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/blkdev.h>
  22 #include <linux/list_sort.h>
  23 #include "ctree.h"
  24 #include "transaction.h"
  25 #include "disk-io.h"
  26 #include "locking.h"
  27 #include "print-tree.h"
  28 #include "backref.h"
  29 #include "compat.h"
  30 #include "tree-log.h"
  31 #include "hash.h"
  32
  33 /* magic values for the inode_only field in btrfs_log_inode:
  34  *
  35  * LOG_INODE_ALL means to log everything
  36  * LOG_INODE_EXISTS means to log just enough to recreate the inode
  37  * during log replay
  38  */
  39 #define LOG_INODE_ALL 0
  40 #define LOG_INODE_EXISTS 1
  41
  42 /*
  43  * directory trouble cases
  44  *
  45  * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
  46  * log, we must force a full commit before doing an fsync of the directory
  47  * where the unlink was done.
  48  * ---> record transid of last unlink/rename per directory
  49  *
  50  * mkdir foo/some_dir
  51  * normal commit
  52  * rename foo/some_dir foo2/some_dir
  53  * mkdir foo/some_dir
  54  * fsync foo/some_dir/some_file
  55  *
  56  * The fsync above will unlink the original some_dir without recording
  57  * it in its new location (foo2).  After a crash, some_dir will be gone
  58  * unless the fsync of some_file forces a full commit
  59  *
  60  * 2) we must log any new names for any file or dir that is in the fsync
  61  * log. ---> check inode while renaming/linking.
  62  *
  63  * 2a) we must log any new names for any file or dir during rename
  64  * when the directory they are being removed from was logged.
  65  * ---> check inode and old parent dir during rename
  66  *
  67  *  2a is actually the more important variant.  With the extra logging
  68  *  a crash might unlink the old name without recreating the new one
  69  *
  70  * 3) after a crash, we must go through any directories with a link count
  71  * of zero and redo the rm -rf
  72  *
  73  * mkdir f1/foo
  74  * normal commit
  75  * rm -rf f1/foo
  76  * fsync(f1)
  77  *
  78  * The directory f1 was fully removed from the FS, but fsync was never
  79  * called on f1, only its parent dir.  After a crash the rm -rf must
  80  * be replayed.  This must be able to recurse down the entire
  81  * directory tree.  The inode link count fixup code takes care of the
  82  * ugly details.
  83  */
  84
  85 /*
  86  * stages for the tree walking.  The first
  87  * stage (0) is to only pin down the blocks we find
  88  * the second stage (1) is to make sure that all the inodes
  89  * we find in the log are created in the subvolume.
  90  *
  91  * The last stage is to deal with directories and links and extents
  92  * and all the other fun semantics
  93  */
  94 #define LOG_WALK_PIN_ONLY 0
  95 #define LOG_WALK_REPLAY_INODES 1
  96 #define LOG_WALK_REPLAY_DIR_INDEX 2
  97 #define LOG_WALK_REPLAY_ALL 3
  98
  99 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
 100                              struct btrfs_root *root, struct inode *inode,
 101                              int inode_only);
 102 static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
 103                              struct btrfs_root *root,
 104                              struct btrfs_path *path, u64 objectid);
 105 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
 106                                        struct btrfs_root *root,
 107                                        struct btrfs_root *log,
 108                                        struct btrfs_path *path,
 109                                        u64 dirid, int del_all);
 110
 111 /*
 112  * tree logging is a special write ahead log used to make sure that
 113  * fsyncs and O_SYNCs can happen without doing full tree commits.
 114  *
 115  * Full tree commits are expensive because they require commonly
 116  * modified blocks to be recowed, creating many dirty pages in the
 117  * extent tree an 4x-6x higher write load than ext3.
 118  *
 119  * Instead of doing a tree commit on every fsync, we use the
 120  * key ranges and transaction ids to find items for a given file or directory
 121  * that have changed in this transaction.  Those items are copied into
 122  * a special tree (one per subvolume root), that tree is written to disk
 123  * and then the fsync is considered complete.
 124  *
 125  * After a crash, items are copied out of the log-tree back into the
 126  * subvolume tree.  Any file data extents found are recorded in the extent
 127  * allocation tree, and the log-tree freed.
 128  *
 129  * The log tree is read three times, once to pin down all the extents it is
 130  * using in ram and once, once to create all the inodes logged in the tree
 131  * and once to do all the other items.
 132  */
 133
 134 /*
 135  * start a sub transaction and setup the log tree
 136  * this increments the log tree writer count to make the people
 137  * syncing the tree wait for us to finish
 138  */
 139 static int start_log_trans(struct btrfs_trans_handle *trans,
 140                            struct btrfs_root *root)
 141 {
 142         int ret;
 143         int err = 0;
 144
 145         mutex_lock(&root->log_mutex);
 146         if (root->log_root) {
 147                 if (!root->log_start_pid) {
 148                         root->log_start_pid = current->pid;
 149                         root->log_multiple_pids = false;
 150                 } else if (root->log_start_pid != current->pid) {
 151                         root->log_multiple_pids = true;
 152                 }
 153
 154                 atomic_inc(&root->log_batch);
 155                 atomic_inc(&root->log_writers);
 156                 mutex_unlock(&root->log_mutex);
 157                 return 0;
 158         }
 159         root->log_multiple_pids = false;
 160         root->log_start_pid = current->pid;
 161         mutex_lock(&root->fs_info->tree_log_mutex);
 162         if (!root->fs_info->log_root_tree) {
 163                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
 164                 if (ret)
 165                         err = ret;
 166         }
 167         if (err == 0 && !root->log_root) {
 168                 ret = btrfs_add_log_tree(trans, root);
 169                 if (ret)
 170                         err = ret;
 171         }
 172         mutex_unlock(&root->fs_info->tree_log_mutex);
 173         atomic_inc(&root->log_batch);
 174         atomic_inc(&root->log_writers);
 175         mutex_unlock(&root->log_mutex);
 176         return err;
 177 }
 178
 179 /*
 180  * returns 0 if there was a log transaction running and we were able
 181  * to join, or returns -ENOENT if there were not transactions
 182  * in progress
 183  */
 184 static int join_running_log_trans(struct btrfs_root *root)
 185 {
 186         int ret = -ENOENT;
 187
 188         smp_mb();
 189         if (!root->log_root)
 190                 return -ENOENT;
 191
 192         mutex_lock(&root->log_mutex);
 193         if (root->log_root) {
 194                 ret = 0;
 195                 atomic_inc(&root->log_writers);
 196         }
 197         mutex_unlock(&root->log_mutex);
 198         return ret;
 199 }
 200
 201 /*
 202  * This either makes the current running log transaction wait
 203  * until you call btrfs_end_log_trans() or it makes any future
 204  * log transactions wait until you call btrfs_end_log_trans()
 205  */
 206 int btrfs_pin_log_trans(struct btrfs_root *root)
 207 {
 208         int ret = -ENOENT;
 209
 210         mutex_lock(&root->log_mutex);
 211         atomic_inc(&root->log_writers);
 212         mutex_unlock(&root->log_mutex);
 213         return ret;
 214 }
 215
 216 /*
 217  * indicate we're done making changes to the log tree
 218  * and wake up anyone waiting to do a sync
 219  */
 220 void btrfs_end_log_trans(struct btrfs_root *root)
 221 {
 222         if (atomic_dec_and_test(&root->log_writers)) {
 223                 smp_mb();
 224                 if (waitqueue_active(&root->log_writer_wait))
 225                         wake_up(&root->log_writer_wait);
 226         }
 227 }
 228
 229
 230 /*
 231  * the walk control struct is used to pass state down the chain when
 232  * processing the log tree.  The stage field tells us which part
 233  * of the log tree processing we are currently doing.  The others
 234  * are state fields used for that specific part
 235  */
 236 struct walk_control {
 237         /* should we free the extent on disk when done?  This is used
 238          * at transaction commit time while freeing a log tree
 239          */
 240         int free;
 241
 242         /* should we write out the extent buffer?  This is used
 243          * while flushing the log tree to disk during a sync
 244          */
 245         int write;
 246
 247         /* should we wait for the extent buffer io to finish?  Also used
 248          * while flushing the log tree to disk for a sync
 249          */
 250         int wait;
 251
 252         /* pin only walk, we record which extents on disk belong to the
 253          * log trees
 254          */
 255         int pin;
 256
 257         /* what stage of the replay code we're currently in */
 258         int stage;
 259
 260         /* the root we are currently replaying */
 261         struct btrfs_root *replay_dest;
 262
 263         /* the trans handle for the current replay */
 264         struct btrfs_trans_handle *trans;
 265
 266         /* the function that gets used to process blocks we find in the
 267          * tree.  Note the extent_buffer might not be up to date when it is
 268          * passed in, and it must be checked or read if you need the data
 269          * inside it
 270          */
 271         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
 272                             struct walk_control *wc, u64 gen);
 273 };
 274
 275 /*
 276  * process_func used to pin down extents, write them or wait on them
 277  */
 278 static int process_one_buffer(struct btrfs_root *log,
 279                               struct extent_buffer *eb,
 280                               struct walk_control *wc, u64 gen)
 281 {
 282         int ret = 0;
 283
 284         /*
 285          * If this fs is mixed then we need to be able to process the leaves to
 286          * pin down any logged extents, so we have to read the block.
 287          */
 288         if (btrfs_fs_incompat(log->fs_info, MIXED_GROUPS)) {
 289                 ret = btrfs_read_buffer(eb, gen);
 290                 if (ret)
 291                         return ret;
 292         }
 293
 294         if (wc->pin)
 295                 ret = btrfs_pin_extent_for_log_replay(log->fs_info->extent_root,
 296                                                       eb->start, eb->len);
 297
 298         if (!ret && btrfs_buffer_uptodate(eb, gen, 0)) {
 299                 if (wc->pin && btrfs_header_level(eb) == 0)
 300                         ret = btrfs_exclude_logged_extents(log, eb);
 301                 if (wc->write)
 302                         btrfs_write_tree_block(eb);
 303                 if (wc->wait)
 304                         btrfs_wait_tree_block_writeback(eb);
 305         }
 306         return ret;
 307 }
 308
 309 /*
 310  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
 311  * to the src data we are copying out.
 312  *
 313  * root is the tree we are copying into, and path is a scratch
 314  * path for use in this function (it should be released on entry and
 315  * will be released on exit).
 316  *
 317  * If the key is already in the destination tree the existing item is
 318  * overwritten.  If the existing item isn't big enough, it is extended.
 319  * If it is too large, it is truncated.
 320  *
 321  * If the key isn't in the destination yet, a new item is inserted.
 322  */
 323 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
 324                                    struct btrfs_root *root,
 325                                    struct btrfs_path *path,
 326                                    struct extent_buffer *eb, int slot,
 327                                    struct btrfs_key *key)
 328 {
 329         int ret;
 330         u32 item_size;
 331         u64 saved_i_size = 0;
 332         int save_old_i_size = 0;
 333         unsigned long src_ptr;
 334         unsigned long dst_ptr;
 335         int overwrite_root = 0;
 336         bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
 337
 338         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
 339                 overwrite_root = 1;
 340
 341         item_size = btrfs_item_size_nr(eb, slot);
 342         src_ptr = btrfs_item_ptr_offset(eb, slot);
 343
 344         /* look for the key in the destination tree */
 345         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
 346         if (ret < 0)
 347                 return ret;
 348
 349         if (ret == 0) {
 350                 char *src_copy;
 351                 char *dst_copy;
 352                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
 353                                                   path->slots[0]);
 354                 if (dst_size != item_size)
 355                         goto insert;
 356
 357                 if (item_size == 0) {
 358                         btrfs_release_path(path);
 359                         return 0;
 360                 }
 361                 dst_copy = kmalloc(item_size, GFP_NOFS);
 362                 src_copy = kmalloc(item_size, GFP_NOFS);
 363                 if (!dst_copy || !src_copy) {
 364                         btrfs_release_path(path);
 365                         kfree(dst_copy);
 366                         kfree(src_copy);
 367                         return -ENOMEM;
 368                 }
 369
 370                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
 371
 372                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 373                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
 374                                    item_size);
 375                 ret = memcmp(dst_copy, src_copy, item_size);
 376
 377                 kfree(dst_copy);
 378                 kfree(src_copy);
 379                 /*
 380                  * they have the same contents, just return, this saves
 381                  * us from cowing blocks in the destination tree and doing
 382                  * extra writes that may not have been done by a previous
 383                  * sync
 384                  */
 385                 if (ret == 0) {
 386                         btrfs_release_path(path);
 387                         return 0;
 388                 }
 389
 390                 /*
 391                  * We need to load the old nbytes into the inode so when we
 392                  * replay the extents we've logged we get the right nbytes.
 393                  */
 394                 if (inode_item) {
 395                         struct btrfs_inode_item *item;
 396                         u64 nbytes;
 397                         u32 mode;
 398
 399                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
 400                                               struct btrfs_inode_item);
 401                         nbytes = btrfs_inode_nbytes(path->nodes[0], item);
 402                         item = btrfs_item_ptr(eb, slot,
 403                                               struct btrfs_inode_item);
 404                         btrfs_set_inode_nbytes(eb, item, nbytes);
 405
 406                         /*
 407                          * If this is a directory we need to reset the i_size to
 408                          * 0 so that we can set it up properly when replaying
 409                          * the rest of the items in this log.
 410                          */
 411                         mode = btrfs_inode_mode(eb, item);
 412                         if (S_ISDIR(mode))
 413                                 btrfs_set_inode_size(eb, item, 0);
 414                 }
 415         } else if (inode_item) {
 416                 struct btrfs_inode_item *item;
 417                 u32 mode;
 418
 419                 /*
 420                  * New inode, set nbytes to 0 so that the nbytes comes out
 421                  * properly when we replay the extents.
 422                  */
 423                 item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
 424                 btrfs_set_inode_nbytes(eb, item, 0);
 425
 426                 /*
 427                  * If this is a directory we need to reset the i_size to 0 so
 428                  * that we can set it up properly when replaying the rest of
 429                  * the items in this log.
 430                  */
 431                 mode = btrfs_inode_mode(eb, item);
 432                 if (S_ISDIR(mode))
 433                         btrfs_set_inode_size(eb, item, 0);
 434         }
 435 insert:
 436         btrfs_release_path(path);
 437         /* try to insert the key into the destination tree */
 438         ret = btrfs_insert_empty_item(trans, root, path,
 439                                       key, item_size);
 440
 441         /* make sure any existing item is the correct size */
 442         if (ret == -EEXIST) {
 443                 u32 found_size;
 444                 found_size = btrfs_item_size_nr(path->nodes[0],
 445                                                 path->slots[0]);
 446                 if (found_size > item_size)
 447                         btrfs_truncate_item(root, path, item_size, 1);
 448                 else if (found_size < item_size)
 449                         btrfs_extend_item(root, path,
 450                                           item_size - found_size);
 451         } else if (ret) {
 452                 return ret;
 453         }
 454         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
 455                                         path->slots[0]);
 456
 457         /* don't overwrite an existing inode if the generation number
 458          * was logged as zero.  This is done when the tree logging code
 459          * is just logging an inode to make sure it exists after recovery.
 460          *
 461          * Also, don't overwrite i_size on directories during replay.
 462          * log replay inserts and removes directory items based on the
 463          * state of the tree found in the subvolume, and i_size is modified
 464          * as it goes
 465          */
 466         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
 467                 struct btrfs_inode_item *src_item;
 468                 struct btrfs_inode_item *dst_item;
 469
 470                 src_item = (struct btrfs_inode_item *)src_ptr;
 471                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 472
 473                 if (btrfs_inode_generation(eb, src_item) == 0)
 474                         goto no_copy;
 475
 476                 if (overwrite_root &&
 477                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
 478                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
 479                         save_old_i_size = 1;
 480                         saved_i_size = btrfs_inode_size(path->nodes[0],
 481                                                         dst_item);
 482                 }
 483         }
 484
 485         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
 486                            src_ptr, item_size);
 487
 488         if (save_old_i_size) {
 489                 struct btrfs_inode_item *dst_item;
 490                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 491                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
 492         }
 493
 494         /* make sure the generation is filled in */
 495         if (key->type == BTRFS_INODE_ITEM_KEY) {
 496                 struct btrfs_inode_item *dst_item;
 497                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 498                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
 499                         btrfs_set_inode_generation(path->nodes[0], dst_item,
 500                                                    trans->transid);
 501                 }
 502         }
 503 no_copy:
 504         btrfs_mark_buffer_dirty(path->nodes[0]);
 505         btrfs_release_path(path);
 506         return 0;
 507 }
 508
 509 /*
 510  * simple helper to read an inode off the disk from a given root
 511  * This can only be called for subvolume roots and not for the log
 512  */
 513 static noinline struct inode *read_one_inode(struct btrfs_root *root,
 514                                              u64 objectid)
 515 {
 516         struct btrfs_key key;
 517         struct inode *inode;
 518
 519         key.objectid = objectid;
 520         key.type = BTRFS_INODE_ITEM_KEY;
 521         key.offset = 0;
 522         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
 523         if (IS_ERR(inode)) {
 524                 inode = NULL;
 525         } else if (is_bad_inode(inode)) {
 526                 iput(inode);
 527                 inode = NULL;
 528         }
 529         return inode;
 530 }
 531
 532 /* replays a single extent in 'eb' at 'slot' with 'key' into the
 533  * subvolume 'root'.  path is released on entry and should be released
 534  * on exit.
