fs/btrfs/backref.c

   1 /*
   2  * Copyright (C) 2011 STRATO.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/mm.h>
  20 #include <linux/rbtree.h>
  21 #include <trace/events/btrfs.h>
  22 #include "ctree.h"
  23 #include "disk-io.h"
  24 #include "backref.h"
  25 #include "ulist.h"
  26 #include "transaction.h"
  27 #include "delayed-ref.h"
  28 #include "locking.h"
  29
  30 /* Just an arbitrary number so we can be sure this happened */
  31 #define BACKREF_FOUND_SHARED 6
  32
  33 struct extent_inode_elem {
  34         u64 inum;
  35         u64 offset;
  36         struct extent_inode_elem *next;
  37 };
  38
  39 static int check_extent_in_eb(const struct btrfs_key *key,
  40                               const struct extent_buffer *eb,
  41                               const struct btrfs_file_extent_item *fi,
  42                               u64 extent_item_pos,
  43                               struct extent_inode_elem **eie)
  44 {
  45         u64 offset = 0;
  46         struct extent_inode_elem *e;
  47
  48         if (!btrfs_file_extent_compression(eb, fi) &&
  49             !btrfs_file_extent_encryption(eb, fi) &&
  50             !btrfs_file_extent_other_encoding(eb, fi)) {
  51                 u64 data_offset;
  52                 u64 data_len;
  53
  54                 data_offset = btrfs_file_extent_offset(eb, fi);
  55                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  56
  57                 if (extent_item_pos < data_offset ||
  58                     extent_item_pos >= data_offset + data_len)
  59                         return 1;
  60                 offset = extent_item_pos - data_offset;
  61         }
  62
  63         e = kmalloc(sizeof(*e), GFP_NOFS);
  64         if (!e)
  65                 return -ENOMEM;
  66
  67         e->next = *eie;
  68         e->inum = key->objectid;
  69         e->offset = key->offset + offset;
  70         *eie = e;
  71
  72         return 0;
  73 }
  74
  75 static void free_inode_elem_list(struct extent_inode_elem *eie)
  76 {
  77         struct extent_inode_elem *eie_next;
  78
  79         for (; eie; eie = eie_next) {
  80                 eie_next = eie->next;
  81                 kfree(eie);
  82         }
  83 }
  84
  85 static int find_extent_in_eb(const struct extent_buffer *eb,
  86                              u64 wanted_disk_byte, u64 extent_item_pos,
  87                              struct extent_inode_elem **eie)
  88 {
  89         u64 disk_byte;
  90         struct btrfs_key key;
  91         struct btrfs_file_extent_item *fi;
  92         int slot;
  93         int nritems;
  94         int extent_type;
  95         int ret;
  96
  97         /*
  98          * from the shared data ref, we only have the leaf but we need
  99          * the key. thus, we must look into all items and see that we
 100          * find one (some) with a reference to our extent item.
 101          */
 102         nritems = btrfs_header_nritems(eb);
 103         for (slot = 0; slot < nritems; ++slot) {
 104                 btrfs_item_key_to_cpu(eb, &key, slot);
 105                 if (key.type != BTRFS_EXTENT_DATA_KEY)
 106                         continue;
 107                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 108                 extent_type = btrfs_file_extent_type(eb, fi);
 109                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 110                         continue;
 111                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 112                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 113                 if (disk_byte != wanted_disk_byte)
 114                         continue;
 115
 116                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
 117                 if (ret < 0)
 118                         return ret;
 119         }
 120
 121         return 0;
 122 }
 123
 124 struct preftree {
 125         struct rb_root root;
 126         unsigned int count;
 127 };
 128
 129 #define PREFTREE_INIT   { .root = RB_ROOT, .count = 0 }
 130
 131 struct preftrees {
 132         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 133         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 134         struct preftree indirect_missing_keys;
 135 };
 136
 137 /*
 138  * Checks for a shared extent during backref search.
 139  *
 140  * The share_count tracks prelim_refs (direct and indirect) having a
 141  * ref->count >0:
 142  *  - incremented when a ref->count transitions to >0
 143  *  - decremented when a ref->count transitions to <1
 144  */
 145 struct share_check {
 146         u64 root_objectid;
 147         u64 inum;
 148         int share_count;
 149         bool have_delayed_delete_refs;
 150 };
 151
 152 static inline int extent_is_shared(struct share_check *sc)
 153 {
 154         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 155 }
 156
 157 static struct kmem_cache *btrfs_prelim_ref_cache;
 158
 159 int __init btrfs_prelim_ref_init(void)
 160 {
 161         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 162                                         sizeof(struct prelim_ref),
 163                                         0,
 164                                         SLAB_MEM_SPREAD,
 165                                         NULL);
 166         if (!btrfs_prelim_ref_cache)
 167                 return -ENOMEM;
 168         return 0;
 169 }
 170
 171 void btrfs_prelim_ref_exit(void)
 172 {
 173         kmem_cache_destroy(btrfs_prelim_ref_cache);
 174 }
 175
 176 static void free_pref(struct prelim_ref *ref)
 177 {
 178         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 179 }
 180
 181 /*
 182  * Return 0 when both refs are for the same block (and can be merged).
 183  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 184  * indicates a 'higher' block.
 185  */
 186 static int prelim_ref_compare(struct prelim_ref *ref1,
 187                               struct prelim_ref *ref2)
 188 {
 189         if (ref1->level < ref2->level)
 190                 return -1;
 191         if (ref1->level > ref2->level)
 192                 return 1;
 193         if (ref1->root_id < ref2->root_id)
 194                 return -1;
 195         if (ref1->root_id > ref2->root_id)
 196                 return 1;
 197         if (ref1->key_for_search.type < ref2->key_for_search.type)
 198                 return -1;
 199         if (ref1->key_for_search.type > ref2->key_for_search.type)
 200                 return 1;
 201         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 202                 return -1;
 203         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 204                 return 1;
 205         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 206                 return -1;
 207         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 208                 return 1;
 209         if (ref1->parent < ref2->parent)
 210                 return -1;
 211         if (ref1->parent > ref2->parent)
 212                 return 1;
 213
 214         return 0;
 215 }
 216
 217 void update_share_count(struct share_check *sc, int oldcount, int newcount)
 218 {
 219         if ((!sc) || (oldcount == 0 && newcount < 1))
 220                 return;
 221
 222         if (oldcount > 0 && newcount < 1)
 223                 sc->share_count--;
 224         else if (oldcount < 1 && newcount > 0)
 225                 sc->share_count++;
 226 }
 227
 228 /*
 229  * Add @newref to the @root rbtree, merging identical refs.
 230  *
 231  * Callers should assume that newref has been freed after calling.
 232  */
 233 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 234                               struct preftree *preftree,
 235                               struct prelim_ref *newref,
 236                               struct share_check *sc)
 237 {
 238         struct rb_root *root;
 239         struct rb_node **p;
 240         struct rb_node *parent = NULL;
 241         struct prelim_ref *ref;
 242         int result;
 243
 244         root = &preftree->root;
 245         p = &root->rb_node;
 246
 247         while (*p) {
 248                 parent = *p;
 249                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 250                 result = prelim_ref_compare(ref, newref);
 251                 if (result < 0) {
 252                         p = &(*p)->rb_left;
 253                 } else if (result > 0) {
 254                         p = &(*p)->rb_right;
 255                 } else {
 256                         /* Identical refs, merge them and free @newref */
 257                         struct extent_inode_elem *eie = ref->inode_list;
 258
 259                         while (eie && eie->next)
 260                                 eie = eie->next;
 261
 262                         if (!eie)
 263                                 ref->inode_list = newref->inode_list;
 264                         else
 265                                 eie->next = newref->inode_list;
 266                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 267                                                      preftree->count);
 268                         /*
 269                          * A delayed ref can have newref->count < 0.
 270                          * The ref->count is updated to follow any
 271                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 272                          */
 273                         update_share_count(sc, ref->count,
 274                                            ref->count + newref->count);
 275                         ref->count += newref->count;
 276                         free_pref(newref);
 277                         return;
 278                 }
 279         }
 280
 281         update_share_count(sc, 0, newref->count);
 282         preftree->count++;
 283         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 284         rb_link_node(&newref->rbnode, parent, p);
 285         rb_insert_color(&newref->rbnode, root);
 286 }
 287
 288 /*
 289  * Release the entire tree.  We don't care about internal consistency so
 290  * just free everything and then reset the tree root.
