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                 blk_finish_plug(&plug);
2639                 mutex_unlock(&log_root_tree->log_mutex);
2640                 ret = root_log_ctx.log_ret;
2641                 goto out;
2642         }
2643
2644         index2 = root_log_ctx.log_transid % 2;
2645         if (atomic_read(&log_root_tree->log_commit[index2])) {
2646                 blk_finish_plug(&plug);
2647                 ret = btrfs_wait_marked_extents(log, &log->dirty_log_pages,
2648                                                 mark);
2649                 btrfs_wait_logged_extents(trans, log, log_transid);
2650                 wait_log_commit(trans, log_root_tree,
2651                                 root_log_ctx.log_transid);
2652                 mutex_unlock(&log_root_tree->log_mutex);
2653                 if (!ret)
2654                         ret = root_log_ctx.log_ret;
2655                 goto out;
2656         }
2657         ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
2658         atomic_set(&log_root_tree->log_commit[index2], 1);
2659
2660         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2661                 wait_log_commit(trans, log_root_tree,
2662                                 root_log_ctx.log_transid - 1);
2663         }
2664
2665         wait_for_writer(trans, log_root_tree);
2666
2667         /*
2668          * now that we've moved on to the tree of log tree roots,
2669          * check the full commit flag again
2670          */
2671         if (btrfs_need_log_full_commit(root->fs_info, trans)) {
2672                 blk_finish_plug(&plug);
2673                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2674                 btrfs_free_logged_extents(log, log_transid);
2675                 mutex_unlock(&log_root_tree->log_mutex);
2676                 ret = -EAGAIN;
2677                 goto out_wake_log_root;
2678         }
2679
2680         ret = btrfs_write_marked_extents(log_root_tree,
2681                                          &log_root_tree->dirty_log_pages,
2682                                          EXTENT_DIRTY | EXTENT_NEW);
2683         blk_finish_plug(&plug);
2684         if (ret) {
2685                 btrfs_set_log_full_commit(root->fs_info, trans);
2686                 btrfs_abort_transaction(trans, root, ret);
2687                 btrfs_free_logged_extents(log, log_transid);
2688                 mutex_unlock(&log_root_tree->log_mutex);
2689                 goto out_wake_log_root;
2690         }
2691         ret = btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2692         if (!ret)
2693                 ret = btrfs_wait_marked_extents(log_root_tree,
2694                                                 &log_root_tree->dirty_log_pages,
2695                                                 EXTENT_NEW | EXTENT_DIRTY);
2696         if (ret) {
2697                 btrfs_set_log_full_commit(root->fs_info, trans);
2698                 btrfs_free_logged_extents(log, log_transid);
2699                 mutex_unlock(&log_root_tree->log_mutex);
2700                 goto out_wake_log_root;
2701         }
2702         btrfs_wait_logged_extents(trans, log, log_transid);
2703
2704         btrfs_set_super_log_root(root->fs_info->super_for_commit,
2705                                 log_root_tree->node->start);
2706         btrfs_set_super_log_root_level(root->fs_info->super_for_commit,
2707                                 btrfs_header_level(log_root_tree->node));
2708
2709         log_root_tree->log_transid++;
2710         mutex_unlock(&log_root_tree->log_mutex);
2711
2712         /*
2713          * nobody else is going to jump in and write the the ctree
2714          * super here because the log_commit atomic below is protecting
2715          * us.  We must be called with a transaction handle pinning
2716          * the running transaction open, so a full commit can't hop
2717          * in and cause problems either.
2718          */
2719         ret = write_ctree_super(trans, root->fs_info->tree_root, 1);
2720         if (ret) {
2721                 btrfs_set_log_full_commit(root->fs_info, trans);
2722                 btrfs_abort_transaction(trans, root, ret);
2723                 goto out_wake_log_root;
2724         }
2725
2726         mutex_lock(&root->log_mutex);
2727         if (root->last_log_commit < log_transid)
2728                 root->last_log_commit = log_transid;
2729         mutex_unlock(&root->log_mutex);
2730
2731 out_wake_log_root:
2732         /*
2733          * We needn't get log_mutex here because we are sure all
2734          * the other tasks are blocked.
