fs/btrfs/tree-log.c

   1 /*
   2  * Copyright (C) 2008 Oracle.  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/sched.h>
  20 #include <linux/slab.h>
  21 #include <linux/list_sort.h>
  22 #include "ctree.h"
  23 #include "transaction.h"
  24 #include "disk-io.h"
  25 #include "locking.h"
  26 #include "print-tree.h"
  27 #include "backref.h"
  28 #include "compat.h"
  29 #include "tree-log.h"
  30 #include "hash.h"
  31
  32 /* magic values for the inode_only field in btrfs_log_inode:
  33  *
  34  * LOG_INODE_ALL means to log everything
  35  * LOG_INODE_EXISTS means to log just enough to recreate the inode
  36  * during log replay
  37  */
  38 #define LOG_INODE_ALL 0
  39 #define LOG_INODE_EXISTS 1
  40
  41 /*
  42  * directory trouble cases
  43  *
  44  * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
  45  * log, we must force a full commit before doing an fsync of the directory
  46  * where the unlink was done.
  47  * ---> record transid of last unlink/rename per directory
  48  *
  49  * mkdir foo/some_dir
  50  * normal commit
  51  * rename foo/some_dir foo2/some_dir
  52  * mkdir foo/some_dir
  53  * fsync foo/some_dir/some_file
  54  *
  55  * The fsync above will unlink the original some_dir without recording
  56  * it in its new location (foo2).  After a crash, some_dir will be gone
  57  * unless the fsync of some_file forces a full commit
  58  *
  59  * 2) we must log any new names for any file or dir that is in the fsync
  60  * log. ---> check inode while renaming/linking.
  61  *
  62  * 2a) we must log any new names for any file or dir during rename
  63  * when the directory they are being removed from was logged.
  64  * ---> check inode and old parent dir during rename
  65  *
  66  *  2a is actually the more important variant.  With the extra logging
  67  *  a crash might unlink the old name without recreating the new one
  68  *
  69  * 3) after a crash, we must go through any directories with a link count
  70  * of zero and redo the rm -rf
  71  *
  72  * mkdir f1/foo
  73  * normal commit
  74  * rm -rf f1/foo
  75  * fsync(f1)
  76  *
  77  * The directory f1 was fully removed from the FS, but fsync was never
  78  * called on f1, only its parent dir.  After a crash the rm -rf must
  79  * be replayed.  This must be able to recurse down the entire
  80  * directory tree.  The inode link count fixup code takes care of the
  81  * ugly details.
  82  */
  83
  84 /*
  85  * stages for the tree walking.  The first
  86  * stage (0) is to only pin down the blocks we find
  87  * the second stage (1) is to make sure that all the inodes
  88  * we find in the log are created in the subvolume.
  89  *
  90  * The last stage is to deal with directories and links and extents
  91  * and all the other fun semantics
  92  */
  93 #define LOG_WALK_PIN_ONLY 0
  94 #define LOG_WALK_REPLAY_INODES 1
  95 #define LOG_WALK_REPLAY_ALL 2
  96
  97 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
  98                              struct btrfs_root *root, struct inode *inode,
  99                              int inode_only);
 100 static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
 101                              struct btrfs_root *root,
 102                              struct btrfs_path *path, u64 objectid);
 103 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
 104                                        struct btrfs_root *root,
 105                                        struct btrfs_root *log,
 106                                        struct btrfs_path *path,
 107                                        u64 dirid, int del_all);
 108
 109 /*
 110  * tree logging is a special write ahead log used to make sure that
 111  * fsyncs and O_SYNCs can happen without doing full tree commits.
 112  *
 113  * Full tree commits are expensive because they require commonly
 114  * modified blocks to be recowed, creating many dirty pages in the
 115  * extent tree an 4x-6x higher write load than ext3.
 116  *
 117  * Instead of doing a tree commit on every fsync, we use the
 118  * key ranges and transaction ids to find items for a given file or directory
 119  * that have changed in this transaction.  Those items are copied into
 120  * a special tree (one per subvolume root), that tree is written to disk
 121  * and then the fsync is considered complete.
 122  *
 123  * After a crash, items are copied out of the log-tree back into the
 124  * subvolume tree.  Any file data extents found are recorded in the extent
 125  * allocation tree, and the log-tree freed.
 126  *
 127  * The log tree is read three times, once to pin down all the extents it is
 128  * using in ram and once, once to create all the inodes logged in the tree
 129  * and once to do all the other items.
 130  */
 131
 132 /*
 133  * start a sub transaction and setup the log tree
 134  * this increments the log tree writer count to make the people
 135  * syncing the tree wait for us to finish
 136  */
 137 static int start_log_trans(struct btrfs_trans_handle *trans,
 138                            struct btrfs_root *root)
 139 {
 140         int ret;
 141         int err = 0;
 142
 143         mutex_lock(&root->log_mutex);
 144         if (root->log_root) {
 145                 if (!root->log_start_pid) {
 146                         root->log_start_pid = current->pid;
 147                         root->log_multiple_pids = false;
 148                 } else if (root->log_start_pid != current->pid) {
 149                         root->log_multiple_pids = true;
 150                 }
 151
 152                 atomic_inc(&root->log_batch);
 153                 atomic_inc(&root->log_writers);
 154                 mutex_unlock(&root->log_mutex);
 155                 return 0;
 156         }
 157         root->log_multiple_pids = false;
 158         root->log_start_pid = current->pid;
 159         mutex_lock(&root->fs_info->tree_log_mutex);
 160         if (!root->fs_info->log_root_tree) {
 161                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
 162                 if (ret)
 163                         err = ret;
 164         }
 165         if (err == 0 && !root->log_root) {
 166                 ret = btrfs_add_log_tree(trans, root);
 167                 if (ret)
 168                         err = ret;
 169         }
 170         mutex_unlock(&root->fs_info->tree_log_mutex);
 171         atomic_inc(&root->log_batch);
 172         atomic_inc(&root->log_writers);
 173         mutex_unlock(&root->log_mutex);
 174         return err;
 175 }
 176
 177 /*
 178  * returns 0 if there was a log transaction running and we were able
 179  * to join, or returns -ENOENT if there were not transactions
 180  * in progress
 181  */
 182 static int join_running_log_trans(struct btrfs_root *root)
 183 {
 184         int ret = -ENOENT;
 185
 186         smp_mb();
 187         if (!root->log_root)
 188                 return -ENOENT;
 189
 190         mutex_lock(&root->log_mutex);
 191         if (root->log_root) {
 192                 ret = 0;
 193                 atomic_inc(&root->log_writers);
 194         }
 195         mutex_unlock(&root->log_mutex);
 196         return ret;
 197 }
 198
 199 /*
 200  * This either makes the current running log transaction wait
 201  * until you call btrfs_end_log_trans() or it makes any future
 202  * log transactions wait until you call btrfs_end_log_trans()
 203  */
 204 int btrfs_pin_log_trans(struct btrfs_root *root)
 205 {
 206         int ret = -ENOENT;
 207
 208         mutex_lock(&root->log_mutex);
 209         atomic_inc(&root->log_writers);
 210         mutex_unlock(&root->log_mutex);
 211         return ret;
 212 }
 213
 214 /*
 215  * indicate we're done making changes to the log tree
 216  * and wake up anyone waiting to do a sync
 217  */
 218 void btrfs_end_log_trans(struct btrfs_root *root)
 219 {
 220         if (atomic_dec_and_test(&root->log_writers)) {
 221                 smp_mb();
 222                 if (waitqueue_active(&root->log_writer_wait))
 223                         wake_up(&root->log_writer_wait);
 224         }
 225 }
 226
 227
 228 /*
 229  * the walk control struct is used to pass state down the chain when
 230  * processing the log tree.  The stage field tells us which part
 231  * of the log tree processing we are currently doing.  The others
 232  * are state fields used for that specific part
 233  */
 234 struct walk_control {
 235         /* should we free the extent on disk when done?  This is used
 236          * at transaction commit time while freeing a log tree
 237          */
 238         int free;
 239
 240         /* should we write out the extent buffer?  This is used
 241          * while flushing the log tree to disk during a sync
 242          */
 243         int write;
 244
 245         /* should we wait for the extent buffer io to finish?  Also used
 246          * while flushing the log tree to disk for a sync
 247          */
 248         int wait;
 249
 250         /* pin only walk, we record which extents on disk belong to the
 251          * log trees
 252          */
 253         int pin;
 254
 255         /* what stage of the replay code we're currently in */
 256         int stage;
 257
 258         /* the root we are currently replaying */
 259         struct btrfs_root *replay_dest;
 260
 261         /* the trans handle for the current replay */
 262         struct btrfs_trans_handle *trans;
 263
 264         /* the function that gets used to process blocks we find in the
 265          * tree.  Note the extent_buffer might not be up to date when it is
 266          * passed in, and it must be checked or read if you need the data
 267          * inside it
 268          */
 269         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
 270                             struct walk_control *wc, u64 gen);
 271 };
 272
 273 /*
 274  * process_func used to pin down extents, write them or wait on them
 275  */
 276 static int process_one_buffer(struct btrfs_root *log,
 277                               struct extent_buffer *eb,
 278                               struct walk_control *wc, u64 gen)
 279 {
 280         if (wc->pin)
 281                 btrfs_pin_extent_for_log_replay(log->fs_info->extent_root,
 282                                                 eb->start, eb->len);
 283
 284         if (btrfs_buffer_uptodate(eb, gen, 0)) {
 285                 if (wc->write)
 286                         btrfs_write_tree_block(eb);
 287                 if (wc->wait)
 288                         btrfs_wait_tree_block_writeback(eb);
 289         }
 290         return 0;
 291 }
 292
 293 /*
 294  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
 295  * to the src data we are copying out.
 296  *
 297  * root is the tree we are copying into, and path is a scratch
 298  * path for use in this function (it should be released on entry and
 299  * will be released on exit).
 300  *
 301  * If the key is already in the destination tree the existing item is
 302  * overwritten.  If the existing item isn't big enough, it is extended.
 303  * If it is too large, it is truncated.
 304  *
 305  * If the key isn't in the destination yet, a new item is inserted.
 306  */
 307 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
 308                                    struct btrfs_root *root,
 309                                    struct btrfs_path *path,
 310                                    struct extent_buffer *eb, int slot,
 311                                    struct btrfs_key *key)
 312 {
 313         int ret;
 314         u32 item_size;
 315         u64 saved_i_size = 0;
 316         int save_old_i_size = 0;
 317         unsigned long src_ptr;
 318         unsigned long dst_ptr;
 319         int overwrite_root = 0;
 320
 321         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
 322                 overwrite_root = 1;
 323
 324         item_size = btrfs_item_size_nr(eb, slot);
 325         src_ptr = btrfs_item_ptr_offset(eb, slot);
 326
 327         /* look for the key in the destination tree */
 328         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
 329         if (ret == 0) {
 330                 char *src_copy;
 331                 char *dst_copy;
 332                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
 333                                                   path->slots[0]);
 334                 if (dst_size != item_size)
 335                         goto insert;
 336
 337                 if (item_size == 0) {
 338                         btrfs_release_path(path);
 339                         return 0;
 340                 }
 341                 dst_copy = kmalloc(item_size, GFP_NOFS);
 342                 src_copy = kmalloc(item_size, GFP_NOFS);
 343                 if (!dst_copy || !src_copy) {
 344                         btrfs_release_path(path);
 345                         kfree(dst_copy);
 346                         kfree(src_copy);
 347                         return -ENOMEM;
 348                 }
 349
 350                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
 351
 352                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 353                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
 354                                    item_size);
 355                 ret = memcmp(dst_copy, src_copy, item_size);
 356
 357                 kfree(dst_copy);
 358                 kfree(src_copy);
 359                 /*
 360                  * they have the same contents, just return, this saves
 361                  * us from cowing blocks in the destination tree and doing
 362                  * extra writes that may not have been done by a previous
 363                  * sync
 364                  */
 365                 if (ret == 0) {
 366                         btrfs_release_path(path);
 367                         return 0;
 368                 }
 369
 370         }
 371 insert:
 372         btrfs_release_path(path);
 373         /* try to insert the key into the destination tree */
 374         ret = btrfs_insert_empty_item(trans, root, path,
 375                                       key, item_size);
 376
 377         /* make sure any existing item is the correct size */
 378         if (ret == -EEXIST) {
 379                 u32 found_size;
 380                 found_size = btrfs_item_size_nr(path->nodes[0],
 381                                                 path->slots[0]);
 382                 if (found_size > item_size)
 383                         btrfs_truncate_item(trans, root, path, item_size, 1);
 384                 else if (found_size < item_size)
 385                         btrfs_extend_item(trans, root, path,
 386                                           item_size - found_size);
 387         } else if (ret) {
 388                 return ret;
 389         }
 390         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
 391                                         path->slots[0]);
 392
 393         /* don't overwrite an existing inode if the generation number
 394          * was logged as zero.  This is done when the tree logging code
 395          * is just logging an inode to make sure it exists after recovery.
 396          *
 397          * Also, don't overwrite i_size on directories during replay.
 398          * log replay inserts and removes directory items based on the
 399          * state of the tree found in the subvolume, and i_size is modified
 400          * as it goes
 401          */
 402         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
 403                 struct btrfs_inode_item *src_item;
 404                 struct btrfs_inode_item *dst_item;
 405
 406                 src_item = (struct btrfs_inode_item *)src_ptr;
 407                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 408
 409                 if (btrfs_inode_generation(eb, src_item) == 0)
 410                         goto no_copy;
 411
 412                 if (overwrite_root &&
 413                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
 414                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
 415                         save_old_i_size = 1;
 416                         saved_i_size = btrfs_inode_size(path->nodes[0],
 417                                                         dst_item);
 418                 }
 419         }
 420
 421         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
 422                            src_ptr, item_size);
 423
 424         if (save_old_i_size) {
 425                 struct btrfs_inode_item *dst_item;
 426                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 427                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
 428         }
 429
 430         /* make sure the generation is filled in */
 431         if (key->type == BTRFS_INODE_ITEM_KEY) {
 432                 struct btrfs_inode_item *dst_item;
 433                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 434                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
 435                         btrfs_set_inode_generation(path->nodes[0], dst_item,
 436                                                    trans->transid);
 437                 }
 438         }
 439 no_copy:
 440         btrfs_mark_buffer_dirty(path->nodes[0]);
 441         btrfs_release_path(path);
 442         return 0;
 443 }
 444
 445 /*
 446  * simple helper to read an inode off the disk from a given root
 447  * This can only be called for subvolume roots and not for the log
 448  */
 449 static noinline struct inode *read_one_inode(struct btrfs_root *root,
 450                                              u64 objectid)
 451 {
 452         struct btrfs_key key;
 453         struct inode *inode;
 454
 455         key.objectid = objectid;
 456         key.type = BTRFS_INODE_ITEM_KEY;
 457         key.offset = 0;
 458         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
 459         if (IS_ERR(inode)) {
 460                 inode = NULL;
 461         } else if (is_bad_inode(inode)) {
 462                 iput(inode);
 463                 inode = NULL;
 464         }
 465         return inode;
 466 }
 467
 468 /* replays a single extent in 'eb' at 'slot' with 'key' into the
 469  * subvolume 'root'.  path is released on entry and should be released
 470  * on exit.
 471  *
 472  * extents in the log tree have not been allocated out of the extent
 473  * tree yet.  So, this completes the allocation, taking a reference
 474  * as required if the extent already exists or creating a new extent
 475  * if it isn't in the extent allocation tree yet.
 476  *
 477  * The extent is inserted into the file, dropping any existing extents
 478  * from the file that overlap the new one.
 479  */
 480 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
 481                                       struct btrfs_root *root,
 482                                       struct btrfs_path *path,
 483                                       struct extent_buffer *eb, int slot,
 484                                       struct btrfs_key *key)
 485 {
 486         int found_type;
 487         u64 extent_end;
 488         u64 start = key->offset;
 489         u64 saved_nbytes;
 490         struct btrfs_file_extent_item *item;
 491         struct inode *inode = NULL;
 492         unsigned long size;
 493         int ret = 0;
 494
 495         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 496         found_type = btrfs_file_extent_type(eb, item);
 497
 498         if (found_type == BTRFS_FILE_EXTENT_REG ||
 499             found_type == BTRFS_FILE_EXTENT_PREALLOC)
 500                 extent_end = start + btrfs_file_extent_num_bytes(eb, item);
 501         else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 502                 size = btrfs_file_extent_inline_len(eb, item);
 503                 extent_end = ALIGN(start + size, root->sectorsize);
 504         } else {
 505                 ret = 0;
 506                 goto out;
 507         }
 508
 509         inode = read_one_inode(root, key->objectid);
 510         if (!inode) {
 511                 ret = -EIO;
 512                 goto out;
 513         }
 514
 515         /*
 516          * first check to see if we already have this extent in the
 517          * file.  This must be done before the btrfs_drop_extents run
 518          * so we don't try to drop this extent.