 291  */
 292 static void prelim_release(struct preftree *preftree)
 293 {
 294         struct prelim_ref *ref, *next_ref;
 295
 296         rbtree_postorder_for_each_entry_safe(ref, next_ref, &preftree->root,
 297                                              rbnode)
 298                 free_pref(ref);
 299
 300         preftree->root = RB_ROOT;
 301         preftree->count = 0;
 302 }
 303
 304 /*
 305  * the rules for all callers of this function are:
 306  * - obtaining the parent is the goal
 307  * - if you add a key, you must know that it is a correct key
 308  * - if you cannot add the parent or a correct key, then we will look into the
 309  *   block later to set a correct key
 310  *
 311  * delayed refs
 312  * ============
 313  *        backref type | shared | indirect | shared | indirect
 314  * information         |   tree |     tree |   data |     data
 315  * --------------------+--------+----------+--------+----------
 316  *      parent logical |    y   |     -    |    -   |     -
 317  *      key to resolve |    -   |     y    |    y   |     y
 318  *  tree block logical |    -   |     -    |    -   |     -
 319  *  root for resolving |    y   |     y    |    y   |     y
 320  *
 321  * - column 1:       we've the parent -> done
 322  * - column 2, 3, 4: we use the key to find the parent
 323  *
 324  * on disk refs (inline or keyed)
 325  * ==============================
 326  *        backref type | shared | indirect | shared | indirect
 327  * information         |   tree |     tree |   data |     data
 328  * --------------------+--------+----------+--------+----------
 329  *      parent logical |    y   |     -    |    y   |     -
 330  *      key to resolve |    -   |     -    |    -   |     y
 331  *  tree block logical |    y   |     y    |    y   |     y
 332  *  root for resolving |    -   |     y    |    y   |     y
 333  *
 334  * - column 1, 3: we've the parent -> done
 335  * - column 2:    we take the first key from the block to find the parent
 336  *                (see add_missing_keys)
 337  * - column 4:    we use the key to find the parent
 338  *
 339  * additional information that's available but not required to find the parent
 340  * block might help in merging entries to gain some speed.
 341  */
 342 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 343                           struct preftree *preftree, u64 root_id,
 344                           const struct btrfs_key *key, int level, u64 parent,
 345                           u64 wanted_disk_byte, int count,
 346                           struct share_check *sc, gfp_t gfp_mask)
 347 {
 348         struct prelim_ref *ref;
 349
 350         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 351                 return 0;
 352
 353         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 354         if (!ref)
 355                 return -ENOMEM;
 356
 357         ref->root_id = root_id;
 358         if (key) {
 359                 ref->key_for_search = *key;
 360                 /*
 361                  * We can often find data backrefs with an offset that is too
 362                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 363                  * underflows when subtracting a file's offset with the data
 364                  * offset of its corresponding extent data item. This can
 365                  * happen for example in the clone ioctl.
 366                  * So if we detect such case we set the search key's offset to
 367                  * zero to make sure we will find the matching file extent item
 368                  * at add_all_parents(), otherwise we will miss it because the
 369                  * offset taken form the backref is much larger then the offset
 370                  * of the file extent item. This can make us scan a very large
 371                  * number of file extent items, but at least it will not make
 372                  * us miss any.
 373                  * This is an ugly workaround for a behaviour that should have
 374                  * never existed, but it does and a fix for the clone ioctl
 375                  * would touch a lot of places, cause backwards incompatibility
 376                  * and would not fix the problem for extents cloned with older
 377                  * kernels.
 378                  */
 379                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 380                     ref->key_for_search.offset >= LLONG_MAX)
 381                         ref->key_for_search.offset = 0;
 382         } else {
 383                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 384         }
 385
 386         ref->inode_list = NULL;
 387         ref->level = level;
 388         ref->count = count;
 389         ref->parent = parent;
 390         ref->wanted_disk_byte = wanted_disk_byte;
 391         prelim_ref_insert(fs_info, preftree, ref, sc);
 392         return extent_is_shared(sc);
 393 }
 394
 395 /* direct refs use root == 0, key == NULL */
 396 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 397                           struct preftrees *preftrees, int level, u64 parent,
 398                           u64 wanted_disk_byte, int count,
 399                           struct share_check *sc, gfp_t gfp_mask)
 400 {
 401         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 402                               parent, wanted_disk_byte, count, sc, gfp_mask);
 403 }
 404
 405 /* indirect refs use parent == 0 */
 406 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 407                             struct preftrees *preftrees, u64 root_id,
 408                             const struct btrfs_key *key, int level,
 409                             u64 wanted_disk_byte, int count,
 410                             struct share_check *sc, gfp_t gfp_mask)
 411 {
 412         struct preftree *tree = &preftrees->indirect;
 413
 414         if (!key)
 415                 tree = &preftrees->indirect_missing_keys;
 416         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 417                               wanted_disk_byte, count, sc, gfp_mask);
 418 }
 419
 420 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 421                            struct ulist *parents, struct prelim_ref *ref,
 422                            int level, u64 time_seq, const u64 *extent_item_pos,
 423                            u64 total_refs)
 424 {
 425         int ret = 0;
 426         int slot;
 427         struct extent_buffer *eb;
 428         struct btrfs_key key;
 429         struct btrfs_key *key_for_search = &ref->key_for_search;
 430         struct btrfs_file_extent_item *fi;
 431         struct extent_inode_elem *eie = NULL, *old = NULL;
 432         u64 disk_byte;
 433         u64 wanted_disk_byte = ref->wanted_disk_byte;
 434         u64 count = 0;
 435
 436         if (level != 0) {
 437                 eb = path->nodes[level];
 438                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 439                 if (ret < 0)
 440                         return ret;
 441                 return 0;
 442         }
 443
 444         /*
 445          * We normally enter this function with the path already pointing to
 446          * the first item to check. But sometimes, we may enter it with
 447          * slot==nritems. In that case, go to the next leaf before we continue.
 448          */
 449         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 450                 if (time_seq == SEQ_LAST)
 451                         ret = btrfs_next_leaf(root, path);
 452                 else
 453                         ret = btrfs_next_old_leaf(root, path, time_seq);
 454         }
 455
 456         while (!ret && count < total_refs) {
 457                 eb = path->nodes[0];
 458                 slot = path->slots[0];
 459
 460                 btrfs_item_key_to_cpu(eb, &key, slot);
 461
 462                 if (key.objectid != key_for_search->objectid ||
 463                     key.type != BTRFS_EXTENT_DATA_KEY)
 464                         break;
 465
 466                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 467                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 468
 469                 if (disk_byte == wanted_disk_byte) {
 470                         eie = NULL;
 471                         old = NULL;
 472                         count++;
 473                         if (extent_item_pos) {
 474                                 ret = check_extent_in_eb(&key, eb, fi,
 475                                                 *extent_item_pos,
 476                                                 &eie);
 477                                 if (ret < 0)
 478                                         break;
 479                         }
 480                         if (ret > 0)
 481                                 goto next;
 482                         ret = ulist_add_merge_ptr(parents, eb->start,
 483                                                   eie, (void **)&old, GFP_NOFS);
 484                         if (ret < 0)
 485                                 break;
 486                         if (!ret && extent_item_pos) {
 487                                 while (old->next)
 488                                         old = old->next;
 489                                 old->next = eie;
 490                         }
 491                         eie = NULL;
 492                 }
 493 next:
 494                 if (time_seq == SEQ_LAST)
 495                         ret = btrfs_next_item(root, path);
 496                 else
 497                         ret = btrfs_next_old_item(root, path, time_seq);
 498         }
 499
 500         if (ret > 0)
 501                 ret = 0;
 502         else if (ret < 0)
 503                 free_inode_elem_list(eie);
 504         return ret;
 505 }
 506
 507 /*
 508  * resolve an indirect backref in the form (root_id, key, level)
 509  * to a logical address
 510  */
 511 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 512                                 struct btrfs_path *path, u64 time_seq,
 513                                 struct prelim_ref *ref, struct ulist *parents,
 514                                 const u64 *extent_item_pos, u64 total_refs)
 515 {
 516         struct btrfs_root *root;
 517         struct btrfs_key root_key;
 518         struct extent_buffer *eb;
 519         int ret = 0;
 520         int root_level;
 521         int level = ref->level;
 522         int index;
 523
 524         root_key.objectid = ref->root_id;
 525         root_key.type = BTRFS_ROOT_ITEM_KEY;
 526         root_key.offset = (u64)-1;
 527
 528         index = srcu_read_lock(&fs_info->subvol_srcu);
 529
 530         root = btrfs_get_fs_root(fs_info, &root_key, false);
 531         if (IS_ERR(root)) {
 532                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 533                 ret = PTR_ERR(root);
 534                 goto out;
 535         }
 536
 537         if (btrfs_is_testing(fs_info)) {
 538                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 539                 ret = -ENOENT;
 540                 goto out;
 541         }
 542
 543         if (path->search_commit_root)
 544                 root_level = btrfs_header_level(root->commit_root);
 545         else if (time_seq == SEQ_LAST)
 546                 root_level = btrfs_header_level(root->node);
 547         else
 548                 root_level = btrfs_old_root_level(root, time_seq);
 549
 550         if (root_level + 1 == level) {
 551                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 552                 goto out;
 553         }
 554
 555         path->lowest_level = level;
 556         if (time_seq == SEQ_LAST)
 557                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 558                                         0, 0);
 559         else
 560                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 561                                             time_seq);
 562
 563         /* root node has been locked, we can release @subvol_srcu safely here */
 564         srcu_read_unlock(&fs_info->subvol_srcu, index);
 565
 566         btrfs_debug(fs_info,
 567                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 568                  ref->root_id, level, ref->count, ret,
 569                  ref->key_for_search.objectid, ref->key_for_search.type,
 570                  ref->key_for_search.offset);
 571         if (ret < 0)
 572                 goto out;
 573
 574         eb = path->nodes[level];
 575         while (!eb) {
 576                 if (WARN_ON(!level)) {
 577                         ret = 1;
 578                         goto out;
 579                 }
 580                 level--;
 581                 eb = path->nodes[level];
 582         }
 583
 584         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 585                               extent_item_pos, total_refs);
 586 out:
 587         path->lowest_level = 0;
 588         btrfs_release_path(path);
 589         return ret;
 590 }
 591
 592 static struct extent_inode_elem *
 593 unode_aux_to_inode_list(struct ulist_node *node)
 594 {
 595         if (!node)
 596                 return NULL;
 597         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 598 }
 599
 600 static void free_leaf_list(struct ulist *ulist)
 601 {
 602         struct ulist_node *node;
 603         struct ulist_iterator uiter;
 604
 605         ULIST_ITER_INIT(&uiter);
 606         while ((node = ulist_next(ulist, &uiter)))
 607                 free_inode_elem_list(unode_aux_to_inode_list(node));
 608
 609         ulist_free(ulist);
 610 }
 611
 612 /*
 613  * We maintain three seperate rbtrees: one for direct refs, one for
 614  * indirect refs which have a key, and one for indirect refs which do not
 615  * have a key. Each tree does merge on insertion.
