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 "tree-log.h"
  24 #include "disk-io.h"
  25 #include "locking.h"
  26 #include "print-tree.h"
  27 #include "backref.h"
  28 #include "hash.h"
  29
  30 /* magic values for the inode_only field in btrfs_log_inode:
  31  *
  32  * LOG_INODE_ALL means to log everything
  33  * LOG_INODE_EXISTS means to log just enough to recreate the inode
  34  * during log replay
  35  */
  36 #define LOG_INODE_ALL 0
  37 #define LOG_INODE_EXISTS 1
  38
  39 /*
  40  * directory trouble cases
  41  *
  42  * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
  43  * log, we must force a full commit before doing an fsync of the directory
  44  * where the unlink was done.
  45  * ---> record transid of last unlink/rename per directory
  46  *
  47  * mkdir foo/some_dir
  48  * normal commit
  49  * rename foo/some_dir foo2/some_dir
  50  * mkdir foo/some_dir
  51  * fsync foo/some_dir/some_file
  52  *
  53  * The fsync above will unlink the original some_dir without recording
  54  * it in its new location (foo2).  After a crash, some_dir will be gone
  55  * unless the fsync of some_file forces a full commit
  56  *
  57  * 2) we must log any new names for any file or dir that is in the fsync
  58  * log. ---> check inode while renaming/linking.
  59  *
  60  * 2a) we must log any new names for any file or dir during rename
  61  * when the directory they are being removed from was logged.
  62  * ---> check inode and old parent dir during rename
  63  *
  64  *  2a is actually the more important variant.  With the extra logging
  65  *  a crash might unlink the old name without recreating the new one
  66  *
  67  * 3) after a crash, we must go through any directories with a link count
  68  * of zero and redo the rm -rf
  69  *
  70  * mkdir f1/foo
  71  * normal commit
  72  * rm -rf f1/foo
  73  * fsync(f1)
  74  *
  75  * The directory f1 was fully removed from the FS, but fsync was never
  76  * called on f1, only its parent dir.  After a crash the rm -rf must
  77  * be replayed.  This must be able to recurse down the entire
  78  * directory tree.  The inode link count fixup code takes care of the
  79  * ugly details.
  80  */
  81
  82 /*
  83  * stages for the tree walking.  The first
  84  * stage (0) is to only pin down the blocks we find
  85  * the second stage (1) is to make sure that all the inodes
  86  * we find in the log are created in the subvolume.
  87  *
  88  * The last stage is to deal with directories and links and extents
  89  * and all the other fun semantics
  90  */
  91 #define LOG_WALK_PIN_ONLY 0
  92 #define LOG_WALK_REPLAY_INODES 1
  93 #define LOG_WALK_REPLAY_DIR_INDEX 2
  94 #define LOG_WALK_REPLAY_ALL 3
  95
  96 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
  97                            struct btrfs_root *root, struct inode *inode,
  98                            int inode_only,
  99                            const loff_t start,
 100                            const loff_t end,
 101                            struct btrfs_log_ctx *ctx);
 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                            struct btrfs_log_ctx *ctx)
 142 {
 143         int index;
 144         int ret;
 145
 146         mutex_lock(&root->log_mutex);
 147         if (root->log_root) {
 148                 if (btrfs_need_log_full_commit(root->fs_info, trans)) {
 149                         ret = -EAGAIN;
 150                         goto out;
 151                 }
 152                 if (!root->log_start_pid) {
 153                         root->log_start_pid = current->pid;
 154                         clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 155                 } else if (root->log_start_pid != current->pid) {
 156                         set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 157                 }
 158
 159                 atomic_inc(&root->log_batch);
 160                 atomic_inc(&root->log_writers);
 161                 if (ctx) {
 162                         index = root->log_transid % 2;
 163                         list_add_tail(&ctx->list, &root->log_ctxs[index]);
 164                         ctx->log_transid = root->log_transid;
 165                 }
 166                 mutex_unlock(&root->log_mutex);
 167                 return 0;
 168         }
 169
 170         ret = 0;
 171         mutex_lock(&root->fs_info->tree_log_mutex);
 172         if (!root->fs_info->log_root_tree)
 173                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
 174         mutex_unlock(&root->fs_info->tree_log_mutex);
 175         if (ret)
 176                 goto out;
 177
 178         if (!root->log_root) {
 179                 ret = btrfs_add_log_tree(trans, root);
 180                 if (ret)
 181                         goto out;
 182         }
 183         clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 184         root->log_start_pid = current->pid;
 185         atomic_inc(&root->log_batch);
 186         atomic_inc(&root->log_writers);
 187         if (ctx) {
 188                 index = root->log_transid % 2;
 189                 list_add_tail(&ctx->list, &root->log_ctxs[index]);
 190                 ctx->log_transid = root->log_transid;
 191         }
 192 out:
 193         mutex_unlock(&root->log_mutex);
 194         return ret;
 195 }
 196
 197 /*
 198  * returns 0 if there was a log transaction running and we were able
 199  * to join, or returns -ENOENT if there were not transactions
 200  * in progress
 201  */
 202 static int join_running_log_trans(struct btrfs_root *root)
 203 {
 204         int ret = -ENOENT;
 205
 206         smp_mb();
 207         if (!root->log_root)
 208                 return -ENOENT;
 209
 210         mutex_lock(&root->log_mutex);
 211         if (root->log_root) {
 212                 ret = 0;
 213                 atomic_inc(&root->log_writers);
 214         }
 215         mutex_unlock(&root->log_mutex);
 216         return ret;
 217 }
 218
 219 /*
 220  * This either makes the current running log transaction wait
 221  * until you call btrfs_end_log_trans() or it makes any future
 222  * log transactions wait until you call btrfs_end_log_trans()
 223  */
 224 int btrfs_pin_log_trans(struct btrfs_root *root)
 225 {
 226         int ret = -ENOENT;
 227
 228         mutex_lock(&root->log_mutex);
 229         atomic_inc(&root->log_writers);
 230         mutex_unlock(&root->log_mutex);
 231         return ret;
 232 }
 233
 234 /*
 235  * indicate we're done making changes to the log tree
 236  * and wake up anyone waiting to do a sync
 237  */
 238 void btrfs_end_log_trans(struct btrfs_root *root)
 239 {
 240         if (atomic_dec_and_test(&root->log_writers)) {
 241                 smp_mb();
 242                 if (waitqueue_active(&root->log_writer_wait))
 243                         wake_up(&root->log_writer_wait);
 244         }
 245 }
 246
 247
 248 /*
 249  * the walk control struct is used to pass state down the chain when
 250  * processing the log tree.  The stage field tells us which part
 251  * of the log tree processing we are currently doing.  The others
 252  * are state fields used for that specific part
 253  */
 254 struct walk_control {
 255         /* should we free the extent on disk when done?  This is used
 256          * at transaction commit time while freeing a log tree
 257          */
 258         int free;
 259
 260         /* should we write out the extent buffer?  This is used
 261          * while flushing the log tree to disk during a sync
 262          */
 263         int write;
 264
 265         /* should we wait for the extent buffer io to finish?  Also used
 266          * while flushing the log tree to disk for a sync
 267          */
 268         int wait;
 269
 270         /* pin only walk, we record which extents on disk belong to the
 271          * log trees
 272          */
 273         int pin;
 274
 275         /* what stage of the replay code we're currently in */
 276         int stage;
 277
 278         /* the root we are currently replaying */
 279         struct btrfs_root *replay_dest;
 280
 281         /* the trans handle for the current replay */
 282         struct btrfs_trans_handle *trans;
 283
 284         /* the function that gets used to process blocks we find in the
 285          * tree.  Note the extent_buffer might not be up to date when it is
 286          * passed in, and it must be checked or read if you need the data
 287          * inside it
 288          */
 289         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
 290                             struct walk_control *wc, u64 gen);
 291 };
 292
 293 /*
 294  * process_func used to pin down extents, write them or wait on them
 295  */
 296 static int process_one_buffer(struct btrfs_root *log,
 297                               struct extent_buffer *eb,
 298                               struct walk_control *wc, u64 gen)
 299 {
 300         int ret = 0;
 301
 302         /*
 303          * If this fs is mixed then we need to be able to process the leaves to
 304          * pin down any logged extents, so we have to read the block.
 305          */
 306         if (btrfs_fs_incompat(log->fs_info, MIXED_GROUPS)) {
 307                 ret = btrfs_read_buffer(eb, gen);
 308                 if (ret)
 309                         return ret;
 310         }
 311
 312         if (wc->pin)
 313                 ret = btrfs_pin_extent_for_log_replay(log->fs_info->extent_root,
 314                                                       eb->start, eb->len);
 315
 316         if (!ret && btrfs_buffer_uptodate(eb, gen, 0)) {
 317                 if (wc->pin && btrfs_header_level(eb) == 0)
 318                         ret = btrfs_exclude_logged_extents(log, eb);
 319                 if (wc->write)
 320                         btrfs_write_tree_block(eb);
 321                 if (wc->wait)
 322                         btrfs_wait_tree_block_writeback(eb);
 323         }
 324         return ret;
 325 }
 326
 327 /*
 328  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
 329  * to the src data we are copying out.
 330  *
 331  * root is the tree we are copying into, and path is a scratch
 332  * path for use in this function (it should be released on entry and
 333  * will be released on exit).
 334  *
 335  * If the key is already in the destination tree the existing item is
 336  * overwritten.  If the existing item isn't big enough, it is extended.
 337  * If it is too large, it is truncated.
 338  *
 339  * If the key isn't in the destination yet, a new item is inserted.
 340  */
 341 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
 342                                    struct btrfs_root *root,
 343                                    struct btrfs_path *path,
 344                                    struct extent_buffer *eb, int slot,
 345                                    struct btrfs_key *key)
 346 {
 347         int ret;
 348         u32 item_size;
 349         u64 saved_i_size = 0;
 350         int save_old_i_size = 0;
 351         unsigned long src_ptr;
 352         unsigned long dst_ptr;
 353         int overwrite_root = 0;
 354         bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
 355
 356         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
 357                 overwrite_root = 1;
 358
 359         item_size = btrfs_item_size_nr(eb, slot);
 360         src_ptr = btrfs_item_ptr_offset(eb, slot);
 361
 362         /* look for the key in the destination tree */
 363         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
 364         if (ret < 0)
 365                 return ret;
 366
 367         if (ret == 0) {
 368                 char *src_copy;
 369                 char *dst_copy;
 370                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
 371                                                   path->slots[0]);
 372                 if (dst_size != item_size)
 373                         goto insert;
 374
 375                 if (item_size == 0) {
 376                         btrfs_release_path(path);
 377                         return 0;
 378                 }
 379                 dst_copy = kmalloc(item_size, GFP_NOFS);
 380                 src_copy = kmalloc(item_size, GFP_NOFS);
 381                 if (!dst_copy || !src_copy) {
 382                         btrfs_release_path(path);
 383                         kfree(dst_copy);
 384                         kfree(src_copy);
 385                         return -ENOMEM;
 386                 }
 387
 388                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
 389
 390                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 391                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
 392                                    item_size);
 393                 ret = memcmp(dst_copy, src_copy, item_size);
 394
 395                 kfree(dst_copy);
 396                 kfree(src_copy);
 397                 /*
 398                  * they have the same contents, just return, this saves
 399                  * us from cowing blocks in the destination tree and doing
 400                  * extra writes that may not have been done by a previous
 401                  * sync
 402                  */
 403                 if (ret == 0) {
 404                         btrfs_release_path(path);
 405                         return 0;
 406                 }
 407
 408                 /*
 409                  * We need to load the old nbytes into the inode so when we
 410                  * replay the extents we've logged we get the right nbytes.
 411                  */
 412                 if (inode_item) {
 413                         struct btrfs_inode_item *item;
 414                         u64 nbytes;
 415                         u32 mode;
 416
 417                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
 418                                               struct btrfs_inode_item);
 419                         nbytes = btrfs_inode_nbytes(path->nodes[0], item);
 420                         item = btrfs_item_ptr(eb, slot,
 421                                               struct btrfs_inode_item);
 422                         btrfs_set_inode_nbytes(eb, item, nbytes);
 423
 424                         /*
 425                          * If this is a directory we need to reset the i_size to
 426                          * 0 so that we can set it up properly when replaying
 427                          * the rest of the items in this log.
 428                          */
 429                         mode = btrfs_inode_mode(eb, item);
 430                         if (S_ISDIR(mode))
 431                                 btrfs_set_inode_size(eb, item, 0);
 432                 }
 433         } else if (inode_item) {
 434                 struct btrfs_inode_item *item;
 435                 u32 mode;
 436
 437                 /*
 438                  * New inode, set nbytes to 0 so that the nbytes comes out
 439                  * properly when we replay the extents.
 440                  */
 441                 item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
 442                 btrfs_set_inode_nbytes(eb, item, 0);
 443
 444                 /*
 445                  * If this is a directory we need to reset the i_size to 0 so
 446                  * that we can set it up properly when replaying the rest of
 447                  * the items in this log.
 448                  */
 449                 mode = btrfs_inode_mode(eb, item);
 450                 if (S_ISDIR(mode))
 451                         btrfs_set_inode_size(eb, item, 0);
 452         }
 453 insert:
 454         btrfs_release_path(path);
 455         /* try to insert the key into the destination tree */
 456         path->skip_release_on_error = 1;
 457         ret = btrfs_insert_empty_item(trans, root, path,
 458                                       key, item_size);
 459         path->skip_release_on_error = 0;
 460
 461         /* make sure any existing item is the correct size */
 462         if (ret == -EEXIST || ret == -EOVERFLOW) {
 463                 u32 found_size;
 464                 found_size = btrfs_item_size_nr(path->nodes[0],
 465                                                 path->slots[0]);
 466                 if (found_size > item_size)
 467                         btrfs_truncate_item(root, path, item_size, 1);
 468                 else if (found_size < item_size)
 469                         btrfs_extend_item(root, path,
 470                                           item_size - found_size);
 471         } else if (ret) {
 472                 return ret;
 473         }
 474         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
 475                                         path->slots[0]);
 476
 477         /* don't overwrite an existing inode if the generation number
 478          * was logged as zero.  This is done when the tree logging code
 479          * is just logging an inode to make sure it exists after recovery.
 480          *
 481          * Also, don't overwrite i_size on directories during replay.
 482          * log replay inserts and removes directory items based on the
 483          * state of the tree found in the subvolume, and i_size is modified
 484          * as it goes
 485          */
 486         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
 487                 struct btrfs_inode_item *src_item;
 488                 struct btrfs_inode_item *dst_item;
 489
 490                 src_item = (struct btrfs_inode_item *)src_ptr;
 491                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 492
 493                 if (btrfs_inode_generation(eb, src_item) == 0) {
 494                         struct extent_buffer *dst_eb = path->nodes[0];
 495
 496                         if (S_ISREG(btrfs_inode_mode(eb, src_item)) &&
 497                             S_ISREG(btrfs_inode_mode(dst_eb, dst_item))) {
 498                                 struct btrfs_map_token token;
 499                                 u64 ino_size = btrfs_inode_size(eb, src_item);
 500
 501                                 btrfs_init_map_token(&token);
 502                                 btrfs_set_token_inode_size(dst_eb, dst_item,
 503                                                            ino_size, &token);
 504                         }
 505                         goto no_copy;
 506                 }
 507
 508                 if (overwrite_root &&
 509                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
 510                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
 511                         save_old_i_size = 1;
 512                         saved_i_size = btrfs_inode_size(path->nodes[0],
 513                                                         dst_item);
 514                 }
 515         }
 516
 517         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
 518                            src_ptr, item_size);
 519
 520         if (save_old_i_size) {
 521                 struct btrfs_inode_item *dst_item;
 522                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 523                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
 524         }
 525
 526         /* make sure the generation is filled in */
 527         if (key->type == BTRFS_INODE_ITEM_KEY) {
 528                 struct btrfs_inode_item *dst_item;
 529                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 530                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
 531                         btrfs_set_inode_generation(path->nodes[0], dst_item,
 532                                                    trans->transid);
 533                 }
 534         }
 535 no_copy:
 536         btrfs_mark_buffer_dirty(path->nodes[0]);
 537         btrfs_release_path(path);
 538         return 0;
 539 }
 540
 541 /*
 542  * simple helper to read an inode off the disk from a given root
 543  * This can only be called for subvolume roots and not for the log
 544  */
 545 static noinline struct inode *read_one_inode(struct btrfs_root *root,
 546                                              u64 objectid)
 547 {
 548         struct btrfs_key key;
 549         struct inode *inode;
 550
 551         key.objectid = objectid;
 552         key.type = BTRFS_INODE_ITEM_KEY;
 553         key.offset = 0;
 554         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
 555         if (IS_ERR(inode)) {
 556                 inode = NULL;
 557         } else if (is_bad_inode(inode)) {
 558                 iput(inode);
 559                 inode = NULL;
 560         }
 561         return inode;
 562 }
 563
 564 /* replays a single extent in 'eb' at 'slot' with 'key' into the
 565  * subvolume 'root'.  path is released on entry and should be released
 566  * on exit.
 567  *
 568  * extents in the log tree have not been allocated out of the extent
 569  * tree yet.  So, this completes the allocation, taking a reference
 570  * as required if the extent already exists or creating a new extent
 571  * if it isn't in the extent allocation tree yet.
 572  *
 573  * The extent is inserted into the file, dropping any existing extents
 574  * from the file that overlap the new one.
 575  */
 576 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
 577                                       struct btrfs_root *root,
 578                                       struct btrfs_path *path,
 579                                       struct extent_buffer *eb, int slot,
 580                                       struct btrfs_key *key)
 581 {
 582         int found_type;
 583         u64 extent_end;
 584         u64 start = key->offset;
 585         u64 nbytes = 0;
 586         struct btrfs_file_extent_item *item;
 587         struct inode *inode = NULL;
 588         unsigned long size;
 589         int ret = 0;
 590
 591         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 592         found_type = btrfs_file_extent_type(eb, item);
 593
 594         if (found_type == BTRFS_FILE_EXTENT_REG ||
 595             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 596                 nbytes = btrfs_file_extent_num_bytes(eb, item);
 597                 extent_end = start + nbytes;
 598
 599                 /*
 600                  * We don't add to the inodes nbytes if we are prealloc or a
 601                  * hole.
 602                  */
 603                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
 604                         nbytes = 0;
 605         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 606                 size = btrfs_file_extent_inline_len(eb, slot, item);
 607                 nbytes = btrfs_file_extent_ram_bytes(eb, item);
 608                 extent_end = ALIGN(start + size, root->sectorsize);
 609         } else {
 610                 ret = 0;
 611                 goto out;
 612         }
 613
 614         inode = read_one_inode(root, key->objectid);
 615         if (!inode) {
 616                 ret = -EIO;
 617                 goto out;
 618         }
 619
 620         /*
 621          * first check to see if we already have this extent in the
 622          * file.  This must be done before the btrfs_drop_extents run
 623          * so we don't try to drop this extent.