2735          */
2736         btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
2737
2738         mutex_lock(&log_root_tree->log_mutex);
2739         log_root_tree->log_transid_committed++;
2740         atomic_set(&log_root_tree->log_commit[index2], 0);
2741         mutex_unlock(&log_root_tree->log_mutex);
2742
2743         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2744                 wake_up(&log_root_tree->log_commit_wait[index2]);
2745 out:
2746         /* See above. */
2747         btrfs_remove_all_log_ctxs(root, index1, ret);
2748
2749         mutex_lock(&root->log_mutex);
2750         root->log_transid_committed++;
2751         atomic_set(&root->log_commit[index1], 0);
2752         mutex_unlock(&root->log_mutex);
2753
2754         if (waitqueue_active(&root->log_commit_wait[index1]))
2755                 wake_up(&root->log_commit_wait[index1]);
2756         return ret;
2757 }
2758
2759 static void free_log_tree(struct btrfs_trans_handle *trans,
2760                           struct btrfs_root *log)
2761 {
2762         int ret;
2763         u64 start;
2764         u64 end;
2765         struct walk_control wc = {
2766                 .free = 1,
2767                 .process_func = process_one_buffer
2768         };
2769
2770         ret = walk_log_tree(trans, log, &wc);
2771         /* I don't think this can happen but just in case */
2772         if (ret)
2773                 btrfs_abort_transaction(trans, log, ret);
2774
2775         while (1) {
2776                 ret = find_first_extent_bit(&log->dirty_log_pages,
2777                                 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW,
2778                                 NULL);
2779                 if (ret)
2780                         break;
2781
2782                 clear_extent_bits(&log->dirty_log_pages, start, end,
2783                                   EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2784         }
2785
2786         /*
2787          * We may have short-circuited the log tree with the full commit logic
2788          * and left ordered extents on our list, so clear these out to keep us
2789          * from leaking inodes and memory.
2790          */
2791         btrfs_free_logged_extents(log, 0);
2792         btrfs_free_logged_extents(log, 1);
2793
2794         free_extent_buffer(log->node);
2795         kfree(log);
2796 }
2797
2798 /*
2799  * free all the extents used by the tree log.  This should be called
2800  * at commit time of the full transaction
2801  */
2802 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2803 {
2804         if (root->log_root) {
2805                 free_log_tree(trans, root->log_root);
2806                 root->log_root = NULL;
2807         }
2808         return 0;
2809 }
2810
2811 int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2812                              struct btrfs_fs_info *fs_info)
2813 {
2814         if (fs_info->log_root_tree) {
2815                 free_log_tree(trans, fs_info->log_root_tree);
2816                 fs_info->log_root_tree = NULL;
2817         }
2818         return 0;
2819 }
2820
2821 /*
2822  * If both a file and directory are logged, and unlinks or renames are
2823  * mixed in, we have a few interesting corners:
2824  *
2825  * create file X in dir Y
2826  * link file X to X.link in dir Y
2827  * fsync file X
2828  * unlink file X but leave X.link
2829  * fsync dir Y
2830  *
2831  * After a crash we would expect only X.link to exist.  But file X
2832  * didn't get fsync'd again so the log has back refs for X and X.link.
2833  *
2834  * We solve this by removing directory entries and inode backrefs from the
2835  * log when a file that was logged in the current transaction is
2836  * unlinked.  Any later fsync will include the updated log entries, and
2837  * we'll be able to reconstruct the proper directory items from backrefs.
2838  *
2839  * This optimizations allows us to avoid relogging the entire inode
2840  * or the entire directory.
