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