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
  17 /* Just an arbitrary number so we can be sure this happened */
  18 #define BACKREF_FOUND_SHARED 6
  19
  20 struct extent_inode_elem {
  21         u64 inum;
  22         u64 offset;
  23         struct extent_inode_elem *next;
  24 };
  25
  26 static int check_extent_in_eb(const struct btrfs_key *key,
  27                               const struct extent_buffer *eb,
  28                               const struct btrfs_file_extent_item *fi,
  29                               u64 extent_item_pos,
  30                               struct extent_inode_elem **eie,
  31                               bool ignore_offset)
  32 {
  33         u64 offset = 0;
  34         struct extent_inode_elem *e;
  35
  36         if (!ignore_offset &&
  37             !btrfs_file_extent_compression(eb, fi) &&
  38             !btrfs_file_extent_encryption(eb, fi) &&
  39             !btrfs_file_extent_other_encoding(eb, fi)) {
  40                 u64 data_offset;
  41                 u64 data_len;
  42
  43                 data_offset = btrfs_file_extent_offset(eb, fi);
  44                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  45
  46                 if (extent_item_pos < data_offset ||
  47                     extent_item_pos >= data_offset + data_len)
  48                         return 1;
  49                 offset = extent_item_pos - data_offset;
  50         }
  51
  52         e = kmalloc(sizeof(*e), GFP_NOFS);
  53         if (!e)
  54                 return -ENOMEM;
  55
  56         e->next = *eie;
  57         e->inum = key->objectid;
  58         e->offset = key->offset + offset;
  59         *eie = e;
  60
  61         return 0;
  62 }
  63
  64 static void free_inode_elem_list(struct extent_inode_elem *eie)
  65 {
  66         struct extent_inode_elem *eie_next;
  67
  68         for (; eie; eie = eie_next) {
  69                 eie_next = eie->next;
  70                 kfree(eie);
  71         }
  72 }
  73
  74 static int find_extent_in_eb(const struct extent_buffer *eb,
  75                              u64 wanted_disk_byte, u64 extent_item_pos,
  76                              struct extent_inode_elem **eie,
  77                              bool ignore_offset)
  78 {
  79         u64 disk_byte;
  80         struct btrfs_key key;
  81         struct btrfs_file_extent_item *fi;
  82         int slot;
  83         int nritems;
  84         int extent_type;
  85         int ret;
  86
  87         /*
  88          * from the shared data ref, we only have the leaf but we need
  89          * the key. thus, we must look into all items and see that we
  90          * find one (some) with a reference to our extent item.
  91          */
  92         nritems = btrfs_header_nritems(eb);
  93         for (slot = 0; slot < nritems; ++slot) {
  94                 btrfs_item_key_to_cpu(eb, &key, slot);
  95                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  96                         continue;
  97                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
  98                 extent_type = btrfs_file_extent_type(eb, fi);
  99                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 100                         continue;
 101                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 102                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 103                 if (disk_byte != wanted_disk_byte)
 104                         continue;
 105
 106                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 107                 if (ret < 0)
 108                         return ret;
 109         }
 110
 111         return 0;
 112 }
 113
 114 struct preftree {
 115         struct rb_root root;
 116         unsigned int count;
 117 };
 118
 119 #define PREFTREE_INIT   { .root = RB_ROOT, .count = 0 }
 120
 121 struct preftrees {
 122         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 123         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 124         struct preftree indirect_missing_keys;
 125 };
 126
 127 /*
 128  * Checks for a shared extent during backref search.
 129  *
 130  * The share_count tracks prelim_refs (direct and indirect) having a
 131  * ref->count >0:
 132  *  - incremented when a ref->count transitions to >0
 133  *  - decremented when a ref->count transitions to <1
 134  */
 135 struct share_check {
 136         u64 root_objectid;
 137         u64 inum;
 138         int share_count;
 139         bool have_delayed_delete_refs;
 140 };
 141
 142 static inline int extent_is_shared(struct share_check *sc)
 143 {
 144         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 145 }
 146
 147 static struct kmem_cache *btrfs_prelim_ref_cache;
 148
 149 int __init btrfs_prelim_ref_init(void)
 150 {
 151         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 152                                         sizeof(struct prelim_ref),
 153                                         0,
 154                                         SLAB_MEM_SPREAD,
 155                                         NULL);
 156         if (!btrfs_prelim_ref_cache)
 157                 return -ENOMEM;
 158         return 0;
 159 }
 160
 161 void __cold btrfs_prelim_ref_exit(void)
 162 {
 163         kmem_cache_destroy(btrfs_prelim_ref_cache);
 164 }
 165
 166 static void free_pref(struct prelim_ref *ref)
 167 {
 168         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 169 }
 170
 171 /*
 172  * Return 0 when both refs are for the same block (and can be merged).
 173  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 174  * indicates a 'higher' block.
 175  */
 176 static int prelim_ref_compare(struct prelim_ref *ref1,
 177                               struct prelim_ref *ref2)
 178 {
 179         if (ref1->level < ref2->level)
 180                 return -1;
 181         if (ref1->level > ref2->level)
 182                 return 1;
 183         if (ref1->root_id < ref2->root_id)
 184                 return -1;
 185         if (ref1->root_id > ref2->root_id)
 186                 return 1;
 187         if (ref1->key_for_search.type < ref2->key_for_search.type)
 188                 return -1;
 189         if (ref1->key_for_search.type > ref2->key_for_search.type)
 190                 return 1;
 191         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 192                 return -1;
 193         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 194                 return 1;
 195         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 196                 return -1;
 197         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 198                 return 1;
 199         if (ref1->parent < ref2->parent)
 200                 return -1;
 201         if (ref1->parent > ref2->parent)
 202                 return 1;
 203
 204         return 0;
 205 }
 206
 207 static void update_share_count(struct share_check *sc, int oldcount,
 208                                int newcount)
 209 {
 210         if ((!sc) || (oldcount == 0 && newcount < 1))
 211                 return;
 212
 213         if (oldcount > 0 && newcount < 1)
 214                 sc->share_count--;
 215         else if (oldcount < 1 && newcount > 0)
 216                 sc->share_count++;
 217 }
 218
 219 /*
 220  * Add @newref to the @root rbtree, merging identical refs.
 221  *
 222  * Callers should assume that newref has been freed after calling.
 223  */
 224 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 225                               struct preftree *preftree,
 226                               struct prelim_ref *newref,
 227                               struct share_check *sc)
 228 {
 229         struct rb_root *root;
 230         struct rb_node **p;
 231         struct rb_node *parent = NULL;
 232         struct prelim_ref *ref;
 233         int result;
 234
 235         root = &preftree->root;
 236         p = &root->rb_node;
 237
 238         while (*p) {
 239                 parent = *p;
 240                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 241                 result = prelim_ref_compare(ref, newref);
 242                 if (result < 0) {
 243                         p = &(*p)->rb_left;
 244                 } else if (result > 0) {
 245                         p = &(*p)->rb_right;
 246                 } else {
 247                         /* Identical refs, merge them and free @newref */
 248                         struct extent_inode_elem *eie = ref->inode_list;
 249
 250                         while (eie && eie->next)
 251                                 eie = eie->next;
 252
 253                         if (!eie)
 254                                 ref->inode_list = newref->inode_list;
 255                         else
 256                                 eie->next = newref->inode_list;
 257                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 258                                                      preftree->count);
 259                         /*
 260                          * A delayed ref can have newref->count < 0.
 261                          * The ref->count is updated to follow any
 262                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 263                          */
 264                         update_share_count(sc, ref->count,
 265                                            ref->count + newref->count);
 266                         ref->count += newref->count;
 267                         free_pref(newref);
 268                         return;
 269                 }
 270         }
 271
 272         update_share_count(sc, 0, newref->count);
 273         preftree->count++;
 274         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 275         rb_link_node(&newref->rbnode, parent, p);
 276         rb_insert_color(&newref->rbnode, root);
 277 }
 278
 279 /*
 280  * Release the entire tree.  We don't care about internal consistency so
 281  * just free everything and then reset the tree root.
 282  */
 283 static void prelim_release(struct preftree *preftree)
 284 {
 285         struct prelim_ref *ref, *next_ref;
 286
 287         rbtree_postorder_for_each_entry_safe(ref, next_ref, &preftree->root,
 288                                              rbnode)
 289                 free_pref(ref);
 290
 291         preftree->root = RB_ROOT;
 292         preftree->count = 0;
 293 }
 294
 295 /*
 296  * the rules for all callers of this function are:
 297  * - obtaining the parent is the goal
 298  * - if you add a key, you must know that it is a correct key
 299  * - if you cannot add the parent or a correct key, then we will look into the
 300  *   block later to set a correct key
 301  *
 302  * delayed refs
 303  * ============
 304  *        backref type | shared | indirect | shared | indirect
 305  * information         |   tree |     tree |   data |     data
 306  * --------------------+--------+----------+--------+----------
 307  *      parent logical |    y   |     -    |    -   |     -
 308  *      key to resolve |    -   |     y    |    y   |     y
 309  *  tree block logical |    -   |     -    |    -   |     -
 310  *  root for resolving |    y   |     y    |    y   |     y
 311  *
 312  * - column 1:       we've the parent -> done
 313  * - column 2, 3, 4: we use the key to find the parent
 314  *
 315  * on disk refs (inline or keyed)
 316  * ==============================
 317  *        backref type | shared | indirect | shared | indirect
 318  * information         |   tree |     tree |   data |     data
 319  * --------------------+--------+----------+--------+----------
 320  *      parent logical |    y   |     -    |    y   |     -
 321  *      key to resolve |    -   |     -    |    -   |     y
 322  *  tree block logical |    y   |     y    |    y   |     y
 323  *  root for resolving |    -   |     y    |    y   |     y
 324  *
 325  * - column 1, 3: we've the parent -> done
 326  * - column 2:    we take the first key from the block to find the parent
 327  *                (see add_missing_keys)
 328  * - column 4:    we use the key to find the parent
 329  *
 330  * additional information that's available but not required to find the parent
 331  * block might help in merging entries to gain some speed.
