694         return 0;
2695 }
2696
2697 /*
2698  * Handle indirect tree backref
2699  *
2700  * Indirect tree backref means, we only know which tree the node belongs to.
2701  * We still need to do a tree search to find out the parents. This is for
2702  * TREE_BLOCK_REF backref (keyed or inlined).
2703  *
2704  * @ref_key:    The same as @ref_key in  handle_direct_tree_backref()
2705  * @tree_key:   The first key of this tree block.
2706  * @path:       A clean (released) path, to avoid allocating path every time
2707  *              the function get called.
2708  */
2709 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2710                                         struct btrfs_path *path,
2711                                         struct btrfs_key *ref_key,
2712                                         struct btrfs_key *tree_key,
2713                                         struct btrfs_backref_node *cur)
2714 {
2715         struct btrfs_fs_info *fs_info = cache->fs_info;
2716         struct btrfs_backref_node *upper;
2717         struct btrfs_backref_node *lower;
2718         struct btrfs_backref_edge *edge;
2719         struct extent_buffer *eb;
2720         struct btrfs_root *root;
2721         struct rb_node *rb_node;
2722         int level;
2723         bool need_check = true;
2724         int ret;
2725
2726         root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2727         if (IS_ERR(root))
2728                 return PTR_ERR(root);
2729         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2730                 cur->cowonly = 1;
2731
2732         if (btrfs_root_level(&root->root_item) == cur->level) {
2733                 /* Tree root */
2734                 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2735                 /*
2736                  * For reloc backref cache, we may ignore reloc root.  But for
2737                  * general purpose backref cache, we can't rely on
2738                  * btrfs_should_ignore_reloc_root() as it may conflict with
2739                  * current running relocation and lead to missing root.
2740                  *
2741                  * For general purpose backref cache, reloc root detection is
2742                  * completely relying on direct backref (key->offset is parent
2743                  * bytenr), thus only do such check for reloc cache.
2744                  */
2745                 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2746                         btrfs_put_root(root);
2747                         list_add(&cur->list, &cache->useless_node);
2748                 } else {
2749                         cur->root = root;
2750                 }
2751                 return 0;
2752         }
2753
2754         level = cur->level + 1;
2755
2756         /* Search the tree to find parent blocks referring to the block */
2757         path->search_commit_root = 1;
2758         path->skip_locking = 1;
2759         path->lowest_level = level;
2760         ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2761         path->lowest_level = 0;
2762         if (ret < 0) {
2763                 btrfs_put_root(root);
2764                 return ret;
2765         }
2766         if (ret > 0 && path->slots[level] > 0)
2767                 path->slots[level]--;
2768
2769         eb = path->nodes[level];
2770         if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2771                 btrfs_err(fs_info,
2772 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2773                           cur->bytenr, level - 1, root->root_key.objectid,
2774                           tree_key->objectid, tree_key->type, tree_key->offset);
2775                 btrfs_put_root(root);
2776                 ret = -ENOENT;
2777                 goto out;
2778         }
2779         lower = cur;
2780
2781         /* Add all nodes and edges in the path */
2782         for (; level < BTRFS_MAX_LEVEL; level++) {
2783                 if (!path->nodes[level]) {
2784                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
2785                                lower->bytenr);
2786                         /* Same as previous should_ignore_reloc_root() call */
2787                         if (btrfs_should_ignore_reloc_root(root) &&
2788                             cache->is_reloc) {
2789                                 btrfs_put_root(root);
2790                                 list_add(&lower->list, &cache->useless_node);
2791                         } else {
2792                                 lower->root = root;
2793                         }
2794                         break;
2795                 }
2796
2797                 edge = btrfs_backref_alloc_edge(cache);
2798                 if (!edge) {
2799                         btrfs_put_root(root);
2800                         ret = -ENOMEM;
2801                         goto out;
2802                 }
2803
2804                 eb = path->nodes[level];
2805                 rb_node = rb_simple_search(&cache->rb_root, eb->start);
2806                 if (!rb_node) {
2807                         upper = btrfs_backref_alloc_node(cache, eb->start,
2808                                                          lower->level + 1);
2809                         if (!upper) {
2810                                 btrfs_put_root(root);
2811                                 btrfs_backref_free_edge(cache, edge);
2812                                 ret = -ENOMEM;
2813                                 goto out;
2814                         }
2815                         upper->owner = btrfs_header_owner(eb);
2816                         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2817                                 upper->cowonly = 1;
2818
2819                         /*
2820                          * If we know the block isn't shared we can avoid
2821                          * checking its backrefs.