 616  *
 617  * Once all of the references are located, we iterate over the tree of
 618  * indirect refs with missing keys. An appropriate key is located and
 619  * the ref is moved onto the tree for indirect refs. After all missing
 620  * keys are thus located, we iterate over the indirect ref tree, resolve
 621  * each reference, and then insert the resolved reference onto the
 622  * direct tree (merging there too).
 623  *
 624  * New backrefs (i.e., for parent nodes) are added to the appropriate
 625  * rbtree as they are encountered. The new backrefs are subsequently
 626  * resolved as above.
 627  */
 628 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 629                                  struct btrfs_path *path, u64 time_seq,
 630                                  struct preftrees *preftrees,
 631                                  const u64 *extent_item_pos, u64 total_refs,
 632                                  struct share_check *sc)
 633 {
 634         int err;
 635         int ret = 0;
 636         struct ulist *parents;
 637         struct ulist_node *node;
 638         struct ulist_iterator uiter;
 639         struct rb_node *rnode;
 640
 641         parents = ulist_alloc(GFP_NOFS);
 642         if (!parents)
 643                 return -ENOMEM;
 644
 645         /*
 646          * We could trade memory usage for performance here by iterating
 647          * the tree, allocating new refs for each insertion, and then
 648          * freeing the entire indirect tree when we're done.  In some test
 649          * cases, the tree can grow quite large (~200k objects).
 650          */
 651         while ((rnode = rb_first(&preftrees->indirect.root))) {
 652                 struct prelim_ref *ref;
 653
 654                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 655                 if (WARN(ref->parent,
 656                          "BUG: direct ref found in indirect tree")) {
 657                         ret = -EINVAL;
 658                         goto out;
 659                 }
 660
 661                 rb_erase(&ref->rbnode, &preftrees->indirect.root);
 662                 preftrees->indirect.count--;
 663
 664                 if (ref->count == 0) {
 665                         free_pref(ref);
 666                         continue;
 667                 }
 668
 669                 if (sc && sc->root_objectid &&
 670                     ref->root_id != sc->root_objectid) {
 671                         free_pref(ref);
 672                         ret = BACKREF_FOUND_SHARED;
 673                         goto out;
 674                 }
 675                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
 676                                            parents, extent_item_pos,
 677                                            total_refs);
 678                 /*
 679                  * we can only tolerate ENOENT,otherwise,we should catch error
 680                  * and return directly.
 681                  */
 682                 if (err == -ENOENT) {
 683                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 684                                           NULL);
 685                         continue;
 686                 } else if (err) {
 687                         free_pref(ref);
 688                         ret = err;
 689                         goto out;
 690                 }
 691
 692                 /* we put the first parent into the ref at hand */
 693                 ULIST_ITER_INIT(&uiter);
 694                 node = ulist_next(parents, &uiter);
 695                 ref->parent = node ? node->val : 0;
 696                 ref->inode_list = unode_aux_to_inode_list(node);
 697
 698                 /* Add a prelim_ref(s) for any other parent(s). */
 699                 while ((node = ulist_next(parents, &uiter))) {
 700                         struct prelim_ref *new_ref;
 701
 702                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 703                                                    GFP_NOFS);
 704                         if (!new_ref) {
 705                                 free_pref(ref);
 706                                 ret = -ENOMEM;
 707                                 goto out;
 708                         }
 709                         memcpy(new_ref, ref, sizeof(*ref));
 710                         new_ref->parent = node->val;
 711                         new_ref->inode_list = unode_aux_to_inode_list(node);
 712                         prelim_ref_insert(fs_info, &preftrees->direct,
 713                                           new_ref, NULL);
 714                 }
 715
 716                 /*
 717                  * Now it's a direct ref, put it in the the direct tree. We must
 718                  * do this last because the ref could be merged/freed here.
 719                  */
 720                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 721
 722                 ulist_reinit(parents);
 723                 cond_resched();
 724         }
 725 out:
 726         /*
 727          * We may have inode lists attached to refs in the parents ulist, so we
 728          * must free them before freeing the ulist and its refs.
 729          */
 730         free_leaf_list(parents);
 731         return ret;
 732 }
 733
 734 /*
 735  * read tree blocks and add keys where required.
 736  */
 737 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 738                             struct preftrees *preftrees, bool lock)
 739 {
 740         struct prelim_ref *ref;
 741         struct extent_buffer *eb;
 742         struct preftree *tree = &preftrees->indirect_missing_keys;
 743         struct rb_node *node;
 744
 745         while ((node = rb_first(&tree->root))) {
 746                 ref = rb_entry(node, struct prelim_ref, rbnode);
 747                 rb_erase(node, &tree->root);
 748
 749                 BUG_ON(ref->parent);    /* should not be a direct ref */
 750                 BUG_ON(ref->key_for_search.type);
 751                 BUG_ON(!ref->wanted_disk_byte);
 752
 753                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0);
 754                 if (IS_ERR(eb)) {
 755                         free_pref(ref);
 756                         return PTR_ERR(eb);
 757                 } else if (!extent_buffer_uptodate(eb)) {
 758                         free_pref(ref);
 759                         free_extent_buffer(eb);
 760                         return -EIO;
 761                 }
 762                 if (lock)
 763                         btrfs_tree_read_lock(eb);
 764                 if (btrfs_header_level(eb) == 0)
 765                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 766                 else
 767                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 768                 if (lock)
 769                         btrfs_tree_read_unlock(eb);
 770                 free_extent_buffer(eb);
 771                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 772                 cond_resched();
 773         }
 774         return 0;
 775 }
 776
 777 /*
 778  * add all currently queued delayed refs from this head whose seq nr is
 779  * smaller or equal that seq to the list
 780  */
 781 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 782                             struct btrfs_delayed_ref_head *head, u64 seq,
 783                             struct preftrees *preftrees, u64 *total_refs,
 784                             struct share_check *sc)
 785 {
 786         struct btrfs_delayed_ref_node *node;
 787         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 788         struct btrfs_key key;
 789         struct btrfs_key tmp_op_key;
 790         struct btrfs_key *op_key = NULL;
 791         int count;
 792         int ret = 0;
 793
 794         if (extent_op && extent_op->update_key) {
 795                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 796                 op_key = &tmp_op_key;
 797         }
 798
 799         spin_lock(&head->lock);
 800         list_for_each_entry(node, &head->ref_list, list) {
 801                 if (node->seq > seq)
 802                         continue;
 803
 804                 switch (node->action) {
 805                 case BTRFS_ADD_DELAYED_EXTENT:
 806                 case BTRFS_UPDATE_DELAYED_HEAD:
 807                         WARN_ON(1);
 808                         continue;
 809                 case BTRFS_ADD_DELAYED_REF:
 810                         count = node->ref_mod;
 811                         break;
 812                 case BTRFS_DROP_DELAYED_REF:
 813                         count = node->ref_mod * -1;
 814                         break;
 815                 default:
 816                         BUG_ON(1);
 817                 }
 818                 *total_refs += count;
 819                 switch (node->type) {
 820                 case BTRFS_TREE_BLOCK_REF_KEY: {
 821                         /* NORMAL INDIRECT METADATA backref */
 822                         struct btrfs_delayed_tree_ref *ref;
 823
 824                         ref = btrfs_delayed_node_to_tree_ref(node);
 825                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 826                                                &tmp_op_key, ref->level + 1,
 827                                                node->bytenr, count, sc,
 828                                                GFP_ATOMIC);
 829                         break;
 830                 }
 831                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 832                         /* SHARED DIRECT METADATA backref */
 833                         struct btrfs_delayed_tree_ref *ref;
 834
 835                         ref = btrfs_delayed_node_to_tree_ref(node);
 836
 837                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 838                                              ref->parent, node->bytenr, count,
 839                                              sc, GFP_ATOMIC);
 840                         break;
 841                 }
 842                 case BTRFS_EXTENT_DATA_REF_KEY: {
 843                         /* NORMAL INDIRECT DATA backref */
 844                         struct btrfs_delayed_data_ref *ref;
 845                         ref = btrfs_delayed_node_to_data_ref(node);
 846
 847                         key.objectid = ref->objectid;
 848                         key.type = BTRFS_EXTENT_DATA_KEY;
 849                         key.offset = ref->offset;
 850
 851                         /*
 852                          * If we have a share check context and a reference for
 853                          * another inode, we can't exit immediately. This is
 854                          * because even if this is a BTRFS_ADD_DELAYED_REF
 855                          * reference we may find next a BTRFS_DROP_DELAYED_REF
 856                          * which cancels out this ADD reference.