 519          */
 520         ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
 521                                        start, 0);
 522
 523         if (ret == 0 &&
 524             (found_type == BTRFS_FILE_EXTENT_REG ||
 525              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
 526                 struct btrfs_file_extent_item cmp1;
 527                 struct btrfs_file_extent_item cmp2;
 528                 struct btrfs_file_extent_item *existing;
 529                 struct extent_buffer *leaf;
 530
 531                 leaf = path->nodes[0];
 532                 existing = btrfs_item_ptr(leaf, path->slots[0],
 533                                           struct btrfs_file_extent_item);
 534
 535                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
 536                                    sizeof(cmp1));
 537                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
 538                                    sizeof(cmp2));
 539
 540                 /*
 541                  * we already have a pointer to this exact extent,
 542                  * we don't have to do anything
 543                  */
 544                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
 545                         btrfs_release_path(path);
 546                         goto out;
 547                 }
 548         }
 549         btrfs_release_path(path);
 550
 551         saved_nbytes = inode_get_bytes(inode);
 552         /* drop any overlapping extents */
 553         ret = btrfs_drop_extents(trans, root, inode, start, extent_end, 1);
 554         BUG_ON(ret);
 555
 556         if (found_type == BTRFS_FILE_EXTENT_REG ||
 557             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 558                 u64 offset;
 559                 unsigned long dest_offset;
 560                 struct btrfs_key ins;
 561
 562                 ret = btrfs_insert_empty_item(trans, root, path, key,
 563                                               sizeof(*item));
 564                 BUG_ON(ret);
 565                 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
 566                                                     path->slots[0]);
 567                 copy_extent_buffer(path->nodes[0], eb, dest_offset,
 568                                 (unsigned long)item,  sizeof(*item));
 569
 570                 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
 571                 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
 572                 ins.type = BTRFS_EXTENT_ITEM_KEY;
 573                 offset = key->offset - btrfs_file_extent_offset(eb, item);
 574
 575                 if (ins.objectid > 0) {
 576                         u64 csum_start;
 577                         u64 csum_end;
 578                         LIST_HEAD(ordered_sums);
 579                         /*
 580                          * is this extent already allocated in the extent
 581                          * allocation tree?  If so, just add a reference
 582                          */
 583                         ret = btrfs_lookup_extent(root, ins.objectid,
 584                                                 ins.offset);
 585                         if (ret == 0) {
 586                                 ret = btrfs_inc_extent_ref(trans, root,
 587                                                 ins.objectid, ins.offset,
 588                                                 0, root->root_key.objectid,
 589                                                 key->objectid, offset, 0);
 590                                 BUG_ON(ret);
 591                         } else {
 592                                 /*
 593                                  * insert the extent pointer in the extent
 594                                  * allocation tree
 595                                  */
 596                                 ret = btrfs_alloc_logged_file_extent(trans,
 597                                                 root, root->root_key.objectid,
 598                                                 key->objectid, offset, &ins);
 599                                 BUG_ON(ret);
 600                         }
 601                         btrfs_release_path(path);
 602
 603                         if (btrfs_file_extent_compression(eb, item)) {
 604                                 csum_start = ins.objectid;
 605                                 csum_end = csum_start + ins.offset;
 606                         } else {
 607                                 csum_start = ins.objectid +
 608                                         btrfs_file_extent_offset(eb, item);
 609                                 csum_end = csum_start +
 610                                         btrfs_file_extent_num_bytes(eb, item);
 611                         }
 612
 613                         ret = btrfs_lookup_csums_range(root->log_root,
 614                                                 csum_start, csum_end - 1,
 615                                                 &ordered_sums, 0);
 616                         BUG_ON(ret);
 617                         while (!list_empty(&ordered_sums)) {
 618                                 struct btrfs_ordered_sum *sums;
 619                                 sums = list_entry(ordered_sums.next,
 620                                                 struct btrfs_ordered_sum,
 621                                                 list);
 622                                 ret = btrfs_csum_file_blocks(trans,
 623                                                 root->fs_info->csum_root,
 624                                                 sums);
 625                                 BUG_ON(ret);
 626                                 list_del(&sums->list);
 627                                 kfree(sums);
 628                         }
 629                 } else {
 630                         btrfs_release_path(path);
 631                 }
 632         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 633                 /* inline extents are easy, we just overwrite them */
 634                 ret = overwrite_item(trans, root, path, eb, slot, key);
 635                 BUG_ON(ret);
 636         }
 637
 638         inode_set_bytes(inode, saved_nbytes);
 639         ret = btrfs_update_inode(trans, root, inode);
 640 out:
 641         if (inode)
 642                 iput(inode);
 643         return ret;
 644 }
 645
 646 /*
 647  * when cleaning up conflicts between the directory names in the
 648  * subvolume, directory names in the log and directory names in the
 649  * inode back references, we may have to unlink inodes from directories.
 650  *
 651  * This is a helper function to do the unlink of a specific directory
 652  * item
 653  */
 654 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
 655                                       struct btrfs_root *root,
 656                                       struct btrfs_path *path,
 657                                       struct inode *dir,
 658                                       struct btrfs_dir_item *di)
 659 {
 660         struct inode *inode;
 661         char *name;
 662         int name_len;
 663         struct extent_buffer *leaf;
 664         struct btrfs_key location;
 665         int ret;
 666
 667         leaf = path->nodes[0];
 668
 669         btrfs_dir_item_key_to_cpu(leaf, di, &location);
 670         name_len = btrfs_dir_name_len(leaf, di);
 671         name = kmalloc(name_len, GFP_NOFS);
 672         if (!name)
 673                 return -ENOMEM;
 674
 675         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
 676         btrfs_release_path(path);
 677
 678         inode = read_one_inode(root, location.objectid);
 679         if (!inode) {
 680                 kfree(name);
 681                 return -EIO;
 682         }
 683
 684         ret = link_to_fixup_dir(trans, root, path, location.objectid);
 685         BUG_ON(ret);
 686
 687         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
 688         BUG_ON(ret);
 689         kfree(name);
 690
 691         iput(inode);
 692
 693         btrfs_run_delayed_items(trans, root);
 694         return ret;
 695 }
 696
 697 /*
 698  * helper function to see if a given name and sequence number found
 699  * in an inode back reference are already in a directory and correctly
 700  * point to this inode
 701  */
 702 static noinline int inode_in_dir(struct btrfs_root *root,
 703                                  struct btrfs_path *path,
 704                                  u64 dirid, u64 objectid, u64 index,
 705                                  const char *name, int name_len)
 706 {
 707         struct btrfs_dir_item *di;
 708         struct btrfs_key location;
 709         int match = 0;
 710
 711         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
 712                                          index, name, name_len, 0);
 713         if (di && !IS_ERR(di)) {
 714                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 715                 if (location.objectid != objectid)
 716                         goto out;
 717         } else
 718                 goto out;
 719         btrfs_release_path(path);
 720
 721         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
 722         if (di && !IS_ERR(di)) {
 723                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 724                 if (location.objectid != objectid)
 725                         goto out;
 726         } else
 727                 goto out;
 728         match = 1;
 729 out:
 730         btrfs_release_path(path);
 731         return match;
 732 }
 733
 734 /*
 735  * helper function to check a log tree for a named back reference in
 736  * an inode.  This is used to decide if a back reference that is
 737  * found in the subvolume conflicts with what we find in the log.
 738  *
 739  * inode backreferences may have multiple refs in a single item,
 740  * during replay we process one reference at a time, and we don't
 741  * want to delete valid links to a file from the subvolume if that
 742  * link is also in the log.
 743  */
 744 static noinline int backref_in_log(struct btrfs_root *log,
 745                                    struct btrfs_key *key,
 746                                    u64 ref_objectid,
 747                                    char *name, int namelen)
 748 {
 749         struct btrfs_path *path;
 750         struct btrfs_inode_ref *ref;
 751         unsigned long ptr;
 752         unsigned long ptr_end;
 753         unsigned long name_ptr;
 754         int found_name_len;
 755         int item_size;
 756         int ret;
 757         int match = 0;
 758
 759         path = btrfs_alloc_path();
 760         if (!path)
 761                 return -ENOMEM;
 762
 763         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
 764         if (ret != 0)
 765                 goto out;
 766
 767         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 768
 769         if (key->type == BTRFS_INODE_EXTREF_KEY) {
 770                 if (btrfs_find_name_in_ext_backref(path, ref_objectid,
 771                                                    name, namelen, NULL))
 772                         match = 1;
 773
 774                 goto out;
 775         }
 776
 777         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
 778         ptr_end = ptr + item_size;
 779         while (ptr < ptr_end) {
 780                 ref = (struct btrfs_inode_ref *)ptr;
 781                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
 782                 if (found_name_len == namelen) {
 783                         name_ptr = (unsigned long)(ref + 1);
 784                         ret = memcmp_extent_buffer(path->nodes[0], name,
 785                                                    name_ptr, namelen);
 786                         if (ret == 0) {
 787                                 match = 1;
 788                                 goto out;
 789                         }
 790                 }
 791                 ptr = (unsigned long)(ref + 1) + found_name_len;
 792         }
 793 out:
 794         btrfs_free_path(path);
 795         return match;
 796 }
 797
 798 static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
 799                                   struct btrfs_root *root,
 800                                   struct btrfs_path *path,
 801                                   struct btrfs_root *log_root,
 802                                   struct inode *dir, struct inode *inode,
 803                                   struct extent_buffer *eb,
 804                                   u64 inode_objectid, u64 parent_objectid,
 805                                   u64 ref_index, char *name, int namelen,
 806                                   int *search_done)
 807 {
 808         int ret;
 809         char *victim_name;
 810         int victim_name_len;
 811         struct extent_buffer *leaf;
 812         struct btrfs_dir_item *di;
 813         struct btrfs_key search_key;
 814         struct btrfs_inode_extref *extref;
 815
 816 again:
 817         /* Search old style refs */
 818         search_key.objectid = inode_objectid;
 819         search_key.type = BTRFS_INODE_REF_KEY;
 820         search_key.offset = parent_objectid;
 821         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 822         if (ret == 0) {
 823                 struct btrfs_inode_ref *victim_ref;
 824                 unsigned long ptr;
 825                 unsigned long ptr_end;
 826
 827                 leaf = path->nodes[0];
 828
 829                 /* are we trying to overwrite a back ref for the root directory
 830                  * if so, just jump out, we're done
 831                  */
 832                 if (search_key.objectid == search_key.offset)
 833                         return 1;
 834
 835                 /* check all the names in this back reference to see
 836                  * if they are in the log.  if so, we allow them to stay
 837                  * otherwise they must be unlinked as a conflict
 838                  */
 839                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 840                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
 841                 while (ptr < ptr_end) {
 842                         victim_ref = (struct btrfs_inode_ref *)ptr;
 843                         victim_name_len = btrfs_inode_ref_name_len(leaf,
 844                                                                    victim_ref);
 845                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 846                         BUG_ON(!victim_name);
 847
 848                         read_extent_buffer(leaf, victim_name,
 849                                            (unsigned long)(victim_ref + 1),
 850                                            victim_name_len);
 851
 852                         if (!backref_in_log(log_root, &search_key,
 853                                             parent_objectid,
 854                                             victim_name,
 855                                             victim_name_len)) {
 856                                 btrfs_inc_nlink(inode);
 857                                 btrfs_release_path(path);
 858
 859                                 ret = btrfs_unlink_inode(trans, root, dir,
 860                                                          inode, victim_name,
 861                                                          victim_name_len);
 862                                 BUG_ON(ret);
 863                                 btrfs_run_delayed_items(trans, root);
 864                                 kfree(victim_name);
 865                                 *search_done = 1;
 866                                 goto again;
 867                         }
 868                         kfree(victim_name);
 869
 870                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
 871                 }
 872                 BUG_ON(ret);
 873
 874                 /*
 875                  * NOTE: we have searched root tree and checked the
 876                  * coresponding ref, it does not need to check again.
 877                  */
 878                 *search_done = 1;
 879         }
 880         btrfs_release_path(path);
 881
 882         /* Same search but for extended refs */
 883         extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
 884                                            inode_objectid, parent_objectid, 0,
 885                                            0);
 886         if (!IS_ERR_OR_NULL(extref)) {
 887                 u32 item_size;
 888                 u32 cur_offset = 0;
 889                 unsigned long base;
 890                 struct inode *victim_parent;
 891
 892                 leaf = path->nodes[0];
 893
 894                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 895                 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
 896
 897                 while (cur_offset < item_size) {
 898                         extref = (struct btrfs_inode_extref *)base + cur_offset;
 899
 900                         victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
 901
 902                         if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
 903                                 goto next;
 904
 905                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 906                         read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
 907                                            victim_name_len);
 908
 909                         search_key.objectid = inode_objectid;
 910                         search_key.type = BTRFS_INODE_EXTREF_KEY;
 911                         search_key.offset = btrfs_extref_hash(parent_objectid,
 912                                                               victim_name,
 913                                                               victim_name_len);
 914                         ret = 0;
 915                         if (!backref_in_log(log_root, &search_key,
 916                                             parent_objectid, victim_name,
 917                                             victim_name_len)) {
 918                                 ret = -ENOENT;
 919                                 victim_parent = read_one_inode(root,
 920                                                                parent_objectid);
 921                                 if (victim_parent) {
 922                                         btrfs_inc_nlink(inode);
 923                                         btrfs_release_path(path);
 924
 925                                         ret = btrfs_unlink_inode(trans, root,
 926                                                                  victim_parent,
 927                                                                  inode,
 928                                                                  victim_name,
 929                                                                  victim_name_len);
 930                                         btrfs_run_delayed_items(trans, root);
 931                                 }
 932                                 BUG_ON(ret);
 933                                 iput(victim_parent);
 934                                 kfree(victim_name);
 935                                 *search_done = 1;
 936                                 goto again;
 937                         }
 938                         kfree(victim_name);
 939                         BUG_ON(ret);
 940 next:
 941                         cur_offset += victim_name_len + sizeof(*extref);
 942                 }
 943                 *search_done = 1;
 944         }
 945         btrfs_release_path(path);
 946
 947         /* look for a conflicting sequence number */
 948         di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
 949                                          ref_index, name, namelen, 0);
 950         if (di && !IS_ERR(di)) {
 951                 ret = drop_one_dir_item(trans, root, path, dir, di);
 952                 BUG_ON(ret);
 953         }
 954         btrfs_release_path(path);
 955
 956         /* look for a conflicing name */
 957         di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
 958                                    name, namelen, 0);
 959         if (di && !IS_ERR(di)) {
 960                 ret = drop_one_dir_item(trans, root, path, dir, di);
 961                 BUG_ON(ret);
 962         }
 963         btrfs_release_path(path);
 964
 965         return 0;
 966 }
 967
 968 static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
 969                              u32 *namelen, char **name, u64 *index,
 970                              u64 *parent_objectid)
 971 {
 972         struct btrfs_inode_extref *extref;
 973
 974         extref = (struct btrfs_inode_extref *)ref_ptr;
 975
 976         *namelen = btrfs_inode_extref_name_len(eb, extref);
 977         *name = kmalloc(*namelen, GFP_NOFS);
 978         if (*name == NULL)
 979                 return -ENOMEM;
 980
 981         read_extent_buffer(eb, *name, (unsigned long)&extref->name,
 982                            *namelen);
 983
 984         *index = btrfs_inode_extref_index(eb, extref);
 985         if (parent_objectid)
 986                 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
 987
 988         return 0;
 989 }
 990
 991 static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
 992                           u32 *namelen, char **name, u64 *index)
 993 {
 994         struct btrfs_inode_ref *ref;
 995
 996         ref = (struct btrfs_inode_ref *)ref_ptr;
 997
 998         *namelen = btrfs_inode_ref_name_len(eb, ref);
 999         *name = kmalloc(*namelen, GFP_NOFS);
1000         if (*name == NULL)
1001                 return -ENOMEM;
1002
1003         read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1004
1005         *index = btrfs_inode_ref_index(eb, ref);
1006
1007         return 0;
1008 }
1009
1010 /*
1011  * replay one inode back reference item found in the log tree.
1012  * eb, slot and key refer to the buffer and key found in the log tree.
1013  * root is the destination we are replaying into, and path is for temp
1014  * use by this function.  (it should be released on return).
1015  */
1016 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1017                                   struct btrfs_root *root,
1018                                   struct btrfs_root *log,
1019                                   struct btrfs_path *path,
1020                                   struct extent_buffer *eb, int slot,
1021                                   struct btrfs_key *key)
1022 {
1023         struct inode *dir;
1024         struct inode *inode;
1025         unsigned long ref_ptr;
1026         unsigned long ref_end;
1027         char *name;
1028         int namelen;
1029         int ret;
1030         int search_done = 0;
1031         int log_ref_ver = 0;
1032         u64 parent_objectid;
1033         u64 inode_objectid;
1034         u64 ref_index = 0;
1035         int ref_struct_size;
1036
1037         ref_ptr = btrfs_item_ptr_offset(eb, slot);
1038         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1039
1040         if (key->type == BTRFS_INODE_EXTREF_KEY) {
1041                 struct btrfs_inode_extref *r;
1042
1043                 ref_struct_size = sizeof(struct btrfs_inode_extref);
1044                 log_ref_ver = 1;
1045                 r = (struct btrfs_inode_extref *)ref_ptr;
1046                 parent_objectid = btrfs_inode_extref_parent(eb, r);
1047         } else {
1048                 ref_struct_size = sizeof(struct btrfs_inode_ref);
1049                 parent_objectid = key->offset;
1050         }
1051         inode_objectid = key->objectid;
1052
1053         /*
1054          * it is possible that we didn't log all the parent directories
1055          * for a given inode.  If we don't find the dir, just don't
1056          * copy the back ref in.  The link count fixup code will take
1057          * care of the rest
1058          */
1059         dir = read_one_inode(root, parent_objectid);
1060         if (!dir)
1061                 return -ENOENT;
1062
1063         inode = read_one_inode(root, inode_objectid);
1064         if (!inode) {
1065                 iput(dir);
1066                 return -EIO;
1067         }
1068
1069         while (ref_ptr < ref_end) {
1070                 if (log_ref_ver) {
1071                         ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1072                                                 &ref_index, &parent_objectid);
1073                         /*
1074                          * parent object can change from one array
1075                          * item to another.