 535  *
 536  * extents in the log tree have not been allocated out of the extent
 537  * tree yet.  So, this completes the allocation, taking a reference
 538  * as required if the extent already exists or creating a new extent
 539  * if it isn't in the extent allocation tree yet.
 540  *
 541  * The extent is inserted into the file, dropping any existing extents
 542  * from the file that overlap the new one.
 543  */
 544 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
 545                                       struct btrfs_root *root,
 546                                       struct btrfs_path *path,
 547                                       struct extent_buffer *eb, int slot,
 548                                       struct btrfs_key *key)
 549 {
 550         int found_type;
 551         u64 extent_end;
 552         u64 start = key->offset;
 553         u64 nbytes = 0;
 554         struct btrfs_file_extent_item *item;
 555         struct inode *inode = NULL;
 556         unsigned long size;
 557         int ret = 0;
 558
 559         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 560         found_type = btrfs_file_extent_type(eb, item);
 561
 562         if (found_type == BTRFS_FILE_EXTENT_REG ||
 563             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 564                 nbytes = btrfs_file_extent_num_bytes(eb, item);
 565                 extent_end = start + nbytes;
 566
 567                 /*
 568                  * We don't add to the inodes nbytes if we are prealloc or a
 569                  * hole.
 570                  */
 571                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
 572                         nbytes = 0;
 573         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 574                 size = btrfs_file_extent_inline_len(eb, item);
 575                 nbytes = btrfs_file_extent_ram_bytes(eb, item);
 576                 extent_end = ALIGN(start + size, root->sectorsize);
 577         } else {
 578                 ret = 0;
 579                 goto out;
 580         }
 581
 582         inode = read_one_inode(root, key->objectid);
 583         if (!inode) {
 584                 ret = -EIO;
 585                 goto out;
 586         }
 587
 588         /*
 589          * first check to see if we already have this extent in the
 590          * file.  This must be done before the btrfs_drop_extents run
 591          * so we don't try to drop this extent.
 592          */
 593         ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
 594                                        start, 0);
 595
 596         if (ret == 0 &&
 597             (found_type == BTRFS_FILE_EXTENT_REG ||
 598              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
 599                 struct btrfs_file_extent_item cmp1;
 600                 struct btrfs_file_extent_item cmp2;
 601                 struct btrfs_file_extent_item *existing;
 602                 struct extent_buffer *leaf;
 603
 604                 leaf = path->nodes[0];
 605                 existing = btrfs_item_ptr(leaf, path->slots[0],
 606                                           struct btrfs_file_extent_item);
 607
 608                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
 609                                    sizeof(cmp1));
 610                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
 611                                    sizeof(cmp2));
 612
 613                 /*
 614                  * we already have a pointer to this exact extent,
 615                  * we don't have to do anything
 616                  */
 617                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
 618                         btrfs_release_path(path);
 619                         goto out;
 620                 }
 621         }
 622         btrfs_release_path(path);
 623
 624         /* drop any overlapping extents */
 625         ret = btrfs_drop_extents(trans, root, inode, start, extent_end, 1);
 626         if (ret)
 627                 goto out;
 628
 629         if (found_type == BTRFS_FILE_EXTENT_REG ||
 630             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 631                 u64 offset;
 632                 unsigned long dest_offset;
 633                 struct btrfs_key ins;
 634
 635                 ret = btrfs_insert_empty_item(trans, root, path, key,
 636                                               sizeof(*item));
 637                 if (ret)
 638                         goto out;
 639                 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
 640                                                     path->slots[0]);
 641                 copy_extent_buffer(path->nodes[0], eb, dest_offset,
 642                                 (unsigned long)item,  sizeof(*item));
 643
 644                 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
 645                 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
 646                 ins.type = BTRFS_EXTENT_ITEM_KEY;
 647                 offset = key->offset - btrfs_file_extent_offset(eb, item);
 648
 649                 if (ins.objectid > 0) {
 650                         u64 csum_start;
 651                         u64 csum_end;
 652                         LIST_HEAD(ordered_sums);
 653                         /*
 654                          * is this extent already allocated in the extent
 655                          * allocation tree?  If so, just add a reference
 656                          */
 657                         ret = btrfs_lookup_extent(root, ins.objectid,
 658                                                 ins.offset);
 659                         if (ret == 0) {
 660                                 ret = btrfs_inc_extent_ref(trans, root,
 661                                                 ins.objectid, ins.offset,
 662                                                 0, root->root_key.objectid,
 663                                                 key->objectid, offset, 0);
 664                                 if (ret)
 665                                         goto out;
 666                         } else {
 667                                 /*
 668                                  * insert the extent pointer in the extent
 669                                  * allocation tree
 670                                  */
 671                                 ret = btrfs_alloc_logged_file_extent(trans,
 672                                                 root, root->root_key.objectid,
 673                                                 key->objectid, offset, &ins);
 674                                 if (ret)
 675                                         goto out;
 676                         }
 677                         btrfs_release_path(path);
 678
 679                         if (btrfs_file_extent_compression(eb, item)) {
 680                                 csum_start = ins.objectid;
 681                                 csum_end = csum_start + ins.offset;
 682                         } else {
 683                                 csum_start = ins.objectid +
 684                                         btrfs_file_extent_offset(eb, item);
 685                                 csum_end = csum_start +
 686                                         btrfs_file_extent_num_bytes(eb, item);
 687                         }
 688
 689                         ret = btrfs_lookup_csums_range(root->log_root,
 690                                                 csum_start, csum_end - 1,
 691                                                 &ordered_sums, 0);
 692                         if (ret)
 693                                 goto out;
 694                         while (!list_empty(&ordered_sums)) {
 695                                 struct btrfs_ordered_sum *sums;
 696                                 sums = list_entry(ordered_sums.next,
 697                                                 struct btrfs_ordered_sum,
 698                                                 list);
 699                                 if (!ret)
 700                                         ret = btrfs_csum_file_blocks(trans,
 701                                                 root->fs_info->csum_root,
 702                                                 sums);
 703                                 list_del(&sums->list);
 704                                 kfree(sums);
 705                         }
 706                         if (ret)
 707                                 goto out;
 708                 } else {
 709                         btrfs_release_path(path);
 710                 }
 711         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 712                 /* inline extents are easy, we just overwrite them */
 713                 ret = overwrite_item(trans, root, path, eb, slot, key);
 714                 if (ret)
 715                         goto out;
 716         }
 717
 718         inode_add_bytes(inode, nbytes);
 719         ret = btrfs_update_inode(trans, root, inode);
 720 out:
 721         if (inode)
 722                 iput(inode);
 723         return ret;
 724 }
 725
 726 /*
 727  * when cleaning up conflicts between the directory names in the
 728  * subvolume, directory names in the log and directory names in the
 729  * inode back references, we may have to unlink inodes from directories.
 730  *
 731  * This is a helper function to do the unlink of a specific directory
 732  * item
 733  */
 734 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
 735                                       struct btrfs_root *root,
 736                                       struct btrfs_path *path,
 737                                       struct inode *dir,
 738                                       struct btrfs_dir_item *di)
 739 {
 740         struct inode *inode;
 741         char *name;
 742         int name_len;
 743         struct extent_buffer *leaf;
 744         struct btrfs_key location;
 745         int ret;
 746
 747         leaf = path->nodes[0];
 748
 749         btrfs_dir_item_key_to_cpu(leaf, di, &location);
 750         name_len = btrfs_dir_name_len(leaf, di);
 751         name = kmalloc(name_len, GFP_NOFS);
 752         if (!name)
 753                 return -ENOMEM;
 754
 755         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
 756         btrfs_release_path(path);
 757
 758         inode = read_one_inode(root, location.objectid);
 759         if (!inode) {
 760                 ret = -EIO;
 761                 goto out;
 762         }
 763
 764         ret = link_to_fixup_dir(trans, root, path, location.objectid);
 765         if (ret)
 766                 goto out;
 767
 768         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
 769         if (ret)
 770                 goto out;
 771         else
 772                 ret = btrfs_run_delayed_items(trans, root);
 773 out:
 774         kfree(name);
 775         iput(inode);
 776         return ret;
 777 }
 778
 779 /*
 780  * helper function to see if a given name and sequence number found
 781  * in an inode back reference are already in a directory and correctly
 782  * point to this inode
 783  */
 784 static noinline int inode_in_dir(struct btrfs_root *root,
 785                                  struct btrfs_path *path,
 786                                  u64 dirid, u64 objectid, u64 index,
 787                                  const char *name, int name_len)
 788 {
 789         struct btrfs_dir_item *di;
 790         struct btrfs_key location;
 791         int match = 0;
 792
 793         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
 794                                          index, name, name_len, 0);
 795         if (di && !IS_ERR(di)) {
 796                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 797                 if (location.objectid != objectid)
 798                         goto out;
 799         } else
 800                 goto out;
 801         btrfs_release_path(path);
 802
 803         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
 804         if (di && !IS_ERR(di)) {
 805                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 806                 if (location.objectid != objectid)
 807                         goto out;
 808         } else
 809                 goto out;
 810         match = 1;
 811 out:
 812         btrfs_release_path(path);
 813         return match;
 814 }
 815
 816 /*
 817  * helper function to check a log tree for a named back reference in
 818  * an inode.  This is used to decide if a back reference that is
 819  * found in the subvolume conflicts with what we find in the log.
 820  *
 821  * inode backreferences may have multiple refs in a single item,
 822  * during replay we process one reference at a time, and we don't
 823  * want to delete valid links to a file from the subvolume if that
 824  * link is also in the log.
 825  */
 826 static noinline int backref_in_log(struct btrfs_root *log,
 827                                    struct btrfs_key *key,
 828                                    u64 ref_objectid,
 829                                    char *name, int namelen)
 830 {
 831         struct btrfs_path *path;
 832         struct btrfs_inode_ref *ref;
 833         unsigned long ptr;
 834         unsigned long ptr_end;
 835         unsigned long name_ptr;
 836         int found_name_len;
 837         int item_size;
 838         int ret;
 839         int match = 0;
 840
 841         path = btrfs_alloc_path();
 842         if (!path)
 843                 return -ENOMEM;
 844
 845         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
 846         if (ret != 0)
 847                 goto out;
 848
 849         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 850
 851         if (key->type == BTRFS_INODE_EXTREF_KEY) {
 852                 if (btrfs_find_name_in_ext_backref(path, ref_objectid,
 853                                                    name, namelen, NULL))
 854                         match = 1;
 855
 856                 goto out;
 857         }
 858
 859         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
 860         ptr_end = ptr + item_size;
 861         while (ptr < ptr_end) {
 862                 ref = (struct btrfs_inode_ref *)ptr;
 863                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
 864                 if (found_name_len == namelen) {
 865                         name_ptr = (unsigned long)(ref + 1);
 866                         ret = memcmp_extent_buffer(path->nodes[0], name,
 867                                                    name_ptr, namelen);
 868                         if (ret == 0) {
 869                                 match = 1;
 870                                 goto out;
 871                         }
 872                 }
 873                 ptr = (unsigned long)(ref + 1) + found_name_len;
 874         }
 875 out:
 876         btrfs_free_path(path);
 877         return match;
 878 }
 879
 880 static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
 881                                   struct btrfs_root *root,
 882                                   struct btrfs_path *path,
 883                                   struct btrfs_root *log_root,
 884                                   struct inode *dir, struct inode *inode,
 885                                   struct extent_buffer *eb,
 886                                   u64 inode_objectid, u64 parent_objectid,
 887                                   u64 ref_index, char *name, int namelen,
 888                                   int *search_done)
 889 {
 890         int ret;
 891         char *victim_name;
 892         int victim_name_len;
 893         struct extent_buffer *leaf;
 894         struct btrfs_dir_item *di;
 895         struct btrfs_key search_key;
 896         struct btrfs_inode_extref *extref;
 897
 898 again:
 899         /* Search old style refs */
 900         search_key.objectid = inode_objectid;
 901         search_key.type = BTRFS_INODE_REF_KEY;
 902         search_key.offset = parent_objectid;
 903         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 904         if (ret == 0) {
 905                 struct btrfs_inode_ref *victim_ref;
 906                 unsigned long ptr;
 907                 unsigned long ptr_end;
 908
 909                 leaf = path->nodes[0];
 910
 911                 /* are we trying to overwrite a back ref for the root directory
 912                  * if so, just jump out, we're done
 913                  */
 914                 if (search_key.objectid == search_key.offset)
 915                         return 1;
 916
 917                 /* check all the names in this back reference to see
 918                  * if they are in the log.  if so, we allow them to stay
 919                  * otherwise they must be unlinked as a conflict
 920                  */
 921                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 922                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
 923                 while (ptr < ptr_end) {
 924                         victim_ref = (struct btrfs_inode_ref *)ptr;
 925                         victim_name_len = btrfs_inode_ref_name_len(leaf,
 926                                                                    victim_ref);
 927                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 928                         if (!victim_name)
 929                                 return -ENOMEM;
 930
 931                         read_extent_buffer(leaf, victim_name,
 932                                            (unsigned long)(victim_ref + 1),
 933                                            victim_name_len);
 934
 935                         if (!backref_in_log(log_root, &search_key,
 936                                             parent_objectid,
 937                                             victim_name,
 938                                             victim_name_len)) {
 939                                 btrfs_inc_nlink(inode);
 940                                 btrfs_release_path(path);
 941
 942                                 ret = btrfs_unlink_inode(trans, root, dir,
 943                                                          inode, victim_name,
 944                                                          victim_name_len);
 945                                 kfree(victim_name);
 946                                 if (ret)
 947                                         return ret;
 948                                 ret = btrfs_run_delayed_items(trans, root);
 949                                 if (ret)
 950                                         return ret;
 951                                 *search_done = 1;
 952                                 goto again;
 953                         }
 954                         kfree(victim_name);
 955
 956                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
 957                 }
 958
 959                 /*
 960                  * NOTE: we have searched root tree and checked the
 961                  * coresponding ref, it does not need to check again.
 962                  */
 963                 *search_done = 1;
 964         }
 965         btrfs_release_path(path);
 966
 967         /* Same search but for extended refs */
 968         extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
 969                                            inode_objectid, parent_objectid, 0,
 970                                            0);
 971         if (!IS_ERR_OR_NULL(extref)) {
 972                 u32 item_size;
 973                 u32 cur_offset = 0;
 974                 unsigned long base;
 975                 struct inode *victim_parent;
 976
 977                 leaf = path->nodes[0];
 978
 979                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 980                 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
 981
 982                 while (cur_offset < item_size) {
 983                         extref = (struct btrfs_inode_extref *)base + cur_offset;
 984
 985                         victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
 986
 987                         if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
 988                                 goto next;
 989
 990                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 991                         if (!victim_name)
 992                                 return -ENOMEM;
 993                         read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
 994                                            victim_name_len);
 995
 996                         search_key.objectid = inode_objectid;
 997                         search_key.type = BTRFS_INODE_EXTREF_KEY;
 998                         search_key.offset = btrfs_extref_hash(parent_objectid,
 999                                                               victim_name,
1000                                                               victim_name_len);
1001                         ret = 0;
1002                         if (!backref_in_log(log_root, &search_key,
1003                                             parent_objectid, victim_name,
1004                                             victim_name_len)) {
1005                                 ret = -ENOENT;
1006                                 victim_parent = read_one_inode(root,
1007                                                                parent_objectid);
1008                                 if (victim_parent) {
1009                                         btrfs_inc_nlink(inode);
1010                                         btrfs_release_path(path);
1011
1012                                         ret = btrfs_unlink_inode(trans, root,
1013                                                                  victim_parent,
1014                                                                  inode,
1015                                                                  victim_name,
1016                                                                  victim_name_len);
1017                                         if (!ret)
1018                                                 ret = btrfs_run_delayed_items(
1019                                                                   trans, root);
1020                                 }
1021                                 iput(victim_parent);
1022                                 kfree(victim_name);
1023                                 if (ret)
1024                                         return ret;
1025                                 *search_done = 1;
1026                                 goto again;
1027                         }
1028                         kfree(victim_name);
1029                         if (ret)
1030                                 return ret;
1031 next:
1032                         cur_offset += victim_name_len + sizeof(*extref);
1033                 }
1034                 *search_done = 1;
1035         }
1036         btrfs_release_path(path);
1037
1038         /* look for a conflicting sequence number */
1039         di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
1040                                          ref_index, name, namelen, 0);
1041         if (di && !IS_ERR(di)) {
1042                 ret = drop_one_dir_item(trans, root, path, dir, di);
1043                 if (ret)
1044                         return ret;
1045         }
1046         btrfs_release_path(path);
1047
1048         /* look for a conflicing name */
1049         di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
1050                                    name, namelen, 0);
1051         if (di && !IS_ERR(di)) {
1052                 ret = drop_one_dir_item(trans, root, path, dir, di);
1053                 if (ret)
1054                         return ret;
1055         }
1056         btrfs_release_path(path);
1057
1058         return 0;
1059 }
1060
1061 static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1062                              u32 *namelen, char **name, u64 *index,
1063                              u64 *parent_objectid)
1064 {
1065         struct btrfs_inode_extref *extref;
1066
1067         extref = (struct btrfs_inode_extref *)ref_ptr;
1068
1069         *namelen = btrfs_inode_extref_name_len(eb, extref);
1070         *name = kmalloc(*namelen, GFP_NOFS);
1071         if (*name == NULL)
1072                 return -ENOMEM;
1073
1074         read_extent_buffer(eb, *name, (unsigned long)&extref->name,
1075                            *namelen);
1076
1077         *index = btrfs_inode_extref_index(eb, extref);
1078         if (parent_objectid)
1079                 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
1080
1081         return 0;
1082 }
1083
1084 static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1085                           u32 *namelen, char **name, u64 *index)
1086 {
1087         struct btrfs_inode_ref *ref;
1088
1089         ref = (struct btrfs_inode_ref *)ref_ptr;
1090
1091         *namelen = btrfs_inode_ref_name_len(eb, ref);
1092         *name = kmalloc(*namelen, GFP_NOFS);
1093         if (*name == NULL)
1094                 return -ENOMEM;
1095
1096         read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1097
1098         *index = btrfs_inode_ref_index(eb, ref);
1099
1100         return 0;
1101 }
1102
1103 /*
1104  * replay one inode back reference item found in the log tree.