 857                          *
 858                          * If this is a DROP reference and there was no previous
 859                          * ADD reference, then we need to signal that when we
 860                          * process references from the extent tree (through
 861                          * add_inline_refs() and add_keyed_refs()), we should
 862                          * not exit early if we find a reference for another
 863                          * inode, because one of the delayed DROP references
 864                          * may cancel that reference in the extent tree.
 865                          */
 866                         if (sc && count < 0)
 867                                 sc->have_delayed_delete_refs = true;
 868
 869                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 870                                                &key, 0, node->bytenr, count, sc,
 871                                                GFP_ATOMIC);
 872                         break;
 873                 }
 874                 case BTRFS_SHARED_DATA_REF_KEY: {
 875                         /* SHARED DIRECT FULL backref */
 876                         struct btrfs_delayed_data_ref *ref;
 877
 878                         ref = btrfs_delayed_node_to_data_ref(node);
 879
 880                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 881                                              node->bytenr, count, sc,
 882                                              GFP_ATOMIC);
 883                         break;
 884                 }
 885                 default:
 886                         WARN_ON(1);
 887                 }
 888                 /*
 889                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 890                  * refs have been checked.
 891                  */
 892                 if (ret && (ret != BACKREF_FOUND_SHARED))
 893                         break;
 894         }
 895         if (!ret)
 896                 ret = extent_is_shared(sc);
 897
 898         spin_unlock(&head->lock);
 899         return ret;
 900 }
 901
 902 /*
 903  * add all inline backrefs for bytenr to the list
 904  *
 905  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 906  */
 907 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 908                            struct btrfs_path *path, u64 bytenr,
 909                            int *info_level, struct preftrees *preftrees,
 910                            u64 *total_refs, struct share_check *sc)
 911 {
 912         int ret = 0;
 913         int slot;
 914         struct extent_buffer *leaf;
 915         struct btrfs_key key;
 916         struct btrfs_key found_key;
 917         unsigned long ptr;
 918         unsigned long end;
 919         struct btrfs_extent_item *ei;
 920         u64 flags;
 921         u64 item_size;
 922
 923         /*
 924          * enumerate all inline refs
 925          */
 926         leaf = path->nodes[0];
 927         slot = path->slots[0];
 928
 929         item_size = btrfs_item_size_nr(leaf, slot);
 930         BUG_ON(item_size < sizeof(*ei));
 931
 932         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 933         flags = btrfs_extent_flags(leaf, ei);
 934         *total_refs += btrfs_extent_refs(leaf, ei);
 935         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 936
 937         ptr = (unsigned long)(ei + 1);
 938         end = (unsigned long)ei + item_size;
 939
 940         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 941             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 942                 struct btrfs_tree_block_info *info;
 943
 944                 info = (struct btrfs_tree_block_info *)ptr;
 945                 *info_level = btrfs_tree_block_level(leaf, info);
 946                 ptr += sizeof(struct btrfs_tree_block_info);
 947                 BUG_ON(ptr > end);
 948         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 949                 *info_level = found_key.offset;
 950         } else {
 951                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 952         }
 953
 954         while (ptr < end) {
 955                 struct btrfs_extent_inline_ref *iref;
 956                 u64 offset;
 957                 int type;
 958
 959                 iref = (struct btrfs_extent_inline_ref *)ptr;
 960                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 961                                                         BTRFS_REF_TYPE_ANY);
 962                 if (type == BTRFS_REF_TYPE_INVALID)
 963                         return -EINVAL;
 964
 965                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 966
 967                 switch (type) {
 968                 case BTRFS_SHARED_BLOCK_REF_KEY:
 969                         ret = add_direct_ref(fs_info, preftrees,
 970                                              *info_level + 1, offset,
 971                                              bytenr, 1, NULL, GFP_NOFS);
 972                         break;
 973                 case BTRFS_SHARED_DATA_REF_KEY: {
 974                         struct btrfs_shared_data_ref *sdref;
 975                         int count;
 976
 977                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 978                         count = btrfs_shared_data_ref_count(leaf, sdref);
 979
 980                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 981                                              bytenr, count, sc, GFP_NOFS);
 982                         break;
 983                 }
 984                 case BTRFS_TREE_BLOCK_REF_KEY:
 985                         ret = add_indirect_ref(fs_info, preftrees, offset,
 986                                                NULL, *info_level + 1,
 987                                                bytenr, 1, NULL, GFP_NOFS);
 988                         break;
 989                 case BTRFS_EXTENT_DATA_REF_KEY: {
 990                         struct btrfs_extent_data_ref *dref;
 991                         int count;
 992                         u64 root;
 993
 994                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 995                         count = btrfs_extent_data_ref_count(leaf, dref);
 996                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 997                                                                       dref);
 998                         key.type = BTRFS_EXTENT_DATA_KEY;
 999                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1000
1001                         if (sc && sc->inum && key.objectid != sc->inum &&
1002                             !sc->have_delayed_delete_refs) {
1003                                 ret = BACKREF_FOUND_SHARED;
1004                                 break;
1005                         }
1006
1007                         root = btrfs_extent_data_ref_root(leaf, dref);
1008
1009                         ret = add_indirect_ref(fs_info, preftrees, root,
1010                                                &key, 0, bytenr, count,
1011                                                sc, GFP_NOFS);
1012
1013                         break;
1014                 }
1015                 default:
1016                         WARN_ON(1);
1017                 }
1018                 if (ret)
1019                         return ret;
1020                 ptr += btrfs_extent_inline_ref_size(type);
1021         }
1022
1023         return 0;
1024 }
1025
1026 /*
1027  * add all non-inline backrefs for bytenr to the list
1028  *
1029  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1030  */
1031 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1032                           struct btrfs_path *path, u64 bytenr,
1033                           int info_level, struct preftrees *preftrees,
1034                           struct share_check *sc)
1035 {
1036         struct btrfs_root *extent_root = fs_info->extent_root;
1037         int ret;
1038         int slot;
1039         struct extent_buffer *leaf;
1040         struct btrfs_key key;
1041
1042         while (1) {
1043                 ret = btrfs_next_item(extent_root, path);
1044                 if (ret < 0)
1045                         break;
1046                 if (ret) {
1047                         ret = 0;
1048                         break;
1049                 }
1050
1051                 slot = path->slots[0];
1052                 leaf = path->nodes[0];
1053                 btrfs_item_key_to_cpu(leaf, &key, slot);
1054
1055                 if (key.objectid != bytenr)
1056                         break;
1057                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1058                         continue;
1059                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1060                         break;
1061
1062                 switch (key.type) {
1063                 case BTRFS_SHARED_BLOCK_REF_KEY:
1064                         /* SHARED DIRECT METADATA backref */
1065                         ret = add_direct_ref(fs_info, preftrees,
1066                                              info_level + 1, key.offset,
1067                                              bytenr, 1, NULL, GFP_NOFS);
1068                         break;
1069                 case BTRFS_SHARED_DATA_REF_KEY: {
1070                         /* SHARED DIRECT FULL backref */
1071                         struct btrfs_shared_data_ref *sdref;
1072                         int count;
1073
1074                         sdref = btrfs_item_ptr(leaf, slot,
1075                                               struct btrfs_shared_data_ref);
1076                         count = btrfs_shared_data_ref_count(leaf, sdref);
1077                         ret = add_direct_ref(fs_info, preftrees, 0,
1078                                              key.offset, bytenr, count,
1079                                              sc, GFP_NOFS);
1080                         break;
1081                 }
1082                 case BTRFS_TREE_BLOCK_REF_KEY:
1083                         /* NORMAL INDIRECT METADATA backref */
1084                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1085                                                NULL, info_level + 1, bytenr,
1086                                                1, NULL, GFP_NOFS);
1087                         break;
1088                 case BTRFS_EXTENT_DATA_REF_KEY: {
1089                         /* NORMAL INDIRECT DATA backref */
1090                         struct btrfs_extent_data_ref *dref;
1091                         int count;
1092                         u64 root;
1093
1094                         dref = btrfs_item_ptr(leaf, slot,
1095                                               struct btrfs_extent_data_ref);
1096                         count = btrfs_extent_data_ref_count(leaf, dref);
1097                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1098                                                                       dref);
1099                         key.type = BTRFS_EXTENT_DATA_KEY;
1100                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1101
1102                         if (sc && sc->inum && key.objectid != sc->inum &&
1103                             !sc->have_delayed_delete_refs) {
1104                                 ret = BACKREF_FOUND_SHARED;
1105                                 break;
1106                         }
1107
1108                         root = btrfs_extent_data_ref_root(leaf, dref);
1109                         ret = add_indirect_ref(fs_info, preftrees, root,
1110                                                &key, 0, bytenr, count,
1111                                                sc, GFP_NOFS);
1112                         break;
1113                 }
1114                 default:
1115                         WARN_ON(1);
1116                 }
1117                 if (ret)
1118                         return ret;
1119
1120         }
1121
1122         return ret;
1123 }
1124
1125 /*
1126  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1127  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1128  * indirect refs to their parent bytenr.