2822                          */
2823                         if (btrfs_block_can_be_shared(root, eb))
2824                                 upper->checked = 0;
2825                         else
2826                                 upper->checked = 1;
2827
2828                         /*
2829                          * Add the block to pending list if we need to check its
2830                          * backrefs, we only do this once while walking up a
2831                          * tree as we will catch anything else later on.
2832                          */
2833                         if (!upper->checked && need_check) {
2834                                 need_check = false;
2835                                 list_add_tail(&edge->list[UPPER],
2836                                               &cache->pending_edge);
2837                         } else {
2838                                 if (upper->checked)
2839                                         need_check = true;
2840                                 INIT_LIST_HEAD(&edge->list[UPPER]);
2841                         }
2842                 } else {
2843                         upper = rb_entry(rb_node, struct btrfs_backref_node,
2844                                          rb_node);
2845                         ASSERT(upper->checked);
2846                         INIT_LIST_HEAD(&edge->list[UPPER]);
2847                         if (!upper->owner)
2848                                 upper->owner = btrfs_header_owner(eb);
2849                 }
2850                 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2851
2852                 if (rb_node) {
2853                         btrfs_put_root(root);
2854                         break;
2855                 }
2856                 lower = upper;
2857                 upper = NULL;
2858         }
2859 out:
2860         btrfs_release_path(path);
2861         return ret;
2862 }
2863
2864 /*
2865  * Add backref node @cur into @cache.
2866  *
2867  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2868  *       links aren't yet bi-directional. Needs to finish such links.
2869  *       Use btrfs_backref_finish_upper_links() to finish such linkage.
2870  *
2871  * @path:       Released path for indirect tree backref lookup
2872  * @iter:       Released backref iter for extent tree search
2873  * @node_key:   The first key of the tree block
2874  */
2875 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2876                                 struct btrfs_path *path,
2877                                 struct btrfs_backref_iter *iter,
2878                                 struct btrfs_key *node_key,
2879                                 struct btrfs_backref_node *cur)
2880 {
2881         struct btrfs_fs_info *fs_info = cache->fs_info;
2882         struct btrfs_backref_edge *edge;
2883         struct btrfs_backref_node *exist;
2884         int ret;
2885
2886         ret = btrfs_backref_iter_start(iter, cur->bytenr);
2887         if (ret < 0)
2888                 return ret;
2889         /*
2890          * We skip the first btrfs_tree_block_info, as we don't use the key
2891          * stored in it, but fetch it from the tree block
2892          */
2893         if (btrfs_backref_has_tree_block_info(iter)) {
2894                 ret = btrfs_backref_iter_next(iter);
2895                 if (ret < 0)
2896                         goto out;
2897                 /* No extra backref? This means the tree block is corrupted */
2898                 if (ret > 0) {
2899                         ret = -EUCLEAN;
2900                         goto out;
2901                 }
2902         }
2903         WARN_ON(cur->checked);
2904         if (!list_empty(&cur->upper)) {
2905                 /*
2906                  * The backref was added previously when processing backref of
2907                  * type BTRFS_TREE_BLOCK_REF_KEY
2908                  */
2909                 ASSERT(list_is_singular(&cur->upper));
2910                 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2911                                   list[LOWER]);
2912                 ASSERT(list_empty(&edge->list[UPPER]));
2913                 exist = edge->node[UPPER];
2914                 /*
2915                  * Add the upper level block to pending list if we need check
2916                  * its backrefs
2917                  */
2918                 if (!exist->checked)
2919                         list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2920         } else {
2921                 exist = NULL;
2922         }
2923
2924         for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2925                 struct extent_buffer *eb;
2926                 struct btrfs_key key;
2927                 int type;
2928
2929                 cond_resched();
2930                 eb = btrfs_backref_get_eb(iter);
2931
2932                 key.objectid = iter->bytenr;
2933                 if (btrfs_backref_iter_is_inline_ref(iter)) {
2934                         struct btrfs_extent_inline_ref *iref;
2935
2936                         /* Update key for inline backref */
2937                         iref = (struct btrfs_extent_inline_ref *)