1105  * eb, slot and key refer to the buffer and key found in the log tree.
1106  * root is the destination we are replaying into, and path is for temp
1107  * use by this function.  (it should be released on return).
1108  */
1109 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1110                                   struct btrfs_root *root,
1111                                   struct btrfs_root *log,
1112                                   struct btrfs_path *path,
1113                                   struct extent_buffer *eb, int slot,
1114                                   struct btrfs_key *key)
1115 {
1116         struct inode *dir = NULL;
1117         struct inode *inode = NULL;
1118         unsigned long ref_ptr;
1119         unsigned long ref_end;
1120         char *name = NULL;
1121         int namelen;
1122         int ret;
1123         int search_done = 0;
1124         int log_ref_ver = 0;
1125         u64 parent_objectid;
1126         u64 inode_objectid;
1127         u64 ref_index = 0;
1128         int ref_struct_size;
1129
1130         ref_ptr = btrfs_item_ptr_offset(eb, slot);
1131         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1132
1133         if (key->type == BTRFS_INODE_EXTREF_KEY) {
1134                 struct btrfs_inode_extref *r;
1135
1136                 ref_struct_size = sizeof(struct btrfs_inode_extref);
1137                 log_ref_ver = 1;
1138                 r = (struct btrfs_inode_extref *)ref_ptr;
1139                 parent_objectid = btrfs_inode_extref_parent(eb, r);
1140         } else {
1141                 ref_struct_size = sizeof(struct btrfs_inode_ref);
1142                 parent_objectid = key->offset;
1143         }
1144         inode_objectid = key->objectid;
1145
1146         /*
1147          * it is possible that we didn't log all the parent directories
1148          * for a given inode.  If we don't find the dir, just don't
1149          * copy the back ref in.  The link count fixup code will take
1150          * care of the rest
1151          */
1152         dir = read_one_inode(root, parent_objectid);
1153         if (!dir) {
1154                 ret = -ENOENT;
1155                 goto out;
1156         }
1157
1158         inode = read_one_inode(root, inode_objectid);
1159         if (!inode) {
1160                 ret = -EIO;
1161                 goto out;
1162         }
1163
1164         while (ref_ptr < ref_end) {
1165                 if (log_ref_ver) {
1166                         ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1167                                                 &ref_index, &parent_objectid);
1168                         /*
1169                          * parent object can change from one array
1170                          * item to another.
1171                          */
1172                         if (!dir)
1173                                 dir = read_one_inode(root, parent_objectid);
1174                         if (!dir) {
1175                                 ret = -ENOENT;
1176                                 goto out;
1177                         }
1178                 } else {
1179                         ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1180                                              &ref_index);
1181                 }
1182                 if (ret)
1183                         goto out;
1184
1185                 /* if we already have a perfect match, we're done */
1186                 if (!inode_in_dir(root, path, btrfs_ino(dir), btrfs_ino(inode),
1187                                   ref_index, name, namelen)) {
1188                         /*
1189                          * look for a conflicting back reference in the
1190                          * metadata. if we find one we have to unlink that name
1191                          * of the file before we add our new link.  Later on, we
1192                          * overwrite any existing back reference, and we don't
1193                          * want to create dangling pointers in the directory.
1194                          */
1195
1196                         if (!search_done) {
1197                                 ret = __add_inode_ref(trans, root, path, log,
1198                                                       dir, inode, eb,
1199                                                       inode_objectid,
1200                                                       parent_objectid,
1201                                                       ref_index, name, namelen,
1202                                                       &search_done);
1203                                 if (ret) {
1204                                         if (ret == 1)
1205                                                 ret = 0;
1206                                         goto out;
1207                                 }
1208                         }
1209
1210                         /* insert our name */
1211                         ret = btrfs_add_link(trans, dir, inode, name, namelen,
1212                                              0, ref_index);
1213                         if (ret)
1214                                 goto out;
1215
1216                         btrfs_update_inode(trans, root, inode);
1217                 }
1218
1219                 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1220                 kfree(name);
1221                 name = NULL;
1222                 if (log_ref_ver) {
1223                         iput(dir);
1224                         dir = NULL;
1225                 }
1226         }
1227
1228         /* finally write the back reference in the inode */
1229         ret = overwrite_item(trans, root, path, eb, slot, key);
1230 out:
1231         btrfs_release_path(path);
1232         kfree(name);
1233         iput(dir);
1234         iput(inode);
1235         return ret;
1236 }
1237
1238 static int insert_orphan_item(struct btrfs_trans_handle *trans,
1239                               struct btrfs_root *root, u64 offset)
1240 {
1241         int ret;
1242         ret = btrfs_find_orphan_item(root, offset);
1243         if (ret > 0)
1244                 ret = btrfs_insert_orphan_item(trans, root, offset);
1245         return ret;
1246 }
1247
1248 static int count_inode_extrefs(struct btrfs_root *root,
1249                                struct inode *inode, struct btrfs_path *path)
1250 {
1251         int ret = 0;
1252         int name_len;
1253         unsigned int nlink = 0;
1254         u32 item_size;
1255         u32 cur_offset = 0;
1256         u64 inode_objectid = btrfs_ino(inode);
1257         u64 offset = 0;
1258         unsigned long ptr;
1259         struct btrfs_inode_extref *extref;
1260         struct extent_buffer *leaf;
1261
1262         while (1) {
1263                 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1264                                             &extref, &offset);
1265                 if (ret)
1266                         break;
1267
1268                 leaf = path->nodes[0];
1269                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1270                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1271
1272                 while (cur_offset < item_size) {
1273                         extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1274                         name_len = btrfs_inode_extref_name_len(leaf, extref);
1275
1276                         nlink++;
1277
1278                         cur_offset += name_len + sizeof(*extref);
1279                 }
1280
1281                 offset++;
1282                 btrfs_release_path(path);
1283         }
1284         btrfs_release_path(path);
1285
1286         if (ret < 0)
1287                 return ret;
1288         return nlink;
1289 }
1290
1291 static int count_inode_refs(struct btrfs_root *root,
1292                                struct inode *inode, struct btrfs_path *path)
1293 {
1294         int ret;
1295         struct btrfs_key key;
1296         unsigned int nlink = 0;
1297         unsigned long ptr;
1298         unsigned long ptr_end;
1299         int name_len;
1300         u64 ino = btrfs_ino(inode);
1301
1302         key.objectid = ino;
1303         key.type = BTRFS_INODE_REF_KEY;
1304         key.offset = (u64)-1;
1305
1306         while (1) {
1307                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1308                 if (ret < 0)
1309                         break;
1310                 if (ret > 0) {
1311                         if (path->slots[0] == 0)
1312                                 break;
1313                         path->slots[0]--;
1314                 }
1315                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1316                                       path->slots[0]);
1317                 if (key.objectid != ino ||
1318                     key.type != BTRFS_INODE_REF_KEY)
1319                         break;
1320                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1321                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1322                                                    path->slots[0]);
1323                 while (ptr < ptr_end) {
1324                         struct btrfs_inode_ref *ref;
1325
1326                         ref = (struct btrfs_inode_ref *)ptr;
1327                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1328                                                             ref);
1329                         ptr = (unsigned long)(ref + 1) + name_len;
1330                         nlink++;
1331                 }
1332
1333                 if (key.offset == 0)
1334                         break;
1335                 key.offset--;
1336                 btrfs_release_path(path);
1337         }
1338         btrfs_release_path(path);
1339
1340         return nlink;
1341 }
1342
1343 /*
1344  * There are a few corners where the link count of the file can't
1345  * be properly maintained during replay.  So, instead of adding
1346  * lots of complexity to the log code, we just scan the backrefs
1347  * for any file that has been through replay.
1348  *
1349  * The scan will update the link count on the inode to reflect the
1350  * number of back refs found.  If it goes down to zero, the iput
1351  * will free the inode.
1352  */
1353 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1354                                            struct btrfs_root *root,
1355                                            struct inode *inode)
1356 {
1357         struct btrfs_path *path;
1358         int ret;
1359         u64 nlink = 0;
1360         u64 ino = btrfs_ino(inode);
1361
1362         path = btrfs_alloc_path();
1363         if (!path)
1364                 return -ENOMEM;
1365
1366         ret = count_inode_refs(root, inode, path);
1367         if (ret < 0)
1368                 goto out;
1369
1370         nlink = ret;
1371
1372         ret = count_inode_extrefs(root, inode, path);
1373         if (ret == -ENOENT)
1374                 ret = 0;
1375
1376         if (ret < 0)
1377                 goto out;
1378
1379         nlink += ret;
1380
1381         ret = 0;
1382
1383         if (nlink != inode->i_nlink) {
1384                 set_nlink(inode, nlink);
1385                 btrfs_update_inode(trans, root, inode);
1386         }
1387         BTRFS_I(inode)->index_cnt = (u64)-1;
1388
1389         if (inode->i_nlink == 0) {
1390                 if (S_ISDIR(inode->i_mode)) {
1391                         ret = replay_dir_deletes(trans, root, NULL, path,
1392                                                  ino, 1);
1393                         if (ret)
1394                                 goto out;
1395                 }
1396                 ret = insert_orphan_item(trans, root, ino);
1397         }
1398
1399 out:
1400         btrfs_free_path(path);
1401         return ret;
1402 }
1403
1404 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1405                                             struct btrfs_root *root,
1406                                             struct btrfs_path *path)
1407 {
1408         int ret;
1409         struct btrfs_key key;
1410         struct inode *inode;
1411
1412         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1413         key.type = BTRFS_ORPHAN_ITEM_KEY;
1414         key.offset = (u64)-1;
1415         while (1) {
1416                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1417                 if (ret < 0)
1418                         break;
1419
1420                 if (ret == 1) {
1421                         if (path->slots[0] == 0)
1422                                 break;
1423                         path->slots[0]--;
1424                 }
1425
1426                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1427                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1428                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1429                         break;
1430
1431                 ret = btrfs_del_item(trans, root, path);
1432                 if (ret)
1433                         goto out;
1434
1435                 btrfs_release_path(path);
1436                 inode = read_one_inode(root, key.offset);
1437                 if (!inode)
1438                         return -EIO;
1439
1440                 ret = fixup_inode_link_count(trans, root, inode);
1441                 iput(inode);
1442                 if (ret)
1443                         goto out;
1444
1445                 /*
1446                  * fixup on a directory may create new entries,
1447                  * make sure we always look for the highset possible
1448                  * offset
1449                  */
1450                 key.offset = (u64)-1;
1451         }
1452         ret = 0;
1453 out:
1454         btrfs_release_path(path);
1455         return ret;
1456 }
1457
1458
1459 /*
1460  * record a given inode in the fixup dir so we can check its link
1461  * count when replay is done.  The link count is incremented here
1462  * so the inode won't go away until we check it
1463  */
1464 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1465                                       struct btrfs_root *root,
1466                                       struct btrfs_path *path,
1467                                       u64 objectid)
1468 {
1469         struct btrfs_key key;
1470         int ret = 0;
1471         struct inode *inode;
1472
1473         inode = read_one_inode(root, objectid);
1474         if (!inode)
1475                 return -EIO;
1476
1477         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1478         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1479         key.offset = objectid;
1480
1481         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1482
1483         btrfs_release_path(path);
1484         if (ret == 0) {
1485                 if (!inode->i_nlink)
1486                         set_nlink(inode, 1);
1487                 else
1488                         btrfs_inc_nlink(inode);
1489                 ret = btrfs_update_inode(trans, root, inode);
1490         } else if (ret == -EEXIST) {
1491                 ret = 0;
1492         } else {
1493                 BUG(); /* Logic Error */
1494         }
1495         iput(inode);
1496
1497         return ret;
1498 }
1499
1500 /*
1501  * when replaying the log for a directory, we only insert names
1502  * for inodes that actually exist.  This means an fsync on a directory
1503  * does not implicitly fsync all the new files in it
1504  */
1505 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1506                                     struct btrfs_root *root,
1507                                     struct btrfs_path *path,
1508                                     u64 dirid, u64 index,
1509                                     char *name, int name_len, u8 type,
1510                                     struct btrfs_key *location)
1511 {
1512         struct inode *inode;
1513         struct inode *dir;
1514         int ret;
1515
1516         inode = read_one_inode(root, location->objectid);
1517         if (!inode)
1518                 return -ENOENT;
1519
1520         dir = read_one_inode(root, dirid);
1521         if (!dir) {
1522                 iput(inode);
1523                 return -EIO;
1524         }
1525
1526         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1527
1528         /* FIXME, put inode into FIXUP list */
1529
1530         iput(inode);
1531         iput(dir);
1532         return ret;
1533 }
1534
1535 /*
1536  * take a single entry in a log directory item and replay it into
1537  * the subvolume.
1538  *
1539  * if a conflicting item exists in the subdirectory already,
1540  * the inode it points to is unlinked and put into the link count
1541  * fix up tree.
1542  *
1543  * If a name from the log points to a file or directory that does
1544  * not exist in the FS, it is skipped.  fsyncs on directories
1545  * do not force down inodes inside that directory, just changes to the
1546  * names or unlinks in a directory.
1547  */
1548 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1549                                     struct btrfs_root *root,
1550                                     struct btrfs_path *path,
1551                                     struct extent_buffer *eb,
1552                                     struct btrfs_dir_item *di,
1553                                     struct btrfs_key *key)
1554 {
1555         char *name;
1556         int name_len;
1557         struct btrfs_dir_item *dst_di;
1558         struct btrfs_key found_key;
1559         struct btrfs_key log_key;
1560         struct inode *dir;
1561         u8 log_type;
1562         int exists;
1563         int ret = 0;
1564         bool update_size = (key->type == BTRFS_DIR_INDEX_KEY);
1565
1566         dir = read_one_inode(root, key->objectid);
1567         if (!dir)
1568                 return -EIO;
1569
1570         name_len = btrfs_dir_name_len(eb, di);
1571         name = kmalloc(name_len, GFP_NOFS);
1572         if (!name) {
1573                 ret = -ENOMEM;
1574                 goto out;
1575         }
1576
1577         log_type = btrfs_dir_type(eb, di);
1578         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1579                    name_len);
1580
1581         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1582         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1583         if (exists == 0)
1584                 exists = 1;
1585         else
1586                 exists = 0;
1587         btrfs_release_path(path);
1588
1589         if (key->type == BTRFS_DIR_ITEM_KEY) {
1590                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1591                                        name, name_len, 1);
1592         } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1593                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1594                                                      key->objectid,
1595                                                      key->offset, name,
1596                                                      name_len, 1);
1597         } else {
1598                 /* Corruption */
1599                 ret = -EINVAL;
1600                 goto out;
1601         }
1602         if (IS_ERR_OR_NULL(dst_di)) {
1603                 /* we need a sequence number to insert, so we only
1604                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1605                  */
1606                 if (key->type != BTRFS_DIR_INDEX_KEY)
1607                         goto out;
1608                 goto insert;
1609         }
1610
1611         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1612         /* the existing item matches the logged item */
1613         if (found_key.objectid == log_key.objectid &&
1614             found_key.type == log_key.type &&
1615             found_key.offset == log_key.offset &&
1616             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1617                 goto out;
1618         }
1619
1620         /*
1621          * don't drop the conflicting directory entry if the inode
1622          * for the new entry doesn't exist
1623          */
1624         if (!exists)
1625                 goto out;
1626
1627         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1628         if (ret)
1629                 goto out;
1630
1631         if (key->type == BTRFS_DIR_INDEX_KEY)
1632                 goto insert;
1633 out:
1634         btrfs_release_path(path);
1635         if (!ret && update_size) {
1636                 btrfs_i_size_write(dir, dir->i_size + name_len * 2);
1637                 ret = btrfs_update_inode(trans, root, dir);
1638         }
1639         kfree(name);
1640         iput(dir);
1641         return ret;
1642
1643 insert:
1644         btrfs_release_path(path);
1645         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1646                               name, name_len, log_type, &log_key);
1647         if (ret && ret != -ENOENT)
1648                 goto out;
1649         update_size = false;
1650         ret = 0;
1651         goto out;
1652 }
1653
1654 /*
1655  * find all the names in a directory item and reconcile them into
1656  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1657  * one name in a directory item, but the same code gets used for
1658  * both directory index types
1659  */
1660 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1661                                         struct btrfs_root *root,
1662                                         struct btrfs_path *path,
1663                                         struct extent_buffer *eb, int slot,
1664                                         struct btrfs_key *key)
1665 {
1666         int ret;
1667         u32 item_size = btrfs_item_size_nr(eb, slot);
1668         struct btrfs_dir_item *di;
1669         int name_len;
1670         unsigned long ptr;
1671         unsigned long ptr_end;
1672
1673         ptr = btrfs_item_ptr_offset(eb, slot);
1674         ptr_end = ptr + item_size;
1675         while (ptr < ptr_end) {
1676                 di = (struct btrfs_dir_item *)ptr;
1677                 if (verify_dir_item(root, eb, di))
1678                         return -EIO;
1679                 name_len = btrfs_dir_name_len(eb, di);
1680                 ret = replay_one_name(trans, root, path, eb, di, key);
1681                 if (ret)
1682                         return ret;
1683                 ptr = (unsigned long)(di + 1);
1684                 ptr += name_len;
1685         }
1686         return 0;
1687 }
1688
1689 /*
1690  * directory replay has two parts.  There are the standard directory
1691  * items in the log copied from the subvolume, and range items
1692  * created in the log while the subvolume was logged.