1129  * When roots are found, they're added to the roots list
1130  *
1131  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1132  * much like trans == NULL case, the difference only lies in it will not
1133  * commit root.
1134  * The special case is for qgroup to search roots in commit_transaction().
1135  *
1136  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1137  * shared extent is detected.
1138  *
1139  * Otherwise this returns 0 for success and <0 for an error.
1140  *
1141  * FIXME some caching might speed things up
1142  */
1143 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1144                              struct btrfs_fs_info *fs_info, u64 bytenr,
1145                              u64 time_seq, struct ulist *refs,
1146                              struct ulist *roots, const u64 *extent_item_pos,
1147                              struct share_check *sc)
1148 {
1149         struct btrfs_key key;
1150         struct btrfs_path *path;
1151         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1152         struct btrfs_delayed_ref_head *head;
1153         int info_level = 0;
1154         int ret;
1155         struct prelim_ref *ref;
1156         struct rb_node *node;
1157         struct extent_inode_elem *eie = NULL;
1158         /* total of both direct AND indirect refs! */
1159         u64 total_refs = 0;
1160         struct preftrees preftrees = {
1161                 .direct = PREFTREE_INIT,
1162                 .indirect = PREFTREE_INIT,
1163                 .indirect_missing_keys = PREFTREE_INIT
1164         };
1165
1166         key.objectid = bytenr;
1167         key.offset = (u64)-1;
1168         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1169                 key.type = BTRFS_METADATA_ITEM_KEY;
1170         else
1171                 key.type = BTRFS_EXTENT_ITEM_KEY;
1172
1173         path = btrfs_alloc_path();
1174         if (!path)
1175                 return -ENOMEM;
1176         if (!trans) {
1177                 path->search_commit_root = 1;
1178                 path->skip_locking = 1;
1179         }
1180
1181         if (time_seq == SEQ_LAST)
1182                 path->skip_locking = 1;
1183
1184         /*
1185          * grab both a lock on the path and a lock on the delayed ref head.
1186          * We need both to get a consistent picture of how the refs look
1187          * at a specified point in time
1188          */
1189 again:
1190         head = NULL;
1191
1192         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1193         if (ret < 0)
1194                 goto out;
1195         if (ret == 0) {
1196                 /* This shouldn't happen, indicates a bug or fs corruption. */
1197                 ASSERT(ret != 0);
1198                 ret = -EUCLEAN;
1199                 goto out;
1200         }
1201
1202 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1203         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1204             time_seq != SEQ_LAST) {
1205 #else
1206         if (trans && time_seq != SEQ_LAST) {
1207 #endif
1208                 /*
1209                  * look if there are updates for this ref queued and lock the
1210                  * head
1211                  */
1212                 delayed_refs = &trans->transaction->delayed_refs;
1213                 spin_lock(&delayed_refs->lock);
1214                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1215                 if (head) {
1216                         if (!mutex_trylock(&head->mutex)) {
1217                                 refcount_inc(&head->node.refs);
1218                                 spin_unlock(&delayed_refs->lock);
1219
1220                                 btrfs_release_path(path);
1221
1222                                 /*
1223                                  * Mutex was contended, block until it's
1224                                  * released and try again
1225                                  */
1226                                 mutex_lock(&head->mutex);
1227                                 mutex_unlock(&head->mutex);
1228                                 btrfs_put_delayed_ref(&head->node);
1229                                 goto again;
1230                         }
1231                         spin_unlock(&delayed_refs->lock);
1232                         ret = add_delayed_refs(fs_info, head, time_seq,
1233                                                &preftrees, &total_refs, sc);
1234                         mutex_unlock(&head->mutex);
1235                         if (ret)
1236                                 goto out;
1237                 } else {
1238                         spin_unlock(&delayed_refs->lock);
1239                 }
1240         }
1241
1242         if (path->slots[0]) {
1243                 struct extent_buffer *leaf;
1244                 int slot;
1245
1246                 path->slots[0]--;
1247                 leaf = path->nodes[0];
1248                 slot = path->slots[0];
1249                 btrfs_item_key_to_cpu(leaf, &key, slot);
1250                 if (key.objectid == bytenr &&
1251                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1252                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1253                         ret = add_inline_refs(fs_info, path, bytenr,
1254                                               &info_level, &preftrees,
1255                                               &total_refs, sc);
1256                         if (ret)
1257                                 goto out;
1258                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1259                                              &preftrees, sc);
1260                         if (ret)
1261                                 goto out;
1262                 }
1263         }
1264
1265         btrfs_release_path(path);
1266
1267         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1268         if (ret)
1269                 goto out;
1270
1271         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root));
1272
1273         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1274                                     extent_item_pos, total_refs, sc);
1275         if (ret)
1276                 goto out;
1277
1278         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root));
1279
1280         /*
1281          * This walks the tree of merged and resolved refs. Tree blocks are
1282          * read in as needed. Unique entries are added to the ulist, and
1283          * the list of found roots is updated.
1284          *
1285          * We release the entire tree in one go before returning.
1286          */
1287         node = rb_first(&preftrees.direct.root);
1288         while (node) {
1289                 ref = rb_entry(node, struct prelim_ref, rbnode);
1290                 node = rb_next(&ref->rbnode);
1291                 /*
1292                  * ref->count < 0 can happen here if there are delayed
1293                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1294                  * prelim_ref_insert() relies on this when merging
1295                  * identical refs to keep the overall count correct.
1296                  * prelim_ref_insert() will merge only those refs
1297                  * which compare identically.  Any refs having
1298                  * e.g. different offsets would not be merged,
1299                  * and would retain their original ref->count < 0.
1300                  */
1301                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1302                         if (sc && sc->root_objectid &&
1303                             ref->root_id != sc->root_objectid) {
1304                                 ret = BACKREF_FOUND_SHARED;
1305                                 goto out;
1306                         }
1307
1308                         /* no parent == root of tree */
1309                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1310                         if (ret < 0)
1311                                 goto out;
1312                 }
1313                 if (ref->count && ref->parent) {
1314                         if (extent_item_pos && !ref->inode_list &&
1315                             ref->level == 0) {
1316                                 struct extent_buffer *eb;
1317
1318                                 eb = read_tree_block(fs_info, ref->parent, 0);
1319                                 if (IS_ERR(eb)) {
1320                                         ret = PTR_ERR(eb);
1321                                         goto out;
1322                                 } else if (!extent_buffer_uptodate(eb)) {
1323                                         free_extent_buffer(eb);
1324                                         ret = -EIO;
1325                                         goto out;
1326                                 }
1327                                 if (!path->skip_locking) {
1328                                         btrfs_tree_read_lock(eb);
1329                                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1330                                 }
1331                                 ret = find_extent_in_eb(eb, bytenr,
1332                                                         *extent_item_pos, &eie);
1333                                 if (!path->skip_locking)
1334                                         btrfs_tree_read_unlock_blocking(eb);
1335                                 free_extent_buffer(eb);
1336                                 if (ret < 0)
1337                                         goto out;
1338                                 ref->inode_list = eie;
1339                         }
1340                         ret = ulist_add_merge_ptr(refs, ref->parent,
1341                                                   ref->inode_list,
1342                                                   (void **)&eie, GFP_NOFS);
1343                         if (ret < 0)
1344                                 goto out;
1345                         if (!ret && extent_item_pos) {
1346                                 /*
1347                                  * We've recorded that parent, so we must extend
1348                                  * its inode list here.
1349                                  *
1350                                  * However if there was corruption we may not
1351                                  * have found an eie, return an error in this
1352                                  * case.