 332  */
 333 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 334                           struct preftree *preftree, u64 root_id,
 335                           const struct btrfs_key *key, int level, u64 parent,
 336                           u64 wanted_disk_byte, int count,
 337                           struct share_check *sc, gfp_t gfp_mask)
 338 {
 339         struct prelim_ref *ref;
 340
 341         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 342                 return 0;
 343
 344         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 345         if (!ref)
 346                 return -ENOMEM;
 347
 348         ref->root_id = root_id;
 349         if (key) {
 350                 ref->key_for_search = *key;
 351                 /*
 352                  * We can often find data backrefs with an offset that is too
 353                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 354                  * underflows when subtracting a file's offset with the data
 355                  * offset of its corresponding extent data item. This can
 356                  * happen for example in the clone ioctl.
 357                  * So if we detect such case we set the search key's offset to
 358                  * zero to make sure we will find the matching file extent item
 359                  * at add_all_parents(), otherwise we will miss it because the
 360                  * offset taken form the backref is much larger then the offset
 361                  * of the file extent item. This can make us scan a very large
 362                  * number of file extent items, but at least it will not make
 363                  * us miss any.
 364                  * This is an ugly workaround for a behaviour that should have
 365                  * never existed, but it does and a fix for the clone ioctl
 366                  * would touch a lot of places, cause backwards incompatibility
 367                  * and would not fix the problem for extents cloned with older
 368                  * kernels.
 369                  */
 370                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 371                     ref->key_for_search.offset >= LLONG_MAX)
 372                         ref->key_for_search.offset = 0;
 373         } else {
 374                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 375         }
 376
 377         ref->inode_list = NULL;
 378         ref->level = level;
 379         ref->count = count;
 380         ref->parent = parent;
 381         ref->wanted_disk_byte = wanted_disk_byte;
 382         prelim_ref_insert(fs_info, preftree, ref, sc);
 383         return extent_is_shared(sc);
 384 }
 385
 386 /* direct refs use root == 0, key == NULL */
 387 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 388                           struct preftrees *preftrees, int level, u64 parent,
 389                           u64 wanted_disk_byte, int count,
 390                           struct share_check *sc, gfp_t gfp_mask)
 391 {
 392         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 393                               parent, wanted_disk_byte, count, sc, gfp_mask);
 394 }
 395
 396 /* indirect refs use parent == 0 */
 397 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 398                             struct preftrees *preftrees, u64 root_id,
 399                             const struct btrfs_key *key, int level,
 400                             u64 wanted_disk_byte, int count,
 401                             struct share_check *sc, gfp_t gfp_mask)
 402 {
 403         struct preftree *tree = &preftrees->indirect;
 404
 405         if (!key)
 406                 tree = &preftrees->indirect_missing_keys;
 407         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 408                               wanted_disk_byte, count, sc, gfp_mask);
 409 }
 410
 411 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 412                            struct ulist *parents, struct prelim_ref *ref,
 413                            int level, u64 time_seq, const u64 *extent_item_pos,
 414                            u64 total_refs, bool ignore_offset)
 415 {
 416         int ret = 0;
 417         int slot;
 418         struct extent_buffer *eb;
 419         struct btrfs_key key;
 420         struct btrfs_key *key_for_search = &ref->key_for_search;
 421         struct btrfs_file_extent_item *fi;
 422         struct extent_inode_elem *eie = NULL, *old = NULL;
 423         u64 disk_byte;
 424         u64 wanted_disk_byte = ref->wanted_disk_byte;
 425         u64 count = 0;
 426
 427         if (level != 0) {
 428                 eb = path->nodes[level];
 429                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 430                 if (ret < 0)
 431                         return ret;
 432                 return 0;
 433         }
 434
 435         /*
 436          * We normally enter this function with the path already pointing to
 437          * the first item to check. But sometimes, we may enter it with
 438          * slot==nritems. In that case, go to the next leaf before we continue.
 439          */
 440         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 441                 if (time_seq == SEQ_LAST)
 442                         ret = btrfs_next_leaf(root, path);
 443                 else
 444                         ret = btrfs_next_old_leaf(root, path, time_seq);
 445         }
 446
 447         while (!ret && count < total_refs) {
 448                 eb = path->nodes[0];
 449                 slot = path->slots[0];
 450
 451                 btrfs_item_key_to_cpu(eb, &key, slot);
 452
 453                 if (key.objectid != key_for_search->objectid ||
 454                     key.type != BTRFS_EXTENT_DATA_KEY)
 455                         break;
 456
 457                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 458                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 459
 460                 if (disk_byte == wanted_disk_byte) {
 461                         eie = NULL;
 462                         old = NULL;
 463                         count++;
 464                         if (extent_item_pos) {
 465                                 ret = check_extent_in_eb(&key, eb, fi,
 466                                                 *extent_item_pos,
 467                                                 &eie, ignore_offset);
 468                                 if (ret < 0)
 469                                         break;
 470                         }
 471                         if (ret > 0)
 472                                 goto next;
 473                         ret = ulist_add_merge_ptr(parents, eb->start,
 474                                                   eie, (void **)&old, GFP_NOFS);
 475                         if (ret < 0)
 476                                 break;
 477                         if (!ret && extent_item_pos) {
 478                                 while (old->next)
 479                                         old = old->next;
 480                                 old->next = eie;
 481                         }
 482                         eie = NULL;
 483                 }
 484 next:
 485                 if (time_seq == SEQ_LAST)
 486                         ret = btrfs_next_item(root, path);
 487                 else
 488                         ret = btrfs_next_old_item(root, path, time_seq);
 489         }
 490
 491         if (ret > 0)
 492                 ret = 0;
 493         else if (ret < 0)
 494                 free_inode_elem_list(eie);
 495         return ret;
 496 }
 497
 498 /*
 499  * resolve an indirect backref in the form (root_id, key, level)
 500  * to a logical address
 501  */
 502 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 503                                 struct btrfs_path *path, u64 time_seq,
 504                                 struct prelim_ref *ref, struct ulist *parents,
 505                                 const u64 *extent_item_pos, u64 total_refs,
 506                                 bool ignore_offset)
 507 {
 508         struct btrfs_root *root;
 509         struct btrfs_key root_key;
 510         struct extent_buffer *eb;
 511         int ret = 0;
 512         int root_level;
 513         int level = ref->level;
 514         int index;
 515
 516         root_key.objectid = ref->root_id;
 517         root_key.type = BTRFS_ROOT_ITEM_KEY;
 518         root_key.offset = (u64)-1;
 519
 520         index = srcu_read_lock(&fs_info->subvol_srcu);
 521
 522         root = btrfs_get_fs_root(fs_info, &root_key, false);
 523         if (IS_ERR(root)) {
 524                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 525                 ret = PTR_ERR(root);
 526                 goto out;
 527         }
 528
 529         if (btrfs_is_testing(fs_info)) {
 530                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 531                 ret = -ENOENT;
 532                 goto out;
 533         }
 534
 535         if (path->search_commit_root)
 536                 root_level = btrfs_header_level(root->commit_root);
 537         else if (time_seq == SEQ_LAST)
 538                 root_level = btrfs_header_level(root->node);
 539         else
 540                 root_level = btrfs_old_root_level(root, time_seq);
 541
 542         if (root_level + 1 == level) {
 543                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 544                 goto out;
 545         }
 546
 547         path->lowest_level = level;
 548         if (time_seq == SEQ_LAST)
 549                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 550                                         0, 0);
 551         else
 552                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 553                                             time_seq);
 554
 555         /* root node has been locked, we can release @subvol_srcu safely here */
 556         srcu_read_unlock(&fs_info->subvol_srcu, index);
 557
 558         btrfs_debug(fs_info,
 559                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 560                  ref->root_id, level, ref->count, ret,
 561                  ref->key_for_search.objectid, ref->key_for_search.type,
 562                  ref->key_for_search.offset);
 563         if (ret < 0)
 564                 goto out;
 565
 566         eb = path->nodes[level];
 567         while (!eb) {
 568                 if (WARN_ON(!level)) {
 569                         ret = 1;
 570                         goto out;
 571                 }
 572                 level--;
 573                 eb = path->nodes[level];
 574         }
 575
 576         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 577                               extent_item_pos, total_refs, ignore_offset);
 578 out:
 579         path->lowest_level = 0;
 580         btrfs_release_path(path);
 581         return ret;
 582 }
 583
 584 static struct extent_inode_elem *
 585 unode_aux_to_inode_list(struct ulist_node *node)
 586 {
 587         if (!node)
 588                 return NULL;
 589         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 590 }
 591
 592 static void free_leaf_list(struct ulist *ulist)
 593 {
 594         struct ulist_node *node;
 595         struct ulist_iterator uiter;
 596
 597         ULIST_ITER_INIT(&uiter);
 598         while ((node = ulist_next(ulist, &uiter)))
 599                 free_inode_elem_list(unode_aux_to_inode_list(node));
 600
 601         ulist_free(ulist);
 602 }
 603
 604 /*
 605  * We maintain three seperate rbtrees: one for direct refs, one for
 606  * indirect refs which have a key, and one for indirect refs which do not
 607  * have a key. Each tree does merge on insertion.