1076                          */
1077                         if (!dir)
1078                                 dir = read_one_inode(root, parent_objectid);
1079                         if (!dir)
1080                                 return -ENOENT;
1081                 } else {
1082                         ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1083                                              &ref_index);
1084                 }
1085                 if (ret)
1086                         return ret;
1087
1088                 /* if we already have a perfect match, we're done */
1089                 if (!inode_in_dir(root, path, btrfs_ino(dir), btrfs_ino(inode),
1090                                   ref_index, name, namelen)) {
1091                         /*
1092                          * look for a conflicting back reference in the
1093                          * metadata. if we find one we have to unlink that name
1094                          * of the file before we add our new link.  Later on, we
1095                          * overwrite any existing back reference, and we don't
1096                          * want to create dangling pointers in the directory.
1097                          */
1098
1099                         if (!search_done) {
1100                                 ret = __add_inode_ref(trans, root, path, log,
1101                                                       dir, inode, eb,
1102                                                       inode_objectid,
1103                                                       parent_objectid,
1104                                                       ref_index, name, namelen,
1105                                                       &search_done);
1106                                 if (ret == 1)
1107                                         goto out;
1108                                 BUG_ON(ret);
1109                         }
1110
1111                         /* insert our name */
1112                         ret = btrfs_add_link(trans, dir, inode, name, namelen,
1113                                              0, ref_index);
1114                         BUG_ON(ret);
1115
1116                         btrfs_update_inode(trans, root, inode);
1117                 }
1118
1119                 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1120                 kfree(name);
1121                 if (log_ref_ver) {
1122                         iput(dir);
1123                         dir = NULL;
1124                 }
1125         }
1126
1127         /* finally write the back reference in the inode */
1128         ret = overwrite_item(trans, root, path, eb, slot, key);
1129         BUG_ON(ret);
1130
1131 out:
1132         btrfs_release_path(path);
1133         iput(dir);
1134         iput(inode);
1135         return 0;
1136 }
1137
1138 static int insert_orphan_item(struct btrfs_trans_handle *trans,
1139                               struct btrfs_root *root, u64 offset)
1140 {
1141         int ret;
1142         ret = btrfs_find_orphan_item(root, offset);
1143         if (ret > 0)
1144                 ret = btrfs_insert_orphan_item(trans, root, offset);
1145         return ret;
1146 }
1147
1148 static int count_inode_extrefs(struct btrfs_root *root,
1149                                struct inode *inode, struct btrfs_path *path)
1150 {
1151         int ret = 0;
1152         int name_len;
1153         unsigned int nlink = 0;
1154         u32 item_size;
1155         u32 cur_offset = 0;
1156         u64 inode_objectid = btrfs_ino(inode);
1157         u64 offset = 0;
1158         unsigned long ptr;
1159         struct btrfs_inode_extref *extref;
1160         struct extent_buffer *leaf;
1161
1162         while (1) {
1163                 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1164                                             &extref, &offset);
1165                 if (ret)
1166                         break;
1167
1168                 leaf = path->nodes[0];
1169                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1170                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1171
1172                 while (cur_offset < item_size) {
1173                         extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1174                         name_len = btrfs_inode_extref_name_len(leaf, extref);
1175
1176                         nlink++;
1177
1178                         cur_offset += name_len + sizeof(*extref);
1179                 }
1180
1181                 offset++;
1182                 btrfs_release_path(path);
1183         }
1184         btrfs_release_path(path);
1185
1186         if (ret < 0)
1187                 return ret;
1188         return nlink;
1189 }
1190
1191 static int count_inode_refs(struct btrfs_root *root,
1192                                struct inode *inode, struct btrfs_path *path)
1193 {
1194         int ret;
1195         struct btrfs_key key;
1196         unsigned int nlink = 0;
1197         unsigned long ptr;
1198         unsigned long ptr_end;
1199         int name_len;
1200         u64 ino = btrfs_ino(inode);
1201
1202         key.objectid = ino;
1203         key.type = BTRFS_INODE_REF_KEY;
1204         key.offset = (u64)-1;
1205
1206         while (1) {
1207                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1208                 if (ret < 0)
1209                         break;
1210                 if (ret > 0) {
1211                         if (path->slots[0] == 0)
1212                                 break;
1213                         path->slots[0]--;
1214                 }
1215                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1216                                       path->slots[0]);
1217                 if (key.objectid != ino ||
1218                     key.type != BTRFS_INODE_REF_KEY)
1219                         break;
1220                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1221                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1222                                                    path->slots[0]);
1223                 while (ptr < ptr_end) {
1224                         struct btrfs_inode_ref *ref;
1225
1226                         ref = (struct btrfs_inode_ref *)ptr;
1227                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1228                                                             ref);
1229                         ptr = (unsigned long)(ref + 1) + name_len;
1230                         nlink++;
1231                 }
1232
1233                 if (key.offset == 0)
1234                         break;
1235                 key.offset--;
1236                 btrfs_release_path(path);
1237         }
1238         btrfs_release_path(path);
1239
1240         return nlink;
1241 }
1242
1243 /*
1244  * There are a few corners where the link count of the file can't
1245  * be properly maintained during replay.  So, instead of adding
1246  * lots of complexity to the log code, we just scan the backrefs
1247  * for any file that has been through replay.
1248  *
1249  * The scan will update the link count on the inode to reflect the
1250  * number of back refs found.  If it goes down to zero, the iput
1251  * will free the inode.
1252  */
1253 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1254                                            struct btrfs_root *root,
1255                                            struct inode *inode)
1256 {
1257         struct btrfs_path *path;
1258         int ret;
1259         u64 nlink = 0;
1260         u64 ino = btrfs_ino(inode);
1261
1262         path = btrfs_alloc_path();
1263         if (!path)
1264                 return -ENOMEM;
1265
1266         ret = count_inode_refs(root, inode, path);
1267         if (ret < 0)
1268                 goto out;
1269
1270         nlink = ret;
1271
1272         ret = count_inode_extrefs(root, inode, path);
1273         if (ret == -ENOENT)
1274                 ret = 0;
1275
1276         if (ret < 0)
1277                 goto out;
1278
1279         nlink += ret;
1280
1281         ret = 0;
1282
1283         if (nlink != inode->i_nlink) {
1284                 set_nlink(inode, nlink);
1285                 btrfs_update_inode(trans, root, inode);
1286         }
1287         BTRFS_I(inode)->index_cnt = (u64)-1;
1288
1289         if (inode->i_nlink == 0) {
1290                 if (S_ISDIR(inode->i_mode)) {
1291                         ret = replay_dir_deletes(trans, root, NULL, path,
1292                                                  ino, 1);
1293                         BUG_ON(ret);
1294                 }
1295                 ret = insert_orphan_item(trans, root, ino);
1296                 BUG_ON(ret);
1297         }
1298
1299 out:
1300         btrfs_free_path(path);
1301         return ret;
1302 }
1303
1304 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1305                                             struct btrfs_root *root,
1306                                             struct btrfs_path *path)
1307 {
1308         int ret;
1309         struct btrfs_key key;
1310         struct inode *inode;
1311
1312         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1313         key.type = BTRFS_ORPHAN_ITEM_KEY;
1314         key.offset = (u64)-1;
1315         while (1) {
1316                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1317                 if (ret < 0)
1318                         break;
1319
1320                 if (ret == 1) {
1321                         if (path->slots[0] == 0)
1322                                 break;
1323                         path->slots[0]--;
1324                 }
1325
1326                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1327                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1328                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1329                         break;
1330
1331                 ret = btrfs_del_item(trans, root, path);
1332                 if (ret)
1333                         goto out;
1334
1335                 btrfs_release_path(path);
1336                 inode = read_one_inode(root, key.offset);
1337                 if (!inode)
1338                         return -EIO;
1339
1340                 ret = fixup_inode_link_count(trans, root, inode);
1341                 BUG_ON(ret);
1342
1343                 iput(inode);
1344
1345                 /*
1346                  * fixup on a directory may create new entries,
1347                  * make sure we always look for the highset possible
1348                  * offset
1349                  */
1350                 key.offset = (u64)-1;
1351         }
1352         ret = 0;
1353 out:
1354         btrfs_release_path(path);
1355         return ret;
1356 }
1357
1358
1359 /*
1360  * record a given inode in the fixup dir so we can check its link
1361  * count when replay is done.  The link count is incremented here
1362  * so the inode won't go away until we check it
1363  */
1364 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1365                                       struct btrfs_root *root,
1366                                       struct btrfs_path *path,
1367                                       u64 objectid)
1368 {
1369         struct btrfs_key key;
1370         int ret = 0;
1371         struct inode *inode;
1372
1373         inode = read_one_inode(root, objectid);
1374         if (!inode)
1375                 return -EIO;
1376
1377         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1378         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1379         key.offset = objectid;
1380
1381         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1382
1383         btrfs_release_path(path);
1384         if (ret == 0) {
1385                 btrfs_inc_nlink(inode);
1386                 ret = btrfs_update_inode(trans, root, inode);
1387         } else if (ret == -EEXIST) {
1388                 ret = 0;
1389         } else {
1390                 BUG();
1391         }
1392         iput(inode);
1393
1394         return ret;
1395 }
1396
1397 /*
1398  * when replaying the log for a directory, we only insert names
1399  * for inodes that actually exist.  This means an fsync on a directory
1400  * does not implicitly fsync all the new files in it
1401  */
1402 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1403                                     struct btrfs_root *root,
1404                                     struct btrfs_path *path,
1405                                     u64 dirid, u64 index,
1406                                     char *name, int name_len, u8 type,
1407                                     struct btrfs_key *location)
1408 {
1409         struct inode *inode;
1410         struct inode *dir;
1411         int ret;
1412
1413         inode = read_one_inode(root, location->objectid);
1414         if (!inode)
1415                 return -ENOENT;
1416
1417         dir = read_one_inode(root, dirid);
1418         if (!dir) {
1419                 iput(inode);
1420                 return -EIO;
1421         }
1422         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1423
1424         /* FIXME, put inode into FIXUP list */
1425
1426         iput(inode);
1427         iput(dir);
1428         return ret;
1429 }
1430
1431 /*
1432  * take a single entry in a log directory item and replay it into
1433  * the subvolume.
1434  *
1435  * if a conflicting item exists in the subdirectory already,
1436  * the inode it points to is unlinked and put into the link count
1437  * fix up tree.
1438  *
1439  * If a name from the log points to a file or directory that does
1440  * not exist in the FS, it is skipped.  fsyncs on directories
1441  * do not force down inodes inside that directory, just changes to the
1442  * names or unlinks in a directory.
1443  */
1444 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1445                                     struct btrfs_root *root,
1446                                     struct btrfs_path *path,
1447                                     struct extent_buffer *eb,
1448                                     struct btrfs_dir_item *di,
1449                                     struct btrfs_key *key)
1450 {
1451         char *name;
1452         int name_len;
1453         struct btrfs_dir_item *dst_di;
1454         struct btrfs_key found_key;
1455         struct btrfs_key log_key;
1456         struct inode *dir;
1457         u8 log_type;
1458         int exists;
1459         int ret;
1460
1461         dir = read_one_inode(root, key->objectid);
1462         if (!dir)
1463                 return -EIO;
1464
1465         name_len = btrfs_dir_name_len(eb, di);
1466         name = kmalloc(name_len, GFP_NOFS);
1467         if (!name)
1468                 return -ENOMEM;
1469
1470         log_type = btrfs_dir_type(eb, di);
1471         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1472                    name_len);
1473
1474         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1475         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1476         if (exists == 0)
1477                 exists = 1;
1478         else
1479                 exists = 0;
1480         btrfs_release_path(path);
1481
1482         if (key->type == BTRFS_DIR_ITEM_KEY) {
1483                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1484                                        name, name_len, 1);
1485         } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1486                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1487                                                      key->objectid,
1488                                                      key->offset, name,
1489                                                      name_len, 1);
1490         } else {
1491                 BUG();
1492         }
1493         if (IS_ERR_OR_NULL(dst_di)) {
1494                 /* we need a sequence number to insert, so we only
1495                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1496                  */
1497                 if (key->type != BTRFS_DIR_INDEX_KEY)
1498                         goto out;
1499                 goto insert;
1500         }
1501
1502         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1503         /* the existing item matches the logged item */
1504         if (found_key.objectid == log_key.objectid &&
1505             found_key.type == log_key.type &&
1506             found_key.offset == log_key.offset &&
1507             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1508                 goto out;
1509         }
1510
1511         /*
1512          * don't drop the conflicting directory entry if the inode
1513          * for the new entry doesn't exist
1514          */
1515         if (!exists)
1516                 goto out;
1517
1518         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1519         BUG_ON(ret);
1520
1521         if (key->type == BTRFS_DIR_INDEX_KEY)
1522                 goto insert;
1523 out:
1524         btrfs_release_path(path);
1525         kfree(name);
1526         iput(dir);
1527         return 0;
1528
1529 insert:
1530         btrfs_release_path(path);
1531         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1532                               name, name_len, log_type, &log_key);
1533
1534         BUG_ON(ret && ret != -ENOENT);
1535         goto out;
1536 }
1537
1538 /*
1539  * find all the names in a directory item and reconcile them into
1540  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1541  * one name in a directory item, but the same code gets used for
1542  * both directory index types
1543  */
1544 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1545                                         struct btrfs_root *root,
1546                                         struct btrfs_path *path,
1547                                         struct extent_buffer *eb, int slot,
1548                                         struct btrfs_key *key)
1549 {
1550         int ret;
1551         u32 item_size = btrfs_item_size_nr(eb, slot);
1552         struct btrfs_dir_item *di;
1553         int name_len;
1554         unsigned long ptr;
1555         unsigned long ptr_end;
1556
1557         ptr = btrfs_item_ptr_offset(eb, slot);
1558         ptr_end = ptr + item_size;
1559         while (ptr < ptr_end) {
1560                 di = (struct btrfs_dir_item *)ptr;
1561                 if (verify_dir_item(root, eb, di))
1562                         return -EIO;
1563                 name_len = btrfs_dir_name_len(eb, di);
1564                 ret = replay_one_name(trans, root, path, eb, di, key);
1565                 BUG_ON(ret);
1566                 ptr = (unsigned long)(di + 1);
1567                 ptr += name_len;
1568         }
1569         return 0;
1570 }
1571
1572 /*
1573  * directory replay has two parts.  There are the standard directory
1574  * items in the log copied from the subvolume, and range items
1575  * created in the log while the subvolume was logged.
1576  *
1577  * The range items tell us which parts of the key space the log
1578  * is authoritative for.  During replay, if a key in the subvolume
1579  * directory is in a logged range item, but not actually in the log
1580  * that means it was deleted from the directory before the fsync
1581  * and should be removed.