2938                                 ((unsigned long)iter->cur_ptr);
2939                         type = btrfs_get_extent_inline_ref_type(eb, iref,
2940                                                         BTRFS_REF_TYPE_BLOCK);
2941                         if (type == BTRFS_REF_TYPE_INVALID) {
2942                                 ret = -EUCLEAN;
2943                                 goto out;
2944                         }
2945                         key.type = type;
2946                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2947                 } else {
2948                         key.type = iter->cur_key.type;
2949                         key.offset = iter->cur_key.offset;
2950                 }
2951
2952                 /*
2953                  * Parent node found and matches current inline ref, no need to
2954                  * rebuild this node for this inline ref
2955                  */
2956                 if (exist &&
2957                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2958                       exist->owner == key.offset) ||
2959                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2960                       exist->bytenr == key.offset))) {
2961                         exist = NULL;
2962                         continue;
2963                 }
2964
2965                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
2966                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2967                         ret = handle_direct_tree_backref(cache, &key, cur);
2968                         if (ret < 0)
2969                                 goto out;
2970                         continue;
2971                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
2972                         ret = -EINVAL;
2973                         btrfs_print_v0_err(fs_info);
2974                         btrfs_handle_fs_error(fs_info, ret, NULL);
2975                         goto out;
2976                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
2977                         continue;
2978                 }
2979
2980                 /*
2981                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
2982                  * means the root objectid. We need to search the tree to get
2983                  * its parent bytenr.
2984                  */
2985                 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
2986                                                    cur);
2987                 if (ret < 0)
2988                         goto out;
2989         }
2990         ret = 0;
2991         cur->checked = 1;
2992         WARN_ON(exist);
2993 out:
2994         btrfs_backref_iter_release(iter);
2995         return ret;
2996 }
2997
2998 /*
2999  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3000  */
3001 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3002                                      struct btrfs_backref_node *start)
3003 {
3004         struct list_head *useless_node = &cache->useless_node;
3005         struct btrfs_backref_edge *edge;
3006         struct rb_node *rb_node;
3007         LIST_HEAD(pending_edge);
3008
3009         ASSERT(start->checked);
3010
3011         /* Insert this node to cache if it's not COW-only */
3012         if (!start->cowonly) {
3013                 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
3014                                            &start->rb_node);
3015                 if (rb_node)
3016                         btrfs_backref_panic(cache->fs_info, start->bytenr,
3017                                             -EEXIST);
3018                 list_add_tail(&start->lower, &cache->leaves);
3019         }
3020
3021         /*
3022          * Use breadth first search to iterate all related edges.
3023          *
3024          * The starting points are all the edges of this node
3025          */
3026         list_for_each_entry(edge, &start->upper, list[LOWER])
3027                 list_add_tail(&edge->list[UPPER], &pending_edge);
3028
3029         while (!list_empty(&pending_edge)) {
3030                 struct btrfs_backref_node *upper;
3031                 struct btrfs_backref_node *lower;
3032
3033                 edge = list_first_entry(&pending_edge,
3034                                 struct btrfs_backref_edge, list[UPPER]);
3035                 list_del_init(&edge->list[UPPER]);
3036                 upper = edge->node[UPPER];
3037                 lower = edge->node[LOWER];
3038
3039                 /* Parent is detached, no need to keep any edges */
3040                 if (upper->detached) {
3041                         list_del(&edge->list[LOWER]);
3042                         btrfs_backref_free_edge(cache, edge);
3043
3044                         /* Lower node is orphan, queue for cleanup */
3045                         if (list_empty(&lower->upper))
3046                                 list_add(&lower->list, useless_node);
3047                         continue;
3048                 }
3049
3050                 /*
3051                  * All new nodes added in current build_backref_tree() haven't
3052                  * been linked to the cache rb tree.