 624          */
 625         ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
 626                                        start, 0);
 627
 628         if (ret == 0 &&
 629             (found_type == BTRFS_FILE_EXTENT_REG ||
 630              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
 631                 struct btrfs_file_extent_item cmp1;
 632                 struct btrfs_file_extent_item cmp2;
 633                 struct btrfs_file_extent_item *existing;
 634                 struct extent_buffer *leaf;
 635
 636                 leaf = path->nodes[0];
 637                 existing = btrfs_item_ptr(leaf, path->slots[0],
 638                                           struct btrfs_file_extent_item);
 639
 640                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
 641                                    sizeof(cmp1));
 642                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
 643                                    sizeof(cmp2));
 644
 645                 /*
 646                  * we already have a pointer to this exact extent,
 647                  * we don't have to do anything
 648                  */
 649                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
 650                         btrfs_release_path(path);
 651                         goto out;
 652                 }
 653         }
 654         btrfs_release_path(path);
 655
 656         /* drop any overlapping extents */
 657         ret = btrfs_drop_extents(trans, root, inode, start, extent_end, 1);
 658         if (ret)
 659                 goto out;
 660
 661         if (found_type == BTRFS_FILE_EXTENT_REG ||
 662             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 663                 u64 offset;
 664                 unsigned long dest_offset;
 665                 struct btrfs_key ins;
 666
 667                 ret = btrfs_insert_empty_item(trans, root, path, key,
 668                                               sizeof(*item));
 669                 if (ret)
 670                         goto out;
 671                 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
 672                                                     path->slots[0]);
 673                 copy_extent_buffer(path->nodes[0], eb, dest_offset,
 674                                 (unsigned long)item,  sizeof(*item));
 675
 676                 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
 677                 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
 678                 ins.type = BTRFS_EXTENT_ITEM_KEY;
 679                 offset = key->offset - btrfs_file_extent_offset(eb, item);
 680
 681                 if (ins.objectid > 0) {
 682                         u64 csum_start;
 683                         u64 csum_end;
 684                         LIST_HEAD(ordered_sums);
 685                         /*
 686                          * is this extent already allocated in the extent
 687                          * allocation tree?  If so, just add a reference
 688                          */
 689                         ret = btrfs_lookup_data_extent(root, ins.objectid,
 690                                                 ins.offset);
 691                         if (ret == 0) {
 692                                 ret = btrfs_inc_extent_ref(trans, root,
 693                                                 ins.objectid, ins.offset,
 694                                                 0, root->root_key.objectid,
 695                                                 key->objectid, offset, 0);
 696                                 if (ret)
 697                                         goto out;
 698                         } else {
 699                                 /*
 700                                  * insert the extent pointer in the extent
 701                                  * allocation tree
 702                                  */
 703                                 ret = btrfs_alloc_logged_file_extent(trans,
 704                                                 root, root->root_key.objectid,
 705                                                 key->objectid, offset, &ins);
 706                                 if (ret)
 707                                         goto out;
 708                         }
 709                         btrfs_release_path(path);
 710
 711                         if (btrfs_file_extent_compression(eb, item)) {
 712                                 csum_start = ins.objectid;
 713                                 csum_end = csum_start + ins.offset;
 714                         } else {
 715                                 csum_start = ins.objectid +
 716                                         btrfs_file_extent_offset(eb, item);
 717                                 csum_end = csum_start +
 718                                         btrfs_file_extent_num_bytes(eb, item);
 719                         }
 720
 721                         ret = btrfs_lookup_csums_range(root->log_root,
 722                                                 csum_start, csum_end - 1,
 723                                                 &ordered_sums, 0);
 724                         if (ret)
 725                                 goto out;
 726                         while (!list_empty(&ordered_sums)) {
 727                                 struct btrfs_ordered_sum *sums;
 728                                 sums = list_entry(ordered_sums.next,
 729                                                 struct btrfs_ordered_sum,
 730                                                 list);
 731                                 if (!ret)
 732                                         ret = btrfs_csum_file_blocks(trans,
 733                                                 root->fs_info->csum_root,
 734                                                 sums);
 735                                 list_del(&sums->list);
 736                                 kfree(sums);
 737                         }
 738                         if (ret)
 739                                 goto out;
 740                 } else {
 741                         btrfs_release_path(path);
 742                 }
 743         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 744                 /* inline extents are easy, we just overwrite them */
 745                 ret = overwrite_item(trans, root, path, eb, slot, key);
 746                 if (ret)
 747                         goto out;
 748         }
 749
 750         inode_add_bytes(inode, nbytes);
 751         ret = btrfs_update_inode(trans, root, inode);
 752 out:
 753         if (inode)
 754                 iput(inode);
 755         return ret;
 756 }
 757
 758 /*
 759  * when cleaning up conflicts between the directory names in the
 760  * subvolume, directory names in the log and directory names in the
 761  * inode back references, we may have to unlink inodes from directories.
 762  *
 763  * This is a helper function to do the unlink of a specific directory
 764  * item
 765  */
 766 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
 767                                       struct btrfs_root *root,
 768                                       struct btrfs_path *path,
 769                                       struct inode *dir,
 770                                       struct btrfs_dir_item *di)
 771 {
 772         struct inode *inode;
 773         char *name;
 774         int name_len;
 775         struct extent_buffer *leaf;
 776         struct btrfs_key location;
 777         int ret;
 778
 779         leaf = path->nodes[0];
 780
 781         btrfs_dir_item_key_to_cpu(leaf, di, &location);
 782         name_len = btrfs_dir_name_len(leaf, di);
 783         name = kmalloc(name_len, GFP_NOFS);
 784         if (!name)
 785                 return -ENOMEM;
 786
 787         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
 788         btrfs_release_path(path);
 789
 790         inode = read_one_inode(root, location.objectid);
 791         if (!inode) {
 792                 ret = -EIO;
 793                 goto out;
 794         }
 795
 796         ret = link_to_fixup_dir(trans, root, path, location.objectid);
 797         if (ret)
 798                 goto out;
 799
 800         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
 801         if (ret)
 802                 goto out;
 803         else
 804                 ret = btrfs_run_delayed_items(trans, root);
 805 out:
 806         kfree(name);
 807         iput(inode);
 808         return ret;
 809 }
 810
 811 /*
 812  * helper function to see if a given name and sequence number found
 813  * in an inode back reference are already in a directory and correctly
 814  * point to this inode
 815  */
 816 static noinline int inode_in_dir(struct btrfs_root *root,
 817                                  struct btrfs_path *path,
 818                                  u64 dirid, u64 objectid, u64 index,
 819                                  const char *name, int name_len)
 820 {
 821         struct btrfs_dir_item *di;
 822         struct btrfs_key location;
 823         int match = 0;
 824
 825         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
 826                                          index, name, name_len, 0);
 827         if (di && !IS_ERR(di)) {
 828                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 829                 if (location.objectid != objectid)
 830                         goto out;
 831         } else
 832                 goto out;
 833         btrfs_release_path(path);
 834
 835         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
 836         if (di && !IS_ERR(di)) {
 837                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 838                 if (location.objectid != objectid)
 839                         goto out;
 840         } else
 841                 goto out;
 842         match = 1;
 843 out:
 844         btrfs_release_path(path);
 845         return match;
 846 }
 847
 848 /*
 849  * helper function to check a log tree for a named back reference in
 850  * an inode.  This is used to decide if a back reference that is
 851  * found in the subvolume conflicts with what we find in the log.
 852  *
 853  * inode backreferences may have multiple refs in a single item,
 854  * during replay we process one reference at a time, and we don't
 855  * want to delete valid links to a file from the subvolume if that
 856  * link is also in the log.
 857  */
 858 static noinline int backref_in_log(struct btrfs_root *log,
 859                                    struct btrfs_key *key,
 860                                    u64 ref_objectid,
 861                                    const char *name, int namelen)
 862 {
 863         struct btrfs_path *path;
 864         struct btrfs_inode_ref *ref;
 865         unsigned long ptr;
 866         unsigned long ptr_end;
 867         unsigned long name_ptr;
 868         int found_name_len;
 869         int item_size;
 870         int ret;
 871         int match = 0;
 872
 873         path = btrfs_alloc_path();
 874         if (!path)
 875                 return -ENOMEM;
 876
 877         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
 878         if (ret != 0)
 879                 goto out;
 880
 881         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 882
 883         if (key->type == BTRFS_INODE_EXTREF_KEY) {
 884                 if (btrfs_find_name_in_ext_backref(path, ref_objectid,
 885                                                    name, namelen, NULL))
 886                         match = 1;
 887
 888                 goto out;
 889         }
 890
 891         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
 892         ptr_end = ptr + item_size;
 893         while (ptr < ptr_end) {
 894                 ref = (struct btrfs_inode_ref *)ptr;
 895                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
 896                 if (found_name_len == namelen) {
 897                         name_ptr = (unsigned long)(ref + 1);
 898                         ret = memcmp_extent_buffer(path->nodes[0], name,
 899                                                    name_ptr, namelen);
 900                         if (ret == 0) {
 901                                 match = 1;
 902                                 goto out;
 903                         }
 904                 }
 905                 ptr = (unsigned long)(ref + 1) + found_name_len;
 906         }
 907 out:
 908         btrfs_free_path(path);
 909         return match;
 910 }
 911
 912 static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
 913                                   struct btrfs_root *root,
 914                                   struct btrfs_path *path,
 915                                   struct btrfs_root *log_root,
 916                                   struct inode *dir, struct inode *inode,
 917                                   struct extent_buffer *eb,
 918                                   u64 inode_objectid, u64 parent_objectid,
 919                                   u64 ref_index, char *name, int namelen,
 920                                   int *search_done)
 921 {
 922         int ret;
 923         char *victim_name;
 924         int victim_name_len;
 925         struct extent_buffer *leaf;
 926         struct btrfs_dir_item *di;
 927         struct btrfs_key search_key;
 928         struct btrfs_inode_extref *extref;
 929
 930 again:
 931         /* Search old style refs */
 932         search_key.objectid = inode_objectid;
 933         search_key.type = BTRFS_INODE_REF_KEY;
 934         search_key.offset = parent_objectid;
 935         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 936         if (ret == 0) {
 937                 struct btrfs_inode_ref *victim_ref;
 938                 unsigned long ptr;
 939                 unsigned long ptr_end;
 940
 941                 leaf = path->nodes[0];
 942
 943                 /* are we trying to overwrite a back ref for the root directory
 944                  * if so, just jump out, we're done
 945                  */
 946                 if (search_key.objectid == search_key.offset)
 947                         return 1;
 948
 949                 /* check all the names in this back reference to see
 950                  * if they are in the log.  if so, we allow them to stay
 951                  * otherwise they must be unlinked as a conflict
 952                  */
 953                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 954                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
 955                 while (ptr < ptr_end) {
 956                         victim_ref = (struct btrfs_inode_ref *)ptr;
 957                         victim_name_len = btrfs_inode_ref_name_len(leaf,
 958                                                                    victim_ref);
 959                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 960                         if (!victim_name)
 961                                 return -ENOMEM;
 962
 963                         read_extent_buffer(leaf, victim_name,
 964                                            (unsigned long)(victim_ref + 1),
 965                                            victim_name_len);
 966
 967                         if (!backref_in_log(log_root, &search_key,
 968                                             parent_objectid,
 969                                             victim_name,
 970                                             victim_name_len)) {
 971                                 inc_nlink(inode);
 972                                 btrfs_release_path(path);
 973
 974                                 ret = btrfs_unlink_inode(trans, root, dir,
 975                                                          inode, victim_name,
 976                                                          victim_name_len);
 977                                 kfree(victim_name);
 978                                 if (ret)
 979                                         return ret;
 980                                 ret = btrfs_run_delayed_items(trans, root);
 981                                 if (ret)
 982                                         return ret;
 983                                 *search_done = 1;
 984                                 goto again;
 985                         }
 986                         kfree(victim_name);
 987
 988                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
 989                 }
 990
 991                 /*
 992                  * NOTE: we have searched root tree and checked the
 993                  * coresponding ref, it does not need to check again.
 994                  */
 995                 *search_done = 1;
 996         }
 997         btrfs_release_path(path);
 998
 999         /* Same search but for extended refs */
1000         extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
1001                                            inode_objectid, parent_objectid, 0,
1002                                            0);
1003         if (!IS_ERR_OR_NULL(extref)) {
1004                 u32 item_size;
1005                 u32 cur_offset = 0;
1006                 unsigned long base;
1007                 struct inode *victim_parent;
1008
1009                 leaf = path->nodes[0];
1010
1011                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1012                 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
1013
1014                 while (cur_offset < item_size) {
1015                         extref = (struct btrfs_inode_extref *)base + cur_offset;
1016
1017                         victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
1018
1019                         if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
1020                                 goto next;
1021
1022                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
1023                         if (!victim_name)
1024                                 return -ENOMEM;
1025                         read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
1026                                            victim_name_len);
1027
1028                         search_key.objectid = inode_objectid;
1029                         search_key.type = BTRFS_INODE_EXTREF_KEY;
1030                         search_key.offset = btrfs_extref_hash(parent_objectid,
1031                                                               victim_name,
1032                                                               victim_name_len);
1033                         ret = 0;
1034                         if (!backref_in_log(log_root, &search_key,
1035                                             parent_objectid, victim_name,
1036                                             victim_name_len)) {
1037                                 ret = -ENOENT;
1038                                 victim_parent = read_one_inode(root,
1039                                                                parent_objectid);
1040                                 if (victim_parent) {
1041                                         inc_nlink(inode);
1042                                         btrfs_release_path(path);
1043
1044                                         ret = btrfs_unlink_inode(trans, root,
1045                                                                  victim_parent,
1046                                                                  inode,
1047                                                                  victim_name,
1048                                                                  victim_name_len);
1049                                         if (!ret)
1050                                                 ret = btrfs_run_delayed_items(
1051                                                                   trans, root);
1052                                 }
1053                                 iput(victim_parent);
1054                                 kfree(victim_name);
1055                                 if (ret)
1056                                         return ret;
1057                                 *search_done = 1;
1058                                 goto again;
1059                         }
1060                         kfree(victim_name);
1061                         if (ret)
1062                                 return ret;
1063 next:
1064                         cur_offset += victim_name_len + sizeof(*extref);
1065                 }
1066                 *search_done = 1;
1067         }
1068         btrfs_release_path(path);
1069
1070         /* look for a conflicting sequence number */
1071         di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
1072                                          ref_index, name, namelen, 0);
1073         if (di && !IS_ERR(di)) {
1074                 ret = drop_one_dir_item(trans, root, path, dir, di);
1075                 if (ret)
1076                         return ret;
1077         }
1078         btrfs_release_path(path);
1079
1080         /* look for a conflicing name */
1081         di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
1082                                    name, namelen, 0);
1083         if (di && !IS_ERR(di)) {
1084                 ret = drop_one_dir_item(trans, root, path, dir, di);
1085                 if (ret)
1086                         return ret;
1087         }
1088         btrfs_release_path(path);
1089
1090         return 0;
1091 }
1092
1093 static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1094                              u32 *namelen, char **name, u64 *index,
1095                              u64 *parent_objectid)
1096 {
1097         struct btrfs_inode_extref *extref;
1098
1099         extref = (struct btrfs_inode_extref *)ref_ptr;
1100
1101         *namelen = btrfs_inode_extref_name_len(eb, extref);
1102         *name = kmalloc(*namelen, GFP_NOFS);
1103         if (*name == NULL)
1104                 return -ENOMEM;
1105
1106         read_extent_buffer(eb, *name, (unsigned long)&extref->name,
1107                            *namelen);
1108
1109         *index = btrfs_inode_extref_index(eb, extref);
1110         if (parent_objectid)
1111                 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
1112
1113         return 0;
1114 }
1115
1116 static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1117                           u32 *namelen, char **name, u64 *index)
1118 {
1119         struct btrfs_inode_ref *ref;
1120
1121         ref = (struct btrfs_inode_ref *)ref_ptr;
1122
1123         *namelen = btrfs_inode_ref_name_len(eb, ref);
1124         *name = kmalloc(*namelen, GFP_NOFS);
1125         if (*name == NULL)
1126                 return -ENOMEM;
1127
1128         read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1129
1130         *index = btrfs_inode_ref_index(eb, ref);
1131
1132         return 0;
1133 }
1134
1135 /*
1136  * replay one inode back reference item found in the log tree.
1137  * eb, slot and key refer to the buffer and key found in the log tree.
1138  * root is the destination we are replaying into, and path is for temp
1139  * use by this function.  (it should be released on return).
1140  */
1141 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1142                                   struct btrfs_root *root,
1143                                   struct btrfs_root *log,
1144                                   struct btrfs_path *path,
1145                                   struct extent_buffer *eb, int slot,
1146                                   struct btrfs_key *key)
1147 {
1148         struct inode *dir = NULL;
1149         struct inode *inode = NULL;
1150         unsigned long ref_ptr;
1151         unsigned long ref_end;
1152         char *name = NULL;
1153         int namelen;
1154         int ret;
1155         int search_done = 0;
1156         int log_ref_ver = 0;
1157         u64 parent_objectid;
1158         u64 inode_objectid;
1159         u64 ref_index = 0;
1160         int ref_struct_size;
1161
1162         ref_ptr = btrfs_item_ptr_offset(eb, slot);
1163         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1164
1165         if (key->type == BTRFS_INODE_EXTREF_KEY) {
1166                 struct btrfs_inode_extref *r;
1167
1168                 ref_struct_size = sizeof(struct btrfs_inode_extref);
1169                 log_ref_ver = 1;
1170                 r = (struct btrfs_inode_extref *)ref_ptr;
1171                 parent_objectid = btrfs_inode_extref_parent(eb, r);
1172         } else {
1173                 ref_struct_size = sizeof(struct btrfs_inode_ref);
1174                 parent_objectid = key->offset;
1175         }
1176         inode_objectid = key->objectid;
1177
1178         /*
1179          * it is possible that we didn't log all the parent directories
1180          * for a given inode.  If we don't find the dir, just don't
1181          * copy the back ref in.  The link count fixup code will take
1182          * care of the rest
1183          */
1184         dir = read_one_inode(root, parent_objectid);
1185         if (!dir) {
1186                 ret = -ENOENT;
1187                 goto out;
1188         }
1189
1190         inode = read_one_inode(root, inode_objectid);
1191         if (!inode) {
1192                 ret = -EIO;
1193                 goto out;
1194         }
1195
1196         while (ref_ptr < ref_end) {
1197                 if (log_ref_ver) {
1198                         ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1199                                                 &ref_index, &parent_objectid);
1200                         /*
1201                          * parent object can change from one array
1202                          * item to another.
1203                          */
1204                         if (!dir)
1205                                 dir = read_one_inode(root, parent_objectid);
1206                         if (!dir) {
1207                                 ret = -ENOENT;
1208                                 goto out;
1209                         }
1210                 } else {
1211                         ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1212                                              &ref_index);
1213                 }
1214                 if (ret)
1215                         goto out;
1216
1217                 /* if we already have a perfect match, we're done */
1218                 if (!inode_in_dir(root, path, btrfs_ino(dir), btrfs_ino(inode),
1219                                   ref_index, name, namelen)) {
1220                         /*
1221                          * look for a conflicting back reference in the
1222                          * metadata. if we find one we have to unlink that name
1223                          * of the file before we add our new link.  Later on, we
1224                          * overwrite any existing back reference, and we don't
1225                          * want to create dangling pointers in the directory.