1693  *
1694  * The range items tell us which parts of the key space the log
1695  * is authoritative for.  During replay, if a key in the subvolume
1696  * directory is in a logged range item, but not actually in the log
1697  * that means it was deleted from the directory before the fsync
1698  * and should be removed.
1699  */
1700 static noinline int find_dir_range(struct btrfs_root *root,
1701                                    struct btrfs_path *path,
1702                                    u64 dirid, int key_type,
1703                                    u64 *start_ret, u64 *end_ret)
1704 {
1705         struct btrfs_key key;
1706         u64 found_end;
1707         struct btrfs_dir_log_item *item;
1708         int ret;
1709         int nritems;
1710
1711         if (*start_ret == (u64)-1)
1712                 return 1;
1713
1714         key.objectid = dirid;
1715         key.type = key_type;
1716         key.offset = *start_ret;
1717
1718         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1719         if (ret < 0)
1720                 goto out;
1721         if (ret > 0) {
1722                 if (path->slots[0] == 0)
1723                         goto out;
1724                 path->slots[0]--;
1725         }
1726         if (ret != 0)
1727                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1728
1729         if (key.type != key_type || key.objectid != dirid) {
1730                 ret = 1;
1731                 goto next;
1732         }
1733         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1734                               struct btrfs_dir_log_item);
1735         found_end = btrfs_dir_log_end(path->nodes[0], item);
1736
1737         if (*start_ret >= key.offset && *start_ret <= found_end) {
1738                 ret = 0;
1739                 *start_ret = key.offset;
1740                 *end_ret = found_end;
1741                 goto out;
1742         }
1743         ret = 1;
1744 next:
1745         /* check the next slot in the tree to see if it is a valid item */
1746         nritems = btrfs_header_nritems(path->nodes[0]);
1747         if (path->slots[0] >= nritems) {
1748                 ret = btrfs_next_leaf(root, path);
1749                 if (ret)
1750                         goto out;
1751         } else {
1752                 path->slots[0]++;
1753         }
1754
1755         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1756
1757         if (key.type != key_type || key.objectid != dirid) {
1758                 ret = 1;
1759                 goto out;
1760         }
1761         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1762                               struct btrfs_dir_log_item);
1763         found_end = btrfs_dir_log_end(path->nodes[0], item);
1764         *start_ret = key.offset;
1765         *end_ret = found_end;
1766         ret = 0;
1767 out:
1768         btrfs_release_path(path);
1769         return ret;
1770 }
1771
1772 /*
1773  * this looks for a given directory item in the log.  If the directory
1774  * item is not in the log, the item is removed and the inode it points
1775  * to is unlinked
1776  */
1777 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1778                                       struct btrfs_root *root,
1779                                       struct btrfs_root *log,
1780                                       struct btrfs_path *path,
1781                                       struct btrfs_path *log_path,
1782                                       struct inode *dir,
1783                                       struct btrfs_key *dir_key)
1784 {
1785         int ret;
1786         struct extent_buffer *eb;
1787         int slot;
1788         u32 item_size;
1789         struct btrfs_dir_item *di;
1790         struct btrfs_dir_item *log_di;
1791         int name_len;
1792         unsigned long ptr;
1793         unsigned long ptr_end;
1794         char *name;
1795         struct inode *inode;
1796         struct btrfs_key location;
1797
1798 again:
1799         eb = path->nodes[0];
1800         slot = path->slots[0];
1801         item_size = btrfs_item_size_nr(eb, slot);
1802         ptr = btrfs_item_ptr_offset(eb, slot);
1803         ptr_end = ptr + item_size;
1804         while (ptr < ptr_end) {
1805                 di = (struct btrfs_dir_item *)ptr;
1806                 if (verify_dir_item(root, eb, di)) {
1807                         ret = -EIO;
1808                         goto out;
1809                 }
1810
1811                 name_len = btrfs_dir_name_len(eb, di);
1812                 name = kmalloc(name_len, GFP_NOFS);
1813                 if (!name) {
1814                         ret = -ENOMEM;
1815                         goto out;
1816                 }
1817                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1818                                   name_len);
1819                 log_di = NULL;
1820                 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
1821                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1822                                                        dir_key->objectid,
1823                                                        name, name_len, 0);
1824                 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
1825                         log_di = btrfs_lookup_dir_index_item(trans, log,
1826                                                      log_path,
1827                                                      dir_key->objectid,
1828                                                      dir_key->offset,
1829                                                      name, name_len, 0);
1830                 }
1831                 if (IS_ERR_OR_NULL(log_di)) {
1832                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1833                         btrfs_release_path(path);
1834                         btrfs_release_path(log_path);
1835                         inode = read_one_inode(root, location.objectid);
1836                         if (!inode) {
1837                                 kfree(name);
1838                                 return -EIO;
1839                         }
1840
1841                         ret = link_to_fixup_dir(trans, root,
1842                                                 path, location.objectid);
1843                         if (ret) {
1844                                 kfree(name);
1845                                 iput(inode);
1846                                 goto out;
1847                         }
1848
1849                         btrfs_inc_nlink(inode);
1850                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1851                                                  name, name_len);
1852                         if (!ret)
1853                                 ret = btrfs_run_delayed_items(trans, root);
1854                         kfree(name);
1855                         iput(inode);
1856                         if (ret)
1857                                 goto out;
1858
1859                         /* there might still be more names under this key
1860                          * check and repeat if required
1861                          */
1862                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1863                                                 0, 0);
1864                         if (ret == 0)
1865                                 goto again;
1866                         ret = 0;
1867                         goto out;
1868                 }
1869                 btrfs_release_path(log_path);
1870                 kfree(name);
1871
1872                 ptr = (unsigned long)(di + 1);
1873                 ptr += name_len;
1874         }
1875         ret = 0;
1876 out:
1877         btrfs_release_path(path);
1878         btrfs_release_path(log_path);
1879         return ret;
1880 }
1881
1882 /*
1883  * deletion replay happens before we copy any new directory items
1884  * out of the log or out of backreferences from inodes.  It
1885  * scans the log to find ranges of keys that log is authoritative for,
1886  * and then scans the directory to find items in those ranges that are
1887  * not present in the log.
1888  *
1889  * Anything we don't find in the log is unlinked and removed from the
1890  * directory.
1891  */
1892 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1893                                        struct btrfs_root *root,
1894                                        struct btrfs_root *log,
1895                                        struct btrfs_path *path,
1896                                        u64 dirid, int del_all)
1897 {
1898         u64 range_start;
1899         u64 range_end;
1900         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1901         int ret = 0;
1902         struct btrfs_key dir_key;
1903         struct btrfs_key found_key;
1904         struct btrfs_path *log_path;
1905         struct inode *dir;
1906
1907         dir_key.objectid = dirid;
1908         dir_key.type = BTRFS_DIR_ITEM_KEY;
1909         log_path = btrfs_alloc_path();
1910         if (!log_path)
1911                 return -ENOMEM;
1912
1913         dir = read_one_inode(root, dirid);
1914         /* it isn't an error if the inode isn't there, that can happen
1915          * because we replay the deletes before we copy in the inode item
1916          * from the log
1917          */
1918         if (!dir) {
1919                 btrfs_free_path(log_path);
1920                 return 0;
1921         }
1922 again:
1923         range_start = 0;
1924         range_end = 0;
1925         while (1) {
1926                 if (del_all)
1927                         range_end = (u64)-1;
1928                 else {
1929                         ret = find_dir_range(log, path, dirid, key_type,
1930                                              &range_start, &range_end);
1931                         if (ret != 0)
1932                                 break;
1933                 }
1934
1935                 dir_key.offset = range_start;
1936                 while (1) {
1937                         int nritems;
1938                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
1939                                                 0, 0);
1940                         if (ret < 0)
1941                                 goto out;
1942
1943                         nritems = btrfs_header_nritems(path->nodes[0]);
1944                         if (path->slots[0] >= nritems) {
1945                                 ret = btrfs_next_leaf(root, path);
1946                                 if (ret)
1947                                         break;
1948                         }
1949                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1950                                               path->slots[0]);
1951                         if (found_key.objectid != dirid ||
1952                             found_key.type != dir_key.type)
1953                                 goto next_type;
1954
1955                         if (found_key.offset > range_end)
1956                                 break;
1957
1958                         ret = check_item_in_log(trans, root, log, path,
1959                                                 log_path, dir,
1960                                                 &found_key);
1961                         if (ret)
1962                                 goto out;
1963                         if (found_key.offset == (u64)-1)
1964                                 break;
1965                         dir_key.offset = found_key.offset + 1;
1966                 }
1967                 btrfs_release_path(path);
1968                 if (range_end == (u64)-1)
1969                         break;
1970                 range_start = range_end + 1;
1971         }
1972
1973 next_type:
1974         ret = 0;
1975         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1976                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1977                 dir_key.type = BTRFS_DIR_INDEX_KEY;
1978                 btrfs_release_path(path);
1979                 goto again;
1980         }
1981 out:
1982         btrfs_release_path(path);
1983         btrfs_free_path(log_path);
1984         iput(dir);
1985         return ret;
1986 }
1987
1988 /*
1989  * the process_func used to replay items from the log tree.  This
1990  * gets called in two different stages.  The first stage just looks
1991  * for inodes and makes sure they are all copied into the subvolume.
1992  *
1993  * The second stage copies all the other item types from the log into
1994  * the subvolume.  The two stage approach is slower, but gets rid of
1995  * lots of complexity around inodes referencing other inodes that exist
1996  * only in the log (references come from either directory items or inode
1997  * back refs).
1998  */
1999 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
2000                              struct walk_control *wc, u64 gen)
2001 {
2002         int nritems;
2003         struct btrfs_path *path;
2004         struct btrfs_root *root = wc->replay_dest;
2005         struct btrfs_key key;
2006         int level;
2007         int i;
2008         int ret;
2009
2010         ret = btrfs_read_buffer(eb, gen);
2011         if (ret)
2012                 return ret;
2013
2014         level = btrfs_header_level(eb);
2015
2016         if (level != 0)
2017                 return 0;
2018
2019         path = btrfs_alloc_path();
2020         if (!path)
2021                 return -ENOMEM;
2022
2023         nritems = btrfs_header_nritems(eb);
2024         for (i = 0; i < nritems; i++) {
2025                 btrfs_item_key_to_cpu(eb, &key, i);
2026
2027                 /* inode keys are done during the first stage */
2028                 if (key.type == BTRFS_INODE_ITEM_KEY &&
2029                     wc->stage == LOG_WALK_REPLAY_INODES) {
2030                         struct btrfs_inode_item *inode_item;
2031                         u32 mode;
2032
2033                         inode_item = btrfs_item_ptr(eb, i,
2034                                             struct btrfs_inode_item);
2035                         mode = btrfs_inode_mode(eb, inode_item);
2036                         if (S_ISDIR(mode)) {
2037                                 ret = replay_dir_deletes(wc->trans,
2038                                          root, log, path, key.objectid, 0);
2039                                 if (ret)
2040                                         break;
2041                         }
2042                         ret = overwrite_item(wc->trans, root, path,
2043                                              eb, i, &key);
2044                         if (ret)
2045                                 break;
2046
2047                         /* for regular files, make sure corresponding
2048                          * orhpan item exist. extents past the new EOF
2049                          * will be truncated later by orphan cleanup.
2050                          */
2051                         if (S_ISREG(mode)) {
2052                                 ret = insert_orphan_item(wc->trans, root,
2053                                                          key.objectid);
2054                                 if (ret)
2055                                         break;
2056                         }
2057
2058                         ret = link_to_fixup_dir(wc->trans, root,
2059                                                 path, key.objectid);
2060                         if (ret)
2061                                 break;
2062                 }
2063
2064                 if (key.type == BTRFS_DIR_INDEX_KEY &&
2065                     wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
2066                         ret = replay_one_dir_item(wc->trans, root, path,
2067                                                   eb, i, &key);
2068                         if (ret)
2069                                 break;
2070                 }
2071
2072                 if (wc->stage < LOG_WALK_REPLAY_ALL)
2073                         continue;
2074
2075                 /* these keys are simply copied */
2076                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
2077                         ret = overwrite_item(wc->trans, root, path,
2078                                              eb, i, &key);
2079                         if (ret)
2080                                 break;
2081                 } else if (key.type == BTRFS_INODE_REF_KEY ||
2082                            key.type == BTRFS_INODE_EXTREF_KEY) {
2083                         ret = add_inode_ref(wc->trans, root, log, path,
2084                                             eb, i, &key);
2085                         if (ret && ret != -ENOENT)
2086                                 break;
2087                         ret = 0;
2088                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
2089                         ret = replay_one_extent(wc->trans, root, path,
2090                                                 eb, i, &key);
2091                         if (ret)
2092                                 break;
2093                 } else if (key.type == BTRFS_DIR_ITEM_KEY) {
2094                         ret = replay_one_dir_item(wc->trans, root, path,
2095                                                   eb, i, &key);
2096                         if (ret)
2097                                 break;
2098                 }
2099         }
2100         btrfs_free_path(path);
2101         return ret;
2102 }
2103
2104 static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
2105                                    struct btrfs_root *root,
2106                                    struct btrfs_path *path, int *level,
2107                                    struct walk_control *wc)
2108 {
2109         u64 root_owner;
2110         u64 bytenr;
2111         u64 ptr_gen;
2112         struct extent_buffer *next;
2113         struct extent_buffer *cur;
2114         struct extent_buffer *parent;
2115         u32 blocksize;
2116         int ret = 0;
2117
2118         WARN_ON(*level < 0);
2119         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2120
2121         while (*level > 0) {
2122                 WARN_ON(*level < 0);
2123                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
2124                 cur = path->nodes[*level];
2125
2126                 if (btrfs_header_level(cur) != *level)
2127                         WARN_ON(1);
2128
2129                 if (path->slots[*level] >=
2130                     btrfs_header_nritems(cur))
2131                         break;
2132
2133                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
2134                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
2135                 blocksize = btrfs_level_size(root, *level - 1);
2136
2137                 parent = path->nodes[*level];
2138                 root_owner = btrfs_header_owner(parent);
2139
2140                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
2141                 if (!next)
2142                         return -ENOMEM;
2143
2144                 if (*level == 1) {
2145                         ret = wc->process_func(root, next, wc, ptr_gen);
2146                         if (ret) {
2147                                 free_extent_buffer(next);
2148                                 return ret;
2149                         }
2150
2151                         path->slots[*level]++;
2152                         if (wc->free) {
2153                                 ret = btrfs_read_buffer(next, ptr_gen);
2154                                 if (ret) {
2155                                         free_extent_buffer(next);
2156                                         return ret;
2157                                 }
2158
2159                                 if (trans) {
2160                                         btrfs_tree_lock(next);
2161                                         btrfs_set_lock_blocking(next);
2162                                         clean_tree_block(trans, root, next);
2163                                         btrfs_wait_tree_block_writeback(next);
2164                                         btrfs_tree_unlock(next);
2165                                 }
2166
2167                                 WARN_ON(root_owner !=
2168                                         BTRFS_TREE_LOG_OBJECTID);
2169                                 ret = btrfs_free_and_pin_reserved_extent(root,
2170                                                          bytenr, blocksize);
2171                                 if (ret) {
2172                                         free_extent_buffer(next);
2173                                         return ret;
2174                                 }
2175                         }
2176                         free_extent_buffer(next);
2177                         continue;
2178                 }
2179                 ret = btrfs_read_buffer(next, ptr_gen);
2180                 if (ret) {
2181                         free_extent_buffer(next);
2182                         return ret;
2183                 }
2184
2185                 WARN_ON(*level <= 0);
2186                 if (path->nodes[*level-1])
2187                         free_extent_buffer(path->nodes[*level-1]);
2188                 path->nodes[*level-1] = next;
2189                 *level = btrfs_header_level(next);
2190                 path->slots[*level] = 0;
2191                 cond_resched();
2192         }
2193         WARN_ON(*level < 0);
2194         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2195
2196         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2197
2198         cond_resched();
2199         return 0;
2200 }
2201
2202 static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2203                                  struct btrfs_root *root,
2204                                  struct btrfs_path *path, int *level,
2205                                  struct walk_control *wc)
2206 {
2207         u64 root_owner;
2208         int i;
2209         int slot;
2210         int ret;
2211
2212         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2213                 slot = path->slots[i];
2214                 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2215                         path->slots[i]++;
2216                         *level = i;
2217                         WARN_ON(*level == 0);
2218                         return 0;
2219                 } else {
2220                         struct extent_buffer *parent;
2221                         if (path->nodes[*level] == root->node)
2222                                 parent = path->nodes[*level];
2223                         else
2224                                 parent = path->nodes[*level + 1];
2225
2226                         root_owner = btrfs_header_owner(parent);
2227                         ret = wc->process_func(root, path->nodes[*level], wc,
2228                                  btrfs_header_generation(path->nodes[*level]));
2229                         if (ret)
2230                                 return ret;
2231
2232                         if (wc->free) {
2233                                 struct extent_buffer *next;
2234
2235                                 next = path->nodes[*level];
2236
2237                                 if (trans) {
2238                                         btrfs_tree_lock(next);
2239                                         btrfs_set_lock_blocking(next);
2240                                         clean_tree_block(trans, root, next);
2241                                         btrfs_wait_tree_block_writeback(next);
2242                                         btrfs_tree_unlock(next);
2243                                 }
2244
2245                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
2246                                 ret = btrfs_free_and_pin_reserved_extent(root,
2247                                                 path->nodes[*level]->start,
2248                                                 path->nodes[*level]->len);
2249                                 if (ret)
2250                                         return ret;
2251                         }
2252                         free_extent_buffer(path->nodes[*level]);
2253                         path->nodes[*level] = NULL;
2254                         *level = i + 1;
2255                 }
2256         }
2257         return 1;
2258 }
2259
2260 /*
2261  * drop the reference count on the tree rooted at 'snap'.  This traverses
2262  * the tree freeing any blocks that have a ref count of zero after being
2263  * decremented.