 608  *
 609  * Once all of the references are located, we iterate over the tree of
 610  * indirect refs with missing keys. An appropriate key is located and
 611  * the ref is moved onto the tree for indirect refs. After all missing
 612  * keys are thus located, we iterate over the indirect ref tree, resolve
 613  * each reference, and then insert the resolved reference onto the
 614  * direct tree (merging there too).
 615  *
 616  * New backrefs (i.e., for parent nodes) are added to the appropriate
 617  * rbtree as they are encountered. The new backrefs are subsequently
 618  * resolved as above.
 619  */
 620 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 621                                  struct btrfs_path *path, u64 time_seq,
 622                                  struct preftrees *preftrees,
 623                                  const u64 *extent_item_pos, u64 total_refs,
 624                                  struct share_check *sc, bool ignore_offset)
 625 {
 626         int err;
 627         int ret = 0;
 628         struct ulist *parents;
 629         struct ulist_node *node;
 630         struct ulist_iterator uiter;
 631         struct rb_node *rnode;
 632
 633         parents = ulist_alloc(GFP_NOFS);
 634         if (!parents)
 635                 return -ENOMEM;
 636
 637         /*
 638          * We could trade memory usage for performance here by iterating
 639          * the tree, allocating new refs for each insertion, and then
 640          * freeing the entire indirect tree when we're done.  In some test
 641          * cases, the tree can grow quite large (~200k objects).
 642          */
 643         while ((rnode = rb_first(&preftrees->indirect.root))) {
 644                 struct prelim_ref *ref;
 645
 646                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 647                 if (WARN(ref->parent,
 648                          "BUG: direct ref found in indirect tree")) {
 649                         ret = -EINVAL;
 650                         goto out;
 651                 }
 652
 653                 rb_erase(&ref->rbnode, &preftrees->indirect.root);
 654                 preftrees->indirect.count--;
 655
 656                 if (ref->count == 0) {
 657                         free_pref(ref);
 658                         continue;
 659                 }
 660
 661                 if (sc && sc->root_objectid &&
 662                     ref->root_id != sc->root_objectid) {
 663                         free_pref(ref);
 664                         ret = BACKREF_FOUND_SHARED;
 665                         goto out;
 666                 }
 667                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
 668                                            parents, extent_item_pos,
 669                                            total_refs, ignore_offset);
 670                 /*
 671                  * we can only tolerate ENOENT,otherwise,we should catch error
 672                  * and return directly.
 673                  */
 674                 if (err == -ENOENT) {
 675                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 676                                           NULL);
 677                         continue;
 678                 } else if (err) {
 679                         free_pref(ref);
 680                         ret = err;
 681                         goto out;
 682                 }
 683
 684                 /* we put the first parent into the ref at hand */
 685                 ULIST_ITER_INIT(&uiter);
 686                 node = ulist_next(parents, &uiter);
 687                 ref->parent = node ? node->val : 0;
 688                 ref->inode_list = unode_aux_to_inode_list(node);
 689
 690                 /* Add a prelim_ref(s) for any other parent(s). */
 691                 while ((node = ulist_next(parents, &uiter))) {
 692                         struct prelim_ref *new_ref;
 693
 694                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 695                                                    GFP_NOFS);
 696                         if (!new_ref) {
 697                                 free_pref(ref);
 698                                 ret = -ENOMEM;
 699                                 goto out;
 700                         }
 701                         memcpy(new_ref, ref, sizeof(*ref));
 702                         new_ref->parent = node->val;
 703                         new_ref->inode_list = unode_aux_to_inode_list(node);
 704                         prelim_ref_insert(fs_info, &preftrees->direct,
 705                                           new_ref, NULL);
 706                 }
 707
 708                 /*
 709                  * Now it's a direct ref, put it in the the direct tree. We must
 710                  * do this last because the ref could be merged/freed here.
 711                  */
 712                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 713
 714                 ulist_reinit(parents);
 715                 cond_resched();
 716         }
 717 out:
 718         /*
 719          * We may have inode lists attached to refs in the parents ulist, so we
 720          * must free them before freeing the ulist and its refs.
 721          */
 722         free_leaf_list(parents);
 723         return ret;
 724 }
 725
 726 /*
 727  * read tree blocks and add keys where required.
 728  */
 729 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 730                             struct preftrees *preftrees, bool lock)
 731 {
 732         struct prelim_ref *ref;
 733         struct extent_buffer *eb;
 734         struct preftree *tree = &preftrees->indirect_missing_keys;
 735         struct rb_node *node;
 736
 737         while ((node = rb_first(&tree->root))) {
 738                 ref = rb_entry(node, struct prelim_ref, rbnode);
 739                 rb_erase(node, &tree->root);
 740
 741                 BUG_ON(ref->parent);    /* should not be a direct ref */
 742                 BUG_ON(ref->key_for_search.type);
 743                 BUG_ON(!ref->wanted_disk_byte);
 744
 745                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 746                                      ref->level - 1, NULL);
 747                 if (IS_ERR(eb)) {
 748                         free_pref(ref);
 749                         return PTR_ERR(eb);
 750                 } else if (!extent_buffer_uptodate(eb)) {
 751                         free_pref(ref);
 752                         free_extent_buffer(eb);
 753                         return -EIO;
 754                 }
 755                 if (lock)
 756                         btrfs_tree_read_lock(eb);
 757                 if (btrfs_header_level(eb) == 0)
 758                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 759                 else
 760                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 761                 if (lock)
 762                         btrfs_tree_read_unlock(eb);
 763                 free_extent_buffer(eb);
 764                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 765                 cond_resched();
 766         }
 767         return 0;
 768 }
 769
 770 /*
 771  * add all currently queued delayed refs from this head whose seq nr is
 772  * smaller or equal that seq to the list
 773  */
 774 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 775                             struct btrfs_delayed_ref_head *head, u64 seq,
 776                             struct preftrees *preftrees, u64 *total_refs,
 777                             struct share_check *sc)
 778 {
 779         struct btrfs_delayed_ref_node *node;
 780         struct btrfs_key key;
 781         struct rb_node *n;
 782         int count;
 783         int ret = 0;
 784
 785         spin_lock(&head->lock);
 786         for (n = rb_first(&head->ref_tree); n; n = rb_next(n)) {
 787                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 788                                 ref_node);
 789                 if (node->seq > seq)
 790                         continue;
 791
 792                 switch (node->action) {
 793                 case BTRFS_ADD_DELAYED_EXTENT:
 794                 case BTRFS_UPDATE_DELAYED_HEAD:
 795                         WARN_ON(1);
 796                         continue;
 797                 case BTRFS_ADD_DELAYED_REF:
 798                         count = node->ref_mod;
 799                         break;
 800                 case BTRFS_DROP_DELAYED_REF:
 801                         count = node->ref_mod * -1;
 802                         break;
 803                 default:
 804                         BUG_ON(1);
 805                 }
 806                 *total_refs += count;
 807                 switch (node->type) {
 808                 case BTRFS_TREE_BLOCK_REF_KEY: {
 809                         /* NORMAL INDIRECT METADATA backref */
 810                         struct btrfs_delayed_tree_ref *ref;
 811                         struct btrfs_key *key_ptr = NULL;
 812
 813                         if (head->extent_op && head->extent_op->update_key) {
 814                                 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
 815                                 key_ptr = &key;
 816                         }
 817
 818                         ref = btrfs_delayed_node_to_tree_ref(node);
 819                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 820                                                key_ptr, ref->level + 1,
 821                                                node->bytenr, count, sc,
 822                                                GFP_ATOMIC);
 823                         break;
 824                 }
 825                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 826                         /* SHARED DIRECT METADATA backref */
 827                         struct btrfs_delayed_tree_ref *ref;
 828
 829                         ref = btrfs_delayed_node_to_tree_ref(node);
 830
 831                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 832                                              ref->parent, node->bytenr, count,
 833                                              sc, GFP_ATOMIC);
 834                         break;
 835                 }
 836                 case BTRFS_EXTENT_DATA_REF_KEY: {
 837                         /* NORMAL INDIRECT DATA backref */
 838                         struct btrfs_delayed_data_ref *ref;
 839                         ref = btrfs_delayed_node_to_data_ref(node);
 840
 841                         key.objectid = ref->objectid;
 842                         key.type = BTRFS_EXTENT_DATA_KEY;
 843                         key.offset = ref->offset;
 844
 845                         /*
 846                          * If we have a share check context and a reference for
 847                          * another inode, we can't exit immediately. This is
 848                          * because even if this is a BTRFS_ADD_DELAYED_REF
 849                          * reference we may find next a BTRFS_DROP_DELAYED_REF
 850                          * which cancels out this ADD reference.