1226                          */
1227
1228                         if (!search_done) {
1229                                 ret = __add_inode_ref(trans, root, path, log,
1230                                                       dir, inode, eb,
1231                                                       inode_objectid,
1232                                                       parent_objectid,
1233                                                       ref_index, name, namelen,
1234                                                       &search_done);
1235                                 if (ret) {
1236                                         if (ret == 1)
1237                                                 ret = 0;
1238                                         goto out;
1239                                 }
1240                         }
1241
1242                         /* insert our name */
1243                         ret = btrfs_add_link(trans, dir, inode, name, namelen,
1244                                              0, ref_index);
1245                         if (ret)
1246                                 goto out;
1247
1248                         btrfs_update_inode(trans, root, inode);
1249                 }
1250
1251                 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1252                 kfree(name);
1253                 name = NULL;
1254                 if (log_ref_ver) {
1255                         iput(dir);
1256                         dir = NULL;
1257                 }
1258         }
1259
1260         /* finally write the back reference in the inode */
1261         ret = overwrite_item(trans, root, path, eb, slot, key);
1262 out:
1263         btrfs_release_path(path);
1264         kfree(name);
1265         iput(dir);
1266         iput(inode);
1267         return ret;
1268 }
1269
1270 static int insert_orphan_item(struct btrfs_trans_handle *trans,
1271                               struct btrfs_root *root, u64 ino)
1272 {
1273         int ret;
1274
1275         ret = btrfs_insert_orphan_item(trans, root, ino);
1276         if (ret == -EEXIST)
1277                 ret = 0;
1278
1279         return ret;
1280 }
1281
1282 static int count_inode_extrefs(struct btrfs_root *root,
1283                                struct inode *inode, struct btrfs_path *path)
1284 {
1285         int ret = 0;
1286         int name_len;
1287         unsigned int nlink = 0;
1288         u32 item_size;
1289         u32 cur_offset = 0;
1290         u64 inode_objectid = btrfs_ino(inode);
1291         u64 offset = 0;
1292         unsigned long ptr;
1293         struct btrfs_inode_extref *extref;
1294         struct extent_buffer *leaf;
1295
1296         while (1) {
1297                 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1298                                             &extref, &offset);
1299                 if (ret)
1300                         break;
1301
1302                 leaf = path->nodes[0];
1303                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1304                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1305                 cur_offset = 0;
1306
1307                 while (cur_offset < item_size) {
1308                         extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1309                         name_len = btrfs_inode_extref_name_len(leaf, extref);
1310
1311                         nlink++;
1312
1313                         cur_offset += name_len + sizeof(*extref);
1314                 }
1315
1316                 offset++;
1317                 btrfs_release_path(path);
1318         }
1319         btrfs_release_path(path);
1320
1321         if (ret < 0 && ret != -ENOENT)
1322                 return ret;
1323         return nlink;
1324 }
1325
1326 static int count_inode_refs(struct btrfs_root *root,
1327                                struct inode *inode, struct btrfs_path *path)
1328 {
1329         int ret;
1330         struct btrfs_key key;
1331         unsigned int nlink = 0;
1332         unsigned long ptr;
1333         unsigned long ptr_end;
1334         int name_len;
1335         u64 ino = btrfs_ino(inode);
1336
1337         key.objectid = ino;
1338         key.type = BTRFS_INODE_REF_KEY;
1339         key.offset = (u64)-1;
1340
1341         while (1) {
1342                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1343                 if (ret < 0)
1344                         break;
1345                 if (ret > 0) {
1346                         if (path->slots[0] == 0)
1347                                 break;
1348                         path->slots[0]--;
1349                 }
1350 process_slot:
1351                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1352                                       path->slots[0]);
1353                 if (key.objectid != ino ||
1354                     key.type != BTRFS_INODE_REF_KEY)
1355                         break;
1356                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1357                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1358                                                    path->slots[0]);
1359                 while (ptr < ptr_end) {
1360                         struct btrfs_inode_ref *ref;
1361
1362                         ref = (struct btrfs_inode_ref *)ptr;
1363                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1364                                                             ref);
1365                         ptr = (unsigned long)(ref + 1) + name_len;
1366                         nlink++;
1367                 }
1368
1369                 if (key.offset == 0)
1370                         break;
1371                 if (path->slots[0] > 0) {
1372                         path->slots[0]--;
1373                         goto process_slot;
1374                 }
1375                 key.offset--;
1376                 btrfs_release_path(path);
1377         }
1378         btrfs_release_path(path);
1379
1380         return nlink;
1381 }
1382
1383 /*
1384  * There are a few corners where the link count of the file can't
1385  * be properly maintained during replay.  So, instead of adding
1386  * lots of complexity to the log code, we just scan the backrefs
1387  * for any file that has been through replay.
1388  *
1389  * The scan will update the link count on the inode to reflect the
1390  * number of back refs found.  If it goes down to zero, the iput
1391  * will free the inode.
1392  */
1393 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1394                                            struct btrfs_root *root,
1395                                            struct inode *inode)
1396 {
1397         struct btrfs_path *path;
1398         int ret;
1399         u64 nlink = 0;
1400         u64 ino = btrfs_ino(inode);
1401
1402         path = btrfs_alloc_path();
1403         if (!path)
1404                 return -ENOMEM;
1405
1406         ret = count_inode_refs(root, inode, path);
1407         if (ret < 0)
1408                 goto out;
1409
1410         nlink = ret;
1411
1412         ret = count_inode_extrefs(root, inode, path);
1413         if (ret < 0)
1414                 goto out;
1415
1416         nlink += ret;
1417
1418         ret = 0;
1419
1420         if (nlink != inode->i_nlink) {
1421                 set_nlink(inode, nlink);
1422                 btrfs_update_inode(trans, root, inode);
1423         }
1424         BTRFS_I(inode)->index_cnt = (u64)-1;
1425
1426         if (inode->i_nlink == 0) {
1427                 if (S_ISDIR(inode->i_mode)) {
1428                         ret = replay_dir_deletes(trans, root, NULL, path,
1429                                                  ino, 1);
1430                         if (ret)
1431                                 goto out;
1432                 }
1433                 ret = insert_orphan_item(trans, root, ino);
1434         }
1435
1436 out:
1437         btrfs_free_path(path);
1438         return ret;
1439 }
1440
1441 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1442                                             struct btrfs_root *root,
1443                                             struct btrfs_path *path)
1444 {
1445         int ret;
1446         struct btrfs_key key;
1447         struct inode *inode;
1448
1449         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1450         key.type = BTRFS_ORPHAN_ITEM_KEY;
1451         key.offset = (u64)-1;
1452         while (1) {
1453                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1454                 if (ret < 0)
1455                         break;
1456
1457                 if (ret == 1) {
1458                         if (path->slots[0] == 0)
1459                                 break;
1460                         path->slots[0]--;
1461                 }
1462
1463                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1464                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1465                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1466                         break;
1467
1468                 ret = btrfs_del_item(trans, root, path);
1469                 if (ret)
1470                         goto out;
1471
1472                 btrfs_release_path(path);
1473                 inode = read_one_inode(root, key.offset);
1474                 if (!inode)
1475                         return -EIO;
1476
1477                 ret = fixup_inode_link_count(trans, root, inode);
1478                 iput(inode);
1479                 if (ret)
1480                         goto out;
1481
1482                 /*
1483                  * fixup on a directory may create new entries,
1484                  * make sure we always look for the highset possible
1485                  * offset
1486                  */
1487                 key.offset = (u64)-1;
1488         }
1489         ret = 0;
1490 out:
1491         btrfs_release_path(path);
1492         return ret;
1493 }
1494
1495
1496 /*
1497  * record a given inode in the fixup dir so we can check its link
1498  * count when replay is done.  The link count is incremented here
1499  * so the inode won't go away until we check it
1500  */
1501 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1502                                       struct btrfs_root *root,
1503                                       struct btrfs_path *path,
1504                                       u64 objectid)
1505 {
1506         struct btrfs_key key;
1507         int ret = 0;
1508         struct inode *inode;
1509
1510         inode = read_one_inode(root, objectid);
1511         if (!inode)
1512                 return -EIO;
1513
1514         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1515         key.type = BTRFS_ORPHAN_ITEM_KEY;
1516         key.offset = objectid;
1517
1518         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1519
1520         btrfs_release_path(path);
1521         if (ret == 0) {
1522                 if (!inode->i_nlink)
1523                         set_nlink(inode, 1);
1524                 else
1525                         inc_nlink(inode);
1526                 ret = btrfs_update_inode(trans, root, inode);
1527         } else if (ret == -EEXIST) {
1528                 ret = 0;
1529         } else {
1530                 BUG(); /* Logic Error */
1531         }
1532         iput(inode);
1533
1534         return ret;
1535 }
1536
1537 /*
1538  * when replaying the log for a directory, we only insert names
1539  * for inodes that actually exist.  This means an fsync on a directory
1540  * does not implicitly fsync all the new files in it
1541  */
1542 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1543                                     struct btrfs_root *root,
1544                                     struct btrfs_path *path,
1545                                     u64 dirid, u64 index,
1546                                     char *name, int name_len, u8 type,
1547                                     struct btrfs_key *location)
1548 {
1549         struct inode *inode;
1550         struct inode *dir;
1551         int ret;
1552
1553         inode = read_one_inode(root, location->objectid);
1554         if (!inode)
1555                 return -ENOENT;
1556
1557         dir = read_one_inode(root, dirid);
1558         if (!dir) {
1559                 iput(inode);
1560                 return -EIO;
1561         }
1562
1563         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1564
1565         /* FIXME, put inode into FIXUP list */
1566
1567         iput(inode);
1568         iput(dir);
1569         return ret;
1570 }
1571
1572 /*
1573  * Return true if an inode reference exists in the log for the given name,
1574  * inode and parent inode.
1575  */
1576 static bool name_in_log_ref(struct btrfs_root *log_root,
1577                             const char *name, const int name_len,
1578                             const u64 dirid, const u64 ino)
1579 {
1580         struct btrfs_key search_key;
1581
1582         search_key.objectid = ino;
1583         search_key.type = BTRFS_INODE_REF_KEY;
1584         search_key.offset = dirid;
1585         if (backref_in_log(log_root, &search_key, dirid, name, name_len))
1586                 return true;
1587
1588         search_key.type = BTRFS_INODE_EXTREF_KEY;
1589         search_key.offset = btrfs_extref_hash(dirid, name, name_len);
1590         if (backref_in_log(log_root, &search_key, dirid, name, name_len))
1591                 return true;
1592
1593         return false;
1594 }
1595
1596 /*
1597  * take a single entry in a log directory item and replay it into
1598  * the subvolume.
1599  *
1600  * if a conflicting item exists in the subdirectory already,
1601  * the inode it points to is unlinked and put into the link count
1602  * fix up tree.
1603  *
1604  * If a name from the log points to a file or directory that does
1605  * not exist in the FS, it is skipped.  fsyncs on directories
1606  * do not force down inodes inside that directory, just changes to the
1607  * names or unlinks in a directory.
1608  */
1609 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1610                                     struct btrfs_root *root,
1611                                     struct btrfs_path *path,
1612                                     struct extent_buffer *eb,
1613                                     struct btrfs_dir_item *di,
1614                                     struct btrfs_key *key)
1615 {
1616         char *name;
1617         int name_len;
1618         struct btrfs_dir_item *dst_di;
1619         struct btrfs_key found_key;
1620         struct btrfs_key log_key;
1621         struct inode *dir;
1622         u8 log_type;
1623         int exists;
1624         int ret = 0;
1625         bool update_size = (key->type == BTRFS_DIR_INDEX_KEY);
1626
1627         dir = read_one_inode(root, key->objectid);
1628         if (!dir)
1629                 return -EIO;
1630
1631         name_len = btrfs_dir_name_len(eb, di);
1632         name = kmalloc(name_len, GFP_NOFS);
1633         if (!name) {
1634                 ret = -ENOMEM;
1635                 goto out;
1636         }
1637
1638         log_type = btrfs_dir_type(eb, di);
1639         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1640                    name_len);
1641
1642         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1643         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1644         if (exists == 0)
1645                 exists = 1;
1646         else
1647                 exists = 0;
1648         btrfs_release_path(path);
1649
1650         if (key->type == BTRFS_DIR_ITEM_KEY) {
1651                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1652                                        name, name_len, 1);
1653         } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1654                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1655                                                      key->objectid,
1656                                                      key->offset, name,
1657                                                      name_len, 1);
1658         } else {
1659                 /* Corruption */
1660                 ret = -EINVAL;
1661                 goto out;
1662         }
1663         if (IS_ERR_OR_NULL(dst_di)) {
1664                 /* we need a sequence number to insert, so we only
1665                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1666                  */
1667                 if (key->type != BTRFS_DIR_INDEX_KEY)
1668                         goto out;
1669                 goto insert;
1670         }
1671
1672         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1673         /* the existing item matches the logged item */
1674         if (found_key.objectid == log_key.objectid &&
1675             found_key.type == log_key.type &&
1676             found_key.offset == log_key.offset &&
1677             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1678                 update_size = false;
1679                 goto out;
1680         }
1681
1682         /*
1683          * don't drop the conflicting directory entry if the inode
1684          * for the new entry doesn't exist
1685          */
1686         if (!exists)
1687                 goto out;
1688
1689         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1690         if (ret)
1691                 goto out;
1692
1693         if (key->type == BTRFS_DIR_INDEX_KEY)
1694                 goto insert;
1695 out:
1696         btrfs_release_path(path);
1697         if (!ret && update_size) {
1698                 btrfs_i_size_write(dir, dir->i_size + name_len * 2);
1699                 ret = btrfs_update_inode(trans, root, dir);
1700         }
1701         kfree(name);
1702         iput(dir);
1703         return ret;
1704
1705 insert:
1706         if (name_in_log_ref(root->log_root, name, name_len,
1707                             key->objectid, log_key.objectid)) {
1708                 /* The dentry will be added later. */
1709                 ret = 0;
1710                 update_size = false;
1711                 goto out;
1712         }
1713         btrfs_release_path(path);
1714         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1715                               name, name_len, log_type, &log_key);
1716         if (ret && ret != -ENOENT && ret != -EEXIST)
1717                 goto out;
1718         update_size = false;
1719         ret = 0;
1720         goto out;
1721 }
1722
1723 /*
1724  * find all the names in a directory item and reconcile them into
1725  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1726  * one name in a directory item, but the same code gets used for
1727  * both directory index types
1728  */
1729 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1730                                         struct btrfs_root *root,
1731                                         struct btrfs_path *path,
1732                                         struct extent_buffer *eb, int slot,
1733                                         struct btrfs_key *key)
1734 {
1735         int ret;
1736         u32 item_size = btrfs_item_size_nr(eb, slot);
1737         struct btrfs_dir_item *di;
1738         int name_len;
1739         unsigned long ptr;
1740         unsigned long ptr_end;
1741
1742         ptr = btrfs_item_ptr_offset(eb, slot);
1743         ptr_end = ptr + item_size;
1744         while (ptr < ptr_end) {
1745                 di = (struct btrfs_dir_item *)ptr;
1746                 if (verify_dir_item(root, eb, di))
1747                         return -EIO;
1748                 name_len = btrfs_dir_name_len(eb, di);
1749                 ret = replay_one_name(trans, root, path, eb, di, key);
1750                 if (ret)
1751                         return ret;
1752                 ptr = (unsigned long)(di + 1);
1753                 ptr += name_len;
1754         }
1755         return 0;
1756 }
1757
1758 /*
1759  * directory replay has two parts.  There are the standard directory
1760  * items in the log copied from the subvolume, and range items
1761  * created in the log while the subvolume was logged.
1762  *
1763  * The range items tell us which parts of the key space the log
1764  * is authoritative for.  During replay, if a key in the subvolume
1765  * directory is in a logged range item, but not actually in the log
1766  * that means it was deleted from the directory before the fsync
1767  * and should be removed.
1768  */
1769 static noinline int find_dir_range(struct btrfs_root *root,
1770                                    struct btrfs_path *path,
1771                                    u64 dirid, int key_type,
1772                                    u64 *start_ret, u64 *end_ret)
1773 {
1774         struct btrfs_key key;
1775         u64 found_end;
1776         struct btrfs_dir_log_item *item;
1777         int ret;
1778         int nritems;
1779
1780         if (*start_ret == (u64)-1)
1781                 return 1;
1782
1783         key.objectid = dirid;
1784         key.type = key_type;
1785         key.offset = *start_ret;
1786
1787         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1788         if (ret < 0)
1789                 goto out;
1790         if (ret > 0) {
1791                 if (path->slots[0] == 0)
1792                         goto out;
1793                 path->slots[0]--;
1794         }
1795         if (ret != 0)
1796                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1797
1798         if (key.type != key_type || key.objectid != dirid) {
1799                 ret = 1;
1800                 goto next;
1801         }
1802         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1803                               struct btrfs_dir_log_item);
1804         found_end = btrfs_dir_log_end(path->nodes[0], item);
1805
1806         if (*start_ret >= key.offset && *start_ret <= found_end) {
1807                 ret = 0;
1808                 *start_ret = key.offset;
1809                 *end_ret = found_end;
1810                 goto out;
1811         }
1812         ret = 1;
1813 next:
1814         /* check the next slot in the tree to see if it is a valid item */
1815         nritems = btrfs_header_nritems(path->nodes[0]);
1816         if (path->slots[0] >= nritems) {
1817                 ret = btrfs_next_leaf(root, path);
1818                 if (ret)
1819                         goto out;
1820         } else {
1821                 path->slots[0]++;
1822         }
1823
1824         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1825
1826         if (key.type != key_type || key.objectid != dirid) {
1827                 ret = 1;
1828                 goto out;
1829         }
1830         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1831                               struct btrfs_dir_log_item);
1832         found_end = btrfs_dir_log_end(path->nodes[0], item);
1833         *start_ret = key.offset;
1834         *end_ret = found_end;
1835         ret = 0;
1836 out:
1837         btrfs_release_path(path);
1838         return ret;
1839 }
1840
1841 /*
1842  * this looks for a given directory item in the log.  If the directory
1843  * item is not in the log, the item is removed and the inode it points
1844  * to is unlinked
1845  */
1846 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1847                                       struct btrfs_root *root,
1848                                       struct btrfs_root *log,
1849                                       struct btrfs_path *path,
1850                                       struct btrfs_path *log_path,
1851                                       struct inode *dir,
1852                                       struct btrfs_key *dir_key)
1853 {
1854         int ret;
1855         struct extent_buffer *eb;
1856         int slot;
1857         u32 item_size;
1858         struct btrfs_dir_item *di;
1859         struct btrfs_dir_item *log_di;
1860         int name_len;
1861         unsigned long ptr;
1862         unsigned long ptr_end;
1863         char *name;
1864         struct inode *inode;
1865         struct btrfs_key location;
1866
1867 again:
1868         eb = path->nodes[0];
1869         slot = path->slots[0];
1870         item_size = btrfs_item_size_nr(eb, slot);
1871         ptr = btrfs_item_ptr_offset(eb, slot);
1872         ptr_end = ptr + item_size;
1873         while (ptr < ptr_end) {
1874                 di = (struct btrfs_dir_item *)ptr;
1875                 if (verify_dir_item(root, eb, di)) {
1876                         ret = -EIO;
1877                         goto out;
1878                 }
1879
1880                 name_len = btrfs_dir_name_len(eb, di);
1881                 name = kmalloc(name_len, GFP_NOFS);
1882                 if (!name) {
1883                         ret = -ENOMEM;
1884                         goto out;
1885                 }
1886                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1887                                   name_len);
1888                 log_di = NULL;
1889                 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
1890                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1891                                                        dir_key->objectid,
1892                                                        name, name_len, 0);
1893                 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
1894                         log_di = btrfs_lookup_dir_index_item(trans, log,
1895                                                      log_path,
1896                                                      dir_key->objectid,
1897                                                      dir_key->offset,
1898                                                      name, name_len, 0);
1899                 }
1900                 if (!log_di || (IS_ERR(log_di) && PTR_ERR(log_di) == -ENOENT)) {
1901                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1902                         btrfs_release_path(path);
1903                         btrfs_release_path(log_path);
1904                         inode = read_one_inode(root, location.objectid);
1905                         if (!inode) {
1906                                 kfree(name);
1907                                 return -EIO;
1908                         }
1909
1910                         ret = link_to_fixup_dir(trans, root,
1911                                                 path, location.objectid);
1912                         if (ret) {
1913                                 kfree(name);
1914                                 iput(inode);
1915                                 goto out;
1916                         }
1917
1918                         inc_nlink(inode);
1919                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1920                                                  name, name_len);
1921                         if (!ret)
1922                                 ret = btrfs_run_delayed_items(trans, root);
1923                         kfree(name);
1924                         iput(inode);
1925                         if (ret)
1926                                 goto out;
1927
1928                         /* there might still be more names under this key
1929                          * check and repeat if required
1930                          */
1931                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1932                                                 0, 0);
1933                         if (ret == 0)
1934                                 goto again;
1935                         ret = 0;
1936                         goto out;
1937                 } else if (IS_ERR(log_di)) {
1938                         kfree(name);
1939                         return PTR_ERR(log_di);
1940                 }
1941                 btrfs_release_path(log_path);
1942                 kfree(name);
1943
1944                 ptr = (unsigned long)(di + 1);
1945                 ptr += name_len;
1946         }
1947         ret = 0;
1948 out:
1949         btrfs_release_path(path);
1950         btrfs_release_path(log_path);
1951         return ret;
1952 }
1953
1954 /*
1955  * deletion replay happens before we copy any new directory items
1956  * out of the log or out of backreferences from inodes.  It
1957  * scans the log to find ranges of keys that log is authoritative for,
1958  * and then scans the directory to find items in those ranges that are
1959  * not present in the log.