2264  */
2265 static int walk_log_tree(struct btrfs_trans_handle *trans,
2266                          struct btrfs_root *log, struct walk_control *wc)
2267 {
2268         int ret = 0;
2269         int wret;
2270         int level;
2271         struct btrfs_path *path;
2272         int orig_level;
2273
2274         path = btrfs_alloc_path();
2275         if (!path)
2276                 return -ENOMEM;
2277
2278         level = btrfs_header_level(log->node);
2279         orig_level = level;
2280         path->nodes[level] = log->node;
2281         extent_buffer_get(log->node);
2282         path->slots[level] = 0;
2283
2284         while (1) {
2285                 wret = walk_down_log_tree(trans, log, path, &level, wc);
2286                 if (wret > 0)
2287                         break;
2288                 if (wret < 0) {
2289                         ret = wret;
2290                         goto out;
2291                 }
2292
2293                 wret = walk_up_log_tree(trans, log, path, &level, wc);
2294                 if (wret > 0)
2295                         break;
2296                 if (wret < 0) {
2297                         ret = wret;
2298                         goto out;
2299                 }
2300         }
2301
2302         /* was the root node processed? if not, catch it here */
2303         if (path->nodes[orig_level]) {
2304                 ret = wc->process_func(log, path->nodes[orig_level], wc,
2305                          btrfs_header_generation(path->nodes[orig_level]));
2306                 if (ret)
2307                         goto out;
2308                 if (wc->free) {
2309                         struct extent_buffer *next;
2310
2311                         next = path->nodes[orig_level];
2312
2313                         if (trans) {
2314                                 btrfs_tree_lock(next);
2315                                 btrfs_set_lock_blocking(next);
2316                                 clean_tree_block(trans, log, next);
2317                                 btrfs_wait_tree_block_writeback(next);
2318                                 btrfs_tree_unlock(next);
2319                         }
2320
2321                         WARN_ON(log->root_key.objectid !=
2322                                 BTRFS_TREE_LOG_OBJECTID);
2323                         ret = btrfs_free_and_pin_reserved_extent(log, next->start,
2324                                                          next->len);
2325                         if (ret)
2326                                 goto out;
2327                 }
2328         }
2329
2330 out:
2331         btrfs_free_path(path);
2332         return ret;
2333 }
2334
2335 /*
2336  * helper function to update the item for a given subvolumes log root
2337  * in the tree of log roots
2338  */
2339 static int update_log_root(struct btrfs_trans_handle *trans,
2340                            struct btrfs_root *log)
2341 {
2342         int ret;
2343
2344         if (log->log_transid == 1) {
2345                 /* insert root item on the first sync */
2346                 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
2347                                 &log->root_key, &log->root_item);
2348         } else {
2349                 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2350                                 &log->root_key, &log->root_item);
2351         }
2352         return ret;
2353 }
2354
2355 static int wait_log_commit(struct btrfs_trans_handle *trans,
2356                            struct btrfs_root *root, unsigned long transid)
2357 {
2358         DEFINE_WAIT(wait);
2359         int index = transid % 2;
2360
2361         /*
2362          * we only allow two pending log transactions at a time,
2363          * so we know that if ours is more than 2 older than the
2364          * current transaction, we're done
2365          */
2366         do {
2367                 prepare_to_wait(&root->log_commit_wait[index],
2368                                 &wait, TASK_UNINTERRUPTIBLE);
2369                 mutex_unlock(&root->log_mutex);
2370
2371                 if (root->fs_info->last_trans_log_full_commit !=
2372                     trans->transid && root->log_transid < transid + 2 &&
2373                     atomic_read(&root->log_commit[index]))
2374                         schedule();
2375
2376                 finish_wait(&root->log_commit_wait[index], &wait);
2377                 mutex_lock(&root->log_mutex);
2378         } while (root->fs_info->last_trans_log_full_commit !=
2379                  trans->transid && root->log_transid < transid + 2 &&
2380                  atomic_read(&root->log_commit[index]));
2381         return 0;
2382 }
2383
2384 static void wait_for_writer(struct btrfs_trans_handle *trans,
2385                             struct btrfs_root *root)
2386 {
2387         DEFINE_WAIT(wait);
2388         while (root->fs_info->last_trans_log_full_commit !=
2389                trans->transid && atomic_read(&root->log_writers)) {
2390                 prepare_to_wait(&root->log_writer_wait,
2391                                 &wait, TASK_UNINTERRUPTIBLE);
2392                 mutex_unlock(&root->log_mutex);
2393                 if (root->fs_info->last_trans_log_full_commit !=
2394                     trans->transid && atomic_read(&root->log_writers))
2395                         schedule();
2396                 mutex_lock(&root->log_mutex);
2397                 finish_wait(&root->log_writer_wait, &wait);
2398         }
2399 }
2400
2401 /*
2402  * btrfs_sync_log does sends a given tree log down to the disk and
2403  * updates the super blocks to record it.  When this call is done,
2404  * you know that any inodes previously logged are safely on disk only
2405  * if it returns 0.
2406  *
2407  * Any other return value means you need to call btrfs_commit_transaction.
2408  * Some of the edge cases for fsyncing directories that have had unlinks
2409  * or renames done in the past mean that sometimes the only safe
2410  * fsync is to commit the whole FS.  When btrfs_sync_log returns -EAGAIN,
2411  * that has happened.
2412  */
2413 int btrfs_sync_log(struct btrfs_trans_handle *trans,
2414                    struct btrfs_root *root)
2415 {
2416         int index1;
2417         int index2;
2418         int mark;
2419         int ret;
2420         struct btrfs_root *log = root->log_root;
2421         struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
2422         unsigned long log_transid = 0;
2423         struct blk_plug plug;
2424
2425         mutex_lock(&root->log_mutex);
2426         log_transid = root->log_transid;
2427         index1 = root->log_transid % 2;
2428         if (atomic_read(&root->log_commit[index1])) {
2429                 wait_log_commit(trans, root, root->log_transid);
2430                 mutex_unlock(&root->log_mutex);
2431                 return 0;
2432         }
2433         atomic_set(&root->log_commit[index1], 1);
2434
2435         /* wait for previous tree log sync to complete */
2436         if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2437                 wait_log_commit(trans, root, root->log_transid - 1);
2438         while (1) {
2439                 int batch = atomic_read(&root->log_batch);
2440                 /* when we're on an ssd, just kick the log commit out */
2441                 if (!btrfs_test_opt(root, SSD) && root->log_multiple_pids) {
2442                         mutex_unlock(&root->log_mutex);
2443                         schedule_timeout_uninterruptible(1);
2444                         mutex_lock(&root->log_mutex);
2445                 }
2446                 wait_for_writer(trans, root);
2447                 if (batch == atomic_read(&root->log_batch))
2448                         break;
2449         }
2450
2451         /* bail out if we need to do a full commit */
2452         if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2453                 ret = -EAGAIN;
2454                 btrfs_free_logged_extents(log, log_transid);
2455                 mutex_unlock(&root->log_mutex);
2456                 goto out;
2457         }
2458
2459         if (log_transid % 2 == 0)
2460                 mark = EXTENT_DIRTY;
2461         else
2462                 mark = EXTENT_NEW;
2463
2464         /* we start IO on  all the marked extents here, but we don't actually
2465          * wait for them until later.
2466          */
2467         blk_start_plug(&plug);
2468         ret = btrfs_write_marked_extents(log, &log->dirty_log_pages, mark);
2469         if (ret) {
2470                 blk_finish_plug(&plug);
2471                 btrfs_abort_transaction(trans, root, ret);
2472                 btrfs_free_logged_extents(log, log_transid);
2473                 mutex_unlock(&root->log_mutex);
2474                 goto out;
2475         }
2476
2477         btrfs_set_root_node(&log->root_item, log->node);
2478
2479         root->log_transid++;
2480         log->log_transid = root->log_transid;
2481         root->log_start_pid = 0;
2482         smp_mb();
2483         /*
2484          * IO has been started, blocks of the log tree have WRITTEN flag set
2485          * in their headers. new modifications of the log will be written to
2486          * new positions. so it's safe to allow log writers to go in.
2487          */
2488         mutex_unlock(&root->log_mutex);
2489
2490         mutex_lock(&log_root_tree->log_mutex);
2491         atomic_inc(&log_root_tree->log_batch);
2492         atomic_inc(&log_root_tree->log_writers);
2493         mutex_unlock(&log_root_tree->log_mutex);
2494
2495         ret = update_log_root(trans, log);
2496
2497         mutex_lock(&log_root_tree->log_mutex);
2498         if (atomic_dec_and_test(&log_root_tree->log_writers)) {
2499                 smp_mb();
2500                 if (waitqueue_active(&log_root_tree->log_writer_wait))
2501                         wake_up(&log_root_tree->log_writer_wait);
2502         }
2503
2504         if (ret) {
2505                 blk_finish_plug(&plug);
2506                 if (ret != -ENOSPC) {
2507                         btrfs_abort_transaction(trans, root, ret);
2508                         mutex_unlock(&log_root_tree->log_mutex);
2509                         goto out;
2510                 }
2511                 root->fs_info->last_trans_log_full_commit = trans->transid;
2512                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2513                 btrfs_free_logged_extents(log, log_transid);
2514                 mutex_unlock(&log_root_tree->log_mutex);
2515                 ret = -EAGAIN;
2516                 goto out;
2517         }
2518
2519         index2 = log_root_tree->log_transid % 2;
2520         if (atomic_read(&log_root_tree->log_commit[index2])) {
2521                 blk_finish_plug(&plug);
2522                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2523                 wait_log_commit(trans, log_root_tree,
2524                                 log_root_tree->log_transid);
2525                 btrfs_free_logged_extents(log, log_transid);
2526                 mutex_unlock(&log_root_tree->log_mutex);
2527                 ret = 0;
2528                 goto out;
2529         }
2530         atomic_set(&log_root_tree->log_commit[index2], 1);
2531
2532         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2533                 wait_log_commit(trans, log_root_tree,
2534                                 log_root_tree->log_transid - 1);
2535         }
2536
2537         wait_for_writer(trans, log_root_tree);
2538
2539         /*
2540          * now that we've moved on to the tree of log tree roots,
2541          * check the full commit flag again
2542          */
2543         if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2544                 blk_finish_plug(&plug);
2545                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2546                 btrfs_free_logged_extents(log, log_transid);
2547                 mutex_unlock(&log_root_tree->log_mutex);
2548                 ret = -EAGAIN;
2549                 goto out_wake_log_root;
2550         }
2551
2552         ret = btrfs_write_marked_extents(log_root_tree,
2553                                          &log_root_tree->dirty_log_pages,
2554                                          EXTENT_DIRTY | EXTENT_NEW);
2555         blk_finish_plug(&plug);
2556         if (ret) {
2557                 btrfs_abort_transaction(trans, root, ret);
2558                 btrfs_free_logged_extents(log, log_transid);
2559                 mutex_unlock(&log_root_tree->log_mutex);
2560                 goto out_wake_log_root;
2561         }
2562         btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2563         btrfs_wait_marked_extents(log_root_tree,
2564                                   &log_root_tree->dirty_log_pages,
2565                                   EXTENT_NEW | EXTENT_DIRTY);
2566         btrfs_wait_logged_extents(log, log_transid);
2567
2568         btrfs_set_super_log_root(root->fs_info->super_for_commit,
2569                                 log_root_tree->node->start);
2570         btrfs_set_super_log_root_level(root->fs_info->super_for_commit,
2571                                 btrfs_header_level(log_root_tree->node));
2572
2573         log_root_tree->log_transid++;
2574         smp_mb();
2575
2576         mutex_unlock(&log_root_tree->log_mutex);
2577
2578         /*
2579          * nobody else is going to jump in and write the the ctree
2580          * super here because the log_commit atomic below is protecting
2581          * us.  We must be called with a transaction handle pinning
2582          * the running transaction open, so a full commit can't hop
2583          * in and cause problems either.
2584          */
2585         btrfs_scrub_pause_super(root);
2586         ret = write_ctree_super(trans, root->fs_info->tree_root, 1);
2587         btrfs_scrub_continue_super(root);
2588         if (ret) {
2589                 btrfs_abort_transaction(trans, root, ret);
2590                 goto out_wake_log_root;
2591         }
2592
2593         mutex_lock(&root->log_mutex);
2594         if (root->last_log_commit < log_transid)
2595                 root->last_log_commit = log_transid;
2596         mutex_unlock(&root->log_mutex);
2597
2598 out_wake_log_root:
2599         atomic_set(&log_root_tree->log_commit[index2], 0);
2600         smp_mb();
2601         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2602                 wake_up(&log_root_tree->log_commit_wait[index2]);
2603 out:
2604         atomic_set(&root->log_commit[index1], 0);
2605         smp_mb();
2606         if (waitqueue_active(&root->log_commit_wait[index1]))
2607                 wake_up(&root->log_commit_wait[index1]);
2608         return ret;
2609 }
2610
2611 static void free_log_tree(struct btrfs_trans_handle *trans,
2612                           struct btrfs_root *log)
2613 {
2614         int ret;
2615         u64 start;
2616         u64 end;
2617         struct walk_control wc = {
2618                 .free = 1,
2619                 .process_func = process_one_buffer
2620         };
2621
2622         ret = walk_log_tree(trans, log, &wc);
2623         /* I don't think this can happen but just in case */
2624         if (ret)
2625                 btrfs_abort_transaction(trans, log, ret);
2626
2627         while (1) {
2628                 ret = find_first_extent_bit(&log->dirty_log_pages,
2629                                 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW,
2630                                 NULL);
2631                 if (ret)
2632                         break;
2633
2634                 clear_extent_bits(&log->dirty_log_pages, start, end,
2635                                   EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2636         }
2637
2638         /*
2639          * We may have short-circuited the log tree with the full commit logic
2640          * and left ordered extents on our list, so clear these out to keep us
2641          * from leaking inodes and memory.
2642          */
2643         btrfs_free_logged_extents(log, 0);
2644         btrfs_free_logged_extents(log, 1);
2645
2646         free_extent_buffer(log->node);
2647         kfree(log);
2648 }
2649
2650 /*
2651  * free all the extents used by the tree log.  This should be called
2652  * at commit time of the full transaction
2653  */
2654 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2655 {
2656         if (root->log_root) {
2657                 free_log_tree(trans, root->log_root);
2658                 root->log_root = NULL;
2659         }
2660         return 0;
2661 }
2662
2663 int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2664                              struct btrfs_fs_info *fs_info)
2665 {
2666         if (fs_info->log_root_tree) {
2667                 free_log_tree(trans, fs_info->log_root_tree);
2668                 fs_info->log_root_tree = NULL;
2669         }
2670         return 0;
2671 }
2672
2673 /*
2674  * If both a file and directory are logged, and unlinks or renames are
2675  * mixed in, we have a few interesting corners:
2676  *
2677  * create file X in dir Y
2678  * link file X to X.link in dir Y
2679  * fsync file X
2680  * unlink file X but leave X.link
2681  * fsync dir Y
2682  *
2683  * After a crash we would expect only X.link to exist.  But file X
2684  * didn't get fsync'd again so the log has back refs for X and X.link.
