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