1582  */
1583 static noinline int find_dir_range(struct btrfs_root *root,
1584                                    struct btrfs_path *path,
1585                                    u64 dirid, int key_type,
1586                                    u64 *start_ret, u64 *end_ret)
1587 {
1588         struct btrfs_key key;
1589         u64 found_end;
1590         struct btrfs_dir_log_item *item;
1591         int ret;
1592         int nritems;
1593
1594         if (*start_ret == (u64)-1)
1595                 return 1;
1596
1597         key.objectid = dirid;
1598         key.type = key_type;
1599         key.offset = *start_ret;
1600
1601         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1602         if (ret < 0)
1603                 goto out;
1604         if (ret > 0) {
1605                 if (path->slots[0] == 0)
1606                         goto out;
1607                 path->slots[0]--;
1608         }
1609         if (ret != 0)
1610                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1611
1612         if (key.type != key_type || key.objectid != dirid) {
1613                 ret = 1;
1614                 goto next;
1615         }
1616         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1617                               struct btrfs_dir_log_item);
1618         found_end = btrfs_dir_log_end(path->nodes[0], item);
1619
1620         if (*start_ret >= key.offset && *start_ret <= found_end) {
1621                 ret = 0;
1622                 *start_ret = key.offset;
1623                 *end_ret = found_end;
1624                 goto out;
1625         }
1626         ret = 1;
1627 next:
1628         /* check the next slot in the tree to see if it is a valid item */
1629         nritems = btrfs_header_nritems(path->nodes[0]);
1630         if (path->slots[0] >= nritems) {
1631                 ret = btrfs_next_leaf(root, path);
1632                 if (ret)
1633                         goto out;
1634         } else {
1635                 path->slots[0]++;
1636         }
1637
1638         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1639
1640         if (key.type != key_type || key.objectid != dirid) {
1641                 ret = 1;
1642                 goto out;
1643         }
1644         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1645                               struct btrfs_dir_log_item);
1646         found_end = btrfs_dir_log_end(path->nodes[0], item);
1647         *start_ret = key.offset;
1648         *end_ret = found_end;
1649         ret = 0;
1650 out:
1651         btrfs_release_path(path);
1652         return ret;
1653 }
1654
1655 /*
1656  * this looks for a given directory item in the log.  If the directory
1657  * item is not in the log, the item is removed and the inode it points
1658  * to is unlinked
1659  */
1660 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1661                                       struct btrfs_root *root,
1662                                       struct btrfs_root *log,
1663                                       struct btrfs_path *path,
1664                                       struct btrfs_path *log_path,
1665                                       struct inode *dir,
1666                                       struct btrfs_key *dir_key)
1667 {
1668         int ret;
1669         struct extent_buffer *eb;
1670         int slot;
1671         u32 item_size;
1672         struct btrfs_dir_item *di;
1673         struct btrfs_dir_item *log_di;
1674         int name_len;
1675         unsigned long ptr;
1676         unsigned long ptr_end;
1677         char *name;
1678         struct inode *inode;
1679         struct btrfs_key location;
1680
1681 again:
1682         eb = path->nodes[0];
1683         slot = path->slots[0];
1684         item_size = btrfs_item_size_nr(eb, slot);
1685         ptr = btrfs_item_ptr_offset(eb, slot);
1686         ptr_end = ptr + item_size;
1687         while (ptr < ptr_end) {
1688                 di = (struct btrfs_dir_item *)ptr;
1689                 if (verify_dir_item(root, eb, di)) {
1690                         ret = -EIO;
1691                         goto out;
1692                 }
1693
1694                 name_len = btrfs_dir_name_len(eb, di);
1695                 name = kmalloc(name_len, GFP_NOFS);
1696                 if (!name) {
1697                         ret = -ENOMEM;
1698                         goto out;
1699                 }
1700                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1701                                   name_len);
1702                 log_di = NULL;
1703                 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
1704                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1705                                                        dir_key->objectid,
1706                                                        name, name_len, 0);
1707                 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
1708                         log_di = btrfs_lookup_dir_index_item(trans, log,
1709                                                      log_path,
1710                                                      dir_key->objectid,
1711                                                      dir_key->offset,
1712                                                      name, name_len, 0);
1713                 }
1714                 if (IS_ERR_OR_NULL(log_di)) {
1715                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1716                         btrfs_release_path(path);
1717                         btrfs_release_path(log_path);
1718                         inode = read_one_inode(root, location.objectid);
1719                         if (!inode) {
1720                                 kfree(name);
1721                                 return -EIO;
1722                         }
1723
1724                         ret = link_to_fixup_dir(trans, root,
1725                                                 path, location.objectid);
1726                         BUG_ON(ret);
1727                         btrfs_inc_nlink(inode);
1728                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1729                                                  name, name_len);
1730                         BUG_ON(ret);
1731
1732                         btrfs_run_delayed_items(trans, root);
1733
1734                         kfree(name);
1735                         iput(inode);
1736
1737                         /* there might still be more names under this key
1738                          * check and repeat if required
1739                          */
1740                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1741                                                 0, 0);
1742                         if (ret == 0)
1743                                 goto again;
1744                         ret = 0;
1745                         goto out;
1746                 }
1747                 btrfs_release_path(log_path);
1748                 kfree(name);
1749
1750                 ptr = (unsigned long)(di + 1);
1751                 ptr += name_len;
1752         }
1753         ret = 0;
1754 out:
1755         btrfs_release_path(path);
1756         btrfs_release_path(log_path);
1757         return ret;
1758 }
1759
1760 /*
1761  * deletion replay happens before we copy any new directory items
1762  * out of the log or out of backreferences from inodes.  It
1763  * scans the log to find ranges of keys that log is authoritative for,
1764  * and then scans the directory to find items in those ranges that are
1765  * not present in the log.
1766  *
1767  * Anything we don't find in the log is unlinked and removed from the
1768  * directory.
1769  */
1770 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1771                                        struct btrfs_root *root,
1772                                        struct btrfs_root *log,
1773                                        struct btrfs_path *path,
1774                                        u64 dirid, int del_all)
1775 {
1776         u64 range_start;
1777         u64 range_end;
1778         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1779         int ret = 0;
1780         struct btrfs_key dir_key;
1781         struct btrfs_key found_key;
1782         struct btrfs_path *log_path;
1783         struct inode *dir;
1784
1785         dir_key.objectid = dirid;
1786         dir_key.type = BTRFS_DIR_ITEM_KEY;
1787         log_path = btrfs_alloc_path();
1788         if (!log_path)
1789                 return -ENOMEM;
1790
1791         dir = read_one_inode(root, dirid);
1792         /* it isn't an error if the inode isn't there, that can happen
1793          * because we replay the deletes before we copy in the inode item
1794          * from the log
1795          */
1796         if (!dir) {
1797                 btrfs_free_path(log_path);
1798                 return 0;
1799         }
1800 again:
1801         range_start = 0;
1802         range_end = 0;
1803         while (1) {
1804                 if (del_all)
1805                         range_end = (u64)-1;
1806                 else {
1807                         ret = find_dir_range(log, path, dirid, key_type,
1808                                              &range_start, &range_end);
1809                         if (ret != 0)
1810                                 break;
1811                 }
1812
1813                 dir_key.offset = range_start;
1814                 while (1) {
1815                         int nritems;
1816                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
1817                                                 0, 0);
1818                         if (ret < 0)
1819                                 goto out;
1820
1821                         nritems = btrfs_header_nritems(path->nodes[0]);
1822                         if (path->slots[0] >= nritems) {
1823                                 ret = btrfs_next_leaf(root, path);
1824                                 if (ret)
1825                                         break;
1826                         }
1827                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1828                                               path->slots[0]);
1829                         if (found_key.objectid != dirid ||
1830                             found_key.type != dir_key.type)
1831                                 goto next_type;
1832
1833                         if (found_key.offset > range_end)
1834                                 break;
1835
1836                         ret = check_item_in_log(trans, root, log, path,
1837                                                 log_path, dir,
1838                                                 &found_key);
1839                         BUG_ON(ret);
1840                         if (found_key.offset == (u64)-1)
1841                                 break;
1842                         dir_key.offset = found_key.offset + 1;
1843                 }
1844                 btrfs_release_path(path);
1845                 if (range_end == (u64)-1)
1846                         break;
1847                 range_start = range_end + 1;
1848         }
1849
1850 next_type:
1851         ret = 0;
1852         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1853                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1854                 dir_key.type = BTRFS_DIR_INDEX_KEY;
1855                 btrfs_release_path(path);
1856                 goto again;
1857         }
1858 out:
1859         btrfs_release_path(path);
1860         btrfs_free_path(log_path);
1861         iput(dir);
1862         return ret;
1863 }
1864
1865 /*
1866  * the process_func used to replay items from the log tree.  This
1867  * gets called in two different stages.  The first stage just looks
1868  * for inodes and makes sure they are all copied into the subvolume.
1869  *
1870  * The second stage copies all the other item types from the log into
1871  * the subvolume.  The two stage approach is slower, but gets rid of
1872  * lots of complexity around inodes referencing other inodes that exist
1873  * only in the log (references come from either directory items or inode
1874  * back refs).
1875  */
1876 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
1877                              struct walk_control *wc, u64 gen)
1878 {
1879         int nritems;
1880         struct btrfs_path *path;
1881         struct btrfs_root *root = wc->replay_dest;
1882         struct btrfs_key key;
1883         int level;
1884         int i;
1885         int ret;
1886
1887         ret = btrfs_read_buffer(eb, gen);
1888         if (ret)
1889                 return ret;
1890
1891         level = btrfs_header_level(eb);
1892
1893         if (level != 0)
1894                 return 0;
1895
1896         path = btrfs_alloc_path();
1897         if (!path)
1898                 return -ENOMEM;
1899
1900         nritems = btrfs_header_nritems(eb);
1901         for (i = 0; i < nritems; i++) {
1902                 btrfs_item_key_to_cpu(eb, &key, i);
1903
1904                 /* inode keys are done during the first stage */
1905                 if (key.type == BTRFS_INODE_ITEM_KEY &&
1906                     wc->stage == LOG_WALK_REPLAY_INODES) {
1907                         struct btrfs_inode_item *inode_item;
1908                         u32 mode;
1909
1910                         inode_item = btrfs_item_ptr(eb, i,
1911                                             struct btrfs_inode_item);
1912                         mode = btrfs_inode_mode(eb, inode_item);
1913                         if (S_ISDIR(mode)) {
1914                                 ret = replay_dir_deletes(wc->trans,
1915                                          root, log, path, key.objectid, 0);
1916                                 BUG_ON(ret);
1917                         }
1918                         ret = overwrite_item(wc->trans, root, path,
1919                                              eb, i, &key);
1920                         BUG_ON(ret);
1921
1922                         /* for regular files, make sure corresponding
1923                          * orhpan item exist. extents past the new EOF
1924                          * will be truncated later by orphan cleanup.
1925                          */
1926                         if (S_ISREG(mode)) {
1927                                 ret = insert_orphan_item(wc->trans, root,
1928                                                          key.objectid);
1929                                 BUG_ON(ret);
1930                         }
1931
1932                         ret = link_to_fixup_dir(wc->trans, root,
1933                                                 path, key.objectid);
1934                         BUG_ON(ret);
1935                 }
1936                 if (wc->stage < LOG_WALK_REPLAY_ALL)
1937                         continue;
1938
1939                 /* these keys are simply copied */
1940                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
1941                         ret = overwrite_item(wc->trans, root, path,
1942                                              eb, i, &key);
1943                         BUG_ON(ret);
1944                 } else if (key.type == BTRFS_INODE_REF_KEY) {
1945                         ret = add_inode_ref(wc->trans, root, log, path,
1946                                             eb, i, &key);
1947                         BUG_ON(ret && ret != -ENOENT);
1948                 } else if (key.type == BTRFS_INODE_EXTREF_KEY) {
1949                         ret = add_inode_ref(wc->trans, root, log, path,
1950                                             eb, i, &key);
1951                         BUG_ON(ret && ret != -ENOENT);
1952                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
1953                         ret = replay_one_extent(wc->trans, root, path,
1954                                                 eb, i, &key);
1955                         BUG_ON(ret);
1956                 } else if (key.type == BTRFS_DIR_ITEM_KEY ||
1957                            key.type == BTRFS_DIR_INDEX_KEY) {
1958                         ret = replay_one_dir_item(wc->trans, root, path,
1959                                                   eb, i, &key);
1960                         BUG_ON(ret);
1961                 }
1962         }
1963         btrfs_free_path(path);
1964         return 0;
1965 }
1966
1967 static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
1968                                    struct btrfs_root *root,
1969                                    struct btrfs_path *path, int *level,
1970                                    struct walk_control *wc)
1971 {
1972         u64 root_owner;
1973         u64 bytenr;
1974         u64 ptr_gen;
1975         struct extent_buffer *next;
1976         struct extent_buffer *cur;
1977         struct extent_buffer *parent;
1978         u32 blocksize;
1979         int ret = 0;
1980
1981         WARN_ON(*level < 0);
1982         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1983
1984         while (*level > 0) {
1985                 WARN_ON(*level < 0);
1986                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1987                 cur = path->nodes[*level];
1988
1989                 if (btrfs_header_level(cur) != *level)
1990                         WARN_ON(1);
1991
1992                 if (path->slots[*level] >=
1993                     btrfs_header_nritems(cur))
1994                         break;
1995
1996                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1997                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1998                 blocksize = btrfs_level_size(root, *level - 1);
1999
2000                 parent = path->nodes[*level];
2001                 root_owner = btrfs_header_owner(parent);
2002
2003                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
2004                 if (!next)
2005                         return -ENOMEM;
2006
2007                 if (*level == 1) {
2008                         ret = wc->process_func(root, next, wc, ptr_gen);
2009                         if (ret)
2010                                 return ret;
2011
2012                         path->slots[*level]++;
2013                         if (wc->free) {
2014                                 ret = btrfs_read_buffer(next, ptr_gen);
2015                                 if (ret) {
2016                                         free_extent_buffer(next);
2017                                         return ret;
2018                                 }
2019
2020                                 btrfs_tree_lock(next);
2021                                 btrfs_set_lock_blocking(next);
2022                                 clean_tree_block(trans, root, next);
2023                                 btrfs_wait_tree_block_writeback(next);
2024                                 btrfs_tree_unlock(next);
2025
2026                                 WARN_ON(root_owner !=
2027                                         BTRFS_TREE_LOG_OBJECTID);
2028                                 ret = btrfs_free_and_pin_reserved_extent(root,
2029                                                          bytenr, blocksize);
2030                                 BUG_ON(ret); /* -ENOMEM or logic errors */
2031                         }
2032                         free_extent_buffer(next);
2033                         continue;
2034                 }
2035                 ret = btrfs_read_buffer(next, ptr_gen);
2036                 if (ret) {
2037                         free_extent_buffer(next);
2038                         return ret;
2039                 }
2040
2041                 WARN_ON(*level <= 0);
2042                 if (path->nodes[*level-1])
2043                         free_extent_buffer(path->nodes[*level-1]);
2044                 path->nodes[*level-1] = next;
2045                 *level = btrfs_header_level(next);
2046                 path->slots[*level] = 0;
2047                 cond_resched();
2048         }
2049         WARN_ON(*level < 0);
2050         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2051
2052         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2053
2054         cond_resched();
2055         return 0;
2056 }
2057
2058 static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2059                                  struct btrfs_root *root,
2060                                  struct btrfs_path *path, int *level,
2061                                  struct walk_control *wc)
2062 {
2063         u64 root_owner;
2064         int i;
2065         int slot;
2066         int ret;
2067
2068         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2069                 slot = path->slots[i];
2070                 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2071                         path->slots[i]++;
2072                         *level = i;
2073                         WARN_ON(*level == 0);
2074                         return 0;
2075                 } else {
2076                         struct extent_buffer *parent;
2077                         if (path->nodes[*level] == root->node)
2078                                 parent = path->nodes[*level];
2079                         else
2080                                 parent = path->nodes[*level + 1];
2081
2082                         root_owner = btrfs_header_owner(parent);
2083                         ret = wc->process_func(root, path->nodes[*level], wc,
2084                                  btrfs_header_generation(path->nodes[*level]));
2085                         if (ret)
2086                                 return ret;
2087
2088                         if (wc->free) {
2089                                 struct extent_buffer *next;
2090
2091                                 next = path->nodes[*level];
2092
2093                                 btrfs_tree_lock(next);
2094                                 btrfs_set_lock_blocking(next);
2095                                 clean_tree_block(trans, root, next);
2096                                 btrfs_wait_tree_block_writeback(next);
2097                                 btrfs_tree_unlock(next);
2098
2099                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
2100                                 ret = btrfs_free_and_pin_reserved_extent(root,
2101                                                 path->nodes[*level]->start,
2102                                                 path->nodes[*level]->len);
2103                                 BUG_ON(ret);
2104                         }
2105                         free_extent_buffer(path->nodes[*level]);
2106                         path->nodes[*level] = NULL;
2107                         *level = i + 1;
2108                 }
2109         }
2110         return 1;
2111 }
2112
2113 /*
2114  * drop the reference count on the tree rooted at 'snap'.  This traverses
2115  * the tree freeing any blocks that have a ref count of zero after being
2116  * decremented.