 851                          *
 852                          * If this is a DROP reference and there was no previous
 853                          * ADD reference, then we need to signal that when we
 854                          * process references from the extent tree (through
 855                          * add_inline_refs() and add_keyed_refs()), we should
 856                          * not exit early if we find a reference for another
 857                          * inode, because one of the delayed DROP references
 858                          * may cancel that reference in the extent tree.
 859                          */
 860                         if (sc && count < 0)
 861                                 sc->have_delayed_delete_refs = true;
 862
 863                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 864                                                &key, 0, node->bytenr, count, sc,
 865                                                GFP_ATOMIC);
 866                         break;
 867                 }
 868                 case BTRFS_SHARED_DATA_REF_KEY: {
 869                         /* SHARED DIRECT FULL backref */
 870                         struct btrfs_delayed_data_ref *ref;
 871
 872                         ref = btrfs_delayed_node_to_data_ref(node);
 873
 874                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 875                                              node->bytenr, count, sc,
 876                                              GFP_ATOMIC);
 877                         break;
 878                 }
 879                 default:
 880                         WARN_ON(1);
 881                 }
 882                 /*
 883                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 884                  * refs have been checked.
 885                  */
 886                 if (ret && (ret != BACKREF_FOUND_SHARED))
 887                         break;
 888         }
 889         if (!ret)
 890                 ret = extent_is_shared(sc);
 891
 892         spin_unlock(&head->lock);
 893         return ret;
 894 }
 895
 896 /*
 897  * add all inline backrefs for bytenr to the list
 898  *
 899  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 900  */
 901 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 902                            struct btrfs_path *path, u64 bytenr,
 903                            int *info_level, struct preftrees *preftrees,
 904                            u64 *total_refs, struct share_check *sc)
 905 {
 906         int ret = 0;
 907         int slot;
 908         struct extent_buffer *leaf;
 909         struct btrfs_key key;
 910         struct btrfs_key found_key;
 911         unsigned long ptr;
 912         unsigned long end;
 913         struct btrfs_extent_item *ei;
 914         u64 flags;
 915         u64 item_size;
 916
 917         /*
 918          * enumerate all inline refs
 919          */
 920         leaf = path->nodes[0];
 921         slot = path->slots[0];
 922
 923         item_size = btrfs_item_size_nr(leaf, slot);
 924         BUG_ON(item_size < sizeof(*ei));
 925
 926         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 927         flags = btrfs_extent_flags(leaf, ei);
 928         *total_refs += btrfs_extent_refs(leaf, ei);
 929         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 930
 931         ptr = (unsigned long)(ei + 1);
 932         end = (unsigned long)ei + item_size;
 933
 934         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 935             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 936                 struct btrfs_tree_block_info *info;
 937
 938                 info = (struct btrfs_tree_block_info *)ptr;
 939                 *info_level = btrfs_tree_block_level(leaf, info);
 940                 ptr += sizeof(struct btrfs_tree_block_info);
 941                 BUG_ON(ptr > end);
 942         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 943                 *info_level = found_key.offset;
 944         } else {
 945                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 946         }
 947
 948         while (ptr < end) {
 949                 struct btrfs_extent_inline_ref *iref;
 950                 u64 offset;
 951                 int type;
 952
 953                 iref = (struct btrfs_extent_inline_ref *)ptr;
 954                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 955                                                         BTRFS_REF_TYPE_ANY);
 956                 if (type == BTRFS_REF_TYPE_INVALID)
 957                         return -EUCLEAN;
 958
 959                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 960
 961                 switch (type) {
 962                 case BTRFS_SHARED_BLOCK_REF_KEY:
 963                         ret = add_direct_ref(fs_info, preftrees,
 964                                              *info_level + 1, offset,
 965                                              bytenr, 1, NULL, GFP_NOFS);
 966                         break;
 967                 case BTRFS_SHARED_DATA_REF_KEY: {
 968                         struct btrfs_shared_data_ref *sdref;
 969                         int count;
 970
 971                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 972                         count = btrfs_shared_data_ref_count(leaf, sdref);
 973
 974                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 975                                              bytenr, count, sc, GFP_NOFS);
 976                         break;
 977                 }
 978                 case BTRFS_TREE_BLOCK_REF_KEY:
 979                         ret = add_indirect_ref(fs_info, preftrees, offset,
 980                                                NULL, *info_level + 1,
 981                                                bytenr, 1, NULL, GFP_NOFS);
 982                         break;
 983                 case BTRFS_EXTENT_DATA_REF_KEY: {
 984                         struct btrfs_extent_data_ref *dref;
 985                         int count;
 986                         u64 root;
 987
 988                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 989                         count = btrfs_extent_data_ref_count(leaf, dref);
 990                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 991                                                                       dref);
 992                         key.type = BTRFS_EXTENT_DATA_KEY;
 993                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 994
 995                         if (sc && sc->inum && key.objectid != sc->inum &&
 996                             !sc->have_delayed_delete_refs) {
 997                                 ret = BACKREF_FOUND_SHARED;
 998                                 break;
 999                         }
1000
1001                         root = btrfs_extent_data_ref_root(leaf, dref);
1002
1003                         ret = add_indirect_ref(fs_info, preftrees, root,
1004                                                &key, 0, bytenr, count,
1005                                                sc, GFP_NOFS);
1006
1007                         break;
1008                 }
1009                 default:
1010                         WARN_ON(1);
1011                 }
1012                 if (ret)
1013                         return ret;
1014                 ptr += btrfs_extent_inline_ref_size(type);
1015         }
1016
1017         return 0;
1018 }
1019
1020 /*
1021  * add all non-inline backrefs for bytenr to the list
1022  *
1023  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1024  */
1025 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1026                           struct btrfs_path *path, u64 bytenr,
1027                           int info_level, struct preftrees *preftrees,
1028                           struct share_check *sc)
1029 {
1030         struct btrfs_root *extent_root = fs_info->extent_root;
1031         int ret;
1032         int slot;
1033         struct extent_buffer *leaf;
1034         struct btrfs_key key;
1035
1036         while (1) {
1037                 ret = btrfs_next_item(extent_root, path);
1038                 if (ret < 0)
1039                         break;
1040                 if (ret) {
1041                         ret = 0;
1042                         break;
1043                 }
1044
1045                 slot = path->slots[0];
1046                 leaf = path->nodes[0];
1047                 btrfs_item_key_to_cpu(leaf, &key, slot);
1048
1049                 if (key.objectid != bytenr)
1050                         break;
1051                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1052                         continue;
1053                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1054                         break;
1055
1056                 switch (key.type) {
1057                 case BTRFS_SHARED_BLOCK_REF_KEY:
1058                         /* SHARED DIRECT METADATA backref */
1059                         ret = add_direct_ref(fs_info, preftrees,
1060                                              info_level + 1, key.offset,
1061                                              bytenr, 1, NULL, GFP_NOFS);
1062                         break;
1063                 case BTRFS_SHARED_DATA_REF_KEY: {
1064                         /* SHARED DIRECT FULL backref */
1065                         struct btrfs_shared_data_ref *sdref;
1066                         int count;
1067
1068                         sdref = btrfs_item_ptr(leaf, slot,
1069                                               struct btrfs_shared_data_ref);
1070                         count = btrfs_shared_data_ref_count(leaf, sdref);
1071                         ret = add_direct_ref(fs_info, preftrees, 0,
1072                                              key.offset, bytenr, count,
1073                                              sc, GFP_NOFS);
1074                         break;
1075                 }
1076                 case BTRFS_TREE_BLOCK_REF_KEY:
1077                         /* NORMAL INDIRECT METADATA backref */
1078                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1079                                                NULL, info_level + 1, bytenr,
1080                                                1, NULL, GFP_NOFS);
1081                         break;
1082                 case BTRFS_EXTENT_DATA_REF_KEY: {
1083                         /* NORMAL INDIRECT DATA backref */
1084                         struct btrfs_extent_data_ref *dref;
1085                         int count;
1086                         u64 root;
1087
1088                         dref = btrfs_item_ptr(leaf, slot,
1089                                               struct btrfs_extent_data_ref);
1090                         count = btrfs_extent_data_ref_count(leaf, dref);
1091                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1092                                                                       dref);
1093                         key.type = BTRFS_EXTENT_DATA_KEY;
1094                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1095
1096                         if (sc && sc->inum && key.objectid != sc->inum &&
1097                             !sc->have_delayed_delete_refs) {
1098                                 ret = BACKREF_FOUND_SHARED;
1099                                 break;
1100                         }
1101
1102                         root = btrfs_extent_data_ref_root(leaf, dref);
1103                         ret = add_indirect_ref(fs_info, preftrees, root,
1104                                                &key, 0, bytenr, count,
1105                                                sc, GFP_NOFS);
1106                         break;
1107                 }
1108                 default:
1109                         WARN_ON(1);
1110                 }
1111                 if (ret)
1112                         return ret;
1113
1114         }
1115
1116         return ret;
1117 }
1118
1119 /*
1120  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1121  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1122  * indirect refs to their parent bytenr.