1353                                  */
1354                                 ASSERT(eie);
1355                                 if (!eie) {
1356                                         ret = -EUCLEAN;
1357                                         goto out;
1358                                 }
1359                                 while (eie->next)
1360                                         eie = eie->next;
1361                                 eie->next = ref->inode_list;
1362                         }
1363                         eie = NULL;
1364                 }
1365                 cond_resched();
1366         }
1367
1368 out:
1369         btrfs_free_path(path);
1370
1371         prelim_release(&preftrees.direct);
1372         prelim_release(&preftrees.indirect);
1373         prelim_release(&preftrees.indirect_missing_keys);
1374
1375         if (ret < 0)
1376                 free_inode_elem_list(eie);
1377         return ret;
1378 }
1379
1380 /*
1381  * Finds all leafs with a reference to the specified combination of bytenr and
1382  * offset. key_list_head will point to a list of corresponding keys (caller must
1383  * free each list element). The leafs will be stored in the leafs ulist, which
1384  * must be freed with ulist_free.
1385  *
1386  * returns 0 on success, <0 on error
1387  */
1388 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1389                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1390                                 u64 time_seq, struct ulist **leafs,
1391                                 const u64 *extent_item_pos)
1392 {
1393         int ret;
1394
1395         *leafs = ulist_alloc(GFP_NOFS);
1396         if (!*leafs)
1397                 return -ENOMEM;
1398
1399         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1400                                 *leafs, NULL, extent_item_pos, NULL);
1401         if (ret < 0 && ret != -ENOENT) {
1402                 free_leaf_list(*leafs);
1403                 return ret;
1404         }
1405
1406         return 0;
1407 }
1408
1409 /*
1410  * walk all backrefs for a given extent to find all roots that reference this
1411  * extent. Walking a backref means finding all extents that reference this
1412  * extent and in turn walk the backrefs of those, too. Naturally this is a
1413  * recursive process, but here it is implemented in an iterative fashion: We
1414  * find all referencing extents for the extent in question and put them on a
1415  * list. In turn, we find all referencing extents for those, further appending
1416  * to the list. The way we iterate the list allows adding more elements after
1417  * the current while iterating. The process stops when we reach the end of the
1418  * list. Found roots are added to the roots list.
1419  *
1420  * returns 0 on success, < 0 on error.
1421  */
1422 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1423                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1424                                      u64 time_seq, struct ulist **roots)
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);
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 {
1465         int ret;
1466
1467         if (!trans)
1468                 down_read(&fs_info->commit_root_sem);
1469         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1470                                         time_seq, roots);
1471         if (!trans)
1472                 up_read(&fs_info->commit_root_sem);
1473         return ret;
1474 }
1475
1476 /**
1477  * btrfs_check_shared - tell us whether an extent is shared
1478  *
1479  * btrfs_check_shared uses the backref walking code but will short
1480  * circuit as soon as it finds a root or inode that doesn't match the
1481  * one passed in. This provides a significant performance benefit for
1482  * callers (such as fiemap) which want to know whether the extent is
1483  * shared but do not need a ref count.
1484  *
1485  * This attempts to attach to the running transaction in order to account for
1486  * delayed refs, but continues on even when no running transaction exists.
1487  *
1488  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1489  */
1490 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1491 {
1492         struct btrfs_fs_info *fs_info = root->fs_info;
1493         struct btrfs_trans_handle *trans;
1494         struct ulist *tmp = NULL;
1495         struct ulist *roots = NULL;
1496         struct ulist_iterator uiter;
1497         struct ulist_node *node;
1498         struct seq_list elem = SEQ_LIST_INIT(elem);
1499         int ret = 0;
1500         struct share_check shared = {
1501                 .root_objectid = root->objectid,
1502                 .inum = inum,
1503                 .share_count = 0,
1504                 .have_delayed_delete_refs = false,
1505         };
1506
1507         tmp = ulist_alloc(GFP_NOFS);
1508         roots = ulist_alloc(GFP_NOFS);
1509         if (!tmp || !roots) {
1510                 ret = -ENOMEM;
1511                 goto out;
1512         }
1513
1514         trans = btrfs_attach_transaction(root);
1515         if (IS_ERR(trans)) {
1516                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1517                         ret = PTR_ERR(trans);
1518                         goto out;
1519                 }
1520                 trans = NULL;
1521                 down_read(&fs_info->commit_root_sem);
1522         } else {
1523                 btrfs_get_tree_mod_seq(fs_info, &elem);
1524         }
1525
1526         ULIST_ITER_INIT(&uiter);
1527         while (1) {
1528                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1529                                         roots, NULL, &shared);
1530                 if (ret == BACKREF_FOUND_SHARED) {
1531                         /* this is the only condition under which we return 1 */
1532                         ret = 1;
1533                         break;
1534                 }
1535                 if (ret < 0 && ret != -ENOENT)
1536                         break;
1537                 ret = 0;
1538                 node = ulist_next(tmp, &uiter);
1539                 if (!node)
1540                         break;
1541                 bytenr = node->val;
1542                 shared.share_count = 0;
1543                 shared.have_delayed_delete_refs = false;
1544                 cond_resched();
1545         }
1546
1547         if (trans) {
1548                 btrfs_put_tree_mod_seq(fs_info, &elem);
1549                 btrfs_end_transaction(trans);
1550         } else {
1551                 up_read(&fs_info->commit_root_sem);
1552         }
1553 out:
1554         ulist_free(tmp);
1555         ulist_free(roots);
1556         return ret;
1557 }
1558
1559 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1560                           u64 start_off, struct btrfs_path *path,
1561                           struct btrfs_inode_extref **ret_extref,
1562                           u64 *found_off)
1563 {
1564         int ret, slot;
1565         struct btrfs_key key;
1566         struct btrfs_key found_key;
1567         struct btrfs_inode_extref *extref;
1568         const struct extent_buffer *leaf;
1569         unsigned long ptr;
1570
1571         key.objectid = inode_objectid;
1572         key.type = BTRFS_INODE_EXTREF_KEY;
1573         key.offset = start_off;
1574
1575         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1576         if (ret < 0)
1577                 return ret;
1578
1579         while (1) {
1580                 leaf = path->nodes[0];
1581                 slot = path->slots[0];
1582                 if (slot >= btrfs_header_nritems(leaf)) {
1583                         /*
1584                          * If the item at offset is not found,
1585                          * btrfs_search_slot will point us to the slot
1586                          * where it should be inserted. In our case
1587                          * that will be the slot directly before the
1588                          * next INODE_REF_KEY_V2 item. In the case
1589                          * that we're pointing to the last slot in a
1590                          * leaf, we must move one leaf over.
1591                          */
1592                         ret = btrfs_next_leaf(root, path);
1593                         if (ret) {
1594                                 if (ret >= 1)
1595                                         ret = -ENOENT;
1596                                 break;
1597                         }
1598                         continue;
1599                 }
1600
1601                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1602
1603                 /*
1604                  * Check that we're still looking at an extended ref key for
1605                  * this particular objectid. If we have different
1606                  * objectid or type then there are no more to be found
1607                  * in the tree and we can exit.
1608                  */
1609                 ret = -ENOENT;
1610                 if (found_key.objectid != inode_objectid)
1611                         break;
1612                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1613                         break;
1614
1615                 ret = 0;
1616                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1617                 extref = (struct btrfs_inode_extref *)ptr;
1618                 *ret_extref = extref;
1619                 if (found_off)
1620                         *found_off = found_key.offset;
1621                 break;
1622         }
1623
1624         return ret;
1625 }
1626
1627 /*
1628  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1629  * Elements of the path are separated by '/' and the path is guaranteed to be
1630  * 0-terminated. the path is only given within the current file system.
1631  * Therefore, it never starts with a '/'. the caller is responsible to provide
1632  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1633  * the start point of the resulting string is returned. this pointer is within
1634  * dest, normally.
1635  * in case the path buffer would overflow, the pointer is decremented further
1636  * as if output was written to the buffer, though no more output is actually
1637  * generated. that way, the caller can determine how much space would be
1638  * required for the path to fit into the buffer. in that case, the returned
1639  * value will be smaller than dest. callers must check this!
1640  */
1641 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1642                         u32 name_len, unsigned long name_off,
1643                         struct extent_buffer *eb_in, u64 parent,
1644                         char *dest, u32 size)
1645 {
1646         int slot;
1647         u64 next_inum;
1648         int ret;
1649         s64 bytes_left = ((s64)size) - 1;
1650         struct extent_buffer *eb = eb_in;
1651         struct btrfs_key found_key;
1652         int leave_spinning = path->leave_spinning;
1653         struct btrfs_inode_ref *iref;
1654
1655         if (bytes_left >= 0)
1656                 dest[bytes_left] = '\0';
1657
1658         path->leave_spinning = 1;
1659         while (1) {
1660                 bytes_left -= name_len;
1661                 if (bytes_left >= 0)
1662                         read_extent_buffer(eb, dest + bytes_left,
1663                                            name_off, name_len);
1664                 if (eb != eb_in) {
1665                         if (!path->skip_locking)
1666                                 btrfs_tree_read_unlock_blocking(eb);
1667                         free_extent_buffer(eb);
1668                 }
1669                 ret = btrfs_find_item(fs_root, path, parent, 0,
1670                                 BTRFS_INODE_REF_KEY, &found_key);
1671                 if (ret > 0)
1672                         ret = -ENOENT;
1673                 if (ret)
1674                         break;
1675
1676                 next_inum = found_key.offset;
1677
1678                 /* regular exit ahead */
1679                 if (parent == next_inum)
1680                         break;
1681
1682                 slot = path->slots[0];
1683                 eb = path->nodes[0];
1684                 /* make sure we can use eb after releasing the path */
1685                 if (eb != eb_in) {
1686                         if (!path->skip_locking)
1687                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1688                         path->nodes[0] = NULL;
1689                         path->locks[0] = 0;
1690                 }
1691                 btrfs_release_path(path);
1692                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1693
1694                 name_len = btrfs_inode_ref_name_len(eb, iref);
1695                 name_off = (unsigned long)(iref + 1);
1696
1697                 parent = next_inum;
1698                 --bytes_left;
1699                 if (bytes_left >= 0)
1700                         dest[bytes_left] = '/';
1701         }
1702
1703         btrfs_release_path(path);
1704         path->leave_spinning = leave_spinning;
1705
1706         if (ret)
1707                 return ERR_PTR(ret);
1708
1709         return dest + bytes_left;
1710 }
1711
1712 /*
1713  * this makes the path point to (logical EXTENT_ITEM *)
1714  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1715  * tree blocks and <0 on error.