2685  *
2686  * We solve this by removing directory entries and inode backrefs from the
2687  * log when a file that was logged in the current transaction is
2688  * unlinked.  Any later fsync will include the updated log entries, and
2689  * we'll be able to reconstruct the proper directory items from backrefs.
2690  *
2691  * This optimizations allows us to avoid relogging the entire inode
2692  * or the entire directory.
2693  */
2694 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2695                                  struct btrfs_root *root,
2696                                  const char *name, int name_len,
2697                                  struct inode *dir, u64 index)
2698 {
2699         struct btrfs_root *log;
2700         struct btrfs_dir_item *di;
2701         struct btrfs_path *path;
2702         int ret;
2703         int err = 0;
2704         int bytes_del = 0;
2705         u64 dir_ino = btrfs_ino(dir);
2706
2707         if (BTRFS_I(dir)->logged_trans < trans->transid)
2708                 return 0;
2709
2710         ret = join_running_log_trans(root);
2711         if (ret)
2712                 return 0;
2713
2714         mutex_lock(&BTRFS_I(dir)->log_mutex);
2715
2716         log = root->log_root;
2717         path = btrfs_alloc_path();
2718         if (!path) {
2719                 err = -ENOMEM;
2720                 goto out_unlock;
2721         }
2722
2723         di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2724                                    name, name_len, -1);
2725         if (IS_ERR(di)) {
2726                 err = PTR_ERR(di);
2727                 goto fail;
2728         }
2729         if (di) {
2730                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2731                 bytes_del += name_len;
2732                 if (ret) {
2733                         err = ret;
2734                         goto fail;
2735                 }
2736         }
2737         btrfs_release_path(path);
2738         di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2739                                          index, name, name_len, -1);
2740         if (IS_ERR(di)) {
2741                 err = PTR_ERR(di);
2742                 goto fail;
2743         }
2744         if (di) {
2745                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2746                 bytes_del += name_len;
2747                 if (ret) {
2748                         err = ret;
2749                         goto fail;
2750                 }
2751         }
2752
2753         /* update the directory size in the log to reflect the names
2754          * we have removed
2755          */
2756         if (bytes_del) {
2757                 struct btrfs_key key;
2758
2759                 key.objectid = dir_ino;
2760                 key.offset = 0;
2761                 key.type = BTRFS_INODE_ITEM_KEY;
2762                 btrfs_release_path(path);
2763
2764                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2765                 if (ret < 0) {
2766                         err = ret;
2767                         goto fail;
2768                 }
2769                 if (ret == 0) {
2770                         struct btrfs_inode_item *item;
2771                         u64 i_size;
2772
2773                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2774                                               struct btrfs_inode_item);
2775                         i_size = btrfs_inode_size(path->nodes[0], item);
2776                         if (i_size > bytes_del)
2777                                 i_size -= bytes_del;
2778                         else
2779                                 i_size = 0;
2780                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2781                         btrfs_mark_buffer_dirty(path->nodes[0]);
2782                 } else
2783                         ret = 0;
2784                 btrfs_release_path(path);
2785         }
2786 fail:
2787         btrfs_free_path(path);
2788 out_unlock:
2789         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2790         if (ret == -ENOSPC) {
2791                 root->fs_info->last_trans_log_full_commit = trans->transid;
2792                 ret = 0;
2793         } else if (ret < 0)
2794                 btrfs_abort_transaction(trans, root, ret);
2795
2796         btrfs_end_log_trans(root);
2797
2798         return err;
2799 }
2800
2801 /* see comments for btrfs_del_dir_entries_in_log */
2802 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2803                                struct btrfs_root *root,
2804                                const char *name, int name_len,
2805                                struct inode *inode, u64 dirid)
2806 {
2807         struct btrfs_root *log;
2808         u64 index;
2809         int ret;
2810
2811         if (BTRFS_I(inode)->logged_trans < trans->transid)
2812                 return 0;
2813
2814         ret = join_running_log_trans(root);
2815         if (ret)
2816                 return 0;
2817         log = root->log_root;
2818         mutex_lock(&BTRFS_I(inode)->log_mutex);
2819
2820         ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
2821                                   dirid, &index);
2822         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2823         if (ret == -ENOSPC) {
2824                 root->fs_info->last_trans_log_full_commit = trans->transid;
2825                 ret = 0;
2826         } else if (ret < 0 && ret != -ENOENT)
2827                 btrfs_abort_transaction(trans, root, ret);
2828         btrfs_end_log_trans(root);
2829
2830         return ret;
2831 }
2832
2833 /*
2834  * creates a range item in the log for 'dirid'.  first_offset and
2835  * last_offset tell us which parts of the key space the log should
2836  * be considered authoritative for.
2837  */
2838 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2839                                        struct btrfs_root *log,
2840                                        struct btrfs_path *path,
2841                                        int key_type, u64 dirid,
2842                                        u64 first_offset, u64 last_offset)
2843 {
2844         int ret;
2845         struct btrfs_key key;
2846         struct btrfs_dir_log_item *item;
2847
2848         key.objectid = dirid;
2849         key.offset = first_offset;
2850         if (key_type == BTRFS_DIR_ITEM_KEY)
2851                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2852         else
2853                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2854         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2855         if (ret)
2856                 return ret;
2857
2858         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2859                               struct btrfs_dir_log_item);
2860         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2861         btrfs_mark_buffer_dirty(path->nodes[0]);
2862         btrfs_release_path(path);
2863         return 0;
2864 }
2865
2866 /*
2867  * log all the items included in the current transaction for a given
2868  * directory.  This also creates the range items in the log tree required
2869  * to replay anything deleted before the fsync
2870  */
2871 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2872                           struct btrfs_root *root, struct inode *inode,
2873                           struct btrfs_path *path,
2874                           struct btrfs_path *dst_path, int key_type,
2875                           u64 min_offset, u64 *last_offset_ret)
2876 {
2877         struct btrfs_key min_key;
2878         struct btrfs_root *log = root->log_root;
2879         struct extent_buffer *src;
2880         int err = 0;
2881         int ret;
2882         int i;
2883         int nritems;
2884         u64 first_offset = min_offset;
2885         u64 last_offset = (u64)-1;
2886         u64 ino = btrfs_ino(inode);
2887
2888         log = root->log_root;
2889
2890         min_key.objectid = ino;
2891         min_key.type = key_type;
2892         min_key.offset = min_offset;
2893
2894         path->keep_locks = 1;
2895
2896         ret = btrfs_search_forward(root, &min_key, path, trans->transid);
2897
2898         /*
2899          * we didn't find anything from this transaction, see if there
2900          * is anything at all
2901          */
2902         if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
2903                 min_key.objectid = ino;
2904                 min_key.type = key_type;
2905                 min_key.offset = (u64)-1;
2906                 btrfs_release_path(path);
2907                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2908                 if (ret < 0) {
2909                         btrfs_release_path(path);
2910                         return ret;
2911                 }
2912                 ret = btrfs_previous_item(root, path, ino, key_type);
2913
2914                 /* if ret == 0 there are items for this type,
2915                  * create a range to tell us the last key of this type.
2916                  * otherwise, there are no items in this directory after
2917                  * *min_offset, and we create a range to indicate that.
2918                  */
2919                 if (ret == 0) {
2920                         struct btrfs_key tmp;
2921                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2922                                               path->slots[0]);
2923                         if (key_type == tmp.type)
2924                                 first_offset = max(min_offset, tmp.offset) + 1;
2925                 }
2926                 goto done;
2927         }
2928
2929         /* go backward to find any previous key */
2930         ret = btrfs_previous_item(root, path, ino, key_type);
2931         if (ret == 0) {
2932                 struct btrfs_key tmp;
2933                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2934                 if (key_type == tmp.type) {
2935                         first_offset = tmp.offset;
2936                         ret = overwrite_item(trans, log, dst_path,
2937                                              path->nodes[0], path->slots[0],
2938                                              &tmp);
2939                         if (ret) {
2940                                 err = ret;
2941                                 goto done;
2942                         }
2943                 }
2944         }
2945         btrfs_release_path(path);
2946
2947         /* find the first key from this transaction again */
2948         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2949         if (ret != 0) {
2950                 WARN_ON(1);
2951                 goto done;
2952         }
2953
2954         /*
2955          * we have a block from this transaction, log every item in it
2956          * from our directory
2957          */
2958         while (1) {
2959                 struct btrfs_key tmp;
2960                 src = path->nodes[0];
2961                 nritems = btrfs_header_nritems(src);
2962                 for (i = path->slots[0]; i < nritems; i++) {
2963                         btrfs_item_key_to_cpu(src, &min_key, i);
2964
2965                         if (min_key.objectid != ino || min_key.type != key_type)
2966                                 goto done;
2967                         ret = overwrite_item(trans, log, dst_path, src, i,
2968                                              &min_key);
2969                         if (ret) {
2970                                 err = ret;
2971                                 goto done;
2972                         }
2973                 }
2974                 path->slots[0] = nritems;
2975
2976                 /*
2977                  * look ahead to the next item and see if it is also
2978                  * from this directory and from this transaction
2979                  */
2980                 ret = btrfs_next_leaf(root, path);
2981                 if (ret == 1) {
2982                         last_offset = (u64)-1;
2983                         goto done;
2984                 }
2985                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2986                 if (tmp.objectid != ino || tmp.type != key_type) {
2987                         last_offset = (u64)-1;
2988                         goto done;
2989                 }
2990                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2991                         ret = overwrite_item(trans, log, dst_path,
2992                                              path->nodes[0], path->slots[0],
2993                                              &tmp);
2994                         if (ret)
2995                                 err = ret;
2996                         else
2997                                 last_offset = tmp.offset;
2998                         goto done;
2999                 }
3000         }
3001 done:
3002         btrfs_release_path(path);
3003         btrfs_release_path(dst_path);
3004
3005         if (err == 0) {
3006                 *last_offset_ret = last_offset;
3007                 /*
3008                  * insert the log range keys to indicate where the log
3009                  * is valid
3010                  */
3011                 ret = insert_dir_log_key(trans, log, path, key_type,
3012                                          ino, first_offset, last_offset);
3013                 if (ret)
3014                         err = ret;
3015         }
3016         return err;
3017 }
3018
3019 /*
3020  * logging directories is very similar to logging inodes, We find all the items
3021  * from the current transaction and write them to the log.
3022  *
3023  * The recovery code scans the directory in the subvolume, and if it finds a
3024  * key in the range logged that is not present in the log tree, then it means
3025  * that dir entry was unlinked during the transaction.
3026  *
3027  * In order for that scan to work, we must include one key smaller than
3028  * the smallest logged by this transaction and one key larger than the largest
3029  * key logged by this transaction.
3030  */
3031 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
3032                           struct btrfs_root *root, struct inode *inode,
3033                           struct btrfs_path *path,
3034                           struct btrfs_path *dst_path)
3035 {
3036         u64 min_key;
3037         u64 max_key;
3038         int ret;
3039         int key_type = BTRFS_DIR_ITEM_KEY;
3040
3041 again:
3042         min_key = 0;
3043         max_key = 0;
3044         while (1) {
3045                 ret = log_dir_items(trans, root, inode, path,
3046                                     dst_path, key_type, min_key,
3047                                     &max_key);
3048                 if (ret)
3049                         return ret;
3050                 if (max_key == (u64)-1)
3051                         break;
3052                 min_key = max_key + 1;
3053         }
3054
3055         if (key_type == BTRFS_DIR_ITEM_KEY) {
3056                 key_type = BTRFS_DIR_INDEX_KEY;
3057                 goto again;
3058         }
3059         return 0;
3060 }
3061
3062 /*
3063  * a helper function to drop items from the log before we relog an
3064  * inode.  max_key_type indicates the highest item type to remove.
3065  * This cannot be run for file data extents because it does not
3066  * free the extents they point to.
3067  */
3068 static int drop_objectid_items(struct btrfs_trans_handle *trans,
3069                                   struct btrfs_root *log,
3070                                   struct btrfs_path *path,
3071                                   u64 objectid, int max_key_type)
3072 {
3073         int ret;
3074         struct btrfs_key key;
3075         struct btrfs_key found_key;
3076         int start_slot;
3077
3078         key.objectid = objectid;
3079         key.type = max_key_type;
3080         key.offset = (u64)-1;
3081
3082         while (1) {
3083                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
3084                 BUG_ON(ret == 0); /* Logic error */
3085                 if (ret < 0)
3086                         break;
3087
3088                 if (path->slots[0] == 0)
3089                         break;
3090
3091                 path->slots[0]--;
3092                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
3093                                       path->slots[0]);
3094
3095                 if (found_key.objectid != objectid)
3096                         break;
3097
3098                 found_key.offset = 0;
3099                 found_key.type = 0;
3100                 ret = btrfs_bin_search(path->nodes[0], &found_key, 0,
3101                                        &start_slot);
3102
3103                 ret = btrfs_del_items(trans, log, path, start_slot,
3104                                       path->slots[0] - start_slot + 1);
3105                 /*
3106                  * If start slot isn't 0 then we don't need to re-search, we've
3107                  * found the last guy with the objectid in this tree.
3108                  */
3109                 if (ret || start_slot != 0)
3110                         break;
3111                 btrfs_release_path(path);
3112         }
3113         btrfs_release_path(path);
3114         if (ret > 0)
3115                 ret = 0;
3116         return ret;
3117 }
3118
3119 static void fill_inode_item(struct btrfs_trans_handle *trans,
3120                             struct extent_buffer *leaf,
3121                             struct btrfs_inode_item *item,
3122                             struct inode *inode, int log_inode_only)
3123 {
3124         struct btrfs_map_token token;
3125
3126         btrfs_init_map_token(&token);
3127
3128         if (log_inode_only) {
3129                 /* set the generation to zero so the recover code
3130                  * can tell the difference between an logging
3131                  * just to say 'this inode exists' and a logging
3132                  * to say 'update this inode with these values'
3133                  */
3134                 btrfs_set_token_inode_generation(leaf, item, 0, &token);
3135                 btrfs_set_token_inode_size(leaf, item, 0, &token);
3136         } else {
3137                 btrfs_set_token_inode_generation(leaf, item,
3138                                                  BTRFS_I(inode)->generation,
3139                                                  &token);
3140                 btrfs_set_token_inode_size(leaf, item, inode->i_size, &token);
3141         }
3142
3143         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3144         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3145         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3146         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3147
3148         btrfs_set_token_timespec_sec(leaf, btrfs_inode_atime(item),
3149                                      inode->i_atime.tv_sec, &token);
3150         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_atime(item),
3151                                       inode->i_atime.tv_nsec, &token);
3152
3153         btrfs_set_token_timespec_sec(leaf, btrfs_inode_mtime(item),
3154                                      inode->i_mtime.tv_sec, &token);
3155         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_mtime(item),
3156                                       inode->i_mtime.tv_nsec, &token);
3157
3158         btrfs_set_token_timespec_sec(leaf, btrfs_inode_ctime(item),
3159                                      inode->i_ctime.tv_sec, &token);
3160         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_ctime(item),
3161                                       inode->i_ctime.tv_nsec, &token);
3162
3163         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3164                                      &token);
3165
3166         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3167         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3168         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3169         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3170         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3171 }
3172
3173 static int log_inode_item(struct btrfs_trans_handle *trans,
3174                           struct btrfs_root *log, struct btrfs_path *path,
3175                           struct inode *inode)
3176 {
3177         struct btrfs_inode_item *inode_item;
3178         int ret;
3179
3180         ret = btrfs_insert_empty_item(trans, log, path,
3181                                       &BTRFS_I(inode)->location,
3182                                       sizeof(*inode_item));
3183         if (ret && ret != -EEXIST)
3184                 return ret;
3185         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3186                                     struct btrfs_inode_item);
3187         fill_inode_item(trans, path->nodes[0], inode_item, inode, 0);
3188         btrfs_release_path(path);
3189         return 0;
3190 }
3191
3192 static noinline int copy_items(struct btrfs_trans_handle *trans,
3193                                struct inode *inode,
3194                                struct btrfs_path *dst_path,
3195                                struct extent_buffer *src,
3196                                int start_slot, int nr, int inode_only)
3197 {
3198         unsigned long src_offset;
3199         unsigned long dst_offset;
3200         struct btrfs_root *log = BTRFS_I(inode)->root->log_root;
3201         struct btrfs_file_extent_item *extent;
3202         struct btrfs_inode_item *inode_item;
3203         int ret;
3204         struct btrfs_key *ins_keys;
3205         u32 *ins_sizes;
3206         char *ins_data;
3207         int i;
3208         struct list_head ordered_sums;
3209         int skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3210
3211         INIT_LIST_HEAD(&ordered_sums);
3212
3213         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
3214                            nr * sizeof(u32), GFP_NOFS);
3215         if (!ins_data)
3216                 return -ENOMEM;
3217
3218         ins_sizes = (u32 *)ins_data;
3219         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
3220
3221         for (i = 0; i < nr; i++) {
3222                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
3223                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
3224         }
3225         ret = btrfs_insert_empty_items(trans, log, dst_path,
3226                                        ins_keys, ins_sizes, nr);
3227         if (ret) {
3228                 kfree(ins_data);
3229                 return ret;
3230         }
3231
3232         for (i = 0; i < nr; i++, dst_path->slots[0]++) {
3233                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
3234                                                    dst_path->slots[0]);
3235
3236                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
3237
3238                 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
3239                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
3240                                                     dst_path->slots[0],
3241                                                     struct btrfs_inode_item);
3242                         fill_inode_item(trans, dst_path->nodes[0], inode_item,
3243                                         inode, inode_only == LOG_INODE_EXISTS);
3244                 } else {
3245                         copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
3246                                            src_offset, ins_sizes[i]);
3247                 }
3248
3249                 /* take a reference on file data extents so that truncates
3250                  * or deletes of this inode don't have to relog the inode
3251                  * again
3252                  */
3253                 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY &&
3254                     !skip_csum) {
3255                         int found_type;
3256                         extent = btrfs_item_ptr(src, start_slot + i,
3257                                                 struct btrfs_file_extent_item);
3258
3259                         if (btrfs_file_extent_generation(src, extent) < trans->transid)
3260                                 continue;
3261
3262                         found_type = btrfs_file_extent_type(src, extent);
3263                         if (found_type == BTRFS_FILE_EXTENT_REG) {
3264                                 u64 ds, dl, cs, cl;
3265                                 ds = btrfs_file_extent_disk_bytenr(src,
3266                                                                 extent);
3267                                 /* ds == 0 is a hole */
3268                                 if (ds == 0)
3269                                         continue;
3270
3271                                 dl = btrfs_file_extent_disk_num_bytes(src,
3272                                                                 extent);
3273                                 cs = btrfs_file_extent_offset(src, extent);
3274                                 cl = btrfs_file_extent_num_bytes(src,
3275                                                                 extent);
3276                                 if (btrfs_file_extent_compression(src,
3277                                                                   extent)) {
3278                                         cs = 0;
3279                                         cl = dl;
3280                                 }
3281
3282                                 ret = btrfs_lookup_csums_range(
3283                                                 log->fs_info->csum_root,
3284                                                 ds + cs, ds + cs + cl - 1,
3285                                                 &ordered_sums, 0);
3286                                 if (ret) {
3287                                         btrfs_release_path(dst_path);
3288                                         kfree(ins_data);
3289                                         return ret;
3290                                 }
3291                         }
3292                 }
3293         }
3294
3295         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
3296         btrfs_release_path(dst_path);
3297         kfree(ins_data);
3298
3299         /*
3300          * we have to do this after the loop above to avoid changing the
3301          * log tree while trying to change the log tree.