2117  */
2118 static int walk_log_tree(struct btrfs_trans_handle *trans,
2119                          struct btrfs_root *log, struct walk_control *wc)
2120 {
2121         int ret = 0;
2122         int wret;
2123         int level;
2124         struct btrfs_path *path;
2125         int i;
2126         int orig_level;
2127
2128         path = btrfs_alloc_path();
2129         if (!path)
2130                 return -ENOMEM;
2131
2132         level = btrfs_header_level(log->node);
2133         orig_level = level;
2134         path->nodes[level] = log->node;
2135         extent_buffer_get(log->node);
2136         path->slots[level] = 0;
2137
2138         while (1) {
2139                 wret = walk_down_log_tree(trans, log, path, &level, wc);
2140                 if (wret > 0)
2141                         break;
2142                 if (wret < 0) {
2143                         ret = wret;
2144                         goto out;
2145                 }
2146
2147                 wret = walk_up_log_tree(trans, log, path, &level, wc);
2148                 if (wret > 0)
2149                         break;
2150                 if (wret < 0) {
2151                         ret = wret;
2152                         goto out;
2153                 }
2154         }
2155
2156         /* was the root node processed? if not, catch it here */
2157         if (path->nodes[orig_level]) {
2158                 ret = wc->process_func(log, path->nodes[orig_level], wc,
2159                          btrfs_header_generation(path->nodes[orig_level]));
2160                 if (ret)
2161                         goto out;
2162                 if (wc->free) {
2163                         struct extent_buffer *next;
2164
2165                         next = path->nodes[orig_level];
2166
2167                         btrfs_tree_lock(next);
2168                         btrfs_set_lock_blocking(next);
2169                         clean_tree_block(trans, log, next);
2170                         btrfs_wait_tree_block_writeback(next);
2171                         btrfs_tree_unlock(next);
2172
2173                         WARN_ON(log->root_key.objectid !=
2174                                 BTRFS_TREE_LOG_OBJECTID);
2175                         ret = btrfs_free_and_pin_reserved_extent(log, next->start,
2176                                                          next->len);
2177                         BUG_ON(ret); /* -ENOMEM or logic errors */
2178                 }
2179         }
2180
2181 out:
2182         for (i = 0; i <= orig_level; i++) {
2183                 if (path->nodes[i]) {
2184                         free_extent_buffer(path->nodes[i]);
2185                         path->nodes[i] = NULL;
2186                 }
2187         }
2188         btrfs_free_path(path);
2189         return ret;
2190 }
2191
2192 /*
2193  * helper function to update the item for a given subvolumes log root
2194  * in the tree of log roots
2195  */
2196 static int update_log_root(struct btrfs_trans_handle *trans,
2197                            struct btrfs_root *log)
2198 {
2199         int ret;
2200
2201         if (log->log_transid == 1) {
2202                 /* insert root item on the first sync */
2203                 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
2204                                 &log->root_key, &log->root_item);
2205         } else {
2206                 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2207                                 &log->root_key, &log->root_item);
2208         }
2209         return ret;
2210 }
2211
2212 static int wait_log_commit(struct btrfs_trans_handle *trans,
2213                            struct btrfs_root *root, unsigned long transid)
2214 {
2215         DEFINE_WAIT(wait);
2216         int index = transid % 2;
2217
2218         /*
2219          * we only allow two pending log transactions at a time,
2220          * so we know that if ours is more than 2 older than the
2221          * current transaction, we're done
2222          */
2223         do {
2224                 prepare_to_wait(&root->log_commit_wait[index],
2225                                 &wait, TASK_UNINTERRUPTIBLE);
2226                 mutex_unlock(&root->log_mutex);
2227
2228                 if (root->fs_info->last_trans_log_full_commit !=
2229                     trans->transid && root->log_transid < transid + 2 &&
2230                     atomic_read(&root->log_commit[index]))
2231                         schedule();
2232
2233                 finish_wait(&root->log_commit_wait[index], &wait);
2234                 mutex_lock(&root->log_mutex);
2235         } while (root->fs_info->last_trans_log_full_commit !=
2236                  trans->transid && root->log_transid < transid + 2 &&
2237                  atomic_read(&root->log_commit[index]));
2238         return 0;
2239 }
2240
2241 static void wait_for_writer(struct btrfs_trans_handle *trans,
2242                             struct btrfs_root *root)
2243 {
2244         DEFINE_WAIT(wait);
2245         while (root->fs_info->last_trans_log_full_commit !=
2246                trans->transid && atomic_read(&root->log_writers)) {
2247                 prepare_to_wait(&root->log_writer_wait,
2248                                 &wait, TASK_UNINTERRUPTIBLE);
2249                 mutex_unlock(&root->log_mutex);
2250                 if (root->fs_info->last_trans_log_full_commit !=
2251                     trans->transid && atomic_read(&root->log_writers))
2252                         schedule();
2253                 mutex_lock(&root->log_mutex);
2254                 finish_wait(&root->log_writer_wait, &wait);
2255         }
2256 }
2257
2258 /*
2259  * btrfs_sync_log does sends a given tree log down to the disk and
2260  * updates the super blocks to record it.  When this call is done,
2261  * you know that any inodes previously logged are safely on disk only
2262  * if it returns 0.
2263  *
2264  * Any other return value means you need to call btrfs_commit_transaction.
2265  * Some of the edge cases for fsyncing directories that have had unlinks
2266  * or renames done in the past mean that sometimes the only safe
2267  * fsync is to commit the whole FS.  When btrfs_sync_log returns -EAGAIN,
2268  * that has happened.
2269  */
2270 int btrfs_sync_log(struct btrfs_trans_handle *trans,
2271                    struct btrfs_root *root)
2272 {
2273         int index1;
2274         int index2;
2275         int mark;
2276         int ret;
2277         struct btrfs_root *log = root->log_root;
2278         struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
2279         unsigned long log_transid = 0;
2280
2281         mutex_lock(&root->log_mutex);
2282         log_transid = root->log_transid;
2283         index1 = root->log_transid % 2;
2284         if (atomic_read(&root->log_commit[index1])) {
2285                 wait_log_commit(trans, root, root->log_transid);
2286                 mutex_unlock(&root->log_mutex);
2287                 return 0;
2288         }
2289         atomic_set(&root->log_commit[index1], 1);
2290
2291         /* wait for previous tree log sync to complete */
2292         if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2293                 wait_log_commit(trans, root, root->log_transid - 1);
2294         while (1) {
2295                 int batch = atomic_read(&root->log_batch);
2296                 /* when we're on an ssd, just kick the log commit out */
2297                 if (!btrfs_test_opt(root, SSD) && root->log_multiple_pids) {
2298                         mutex_unlock(&root->log_mutex);
2299                         schedule_timeout_uninterruptible(1);
2300                         mutex_lock(&root->log_mutex);
2301                 }
2302                 wait_for_writer(trans, root);
2303                 if (batch == atomic_read(&root->log_batch))
2304                         break;
2305         }
2306
2307         /* bail out if we need to do a full commit */
2308         if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2309                 ret = -EAGAIN;
2310                 btrfs_free_logged_extents(log, log_transid);
2311                 mutex_unlock(&root->log_mutex);
2312                 goto out;
2313         }
2314
2315         if (log_transid % 2 == 0)
2316                 mark = EXTENT_DIRTY;
2317         else
2318                 mark = EXTENT_NEW;
2319
2320         /* we start IO on  all the marked extents here, but we don't actually
2321          * wait for them until later.
2322          */
2323         ret = btrfs_write_marked_extents(log, &log->dirty_log_pages, mark);
2324         if (ret) {
2325                 btrfs_abort_transaction(trans, root, ret);
2326                 btrfs_free_logged_extents(log, log_transid);
2327                 mutex_unlock(&root->log_mutex);
2328                 goto out;
2329         }
2330
2331         btrfs_set_root_node(&log->root_item, log->node);
2332
2333         root->log_transid++;
2334         log->log_transid = root->log_transid;
2335         root->log_start_pid = 0;
2336         smp_mb();
2337         /*
2338          * IO has been started, blocks of the log tree have WRITTEN flag set
2339          * in their headers. new modifications of the log will be written to
2340          * new positions. so it's safe to allow log writers to go in.
2341          */
2342         mutex_unlock(&root->log_mutex);
2343
2344         mutex_lock(&log_root_tree->log_mutex);
2345         atomic_inc(&log_root_tree->log_batch);
2346         atomic_inc(&log_root_tree->log_writers);
2347         mutex_unlock(&log_root_tree->log_mutex);
2348
2349         ret = update_log_root(trans, log);
2350
2351         mutex_lock(&log_root_tree->log_mutex);
2352         if (atomic_dec_and_test(&log_root_tree->log_writers)) {
2353                 smp_mb();
2354                 if (waitqueue_active(&log_root_tree->log_writer_wait))
2355                         wake_up(&log_root_tree->log_writer_wait);
2356         }
2357
2358         if (ret) {
2359                 if (ret != -ENOSPC) {
2360                         btrfs_abort_transaction(trans, root, ret);
2361                         mutex_unlock(&log_root_tree->log_mutex);
2362                         goto out;
2363                 }
2364                 root->fs_info->last_trans_log_full_commit = trans->transid;
2365                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2366                 btrfs_free_logged_extents(log, log_transid);
2367                 mutex_unlock(&log_root_tree->log_mutex);
2368                 ret = -EAGAIN;
2369                 goto out;
2370         }
2371
2372         index2 = log_root_tree->log_transid % 2;
2373         if (atomic_read(&log_root_tree->log_commit[index2])) {
2374                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2375                 wait_log_commit(trans, log_root_tree,
2376                                 log_root_tree->log_transid);
2377                 btrfs_free_logged_extents(log, log_transid);
2378                 mutex_unlock(&log_root_tree->log_mutex);
2379                 ret = 0;
2380                 goto out;
2381         }
2382         atomic_set(&log_root_tree->log_commit[index2], 1);
2383
2384         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2385                 wait_log_commit(trans, log_root_tree,
2386                                 log_root_tree->log_transid - 1);
2387         }
2388
2389         wait_for_writer(trans, log_root_tree);
2390
2391         /*
2392          * now that we've moved on to the tree of log tree roots,
2393          * check the full commit flag again
2394          */
2395         if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2396                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2397                 btrfs_free_logged_extents(log, log_transid);
2398                 mutex_unlock(&log_root_tree->log_mutex);
2399                 ret = -EAGAIN;
2400                 goto out_wake_log_root;
2401         }
2402
2403         ret = btrfs_write_and_wait_marked_extents(log_root_tree,
2404                                 &log_root_tree->dirty_log_pages,
2405                                 EXTENT_DIRTY | EXTENT_NEW);
2406         if (ret) {
2407                 btrfs_abort_transaction(trans, root, ret);
2408                 btrfs_free_logged_extents(log, log_transid);
2409                 mutex_unlock(&log_root_tree->log_mutex);
2410                 goto out_wake_log_root;
2411         }
2412         btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2413         btrfs_wait_logged_extents(log, log_transid);
2414
2415         btrfs_set_super_log_root(root->fs_info->super_for_commit,
2416                                 log_root_tree->node->start);
2417         btrfs_set_super_log_root_level(root->fs_info->super_for_commit,
2418                                 btrfs_header_level(log_root_tree->node));
2419
2420         log_root_tree->log_transid++;
2421         smp_mb();
2422
2423         mutex_unlock(&log_root_tree->log_mutex);
2424
2425         /*
2426          * nobody else is going to jump in and write the the ctree
2427          * super here because the log_commit atomic below is protecting
2428          * us.  We must be called with a transaction handle pinning
2429          * the running transaction open, so a full commit can't hop
2430          * in and cause problems either.
2431          */
2432         btrfs_scrub_pause_super(root);
2433         ret = write_ctree_super(trans, root->fs_info->tree_root, 1);
2434         btrfs_scrub_continue_super(root);
2435         if (ret) {
2436                 btrfs_abort_transaction(trans, root, ret);
2437                 goto out_wake_log_root;
2438         }
2439
2440         mutex_lock(&root->log_mutex);
2441         if (root->last_log_commit < log_transid)
2442                 root->last_log_commit = log_transid;
2443         mutex_unlock(&root->log_mutex);
2444
2445 out_wake_log_root:
2446         atomic_set(&log_root_tree->log_commit[index2], 0);
2447         smp_mb();
2448         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2449                 wake_up(&log_root_tree->log_commit_wait[index2]);
2450 out:
2451         atomic_set(&root->log_commit[index1], 0);
2452         smp_mb();
2453         if (waitqueue_active(&root->log_commit_wait[index1]))
2454                 wake_up(&root->log_commit_wait[index1]);
2455         return ret;
2456 }
2457
2458 static void free_log_tree(struct btrfs_trans_handle *trans,
2459                           struct btrfs_root *log)
2460 {
2461         int ret;
2462         u64 start;
2463         u64 end;
2464         struct walk_control wc = {
2465                 .free = 1,
2466                 .process_func = process_one_buffer
2467         };
2468
2469         if (trans) {
2470                 ret = walk_log_tree(trans, log, &wc);
2471                 BUG_ON(ret);
2472         }
2473
2474         while (1) {
2475                 ret = find_first_extent_bit(&log->dirty_log_pages,
2476                                 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW,
2477                                 NULL);
2478                 if (ret)
2479                         break;
2480
2481                 clear_extent_bits(&log->dirty_log_pages, start, end,
2482                                   EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2483         }
2484
2485         /*
2486          * We may have short-circuited the log tree with the full commit logic
2487          * and left ordered extents on our list, so clear these out to keep us
2488          * from leaking inodes and memory.
2489          */
2490         btrfs_free_logged_extents(log, 0);
2491         btrfs_free_logged_extents(log, 1);
2492
2493         free_extent_buffer(log->node);
2494         kfree(log);
2495 }
2496
2497 /*
2498  * free all the extents used by the tree log.  This should be called
2499  * at commit time of the full transaction
2500  */
2501 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2502 {
2503         if (root->log_root) {
2504                 free_log_tree(trans, root->log_root);
2505                 root->log_root = NULL;
2506         }
2507         return 0;
2508 }
2509
2510 int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2511                              struct btrfs_fs_info *fs_info)
2512 {
2513         if (fs_info->log_root_tree) {
2514                 free_log_tree(trans, fs_info->log_root_tree);
2515                 fs_info->log_root_tree = NULL;
2516         }
2517         return 0;
2518 }
2519
2520 /*
2521  * If both a file and directory are logged, and unlinks or renames are
2522  * mixed in, we have a few interesting corners:
2523  *
2524  * create file X in dir Y
2525  * link file X to X.link in dir Y
2526  * fsync file X
2527  * unlink file X but leave X.link
2528  * fsync dir Y
2529  *
2530  * After a crash we would expect only X.link to exist.  But file X
2531  * didn't get fsync'd again so the log has back refs for X and X.link.
2532  *
2533  * We solve this by removing directory entries and inode backrefs from the
2534  * log when a file that was logged in the current transaction is
2535  * unlinked.  Any later fsync will include the updated log entries, and
2536  * we'll be able to reconstruct the proper directory items from backrefs.
2537  *
2538  * This optimizations allows us to avoid relogging the entire inode
2539  * or the entire directory.
2540  */
2541 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2542                                  struct btrfs_root *root,
2543                                  const char *name, int name_len,
2544                                  struct inode *dir, u64 index)
2545 {
2546         struct btrfs_root *log;
2547         struct btrfs_dir_item *di;
2548         struct btrfs_path *path;
2549         int ret;
2550         int err = 0;
2551         int bytes_del = 0;
2552         u64 dir_ino = btrfs_ino(dir);
2553
2554         if (BTRFS_I(dir)->logged_trans < trans->transid)
2555                 return 0;
2556
2557         ret = join_running_log_trans(root);
2558         if (ret)
2559                 return 0;
2560
2561         mutex_lock(&BTRFS_I(dir)->log_mutex);
2562
2563         log = root->log_root;
2564         path = btrfs_alloc_path();
2565         if (!path) {
2566                 err = -ENOMEM;
2567                 goto out_unlock;
2568         }
2569
2570         di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2571                                    name, name_len, -1);
2572         if (IS_ERR(di)) {
2573                 err = PTR_ERR(di);
2574                 goto fail;
2575         }
2576         if (di) {
2577                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2578                 bytes_del += name_len;
2579                 BUG_ON(ret);
2580         }
2581         btrfs_release_path(path);
2582         di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2583                                          index, name, name_len, -1);
2584         if (IS_ERR(di)) {
2585                 err = PTR_ERR(di);
2586                 goto fail;
2587         }
2588         if (di) {
2589                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2590                 bytes_del += name_len;
2591                 BUG_ON(ret);
2592         }
2593
2594         /* update the directory size in the log to reflect the names
2595          * we have removed
2596          */
2597         if (bytes_del) {
2598                 struct btrfs_key key;
2599
2600                 key.objectid = dir_ino;
2601                 key.offset = 0;
2602                 key.type = BTRFS_INODE_ITEM_KEY;
2603                 btrfs_release_path(path);
2604
2605                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2606                 if (ret < 0) {
2607                         err = ret;
2608                         goto fail;
2609                 }
2610                 if (ret == 0) {
2611                         struct btrfs_inode_item *item;
2612                         u64 i_size;
2613
2614                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2615                                               struct btrfs_inode_item);
2616                         i_size = btrfs_inode_size(path->nodes[0], item);
2617                         if (i_size > bytes_del)
2618                                 i_size -= bytes_del;
2619                         else
2620                                 i_size = 0;
2621                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2622                         btrfs_mark_buffer_dirty(path->nodes[0]);
2623                 } else
2624                         ret = 0;
2625                 btrfs_release_path(path);
2626         }
2627 fail:
2628         btrfs_free_path(path);
2629 out_unlock:
2630         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2631         if (ret == -ENOSPC) {
2632                 root->fs_info->last_trans_log_full_commit = trans->transid;
2633                 ret = 0;
2634         } else if (ret < 0)
2635                 btrfs_abort_transaction(trans, root, ret);
2636
2637         btrfs_end_log_trans(root);
2638
2639         return err;
2640 }
2641
2642 /* see comments for btrfs_del_dir_entries_in_log */
2643 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2644                                struct btrfs_root *root,
2645                                const char *name, int name_len,
2646                                struct inode *inode, u64 dirid)
2647 {
2648         struct btrfs_root *log;
2649         u64 index;
2650         int ret;
2651
2652         if (BTRFS_I(inode)->logged_trans < trans->transid)
2653                 return 0;
2654
2655         ret = join_running_log_trans(root);
2656         if (ret)
2657                 return 0;
2658         log = root->log_root;
2659         mutex_lock(&BTRFS_I(inode)->log_mutex);
2660
2661         ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
2662                                   dirid, &index);
2663         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2664         if (ret == -ENOSPC) {
2665                 root->fs_info->last_trans_log_full_commit = trans->transid;
2666                 ret = 0;
2667         } else if (ret < 0 && ret != -ENOENT)
2668                 btrfs_abort_transaction(trans, root, ret);
2669         btrfs_end_log_trans(root);
2670
2671         return ret;
2672 }
2673
2674 /*
2675  * creates a range item in the log for 'dirid'.  first_offset and
2676  * last_offset tell us which parts of the key space the log should
2677  * be considered authoritative for.