1960  *
1961  * Anything we don't find in the log is unlinked and removed from the
1962  * directory.
1963  */
1964 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1965                                        struct btrfs_root *root,
1966                                        struct btrfs_root *log,
1967                                        struct btrfs_path *path,
1968                                        u64 dirid, int del_all)
1969 {
1970         u64 range_start;
1971         u64 range_end;
1972         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1973         int ret = 0;
1974         struct btrfs_key dir_key;
1975         struct btrfs_key found_key;
1976         struct btrfs_path *log_path;
1977         struct inode *dir;
1978
1979         dir_key.objectid = dirid;
1980         dir_key.type = BTRFS_DIR_ITEM_KEY;
1981         log_path = btrfs_alloc_path();
1982         if (!log_path)
1983                 return -ENOMEM;
1984
1985         dir = read_one_inode(root, dirid);
1986         /* it isn't an error if the inode isn't there, that can happen
1987          * because we replay the deletes before we copy in the inode item
1988          * from the log
1989          */
1990         if (!dir) {
1991                 btrfs_free_path(log_path);
1992                 return 0;
1993         }
1994 again:
1995         range_start = 0;
1996         range_end = 0;
1997         while (1) {
1998                 if (del_all)
1999                         range_end = (u64)-1;
2000                 else {
2001                         ret = find_dir_range(log, path, dirid, key_type,
2002                                              &range_start, &range_end);
2003                         if (ret != 0)
2004                                 break;
2005                 }
2006
2007                 dir_key.offset = range_start;
2008                 while (1) {
2009                         int nritems;
2010                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
2011                                                 0, 0);
2012                         if (ret < 0)
2013                                 goto out;
2014
2015                         nritems = btrfs_header_nritems(path->nodes[0]);
2016                         if (path->slots[0] >= nritems) {
2017                                 ret = btrfs_next_leaf(root, path);
2018                                 if (ret)
2019                                         break;
2020                         }
2021                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2022                                               path->slots[0]);
2023                         if (found_key.objectid != dirid ||
2024                             found_key.type != dir_key.type)
2025                                 goto next_type;
2026
2027                         if (found_key.offset > range_end)
2028                                 break;
2029
2030                         ret = check_item_in_log(trans, root, log, path,
2031                                                 log_path, dir,
2032                                                 &found_key);
2033                         if (ret)
2034                                 goto out;
2035                         if (found_key.offset == (u64)-1)
2036                                 break;
2037                         dir_key.offset = found_key.offset + 1;
2038                 }
2039                 btrfs_release_path(path);
2040                 if (range_end == (u64)-1)
2041                         break;
2042                 range_start = range_end + 1;
2043         }
2044
2045 next_type:
2046         ret = 0;
2047         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
2048                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
2049                 dir_key.type = BTRFS_DIR_INDEX_KEY;
2050                 btrfs_release_path(path);
2051                 goto again;
2052         }
2053 out:
2054         btrfs_release_path(path);
2055         btrfs_free_path(log_path);
2056         iput(dir);
2057         return ret;
2058 }
2059
2060 /*
2061  * the process_func used to replay items from the log tree.  This
2062  * gets called in two different stages.  The first stage just looks
2063  * for inodes and makes sure they are all copied into the subvolume.
2064  *
2065  * The second stage copies all the other item types from the log into
2066  * the subvolume.  The two stage approach is slower, but gets rid of
2067  * lots of complexity around inodes referencing other inodes that exist
2068  * only in the log (references come from either directory items or inode
2069  * back refs).
2070  */
2071 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
2072                              struct walk_control *wc, u64 gen)
2073 {
2074         int nritems;
2075         struct btrfs_path *path;
2076         struct btrfs_root *root = wc->replay_dest;
2077         struct btrfs_key key;
2078         int level;
2079         int i;
2080         int ret;
2081
2082         ret = btrfs_read_buffer(eb, gen);
2083         if (ret)
2084                 return ret;
2085
2086         level = btrfs_header_level(eb);
2087
2088         if (level != 0)
2089                 return 0;
2090
2091         path = btrfs_alloc_path();
2092         if (!path)
2093                 return -ENOMEM;
2094
2095         nritems = btrfs_header_nritems(eb);
2096         for (i = 0; i < nritems; i++) {
2097                 btrfs_item_key_to_cpu(eb, &key, i);
2098
2099                 /* inode keys are done during the first stage */
2100                 if (key.type == BTRFS_INODE_ITEM_KEY &&
2101                     wc->stage == LOG_WALK_REPLAY_INODES) {
2102                         struct btrfs_inode_item *inode_item;
2103                         u32 mode;
2104
2105                         inode_item = btrfs_item_ptr(eb, i,
2106                                             struct btrfs_inode_item);
2107                         mode = btrfs_inode_mode(eb, inode_item);
2108                         if (S_ISDIR(mode)) {
2109                                 ret = replay_dir_deletes(wc->trans,
2110                                          root, log, path, key.objectid, 0);
2111                                 if (ret)
2112                                         break;
2113                         }
2114                         ret = overwrite_item(wc->trans, root, path,
2115                                              eb, i, &key);
2116                         if (ret)
2117                                 break;
2118
2119                         /* for regular files, make sure corresponding
2120                          * orhpan item exist. extents past the new EOF
2121                          * will be truncated later by orphan cleanup.
2122                          */
2123                         if (S_ISREG(mode)) {
2124                                 ret = insert_orphan_item(wc->trans, root,
2125                                                          key.objectid);
2126                                 if (ret)
2127                                         break;
2128                         }
2129
2130                         ret = link_to_fixup_dir(wc->trans, root,
2131                                                 path, key.objectid);
2132                         if (ret)
2133                                 break;
2134                 }
2135
2136                 if (key.type == BTRFS_DIR_INDEX_KEY &&
2137                     wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
2138                         ret = replay_one_dir_item(wc->trans, root, path,
2139                                                   eb, i, &key);
2140                         if (ret)
2141                                 break;
2142                 }
2143
2144                 if (wc->stage < LOG_WALK_REPLAY_ALL)
2145                         continue;
2146
2147                 /* these keys are simply copied */
2148                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
2149                         ret = overwrite_item(wc->trans, root, path,
2150                                              eb, i, &key);
2151                         if (ret)
2152                                 break;
2153                 } else if (key.type == BTRFS_INODE_REF_KEY ||
2154                            key.type == BTRFS_INODE_EXTREF_KEY) {
2155                         ret = add_inode_ref(wc->trans, root, log, path,
2156                                             eb, i, &key);
2157                         if (ret && ret != -ENOENT)
2158                                 break;
2159                         ret = 0;
2160                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
2161                         ret = replay_one_extent(wc->trans, root, path,
2162                                                 eb, i, &key);
2163                         if (ret)
2164                                 break;
2165                 } else if (key.type == BTRFS_DIR_ITEM_KEY) {
2166                         ret = replay_one_dir_item(wc->trans, root, path,
2167                                                   eb, i, &key);
2168                         if (ret)
2169                                 break;
2170                 }
2171         }
2172         btrfs_free_path(path);
2173         return ret;
2174 }
2175
2176 static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
2177                                    struct btrfs_root *root,
2178                                    struct btrfs_path *path, int *level,
2179                                    struct walk_control *wc)
2180 {
2181         u64 root_owner;
2182         u64 bytenr;
2183         u64 ptr_gen;
2184         struct extent_buffer *next;
2185         struct extent_buffer *cur;
2186         struct extent_buffer *parent;
2187         u32 blocksize;
2188         int ret = 0;
2189
2190         WARN_ON(*level < 0);
2191         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2192
2193         while (*level > 0) {
2194                 WARN_ON(*level < 0);
2195                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
2196                 cur = path->nodes[*level];
2197
2198                 WARN_ON(btrfs_header_level(cur) != *level);
2199
2200                 if (path->slots[*level] >=
2201                     btrfs_header_nritems(cur))
2202                         break;
2203
2204                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
2205                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
2206                 blocksize = root->nodesize;
2207
2208                 parent = path->nodes[*level];
2209                 root_owner = btrfs_header_owner(parent);
2210
2211                 next = btrfs_find_create_tree_block(root, bytenr);
2212                 if (!next)
2213                         return -ENOMEM;
2214
2215                 if (*level == 1) {
2216                         ret = wc->process_func(root, next, wc, ptr_gen);
2217                         if (ret) {
2218                                 free_extent_buffer(next);
2219                                 return ret;
2220                         }
2221
2222                         path->slots[*level]++;
2223                         if (wc->free) {
2224                                 ret = btrfs_read_buffer(next, ptr_gen);
2225                                 if (ret) {
2226                                         free_extent_buffer(next);
2227                                         return ret;
2228                                 }
2229
2230                                 if (trans) {
2231                                         btrfs_tree_lock(next);
2232                                         btrfs_set_lock_blocking(next);
2233                                         clean_tree_block(trans, root, next);
2234                                         btrfs_wait_tree_block_writeback(next);
2235                                         btrfs_tree_unlock(next);
2236                                 }
2237
2238                                 WARN_ON(root_owner !=
2239                                         BTRFS_TREE_LOG_OBJECTID);
2240                                 ret = btrfs_free_and_pin_reserved_extent(root,
2241                                                          bytenr, blocksize);
2242                                 if (ret) {
2243                                         free_extent_buffer(next);
2244                                         return ret;
2245                                 }
2246                         }
2247                         free_extent_buffer(next);
2248                         continue;
2249                 }
2250                 ret = btrfs_read_buffer(next, ptr_gen);
2251                 if (ret) {
2252                         free_extent_buffer(next);
2253                         return ret;
2254                 }
2255
2256                 WARN_ON(*level <= 0);
2257                 if (path->nodes[*level-1])
2258                         free_extent_buffer(path->nodes[*level-1]);
2259                 path->nodes[*level-1] = next;
2260                 *level = btrfs_header_level(next);
2261                 path->slots[*level] = 0;
2262                 cond_resched();
2263         }
2264         WARN_ON(*level < 0);
2265         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2266
2267         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2268
2269         cond_resched();
2270         return 0;
2271 }
2272
2273 static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2274                                  struct btrfs_root *root,
2275                                  struct btrfs_path *path, int *level,
2276                                  struct walk_control *wc)
2277 {
2278         u64 root_owner;
2279         int i;
2280         int slot;
2281         int ret;
2282
2283         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2284                 slot = path->slots[i];
2285                 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2286                         path->slots[i]++;
2287                         *level = i;
2288                         WARN_ON(*level == 0);
2289                         return 0;
2290                 } else {
2291                         struct extent_buffer *parent;
2292                         if (path->nodes[*level] == root->node)
2293                                 parent = path->nodes[*level];
2294                         else
2295                                 parent = path->nodes[*level + 1];
2296
2297                         root_owner = btrfs_header_owner(parent);
2298                         ret = wc->process_func(root, path->nodes[*level], wc,
2299                                  btrfs_header_generation(path->nodes[*level]));
2300                         if (ret)
2301                                 return ret;
2302
2303                         if (wc->free) {
2304                                 struct extent_buffer *next;
2305
2306                                 next = path->nodes[*level];
2307
2308                                 if (trans) {
2309                                         btrfs_tree_lock(next);
2310                                         btrfs_set_lock_blocking(next);
2311                                         clean_tree_block(trans, root, next);
2312                                         btrfs_wait_tree_block_writeback(next);
2313                                         btrfs_tree_unlock(next);
2314                                 }
2315
2316                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
2317                                 ret = btrfs_free_and_pin_reserved_extent(root,
2318                                                 path->nodes[*level]->start,
2319                                                 path->nodes[*level]->len);
2320                                 if (ret)
2321                                         return ret;
2322                         }
2323                         free_extent_buffer(path->nodes[*level]);
2324                         path->nodes[*level] = NULL;
2325                         *level = i + 1;
2326                 }
2327         }
2328         return 1;
2329 }
2330
2331 /*
2332  * drop the reference count on the tree rooted at 'snap'.  This traverses
2333  * the tree freeing any blocks that have a ref count of zero after being
2334  * decremented.
2335  */
2336 static int walk_log_tree(struct btrfs_trans_handle *trans,
2337                          struct btrfs_root *log, struct walk_control *wc)
2338 {
2339         int ret = 0;
2340         int wret;
2341         int level;
2342         struct btrfs_path *path;
2343         int orig_level;
2344
2345         path = btrfs_alloc_path();
2346         if (!path)
2347                 return -ENOMEM;
2348
2349         level = btrfs_header_level(log->node);
2350         orig_level = level;
2351         path->nodes[level] = log->node;
2352         extent_buffer_get(log->node);
2353         path->slots[level] = 0;
2354
2355         while (1) {
2356                 wret = walk_down_log_tree(trans, log, path, &level, wc);
2357                 if (wret > 0)
2358                         break;
2359                 if (wret < 0) {
2360                         ret = wret;
2361                         goto out;
2362                 }
2363
2364                 wret = walk_up_log_tree(trans, log, path, &level, wc);
2365                 if (wret > 0)
2366                         break;
2367                 if (wret < 0) {
2368                         ret = wret;
2369                         goto out;
2370                 }
2371         }
2372
2373         /* was the root node processed? if not, catch it here */
2374         if (path->nodes[orig_level]) {
2375                 ret = wc->process_func(log, path->nodes[orig_level], wc,
2376                          btrfs_header_generation(path->nodes[orig_level]));
2377                 if (ret)
2378                         goto out;
2379                 if (wc->free) {
2380                         struct extent_buffer *next;
2381
2382                         next = path->nodes[orig_level];
2383
2384                         if (trans) {
2385                                 btrfs_tree_lock(next);
2386                                 btrfs_set_lock_blocking(next);
2387                                 clean_tree_block(trans, log, next);
2388                                 btrfs_wait_tree_block_writeback(next);
2389                                 btrfs_tree_unlock(next);
2390                         }
2391
2392                         WARN_ON(log->root_key.objectid !=
2393                                 BTRFS_TREE_LOG_OBJECTID);
2394                         ret = btrfs_free_and_pin_reserved_extent(log, next->start,
2395                                                          next->len);
2396                         if (ret)
2397                                 goto out;
2398                 }
2399         }
2400
2401 out:
2402         btrfs_free_path(path);
2403         return ret;
2404 }
2405
2406 /*
2407  * helper function to update the item for a given subvolumes log root
2408  * in the tree of log roots
2409  */
2410 static int update_log_root(struct btrfs_trans_handle *trans,
2411                            struct btrfs_root *log)
2412 {
2413         int ret;
2414
2415         if (log->log_transid == 1) {
2416                 /* insert root item on the first sync */
2417                 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
2418                                 &log->root_key, &log->root_item);
2419         } else {
2420                 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2421                                 &log->root_key, &log->root_item);
2422         }
2423         return ret;
2424 }
2425
2426 static void wait_log_commit(struct btrfs_trans_handle *trans,
2427                             struct btrfs_root *root, int transid)
2428 {
2429         DEFINE_WAIT(wait);
2430         int index = transid % 2;
2431
2432         /*
2433          * we only allow two pending log transactions at a time,
2434          * so we know that if ours is more than 2 older than the
2435          * current transaction, we're done
2436          */
2437         do {
2438                 prepare_to_wait(&root->log_commit_wait[index],
2439                                 &wait, TASK_UNINTERRUPTIBLE);
2440                 mutex_unlock(&root->log_mutex);
2441
2442                 if (root->log_transid_committed < transid &&
2443                     atomic_read(&root->log_commit[index]))
2444                         schedule();
2445
2446                 finish_wait(&root->log_commit_wait[index], &wait);
2447                 mutex_lock(&root->log_mutex);
2448         } while (root->log_transid_committed < transid &&
2449                  atomic_read(&root->log_commit[index]));
2450 }
2451
2452 static void wait_for_writer(struct btrfs_trans_handle *trans,
2453                             struct btrfs_root *root)
2454 {
2455         DEFINE_WAIT(wait);
2456
2457         while (atomic_read(&root->log_writers)) {
2458                 prepare_to_wait(&root->log_writer_wait,
2459                                 &wait, TASK_UNINTERRUPTIBLE);
2460                 mutex_unlock(&root->log_mutex);
2461                 if (atomic_read(&root->log_writers))
2462                         schedule();
2463                 finish_wait(&root->log_writer_wait, &wait);
2464                 mutex_lock(&root->log_mutex);
2465         }
2466 }
2467
2468 static inline void btrfs_remove_log_ctx(struct btrfs_root *root,
2469                                         struct btrfs_log_ctx *ctx)
2470 {
2471         if (!ctx)
2472                 return;
2473
2474         mutex_lock(&root->log_mutex);
2475         list_del_init(&ctx->list);
2476         mutex_unlock(&root->log_mutex);
2477 }
2478
2479 /*
2480  * Invoked in log mutex context, or be sure there is no other task which
2481  * can access the list.