3302          */
3303         ret = 0;
3304         while (!list_empty(&ordered_sums)) {
3305                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3306                                                    struct btrfs_ordered_sum,
3307                                                    list);
3308                 if (!ret)
3309                         ret = btrfs_csum_file_blocks(trans, log, sums);
3310                 list_del(&sums->list);
3311                 kfree(sums);
3312         }
3313         return ret;
3314 }
3315
3316 static int extent_cmp(void *priv, struct list_head *a, struct list_head *b)
3317 {
3318         struct extent_map *em1, *em2;
3319
3320         em1 = list_entry(a, struct extent_map, list);
3321         em2 = list_entry(b, struct extent_map, list);
3322
3323         if (em1->start < em2->start)
3324                 return -1;
3325         else if (em1->start > em2->start)
3326                 return 1;
3327         return 0;
3328 }
3329
3330 static int log_one_extent(struct btrfs_trans_handle *trans,
3331                           struct inode *inode, struct btrfs_root *root,
3332                           struct extent_map *em, struct btrfs_path *path)
3333 {
3334         struct btrfs_root *log = root->log_root;
3335         struct btrfs_file_extent_item *fi;
3336         struct extent_buffer *leaf;
3337         struct btrfs_ordered_extent *ordered;
3338         struct list_head ordered_sums;
3339         struct btrfs_map_token token;
3340         struct btrfs_key key;
3341         u64 mod_start = em->mod_start;
3342         u64 mod_len = em->mod_len;
3343         u64 csum_offset;
3344         u64 csum_len;
3345         u64 extent_offset = em->start - em->orig_start;
3346         u64 block_len;
3347         int ret;
3348         int index = log->log_transid % 2;
3349         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3350
3351         ret = __btrfs_drop_extents(trans, log, inode, path, em->start,
3352                                    em->start + em->len, NULL, 0);
3353         if (ret)
3354                 return ret;
3355
3356         INIT_LIST_HEAD(&ordered_sums);
3357         btrfs_init_map_token(&token);
3358         key.objectid = btrfs_ino(inode);
3359         key.type = BTRFS_EXTENT_DATA_KEY;
3360         key.offset = em->start;
3361
3362         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*fi));
3363         if (ret)
3364                 return ret;
3365         leaf = path->nodes[0];
3366         fi = btrfs_item_ptr(leaf, path->slots[0],
3367                             struct btrfs_file_extent_item);
3368
3369         btrfs_set_token_file_extent_generation(leaf, fi, em->generation,
3370                                                &token);
3371         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3372                 skip_csum = true;
3373                 btrfs_set_token_file_extent_type(leaf, fi,
3374                                                  BTRFS_FILE_EXTENT_PREALLOC,
3375                                                  &token);
3376         } else {
3377                 btrfs_set_token_file_extent_type(leaf, fi,
3378                                                  BTRFS_FILE_EXTENT_REG,
3379                                                  &token);
3380                 if (em->block_start == EXTENT_MAP_HOLE)
3381                         skip_csum = true;
3382         }
3383
3384         block_len = max(em->block_len, em->orig_block_len);
3385         if (em->compress_type != BTRFS_COMPRESS_NONE) {
3386                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3387                                                         em->block_start,
3388                                                         &token);
3389                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3390                                                            &token);
3391         } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
3392                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3393                                                         em->block_start -
3394                                                         extent_offset, &token);
3395                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3396                                                            &token);
3397         } else {
3398                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi, 0, &token);
3399                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, 0,
3400                                                            &token);
3401         }
3402
3403         btrfs_set_token_file_extent_offset(leaf, fi,
3404                                            em->start - em->orig_start,
3405                                            &token);
3406         btrfs_set_token_file_extent_num_bytes(leaf, fi, em->len, &token);
3407         btrfs_set_token_file_extent_ram_bytes(leaf, fi, em->ram_bytes, &token);
3408         btrfs_set_token_file_extent_compression(leaf, fi, em->compress_type,
3409                                                 &token);
3410         btrfs_set_token_file_extent_encryption(leaf, fi, 0, &token);
3411         btrfs_set_token_file_extent_other_encoding(leaf, fi, 0, &token);
3412         btrfs_mark_buffer_dirty(leaf);
3413
3414         btrfs_release_path(path);
3415         if (ret) {
3416                 return ret;
3417         }
3418
3419         if (skip_csum)
3420                 return 0;
3421
3422         if (em->compress_type) {
3423                 csum_offset = 0;
3424                 csum_len = block_len;
3425         }
3426
3427         /*
3428          * First check and see if our csums are on our outstanding ordered
3429          * extents.
3430          */
3431 again:
3432         spin_lock_irq(&log->log_extents_lock[index]);
3433         list_for_each_entry(ordered, &log->logged_list[index], log_list) {
3434                 struct btrfs_ordered_sum *sum;
3435
3436                 if (!mod_len)
3437                         break;
3438
3439                 if (ordered->inode != inode)
3440                         continue;
3441
3442                 if (ordered->file_offset + ordered->len <= mod_start ||
3443                     mod_start + mod_len <= ordered->file_offset)
3444                         continue;
3445
3446                 /*
3447                  * We are going to copy all the csums on this ordered extent, so
3448                  * go ahead and adjust mod_start and mod_len in case this
3449                  * ordered extent has already been logged.
3450                  */
3451                 if (ordered->file_offset > mod_start) {
3452                         if (ordered->file_offset + ordered->len >=
3453                             mod_start + mod_len)
3454                                 mod_len = ordered->file_offset - mod_start;
3455                         /*
3456                          * If we have this case
3457                          *
3458                          * |--------- logged extent ---------|
3459                          *       |----- ordered extent ----|
3460                          *
3461                          * Just don't mess with mod_start and mod_len, we'll
3462                          * just end up logging more csums than we need and it
3463                          * will be ok.
3464                          */
3465                 } else {
3466                         if (ordered->file_offset + ordered->len <
3467                             mod_start + mod_len) {
3468                                 mod_len = (mod_start + mod_len) -
3469                                         (ordered->file_offset + ordered->len);
3470                                 mod_start = ordered->file_offset +
3471                                         ordered->len;
3472                         } else {
3473                                 mod_len = 0;
3474                         }
3475                 }
3476
3477                 /*
3478                  * To keep us from looping for the above case of an ordered
3479                  * extent that falls inside of the logged extent.
3480                  */
3481                 if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM,
3482                                      &ordered->flags))
3483                         continue;
3484                 atomic_inc(&ordered->refs);
3485                 spin_unlock_irq(&log->log_extents_lock[index]);
3486                 /*
3487                  * we've dropped the lock, we must either break or
3488                  * start over after this.
3489                  */
3490
3491                 wait_event(ordered->wait, ordered->csum_bytes_left == 0);
3492
3493                 list_for_each_entry(sum, &ordered->list, list) {
3494                         ret = btrfs_csum_file_blocks(trans, log, sum);
3495                         if (ret) {
3496                                 btrfs_put_ordered_extent(ordered);
3497                                 goto unlocked;
3498                         }
3499                 }
3500                 btrfs_put_ordered_extent(ordered);
3501                 goto again;
3502
3503         }
3504         spin_unlock_irq(&log->log_extents_lock[index]);
3505 unlocked:
3506
3507         if (!mod_len || ret)
3508                 return ret;
3509
3510         csum_offset = mod_start - em->start;
3511         csum_len = mod_len;
3512
3513         /* block start is already adjusted for the file extent offset. */
3514         ret = btrfs_lookup_csums_range(log->fs_info->csum_root,
3515                                        em->block_start + csum_offset,
3516                                        em->block_start + csum_offset +
3517                                        csum_len - 1, &ordered_sums, 0);
3518         if (ret)
3519                 return ret;
3520
3521         while (!list_empty(&ordered_sums)) {
3522                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3523                                                    struct btrfs_ordered_sum,
3524                                                    list);
3525                 if (!ret)
3526                         ret = btrfs_csum_file_blocks(trans, log, sums);
3527                 list_del(&sums->list);
3528                 kfree(sums);
3529         }
3530
3531         return ret;
3532 }
3533
3534 static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
3535                                      struct btrfs_root *root,
3536                                      struct inode *inode,
3537                                      struct btrfs_path *path)
3538 {
3539         struct extent_map *em, *n;
3540         struct list_head extents;
3541         struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3542         u64 test_gen;
3543         int ret = 0;
3544         int num = 0;
3545
3546         INIT_LIST_HEAD(&extents);
3547
3548         write_lock(&tree->lock);
3549         test_gen = root->fs_info->last_trans_committed;
3550
3551         list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
3552                 list_del_init(&em->list);
3553
3554                 /*
3555                  * Just an arbitrary number, this can be really CPU intensive
3556                  * once we start getting a lot of extents, and really once we
3557                  * have a bunch of extents we just want to commit since it will
3558                  * be faster.
3559                  */
3560                 if (++num > 32768) {
3561                         list_del_init(&tree->modified_extents);
3562                         ret = -EFBIG;
3563                         goto process;
3564                 }
3565
3566                 if (em->generation <= test_gen)
3567                         continue;
3568                 /* Need a ref to keep it from getting evicted from cache */
3569                 atomic_inc(&em->refs);
3570                 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
3571                 list_add_tail(&em->list, &extents);
3572                 num++;
3573         }
3574
3575         list_sort(NULL, &extents, extent_cmp);
3576
3577 process:
3578         while (!list_empty(&extents)) {
3579                 em = list_entry(extents.next, struct extent_map, list);
3580
3581                 list_del_init(&em->list);
3582
3583                 /*
3584                  * If we had an error we just need to delete everybody from our
3585                  * private list.
3586                  */
3587                 if (ret) {
3588                         clear_em_logging(tree, em);
3589                         free_extent_map(em);
3590                         continue;
3591                 }
3592
3593                 write_unlock(&tree->lock);
3594
3595                 ret = log_one_extent(trans, inode, root, em, path);
3596                 write_lock(&tree->lock);
3597                 clear_em_logging(tree, em);
3598                 free_extent_map(em);
3599         }
3600         WARN_ON(!list_empty(&extents));
3601         write_unlock(&tree->lock);
3602
3603         btrfs_release_path(path);
3604         return ret;
3605 }
3606
3607 /* log a single inode in the tree log.
3608  * At least one parent directory for this inode must exist in the tree
3609  * or be logged already.
3610  *
3611  * Any items from this inode changed by the current transaction are copied
3612  * to the log tree.  An extra reference is taken on any extents in this
3613  * file, allowing us to avoid a whole pile of corner cases around logging
3614  * blocks that have been removed from the tree.
3615  *
3616  * See LOG_INODE_ALL and related defines for a description of what inode_only
3617  * does.
3618  *
3619  * This handles both files and directories.
3620  */
3621 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
3622                              struct btrfs_root *root, struct inode *inode,
3623                              int inode_only)
3624 {
3625         struct btrfs_path *path;
3626         struct btrfs_path *dst_path;
3627         struct btrfs_key min_key;
3628         struct btrfs_key max_key;
3629         struct btrfs_root *log = root->log_root;
3630         struct extent_buffer *src = NULL;
3631         int err = 0;
3632         int ret;
3633         int nritems;
3634         int ins_start_slot = 0;
3635         int ins_nr;
3636         bool fast_search = false;
3637         u64 ino = btrfs_ino(inode);
3638
3639         path = btrfs_alloc_path();
3640         if (!path)
3641                 return -ENOMEM;
3642         dst_path = btrfs_alloc_path();
3643         if (!dst_path) {
3644                 btrfs_free_path(path);
3645                 return -ENOMEM;
3646         }
3647
3648         min_key.objectid = ino;
3649         min_key.type = BTRFS_INODE_ITEM_KEY;
3650         min_key.offset = 0;
3651
3652         max_key.objectid = ino;
3653
3654
3655         /* today the code can only do partial logging of directories */
3656         if (S_ISDIR(inode->i_mode) ||
3657             (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3658                        &BTRFS_I(inode)->runtime_flags) &&
3659              inode_only == LOG_INODE_EXISTS))
3660                 max_key.type = BTRFS_XATTR_ITEM_KEY;
3661         else
3662                 max_key.type = (u8)-1;
3663         max_key.offset = (u64)-1;
3664
3665         /* Only run delayed items if we are a dir or a new file */
3666         if (S_ISDIR(inode->i_mode) ||
3667             BTRFS_I(inode)->generation > root->fs_info->last_trans_committed) {
3668                 ret = btrfs_commit_inode_delayed_items(trans, inode);
3669                 if (ret) {
3670                         btrfs_free_path(path);
3671                         btrfs_free_path(dst_path);
3672                         return ret;
3673                 }
3674         }
3675
3676         mutex_lock(&BTRFS_I(inode)->log_mutex);
3677
3678         btrfs_get_logged_extents(log, inode);
3679
3680         /*
3681          * a brute force approach to making sure we get the most uptodate
3682          * copies of everything.