1123  * When roots are found, they're added to the roots list
1124  *
1125  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1126  * much like trans == NULL case, the difference only lies in it will not
1127  * commit root.
1128  * The special case is for qgroup to search roots in commit_transaction().
1129  *
1130  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1131  * shared extent is detected.
1132  *
1133  * Otherwise this returns 0 for success and <0 for an error.
1134  *
1135  * If ignore_offset is set to false, only extent refs whose offsets match
1136  * extent_item_pos are returned.  If true, every extent ref is returned
1137  * and extent_item_pos is ignored.
1138  *
1139  * FIXME some caching might speed things up
1140  */
1141 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1142                              struct btrfs_fs_info *fs_info, u64 bytenr,
1143                              u64 time_seq, struct ulist *refs,
1144                              struct ulist *roots, const u64 *extent_item_pos,
1145                              struct share_check *sc, bool ignore_offset)
1146 {
1147         struct btrfs_key key;
1148         struct btrfs_path *path;
1149         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1150         struct btrfs_delayed_ref_head *head;
1151         int info_level = 0;
1152         int ret;
1153         struct prelim_ref *ref;
1154         struct rb_node *node;
1155         struct extent_inode_elem *eie = NULL;
1156         /* total of both direct AND indirect refs! */
1157         u64 total_refs = 0;
1158         struct preftrees preftrees = {
1159                 .direct = PREFTREE_INIT,
1160                 .indirect = PREFTREE_INIT,
1161                 .indirect_missing_keys = PREFTREE_INIT
1162         };
1163
1164         key.objectid = bytenr;
1165         key.offset = (u64)-1;
1166         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1167                 key.type = BTRFS_METADATA_ITEM_KEY;
1168         else
1169                 key.type = BTRFS_EXTENT_ITEM_KEY;
1170
1171         path = btrfs_alloc_path();
1172         if (!path)
1173                 return -ENOMEM;
1174         if (!trans) {
1175                 path->search_commit_root = 1;
1176                 path->skip_locking = 1;
1177         }
1178
1179         if (time_seq == SEQ_LAST)
1180                 path->skip_locking = 1;
1181
1182         /*
1183          * grab both a lock on the path and a lock on the delayed ref head.
1184          * We need both to get a consistent picture of how the refs look
1185          * at a specified point in time
1186          */
1187 again:
1188         head = NULL;
1189
1190         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1191         if (ret < 0)
1192                 goto out;
1193         if (ret == 0) {
1194                 /* This shouldn't happen, indicates a bug or fs corruption. */
1195                 ASSERT(ret != 0);
1196                 ret = -EUCLEAN;
1197                 goto out;
1198         }
1199
1200 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1201         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1202             time_seq != SEQ_LAST) {
1203 #else
1204         if (trans && time_seq != SEQ_LAST) {
1205 #endif
1206                 /*
1207                  * look if there are updates for this ref queued and lock the
1208                  * head
1209                  */
1210                 delayed_refs = &trans->transaction->delayed_refs;
1211                 spin_lock(&delayed_refs->lock);
1212                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1213                 if (head) {
1214                         if (!mutex_trylock(&head->mutex)) {
1215                                 refcount_inc(&head->refs);
1216                                 spin_unlock(&delayed_refs->lock);
1217
1218                                 btrfs_release_path(path);
1219
1220                                 /*
1221                                  * Mutex was contended, block until it's
1222                                  * released and try again
1223                                  */
1224                                 mutex_lock(&head->mutex);
1225                                 mutex_unlock(&head->mutex);
1226                                 btrfs_put_delayed_ref_head(head);
1227                                 goto again;
1228                         }
1229                         spin_unlock(&delayed_refs->lock);
1230                         ret = add_delayed_refs(fs_info, head, time_seq,
1231                                                &preftrees, &total_refs, sc);
1232                         mutex_unlock(&head->mutex);
1233                         if (ret)
1234                                 goto out;
1235                 } else {
1236                         spin_unlock(&delayed_refs->lock);
1237                 }
1238         }
1239
1240         if (path->slots[0]) {
1241                 struct extent_buffer *leaf;
1242                 int slot;
1243
1244                 path->slots[0]--;
1245                 leaf = path->nodes[0];
1246                 slot = path->slots[0];
1247                 btrfs_item_key_to_cpu(leaf, &key, slot);
1248                 if (key.objectid == bytenr &&
1249                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1250                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1251                         ret = add_inline_refs(fs_info, path, bytenr,
1252                                               &info_level, &preftrees,
1253                                               &total_refs, sc);
1254                         if (ret)
1255                                 goto out;
1256                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1257                                              &preftrees, sc);
1258                         if (ret)
1259                                 goto out;
1260                 }
1261         }
1262
1263         btrfs_release_path(path);
1264
1265         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1266         if (ret)
1267                 goto out;
1268
1269         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root));
1270
1271         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1272                                     extent_item_pos, total_refs, sc, ignore_offset);
1273         if (ret)
1274                 goto out;
1275
1276         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root));
1277
1278         /*
1279          * This walks the tree of merged and resolved refs. Tree blocks are
1280          * read in as needed. Unique entries are added to the ulist, and
1281          * the list of found roots is updated.
1282          *
1283          * We release the entire tree in one go before returning.
1284          */
1285         node = rb_first(&preftrees.direct.root);
1286         while (node) {
1287                 ref = rb_entry(node, struct prelim_ref, rbnode);
1288                 node = rb_next(&ref->rbnode);
1289                 /*
1290                  * ref->count < 0 can happen here if there are delayed
1291                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1292                  * prelim_ref_insert() relies on this when merging
1293                  * identical refs to keep the overall count correct.
1294                  * prelim_ref_insert() will merge only those refs
1295                  * which compare identically.  Any refs having
1296                  * e.g. different offsets would not be merged,
1297                  * and would retain their original ref->count < 0.
1298                  */
1299                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1300                         if (sc && sc->root_objectid &&
1301                             ref->root_id != sc->root_objectid) {
1302                                 ret = BACKREF_FOUND_SHARED;
1303                                 goto out;
1304                         }
1305
1306                         /* no parent == root of tree */
1307                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1308                         if (ret < 0)
1309                                 goto out;
1310                 }
1311                 if (ref->count && ref->parent) {
1312                         if (extent_item_pos && !ref->inode_list &&
1313                             ref->level == 0) {
1314                                 struct extent_buffer *eb;
1315
1316                                 eb = read_tree_block(fs_info, ref->parent, 0,
1317                                                      ref->level, NULL);
1318                                 if (IS_ERR(eb)) {
1319                                         ret = PTR_ERR(eb);
1320                                         goto out;
1321                                 } else if (!extent_buffer_uptodate(eb)) {
1322                                         free_extent_buffer(eb);
1323                                         ret = -EIO;
1324                                         goto out;
1325                                 }
1326                                 if (!path->skip_locking) {
1327                                         btrfs_tree_read_lock(eb);
1328                                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1329                                 }
1330                                 ret = find_extent_in_eb(eb, bytenr,
1331                                                         *extent_item_pos, &eie, ignore_offset);
1332                                 if (!path->skip_locking)
1333                                         btrfs_tree_read_unlock_blocking(eb);
1334                                 free_extent_buffer(eb);
1335                                 if (ret < 0)
1336                                         goto out;
1337                                 ref->inode_list = eie;
1338                         }
1339                         ret = ulist_add_merge_ptr(refs, ref->parent,
1340                                                   ref->inode_list,
1341                                                   (void **)&eie, GFP_NOFS);
1342                         if (ret < 0)
1343                                 goto out;
1344                         if (!ret && extent_item_pos) {
1345                                 /*
1346                                  * We've recorded that parent, so we must extend
1347                                  * its inode list here.
1348                                  *
1349                                  * However if there was corruption we may not
1350                                  * have found an eie, return an error in this
1351                                  * case.
1352                                  */
1353                                 ASSERT(eie);
1354                                 if (!eie) {
1355                                         ret = -EUCLEAN;
1356                                         goto out;
1357                                 }
1358                                 while (eie->next)
1359                                         eie = eie->next;
1360                                 eie->next = ref->inode_list;
1361                         }
1362                         eie = NULL;
1363                 }
1364                 cond_resched();
1365         }
1366
1367 out:
1368         btrfs_free_path(path);
1369
1370         prelim_release(&preftrees.direct);
1371         prelim_release(&preftrees.indirect);
1372         prelim_release(&preftrees.indirect_missing_keys);
1373
1374         if (ret < 0)
1375                 free_inode_elem_list(eie);
1376         return ret;
1377 }
1378
1379 /*
1380  * Finds all leafs with a reference to the specified combination of bytenr and
1381  * offset. key_list_head will point to a list of corresponding keys (caller must
1382  * free each list element). The leafs will be stored in the leafs ulist, which
1383  * must be freed with ulist_free.