2482  */
2483 static inline void btrfs_remove_all_log_ctxs(struct btrfs_root *root,
2484                                              int index, int error)
2485 {
2486         struct btrfs_log_ctx *ctx;
2487
2488         if (!error) {
2489                 INIT_LIST_HEAD(&root->log_ctxs[index]);
2490                 return;
2491         }
2492
2493         list_for_each_entry(ctx, &root->log_ctxs[index], list)
2494                 ctx->log_ret = error;
2495
2496         INIT_LIST_HEAD(&root->log_ctxs[index]);
2497 }
2498
2499 /*
2500  * btrfs_sync_log does sends a given tree log down to the disk and
2501  * updates the super blocks to record it.  When this call is done,
2502  * you know that any inodes previously logged are safely on disk only
2503  * if it returns 0.
2504  *
2505  * Any other return value means you need to call btrfs_commit_transaction.
2506  * Some of the edge cases for fsyncing directories that have had unlinks
2507  * or renames done in the past mean that sometimes the only safe
2508  * fsync is to commit the whole FS.  When btrfs_sync_log returns -EAGAIN,
2509  * that has happened.
2510  */
2511 int btrfs_sync_log(struct btrfs_trans_handle *trans,
2512                    struct btrfs_root *root, struct btrfs_log_ctx *ctx)
2513 {
2514         int index1;
2515         int index2;
2516         int mark;
2517         int ret;
2518         struct btrfs_root *log = root->log_root;
2519         struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
2520         int log_transid = 0;
2521         struct btrfs_log_ctx root_log_ctx;
2522         struct blk_plug plug;
2523
2524         mutex_lock(&root->log_mutex);
2525         log_transid = ctx->log_transid;
2526         if (root->log_transid_committed >= log_transid) {
2527                 mutex_unlock(&root->log_mutex);
2528                 return ctx->log_ret;
2529         }
2530
2531         index1 = log_transid % 2;
2532         if (atomic_read(&root->log_commit[index1])) {
2533                 wait_log_commit(trans, root, log_transid);
2534                 mutex_unlock(&root->log_mutex);
2535                 return ctx->log_ret;
2536         }
2537         ASSERT(log_transid == root->log_transid);
2538         atomic_set(&root->log_commit[index1], 1);
2539
2540         /* wait for previous tree log sync to complete */
2541         if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2542                 wait_log_commit(trans, root, log_transid - 1);
2543
2544         while (1) {
2545                 int batch = atomic_read(&root->log_batch);
2546                 /* when we're on an ssd, just kick the log commit out */
2547                 if (!btrfs_test_opt(root, SSD) &&
2548                     test_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state)) {
2549                         mutex_unlock(&root->log_mutex);
2550                         schedule_timeout_uninterruptible(1);
2551                         mutex_lock(&root->log_mutex);
2552                 }
2553                 wait_for_writer(trans, root);
2554                 if (batch == atomic_read(&root->log_batch))
2555                         break;
2556         }
2557
2558         /* bail out if we need to do a full commit */
2559         if (btrfs_need_log_full_commit(root->fs_info, trans)) {
2560                 ret = -EAGAIN;
2561                 btrfs_free_logged_extents(log, log_transid);
2562                 mutex_unlock(&root->log_mutex);
2563                 goto out;
2564         }
2565
2566         if (log_transid % 2 == 0)
2567                 mark = EXTENT_DIRTY;
2568         else
2569                 mark = EXTENT_NEW;
2570
2571         /* we start IO on  all the marked extents here, but we don't actually
2572          * wait for them until later.
2573          */
2574         blk_start_plug(&plug);
2575         ret = btrfs_write_marked_extents(log, &log->dirty_log_pages, mark);
2576         if (ret) {
2577                 blk_finish_plug(&plug);
2578                 btrfs_abort_transaction(trans, root, ret);
2579                 btrfs_free_logged_extents(log, log_transid);
2580                 btrfs_set_log_full_commit(root->fs_info, trans);
2581                 mutex_unlock(&root->log_mutex);
2582                 goto out;
2583         }
2584
2585         btrfs_set_root_node(&log->root_item, log->node);
2586
2587         root->log_transid++;
2588         log->log_transid = root->log_transid;
2589         root->log_start_pid = 0;
2590         /*
2591          * IO has been started, blocks of the log tree have WRITTEN flag set
2592          * in their headers. new modifications of the log will be written to
2593          * new positions. so it's safe to allow log writers to go in.
2594          */
2595         mutex_unlock(&root->log_mutex);
2596
2597         btrfs_init_log_ctx(&root_log_ctx);
2598
2599         mutex_lock(&log_root_tree->log_mutex);
2600         atomic_inc(&log_root_tree->log_batch);
2601         atomic_inc(&log_root_tree->log_writers);
2602
2603         index2 = log_root_tree->log_transid % 2;
2604         list_add_tail(&root_log_ctx.list, &log_root_tree->log_ctxs[index2]);
2605         root_log_ctx.log_transid = log_root_tree->log_transid;
2606
2607         mutex_unlock(&log_root_tree->log_mutex);
2608
2609         ret = update_log_root(trans, log);
2610
2611         mutex_lock(&log_root_tree->log_mutex);
2612         if (atomic_dec_and_test(&log_root_tree->log_writers)) {
2613                 smp_mb();
2614                 if (waitqueue_active(&log_root_tree->log_writer_wait))
2615                         wake_up(&log_root_tree->log_writer_wait);
2616         }
2617
2618         if (ret) {
2619                 if (!list_empty(&root_log_ctx.list))
2620                         list_del_init(&root_log_ctx.list);
2621
2622                 blk_finish_plug(&plug);
2623                 btrfs_set_log_full_commit(root->fs_info, trans);
2624
2625                 if (ret != -ENOSPC) {
2626                         btrfs_abort_transaction(trans, root, ret);
2627                         mutex_unlock(&log_root_tree->log_mutex);
2628                         goto out;
2629                 }
2630                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2631                 btrfs_free_logged_extents(log, log_transid);
2632                 mutex_unlock(&log_root_tree->log_mutex);
2633                 ret = -EAGAIN;
2634                 goto out;
2635         }
2636
2637         if (log_root_tree->log_transid_committed >= root_log_ctx.log_transid) {
2638                 mutex_unlock(&log_root_tree->log_mutex);
2639                 ret = root_log_ctx.log_ret;
2640                 goto out;
2641         }
2642
2643         index2 = root_log_ctx.log_transid % 2;
2644         if (atomic_read(&log_root_tree->log_commit[index2])) {
2645                 blk_finish_plug(&plug);
2646                 ret = btrfs_wait_marked_extents(log, &log->dirty_log_pages,
2647                                                 mark);
2648                 btrfs_wait_logged_extents(trans, log, log_transid);
2649                 wait_log_commit(trans, log_root_tree,
2650                                 root_log_ctx.log_transid);
2651                 mutex_unlock(&log_root_tree->log_mutex);
2652                 if (!ret)
2653                         ret = root_log_ctx.log_ret;
2654                 goto out;
2655         }
2656         ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
2657         atomic_set(&log_root_tree->log_commit[index2], 1);
2658
2659         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2660                 wait_log_commit(trans, log_root_tree,
2661                                 root_log_ctx.log_transid - 1);
2662         }
2663
2664         wait_for_writer(trans, log_root_tree);
2665
2666         /*
2667          * now that we've moved on to the tree of log tree roots,
2668          * check the full commit flag again
2669          */
2670         if (btrfs_need_log_full_commit(root->fs_info, trans)) {
2671                 blk_finish_plug(&plug);
2672                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2673                 btrfs_free_logged_extents(log, log_transid);
2674                 mutex_unlock(&log_root_tree->log_mutex);
2675                 ret = -EAGAIN;
2676                 goto out_wake_log_root;
2677         }
2678
2679         ret = btrfs_write_marked_extents(log_root_tree,
2680                                          &log_root_tree->dirty_log_pages,
2681                                          EXTENT_DIRTY | EXTENT_NEW);
2682         blk_finish_plug(&plug);
2683         if (ret) {
2684                 btrfs_set_log_full_commit(root->fs_info, trans);
2685                 btrfs_abort_transaction(trans, root, ret);
2686                 btrfs_free_logged_extents(log, log_transid);
2687                 mutex_unlock(&log_root_tree->log_mutex);
2688                 goto out_wake_log_root;
2689         }
2690         ret = btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2691         if (!ret)
2692                 ret = btrfs_wait_marked_extents(log_root_tree,
2693                                                 &log_root_tree->dirty_log_pages,
2694                                                 EXTENT_NEW | EXTENT_DIRTY);
2695         if (ret) {
2696                 btrfs_set_log_full_commit(root->fs_info, trans);
2697                 btrfs_free_logged_extents(log, log_transid);
2698                 mutex_unlock(&log_root_tree->log_mutex);
2699                 goto out_wake_log_root;
2700         }
2701         btrfs_wait_logged_extents(trans, log, log_transid);
2702
2703         btrfs_set_super_log_root(root->fs_info->super_for_commit,
2704                                 log_root_tree->node->start);
2705         btrfs_set_super_log_root_level(root->fs_info->super_for_commit,
2706                                 btrfs_header_level(log_root_tree->node));
2707
2708         log_root_tree->log_transid++;
2709         mutex_unlock(&log_root_tree->log_mutex);
2710
2711         /*
2712          * nobody else is going to jump in and write the the ctree
2713          * super here because the log_commit atomic below is protecting
2714          * us.  We must be called with a transaction handle pinning
2715          * the running transaction open, so a full commit can't hop
2716          * in and cause problems either.
2717          */
2718         ret = write_ctree_super(trans, root->fs_info->tree_root, 1);
2719         if (ret) {
2720                 btrfs_set_log_full_commit(root->fs_info, trans);
2721                 btrfs_abort_transaction(trans, root, ret);
2722                 goto out_wake_log_root;
2723         }
2724
2725         mutex_lock(&root->log_mutex);
2726         if (root->last_log_commit < log_transid)
2727                 root->last_log_commit = log_transid;
2728         mutex_unlock(&root->log_mutex);
2729
2730 out_wake_log_root:
2731         /*
2732          * We needn't get log_mutex here because we are sure all
2733          * the other tasks are blocked.
2734          */
2735         btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
2736
2737         mutex_lock(&log_root_tree->log_mutex);
2738         log_root_tree->log_transid_committed++;
2739         atomic_set(&log_root_tree->log_commit[index2], 0);
2740         mutex_unlock(&log_root_tree->log_mutex);
2741
2742         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2743                 wake_up(&log_root_tree->log_commit_wait[index2]);
2744 out:
2745         /* See above. */
2746         btrfs_remove_all_log_ctxs(root, index1, ret);
2747
2748         mutex_lock(&root->log_mutex);
2749         root->log_transid_committed++;
2750         atomic_set(&root->log_commit[index1], 0);
2751         mutex_unlock(&root->log_mutex);
2752
2753         if (waitqueue_active(&root->log_commit_wait[index1]))
2754                 wake_up(&root->log_commit_wait[index1]);
2755         return ret;
2756 }
2757
2758 static void free_log_tree(struct btrfs_trans_handle *trans,
2759                           struct btrfs_root *log)
2760 {
2761         int ret;
2762         u64 start;
2763         u64 end;
2764         struct walk_control wc = {
2765                 .free = 1,
2766                 .process_func = process_one_buffer
2767         };
2768
2769         ret = walk_log_tree(trans, log, &wc);
2770         /* I don't think this can happen but just in case */
2771         if (ret)
2772                 btrfs_abort_transaction(trans, log, ret);
2773
2774         while (1) {
2775                 ret = find_first_extent_bit(&log->dirty_log_pages,
2776                                 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW,
2777                                 NULL);
2778                 if (ret)
2779                         break;
2780
2781                 clear_extent_bits(&log->dirty_log_pages, start, end,
2782                                   EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2783         }
2784
2785         /*
2786          * We may have short-circuited the log tree with the full commit logic
2787          * and left ordered extents on our list, so clear these out to keep us
2788          * from leaking inodes and memory.
2789          */
2790         btrfs_free_logged_extents(log, 0);
2791         btrfs_free_logged_extents(log, 1);
2792
2793         free_extent_buffer(log->node);
2794         kfree(log);
2795 }
2796
2797 /*
2798  * free all the extents used by the tree log.  This should be called
2799  * at commit time of the full transaction
2800  */
2801 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2802 {
2803         if (root->log_root) {
2804                 free_log_tree(trans, root->log_root);
2805                 root->log_root = NULL;
2806         }
2807         return 0;
2808 }
2809
2810 int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2811                              struct btrfs_fs_info *fs_info)
2812 {
2813         if (fs_info->log_root_tree) {
2814                 free_log_tree(trans, fs_info->log_root_tree);
2815                 fs_info->log_root_tree = NULL;
2816         }
2817         return 0;
2818 }
2819
2820 /*
2821  * If both a file and directory are logged, and unlinks or renames are
2822  * mixed in, we have a few interesting corners:
2823  *
2824  * create file X in dir Y
2825  * link file X to X.link in dir Y
2826  * fsync file X
2827  * unlink file X but leave X.link
2828  * fsync dir Y
2829  *
2830  * After a crash we would expect only X.link to exist.  But file X
2831  * didn't get fsync'd again so the log has back refs for X and X.link.
2832  *
2833  * We solve this by removing directory entries and inode backrefs from the
2834  * log when a file that was logged in the current transaction is
2835  * unlinked.  Any later fsync will include the updated log entries, and
2836  * we'll be able to reconstruct the proper directory items from backrefs.
2837  *
2838  * This optimizations allows us to avoid relogging the entire inode
2839  * or the entire directory.
2840  */
2841 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2842                                  struct btrfs_root *root,
2843                                  const char *name, int name_len,
2844                                  struct inode *dir, u64 index)
2845 {
2846         struct btrfs_root *log;
2847         struct btrfs_dir_item *di;
2848         struct btrfs_path *path;
2849         int ret;
2850         int err = 0;
2851         int bytes_del = 0;
2852         u64 dir_ino = btrfs_ino(dir);
2853
2854         if (BTRFS_I(dir)->logged_trans < trans->transid)
2855                 return 0;
2856
2857         ret = join_running_log_trans(root);
2858         if (ret)
2859                 return 0;
2860
2861         mutex_lock(&BTRFS_I(dir)->log_mutex);
2862
2863         log = root->log_root;
2864         path = btrfs_alloc_path();
2865         if (!path) {
2866                 err = -ENOMEM;
2867                 goto out_unlock;
2868         }
2869
2870         di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2871                                    name, name_len, -1);
2872         if (IS_ERR(di)) {
2873                 err = PTR_ERR(di);
2874                 goto fail;
2875         }
2876         if (di) {
2877                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2878                 bytes_del += name_len;
2879                 if (ret) {
2880                         err = ret;
2881                         goto fail;
2882                 }
2883         }
2884         btrfs_release_path(path);
2885         di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2886                                          index, name, name_len, -1);
2887         if (IS_ERR(di)) {
2888                 err = PTR_ERR(di);
2889                 goto fail;
2890         }
2891         if (di) {
2892                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2893                 bytes_del += name_len;
2894                 if (ret) {
2895                         err = ret;
2896                         goto fail;
2897                 }
2898         }
2899
2900         /* update the directory size in the log to reflect the names
2901          * we have removed
2902          */
2903         if (bytes_del) {
2904                 struct btrfs_key key;
2905
2906                 key.objectid = dir_ino;
2907                 key.offset = 0;
2908                 key.type = BTRFS_INODE_ITEM_KEY;
2909                 btrfs_release_path(path);
2910
2911                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2912                 if (ret < 0) {
2913                         err = ret;
2914                         goto fail;
2915                 }
2916                 if (ret == 0) {
2917                         struct btrfs_inode_item *item;
2918                         u64 i_size;
2919
2920                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2921                                               struct btrfs_inode_item);
2922                         i_size = btrfs_inode_size(path->nodes[0], item);
2923                         if (i_size > bytes_del)
2924                                 i_size -= bytes_del;
2925                         else
2926                                 i_size = 0;
2927                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2928                         btrfs_mark_buffer_dirty(path->nodes[0]);
2929                 } else
2930                         ret = 0;
2931                 btrfs_release_path(path);
2932         }
2933 fail:
2934         btrfs_free_path(path);
2935 out_unlock:
2936         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2937         if (ret == -ENOSPC) {
2938                 btrfs_set_log_full_commit(root->fs_info, trans);
2939                 ret = 0;
2940         } else if (ret < 0)
2941                 btrfs_abort_transaction(trans, root, ret);
2942
2943         btrfs_end_log_trans(root);
2944
2945         return err;
2946 }
2947
2948 /* see comments for btrfs_del_dir_entries_in_log */
2949 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2950                                struct btrfs_root *root,
2951                                const char *name, int name_len,
2952                                struct inode *inode, u64 dirid)
2953 {
2954         struct btrfs_root *log;
2955         u64 index;
2956         int ret;
2957
2958         if (BTRFS_I(inode)->logged_trans < trans->transid)
2959                 return 0;
2960
2961         ret = join_running_log_trans(root);
2962         if (ret)
2963                 return 0;
2964         log = root->log_root;
2965         mutex_lock(&BTRFS_I(inode)->log_mutex);
2966
2967         ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
2968                                   dirid, &index);
2969         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2970         if (ret == -ENOSPC) {
2971                 btrfs_set_log_full_commit(root->fs_info, trans);
2972                 ret = 0;
2973         } else if (ret < 0 && ret != -ENOENT)
2974                 btrfs_abort_transaction(trans, root, ret);
2975         btrfs_end_log_trans(root);
2976
2977         return ret;
2978 }
2979
2980 /*
2981  * creates a range item in the log for 'dirid'.  first_offset and
2982  * last_offset tell us which parts of the key space the log should
2983  * be considered authoritative for.
2984  */
2985 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2986                                        struct btrfs_root *log,
2987                                        struct btrfs_path *path,
2988                                        int key_type, u64 dirid,
2989                                        u64 first_offset, u64 last_offset)
2990 {
2991         int ret;
2992         struct btrfs_key key;
2993         struct btrfs_dir_log_item *item;
2994
2995         key.objectid = dirid;
2996         key.offset = first_offset;
2997         if (key_type == BTRFS_DIR_ITEM_KEY)
2998                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2999         else
3000                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
3001         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
3002         if (ret)
3003                 return ret;
3004
3005         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3006                               struct btrfs_dir_log_item);
3007         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
3008         btrfs_mark_buffer_dirty(path->nodes[0]);
3009         btrfs_release_path(path);
3010         return 0;
3011 }
3012
3013 /*
3014  * log all the items included in the current transaction for a given
3015  * directory.  This also creates the range items in the log tree required
3016  * to replay anything deleted before the fsync
3017  */
3018 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
3019                           struct btrfs_root *root, struct inode *inode,
3020                           struct btrfs_path *path,
3021                           struct btrfs_path *dst_path, int key_type,
3022                           u64 min_offset, u64 *last_offset_ret)
3023 {
3024         struct btrfs_key min_key;
3025         struct btrfs_root *log = root->log_root;
3026         struct extent_buffer *src;
3027         int err = 0;
3028         int ret;
3029         int i;
3030         int nritems;
3031         u64 first_offset = min_offset;
3032         u64 last_offset = (u64)-1;
3033         u64 ino = btrfs_ino(inode);
3034
3035         log = root->log_root;
3036
3037         min_key.objectid = ino;
3038         min_key.type = key_type;
3039         min_key.offset = min_offset;
3040
3041         ret = btrfs_search_forward(root, &min_key, path, trans->transid);
3042
3043         /*
3044          * we didn't find anything from this transaction, see if there
3045          * is anything at all
3046          */
3047         if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
3048                 min_key.objectid = ino;
3049                 min_key.type = key_type;
3050                 min_key.offset = (u64)-1;
3051                 btrfs_release_path(path);
3052                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3053                 if (ret < 0) {
3054                         btrfs_release_path(path);
3055                         return ret;
3056                 }
3057                 ret = btrfs_previous_item(root, path, ino, key_type);
3058
3059                 /* if ret == 0 there are items for this type,
3060                  * create a range to tell us the last key of this type.