2678  */
2679 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2680                                        struct btrfs_root *log,
2681                                        struct btrfs_path *path,
2682                                        int key_type, u64 dirid,
2683                                        u64 first_offset, u64 last_offset)
2684 {
2685         int ret;
2686         struct btrfs_key key;
2687         struct btrfs_dir_log_item *item;
2688
2689         key.objectid = dirid;
2690         key.offset = first_offset;
2691         if (key_type == BTRFS_DIR_ITEM_KEY)
2692                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2693         else
2694                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2695         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2696         if (ret)
2697                 return ret;
2698
2699         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2700                               struct btrfs_dir_log_item);
2701         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2702         btrfs_mark_buffer_dirty(path->nodes[0]);
2703         btrfs_release_path(path);
2704         return 0;
2705 }
2706
2707 /*
2708  * log all the items included in the current transaction for a given
2709  * directory.  This also creates the range items in the log tree required
2710  * to replay anything deleted before the fsync
2711  */
2712 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2713                           struct btrfs_root *root, struct inode *inode,
2714                           struct btrfs_path *path,
2715                           struct btrfs_path *dst_path, int key_type,
2716                           u64 min_offset, u64 *last_offset_ret)
2717 {
2718         struct btrfs_key min_key;
2719         struct btrfs_key max_key;
2720         struct btrfs_root *log = root->log_root;
2721         struct extent_buffer *src;
2722         int err = 0;
2723         int ret;
2724         int i;
2725         int nritems;
2726         u64 first_offset = min_offset;
2727         u64 last_offset = (u64)-1;
2728         u64 ino = btrfs_ino(inode);
2729
2730         log = root->log_root;
2731         max_key.objectid = ino;
2732         max_key.offset = (u64)-1;
2733         max_key.type = key_type;
2734
2735         min_key.objectid = ino;
2736         min_key.type = key_type;
2737         min_key.offset = min_offset;
2738
2739         path->keep_locks = 1;
2740
2741         ret = btrfs_search_forward(root, &min_key, &max_key,
2742                                    path, trans->transid);
2743
2744         /*
2745          * we didn't find anything from this transaction, see if there
2746          * is anything at all
2747          */
2748         if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
2749                 min_key.objectid = ino;
2750                 min_key.type = key_type;
2751                 min_key.offset = (u64)-1;
2752                 btrfs_release_path(path);
2753                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2754                 if (ret < 0) {
2755                         btrfs_release_path(path);
2756                         return ret;
2757                 }
2758                 ret = btrfs_previous_item(root, path, ino, key_type);
2759
2760                 /* if ret == 0 there are items for this type,
2761                  * create a range to tell us the last key of this type.
2762                  * otherwise, there are no items in this directory after
2763                  * *min_offset, and we create a range to indicate that.
2764                  */
2765                 if (ret == 0) {
2766                         struct btrfs_key tmp;
2767                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2768                                               path->slots[0]);
2769                         if (key_type == tmp.type)
2770                                 first_offset = max(min_offset, tmp.offset) + 1;
2771                 }
2772                 goto done;
2773         }
2774
2775         /* go backward to find any previous key */
2776         ret = btrfs_previous_item(root, path, ino, key_type);
2777         if (ret == 0) {
2778                 struct btrfs_key tmp;
2779                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2780                 if (key_type == tmp.type) {
2781                         first_offset = tmp.offset;
2782                         ret = overwrite_item(trans, log, dst_path,
2783                                              path->nodes[0], path->slots[0],
2784                                              &tmp);
2785                         if (ret) {
2786                                 err = ret;
2787                                 goto done;
2788                         }
2789                 }
2790         }
2791         btrfs_release_path(path);
2792
2793         /* find the first key from this transaction again */
2794         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2795         if (ret != 0) {
2796                 WARN_ON(1);
2797                 goto done;
2798         }
2799
2800         /*
2801          * we have a block from this transaction, log every item in it
2802          * from our directory
2803          */
2804         while (1) {
2805                 struct btrfs_key tmp;
2806                 src = path->nodes[0];
2807                 nritems = btrfs_header_nritems(src);
2808                 for (i = path->slots[0]; i < nritems; i++) {
2809                         btrfs_item_key_to_cpu(src, &min_key, i);
2810
2811                         if (min_key.objectid != ino || min_key.type != key_type)
2812                                 goto done;
2813                         ret = overwrite_item(trans, log, dst_path, src, i,
2814                                              &min_key);
2815                         if (ret) {
2816                                 err = ret;
2817                                 goto done;
2818                         }
2819                 }
2820                 path->slots[0] = nritems;
2821
2822                 /*
2823                  * look ahead to the next item and see if it is also
2824                  * from this directory and from this transaction
2825                  */
2826                 ret = btrfs_next_leaf(root, path);
2827                 if (ret == 1) {
2828                         last_offset = (u64)-1;
2829                         goto done;
2830                 }
2831                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2832                 if (tmp.objectid != ino || tmp.type != key_type) {
2833                         last_offset = (u64)-1;
2834                         goto done;
2835                 }
2836                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2837                         ret = overwrite_item(trans, log, dst_path,
2838                                              path->nodes[0], path->slots[0],
2839                                              &tmp);
2840                         if (ret)
2841                                 err = ret;
2842                         else
2843                                 last_offset = tmp.offset;
2844                         goto done;
2845                 }
2846         }
2847 done:
2848         btrfs_release_path(path);
2849         btrfs_release_path(dst_path);
2850
2851         if (err == 0) {
2852                 *last_offset_ret = last_offset;
2853                 /*
2854                  * insert the log range keys to indicate where the log
2855                  * is valid
2856                  */
2857                 ret = insert_dir_log_key(trans, log, path, key_type,
2858                                          ino, first_offset, last_offset);
2859                 if (ret)
2860                         err = ret;
2861         }
2862         return err;
2863 }
2864
2865 /*
2866  * logging directories is very similar to logging inodes, We find all the items
2867  * from the current transaction and write them to the log.
2868  *
2869  * The recovery code scans the directory in the subvolume, and if it finds a
2870  * key in the range logged that is not present in the log tree, then it means
2871  * that dir entry was unlinked during the transaction.
2872  *
2873  * In order for that scan to work, we must include one key smaller than
2874  * the smallest logged by this transaction and one key larger than the largest
2875  * key logged by this transaction.
2876  */
2877 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
2878                           struct btrfs_root *root, struct inode *inode,
2879                           struct btrfs_path *path,
2880                           struct btrfs_path *dst_path)
2881 {
2882         u64 min_key;
2883         u64 max_key;
2884         int ret;
2885         int key_type = BTRFS_DIR_ITEM_KEY;
2886
2887 again:
2888         min_key = 0;
2889         max_key = 0;
2890         while (1) {
2891                 ret = log_dir_items(trans, root, inode, path,
2892                                     dst_path, key_type, min_key,
2893                                     &max_key);
2894                 if (ret)
2895                         return ret;
2896                 if (max_key == (u64)-1)
2897                         break;
2898                 min_key = max_key + 1;
2899         }
2900
2901         if (key_type == BTRFS_DIR_ITEM_KEY) {
2902                 key_type = BTRFS_DIR_INDEX_KEY;
2903                 goto again;
2904         }
2905         return 0;
2906 }
2907
2908 /*
2909  * a helper function to drop items from the log before we relog an
2910  * inode.  max_key_type indicates the highest item type to remove.
2911  * This cannot be run for file data extents because it does not
2912  * free the extents they point to.
2913  */
2914 static int drop_objectid_items(struct btrfs_trans_handle *trans,
2915                                   struct btrfs_root *log,
2916                                   struct btrfs_path *path,
2917                                   u64 objectid, int max_key_type)
2918 {
2919         int ret;
2920         struct btrfs_key key;
2921         struct btrfs_key found_key;
2922         int start_slot;
2923
2924         key.objectid = objectid;
2925         key.type = max_key_type;
2926         key.offset = (u64)-1;
2927
2928         while (1) {
2929                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
2930                 BUG_ON(ret == 0);
2931                 if (ret < 0)
2932                         break;
2933
2934                 if (path->slots[0] == 0)
2935                         break;
2936
2937                 path->slots[0]--;
2938                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2939                                       path->slots[0]);
2940
2941                 if (found_key.objectid != objectid)
2942                         break;
2943
2944                 found_key.offset = 0;
2945                 found_key.type = 0;
2946                 ret = btrfs_bin_search(path->nodes[0], &found_key, 0,
2947                                        &start_slot);
2948
2949                 ret = btrfs_del_items(trans, log, path, start_slot,
2950                                       path->slots[0] - start_slot + 1);
2951                 /*
2952                  * If start slot isn't 0 then we don't need to re-search, we've
2953                  * found the last guy with the objectid in this tree.
2954                  */
2955                 if (ret || start_slot != 0)
2956                         break;
2957                 btrfs_release_path(path);
2958         }
2959         btrfs_release_path(path);
2960         if (ret > 0)
2961                 ret = 0;
2962         return ret;
2963 }
2964
2965 static void fill_inode_item(struct btrfs_trans_handle *trans,
2966                             struct extent_buffer *leaf,
2967                             struct btrfs_inode_item *item,
2968                             struct inode *inode, int log_inode_only)
2969 {
2970         struct btrfs_map_token token;
2971
2972         btrfs_init_map_token(&token);
2973
2974         if (log_inode_only) {
2975                 /* set the generation to zero so the recover code
2976                  * can tell the difference between an logging
2977                  * just to say 'this inode exists' and a logging
2978                  * to say 'update this inode with these values'
2979                  */
2980                 btrfs_set_token_inode_generation(leaf, item, 0, &token);
2981                 btrfs_set_token_inode_size(leaf, item, 0, &token);
2982         } else {
2983                 btrfs_set_token_inode_generation(leaf, item,
2984                                                  BTRFS_I(inode)->generation,
2985                                                  &token);
2986                 btrfs_set_token_inode_size(leaf, item, inode->i_size, &token);
2987         }
2988
2989         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
2990         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
2991         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
2992         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
2993
2994         btrfs_set_token_timespec_sec(leaf, btrfs_inode_atime(item),
2995                                      inode->i_atime.tv_sec, &token);
2996         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_atime(item),
2997                                       inode->i_atime.tv_nsec, &token);
2998
2999         btrfs_set_token_timespec_sec(leaf, btrfs_inode_mtime(item),
3000                                      inode->i_mtime.tv_sec, &token);
3001         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_mtime(item),
3002                                       inode->i_mtime.tv_nsec, &token);
3003
3004         btrfs_set_token_timespec_sec(leaf, btrfs_inode_ctime(item),
3005                                      inode->i_ctime.tv_sec, &token);
3006         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_ctime(item),
3007                                       inode->i_ctime.tv_nsec, &token);
3008
3009         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3010                                      &token);
3011
3012         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3013         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3014         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3015         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3016         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3017 }
3018
3019 static int log_inode_item(struct btrfs_trans_handle *trans,
3020                           struct btrfs_root *log, struct btrfs_path *path,
3021                           struct inode *inode)
3022 {
3023         struct btrfs_inode_item *inode_item;
3024         struct btrfs_key key;
3025         int ret;
3026
3027         memcpy(&key, &BTRFS_I(inode)->location, sizeof(key));
3028         ret = btrfs_insert_empty_item(trans, log, path, &key,
3029                                       sizeof(*inode_item));
3030         if (ret && ret != -EEXIST)
3031                 return ret;
3032         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3033                                     struct btrfs_inode_item);
3034         fill_inode_item(trans, path->nodes[0], inode_item, inode, 0);
3035         btrfs_release_path(path);
3036         return 0;
3037 }
3038
3039 static noinline int copy_items(struct btrfs_trans_handle *trans,
3040                                struct inode *inode,
3041                                struct btrfs_path *dst_path,
3042                                struct extent_buffer *src,
3043                                int start_slot, int nr, int inode_only)
3044 {
3045         unsigned long src_offset;
3046         unsigned long dst_offset;
3047         struct btrfs_root *log = BTRFS_I(inode)->root->log_root;
3048         struct btrfs_file_extent_item *extent;
3049         struct btrfs_inode_item *inode_item;
3050         int ret;
3051         struct btrfs_key *ins_keys;
3052         u32 *ins_sizes;
3053         char *ins_data;
3054         int i;
3055         struct list_head ordered_sums;
3056         int skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3057
3058         INIT_LIST_HEAD(&ordered_sums);
3059
3060         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
3061                            nr * sizeof(u32), GFP_NOFS);
3062         if (!ins_data)
3063                 return -ENOMEM;
3064
3065         ins_sizes = (u32 *)ins_data;
3066         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
3067
3068         for (i = 0; i < nr; i++) {
3069                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
3070                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
3071         }
3072         ret = btrfs_insert_empty_items(trans, log, dst_path,
3073                                        ins_keys, ins_sizes, nr);
3074         if (ret) {
3075                 kfree(ins_data);
3076                 return ret;
3077         }
3078
3079         for (i = 0; i < nr; i++, dst_path->slots[0]++) {
3080                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
3081                                                    dst_path->slots[0]);
3082
3083                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
3084
3085                 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
3086                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
3087                                                     dst_path->slots[0],
3088                                                     struct btrfs_inode_item);
3089                         fill_inode_item(trans, dst_path->nodes[0], inode_item,
3090                                         inode, inode_only == LOG_INODE_EXISTS);
3091                 } else {
3092                         copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
3093                                            src_offset, ins_sizes[i]);
3094                 }
3095
3096                 /* take a reference on file data extents so that truncates
3097                  * or deletes of this inode don't have to relog the inode
3098                  * again
3099                  */
3100                 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY &&
3101                     !skip_csum) {
3102                         int found_type;
3103                         extent = btrfs_item_ptr(src, start_slot + i,
3104                                                 struct btrfs_file_extent_item);
3105
3106                         if (btrfs_file_extent_generation(src, extent) < trans->transid)
3107                                 continue;
3108
3109                         found_type = btrfs_file_extent_type(src, extent);
3110                         if (found_type == BTRFS_FILE_EXTENT_REG) {
3111                                 u64 ds, dl, cs, cl;
3112                                 ds = btrfs_file_extent_disk_bytenr(src,
3113                                                                 extent);
3114                                 /* ds == 0 is a hole */
3115                                 if (ds == 0)
3116                                         continue;
3117
3118                                 dl = btrfs_file_extent_disk_num_bytes(src,
3119                                                                 extent);
3120                                 cs = btrfs_file_extent_offset(src, extent);
3121                                 cl = btrfs_file_extent_num_bytes(src,
3122                                                                 extent);
3123                                 if (btrfs_file_extent_compression(src,
3124                                                                   extent)) {
3125                                         cs = 0;
3126                                         cl = dl;
3127                                 }
3128
3129                                 ret = btrfs_lookup_csums_range(
3130                                                 log->fs_info->csum_root,
3131                                                 ds + cs, ds + cs + cl - 1,
3132                                                 &ordered_sums, 0);
3133                                 BUG_ON(ret);
3134                         }
3135                 }
3136         }
3137
3138         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
3139         btrfs_release_path(dst_path);
3140         kfree(ins_data);
3141
3142         /*
3143          * we have to do this after the loop above to avoid changing the
3144          * log tree while trying to change the log tree.