2841  */
2842 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2843                                  struct btrfs_root *root,
2844                                  const char *name, int name_len,
2845                                  struct inode *dir, u64 index)
2846 {
2847         struct btrfs_root *log;
2848         struct btrfs_dir_item *di;
2849         struct btrfs_path *path;
2850         int ret;
2851         int err = 0;
2852         int bytes_del = 0;
2853         u64 dir_ino = btrfs_ino(dir);
2854
2855         if (BTRFS_I(dir)->logged_trans < trans->transid)
2856                 return 0;
2857
2858         ret = join_running_log_trans(root);
2859         if (ret)
2860                 return 0;
2861
2862         mutex_lock(&BTRFS_I(dir)->log_mutex);
2863
2864         log = root->log_root;
2865         path = btrfs_alloc_path();
2866         if (!path) {
2867                 err = -ENOMEM;
2868                 goto out_unlock;
2869         }
2870
2871         di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2872                                    name, name_len, -1);
2873         if (IS_ERR(di)) {
2874                 err = PTR_ERR(di);
2875                 goto fail;
2876         }
2877         if (di) {
2878                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2879                 bytes_del += name_len;
2880                 if (ret) {
2881                         err = ret;
2882                         goto fail;
2883                 }
2884         }
2885         btrfs_release_path(path);
2886         di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2887                                          index, name, name_len, -1);
2888         if (IS_ERR(di)) {
2889                 err = PTR_ERR(di);
2890                 goto fail;
2891         }
2892         if (di) {
2893                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2894                 bytes_del += name_len;
2895                 if (ret) {
2896                         err = ret;
2897                         goto fail;
2898                 }
2899         }
2900
2901         /* update the directory size in the log to reflect the names
2902          * we have removed
2903          */
2904         if (bytes_del) {
2905                 struct btrfs_key key;
2906
2907                 key.objectid = dir_ino;
2908                 key.offset = 0;
2909                 key.type = BTRFS_INODE_ITEM_KEY;
2910                 btrfs_release_path(path);
2911
2912                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2913                 if (ret < 0) {
2914                         err = ret;
2915                         goto fail;
2916                 }
2917                 if (ret == 0) {
2918                         struct btrfs_inode_item *item;
2919                         u64 i_size;
2920
2921                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2922                                               struct btrfs_inode_item);
2923                         i_size = btrfs_inode_size(path->nodes[0], item);
2924                         if (i_size > bytes_del)
2925                                 i_size -= bytes_del;
2926                         else
2927                                 i_size = 0;
2928                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2929                         btrfs_mark_buffer_dirty(path->nodes[0]);
2930                 } else
2931                         ret = 0;
2932                 btrfs_release_path(path);
2933         }
2934 fail:
2935         btrfs_free_path(path);
2936 out_unlock:
2937         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2938         if (ret == -ENOSPC) {
2939                 btrfs_set_log_full_commit(root->fs_info, trans);
2940                 ret = 0;
2941         } else if (ret < 0)
2942                 btrfs_abort_transaction(trans, root, ret);
2943
2944         btrfs_end_log_trans(root);
2945
2946         return err;
2947 }
2948
2949 /* see comments for btrfs_del_dir_entries_in_log */
2950 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2951                                struct btrfs_root *root,
2952                                const char *name, int name_len,
2953                                struct inode *inode, u64 dirid)
2954 {
2955         struct btrfs_root *log;
2956         u64 index;
2957         int ret;
2958
2959         if (BTRFS_I(inode)->logged_trans < trans->transid)
2960                 return 0;
2961
2962         ret = join_running_log_trans(root);
2963         if (ret)
2964                 return 0;
2965         log = root->log_root;
2966         mutex_lock(&BTRFS_I(inode)->log_mutex);
2967
2968         ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
2969                                   dirid, &index);
2970         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2971         if (ret == -ENOSPC) {
2972                 btrfs_set_log_full_commit(root->fs_info, trans);
2973                 ret = 0;
2974         } else if (ret < 0 && ret != -ENOENT)
2975                 btrfs_abort_transaction(trans, root, ret);
2976         btrfs_end_log_trans(root);
2977
2978         return ret;
2979 }
2980
2981 /*
2982  * creates a range item in the log for 'dirid'.  first_offset and
2983  * last_offset tell us which parts of the key space the log should
2984  * be considered authoritative for.