3061                  * otherwise, there are no items in this directory after
3062                  * *min_offset, and we create a range to indicate that.
3063                  */
3064                 if (ret == 0) {
3065                         struct btrfs_key tmp;
3066                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
3067                                               path->slots[0]);
3068                         if (key_type == tmp.type)
3069                                 first_offset = max(min_offset, tmp.offset) + 1;
3070                 }
3071                 goto done;
3072         }
3073
3074         /* go backward to find any previous key */
3075         ret = btrfs_previous_item(root, path, ino, key_type);
3076         if (ret == 0) {
3077                 struct btrfs_key tmp;
3078                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3079                 if (key_type == tmp.type) {
3080                         first_offset = tmp.offset;
3081                         ret = overwrite_item(trans, log, dst_path,
3082                                              path->nodes[0], path->slots[0],
3083                                              &tmp);
3084                         if (ret) {
3085                                 err = ret;
3086                                 goto done;
3087                         }
3088                 }
3089         }
3090         btrfs_release_path(path);
3091
3092         /* find the first key from this transaction again */
3093         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3094         if (WARN_ON(ret != 0))
3095                 goto done;
3096
3097         /*
3098          * we have a block from this transaction, log every item in it
3099          * from our directory
3100          */
3101         while (1) {
3102                 struct btrfs_key tmp;
3103                 src = path->nodes[0];
3104                 nritems = btrfs_header_nritems(src);
3105                 for (i = path->slots[0]; i < nritems; i++) {
3106                         btrfs_item_key_to_cpu(src, &min_key, i);
3107
3108                         if (min_key.objectid != ino || min_key.type != key_type)
3109                                 goto done;
3110                         ret = overwrite_item(trans, log, dst_path, src, i,
3111                                              &min_key);
3112                         if (ret) {
3113                                 err = ret;
3114                                 goto done;
3115                         }
3116                 }
3117                 path->slots[0] = nritems;
3118
3119                 /*
3120                  * look ahead to the next item and see if it is also
3121                  * from this directory and from this transaction
3122                  */
3123                 ret = btrfs_next_leaf(root, path);
3124                 if (ret == 1) {
3125                         last_offset = (u64)-1;
3126                         goto done;
3127                 }
3128                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3129                 if (tmp.objectid != ino || tmp.type != key_type) {
3130                         last_offset = (u64)-1;
3131                         goto done;
3132                 }
3133                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
3134                         ret = overwrite_item(trans, log, dst_path,
3135                                              path->nodes[0], path->slots[0],
3136                                              &tmp);
3137                         if (ret)
3138                                 err = ret;
3139                         else
3140                                 last_offset = tmp.offset;
3141                         goto done;
3142                 }
3143         }
3144 done:
3145         btrfs_release_path(path);
3146         btrfs_release_path(dst_path);
3147
3148         if (err == 0) {
3149                 *last_offset_ret = last_offset;
3150                 /*
3151                  * insert the log range keys to indicate where the log
3152                  * is valid
3153                  */
3154                 ret = insert_dir_log_key(trans, log, path, key_type,
3155                                          ino, first_offset, last_offset);
3156                 if (ret)
3157                         err = ret;
3158         }
3159         return err;
3160 }
3161
3162 /*
3163  * logging directories is very similar to logging inodes, We find all the items
3164  * from the current transaction and write them to the log.
3165  *
3166  * The recovery code scans the directory in the subvolume, and if it finds a
3167  * key in the range logged that is not present in the log tree, then it means
3168  * that dir entry was unlinked during the transaction.
3169  *
3170  * In order for that scan to work, we must include one key smaller than
3171  * the smallest logged by this transaction and one key larger than the largest
3172  * key logged by this transaction.
3173  */
3174 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
3175                           struct btrfs_root *root, struct inode *inode,
3176                           struct btrfs_path *path,
3177                           struct btrfs_path *dst_path)
3178 {
3179         u64 min_key;
3180         u64 max_key;
3181         int ret;
3182         int key_type = BTRFS_DIR_ITEM_KEY;
3183
3184 again:
3185         min_key = 0;
3186         max_key = 0;
3187         while (1) {
3188                 ret = log_dir_items(trans, root, inode, path,
3189                                     dst_path, key_type, min_key,
3190                                     &max_key);
3191                 if (ret)
3192                         return ret;
3193                 if (max_key == (u64)-1)
3194                         break;
3195                 min_key = max_key + 1;
3196         }
3197
3198         if (key_type == BTRFS_DIR_ITEM_KEY) {
3199                 key_type = BTRFS_DIR_INDEX_KEY;
3200                 goto again;
3201         }
3202         return 0;
3203 }
3204
3205 /*
3206  * a helper function to drop items from the log before we relog an
3207  * inode.  max_key_type indicates the highest item type to remove.
3208  * This cannot be run for file data extents because it does not
3209  * free the extents they point to.
3210  */
3211 static int drop_objectid_items(struct btrfs_trans_handle *trans,
3212                                   struct btrfs_root *log,
3213                                   struct btrfs_path *path,
3214                                   u64 objectid, int max_key_type)
3215 {
3216         int ret;
3217         struct btrfs_key key;
3218         struct btrfs_key found_key;
3219         int start_slot;
3220
3221         key.objectid = objectid;
3222         key.type = max_key_type;
3223         key.offset = (u64)-1;
3224
3225         while (1) {
3226                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
3227                 BUG_ON(ret == 0); /* Logic error */
3228                 if (ret < 0)
3229                         break;
3230
3231                 if (path->slots[0] == 0)
3232                         break;
3233
3234                 path->slots[0]--;
3235                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
3236                                       path->slots[0]);
3237
3238                 if (found_key.objectid != objectid)
3239                         break;
3240
3241                 found_key.offset = 0;
3242                 found_key.type = 0;
3243                 ret = btrfs_bin_search(path->nodes[0], &found_key, 0,
3244                                        &start_slot);
3245
3246                 ret = btrfs_del_items(trans, log, path, start_slot,
3247                                       path->slots[0] - start_slot + 1);
3248                 /*
3249                  * If start slot isn't 0 then we don't need to re-search, we've
3250                  * found the last guy with the objectid in this tree.
3251                  */
3252                 if (ret || start_slot != 0)
3253                         break;
3254                 btrfs_release_path(path);
3255         }
3256         btrfs_release_path(path);
3257         if (ret > 0)
3258                 ret = 0;
3259         return ret;
3260 }
3261
3262 static void fill_inode_item(struct btrfs_trans_handle *trans,
3263                             struct extent_buffer *leaf,
3264                             struct btrfs_inode_item *item,
3265                             struct inode *inode, int log_inode_only,
3266                             u64 logged_isize)
3267 {
3268         struct btrfs_map_token token;
3269
3270         btrfs_init_map_token(&token);
3271
3272         if (log_inode_only) {
3273                 /* set the generation to zero so the recover code
3274                  * can tell the difference between an logging
3275                  * just to say 'this inode exists' and a logging
3276                  * to say 'update this inode with these values'
3277                  */
3278                 btrfs_set_token_inode_generation(leaf, item, 0, &token);
3279                 btrfs_set_token_inode_size(leaf, item, logged_isize, &token);
3280         } else {
3281                 btrfs_set_token_inode_generation(leaf, item,
3282                                                  BTRFS_I(inode)->generation,
3283                                                  &token);
3284                 btrfs_set_token_inode_size(leaf, item, inode->i_size, &token);
3285         }
3286
3287         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3288         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3289         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3290         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3291
3292         btrfs_set_token_timespec_sec(leaf, &item->atime,
3293                                      inode->i_atime.tv_sec, &token);
3294         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3295                                       inode->i_atime.tv_nsec, &token);
3296
3297         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3298                                      inode->i_mtime.tv_sec, &token);
3299         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3300                                       inode->i_mtime.tv_nsec, &token);
3301
3302         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3303                                      inode->i_ctime.tv_sec, &token);
3304         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3305                                       inode->i_ctime.tv_nsec, &token);
3306
3307         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3308                                      &token);
3309
3310         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3311         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3312         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3313         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3314         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3315 }
3316
3317 static int log_inode_item(struct btrfs_trans_handle *trans,
3318                           struct btrfs_root *log, struct btrfs_path *path,
3319                           struct inode *inode)
3320 {
3321         struct btrfs_inode_item *inode_item;
3322         int ret;
3323
3324         ret = btrfs_insert_empty_item(trans, log, path,
3325                                       &BTRFS_I(inode)->location,
3326                                       sizeof(*inode_item));
3327         if (ret && ret != -EEXIST)
3328                 return ret;
3329         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3330                                     struct btrfs_inode_item);
3331         fill_inode_item(trans, path->nodes[0], inode_item, inode, 0, 0);
3332         btrfs_release_path(path);
3333         return 0;
3334 }
3335
3336 static noinline int copy_items(struct btrfs_trans_handle *trans,
3337                                struct inode *inode,
3338                                struct btrfs_path *dst_path,
3339                                struct btrfs_path *src_path, u64 *last_extent,
3340                                int start_slot, int nr, int inode_only,
3341                                u64 logged_isize)
3342 {
3343         unsigned long src_offset;
3344         unsigned long dst_offset;
3345         struct btrfs_root *log = BTRFS_I(inode)->root->log_root;
3346         struct btrfs_file_extent_item *extent;
3347         struct btrfs_inode_item *inode_item;
3348         struct extent_buffer *src = src_path->nodes[0];
3349         struct btrfs_key first_key, last_key, key;
3350         int ret;
3351         struct btrfs_key *ins_keys;
3352         u32 *ins_sizes;
3353         char *ins_data;
3354         int i;
3355         struct list_head ordered_sums;
3356         int skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3357         bool has_extents = false;
3358         bool need_find_last_extent = true;
3359         bool done = false;
3360
3361         INIT_LIST_HEAD(&ordered_sums);
3362
3363         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
3364                            nr * sizeof(u32), GFP_NOFS);
3365         if (!ins_data)
3366                 return -ENOMEM;
3367
3368         first_key.objectid = (u64)-1;
3369
3370         ins_sizes = (u32 *)ins_data;
3371         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
3372
3373         for (i = 0; i < nr; i++) {
3374                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
3375                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
3376         }
3377         ret = btrfs_insert_empty_items(trans, log, dst_path,
3378                                        ins_keys, ins_sizes, nr);
3379         if (ret) {
3380                 kfree(ins_data);
3381                 return ret;
3382         }
3383
3384         for (i = 0; i < nr; i++, dst_path->slots[0]++) {
3385                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
3386                                                    dst_path->slots[0]);
3387
3388                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
3389
3390                 if ((i == (nr - 1)))
3391                         last_key = ins_keys[i];
3392
3393                 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
3394                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
3395                                                     dst_path->slots[0],
3396                                                     struct btrfs_inode_item);
3397                         fill_inode_item(trans, dst_path->nodes[0], inode_item,
3398                                         inode, inode_only == LOG_INODE_EXISTS,
3399                                         logged_isize);
3400                 } else {
3401                         copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
3402                                            src_offset, ins_sizes[i]);
3403                 }
3404
3405                 /*
3406                  * We set need_find_last_extent here in case we know we were
3407                  * processing other items and then walk into the first extent in
3408                  * the inode.  If we don't hit an extent then nothing changes,
3409                  * we'll do the last search the next time around.
3410                  */
3411                 if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY) {
3412                         has_extents = true;
3413                         if (first_key.objectid == (u64)-1)
3414                                 first_key = ins_keys[i];
3415                 } else {
3416                         need_find_last_extent = false;
3417                 }
3418
3419                 /* take a reference on file data extents so that truncates
3420                  * or deletes of this inode don't have to relog the inode
3421                  * again
3422                  */
3423                 if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY &&
3424                     !skip_csum) {
3425                         int found_type;
3426                         extent = btrfs_item_ptr(src, start_slot + i,
3427                                                 struct btrfs_file_extent_item);
3428
3429                         if (btrfs_file_extent_generation(src, extent) < trans->transid)
3430                                 continue;
3431
3432                         found_type = btrfs_file_extent_type(src, extent);
3433                         if (found_type == BTRFS_FILE_EXTENT_REG) {
3434                                 u64 ds, dl, cs, cl;
3435                                 ds = btrfs_file_extent_disk_bytenr(src,
3436                                                                 extent);
3437                                 /* ds == 0 is a hole */
3438                                 if (ds == 0)
3439                                         continue;
3440
3441                                 dl = btrfs_file_extent_disk_num_bytes(src,
3442                                                                 extent);
3443                                 cs = btrfs_file_extent_offset(src, extent);
3444                                 cl = btrfs_file_extent_num_bytes(src,
3445                                                                 extent);
3446                                 if (btrfs_file_extent_compression(src,
3447                                                                   extent)) {
3448                                         cs = 0;
3449                                         cl = dl;
3450                                 }
3451
3452                                 ret = btrfs_lookup_csums_range(
3453                                                 log->fs_info->csum_root,
3454                                                 ds + cs, ds + cs + cl - 1,
3455                                                 &ordered_sums, 0);
3456                                 if (ret) {
3457                                         btrfs_release_path(dst_path);
3458                                         kfree(ins_data);
3459                                         return ret;
3460                                 }
3461                         }
3462                 }
3463         }
3464
3465         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
3466         btrfs_release_path(dst_path);
3467         kfree(ins_data);
3468
3469         /*
3470          * we have to do this after the loop above to avoid changing the
3471          * log tree while trying to change the log tree.
3472          */
3473         ret = 0;
3474         while (!list_empty(&ordered_sums)) {
3475                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3476                                                    struct btrfs_ordered_sum,
3477                                                    list);
3478                 if (!ret)
3479                         ret = btrfs_csum_file_blocks(trans, log, sums);
3480                 list_del(&sums->list);
3481                 kfree(sums);
3482         }
3483
3484         if (!has_extents)
3485                 return ret;
3486
3487         if (need_find_last_extent && *last_extent == first_key.offset) {
3488                 /*
3489                  * We don't have any leafs between our current one and the one
3490                  * we processed before that can have file extent items for our
3491                  * inode (and have a generation number smaller than our current
3492                  * transaction id).
3493                  */
3494                 need_find_last_extent = false;
3495         }
3496
3497         /*
3498          * Because we use btrfs_search_forward we could skip leaves that were
3499          * not modified and then assume *last_extent is valid when it really
3500          * isn't.  So back up to the previous leaf and read the end of the last
3501          * extent before we go and fill in holes.
3502          */
3503         if (need_find_last_extent) {
3504                 u64 len;
3505
3506                 ret = btrfs_prev_leaf(BTRFS_I(inode)->root, src_path);
3507                 if (ret < 0)
3508                         return ret;
3509                 if (ret)
3510                         goto fill_holes;
3511                 if (src_path->slots[0])
3512                         src_path->slots[0]--;
3513                 src = src_path->nodes[0];
3514                 btrfs_item_key_to_cpu(src, &key, src_path->slots[0]);
3515                 if (key.objectid != btrfs_ino(inode) ||
3516                     key.type != BTRFS_EXTENT_DATA_KEY)
3517                         goto fill_holes;
3518                 extent = btrfs_item_ptr(src, src_path->slots[0],
3519                                         struct btrfs_file_extent_item);
3520                 if (btrfs_file_extent_type(src, extent) ==
3521                     BTRFS_FILE_EXTENT_INLINE) {
3522                         len = btrfs_file_extent_inline_len(src,
3523                                                            src_path->slots[0],
3524                                                            extent);
3525                         *last_extent = ALIGN(key.offset + len,
3526                                              log->sectorsize);
3527                 } else {
3528                         len = btrfs_file_extent_num_bytes(src, extent);
3529                         *last_extent = key.offset + len;
3530                 }
3531         }
3532 fill_holes:
3533         /* So we did prev_leaf, now we need to move to the next leaf, but a few
3534          * things could have happened
3535          *
3536          * 1) A merge could have happened, so we could currently be on a leaf
3537          * that holds what we were copying in the first place.
3538          * 2) A split could have happened, and now not all of the items we want
3539          * are on the same leaf.
3540          *
3541          * So we need to adjust how we search for holes, we need to drop the
3542          * path and re-search for the first extent key we found, and then walk
3543          * forward until we hit the last one we copied.
3544          */
3545         if (need_find_last_extent) {
3546                 /* btrfs_prev_leaf could return 1 without releasing the path */
3547                 btrfs_release_path(src_path);
3548                 ret = btrfs_search_slot(NULL, BTRFS_I(inode)->root, &first_key,
3549                                         src_path, 0, 0);
3550                 if (ret < 0)
3551                         return ret;
3552                 ASSERT(ret == 0);
3553                 src = src_path->nodes[0];
3554                 i = src_path->slots[0];
3555         } else {
3556                 i = start_slot;
3557         }
3558
3559         /*
3560          * Ok so here we need to go through and fill in any holes we may have
3561          * to make sure that holes are punched for those areas in case they had
3562          * extents previously.
3563          */
3564         while (!done) {
3565                 u64 offset, len;
3566                 u64 extent_end;
3567
3568                 if (i >= btrfs_header_nritems(src_path->nodes[0])) {
3569                         ret = btrfs_next_leaf(BTRFS_I(inode)->root, src_path);
3570                         if (ret < 0)
3571                                 return ret;
3572                         ASSERT(ret == 0);
3573                         src = src_path->nodes[0];
3574                         i = 0;
3575                 }
3576
3577                 btrfs_item_key_to_cpu(src, &key, i);
3578                 if (!btrfs_comp_cpu_keys(&key, &last_key))
3579                         done = true;
3580                 if (key.objectid != btrfs_ino(inode) ||
3581                     key.type != BTRFS_EXTENT_DATA_KEY) {
3582                         i++;
3583                         continue;
3584                 }
3585                 extent = btrfs_item_ptr(src, i, struct btrfs_file_extent_item);
3586                 if (btrfs_file_extent_type(src, extent) ==
3587                     BTRFS_FILE_EXTENT_INLINE) {
3588                         len = btrfs_file_extent_inline_len(src, i, extent);
3589                         extent_end = ALIGN(key.offset + len, log->sectorsize);
3590                 } else {
3591                         len = btrfs_file_extent_num_bytes(src, extent);
3592                         extent_end = key.offset + len;
3593                 }
3594                 i++;
3595
3596                 if (*last_extent == key.offset) {
3597                         *last_extent = extent_end;
3598                         continue;
3599                 }
3600                 offset = *last_extent;
3601                 len = key.offset - *last_extent;
3602                 ret = btrfs_insert_file_extent(trans, log, btrfs_ino(inode),
3603                                                offset, 0, 0, len, 0, len, 0,
3604                                                0, 0);
3605                 if (ret)
3606                         break;
3607                 *last_extent = extent_end;
3608         }
3609         /*
3610          * Need to let the callers know we dropped the path so they should
3611          * re-search.