3145          */
3146         ret = 0;
3147         while (!list_empty(&ordered_sums)) {
3148                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3149                                                    struct btrfs_ordered_sum,
3150                                                    list);
3151                 if (!ret)
3152                         ret = btrfs_csum_file_blocks(trans, log, sums);
3153                 list_del(&sums->list);
3154                 kfree(sums);
3155         }
3156         return ret;
3157 }
3158
3159 static int extent_cmp(void *priv, struct list_head *a, struct list_head *b)
3160 {
3161         struct extent_map *em1, *em2;
3162
3163         em1 = list_entry(a, struct extent_map, list);
3164         em2 = list_entry(b, struct extent_map, list);
3165
3166         if (em1->start < em2->start)
3167                 return -1;
3168         else if (em1->start > em2->start)
3169                 return 1;
3170         return 0;
3171 }
3172
3173 static int drop_adjacent_extents(struct btrfs_trans_handle *trans,
3174                                  struct btrfs_root *root, struct inode *inode,
3175                                  struct extent_map *em,
3176                                  struct btrfs_path *path)
3177 {
3178         struct btrfs_file_extent_item *fi;
3179         struct extent_buffer *leaf;
3180         struct btrfs_key key, new_key;
3181         struct btrfs_map_token token;
3182         u64 extent_end;
3183         u64 extent_offset = 0;
3184         int extent_type;
3185         int del_slot = 0;
3186         int del_nr = 0;
3187         int ret = 0;
3188
3189         while (1) {
3190                 btrfs_init_map_token(&token);
3191                 leaf = path->nodes[0];
3192                 path->slots[0]++;
3193                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3194                         if (del_nr) {
3195                                 ret = btrfs_del_items(trans, root, path,
3196                                                       del_slot, del_nr);
3197                                 if (ret)
3198                                         return ret;
3199                                 del_nr = 0;
3200                         }
3201
3202                         ret = btrfs_next_leaf_write(trans, root, path, 1);
3203                         if (ret < 0)
3204                                 return ret;
3205                         if (ret > 0)
3206                                 return 0;
3207                         leaf = path->nodes[0];
3208                 }
3209
3210                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3211                 if (key.objectid != btrfs_ino(inode) ||
3212                     key.type != BTRFS_EXTENT_DATA_KEY ||
3213                     key.offset >= em->start + em->len)
3214                         break;
3215
3216                 fi = btrfs_item_ptr(leaf, path->slots[0],
3217                                     struct btrfs_file_extent_item);
3218                 extent_type = btrfs_token_file_extent_type(leaf, fi, &token);
3219                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
3220                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3221                         extent_offset = btrfs_token_file_extent_offset(leaf,
3222                                                                 fi, &token);
3223                         extent_end = key.offset +
3224                                 btrfs_token_file_extent_num_bytes(leaf, fi,
3225                                                                   &token);
3226                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3227                         extent_end = key.offset +
3228                                 btrfs_file_extent_inline_len(leaf, fi);
3229                 } else {
3230                         BUG();
3231                 }
3232
3233                 if (extent_end <= em->len + em->start) {
3234                         if (!del_nr) {
3235                                 del_slot = path->slots[0];
3236                         }
3237                         del_nr++;
3238                         continue;
3239                 }
3240
3241                 /*
3242                  * Ok so we'll ignore previous items if we log a new extent,
3243                  * which can lead to overlapping extents, so if we have an
3244                  * existing extent we want to adjust we _have_ to check the next
3245                  * guy to make sure we even need this extent anymore, this keeps
3246                  * us from panicing in set_item_key_safe.
3247                  */
3248                 if (path->slots[0] < btrfs_header_nritems(leaf) - 1) {
3249                         struct btrfs_key tmp_key;
3250
3251                         btrfs_item_key_to_cpu(leaf, &tmp_key,
3252                                               path->slots[0] + 1);
3253                         if (tmp_key.objectid == btrfs_ino(inode) &&
3254                             tmp_key.type == BTRFS_EXTENT_DATA_KEY &&
3255                             tmp_key.offset <= em->start + em->len) {
3256                                 if (!del_nr)
3257                                         del_slot = path->slots[0];
3258                                 del_nr++;
3259                                 continue;
3260                         }
3261                 }
3262
3263                 BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
3264                 memcpy(&new_key, &key, sizeof(new_key));
3265                 new_key.offset = em->start + em->len;
3266                 btrfs_set_item_key_safe(trans, root, path, &new_key);
3267                 extent_offset += em->start + em->len - key.offset;
3268                 btrfs_set_token_file_extent_offset(leaf, fi, extent_offset,
3269                                                    &token);
3270                 btrfs_set_token_file_extent_num_bytes(leaf, fi, extent_end -
3271                                                       (em->start + em->len),
3272                                                       &token);
3273                 btrfs_mark_buffer_dirty(leaf);
3274         }
3275
3276         if (del_nr)
3277                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
3278
3279         return ret;
3280 }
3281
3282 static int log_one_extent(struct btrfs_trans_handle *trans,
3283                           struct inode *inode, struct btrfs_root *root,
3284                           struct extent_map *em, struct btrfs_path *path)
3285 {
3286         struct btrfs_root *log = root->log_root;
3287         struct btrfs_file_extent_item *fi;
3288         struct extent_buffer *leaf;
3289         struct btrfs_ordered_extent *ordered;
3290         struct list_head ordered_sums;
3291         struct btrfs_map_token token;
3292         struct btrfs_key key;
3293         u64 mod_start = em->mod_start;
3294         u64 mod_len = em->mod_len;
3295         u64 csum_offset;
3296         u64 csum_len;
3297         u64 extent_offset = em->start - em->orig_start;
3298         u64 block_len;
3299         int ret;
3300         int index = log->log_transid % 2;
3301         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3302
3303 insert:
3304         INIT_LIST_HEAD(&ordered_sums);
3305         btrfs_init_map_token(&token);
3306         key.objectid = btrfs_ino(inode);
3307         key.type = BTRFS_EXTENT_DATA_KEY;
3308         key.offset = em->start;
3309         path->really_keep_locks = 1;
3310
3311         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*fi));
3312         if (ret && ret != -EEXIST) {
3313                 path->really_keep_locks = 0;
3314                 return ret;
3315         }
3316         leaf = path->nodes[0];
3317         fi = btrfs_item_ptr(leaf, path->slots[0],
3318                             struct btrfs_file_extent_item);
3319
3320         /*
3321          * If we are overwriting an inline extent with a real one then we need
3322          * to just delete the inline extent as it may not be large enough to
3323          * have the entire file_extent_item.
3324          */
3325         if (ret && btrfs_token_file_extent_type(leaf, fi, &token) ==
3326             BTRFS_FILE_EXTENT_INLINE) {
3327                 ret = btrfs_del_item(trans, log, path);
3328                 btrfs_release_path(path);
3329                 if (ret) {
3330                         path->really_keep_locks = 0;
3331                         return ret;
3332                 }
3333                 goto insert;
3334         }
3335
3336         btrfs_set_token_file_extent_generation(leaf, fi, em->generation,
3337                                                &token);
3338         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3339                 skip_csum = true;
3340                 btrfs_set_token_file_extent_type(leaf, fi,
3341                                                  BTRFS_FILE_EXTENT_PREALLOC,
3342                                                  &token);
3343         } else {
3344                 btrfs_set_token_file_extent_type(leaf, fi,
3345                                                  BTRFS_FILE_EXTENT_REG,
3346                                                  &token);
3347                 if (em->block_start == 0)
3348                         skip_csum = true;
3349         }
3350
3351         block_len = max(em->block_len, em->orig_block_len);
3352         if (em->compress_type != BTRFS_COMPRESS_NONE) {
3353                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3354                                                         em->block_start,
3355                                                         &token);
3356                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3357                                                            &token);
3358         } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
3359                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3360                                                         em->block_start -
3361                                                         extent_offset, &token);
3362                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3363                                                            &token);
3364         } else {
3365                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi, 0, &token);
3366                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, 0,
3367                                                            &token);
3368         }
3369
3370         btrfs_set_token_file_extent_offset(leaf, fi,
3371                                            em->start - em->orig_start,
3372                                            &token);
3373         btrfs_set_token_file_extent_num_bytes(leaf, fi, em->len, &token);
3374         btrfs_set_token_file_extent_ram_bytes(leaf, fi, em->len, &token);
3375         btrfs_set_token_file_extent_compression(leaf, fi, em->compress_type,
3376                                                 &token);
3377         btrfs_set_token_file_extent_encryption(leaf, fi, 0, &token);
3378         btrfs_set_token_file_extent_other_encoding(leaf, fi, 0, &token);
3379         btrfs_mark_buffer_dirty(leaf);
3380
3381         /*
3382          * Have to check the extent to the right of us to make sure it doesn't
3383          * fall in our current range.  We're ok if the previous extent is in our
3384          * range since the recovery stuff will run us in key order and thus just
3385          * drop the part we overwrote.
3386          */
3387         ret = drop_adjacent_extents(trans, log, inode, em, path);
3388         btrfs_release_path(path);
3389         path->really_keep_locks = 0;
3390         if (ret) {
3391                 return ret;
3392         }
3393
3394         if (skip_csum)
3395                 return 0;
3396
3397         if (em->compress_type) {
3398                 csum_offset = 0;
3399                 csum_len = block_len;
3400         }
3401
3402         /*
3403          * First check and see if our csums are on our outstanding ordered
3404          * extents.
3405          */
3406 again:
3407         spin_lock_irq(&log->log_extents_lock[index]);
3408         list_for_each_entry(ordered, &log->logged_list[index], log_list) {
3409                 struct btrfs_ordered_sum *sum;
3410
3411                 if (!mod_len)
3412                         break;
3413
3414                 if (ordered->inode != inode)
3415                         continue;
3416
3417                 if (ordered->file_offset + ordered->len <= mod_start ||
3418                     mod_start + mod_len <= ordered->file_offset)
3419                         continue;
3420
3421                 /*
3422                  * We are going to copy all the csums on this ordered extent, so
3423                  * go ahead and adjust mod_start and mod_len in case this
3424                  * ordered extent has already been logged.
3425                  */
3426                 if (ordered->file_offset > mod_start) {
3427                         if (ordered->file_offset + ordered->len >=
3428                             mod_start + mod_len)
3429                                 mod_len = ordered->file_offset - mod_start;
3430                         /*
3431                          * If we have this case
3432                          *
3433                          * |--------- logged extent ---------|
3434                          *       |----- ordered extent ----|
3435                          *
3436                          * Just don't mess with mod_start and mod_len, we'll
3437                          * just end up logging more csums than we need and it
3438                          * will be ok.
3439                          */
3440                 } else {
3441                         if (ordered->file_offset + ordered->len <
3442                             mod_start + mod_len) {
3443                                 mod_len = (mod_start + mod_len) -
3444                                         (ordered->file_offset + ordered->len);
3445                                 mod_start = ordered->file_offset +
3446                                         ordered->len;
3447                         } else {
3448                                 mod_len = 0;
3449                         }
3450                 }
3451
3452                 /*
3453                  * To keep us from looping for the above case of an ordered
3454                  * extent that falls inside of the logged extent.
3455                  */
3456                 if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM,
3457                                      &ordered->flags))
3458                         continue;
3459                 atomic_inc(&ordered->refs);
3460                 spin_unlock_irq(&log->log_extents_lock[index]);
3461                 /*
3462                  * we've dropped the lock, we must either break or
3463                  * start over after this.
3464                  */
3465
3466                 wait_event(ordered->wait, ordered->csum_bytes_left == 0);
3467
3468                 list_for_each_entry(sum, &ordered->list, list) {
3469                         ret = btrfs_csum_file_blocks(trans, log, sum);
3470                         if (ret) {
3471                                 btrfs_put_ordered_extent(ordered);
3472                                 goto unlocked;
3473                         }
3474                 }
3475                 btrfs_put_ordered_extent(ordered);
3476                 goto again;
3477
3478         }
3479         spin_unlock_irq(&log->log_extents_lock[index]);
3480 unlocked:
3481
3482         if (!mod_len || ret)
3483                 return ret;
3484
3485         csum_offset = mod_start - em->start;
3486         csum_len = mod_len;
3487
3488         /* block start is already adjusted for the file extent offset. */
3489         ret = btrfs_lookup_csums_range(log->fs_info->csum_root,
3490                                        em->block_start + csum_offset,
3491                                        em->block_start + csum_offset +
3492                                        csum_len - 1, &ordered_sums, 0);
3493         if (ret)
3494                 return ret;
3495
3496         while (!list_empty(&ordered_sums)) {
3497                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3498                                                    struct btrfs_ordered_sum,
3499                                                    list);
3500                 if (!ret)
3501                         ret = btrfs_csum_file_blocks(trans, log, sums);
3502                 list_del(&sums->list);
3503                 kfree(sums);
3504         }
3505
3506         return ret;
3507 }
3508
3509 static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
3510                                      struct btrfs_root *root,
3511                                      struct inode *inode,
3512                                      struct btrfs_path *path)
3513 {
3514         struct extent_map *em, *n;
3515         struct list_head extents;
3516         struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3517         u64 test_gen;
3518         int ret = 0;
3519         int num = 0;
3520
3521         INIT_LIST_HEAD(&extents);
3522
3523         write_lock(&tree->lock);
3524         test_gen = root->fs_info->last_trans_committed;
3525
3526         list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
3527                 list_del_init(&em->list);
3528
3529                 /*
3530                  * Just an arbitrary number, this can be really CPU intensive
3531                  * once we start getting a lot of extents, and really once we
3532                  * have a bunch of extents we just want to commit since it will
3533                  * be faster.
3534                  */
3535                 if (++num > 32768) {
3536                         list_del_init(&tree->modified_extents);
3537                         ret = -EFBIG;
3538                         goto process;
3539                 }
3540
3541                 if (em->generation <= test_gen)
3542                         continue;
3543                 /* Need a ref to keep it from getting evicted from cache */
3544                 atomic_inc(&em->refs);
3545                 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
3546                 list_add_tail(&em->list, &extents);
3547                 num++;
3548         }
3549
3550         list_sort(NULL, &extents, extent_cmp);
3551
3552 process:
3553         while (!list_empty(&extents)) {
3554                 em = list_entry(extents.next, struct extent_map, list);
3555
3556                 list_del_init(&em->list);
3557
3558                 /*
3559                  * If we had an error we just need to delete everybody from our
3560                  * private list.
3561                  */
3562                 if (ret) {
3563                         clear_em_logging(tree, em);
3564                         free_extent_map(em);
3565                         continue;
3566                 }
3567
3568                 write_unlock(&tree->lock);
3569
3570                 ret = log_one_extent(trans, inode, root, em, path);
3571                 write_lock(&tree->lock);
3572                 clear_em_logging(tree, em);
3573                 free_extent_map(em);
3574         }
3575         WARN_ON(!list_empty(&extents));
3576         write_unlock(&tree->lock);
3577
3578         btrfs_release_path(path);
3579         return ret;
3580 }
3581
3582 /* log a single inode in the tree log.
3583  * At least one parent directory for this inode must exist in the tree
3584  * or be logged already.
3585  *
3586  * Any items from this inode changed by the current transaction are copied
3587  * to the log tree.  An extra reference is taken on any extents in this
3588  * file, allowing us to avoid a whole pile of corner cases around logging
3589  * blocks that have been removed from the tree.
3590  *
3591  * See LOG_INODE_ALL and related defines for a description of what inode_only
3592  * does.
3593  *
3594  * This handles both files and directories.
3595  */
3596 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
3597                              struct btrfs_root *root, struct inode *inode,
3598                              int inode_only)
3599 {
3600         struct btrfs_path *path;
3601         struct btrfs_path *dst_path;
3602         struct btrfs_key min_key;
3603         struct btrfs_key max_key;
3604         struct btrfs_root *log = root->log_root;
3605         struct extent_buffer *src = NULL;
3606         int err = 0;
3607         int ret;
3608         int nritems;
3609         int ins_start_slot = 0;
3610         int ins_nr;
3611         bool fast_search = false;
3612         u64 ino = btrfs_ino(inode);
3613
3614         log = root->log_root;
3615
3616         path = btrfs_alloc_path();
3617         if (!path)
3618                 return -ENOMEM;
3619         dst_path = btrfs_alloc_path();
3620         if (!dst_path) {
3621                 btrfs_free_path(path);
3622                 return -ENOMEM;
3623         }
3624
3625         min_key.objectid = ino;
3626         min_key.type = BTRFS_INODE_ITEM_KEY;
3627         min_key.offset = 0;
3628
3629         max_key.objectid = ino;
3630
3631
3632         /* today the code can only do partial logging of directories */
3633         if (S_ISDIR(inode->i_mode) ||
3634             (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3635                        &BTRFS_I(inode)->runtime_flags) &&
3636              inode_only == LOG_INODE_EXISTS))
3637                 max_key.type = BTRFS_XATTR_ITEM_KEY;
3638         else
3639                 max_key.type = (u8)-1;
3640         max_key.offset = (u64)-1;
3641
3642         /* Only run delayed items if we are a dir or a new file */
3643         if (S_ISDIR(inode->i_mode) ||
3644             BTRFS_I(inode)->generation > root->fs_info->last_trans_committed) {
3645                 ret = btrfs_commit_inode_delayed_items(trans, inode);
3646                 if (ret) {
3647                         btrfs_free_path(path);
3648                         btrfs_free_path(dst_path);
3649                         return ret;
3650                 }
3651         }
3652
3653         mutex_lock(&BTRFS_I(inode)->log_mutex);
3654
3655         btrfs_get_logged_extents(log, inode);
3656
3657         /*
3658          * a brute force approach to making sure we get the most uptodate
3659          * copies of everything.