1716  */
1717 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1718                         struct btrfs_path *path, struct btrfs_key *found_key,
1719                         u64 *flags_ret)
1720 {
1721         int ret;
1722         u64 flags;
1723         u64 size = 0;
1724         u32 item_size;
1725         const struct extent_buffer *eb;
1726         struct btrfs_extent_item *ei;
1727         struct btrfs_key key;
1728
1729         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1730                 key.type = BTRFS_METADATA_ITEM_KEY;
1731         else
1732                 key.type = BTRFS_EXTENT_ITEM_KEY;
1733         key.objectid = logical;
1734         key.offset = (u64)-1;
1735
1736         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1737         if (ret < 0)
1738                 return ret;
1739
1740         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1741         if (ret) {
1742                 if (ret > 0)
1743                         ret = -ENOENT;
1744                 return ret;
1745         }
1746         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1747         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1748                 size = fs_info->nodesize;
1749         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1750                 size = found_key->offset;
1751
1752         if (found_key->objectid > logical ||
1753             found_key->objectid + size <= logical) {
1754                 btrfs_debug(fs_info,
1755                         "logical %llu is not within any extent", logical);
1756                 return -ENOENT;
1757         }
1758
1759         eb = path->nodes[0];
1760         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1761         BUG_ON(item_size < sizeof(*ei));
1762
1763         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1764         flags = btrfs_extent_flags(eb, ei);
1765
1766         btrfs_debug(fs_info,
1767                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1768                  logical, logical - found_key->objectid, found_key->objectid,
1769                  found_key->offset, flags, item_size);
1770
1771         WARN_ON(!flags_ret);
1772         if (flags_ret) {
1773                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1774                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1775                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1776                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1777                 else
1778                         BUG_ON(1);
1779                 return 0;
1780         }
1781
1782         return -EIO;
1783 }
1784
1785 /*
1786  * helper function to iterate extent inline refs. ptr must point to a 0 value
1787  * for the first call and may be modified. it is used to track state.
1788  * if more refs exist, 0 is returned and the next call to
1789  * get_extent_inline_ref must pass the modified ptr parameter to get the
1790  * next ref. after the last ref was processed, 1 is returned.
1791  * returns <0 on error
1792  */
1793 static int get_extent_inline_ref(unsigned long *ptr,
1794                                  const struct extent_buffer *eb,
1795                                  const struct btrfs_key *key,
1796                                  const struct btrfs_extent_item *ei,
1797                                  u32 item_size,
1798                                  struct btrfs_extent_inline_ref **out_eiref,
1799                                  int *out_type)
1800 {
1801         unsigned long end;
1802         u64 flags;
1803         struct btrfs_tree_block_info *info;
1804
1805         if (!*ptr) {
1806                 /* first call */
1807                 flags = btrfs_extent_flags(eb, ei);
1808                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1809                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1810                                 /* a skinny metadata extent */
1811                                 *out_eiref =
1812                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1813                         } else {
1814                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1815                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1816                                 *out_eiref =
1817                                    (struct btrfs_extent_inline_ref *)(info + 1);
1818                         }
1819                 } else {
1820                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1821                 }
1822                 *ptr = (unsigned long)*out_eiref;
1823                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1824                         return -ENOENT;
1825         }
1826
1827         end = (unsigned long)ei + item_size;
1828         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1829         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1830                                                      BTRFS_REF_TYPE_ANY);
1831         if (*out_type == BTRFS_REF_TYPE_INVALID)
1832                 return -EINVAL;
1833
1834         *ptr += btrfs_extent_inline_ref_size(*out_type);
1835         WARN_ON(*ptr > end);
1836         if (*ptr == end)
1837                 return 1; /* last */
1838
1839         return 0;
1840 }
1841
1842 /*
1843  * reads the tree block backref for an extent. tree level and root are returned
1844  * through out_level and out_root. ptr must point to a 0 value for the first
1845  * call and may be modified (see get_extent_inline_ref comment).
1846  * returns 0 if data was provided, 1 if there was no more data to provide or
1847  * <0 on error.
1848  */
1849 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1850                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1851                             u32 item_size, u64 *out_root, u8 *out_level)
1852 {
1853         int ret;
1854         int type;
1855         struct btrfs_extent_inline_ref *eiref;
1856
1857         if (*ptr == (unsigned long)-1)
1858                 return 1;
1859
1860         while (1) {
1861                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1862                                               &eiref, &type);
1863                 if (ret < 0)
1864                         return ret;
1865
1866                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1867                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1868                         break;
1869
1870                 if (ret == 1)
1871                         return 1;
1872         }
1873
1874         /* we can treat both ref types equally here */
1875         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1876
1877         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1878                 struct btrfs_tree_block_info *info;
1879
1880                 info = (struct btrfs_tree_block_info *)(ei + 1);
1881                 *out_level = btrfs_tree_block_level(eb, info);
1882         } else {
1883                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1884                 *out_level = (u8)key->offset;
1885         }
1886
1887         if (ret == 1)
1888                 *ptr = (unsigned long)-1;
1889
1890         return 0;
1891 }
1892
1893 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1894                              struct extent_inode_elem *inode_list,
1895                              u64 root, u64 extent_item_objectid,
1896                              iterate_extent_inodes_t *iterate, void *ctx)
1897 {
1898         struct extent_inode_elem *eie;
1899         int ret = 0;
1900
1901         for (eie = inode_list; eie; eie = eie->next) {
1902                 btrfs_debug(fs_info,
1903                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1904                             extent_item_objectid, eie->inum,
1905                             eie->offset, root);
1906                 ret = iterate(eie->inum, eie->offset, root, ctx);
1907                 if (ret) {
1908                         btrfs_debug(fs_info,
1909                                     "stopping iteration for %llu due to ret=%d",
1910                                     extent_item_objectid, ret);
1911                         break;
1912                 }
1913         }
1914
1915         return ret;
1916 }
1917
1918 /*
1919  * calls iterate() for every inode that references the extent identified by
1920  * the given parameters.
1921  * when the iterator function returns a non-zero value, iteration stops.