3612          */
3613         if (!ret && need_find_last_extent)
3614                 ret = 1;
3615         return ret;
3616 }
3617
3618 static int extent_cmp(void *priv, struct list_head *a, struct list_head *b)
3619 {
3620         struct extent_map *em1, *em2;
3621
3622         em1 = list_entry(a, struct extent_map, list);
3623         em2 = list_entry(b, struct extent_map, list);
3624
3625         if (em1->start < em2->start)
3626                 return -1;
3627         else if (em1->start > em2->start)
3628                 return 1;
3629         return 0;
3630 }
3631
3632 static int wait_ordered_extents(struct btrfs_trans_handle *trans,
3633                                 struct inode *inode,
3634                                 struct btrfs_root *root,
3635                                 const struct extent_map *em,
3636                                 const struct list_head *logged_list,
3637                                 bool *ordered_io_error)
3638 {
3639         struct btrfs_ordered_extent *ordered;
3640         struct btrfs_root *log = root->log_root;
3641         u64 mod_start = em->mod_start;
3642         u64 mod_len = em->mod_len;
3643         const bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3644         u64 csum_offset;
3645         u64 csum_len;
3646         LIST_HEAD(ordered_sums);
3647         int ret = 0;
3648
3649         *ordered_io_error = false;
3650
3651         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
3652             em->block_start == EXTENT_MAP_HOLE)
3653                 return 0;
3654
3655         /*
3656          * Wait far any ordered extent that covers our extent map. If it
3657          * finishes without an error, first check and see if our csums are on
3658          * our outstanding ordered extents.
3659          */
3660         list_for_each_entry(ordered, logged_list, log_list) {
3661                 struct btrfs_ordered_sum *sum;
3662
3663                 if (!mod_len)
3664                         break;
3665
3666                 if (ordered->file_offset + ordered->len <= mod_start ||
3667                     mod_start + mod_len <= ordered->file_offset)
3668                         continue;
3669
3670                 if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
3671                     !test_bit(BTRFS_ORDERED_IOERR, &ordered->flags) &&
3672                     !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
3673                         const u64 start = ordered->file_offset;
3674                         const u64 end = ordered->file_offset + ordered->len - 1;
3675
3676                         WARN_ON(ordered->inode != inode);
3677                         filemap_fdatawrite_range(inode->i_mapping, start, end);
3678                 }
3679
3680                 wait_event(ordered->wait,
3681                            (test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) ||
3682                             test_bit(BTRFS_ORDERED_IOERR, &ordered->flags)));
3683
3684                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags)) {
3685                         /*
3686                          * Clear the AS_EIO/AS_ENOSPC flags from the inode's
3687                          * i_mapping flags, so that the next fsync won't get
3688                          * an outdated io error too.
3689                          */
3690                         btrfs_inode_check_errors(inode);
3691                         *ordered_io_error = true;
3692                         break;
3693                 }
3694                 /*
3695                  * We are going to copy all the csums on this ordered extent, so
3696                  * go ahead and adjust mod_start and mod_len in case this
3697                  * ordered extent has already been logged.
3698                  */
3699                 if (ordered->file_offset > mod_start) {
3700                         if (ordered->file_offset + ordered->len >=
3701                             mod_start + mod_len)
3702                                 mod_len = ordered->file_offset - mod_start;
3703                         /*
3704                          * If we have this case
3705                          *
3706                          * |--------- logged extent ---------|
3707                          *       |----- ordered extent ----|
3708                          *
3709                          * Just don't mess with mod_start and mod_len, we'll
3710                          * just end up logging more csums than we need and it
3711                          * will be ok.
3712                          */
3713                 } else {
3714                         if (ordered->file_offset + ordered->len <
3715                             mod_start + mod_len) {
3716                                 mod_len = (mod_start + mod_len) -
3717                                         (ordered->file_offset + ordered->len);
3718                                 mod_start = ordered->file_offset +
3719                                         ordered->len;
3720                         } else {
3721                                 mod_len = 0;
3722                         }
3723                 }
3724
3725                 if (skip_csum)
3726                         continue;
3727
3728                 /*
3729                  * To keep us from looping for the above case of an ordered
3730                  * extent that falls inside of the logged extent.
3731                  */
3732                 if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM,
3733                                      &ordered->flags))
3734                         continue;
3735
3736                 if (ordered->csum_bytes_left) {
3737                         btrfs_start_ordered_extent(inode, ordered, 0);
3738                         wait_event(ordered->wait,
3739                                    ordered->csum_bytes_left == 0);
3740                 }
3741
3742                 list_for_each_entry(sum, &ordered->list, list) {
3743                         ret = btrfs_csum_file_blocks(trans, log, sum);
3744                         if (ret)
3745                                 break;
3746                 }
3747         }
3748
3749         if (*ordered_io_error || !mod_len || ret || skip_csum)
3750                 return ret;
3751
3752         if (em->compress_type) {
3753                 csum_offset = 0;
3754                 csum_len = max(em->block_len, em->orig_block_len);
3755         } else {
3756                 csum_offset = mod_start - em->start;
3757                 csum_len = mod_len;
3758         }
3759
3760         /* block start is already adjusted for the file extent offset. */
3761         ret = btrfs_lookup_csums_range(log->fs_info->csum_root,
3762                                        em->block_start + csum_offset,
3763                                        em->block_start + csum_offset +
3764                                        csum_len - 1, &ordered_sums, 0);
3765         if (ret)
3766                 return ret;
3767
3768         while (!list_empty(&ordered_sums)) {
3769                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3770                                                    struct btrfs_ordered_sum,
3771                                                    list);
3772                 if (!ret)
3773                         ret = btrfs_csum_file_blocks(trans, log, sums);
3774                 list_del(&sums->list);
3775                 kfree(sums);
3776         }
3777
3778         return ret;
3779 }
3780
3781 static int log_one_extent(struct btrfs_trans_handle *trans,
3782                           struct inode *inode, struct btrfs_root *root,
3783                           const struct extent_map *em,
3784                           struct btrfs_path *path,
3785                           const struct list_head *logged_list,
3786                           struct btrfs_log_ctx *ctx)
3787 {
3788         struct btrfs_root *log = root->log_root;
3789         struct btrfs_file_extent_item *fi;
3790         struct extent_buffer *leaf;
3791         struct btrfs_map_token token;
3792         struct btrfs_key key;
3793         u64 extent_offset = em->start - em->orig_start;
3794         u64 block_len;
3795         int ret;
3796         int extent_inserted = 0;
3797         bool ordered_io_err = false;
3798
3799         ret = wait_ordered_extents(trans, inode, root, em, logged_list,
3800                                    &ordered_io_err);
3801         if (ret)
3802                 return ret;
3803
3804         if (ordered_io_err) {
3805                 ctx->io_err = -EIO;
3806                 return 0;
3807         }
3808
3809         btrfs_init_map_token(&token);
3810
3811         ret = __btrfs_drop_extents(trans, log, inode, path, em->start,
3812                                    em->start + em->len, NULL, 0, 1,
3813                                    sizeof(*fi), &extent_inserted);
3814         if (ret)
3815                 return ret;
3816
3817         if (!extent_inserted) {
3818                 key.objectid = btrfs_ino(inode);
3819                 key.type = BTRFS_EXTENT_DATA_KEY;
3820                 key.offset = em->start;
3821
3822                 ret = btrfs_insert_empty_item(trans, log, path, &key,
3823                                               sizeof(*fi));
3824                 if (ret)
3825                         return ret;
3826         }
3827         leaf = path->nodes[0];
3828         fi = btrfs_item_ptr(leaf, path->slots[0],
3829                             struct btrfs_file_extent_item);
3830
3831         btrfs_set_token_file_extent_generation(leaf, fi, trans->transid,
3832                                                &token);
3833         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3834                 btrfs_set_token_file_extent_type(leaf, fi,
3835                                                  BTRFS_FILE_EXTENT_PREALLOC,
3836                                                  &token);
3837         else
3838                 btrfs_set_token_file_extent_type(leaf, fi,
3839                                                  BTRFS_FILE_EXTENT_REG,
3840                                                  &token);
3841
3842         block_len = max(em->block_len, em->orig_block_len);
3843         if (em->compress_type != BTRFS_COMPRESS_NONE) {
3844                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3845                                                         em->block_start,
3846                                                         &token);
3847                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3848                                                            &token);
3849         } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
3850                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3851                                                         em->block_start -
3852                                                         extent_offset, &token);
3853                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3854                                                            &token);
3855         } else {
3856                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi, 0, &token);
3857                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, 0,
3858                                                            &token);
3859         }
3860
3861         btrfs_set_token_file_extent_offset(leaf, fi, extent_offset, &token);
3862         btrfs_set_token_file_extent_num_bytes(leaf, fi, em->len, &token);
3863         btrfs_set_token_file_extent_ram_bytes(leaf, fi, em->ram_bytes, &token);
3864         btrfs_set_token_file_extent_compression(leaf, fi, em->compress_type,
3865                                                 &token);
3866         btrfs_set_token_file_extent_encryption(leaf, fi, 0, &token);
3867         btrfs_set_token_file_extent_other_encoding(leaf, fi, 0, &token);
3868         btrfs_mark_buffer_dirty(leaf);
3869
3870         btrfs_release_path(path);
3871
3872         return ret;
3873 }
3874
3875 static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
3876                                      struct btrfs_root *root,
3877                                      struct inode *inode,
3878                                      struct btrfs_path *path,
3879                                      struct list_head *logged_list,
3880                                      struct btrfs_log_ctx *ctx)
3881 {
3882         struct extent_map *em, *n;
3883         struct list_head extents;
3884         struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3885         u64 test_gen;
3886         int ret = 0;
3887         int num = 0;
3888
3889         INIT_LIST_HEAD(&extents);
3890
3891         write_lock(&tree->lock);
3892         test_gen = root->fs_info->last_trans_committed;
3893
3894         list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
3895                 list_del_init(&em->list);
3896
3897                 /*
3898                  * Just an arbitrary number, this can be really CPU intensive
3899                  * once we start getting a lot of extents, and really once we
3900                  * have a bunch of extents we just want to commit since it will
3901                  * be faster.
3902                  */
3903                 if (++num > 32768) {
3904                         list_del_init(&tree->modified_extents);
3905                         ret = -EFBIG;
3906                         goto process;
3907                 }
3908
3909                 if (em->generation <= test_gen)
3910                         continue;
3911                 /* Need a ref to keep it from getting evicted from cache */
3912                 atomic_inc(&em->refs);
3913                 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
3914                 list_add_tail(&em->list, &extents);
3915                 num++;
3916         }
3917
3918         list_sort(NULL, &extents, extent_cmp);
3919
3920 process:
3921         while (!list_empty(&extents)) {
3922                 em = list_entry(extents.next, struct extent_map, list);
3923
3924                 list_del_init(&em->list);
3925
3926                 /*
3927                  * If we had an error we just need to delete everybody from our
3928                  * private list.
3929                  */
3930                 if (ret) {
3931                         clear_em_logging(tree, em);
3932                         free_extent_map(em);
3933                         continue;
3934                 }
3935
3936                 write_unlock(&tree->lock);
3937
3938                 ret = log_one_extent(trans, inode, root, em, path, logged_list,
3939                                      ctx);
3940                 write_lock(&tree->lock);
3941                 clear_em_logging(tree, em);
3942                 free_extent_map(em);
3943         }
3944         WARN_ON(!list_empty(&extents));
3945         write_unlock(&tree->lock);
3946
3947         btrfs_release_path(path);
3948         return ret;
3949 }
3950
3951 static int logged_inode_size(struct btrfs_root *log, struct inode *inode,
3952                              struct btrfs_path *path, u64 *size_ret)
3953 {
3954         struct btrfs_key key;
3955         int ret;
3956
3957         key.objectid = btrfs_ino(inode);
3958         key.type = BTRFS_INODE_ITEM_KEY;
3959         key.offset = 0;
3960
3961         ret = btrfs_search_slot(NULL, log, &key, path, 0, 0);
3962         if (ret < 0) {
3963                 return ret;
3964         } else if (ret > 0) {
3965                 *size_ret = i_size_read(inode);
3966         } else {
3967                 struct btrfs_inode_item *item;
3968
3969                 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3970                                       struct btrfs_inode_item);
3971                 *size_ret = btrfs_inode_size(path->nodes[0], item);
3972         }
3973
3974         btrfs_release_path(path);
3975         return 0;
3976 }
3977
3978 /* log a single inode in the tree log.
3979  * At least one parent directory for this inode must exist in the tree
3980  * or be logged already.
3981  *
3982  * Any items from this inode changed by the current transaction are copied
3983  * to the log tree.  An extra reference is taken on any extents in this
3984  * file, allowing us to avoid a whole pile of corner cases around logging
3985  * blocks that have been removed from the tree.
3986  *
3987  * See LOG_INODE_ALL and related defines for a description of what inode_only
3988  * does.
3989  *
3990  * This handles both files and directories.
3991  */
3992 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
3993                            struct btrfs_root *root, struct inode *inode,
3994                            int inode_only,
3995                            const loff_t start,
3996                            const loff_t end,
3997                            struct btrfs_log_ctx *ctx)
3998 {
3999         struct btrfs_path *path;
4000         struct btrfs_path *dst_path;
4001         struct btrfs_key min_key;
4002         struct btrfs_key max_key;
4003         struct btrfs_root *log = root->log_root;
4004         struct extent_buffer *src = NULL;
4005         LIST_HEAD(logged_list);
4006         u64 last_extent = 0;
4007         int err = 0;
4008         int ret;
4009         int nritems;
4010         int ins_start_slot = 0;
4011         int ins_nr;
4012         bool fast_search = false;
4013         u64 ino = btrfs_ino(inode);
4014         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4015         u64 logged_isize = 0;
4016
4017         path = btrfs_alloc_path();
4018         if (!path)
4019                 return -ENOMEM;
4020         dst_path = btrfs_alloc_path();
4021         if (!dst_path) {
4022                 btrfs_free_path(path);
4023                 return -ENOMEM;
4024         }
4025
4026         min_key.objectid = ino;
4027         min_key.type = BTRFS_INODE_ITEM_KEY;
4028         min_key.offset = 0;
4029
4030         max_key.objectid = ino;
4031
4032
4033         /* today the code can only do partial logging of directories */
4034         if (S_ISDIR(inode->i_mode) ||
4035             (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4036                        &BTRFS_I(inode)->runtime_flags) &&
4037              inode_only == LOG_INODE_EXISTS))
4038                 max_key.type = BTRFS_XATTR_ITEM_KEY;
4039         else
4040                 max_key.type = (u8)-1;
4041         max_key.offset = (u64)-1;
4042
4043         /*
4044          * Only run delayed items if we are a dir or a new file.
4045          * Otherwise commit the delayed inode only, which is needed in
4046          * order for the log replay code to mark inodes for link count
4047          * fixup (create temporary BTRFS_TREE_LOG_FIXUP_OBJECTID items).
4048          */
4049         if (S_ISDIR(inode->i_mode) ||
4050             BTRFS_I(inode)->generation > root->fs_info->last_trans_committed)
4051                 ret = btrfs_commit_inode_delayed_items(trans, inode);
4052         else
4053                 ret = btrfs_commit_inode_delayed_inode(inode);
4054
4055         if (ret) {
4056                 btrfs_free_path(path);
4057                 btrfs_free_path(dst_path);
4058                 return ret;
4059         }
4060
4061         mutex_lock(&BTRFS_I(inode)->log_mutex);
4062
4063         btrfs_get_logged_extents(inode, &logged_list, start, end);
4064
4065         /*
4066          * a brute force approach to making sure we get the most uptodate
4067          * copies of everything.
4068          */
4069         if (S_ISDIR(inode->i_mode)) {
4070                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
4071
4072                 if (inode_only == LOG_INODE_EXISTS)
4073                         max_key_type = BTRFS_XATTR_ITEM_KEY;
4074                 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
4075         } else {
4076                 if (inode_only == LOG_INODE_EXISTS) {
4077                         /*
4078                          * Make sure the new inode item we write to the log has
4079                          * the same isize as the current one (if it exists).
4080                          * This is necessary to prevent data loss after log
4081                          * replay, and also to prevent doing a wrong expanding
4082                          * truncate - for e.g. create file, write 4K into offset
4083                          * 0, fsync, write 4K into offset 4096, add hard link,
4084                          * fsync some other file (to sync log), power fail - if
4085                          * we use the inode's current i_size, after log replay
4086                          * we get a 8Kb file, with the last 4Kb extent as a hole
4087                          * (zeroes), as if an expanding truncate happened,
4088                          * instead of getting a file of 4Kb only.
4089                          */
4090                         err = logged_inode_size(log, inode, path,
4091                                                 &logged_isize);
4092                         if (err)
4093                                 goto out_unlock;
4094                 }
4095                 if (test_and_clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4096                                        &BTRFS_I(inode)->runtime_flags)) {
4097                         clear_bit(BTRFS_INODE_COPY_EVERYTHING,
4098                                   &BTRFS_I(inode)->runtime_flags);
4099                         ret = btrfs_truncate_inode_items(trans, log,
4100                                                          inode, 0, 0);
4101                 } else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
4102                                               &BTRFS_I(inode)->runtime_flags) ||
4103                            inode_only == LOG_INODE_EXISTS) {
4104                         if (inode_only == LOG_INODE_ALL)
4105                                 fast_search = true;
4106                         max_key.type = BTRFS_XATTR_ITEM_KEY;
4107                         ret = drop_objectid_items(trans, log, path, ino,
4108                                                   max_key.type);
4109                 } else {
4110                         if (inode_only == LOG_INODE_ALL)
4111                                 fast_search = true;
4112                         ret = log_inode_item(trans, log, dst_path, inode);
4113                         if (ret) {
4114                                 err = ret;
4115                                 goto out_unlock;
4116                         }
4117                         goto log_extents;
4118                 }
4119
4120         }
4121         if (ret) {
4122                 err = ret;
4123                 goto out_unlock;
4124         }
4125
4126         while (1) {
4127                 ins_nr = 0;
4128                 ret = btrfs_search_forward(root, &min_key,
4129                                            path, trans->transid);
4130                 if (ret != 0)
4131                         break;
4132 again:
4133                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
4134                 if (min_key.objectid != ino)
4135                         break;
4136                 if (min_key.type > max_key.type)
4137                         break;
4138
4139                 src = path->nodes[0];
4140                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
4141                         ins_nr++;
4142                         goto next_slot;
4143                 } else if (!ins_nr) {
4144                         ins_start_slot = path->slots[0];
4145                         ins_nr = 1;
4146                         goto next_slot;
4147                 }
4148
4149                 ret = copy_items(trans, inode, dst_path, path, &last_extent,
4150                                  ins_start_slot, ins_nr, inode_only,
4151                                  logged_isize);
4152                 if (ret < 0) {
4153                         err = ret;
4154                         goto out_unlock;
4155                 }
4156                 if (ret) {
4157                         ins_nr = 0;
4158                         btrfs_release_path(path);
4159                         continue;
4160                 }
4161                 ins_nr = 1;
4162                 ins_start_slot = path->slots[0];
4163 next_slot:
4164
4165                 nritems = btrfs_header_nritems(path->nodes[0]);
4166                 path->slots[0]++;
4167                 if (path->slots[0] < nritems) {
4168                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
4169                                               path->slots[0]);
4170                         goto again;
4171                 }
4172                 if (ins_nr) {
4173                         ret = copy_items(trans, inode, dst_path, path,
4174                                          &last_extent, ins_start_slot,
4175                                          ins_nr, inode_only, logged_isize);
4176                         if (ret < 0) {
4177                                 err = ret;
4178                                 goto out_unlock;
4179                         }
4180                         ret = 0;
4181                         ins_nr = 0;
4182                 }
4183                 btrfs_release_path(path);
4184
4185                 if (min_key.offset < (u64)-1) {
4186                         min_key.offset++;
4187                 } else if (min_key.type < max_key.type) {
4188                         min_key.type++;
4189                         min_key.offset = 0;
4190                 } else {
4191                         break;
4192                 }
4193         }
4194         if (ins_nr) {
4195                 ret = copy_items(trans, inode, dst_path, path, &last_extent,
4196                                  ins_start_slot, ins_nr, inode_only,
4197                                  logged_isize);
4198                 if (ret < 0) {
4199                         err = ret;
4200                         goto out_unlock;
4201                 }
4202                 ret = 0;
4203                 ins_nr = 0;
4204         }
4205
4206 log_extents:
4207         btrfs_release_path(path);
4208         btrfs_release_path(dst_path);
4209         if (fast_search) {
4210                 /*
4211                  * Some ordered extents started by fsync might have completed
4212                  * before we collected the ordered extents in logged_list, which
4213                  * means they're gone, not in our logged_list nor in the inode's
4214                  * ordered tree. We want the application/user space to know an
4215                  * error happened while attempting to persist file data so that
4216                  * it can take proper action. If such error happened, we leave
4217                  * without writing to the log tree and the fsync must report the
4218                  * file data write error and not commit the current transaction.