1384  *
1385  * returns 0 on success, <0 on error
1386  */
1387 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1388                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1389                                 u64 time_seq, struct ulist **leafs,
1390                                 const u64 *extent_item_pos, bool ignore_offset)
1391 {
1392         int ret;
1393
1394         *leafs = ulist_alloc(GFP_NOFS);
1395         if (!*leafs)
1396                 return -ENOMEM;
1397
1398         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1399                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1400         if (ret < 0 && ret != -ENOENT) {
1401                 free_leaf_list(*leafs);
1402                 return ret;
1403         }
1404
1405         return 0;
1406 }
1407
1408 /*
1409  * walk all backrefs for a given extent to find all roots that reference this
1410  * extent. Walking a backref means finding all extents that reference this
1411  * extent and in turn walk the backrefs of those, too. Naturally this is a
1412  * recursive process, but here it is implemented in an iterative fashion: We
1413  * find all referencing extents for the extent in question and put them on a
1414  * list. In turn, we find all referencing extents for those, further appending
1415  * to the list. The way we iterate the list allows adding more elements after
1416  * the current while iterating. The process stops when we reach the end of the
1417  * list. Found roots are added to the roots list.
1418  *
1419  * returns 0 on success, < 0 on error.
1420  */
1421 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1422                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1423                                      u64 time_seq, struct ulist **roots,
1424                                      bool ignore_offset)
1425 {
1426         struct ulist *tmp;
1427         struct ulist_node *node = NULL;
1428         struct ulist_iterator uiter;
1429         int ret;
1430
1431         tmp = ulist_alloc(GFP_NOFS);
1432         if (!tmp)
1433                 return -ENOMEM;
1434         *roots = ulist_alloc(GFP_NOFS);
1435         if (!*roots) {
1436                 ulist_free(tmp);
1437                 return -ENOMEM;
1438         }
1439
1440         ULIST_ITER_INIT(&uiter);
1441         while (1) {
1442                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1443                                         tmp, *roots, NULL, NULL, ignore_offset);
1444                 if (ret < 0 && ret != -ENOENT) {
1445                         ulist_free(tmp);
1446                         ulist_free(*roots);
1447                         *roots = NULL;
1448                         return ret;
1449                 }
1450                 node = ulist_next(tmp, &uiter);
1451                 if (!node)
1452                         break;
1453                 bytenr = node->val;
1454                 cond_resched();
1455         }
1456
1457         ulist_free(tmp);
1458         return 0;
1459 }
1460
1461 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1462                          struct btrfs_fs_info *fs_info, u64 bytenr,
1463                          u64 time_seq, struct ulist **roots,
1464                          bool ignore_offset)
1465 {
1466         int ret;
1467
1468         if (!trans)
1469                 down_read(&fs_info->commit_root_sem);
1470         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1471                                         time_seq, roots, ignore_offset);
1472         if (!trans)
1473                 up_read(&fs_info->commit_root_sem);
1474         return ret;
1475 }
1476
1477 /**
1478  * btrfs_check_shared - tell us whether an extent is shared
1479  *
1480  * btrfs_check_shared uses the backref walking code but will short
1481  * circuit as soon as it finds a root or inode that doesn't match the
1482  * one passed in. This provides a significant performance benefit for
1483  * callers (such as fiemap) which want to know whether the extent is
1484  * shared but do not need a ref count.
1485  *
1486  * This attempts to attach to the running transaction in order to account for
1487  * delayed refs, but continues on even when no running transaction exists.
1488  *
1489  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1490  */
1491 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1492 {
1493         struct btrfs_fs_info *fs_info = root->fs_info;
1494         struct btrfs_trans_handle *trans;
1495         struct ulist *tmp = NULL;
1496         struct ulist *roots = NULL;
1497         struct ulist_iterator uiter;
1498         struct ulist_node *node;
1499         struct seq_list elem = SEQ_LIST_INIT(elem);
1500         int ret = 0;
1501         struct share_check shared = {
1502                 .root_objectid = root->objectid,
1503                 .inum = inum,
1504                 .share_count = 0,
1505                 .have_delayed_delete_refs = false,
1506         };
1507
1508         tmp = ulist_alloc(GFP_NOFS);
1509         roots = ulist_alloc(GFP_NOFS);
1510         if (!tmp || !roots) {
1511                 ret = -ENOMEM;
1512                 goto out;
1513         }
1514
1515         trans = btrfs_join_transaction_nostart(root);
1516         if (IS_ERR(trans)) {
1517                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1518                         ret = PTR_ERR(trans);
1519                         goto out;
1520                 }
1521                 trans = NULL;
1522                 down_read(&fs_info->commit_root_sem);
1523         } else {
1524                 btrfs_get_tree_mod_seq(fs_info, &elem);
1525         }
1526
1527         ULIST_ITER_INIT(&uiter);
1528         while (1) {
1529                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1530                                         roots, NULL, &shared, false);
1531                 if (ret == BACKREF_FOUND_SHARED) {
1532                         /* this is the only condition under which we return 1 */
1533                         ret = 1;
1534                         break;
1535                 }
1536                 if (ret < 0 && ret != -ENOENT)
1537                         break;
1538                 ret = 0;
1539                 node = ulist_next(tmp, &uiter);
1540                 if (!node)
1541                         break;
1542                 bytenr = node->val;
1543                 shared.share_count = 0;
1544                 shared.have_delayed_delete_refs = false;
1545                 cond_resched();
1546         }
1547
1548         if (trans) {
1549                 btrfs_put_tree_mod_seq(fs_info, &elem);
1550                 btrfs_end_transaction(trans);
1551         } else {
1552                 up_read(&fs_info->commit_root_sem);
1553         }
1554 out:
1555         ulist_free(tmp);
1556         ulist_free(roots);
1557         return ret;
1558 }
1559
1560 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1561                           u64 start_off, struct btrfs_path *path,
1562                           struct btrfs_inode_extref **ret_extref,
1563                           u64 *found_off)
1564 {
1565         int ret, slot;
1566         struct btrfs_key key;
1567         struct btrfs_key found_key;
1568         struct btrfs_inode_extref *extref;
1569         const struct extent_buffer *leaf;
1570         unsigned long ptr;
1571
1572         key.objectid = inode_objectid;
1573         key.type = BTRFS_INODE_EXTREF_KEY;
1574         key.offset = start_off;
1575
1576         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1577         if (ret < 0)
1578                 return ret;
1579
1580         while (1) {
1581                 leaf = path->nodes[0];
1582                 slot = path->slots[0];
1583                 if (slot >= btrfs_header_nritems(leaf)) {
1584                         /*
1585                          * If the item at offset is not found,
1586                          * btrfs_search_slot will point us to the slot
1587                          * where it should be inserted. In our case
1588                          * that will be the slot directly before the
1589                          * next INODE_REF_KEY_V2 item. In the case
1590                          * that we're pointing to the last slot in a
1591                          * leaf, we must move one leaf over.
1592                          */
1593                         ret = btrfs_next_leaf(root, path);
1594                         if (ret) {
1595                                 if (ret >= 1)
1596                                         ret = -ENOENT;
1597                                 break;
1598                         }
1599                         continue;
1600                 }
1601
1602                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1603
1604                 /*
1605                  * Check that we're still looking at an extended ref key for
1606                  * this particular objectid. If we have different
1607                  * objectid or type then there are no more to be found
1608                  * in the tree and we can exit.
1609                  */
1610                 ret = -ENOENT;
1611                 if (found_key.objectid != inode_objectid)
1612                         break;
1613                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1614                         break;
1615
1616                 ret = 0;
1617                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1618                 extref = (struct btrfs_inode_extref *)ptr;
1619                 *ret_extref = extref;
1620                 if (found_off)
1621                         *found_off = found_key.offset;
1622                 break;
1623         }
1624
1625         return ret;
1626 }
1627
1628 /*
1629  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1630  * Elements of the path are separated by '/' and the path is guaranteed to be
1631  * 0-terminated. the path is only given within the current file system.
1632  * Therefore, it never starts with a '/'. the caller is responsible to provide
1633  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1634  * the start point of the resulting string is returned. this pointer is within
1635  * dest, normally.
1636  * in case the path buffer would overflow, the pointer is decremented further
1637  * as if output was written to the buffer, though no more output is actually
1638  * generated. that way, the caller can determine how much space would be
1639  * required for the path to fit into the buffer. in that case, the returned
1640  * value will be smaller than dest. callers must check this!