2985  */
2986 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2987                                        struct btrfs_root *log,
2988                                        struct btrfs_path *path,
2989                                        int key_type, u64 dirid,
2990                                        u64 first_offset, u64 last_offset)
2991 {
2992         int ret;
2993         struct btrfs_key key;
2994         struct btrfs_dir_log_item *item;
2995
2996         key.objectid = dirid;
2997         key.offset = first_offset;
2998         if (key_type == BTRFS_DIR_ITEM_KEY)
2999                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
3000         else
3001                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
3002         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
3003         if (ret)
3004                 return ret;
3005
3006         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3007                               struct btrfs_dir_log_item);
3008         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
3009         btrfs_mark_buffer_dirty(path->nodes[0]);
3010         btrfs_release_path(path);
3011         return 0;
3012 }
3013
3014 /*
3015  * log all the items included in the current transaction for a given
3016  * directory.  This also creates the range items in the log tree required
3017  * to replay anything deleted before the fsync
3018  */
3019 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
3020                           struct btrfs_root *root, struct inode *inode,
3021                           struct btrfs_path *path,
3022                           struct btrfs_path *dst_path, int key_type,
3023                           u64 min_offset, u64 *last_offset_ret)
3024 {
3025         struct btrfs_key min_key;
3026         struct btrfs_root *log = root->log_root;
3027         struct extent_buffer *src;
3028         int err = 0;
3029         int ret;
3030         int i;
3031         int nritems;
3032         u64 first_offset = min_offset;
3033         u64 last_offset = (u64)-1;
3034         u64 ino = btrfs_ino(inode);
3035
3036         log = root->log_root;
3037
3038         min_key.objectid = ino;
3039         min_key.type = key_type;
3040         min_key.offset = min_offset;
3041
3042         ret = btrfs_search_forward(root, &min_key, path, trans->transid);
3043
3044         /*
3045          * we didn't find anything from this transaction, see if there
3046          * is anything at all
3047          */
3048         if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
3049                 min_key.objectid = ino;
3050                 min_key.type = key_type;
3051                 min_key.offset = (u64)-1;
3052                 btrfs_release_path(path);
3053                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3054                 if (ret < 0) {
3055                         btrfs_release_path(path);
3056                         return ret;
3057                 }
3058                 ret = btrfs_previous_item(root, path, ino, key_type);
3059
3060                 /* if ret == 0 there are items for this type,
3061                  * create a range to tell us the last key of this type.
3062                  * otherwise, there are no items in this directory after
3063                  * *min_offset, and we create a range to indicate that.
3064                  */
3065                 if (ret == 0) {
3066                         struct btrfs_key tmp;
3067                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
3068                                               path->slots[0]);
3069                         if (key_type == tmp.type)
3070                                 first_offset = max(min_offset, tmp.offset) + 1;
3071                 }
3072                 goto done;
3073         }
3074
3075         /* go backward to find any previous key */
3076         ret = btrfs_previous_item(root, path, ino, key_type);
3077         if (ret == 0) {
3078                 struct btrfs_key tmp;
3079                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3080                 if (key_type == tmp.type) {
3081                         first_offset = tmp.offset;
3082                         ret = overwrite_item(trans, log, dst_path,
3083                                              path->nodes[0], path->slots[0],
3084                                              &tmp);
3085                         if (ret) {
3086                                 err = ret;
3087                                 goto done;
3088                         }
3089                 }
3090         }
3091         btrfs_release_path(path);
3092
3093         /* find the first key from this transaction again */
3094         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3095         if (WARN_ON(ret != 0))
3096                 goto done;
3097
3098         /*
3099          * we have a block from this transaction, log every item in it
3100          * from our directory
3101          */
3102         while (1) {
3103                 struct btrfs_key tmp;
3104                 src = path->nodes[0];
3105                 nritems = btrfs_header_nritems(src);
3106                 for (i = path->slots[0]; i < nritems; i++) {
3107                         btrfs_item_key_to_cpu(src, &min_key, i);
3108
3109                         if (min_key.objectid != ino || min_key.type != key_type)
3110                                 goto done;
3111                         ret = overwrite_item(trans, log, dst_path, src, i,
3112                                              &min_key);
3113                         if (ret) {