3660          */
3661         if (S_ISDIR(inode->i_mode)) {
3662                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
3663
3664                 if (inode_only == LOG_INODE_EXISTS)
3665                         max_key_type = BTRFS_XATTR_ITEM_KEY;
3666                 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
3667         } else {
3668                 if (test_and_clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3669                                        &BTRFS_I(inode)->runtime_flags)) {
3670                         clear_bit(BTRFS_INODE_COPY_EVERYTHING,
3671                                   &BTRFS_I(inode)->runtime_flags);
3672                         ret = btrfs_truncate_inode_items(trans, log,
3673                                                          inode, 0, 0);
3674                 } else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
3675                                               &BTRFS_I(inode)->runtime_flags)) {
3676                         if (inode_only == LOG_INODE_ALL)
3677                                 fast_search = true;
3678                         max_key.type = BTRFS_XATTR_ITEM_KEY;
3679                         ret = drop_objectid_items(trans, log, path, ino,
3680                                                   max_key.type);
3681                 } else {
3682                         if (inode_only == LOG_INODE_ALL)
3683                                 fast_search = true;
3684                         ret = log_inode_item(trans, log, dst_path, inode);
3685                         if (ret) {
3686                                 err = ret;
3687                                 goto out_unlock;
3688                         }
3689                         goto log_extents;
3690                 }
3691
3692         }
3693         if (ret) {
3694                 err = ret;
3695                 goto out_unlock;
3696         }
3697         path->keep_locks = 1;
3698
3699         while (1) {
3700                 ins_nr = 0;
3701                 ret = btrfs_search_forward(root, &min_key, &max_key,
3702                                            path, trans->transid);
3703                 if (ret != 0)
3704                         break;
3705 again:
3706                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
3707                 if (min_key.objectid != ino)
3708                         break;
3709                 if (min_key.type > max_key.type)
3710                         break;
3711
3712                 src = path->nodes[0];
3713                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
3714                         ins_nr++;
3715                         goto next_slot;
3716                 } else if (!ins_nr) {
3717                         ins_start_slot = path->slots[0];
3718                         ins_nr = 1;
3719                         goto next_slot;
3720                 }
3721
3722                 ret = copy_items(trans, inode, dst_path, src, ins_start_slot,
3723                                  ins_nr, inode_only);
3724                 if (ret) {
3725                         err = ret;
3726                         goto out_unlock;
3727                 }
3728                 ins_nr = 1;
3729                 ins_start_slot = path->slots[0];
3730 next_slot:
3731
3732                 nritems = btrfs_header_nritems(path->nodes[0]);
3733                 path->slots[0]++;
3734                 if (path->slots[0] < nritems) {
3735                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
3736                                               path->slots[0]);
3737                         goto again;
3738                 }
3739                 if (ins_nr) {
3740                         ret = copy_items(trans, inode, dst_path, src,
3741                                          ins_start_slot,
3742                                          ins_nr, inode_only);
3743                         if (ret) {
3744                                 err = ret;
3745                                 goto out_unlock;
3746                         }
3747                         ins_nr = 0;
3748                 }
3749                 btrfs_release_path(path);
3750
3751                 if (min_key.offset < (u64)-1)
3752                         min_key.offset++;
3753                 else if (min_key.type < (u8)-1)
3754                         min_key.type++;
3755                 else if (min_key.objectid < (u64)-1)
3756                         min_key.objectid++;
3757                 else
3758                         break;
3759         }
3760         if (ins_nr) {
3761                 ret = copy_items(trans, inode, dst_path, src, ins_start_slot,
3762                                  ins_nr, inode_only);
3763                 if (ret) {
3764                         err = ret;
3765                         goto out_unlock;
3766                 }
3767                 ins_nr = 0;
3768         }
3769
3770 log_extents:
3771         if (fast_search) {
3772                 btrfs_release_path(dst_path);
3773                 ret = btrfs_log_changed_extents(trans, root, inode, dst_path);
3774                 if (ret) {
3775                         err = ret;
3776                         goto out_unlock;
3777                 }
3778         } else {
3779                 struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3780                 struct extent_map *em, *n;
3781
3782                 write_lock(&tree->lock);
3783                 list_for_each_entry_safe(em, n, &tree->modified_extents, list)
3784                         list_del_init(&em->list);
3785                 write_unlock(&tree->lock);
3786         }
3787
3788         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
3789                 btrfs_release_path(path);
3790                 btrfs_release_path(dst_path);
3791                 ret = log_directory_changes(trans, root, inode, path, dst_path);
3792                 if (ret) {
3793                         err = ret;
3794                         goto out_unlock;
3795                 }
3796         }
3797         BTRFS_I(inode)->logged_trans = trans->transid;
3798         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->last_sub_trans;
3799 out_unlock:
3800         if (err)
3801                 btrfs_free_logged_extents(log, log->log_transid);
3802         mutex_unlock(&BTRFS_I(inode)->log_mutex);
3803
3804         btrfs_free_path(path);
3805         btrfs_free_path(dst_path);
3806         return err;
3807 }
3808
3809 /*
3810  * follow the dentry parent pointers up the chain and see if any
3811  * of the directories in it require a full commit before they can
3812  * be logged.  Returns zero if nothing special needs to be done or 1 if
3813  * a full commit is required.
3814  */
3815 static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
3816                                                struct inode *inode,
3817                                                struct dentry *parent,
3818                                                struct super_block *sb,
3819                                                u64 last_committed)
3820 {
3821         int ret = 0;
3822         struct btrfs_root *root;
3823         struct dentry *old_parent = NULL;
3824
3825         /*
3826          * for regular files, if its inode is already on disk, we don't
3827          * have to worry about the parents at all.  This is because
3828          * we can use the last_unlink_trans field to record renames
3829          * and other fun in this file.
3830          */
3831         if (S_ISREG(inode->i_mode) &&
3832             BTRFS_I(inode)->generation <= last_committed &&
3833             BTRFS_I(inode)->last_unlink_trans <= last_committed)
3834                         goto out;
3835
3836         if (!S_ISDIR(inode->i_mode)) {
3837                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3838                         goto out;
3839                 inode = parent->d_inode;
3840         }
3841
3842         while (1) {
3843                 BTRFS_I(inode)->logged_trans = trans->transid;
3844                 smp_mb();
3845
3846                 if (BTRFS_I(inode)->last_unlink_trans > last_committed) {
3847                         root = BTRFS_I(inode)->root;
3848
3849                         /*
3850                          * make sure any commits to the log are forced
3851                          * to be full commits
3852                          */
3853                         root->fs_info->last_trans_log_full_commit =
3854                                 trans->transid;
3855                         ret = 1;
3856                         break;
3857                 }
3858
3859                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3860                         break;
3861
3862                 if (IS_ROOT(parent))
3863                         break;
3864
3865                 parent = dget_parent(parent);
3866                 dput(old_parent);
3867                 old_parent = parent;
3868                 inode = parent->d_inode;
3869
3870         }
3871         dput(old_parent);
3872 out:
3873         return ret;
3874 }
3875
3876 /*
3877  * helper function around btrfs_log_inode to make sure newly created
3878  * parent directories also end up in the log.  A minimal inode and backref
3879  * only logging is done of any parent directories that are older than
3880  * the last committed transaction
3881  */
3882 int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
3883                     struct btrfs_root *root, struct inode *inode,
3884                     struct dentry *parent, int exists_only)
3885 {
3886         int inode_only = exists_only ? LOG_INODE_EXISTS : LOG_INODE_ALL;
3887         struct super_block *sb;
3888         struct dentry *old_parent = NULL;
3889         int ret = 0;
3890         u64 last_committed = root->fs_info->last_trans_committed;
3891
3892         sb = inode->i_sb;
3893
3894         if (btrfs_test_opt(root, NOTREELOG)) {
3895                 ret = 1;
3896                 goto end_no_trans;
3897         }
3898
3899         if (root->fs_info->last_trans_log_full_commit >
3900             root->fs_info->last_trans_committed) {
3901                 ret = 1;
3902                 goto end_no_trans;
3903         }
3904
3905         if (root != BTRFS_I(inode)->root ||
3906             btrfs_root_refs(&root->root_item) == 0) {
3907                 ret = 1;
3908                 goto end_no_trans;
3909         }
3910
3911         ret = check_parent_dirs_for_sync(trans, inode, parent,
3912                                          sb, last_committed);
3913         if (ret)
3914                 goto end_no_trans;
3915
3916         if (btrfs_inode_in_log(inode, trans->transid)) {
3917                 ret = BTRFS_NO_LOG_SYNC;
3918                 goto end_no_trans;
3919         }
3920
3921         ret = start_log_trans(trans, root);
3922         if (ret)
3923                 goto end_trans;
3924
3925         ret = btrfs_log_inode(trans, root, inode, inode_only);
3926         if (ret)
3927                 goto end_trans;
3928
3929         /*
3930          * for regular files, if its inode is already on disk, we don't
3931          * have to worry about the parents at all.  This is because
3932          * we can use the last_unlink_trans field to record renames
3933          * and other fun in this file.
3934          */
3935         if (S_ISREG(inode->i_mode) &&
3936             BTRFS_I(inode)->generation <= last_committed &&
3937             BTRFS_I(inode)->last_unlink_trans <= last_committed) {
3938                 ret = 0;
3939                 goto end_trans;
3940         }
3941
3942         inode_only = LOG_INODE_EXISTS;
3943         while (1) {
3944                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3945                         break;
3946
3947                 inode = parent->d_inode;
3948                 if (root != BTRFS_I(inode)->root)
3949                         break;
3950
3951                 if (BTRFS_I(inode)->generation >
3952                     root->fs_info->last_trans_committed) {
3953                         ret = btrfs_log_inode(trans, root, inode, inode_only);
3954                         if (ret)
3955                                 goto end_trans;
3956                 }
3957                 if (IS_ROOT(parent))
3958                         break;
3959
3960                 parent = dget_parent(parent);
3961                 dput(old_parent);
3962                 old_parent = parent;
3963         }
3964         ret = 0;
3965 end_trans:
3966         dput(old_parent);
3967         if (ret < 0) {
3968                 root->fs_info->last_trans_log_full_commit = trans->transid;
3969                 ret = 1;
3970         }
3971         btrfs_end_log_trans(root);
3972 end_no_trans:
3973         return ret;
3974 }
3975
3976 /*
3977  * it is not safe to log dentry if the chunk root has added new
3978  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
3979  * If this returns 1, you must commit the transaction to safely get your
3980  * data on disk.
3981  */
3982 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
3983                           struct btrfs_root *root, struct dentry *dentry)
3984 {
3985         struct dentry *parent = dget_parent(dentry);
3986         int ret;
3987
3988         ret = btrfs_log_inode_parent(trans, root, dentry->d_inode, parent, 0);
3989         dput(parent);
3990
3991         return ret;
3992 }
3993
3994 /*
3995  * should be called during mount to recover any replay any log trees
3996  * from the FS
3997  */
3998 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
3999 {
4000         int ret;
4001         struct btrfs_path *path;
4002         struct btrfs_trans_handle *trans;
4003         struct btrfs_key key;
4004         struct btrfs_key found_key;
4005         struct btrfs_key tmp_key;
4006         struct btrfs_root *log;
4007         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
4008         struct walk_control wc = {
4009                 .process_func = process_one_buffer,
4010                 .stage = 0,
4011         };
4012
4013         path = btrfs_alloc_path();
4014         if (!path)
4015                 return -ENOMEM;
4016
4017         fs_info->log_root_recovering = 1;
4018
4019         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4020         if (IS_ERR(trans)) {
4021                 ret = PTR_ERR(trans);
4022                 goto error;
4023         }
4024
4025         wc.trans = trans;
4026         wc.pin = 1;
4027
4028         ret = walk_log_tree(trans, log_root_tree, &wc);
4029         if (ret) {
4030                 btrfs_error(fs_info, ret, "Failed to pin buffers while "
4031                             "recovering log root tree.");
4032                 goto error;
4033         }
4034
4035 again:
4036         key.objectid = BTRFS_TREE_LOG_OBJECTID;
4037         key.offset = (u64)-1;
4038         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
4039
4040         while (1) {
4041                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
4042
4043                 if (ret < 0) {
4044                         btrfs_error(fs_info, ret,
4045                                     "Couldn't find tree log root.");
4046                         goto error;
4047                 }
4048                 if (ret > 0) {
4049                         if (path->slots[0] == 0)
4050                                 break;
4051                         path->slots[0]--;
4052                 }
4053                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
4054                                       path->slots[0]);
4055                 btrfs_release_path(path);
4056                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4057                         break;
4058
4059                 log = btrfs_read_fs_root_no_radix(log_root_tree,
4060                                                   &found_key);
4061                 if (IS_ERR(log)) {
4062                         ret = PTR_ERR(log);
4063                         btrfs_error(fs_info, ret,
4064                                     "Couldn't read tree log root.");
4065                         goto error;
4066                 }
4067
4068                 tmp_key.objectid = found_key.offset;
4069                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
4070                 tmp_key.offset = (u64)-1;
4071
4072                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
4073                 if (IS_ERR(wc.replay_dest)) {
4074                         ret = PTR_ERR(wc.replay_dest);
4075                         btrfs_error(fs_info, ret, "Couldn't read target root "
4076                                     "for tree log recovery.");
4077                         goto error;
4078                 }
4079
4080                 wc.replay_dest->log_root = log;
4081                 btrfs_record_root_in_trans(trans, wc.replay_dest);
4082                 ret = walk_log_tree(trans, log, &wc);
4083                 BUG_ON(ret);
4084
4085                 if (wc.stage == LOG_WALK_REPLAY_ALL) {
4086                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
4087                                                       path);
4088                         BUG_ON(ret);
4089                 }
4090
4091                 key.offset = found_key.offset - 1;
4092                 wc.replay_dest->log_root = NULL;
4093                 free_extent_buffer(log->node);
4094                 free_extent_buffer(log->commit_root);
4095                 kfree(log);
4096
4097                 if (found_key.offset == 0)
4098                         break;
4099         }
4100         btrfs_release_path(path);
4101
4102         /* step one is to pin it all, step two is to replay just inodes */
4103         if (wc.pin) {
4104                 wc.pin = 0;
4105                 wc.process_func = replay_one_buffer;
4106                 wc.stage = LOG_WALK_REPLAY_INODES;
4107                 goto again;
4108         }
4109         /* step three is to replay everything */
4110         if (wc.stage < LOG_WALK_REPLAY_ALL) {
4111                 wc.stage++;
4112                 goto again;
4113         }
4114
4115         btrfs_free_path(path);
4116
4117         free_extent_buffer(log_root_tree->node);
4118         log_root_tree->log_root = NULL;
4119         fs_info->log_root_recovering = 0;
4120
4121         /* step 4: commit the transaction, which also unpins the blocks */
4122         btrfs_commit_transaction(trans, fs_info->tree_root);
4123
4124         kfree(log_root_tree);
4125         return 0;
4126
4127 error:
4128         btrfs_free_path(path);
4129         return ret;
4130 }
4131
4132 /*
4133  * there are some corner cases where we want to force a full
4134  * commit instead of allowing a directory to be logged.
4135  *
4136  * They revolve around files there were unlinked from the directory, and
4137  * this function updates the parent directory so that a full commit is
4138  * properly done if it is fsync'd later after the unlinks are done.
4139  */
4140 void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
4141                              struct inode *dir, struct inode *inode,
4142                              int for_rename)
4143 {
4144         /*
4145          * when we're logging a file, if it hasn't been renamed
4146          * or unlinked, and its inode is fully committed on disk,
4147          * we don't have to worry about walking up the directory chain
4148          * to log its parents.
4149          *
4150          * So, we use the last_unlink_trans field to put this transid
4151          * into the file.  When the file is logged we check it and
4152          * don't log the parents if the file is fully on disk.
4153          */
4154         if (S_ISREG(inode->i_mode))
4155                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4156
4157         /*
4158          * if this directory was already logged any new
4159          * names for this file/dir will get recorded
4160          */
4161         smp_mb();
4162         if (BTRFS_I(dir)->logged_trans == trans->transid)
4163                 return;
4164
4165         /*
4166          * if the inode we're about to unlink was logged,
4167          * the log will be properly updated for any new names
4168          */
4169         if (BTRFS_I(inode)->logged_trans == trans->transid)
4170                 return;
4171
4172         /*
4173          * when renaming files across directories, if the directory
4174          * there we're unlinking from gets fsync'd later on, there's
4175          * no way to find the destination directory later and fsync it
4176          * properly.  So, we have to be conservative and force commits
4177          * so the new name gets discovered.
4178          */
4179         if (for_rename)
4180                 goto record;
4181
4182         /* we can safely do the unlink without any special recording */
4183         return;
4184
4185 record:
4186         BTRFS_I(dir)->last_unlink_trans = trans->transid;
4187 }
4188
4189 /*
4190  * Call this after adding a new name for a file and it will properly
4191  * update the log to reflect the new name.
4192  *
4193  * It will return zero if all goes well, and it will return 1 if a
4194  * full transaction commit is required.
4195  */
4196 int btrfs_log_new_name(struct btrfs_trans_handle *trans,
4197                         struct inode *inode, struct inode *old_dir,
4198                         struct dentry *parent)
4199 {
4200         struct btrfs_root * root = BTRFS_I(inode)->root;
4201
4202         /*
4203          * this will force the logging code to walk the dentry chain
4204          * up for the file
4205          */
4206         if (S_ISREG(inode->i_mode))
4207                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4208
4209         /*
4210          * if this inode hasn't been logged and directory we're renaming it
4211          * from hasn't been logged, we don't need to log it
4212          */
4213         if (BTRFS_I(inode)->logged_trans <=
4214             root->fs_info->last_trans_committed &&
4215             (!old_dir || BTRFS_I(old_dir)->logged_trans <=
4216                     root->fs_info->last_trans_committed))
4217                 return 0;
4218
4219         return btrfs_log_inode_parent(trans, root, inode, parent, 1);
4220 }
4221