3053                  * So if we have upper->rb_node populated, this means a cache
3054                  * hit. We only need to link the edge, as @upper and all its
3055                  * parents have already been linked.
3056                  */
3057                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3058                         if (upper->lowest) {
3059                                 list_del_init(&upper->lower);
3060                                 upper->lowest = 0;
3061                         }
3062
3063                         list_add_tail(&edge->list[UPPER], &upper->lower);
3064                         continue;
3065                 }
3066
3067                 /* Sanity check, we shouldn't have any unchecked nodes */
3068                 if (!upper->checked) {
3069                         ASSERT(0);
3070                         return -EUCLEAN;
3071                 }
3072
3073                 /* Sanity check, COW-only node has non-COW-only parent */
3074                 if (start->cowonly != upper->cowonly) {
3075                         ASSERT(0);
3076                         return -EUCLEAN;
3077                 }
3078
3079                 /* Only cache non-COW-only (subvolume trees) tree blocks */
3080                 if (!upper->cowonly) {
3081                         rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3082                                                    &upper->rb_node);
3083                         if (rb_node) {
3084                                 btrfs_backref_panic(cache->fs_info,
3085                                                 upper->bytenr, -EEXIST);
3086                                 return -EUCLEAN;
3087                         }
3088                 }
3089
3090                 list_add_tail(&edge->list[UPPER], &upper->lower);
3091
3092                 /*
3093                  * Also queue all the parent edges of this uncached node
3094                  * to finish the upper linkage
3095                  */
3096                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3097                         list_add_tail(&edge->list[UPPER], &pending_edge);
3098         }
3099         return 0;
3100 }
3101
3102 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3103                                  struct btrfs_backref_node *node)
3104 {
3105         struct btrfs_backref_node *lower;
3106         struct btrfs_backref_node *upper;
3107         struct btrfs_backref_edge *edge;
3108
3109         while (!list_empty(&cache->useless_node)) {
3110                 lower = list_first_entry(&cache->useless_node,
3111                                    struct btrfs_backref_node, list);
3112                 list_del_init(&lower->list);
3113         }
3114         while (!list_empty(&cache->pending_edge)) {
3115                 edge = list_first_entry(&cache->pending_edge,
3116                                 struct btrfs_backref_edge, list[UPPER]);
3117                 list_del(&edge->list[UPPER]);
3118                 list_del(&edge->list[LOWER]);
3119                 lower = edge->node[LOWER];
3120                 upper = edge->node[UPPER];
3121                 btrfs_backref_free_edge(cache, edge);
3122
3123                 /*
3124                  * Lower is no longer linked to any upper backref nodes and
3125                  * isn't in the cache, we can free it ourselves.
3126                  */
3127                 if (list_empty(&lower->upper) &&
3128                     RB_EMPTY_NODE(&lower->rb_node))
3129                         list_add(&lower->list, &cache->useless_node);
3130
3131                 if (!RB_EMPTY_NODE(&upper->rb_node))
3132                         continue;
3133
3134                 /* Add this guy's upper edges to the list to process */
3135                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3136                         list_add_tail(&edge->list[UPPER],
3137                                       &cache->pending_edge);
3138                 if (list_empty(&upper->upper))
3139                         list_add(&upper->list, &cache->useless_node);
3140         }
3141
3142         while (!list_empty(&cache->useless_node)) {
3143                 lower = list_first_entry(&cache->useless_node,
3144                                    struct btrfs_backref_node, list);
3145                 list_del_init(&lower->list);
3146                 if (lower == node)
3147                         node = NULL;
3148                 btrfs_backref_drop_node(cache, lower);
3149         }
3150
3151         btrfs_backref_cleanup_node(cache, node);
3152         ASSERT(list_empty(&cache->useless_node) &&
3153                list_empty(&cache->pending_edge));
3154 }