3114                                 err = ret;
3115                                 goto done;
3116                         }
3117                 }
3118                 path->slots[0] = nritems;
3119
3120                 /*
3121                  * look ahead to the next item and see if it is also
3122                  * from this directory and from this transaction
3123                  */
3124                 ret = btrfs_next_leaf(root, path);
3125                 if (ret == 1) {
3126                         last_offset = (u64)-1;
3127                         goto done;
3128                 }
3129                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3130                 if (tmp.objectid != ino || tmp.type != key_type) {
3131                         last_offset = (u64)-1;
3132                         goto done;
3133                 }
3134                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
3135                         ret = overwrite_item(trans, log, dst_path,
3136                                              path->nodes[0], path->slots[0],
3137                                              &tmp);
3138                         if (ret)
3139                                 err = ret;
3140                         else
3141                                 last_offset = tmp.offset;
3142                         goto done;
3143                 }
3144         }
3145 done:
3146         btrfs_release_path(path);
3147         btrfs_release_path(dst_path);
3148
3149         if (err == 0) {
3150                 *last_offset_ret = last_offset;
3151                 /*
3152                  * insert the log range keys to indicate where the log
3153                  * is valid
3154                  */
3155                 ret = insert_dir_log_key(trans, log, path, key_type,
3156                                          ino, first_offset, last_offset);
3157                 if (ret)
3158                         err = ret;
3159         }
3160         return err;
3161 }
3162
3163 /*
3164  * logging directories is very similar to logging inodes, We find all the items
3165  * from the current transaction and write them to the log.
3166  *
3167  * The recovery code scans the directory in the subvolume, and if it finds a
3168  * key in the range logged that is not present in the log tree, then it means
3169  * that dir entry was unlinked during the transaction.
3170  *
3171  * In order for that scan to work, we must include one key smaller than
3172  * the smallest logged by this transaction and one key larger than the largest
3173  * key logged by this transaction.
3174  */
3175 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
3176                           struct btrfs_root *root, struct inode *inode,
3177                           struct btrfs_path *path,
3178                           struct btrfs_path *dst_path)
3179 {
3180         u64 min_key;
3181         u64 max_key;
3182         int ret;
3183         int key_type = BTRFS_DIR_ITEM_KEY;
3184
3185 again:
3186         min_key = 0;
3187         max_key = 0;
3188         while (1) {
3189                 ret = log_dir_items(trans, root, inode, path,
3190                                     dst_path, key_type, min_key,
3191                                     &max_key);
3192                 if (ret)
3193                         return ret;
3194                 if (max_key == (u64)-1)
3195                         break;
3196                 min_key = max_key + 1;
3197         }
3198
3199         if (key_type == BTRFS_DIR_ITEM_KEY) {
3200                 key_type = BTRFS_DIR_INDEX_KEY;
3201                 goto again;
3202         }
3203         return 0;
3204 }
3205
3206 /*
3207  * a helper function to drop items from the log before we relog an
3208  * inode.  max_key_type indicates the highest item type to remove.
3209  * This cannot be run for file data extents because it does not
3210  * free the extents they point to.
3211  */
3212 static int drop_objectid_items(struct btrfs_trans_handle *trans,
3213                                   struct btrfs_root *log,
3214                                   struct btrfs_path *path,
3215                                   u64 objectid, int max_key_type)
3216 {
3217         int ret;
3218         struct btrfs_key key;
3219         struct btrfs_key found_key;
3220         int start_slot;
3221
3222         key.objectid = objectid;
3223         key.type = max_key_type;
3224         key.offset = (u64)-1;
3225
3226         while (1) {
3227                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
3228                 BUG_ON(ret == 0); /* Logic error */
3229                 if (ret < 0)
3230                         break;
3231
3232                 if (path->slots[0] == 0)
3233                         break;
3234
3235                 path->slots[0]--;
3236                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
3237                                       path->slots[0]);
3238
3239                 if (found_key.objectid != objectid)
3240                         break;
3241
3242                 found_key.offset = 0;
3243                 found_key.type = 0;
3244                 ret = btrfs_bin_search(path->nodes[0], &found_key, 0,
3245                                        &start_slot);
3246
3247                 ret = btrfs_del_items(trans, log, path, start_slot,
3248                                       path->slots[0] - start_slot + 1);
3249                 /*
3250                  * If start slot isn't 0 then we don't need to re-search, we've
3251                  * found the last guy with the objectid in this tree.