1641  */
1642 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1643                         u32 name_len, unsigned long name_off,
1644                         struct extent_buffer *eb_in, u64 parent,
1645                         char *dest, u32 size)
1646 {
1647         int slot;
1648         u64 next_inum;
1649         int ret;
1650         s64 bytes_left = ((s64)size) - 1;
1651         struct extent_buffer *eb = eb_in;
1652         struct btrfs_key found_key;
1653         int leave_spinning = path->leave_spinning;
1654         struct btrfs_inode_ref *iref;
1655
1656         if (bytes_left >= 0)
1657                 dest[bytes_left] = '\0';
1658
1659         path->leave_spinning = 1;
1660         while (1) {
1661                 bytes_left -= name_len;
1662                 if (bytes_left >= 0)
1663                         read_extent_buffer(eb, dest + bytes_left,
1664                                            name_off, name_len);
1665                 if (eb != eb_in) {
1666                         if (!path->skip_locking)
1667                                 btrfs_tree_read_unlock_blocking(eb);
1668                         free_extent_buffer(eb);
1669                 }
1670                 ret = btrfs_find_item(fs_root, path, parent, 0,
1671                                 BTRFS_INODE_REF_KEY, &found_key);
1672                 if (ret > 0)
1673                         ret = -ENOENT;
1674                 if (ret)
1675                         break;
1676
1677                 next_inum = found_key.offset;
1678
1679                 /* regular exit ahead */
1680                 if (parent == next_inum)
1681                         break;
1682
1683                 slot = path->slots[0];
1684                 eb = path->nodes[0];
1685                 /* make sure we can use eb after releasing the path */
1686                 if (eb != eb_in) {
1687                         if (!path->skip_locking)
1688                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1689                         path->nodes[0] = NULL;
1690                         path->locks[0] = 0;
1691                 }
1692                 btrfs_release_path(path);
1693                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1694
1695                 name_len = btrfs_inode_ref_name_len(eb, iref);
1696                 name_off = (unsigned long)(iref + 1);
1697
1698                 parent = next_inum;
1699                 --bytes_left;
1700                 if (bytes_left >= 0)
1701                         dest[bytes_left] = '/';
1702         }
1703
1704         btrfs_release_path(path);
1705         path->leave_spinning = leave_spinning;
1706
1707         if (ret)
1708                 return ERR_PTR(ret);
1709
1710         return dest + bytes_left;
1711 }
1712
1713 /*
1714  * this makes the path point to (logical EXTENT_ITEM *)
1715  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1716  * tree blocks and <0 on error.
1717  */
1718 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1719                         struct btrfs_path *path, struct btrfs_key *found_key,
1720                         u64 *flags_ret)
1721 {
1722         int ret;
1723         u64 flags;
1724         u64 size = 0;
1725         u32 item_size;
1726         const struct extent_buffer *eb;
1727         struct btrfs_extent_item *ei;
1728         struct btrfs_key key;
1729
1730         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1731                 key.type = BTRFS_METADATA_ITEM_KEY;
1732         else
1733                 key.type = BTRFS_EXTENT_ITEM_KEY;
1734         key.objectid = logical;
1735         key.offset = (u64)-1;
1736
1737         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1738         if (ret < 0)
1739                 return ret;
1740
1741         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1742         if (ret) {
1743                 if (ret > 0)
1744                         ret = -ENOENT;
1745                 return ret;
1746         }
1747         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1748         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1749                 size = fs_info->nodesize;
1750         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1751                 size = found_key->offset;
1752
1753         if (found_key->objectid > logical ||
1754             found_key->objectid + size <= logical) {
1755                 btrfs_debug(fs_info,
1756                         "logical %llu is not within any extent", logical);
1757                 return -ENOENT;
1758         }
1759
1760         eb = path->nodes[0];
1761         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1762         BUG_ON(item_size < sizeof(*ei));
1763
1764         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1765         flags = btrfs_extent_flags(eb, ei);
1766
1767         btrfs_debug(fs_info,
1768                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1769                  logical, logical - found_key->objectid, found_key->objectid,
1770                  found_key->offset, flags, item_size);
1771
1772         WARN_ON(!flags_ret);
1773         if (flags_ret) {
1774                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1775                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1776                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1777                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1778                 else
1779                         BUG_ON(1);
1780                 return 0;
1781         }
1782
1783         return -EIO;
1784 }
1785
1786 /*
1787  * helper function to iterate extent inline refs. ptr must point to a 0 value
1788  * for the first call and may be modified. it is used to track state.
1789  * if more refs exist, 0 is returned and the next call to
1790  * get_extent_inline_ref must pass the modified ptr parameter to get the
1791  * next ref. after the last ref was processed, 1 is returned.
1792  * returns <0 on error
1793  */
1794 static int get_extent_inline_ref(unsigned long *ptr,
1795                                  const struct extent_buffer *eb,
1796                                  const struct btrfs_key *key,
1797                                  const struct btrfs_extent_item *ei,
1798                                  u32 item_size,
1799                                  struct btrfs_extent_inline_ref **out_eiref,
1800                                  int *out_type)
1801 {
1802         unsigned long end;
1803         u64 flags;
1804         struct btrfs_tree_block_info *info;
1805
1806         if (!*ptr) {
1807                 /* first call */
1808                 flags = btrfs_extent_flags(eb, ei);
1809                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1810                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1811                                 /* a skinny metadata extent */
1812                                 *out_eiref =
1813                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1814                         } else {
1815                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1816                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1817                                 *out_eiref =
1818                                    (struct btrfs_extent_inline_ref *)(info + 1);
1819                         }
1820                 } else {
1821                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1822                 }
1823                 *ptr = (unsigned long)*out_eiref;
1824                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1825                         return -ENOENT;
1826         }
1827
1828         end = (unsigned long)ei + item_size;
1829         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1830         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1831                                                      BTRFS_REF_TYPE_ANY);
1832         if (*out_type == BTRFS_REF_TYPE_INVALID)
1833                 return -EUCLEAN;
1834
1835         *ptr += btrfs_extent_inline_ref_size(*out_type);
1836         WARN_ON(*ptr > end);
1837         if (*ptr == end)
1838                 return 1; /* last */
1839
1840         return 0;
1841 }
1842
1843 /*
1844  * reads the tree block backref for an extent. tree level and root are returned
1845  * through out_level and out_root. ptr must point to a 0 value for the first
1846  * call and may be modified (see get_extent_inline_ref comment).
1847  * returns 0 if data was provided, 1 if there was no more data to provide or
1848  * <0 on error.
1849  */
1850 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1851                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1852                             u32 item_size, u64 *out_root, u8 *out_level)
1853 {
1854         int ret;
1855         int type;
1856         struct btrfs_extent_inline_ref *eiref;
1857
1858         if (*ptr == (unsigned long)-1)
1859                 return 1;
1860
1861         while (1) {
1862                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1863                                               &eiref, &type);
1864                 if (ret < 0)
1865                         return ret;
1866
1867                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1868                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1869                         break;
1870
1871                 if (ret == 1)
1872                         return 1;
1873         }
1874
1875         /* we can treat both ref types equally here */
1876         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1877
1878         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1879                 struct btrfs_tree_block_info *info;
1880
1881                 info = (struct btrfs_tree_block_info *)(ei + 1);
1882                 *out_level = btrfs_tree_block_level(eb, info);
1883         } else {
1884                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1885                 *out_level = (u8)key->offset;
1886         }
1887
1888         if (ret == 1)
1889                 *ptr = (unsigned long)-1;
1890
1891         return 0;
1892 }
1893
1894 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1895                              struct extent_inode_elem *inode_list,
1896                              u64 root, u64 extent_item_objectid,
1897                              iterate_extent_inodes_t *iterate, void *ctx)
1898 {
1899         struct extent_inode_elem *eie;
1900         int ret = 0;
1901
1902         for (eie = inode_list; eie; eie = eie->next) {
1903                 btrfs_debug(fs_info,
1904                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1905                             extent_item_objectid, eie->inum,
1906                             eie->offset, root);
1907                 ret = iterate(eie->inum, eie->offset, root, ctx);
1908                 if (ret) {
1909                         btrfs_debug(fs_info,
1910                                     "stopping iteration for %llu due to ret=%d",
1911                                     extent_item_objectid, ret);
1912                         break;
1913                 }
1914         }
1915
1916         return ret;
1917 }
1918
1919 /*
1920  * calls iterate() for every inode that references the extent identified by
1921  * the given parameters.
1922  * when the iterator function returns a non-zero value, iteration stops.