4219                  */
4220                 err = btrfs_inode_check_errors(inode);
4221                 if (err) {
4222                         ctx->io_err = err;
4223                         goto out_unlock;
4224                 }
4225                 ret = btrfs_log_changed_extents(trans, root, inode, dst_path,
4226                                                 &logged_list, ctx);
4227                 if (ret) {
4228                         err = ret;
4229                         goto out_unlock;
4230                 }
4231         } else if (inode_only == LOG_INODE_ALL) {
4232                 struct extent_map *em, *n;
4233
4234                 write_lock(&em_tree->lock);
4235                 /*
4236                  * We can't just remove every em if we're called for a ranged
4237                  * fsync - that is, one that doesn't cover the whole possible
4238                  * file range (0 to LLONG_MAX). This is because we can have
4239                  * em's that fall outside the range we're logging and therefore
4240                  * their ordered operations haven't completed yet
4241                  * (btrfs_finish_ordered_io() not invoked yet). This means we
4242                  * didn't get their respective file extent item in the fs/subvol
4243                  * tree yet, and need to let the next fast fsync (one which
4244                  * consults the list of modified extent maps) find the em so
4245                  * that it logs a matching file extent item and waits for the
4246                  * respective ordered operation to complete (if it's still
4247                  * running).
4248                  *
4249                  * Removing every em outside the range we're logging would make
4250                  * the next fast fsync not log their matching file extent items,
4251                  * therefore making us lose data after a log replay.
4252                  */
4253                 list_for_each_entry_safe(em, n, &em_tree->modified_extents,
4254                                          list) {
4255                         const u64 mod_end = em->mod_start + em->mod_len - 1;
4256
4257                         if (em->mod_start >= start && mod_end <= end)
4258                                 list_del_init(&em->list);
4259                 }
4260                 write_unlock(&em_tree->lock);
4261         }
4262
4263         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
4264                 ret = log_directory_changes(trans, root, inode, path, dst_path);
4265                 if (ret) {
4266                         err = ret;
4267                         goto out_unlock;
4268                 }
4269         }
4270
4271         BTRFS_I(inode)->logged_trans = trans->transid;
4272         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->last_sub_trans;
4273 out_unlock:
4274         if (unlikely(err))
4275                 btrfs_put_logged_extents(&logged_list);
4276         else
4277                 btrfs_submit_logged_extents(&logged_list, log);
4278         mutex_unlock(&BTRFS_I(inode)->log_mutex);
4279
4280         btrfs_free_path(path);
4281         btrfs_free_path(dst_path);
4282         return err;
4283 }
4284
4285 /*
4286  * follow the dentry parent pointers up the chain and see if any
4287  * of the directories in it require a full commit before they can
4288  * be logged.  Returns zero if nothing special needs to be done or 1 if
4289  * a full commit is required.
4290  */
4291 static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
4292                                                struct inode *inode,
4293                                                struct dentry *parent,
4294                                                struct super_block *sb,
4295                                                u64 last_committed)
4296 {
4297         int ret = 0;
4298         struct btrfs_root *root;
4299         struct dentry *old_parent = NULL;
4300         struct inode *orig_inode = inode;
4301
4302         /*
4303          * for regular files, if its inode is already on disk, we don't
4304          * have to worry about the parents at all.  This is because
4305          * we can use the last_unlink_trans field to record renames
4306          * and other fun in this file.
4307          */
4308         if (S_ISREG(inode->i_mode) &&
4309             BTRFS_I(inode)->generation <= last_committed &&
4310             BTRFS_I(inode)->last_unlink_trans <= last_committed)
4311                         goto out;
4312
4313         if (!S_ISDIR(inode->i_mode)) {
4314                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
4315                         goto out;
4316                 inode = parent->d_inode;
4317         }
4318
4319         while (1) {
4320                 /*
4321                  * If we are logging a directory then we start with our inode,
4322                  * not our parents inode, so we need to skipp setting the
4323                  * logged_trans so that further down in the log code we don't
4324                  * think this inode has already been logged.
4325                  */
4326                 if (inode != orig_inode)
4327                         BTRFS_I(inode)->logged_trans = trans->transid;
4328                 smp_mb();
4329
4330                 if (BTRFS_I(inode)->last_unlink_trans > last_committed) {
4331                         root = BTRFS_I(inode)->root;
4332
4333                         /*
4334                          * make sure any commits to the log are forced
4335                          * to be full commits
4336                          */
4337                         btrfs_set_log_full_commit(root->fs_info, trans);
4338                         ret = 1;
4339                         break;
4340                 }
4341
4342                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
4343                         break;
4344
4345                 if (IS_ROOT(parent))
4346                         break;
4347
4348                 parent = dget_parent(parent);
4349                 dput(old_parent);
4350                 old_parent = parent;
4351                 inode = parent->d_inode;
4352
4353         }
4354         dput(old_parent);
4355 out:
4356         return ret;
4357 }
4358
4359 /*
4360  * helper function around btrfs_log_inode to make sure newly created
4361  * parent directories also end up in the log.  A minimal inode and backref
4362  * only logging is done of any parent directories that are older than
4363  * the last committed transaction
4364  */
4365 static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
4366                                   struct btrfs_root *root, struct inode *inode,
4367                                   struct dentry *parent,
4368                                   const loff_t start,
4369                                   const loff_t end,
4370                                   int exists_only,
4371                                   struct btrfs_log_ctx *ctx)
4372 {
4373         int inode_only = exists_only ? LOG_INODE_EXISTS : LOG_INODE_ALL;
4374         struct super_block *sb;
4375         struct dentry *old_parent = NULL;
4376         int ret = 0;
4377         u64 last_committed = root->fs_info->last_trans_committed;
4378         const struct dentry * const first_parent = parent;
4379         const bool did_unlink = (BTRFS_I(inode)->last_unlink_trans >
4380                                  last_committed);
4381
4382         sb = inode->i_sb;
4383
4384         if (btrfs_test_opt(root, NOTREELOG)) {
4385                 ret = 1;
4386                 goto end_no_trans;
4387         }
4388
4389         /*
4390          * The prev transaction commit doesn't complete, we need do
4391          * full commit by ourselves.
4392          */
4393         if (root->fs_info->last_trans_log_full_commit >
4394             root->fs_info->last_trans_committed) {
4395                 ret = 1;
4396                 goto end_no_trans;
4397         }
4398
4399         if (root != BTRFS_I(inode)->root ||
4400             btrfs_root_refs(&root->root_item) == 0) {
4401                 ret = 1;
4402                 goto end_no_trans;
4403         }
4404
4405         ret = check_parent_dirs_for_sync(trans, inode, parent,
4406                                          sb, last_committed);
4407         if (ret)
4408                 goto end_no_trans;
4409
4410         if (btrfs_inode_in_log(inode, trans->transid)) {
4411                 ret = BTRFS_NO_LOG_SYNC;
4412                 goto end_no_trans;
4413         }
4414
4415         ret = start_log_trans(trans, root, ctx);
4416         if (ret)
4417                 goto end_no_trans;
4418
4419         ret = btrfs_log_inode(trans, root, inode, inode_only, start, end, ctx);
4420         if (ret)
4421                 goto end_trans;
4422
4423         /*
4424          * for regular files, if its inode is already on disk, we don't
4425          * have to worry about the parents at all.  This is because
4426          * we can use the last_unlink_trans field to record renames
4427          * and other fun in this file.
4428          */
4429         if (S_ISREG(inode->i_mode) &&
4430             BTRFS_I(inode)->generation <= last_committed &&
4431             BTRFS_I(inode)->last_unlink_trans <= last_committed) {
4432                 ret = 0;
4433                 goto end_trans;
4434         }
4435
4436         while (1) {
4437                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
4438                         break;
4439
4440                 inode = parent->d_inode;
4441                 if (root != BTRFS_I(inode)->root)
4442                         break;
4443
4444                 /*
4445                  * On unlink we must make sure our immediate parent directory
4446                  * inode is fully logged. This is to prevent leaving dangling
4447                  * directory index entries and a wrong directory inode's i_size.
4448                  * Not doing so can result in a directory being impossible to
4449                  * delete after log replay (rmdir will always fail with error
4450                  * -ENOTEMPTY).
4451                  */
4452                 if (did_unlink && parent == first_parent)
4453                         inode_only = LOG_INODE_ALL;
4454                 else
4455                         inode_only = LOG_INODE_EXISTS;
4456
4457                 if (BTRFS_I(inode)->generation >
4458                     root->fs_info->last_trans_committed ||
4459                     inode_only == LOG_INODE_ALL) {
4460                         ret = btrfs_log_inode(trans, root, inode, inode_only,
4461                                               0, LLONG_MAX, ctx);
4462                         if (ret)
4463                                 goto end_trans;
4464                 }
4465                 if (IS_ROOT(parent))
4466                         break;
4467
4468                 parent = dget_parent(parent);
4469                 dput(old_parent);
4470                 old_parent = parent;
4471         }
4472         ret = 0;
4473 end_trans:
4474         dput(old_parent);
4475         if (ret < 0) {
4476                 btrfs_set_log_full_commit(root->fs_info, trans);
4477                 ret = 1;
4478         }
4479
4480         if (ret)
4481                 btrfs_remove_log_ctx(root, ctx);
4482         btrfs_end_log_trans(root);
4483 end_no_trans:
4484         return ret;
4485 }
4486
4487 /*
4488  * it is not safe to log dentry if the chunk root has added new
4489  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
4490  * If this returns 1, you must commit the transaction to safely get your
4491  * data on disk.
4492  */
4493 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
4494                           struct btrfs_root *root, struct dentry *dentry,
4495                           const loff_t start,
4496                           const loff_t end,
4497                           struct btrfs_log_ctx *ctx)
4498 {
4499         struct dentry *parent = dget_parent(dentry);
4500         int ret;
4501
4502         ret = btrfs_log_inode_parent(trans, root, dentry->d_inode, parent,
4503                                      start, end, 0, ctx);
4504         dput(parent);
4505
4506         return ret;
4507 }
4508
4509 /*
4510  * should be called during mount to recover any replay any log trees
4511  * from the FS
4512  */
4513 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
4514 {
4515         int ret;
4516         struct btrfs_path *path;
4517         struct btrfs_trans_handle *trans;
4518         struct btrfs_key key;
4519         struct btrfs_key found_key;
4520         struct btrfs_key tmp_key;
4521         struct btrfs_root *log;
4522         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
4523         struct walk_control wc = {
4524                 .process_func = process_one_buffer,
4525                 .stage = 0,
4526         };
4527
4528         path = btrfs_alloc_path();
4529         if (!path)
4530                 return -ENOMEM;
4531
4532         fs_info->log_root_recovering = 1;
4533
4534         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4535         if (IS_ERR(trans)) {
4536                 ret = PTR_ERR(trans);
4537                 goto error;
4538         }
4539
4540         wc.trans = trans;
4541         wc.pin = 1;
4542
4543         ret = walk_log_tree(trans, log_root_tree, &wc);
4544         if (ret) {
4545                 btrfs_error(fs_info, ret, "Failed to pin buffers while "
4546                             "recovering log root tree.");
4547                 goto error;
4548         }
4549
4550 again:
4551         key.objectid = BTRFS_TREE_LOG_OBJECTID;
4552         key.offset = (u64)-1;
4553         key.type = BTRFS_ROOT_ITEM_KEY;
4554
4555         while (1) {
4556                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
4557
4558                 if (ret < 0) {
4559                         btrfs_error(fs_info, ret,
4560                                     "Couldn't find tree log root.");
4561                         goto error;
4562                 }
4563                 if (ret > 0) {
4564                         if (path->slots[0] == 0)
4565                                 break;
4566                         path->slots[0]--;
4567                 }
4568                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
4569                                       path->slots[0]);
4570                 btrfs_release_path(path);
4571                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4572                         break;
4573
4574                 log = btrfs_read_fs_root(log_root_tree, &found_key);
4575                 if (IS_ERR(log)) {
4576                         ret = PTR_ERR(log);
4577                         btrfs_error(fs_info, ret,
4578                                     "Couldn't read tree log root.");
4579                         goto error;
4580                 }
4581
4582                 tmp_key.objectid = found_key.offset;
4583                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
4584                 tmp_key.offset = (u64)-1;
4585
4586                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
4587                 if (IS_ERR(wc.replay_dest)) {
4588                         ret = PTR_ERR(wc.replay_dest);
4589                         free_extent_buffer(log->node);
4590                         free_extent_buffer(log->commit_root);
4591                         kfree(log);
4592                         btrfs_error(fs_info, ret, "Couldn't read target root "
4593                                     "for tree log recovery.");
4594                         goto error;
4595                 }
4596
4597                 wc.replay_dest->log_root = log;
4598                 btrfs_record_root_in_trans(trans, wc.replay_dest);
4599                 ret = walk_log_tree(trans, log, &wc);
4600
4601                 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
4602                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
4603                                                       path);
4604                 }
4605
4606                 key.offset = found_key.offset - 1;
4607                 wc.replay_dest->log_root = NULL;
4608                 free_extent_buffer(log->node);
4609                 free_extent_buffer(log->commit_root);
4610                 kfree(log);
4611
4612                 if (ret)
4613                         goto error;
4614
4615                 if (found_key.offset == 0)
4616                         break;
4617         }
4618         btrfs_release_path(path);
4619
4620         /* step one is to pin it all, step two is to replay just inodes */
4621         if (wc.pin) {
4622                 wc.pin = 0;
4623                 wc.process_func = replay_one_buffer;
4624                 wc.stage = LOG_WALK_REPLAY_INODES;
4625                 goto again;
4626         }
4627         /* step three is to replay everything */
4628         if (wc.stage < LOG_WALK_REPLAY_ALL) {
4629                 wc.stage++;
4630                 goto again;
4631         }
4632
4633         btrfs_free_path(path);
4634
4635         /* step 4: commit the transaction, which also unpins the blocks */
4636         ret = btrfs_commit_transaction(trans, fs_info->tree_root);
4637         if (ret)
4638                 return ret;
4639
4640         free_extent_buffer(log_root_tree->node);
4641         log_root_tree->log_root = NULL;
4642         fs_info->log_root_recovering = 0;
4643         kfree(log_root_tree);
4644
4645         return 0;
4646 error:
4647         if (wc.trans)
4648                 btrfs_end_transaction(wc.trans, fs_info->tree_root);
4649         btrfs_free_path(path);
4650         return ret;
4651 }
4652
4653 /*
4654  * there are some corner cases where we want to force a full
4655  * commit instead of allowing a directory to be logged.
4656  *
4657  * They revolve around files there were unlinked from the directory, and
4658  * this function updates the parent directory so that a full commit is
4659  * properly done if it is fsync'd later after the unlinks are done.
4660  */
4661 void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
4662                              struct inode *dir, struct inode *inode,
4663                              int for_rename)
4664 {
4665         /*
4666          * when we're logging a file, if it hasn't been renamed
4667          * or unlinked, and its inode is fully committed on disk,
4668          * we don't have to worry about walking up the directory chain
4669          * to log its parents.
4670          *
4671          * So, we use the last_unlink_trans field to put this transid
4672          * into the file.  When the file is logged we check it and
4673          * don't log the parents if the file is fully on disk.
4674          */
4675         if (S_ISREG(inode->i_mode))
4676                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4677
4678         /*
4679          * if this directory was already logged any new
4680          * names for this file/dir will get recorded
4681          */
4682         smp_mb();
4683         if (BTRFS_I(dir)->logged_trans == trans->transid)
4684                 return;
4685
4686         /*
4687          * if the inode we're about to unlink was logged,
4688          * the log will be properly updated for any new names
4689          */
4690         if (BTRFS_I(inode)->logged_trans == trans->transid)
4691                 return;
4692
4693         /*
4694          * when renaming files across directories, if the directory
4695          * there we're unlinking from gets fsync'd later on, there's
4696          * no way to find the destination directory later and fsync it
4697          * properly.  So, we have to be conservative and force commits
4698          * so the new name gets discovered.
4699          */
4700         if (for_rename)
4701                 goto record;
4702
4703         /* we can safely do the unlink without any special recording */
4704         return;
4705
4706 record:
4707         BTRFS_I(dir)->last_unlink_trans = trans->transid;
4708 }
4709
4710 /*
4711  * Call this after adding a new name for a file and it will properly
4712  * update the log to reflect the new name.
4713  *
4714  * It will return zero if all goes well, and it will return 1 if a
4715  * full transaction commit is required.
4716  */
4717 int btrfs_log_new_name(struct btrfs_trans_handle *trans,
4718                         struct inode *inode, struct inode *old_dir,
4719                         struct dentry *parent)
4720 {
4721         struct btrfs_root * root = BTRFS_I(inode)->root;
4722
4723         /*
4724          * this will force the logging code to walk the dentry chain
4725          * up for the file
4726          */
4727         if (S_ISREG(inode->i_mode))
4728                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4729
4730         /*
4731          * if this inode hasn't been logged and directory we're renaming it
4732          * from hasn't been logged, we don't need to log it
4733          */
4734         if (BTRFS_I(inode)->logged_trans <=
4735             root->fs_info->last_trans_committed &&
4736             (!old_dir || BTRFS_I(old_dir)->logged_trans <=
4737                     root->fs_info->last_trans_committed))
4738                 return 0;
4739
4740         return btrfs_log_inode_parent(trans, root, inode, parent, 0,
4741                                       LLONG_MAX, 1, NULL);
4742 }
4743