1922  */
1923 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1924                                 u64 extent_item_objectid, u64 extent_item_pos,
1925                                 int search_commit_root,
1926                                 iterate_extent_inodes_t *iterate, void *ctx)
1927 {
1928         int ret;
1929         struct btrfs_trans_handle *trans = NULL;
1930         struct ulist *refs = NULL;
1931         struct ulist *roots = NULL;
1932         struct ulist_node *ref_node = NULL;
1933         struct ulist_node *root_node = NULL;
1934         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1935         struct ulist_iterator ref_uiter;
1936         struct ulist_iterator root_uiter;
1937
1938         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1939                         extent_item_objectid);
1940
1941         if (!search_commit_root) {
1942                 trans = btrfs_attach_transaction(fs_info->extent_root);
1943                 if (IS_ERR(trans)) {
1944                         if (PTR_ERR(trans) != -ENOENT &&
1945                             PTR_ERR(trans) != -EROFS)
1946                                 return PTR_ERR(trans);
1947                         trans = NULL;
1948                 }
1949         }
1950
1951         if (trans)
1952                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1953         else
1954                 down_read(&fs_info->commit_root_sem);
1955
1956         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1957                                    tree_mod_seq_elem.seq, &refs,
1958                                    &extent_item_pos);
1959         if (ret)
1960                 goto out;
1961
1962         ULIST_ITER_INIT(&ref_uiter);
1963         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1964                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1965                                                 tree_mod_seq_elem.seq, &roots);
1966                 if (ret)
1967                         break;
1968                 ULIST_ITER_INIT(&root_uiter);
1969                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1970                         btrfs_debug(fs_info,
1971                                     "root %llu references leaf %llu, data list %#llx",
1972                                     root_node->val, ref_node->val,
1973                                     ref_node->aux);
1974                         ret = iterate_leaf_refs(fs_info,
1975                                                 (struct extent_inode_elem *)
1976                                                 (uintptr_t)ref_node->aux,
1977                                                 root_node->val,
1978                                                 extent_item_objectid,
1979                                                 iterate, ctx);
1980                 }
1981                 ulist_free(roots);
1982         }
1983
1984         free_leaf_list(refs);
1985 out:
1986         if (trans) {
1987                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1988                 btrfs_end_transaction(trans);
1989         } else {
1990                 up_read(&fs_info->commit_root_sem);
1991         }
1992
1993         return ret;
1994 }
1995
1996 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1997                                 struct btrfs_path *path,
1998                                 iterate_extent_inodes_t *iterate, void *ctx)
1999 {
2000         int ret;
2001         u64 extent_item_pos;
2002         u64 flags = 0;
2003         struct btrfs_key found_key;
2004         int search_commit_root = path->search_commit_root;
2005
2006         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2007         btrfs_release_path(path);
2008         if (ret < 0)
2009                 return ret;
2010         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2011                 return -EINVAL;
2012
2013         extent_item_pos = logical - found_key.objectid;
2014         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2015                                         extent_item_pos, search_commit_root,
2016                                         iterate, ctx);
2017
2018         return ret;
2019 }
2020
2021 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2022                               struct extent_buffer *eb, void *ctx);
2023
2024 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2025                               struct btrfs_path *path,
2026                               iterate_irefs_t *iterate, void *ctx)
2027 {
2028         int ret = 0;
2029         int slot;
2030         u32 cur;
2031         u32 len;
2032         u32 name_len;
2033         u64 parent = 0;
2034         int found = 0;
2035         struct extent_buffer *eb;
2036         struct btrfs_item *item;
2037         struct btrfs_inode_ref *iref;
2038         struct btrfs_key found_key;
2039
2040         while (!ret) {
2041                 ret = btrfs_find_item(fs_root, path, inum,
2042                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2043                                 &found_key);
2044
2045                 if (ret < 0)
2046                         break;
2047                 if (ret) {
2048                         ret = found ? 0 : -ENOENT;
2049                         break;
2050                 }
2051                 ++found;
2052
2053                 parent = found_key.offset;
2054                 slot = path->slots[0];
2055                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2056                 if (!eb) {
2057                         ret = -ENOMEM;
2058                         break;
2059                 }
2060                 extent_buffer_get(eb);
2061                 btrfs_tree_read_lock(eb);
2062                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2063                 btrfs_release_path(path);
2064
2065                 item = btrfs_item_nr(slot);
2066                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2067
2068                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2069                         name_len = btrfs_inode_ref_name_len(eb, iref);
2070                         /* path must be released before calling iterate()! */
2071                         btrfs_debug(fs_root->fs_info,
2072                                 "following ref at offset %u for inode %llu in tree %llu",
2073                                 cur, found_key.objectid, fs_root->objectid);
2074                         ret = iterate(parent, name_len,
2075                                       (unsigned long)(iref + 1), eb, ctx);
2076                         if (ret)
2077                                 break;
2078                         len = sizeof(*iref) + name_len;
2079                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2080                 }
2081                 btrfs_tree_read_unlock_blocking(eb);
2082                 free_extent_buffer(eb);
2083         }
2084
2085         btrfs_release_path(path);
2086
2087         return ret;
2088 }
2089
2090 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2091                                  struct btrfs_path *path,
2092                                  iterate_irefs_t *iterate, void *ctx)
2093 {
2094         int ret;
2095         int slot;
2096         u64 offset = 0;
2097         u64 parent;
2098         int found = 0;
2099         struct extent_buffer *eb;
2100         struct btrfs_inode_extref *extref;
2101         u32 item_size;
2102         u32 cur_offset;
2103         unsigned long ptr;
2104
2105         while (1) {
2106                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2107                                             &offset);
2108                 if (ret < 0)
2109                         break;
2110                 if (ret) {
2111                         ret = found ? 0 : -ENOENT;
2112                         break;
2113                 }
2114                 ++found;
2115
2116                 slot = path->slots[0];
2117                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2118                 if (!eb) {
2119                         ret = -ENOMEM;
2120                         break;
2121                 }
2122                 extent_buffer_get(eb);
2123
2124                 btrfs_tree_read_lock(eb);
2125                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2126                 btrfs_release_path(path);
2127
2128                 item_size = btrfs_item_size_nr(eb, slot);
2129                 ptr = btrfs_item_ptr_offset(eb, slot);
2130                 cur_offset = 0;
2131
2132                 while (cur_offset < item_size) {
2133                         u32 name_len;
2134
2135                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2136                         parent = btrfs_inode_extref_parent(eb, extref);
2137                         name_len = btrfs_inode_extref_name_len(eb, extref);
2138                         ret = iterate(parent, name_len,
2139                                       (unsigned long)&extref->name, eb, ctx);
2140                         if (ret)
2141                                 break;
2142
2143                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2144                         cur_offset += sizeof(*extref);
2145                 }
2146                 btrfs_tree_read_unlock_blocking(eb);
2147                 free_extent_buffer(eb);
2148
2149                 offset++;
2150         }
2151
2152         btrfs_release_path(path);
2153
2154         return ret;
2155 }
2156
2157 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2158                          struct btrfs_path *path, iterate_irefs_t *iterate,
2159                          void *ctx)
2160 {
2161         int ret;
2162         int found_refs = 0;
2163
2164         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2165         if (!ret)
2166                 ++found_refs;
2167         else if (ret != -ENOENT)
2168                 return ret;
2169
2170         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2171         if (ret == -ENOENT && found_refs)
2172                 return 0;
2173
2174         return ret;
2175 }
2176
2177 /*
2178  * returns 0 if the path could be dumped (probably truncated)
2179  * returns <0 in case of an error
2180  */
2181 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2182                          struct extent_buffer *eb, void *ctx)
2183 {
2184         struct inode_fs_paths *ipath = ctx;
2185         char *fspath;
2186         char *fspath_min;
2187         int i = ipath->fspath->elem_cnt;
2188         const int s_ptr = sizeof(char *);
2189         u32 bytes_left;
2190
2191         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2192                                         ipath->fspath->bytes_left - s_ptr : 0;
2193
2194         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2195         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2196                                    name_off, eb, inum, fspath_min, bytes_left);
2197         if (IS_ERR(fspath))
2198                 return PTR_ERR(fspath);
2199
2200         if (fspath > fspath_min) {
2201                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2202                 ++ipath->fspath->elem_cnt;
2203                 ipath->fspath->bytes_left = fspath - fspath_min;
2204         } else {
2205                 ++ipath->fspath->elem_missed;
2206                 ipath->fspath->bytes_missing += fspath_min - fspath;
2207                 ipath->fspath->bytes_left = 0;
2208         }
2209
2210         return 0;
2211 }
2212
2213 /*
2214  * this dumps all file system paths to the inode into the ipath struct, provided
2215  * is has been created large enough. each path is zero-terminated and accessed
2216  * from ipath->fspath->val[i].
2217  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2218  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2219  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2220  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2221  * have been needed to return all paths.
2222  */
2223 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2224 {
2225         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2226                              inode_to_path, ipath);
2227 }
2228
2229 struct btrfs_data_container *init_data_container(u32 total_bytes)
2230 {
2231         struct btrfs_data_container *data;
2232         size_t alloc_bytes;
2233
2234         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2235         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2236         if (!data)
2237                 return ERR_PTR(-ENOMEM);
2238
2239         if (total_bytes >= sizeof(*data)) {
2240                 data->bytes_left = total_bytes - sizeof(*data);
2241                 data->bytes_missing = 0;
2242         } else {
2243                 data->bytes_missing = sizeof(*data) - total_bytes;
2244                 data->bytes_left = 0;
2245         }
2246
2247         data->elem_cnt = 0;
2248         data->elem_missed = 0;
2249
2250         return data;
2251 }
2252
2253 /*
2254  * allocates space to return multiple file system paths for an inode.
2255  * total_bytes to allocate are passed, note that space usable for actual path
2256  * information will be total_bytes - sizeof(struct inode_fs_paths).
2257  * the returned pointer must be freed with free_ipath() in the end.
2258  */
2259 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2260                                         struct btrfs_path *path)
2261 {
2262         struct inode_fs_paths *ifp;
2263         struct btrfs_data_container *fspath;
2264
2265         fspath = init_data_container(total_bytes);
2266         if (IS_ERR(fspath))
2267                 return (void *)fspath;
2268
2269         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2270         if (!ifp) {
2271                 kvfree(fspath);
2272                 return ERR_PTR(-ENOMEM);
2273         }
2274
2275         ifp->btrfs_path = path;
2276         ifp->fspath = fspath;
2277         ifp->fs_root = fs_root;
2278
2279         return ifp;
2280 }
2281
2282 void free_ipath(struct inode_fs_paths *ipath)
2283 {
2284         if (!ipath)
2285                 return;
2286         kvfree(ipath->fspath);
2287         kfree(ipath);
2288 }