3683          */
3684         if (S_ISDIR(inode->i_mode)) {
3685                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
3686
3687                 if (inode_only == LOG_INODE_EXISTS)
3688                         max_key_type = BTRFS_XATTR_ITEM_KEY;
3689                 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
3690         } else {
3691                 if (test_and_clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3692                                        &BTRFS_I(inode)->runtime_flags)) {
3693                         clear_bit(BTRFS_INODE_COPY_EVERYTHING,
3694                                   &BTRFS_I(inode)->runtime_flags);
3695                         ret = btrfs_truncate_inode_items(trans, log,
3696                                                          inode, 0, 0);
3697                 } else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
3698                                               &BTRFS_I(inode)->runtime_flags)) {
3699                         if (inode_only == LOG_INODE_ALL)
3700                                 fast_search = true;
3701                         max_key.type = BTRFS_XATTR_ITEM_KEY;
3702                         ret = drop_objectid_items(trans, log, path, ino,
3703                                                   max_key.type);
3704                 } else {
3705                         if (inode_only == LOG_INODE_ALL)
3706                                 fast_search = true;
3707                         ret = log_inode_item(trans, log, dst_path, inode);
3708                         if (ret) {
3709                                 err = ret;
3710                                 goto out_unlock;
3711                         }
3712                         goto log_extents;
3713                 }
3714
3715         }
3716         if (ret) {
3717                 err = ret;
3718                 goto out_unlock;
3719         }
3720         path->keep_locks = 1;
3721
3722         while (1) {
3723                 ins_nr = 0;
3724                 ret = btrfs_search_forward(root, &min_key,
3725                                            path, trans->transid);
3726                 if (ret != 0)
3727                         break;
3728 again:
3729                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
3730                 if (min_key.objectid != ino)
3731                         break;
3732                 if (min_key.type > max_key.type)
3733                         break;
3734
3735                 src = path->nodes[0];
3736                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
3737                         ins_nr++;
3738                         goto next_slot;
3739                 } else if (!ins_nr) {
3740                         ins_start_slot = path->slots[0];
3741                         ins_nr = 1;
3742                         goto next_slot;
3743                 }
3744
3745                 ret = copy_items(trans, inode, dst_path, src, ins_start_slot,
3746                                  ins_nr, inode_only);
3747                 if (ret) {
3748                         err = ret;
3749                         goto out_unlock;
3750                 }
3751                 ins_nr = 1;
3752                 ins_start_slot = path->slots[0];
3753 next_slot:
3754
3755                 nritems = btrfs_header_nritems(path->nodes[0]);
3756                 path->slots[0]++;
3757                 if (path->slots[0] < nritems) {
3758                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
3759                                               path->slots[0]);
3760                         goto again;
3761                 }
3762                 if (ins_nr) {
3763                         ret = copy_items(trans, inode, dst_path, src,
3764                                          ins_start_slot,
3765                                          ins_nr, inode_only);
3766                         if (ret) {
3767                                 err = ret;
3768                                 goto out_unlock;
3769                         }
3770                         ins_nr = 0;
3771                 }
3772                 btrfs_release_path(path);
3773
3774                 if (min_key.offset < (u64)-1) {
3775                         min_key.offset++;
3776                 } else if (min_key.type < max_key.type) {
3777                         min_key.type++;
3778                         min_key.offset = 0;
3779                 } else {
3780                         break;
3781                 }
3782         }
3783         if (ins_nr) {
3784                 ret = copy_items(trans, inode, dst_path, src, ins_start_slot,
3785                                  ins_nr, inode_only);
3786                 if (ret) {
3787                         err = ret;
3788                         goto out_unlock;
3789                 }
3790                 ins_nr = 0;
3791         }
3792
3793 log_extents:
3794         btrfs_release_path(path);
3795         btrfs_release_path(dst_path);
3796         if (fast_search) {
3797                 ret = btrfs_log_changed_extents(trans, root, inode, dst_path);
3798                 if (ret) {
3799                         err = ret;
3800                         goto out_unlock;
3801                 }
3802         } else {
3803                 struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3804                 struct extent_map *em, *n;
3805
3806                 write_lock(&tree->lock);
3807                 list_for_each_entry_safe(em, n, &tree->modified_extents, list)
3808                         list_del_init(&em->list);
3809                 write_unlock(&tree->lock);
3810         }
3811
3812         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
3813                 ret = log_directory_changes(trans, root, inode, path, dst_path);
3814                 if (ret) {
3815                         err = ret;
3816                         goto out_unlock;
3817                 }
3818         }
3819         BTRFS_I(inode)->logged_trans = trans->transid;
3820         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->last_sub_trans;
3821 out_unlock:
3822         if (err)
3823                 btrfs_free_logged_extents(log, log->log_transid);
3824         mutex_unlock(&BTRFS_I(inode)->log_mutex);
3825
3826         btrfs_free_path(path);
3827         btrfs_free_path(dst_path);
3828         return err;
3829 }
3830
3831 /*
3832  * follow the dentry parent pointers up the chain and see if any
3833  * of the directories in it require a full commit before they can
3834  * be logged.  Returns zero if nothing special needs to be done or 1 if
3835  * a full commit is required.
3836  */
3837 static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
3838                                                struct inode *inode,
3839                                                struct dentry *parent,
3840                                                struct super_block *sb,
3841                                                u64 last_committed)
3842 {
3843         int ret = 0;
3844         struct btrfs_root *root;
3845         struct dentry *old_parent = NULL;
3846         struct inode *orig_inode = inode;
3847
3848         /*
3849          * for regular files, if its inode is already on disk, we don't
3850          * have to worry about the parents at all.  This is because
3851          * we can use the last_unlink_trans field to record renames
3852          * and other fun in this file.
3853          */
3854         if (S_ISREG(inode->i_mode) &&
3855             BTRFS_I(inode)->generation <= last_committed &&
3856             BTRFS_I(inode)->last_unlink_trans <= last_committed)
3857                         goto out;
3858
3859         if (!S_ISDIR(inode->i_mode)) {
3860                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3861                         goto out;
3862                 inode = parent->d_inode;
3863         }
3864
3865         while (1) {
3866                 /*
3867                  * If we are logging a directory then we start with our inode,
3868                  * not our parents inode, so we need to skipp setting the
3869                  * logged_trans so that further down in the log code we don't
3870                  * think this inode has already been logged.
3871                  */
3872                 if (inode != orig_inode)
3873                         BTRFS_I(inode)->logged_trans = trans->transid;
3874                 smp_mb();
3875
3876                 if (BTRFS_I(inode)->last_unlink_trans > last_committed) {
3877                         root = BTRFS_I(inode)->root;
3878
3879                         /*
3880                          * make sure any commits to the log are forced
3881                          * to be full commits
3882                          */
3883                         root->fs_info->last_trans_log_full_commit =
3884                                 trans->transid;
3885                         ret = 1;
3886                         break;
3887                 }
3888
3889                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3890                         break;
3891
3892                 if (IS_ROOT(parent))
3893                         break;
3894
3895                 parent = dget_parent(parent);
3896                 dput(old_parent);
3897                 old_parent = parent;
3898                 inode = parent->d_inode;
3899
3900         }
3901         dput(old_parent);
3902 out:
3903         return ret;
3904 }
3905
3906 /*
3907  * helper function around btrfs_log_inode to make sure newly created
3908  * parent directories also end up in the log.  A minimal inode and backref
3909  * only logging is done of any parent directories that are older than
3910  * the last committed transaction
3911  */
3912 static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
3913                                   struct btrfs_root *root, struct inode *inode,
3914                                   struct dentry *parent, int exists_only)
3915 {
3916         int inode_only = exists_only ? LOG_INODE_EXISTS : LOG_INODE_ALL;
3917         struct super_block *sb;
3918         struct dentry *old_parent = NULL;
3919         int ret = 0;
3920         u64 last_committed = root->fs_info->last_trans_committed;
3921
3922         sb = inode->i_sb;
3923
3924         if (btrfs_test_opt(root, NOTREELOG)) {
3925                 ret = 1;
3926                 goto end_no_trans;
3927         }
3928
3929         if (root->fs_info->last_trans_log_full_commit >
3930             root->fs_info->last_trans_committed) {
3931                 ret = 1;
3932                 goto end_no_trans;
3933         }
3934
3935         if (root != BTRFS_I(inode)->root ||
3936             btrfs_root_refs(&root->root_item) == 0) {
3937                 ret = 1;
3938                 goto end_no_trans;
3939         }
3940
3941         ret = check_parent_dirs_for_sync(trans, inode, parent,
3942                                          sb, last_committed);
3943         if (ret)
3944                 goto end_no_trans;
3945
3946         if (btrfs_inode_in_log(inode, trans->transid)) {
3947                 ret = BTRFS_NO_LOG_SYNC;
3948                 goto end_no_trans;
3949         }
3950
3951         ret = start_log_trans(trans, root);
3952         if (ret)
3953                 goto end_trans;
3954
3955         ret = btrfs_log_inode(trans, root, inode, inode_only);
3956         if (ret)
3957                 goto end_trans;
3958
3959         /*
3960          * for regular files, if its inode is already on disk, we don't
3961          * have to worry about the parents at all.  This is because
3962          * we can use the last_unlink_trans field to record renames
3963          * and other fun in this file.
3964          */
3965         if (S_ISREG(inode->i_mode) &&
3966             BTRFS_I(inode)->generation <= last_committed &&
3967             BTRFS_I(inode)->last_unlink_trans <= last_committed) {
3968                 ret = 0;
3969                 goto end_trans;
3970         }
3971
3972         inode_only = LOG_INODE_EXISTS;
3973         while (1) {
3974                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3975                         break;
3976
3977                 inode = parent->d_inode;
3978                 if (root != BTRFS_I(inode)->root)
3979                         break;
3980
3981                 if (BTRFS_I(inode)->generation >
3982                     root->fs_info->last_trans_committed) {
3983                         ret = btrfs_log_inode(trans, root, inode, inode_only);
3984                         if (ret)
3985                                 goto end_trans;
3986                 }
3987                 if (IS_ROOT(parent))
3988                         break;
3989
3990                 parent = dget_parent(parent);
3991                 dput(old_parent);
3992                 old_parent = parent;
3993         }
3994         ret = 0;
3995 end_trans:
3996         dput(old_parent);
3997         if (ret < 0) {
3998                 root->fs_info->last_trans_log_full_commit = trans->transid;
3999                 ret = 1;
4000         }
4001         btrfs_end_log_trans(root);
4002 end_no_trans:
4003         return ret;
4004 }
4005
4006 /*
4007  * it is not safe to log dentry if the chunk root has added new
4008  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
4009  * If this returns 1, you must commit the transaction to safely get your
4010  * data on disk.
4011  */
4012 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
4013                           struct btrfs_root *root, struct dentry *dentry)
4014 {
4015         struct dentry *parent = dget_parent(dentry);
4016         int ret;
4017
4018         ret = btrfs_log_inode_parent(trans, root, dentry->d_inode, parent, 0);
4019         dput(parent);
4020
4021         return ret;
4022 }
4023
4024 /*
4025  * should be called during mount to recover any replay any log trees
4026  * from the FS
4027  */
4028 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
4029 {
4030         int ret;
4031         struct btrfs_path *path;
4032         struct btrfs_trans_handle *trans;
4033         struct btrfs_key key;
4034         struct btrfs_key found_key;
4035         struct btrfs_key tmp_key;
4036         struct btrfs_root *log;
4037         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
4038         struct walk_control wc = {
4039                 .process_func = process_one_buffer,
4040                 .stage = 0,
4041         };
4042
4043         path = btrfs_alloc_path();
4044         if (!path)
4045                 return -ENOMEM;
4046
4047         fs_info->log_root_recovering = 1;
4048
4049         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4050         if (IS_ERR(trans)) {
4051                 ret = PTR_ERR(trans);
4052                 goto error;
4053         }
4054
4055         wc.trans = trans;
4056         wc.pin = 1;
4057
4058         ret = walk_log_tree(trans, log_root_tree, &wc);
4059         if (ret) {
4060                 btrfs_error(fs_info, ret, "Failed to pin buffers while "
4061                             "recovering log root tree.");
4062                 goto error;
4063         }
4064
4065 again:
4066         key.objectid = BTRFS_TREE_LOG_OBJECTID;
4067         key.offset = (u64)-1;
4068         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
4069
4070         while (1) {
4071                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
4072
4073                 if (ret < 0) {
4074                         btrfs_error(fs_info, ret,
4075                                     "Couldn't find tree log root.");
4076                         goto error;
4077                 }
4078                 if (ret > 0) {
4079                         if (path->slots[0] == 0)
4080                                 break;
4081                         path->slots[0]--;
4082                 }
4083                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
4084                                       path->slots[0]);
4085                 btrfs_release_path(path);
4086                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4087                         break;
4088
4089                 log = btrfs_read_fs_root(log_root_tree, &found_key);
4090                 if (IS_ERR(log)) {
4091                         ret = PTR_ERR(log);
4092                         btrfs_error(fs_info, ret,
4093                                     "Couldn't read tree log root.");
4094                         goto error;
4095                 }
4096
4097                 tmp_key.objectid = found_key.offset;
4098                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
4099                 tmp_key.offset = (u64)-1;
4100
4101                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
4102                 if (IS_ERR(wc.replay_dest)) {
4103                         ret = PTR_ERR(wc.replay_dest);
4104                         free_extent_buffer(log->node);
4105                         free_extent_buffer(log->commit_root);
4106                         kfree(log);
4107                         btrfs_error(fs_info, ret, "Couldn't read target root "
4108                                     "for tree log recovery.");
4109                         goto error;
4110                 }
4111
4112                 wc.replay_dest->log_root = log;
4113                 btrfs_record_root_in_trans(trans, wc.replay_dest);
4114                 ret = walk_log_tree(trans, log, &wc);
4115
4116                 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
4117                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
4118                                                       path);
4119                 }
4120
4121                 key.offset = found_key.offset - 1;
4122                 wc.replay_dest->log_root = NULL;
4123                 free_extent_buffer(log->node);
4124                 free_extent_buffer(log->commit_root);
4125                 kfree(log);
4126
4127                 if (ret)
4128                         goto error;
4129
4130                 if (found_key.offset == 0)
4131                         break;
4132         }
4133         btrfs_release_path(path);
4134
4135         /* step one is to pin it all, step two is to replay just inodes */
4136         if (wc.pin) {
4137                 wc.pin = 0;
4138                 wc.process_func = replay_one_buffer;
4139                 wc.stage = LOG_WALK_REPLAY_INODES;
4140                 goto again;
4141         }
4142         /* step three is to replay everything */
4143         if (wc.stage < LOG_WALK_REPLAY_ALL) {
4144                 wc.stage++;
4145                 goto again;
4146         }
4147
4148         btrfs_free_path(path);
4149
4150         /* step 4: commit the transaction, which also unpins the blocks */
4151         ret = btrfs_commit_transaction(trans, fs_info->tree_root);
4152         if (ret)
4153                 return ret;
4154
4155         free_extent_buffer(log_root_tree->node);
4156         log_root_tree->log_root = NULL;
4157         fs_info->log_root_recovering = 0;
4158         kfree(log_root_tree);
4159
4160         return 0;
4161 error:
4162         if (wc.trans)
4163                 btrfs_end_transaction(wc.trans, fs_info->tree_root);
4164         btrfs_free_path(path);
4165         return ret;
4166 }
4167
4168 /*
4169  * there are some corner cases where we want to force a full
4170  * commit instead of allowing a directory to be logged.
4171  *
4172  * They revolve around files there were unlinked from the directory, and
4173  * this function updates the parent directory so that a full commit is
4174  * properly done if it is fsync'd later after the unlinks are done.
4175  */
4176 void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
4177                              struct inode *dir, struct inode *inode,
4178                              int for_rename)
4179 {
4180         /*
4181          * when we're logging a file, if it hasn't been renamed
4182          * or unlinked, and its inode is fully committed on disk,
4183          * we don't have to worry about walking up the directory chain
4184          * to log its parents.
4185          *
4186          * So, we use the last_unlink_trans field to put this transid
4187          * into the file.  When the file is logged we check it and
4188          * don't log the parents if the file is fully on disk.
4189          */
4190         if (S_ISREG(inode->i_mode))
4191                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4192
4193         /*
4194          * if this directory was already logged any new
4195          * names for this file/dir will get recorded
4196          */
4197         smp_mb();
4198         if (BTRFS_I(dir)->logged_trans == trans->transid)
4199                 return;
4200
4201         /*
4202          * if the inode we're about to unlink was logged,
4203          * the log will be properly updated for any new names
4204          */
4205         if (BTRFS_I(inode)->logged_trans == trans->transid)
4206                 return;
4207
4208         /*
4209          * when renaming files across directories, if the directory
4210          * there we're unlinking from gets fsync'd later on, there's
4211          * no way to find the destination directory later and fsync it
4212          * properly.  So, we have to be conservative and force commits
4213          * so the new name gets discovered.
4214          */
4215         if (for_rename)
4216                 goto record;
4217
4218         /* we can safely do the unlink without any special recording */
4219         return;
4220
4221 record:
4222         BTRFS_I(dir)->last_unlink_trans = trans->transid;
4223 }
4224
4225 /*
4226  * Call this after adding a new name for a file and it will properly
4227  * update the log to reflect the new name.
4228  *
4229  * It will return zero if all goes well, and it will return 1 if a
4230  * full transaction commit is required.
4231  */
4232 int btrfs_log_new_name(struct btrfs_trans_handle *trans,
4233                         struct inode *inode, struct inode *old_dir,
4234                         struct dentry *parent)
4235 {
4236         struct btrfs_root * root = BTRFS_I(inode)->root;
4237
4238         /*
4239          * this will force the logging code to walk the dentry chain
4240          * up for the file
4241          */
4242         if (S_ISREG(inode->i_mode))
4243                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4244
4245         /*
4246          * if this inode hasn't been logged and directory we're renaming it
4247          * from hasn't been logged, we don't need to log it
4248          */
4249         if (BTRFS_I(inode)->logged_trans <=
4250             root->fs_info->last_trans_committed &&
4251             (!old_dir || BTRFS_I(old_dir)->logged_trans <=
4252                     root->fs_info->last_trans_committed))
4253                 return 0;
4254
4255         return btrfs_log_inode_parent(trans, root, inode, parent, 1);
4256 }
4257