1923  */
1924 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1925                                 u64 extent_item_objectid, u64 extent_item_pos,
1926                                 int search_commit_root,
1927                                 iterate_extent_inodes_t *iterate, void *ctx,
1928                                 bool ignore_offset)
1929 {
1930         int ret;
1931         struct btrfs_trans_handle *trans = NULL;
1932         struct ulist *refs = NULL;
1933         struct ulist *roots = NULL;
1934         struct ulist_node *ref_node = NULL;
1935         struct ulist_node *root_node = NULL;
1936         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1937         struct ulist_iterator ref_uiter;
1938         struct ulist_iterator root_uiter;
1939
1940         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1941                         extent_item_objectid);
1942
1943         if (!search_commit_root) {
1944                 trans = btrfs_attach_transaction(fs_info->extent_root);
1945                 if (IS_ERR(trans)) {
1946                         if (PTR_ERR(trans) != -ENOENT &&
1947                             PTR_ERR(trans) != -EROFS)
1948                                 return PTR_ERR(trans);
1949                         trans = NULL;
1950                 }
1951         }
1952
1953         if (trans)
1954                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1955         else
1956                 down_read(&fs_info->commit_root_sem);
1957
1958         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1959                                    tree_mod_seq_elem.seq, &refs,
1960                                    &extent_item_pos, ignore_offset);
1961         if (ret)
1962                 goto out;
1963
1964         ULIST_ITER_INIT(&ref_uiter);
1965         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1966                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1967                                                 tree_mod_seq_elem.seq, &roots,
1968                                                 ignore_offset);
1969                 if (ret)
1970                         break;
1971                 ULIST_ITER_INIT(&root_uiter);
1972                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1973                         btrfs_debug(fs_info,
1974                                     "root %llu references leaf %llu, data list %#llx",
1975                                     root_node->val, ref_node->val,
1976                                     ref_node->aux);
1977                         ret = iterate_leaf_refs(fs_info,
1978                                                 (struct extent_inode_elem *)
1979                                                 (uintptr_t)ref_node->aux,
1980                                                 root_node->val,
1981                                                 extent_item_objectid,
1982                                                 iterate, ctx);
1983                 }
1984                 ulist_free(roots);
1985         }
1986
1987         free_leaf_list(refs);
1988 out:
1989         if (trans) {
1990                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1991                 btrfs_end_transaction(trans);
1992         } else {
1993                 up_read(&fs_info->commit_root_sem);
1994         }
1995
1996         return ret;
1997 }
1998
1999 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2000                                 struct btrfs_path *path,
2001                                 iterate_extent_inodes_t *iterate, void *ctx,
2002                                 bool ignore_offset)
2003 {
2004         int ret;
2005         u64 extent_item_pos;
2006         u64 flags = 0;
2007         struct btrfs_key found_key;
2008         int search_commit_root = path->search_commit_root;
2009
2010         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2011         btrfs_release_path(path);
2012         if (ret < 0)
2013                 return ret;
2014         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2015                 return -EINVAL;
2016
2017         extent_item_pos = logical - found_key.objectid;
2018         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2019                                         extent_item_pos, search_commit_root,
2020                                         iterate, ctx, ignore_offset);
2021
2022         return ret;
2023 }
2024
2025 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2026                               struct extent_buffer *eb, void *ctx);
2027
2028 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2029                               struct btrfs_path *path,
2030                               iterate_irefs_t *iterate, void *ctx)
2031 {
2032         int ret = 0;
2033         int slot;
2034         u32 cur;
2035         u32 len;
2036         u32 name_len;
2037         u64 parent = 0;
2038         int found = 0;
2039         struct extent_buffer *eb;
2040         struct btrfs_item *item;
2041         struct btrfs_inode_ref *iref;
2042         struct btrfs_key found_key;
2043
2044         while (!ret) {
2045                 ret = btrfs_find_item(fs_root, path, inum,
2046                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2047                                 &found_key);
2048
2049                 if (ret < 0)
2050                         break;
2051                 if (ret) {
2052                         ret = found ? 0 : -ENOENT;
2053                         break;
2054                 }
2055                 ++found;
2056
2057                 parent = found_key.offset;
2058                 slot = path->slots[0];
2059                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2060                 if (!eb) {
2061                         ret = -ENOMEM;
2062                         break;
2063                 }
2064                 extent_buffer_get(eb);
2065                 btrfs_tree_read_lock(eb);
2066                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2067                 btrfs_release_path(path);
2068
2069                 item = btrfs_item_nr(slot);
2070                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2071
2072                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2073                         name_len = btrfs_inode_ref_name_len(eb, iref);
2074                         /* path must be released before calling iterate()! */
2075                         btrfs_debug(fs_root->fs_info,
2076                                 "following ref at offset %u for inode %llu in tree %llu",
2077                                 cur, found_key.objectid, fs_root->objectid);
2078                         ret = iterate(parent, name_len,
2079                                       (unsigned long)(iref + 1), eb, ctx);
2080                         if (ret)
2081                                 break;
2082                         len = sizeof(*iref) + name_len;
2083                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2084                 }
2085                 btrfs_tree_read_unlock_blocking(eb);
2086                 free_extent_buffer(eb);
2087         }
2088
2089         btrfs_release_path(path);
2090
2091         return ret;
2092 }
2093
2094 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2095                                  struct btrfs_path *path,
2096                                  iterate_irefs_t *iterate, void *ctx)
2097 {
2098         int ret;
2099         int slot;
2100         u64 offset = 0;
2101         u64 parent;
2102         int found = 0;
2103         struct extent_buffer *eb;
2104         struct btrfs_inode_extref *extref;
2105         u32 item_size;
2106         u32 cur_offset;
2107         unsigned long ptr;
2108
2109         while (1) {
2110                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2111                                             &offset);
2112                 if (ret < 0)
2113                         break;
2114                 if (ret) {
2115                         ret = found ? 0 : -ENOENT;
2116                         break;
2117                 }
2118                 ++found;
2119
2120                 slot = path->slots[0];
2121                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2122                 if (!eb) {
2123                         ret = -ENOMEM;
2124                         break;
2125                 }
2126                 extent_buffer_get(eb);
2127
2128                 btrfs_tree_read_lock(eb);
2129                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2130                 btrfs_release_path(path);
2131
2132                 item_size = btrfs_item_size_nr(eb, slot);
2133                 ptr = btrfs_item_ptr_offset(eb, slot);
2134                 cur_offset = 0;
2135
2136                 while (cur_offset < item_size) {
2137                         u32 name_len;
2138
2139                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2140                         parent = btrfs_inode_extref_parent(eb, extref);
2141                         name_len = btrfs_inode_extref_name_len(eb, extref);
2142                         ret = iterate(parent, name_len,
2143                                       (unsigned long)&extref->name, eb, ctx);
2144                         if (ret)
2145                                 break;
2146
2147                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2148                         cur_offset += sizeof(*extref);
2149                 }
2150                 btrfs_tree_read_unlock_blocking(eb);
2151                 free_extent_buffer(eb);
2152
2153                 offset++;
2154         }
2155
2156         btrfs_release_path(path);
2157
2158         return ret;
2159 }
2160
2161 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2162                          struct btrfs_path *path, iterate_irefs_t *iterate,
2163                          void *ctx)
2164 {
2165         int ret;
2166         int found_refs = 0;
2167
2168         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2169         if (!ret)
2170                 ++found_refs;
2171         else if (ret != -ENOENT)
2172                 return ret;
2173
2174         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2175         if (ret == -ENOENT && found_refs)
2176                 return 0;
2177
2178         return ret;
2179 }
2180
2181 /*
2182  * returns 0 if the path could be dumped (probably truncated)
2183  * returns <0 in case of an error
2184  */
2185 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2186                          struct extent_buffer *eb, void *ctx)
2187 {
2188         struct inode_fs_paths *ipath = ctx;
2189         char *fspath;
2190         char *fspath_min;
2191         int i = ipath->fspath->elem_cnt;
2192         const int s_ptr = sizeof(char *);
2193         u32 bytes_left;
2194
2195         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2196                                         ipath->fspath->bytes_left - s_ptr : 0;
2197
2198         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2199         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2200                                    name_off, eb, inum, fspath_min, bytes_left);
2201         if (IS_ERR(fspath))
2202                 return PTR_ERR(fspath);
2203
2204         if (fspath > fspath_min) {
2205                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2206                 ++ipath->fspath->elem_cnt;
2207                 ipath->fspath->bytes_left = fspath - fspath_min;
2208         } else {
2209                 ++ipath->fspath->elem_missed;
2210                 ipath->fspath->bytes_missing += fspath_min - fspath;
2211                 ipath->fspath->bytes_left = 0;
2212         }
2213
2214         return 0;
2215 }
2216
2217 /*
2218  * this dumps all file system paths to the inode into the ipath struct, provided
2219  * is has been created large enough. each path is zero-terminated and accessed
2220  * from ipath->fspath->val[i].
2221  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2222  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2223  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2224  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2225  * have been needed to return all paths.
2226  */
2227 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2228 {
2229         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2230                              inode_to_path, ipath);
2231 }
2232
2233 struct btrfs_data_container *init_data_container(u32 total_bytes)
2234 {
2235         struct btrfs_data_container *data;
2236         size_t alloc_bytes;
2237
2238         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2239         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2240         if (!data)
2241                 return ERR_PTR(-ENOMEM);
2242
2243         if (total_bytes >= sizeof(*data)) {
2244                 data->bytes_left = total_bytes - sizeof(*data);
2245                 data->bytes_missing = 0;
2246         } else {
2247                 data->bytes_missing = sizeof(*data) - total_bytes;
2248                 data->bytes_left = 0;
2249         }
2250
2251         data->elem_cnt = 0;
2252         data->elem_missed = 0;
2253
2254         return data;
2255 }
2256
2257 /*
2258  * allocates space to return multiple file system paths for an inode.
2259  * total_bytes to allocate are passed, note that space usable for actual path
2260  * information will be total_bytes - sizeof(struct inode_fs_paths).
2261  * the returned pointer must be freed with free_ipath() in the end.
2262  */
2263 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2264                                         struct btrfs_path *path)
2265 {
2266         struct inode_fs_paths *ifp;
2267         struct btrfs_data_container *fspath;
2268
2269         fspath = init_data_container(total_bytes);
2270         if (IS_ERR(fspath))
2271                 return ERR_CAST(fspath);
2272
2273         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2274         if (!ifp) {
2275                 kvfree(fspath);
2276                 return ERR_PTR(-ENOMEM);
2277         }
2278
2279         ifp->btrfs_path = path;
2280         ifp->fspath = fspath;
2281         ifp->fs_root = fs_root;
2282
2283         return ifp;
2284 }
2285
2286 void free_ipath(struct inode_fs_paths *ipath)
2287 {
2288         if (!ipath)
2289                 return;
2290         kvfree(ipath->fspath);
2291         kfree(ipath);
2292 }