3252                  */
3253                 if (ret || start_slot != 0)
3254                         break;
3255                 btrfs_release_path(path);
3256         }
3257         btrfs_release_path(path);
3258         if (ret > 0)
3259                 ret = 0;
3260         return ret;
3261 }
3262
3263 static void fill_inode_item(struct btrfs_trans_handle *trans,
3264                             struct extent_buffer *leaf,
3265                             struct btrfs_inode_item *item,
3266                             struct inode *inode, int log_inode_only,
3267                             u64 logged_isize)
3268 {
3269         struct btrfs_map_token token;
3270
3271         btrfs_init_map_token(&token);
3272
3273         if (log_inode_only) {
3274                 /* set the generation to zero so the recover code
3275                  * can tell the difference between an logging
3276                  * just to say 'this inode exists' and a logging
3277                  * to say 'update this inode with these values'
3278                  */
3279                 btrfs_set_token_inode_generation(leaf, item, 0, &token);
3280                 btrfs_set_token_inode_size(leaf, item, logged_isize, &token);
3281         } else {
3282                 btrfs_set_token_inode_generation(leaf, item,
3283                                                  BTRFS_I(inode)->generation,
3284                                                  &token);
3285                 btrfs_set_token_inode_size(leaf, item, inode->i_size, &token);
3286         }
3287
3288         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3289         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3290         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3291         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3292
3293         btrfs_set_token_timespec_sec(leaf, &item->atime,
3294                                      inode->i_atime.tv_sec, &token);
3295         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3296                                       inode->i_atime.tv_nsec, &token);
3297
3298         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3299                                      inode->i_mtime.tv_sec, &token);
3300         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3301                                       inode->i_mtime.tv_nsec, &token);
3302
3303         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3304                                      inode->i_ctime.tv_sec, &token);
3305         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3306                                       inode->i_ctime.tv_nsec, &token);
3307
3308         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3309                                      &token);
3310
3311         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3312         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3313         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3314         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3315         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3316 }
3317
3318 static int log_inode_item(struct btrfs_trans_handle *trans,
3319                           struct btrfs_root *log, struct btrfs_path *path,
3320                           struct inode *inode)
3321 {
3322         struct btrfs_inode_item *inode_item;
3323         int ret;
3324
3325         ret = btrfs_insert_empty_item(trans, log, path,
3326                                       &BTRFS_I(inode)->location,
3327                                       sizeof(*inode_item));
3328         if (ret && ret != -EEXIST)
3329                 return ret;
3330         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3331                                     struct btrfs_inode_item);
3332         fill_inode_item(trans, path->nodes[0], inode_item, inode, 0, 0);
3333         btrfs_release_path(path);
3334         return 0;
3335 }
3336
3337 static noinline int copy_items(struct btrfs_trans_handle *trans,
3338                                struct inode *inode,
3339                                struct btrfs_path *dst_path,
3340                                struct btrfs_path *src_path, u64 *last_extent,
3341                                int start_slot, int nr, int inode_only,
3342                                u64 logged_isize)
3343 {
3344         unsigned long src_offset;
3345         unsigned long dst_offset;
3346         struct btrfs_root *log = BTRFS_I(inode)->root->log_root;
3347         struct btrfs_file_extent_item *extent;
3348         struct btrfs_inode_item *inode_item;
3349         struct extent_buffer *src = src_path->nodes[0];
3350         struct btrfs_key first_key, last_key, key;
3351         int ret;
3352         struct btrfs_key *ins_keys;
3353         u32 *ins_sizes;
3354         char *ins_data;
3355         int i;
3356         struct list_head ordered_sums;
3357         int skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3358         bool has_extents = false;
3359         bool need_find_last_extent = true;
3360         bool done = false;
3361
3362         INIT_LIST_HEAD(&ordered_sums);
3363
3364         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
3365                            nr * sizeof(u32), GFP_NOFS);
3366         if (!ins_data)
3367                 return -ENOMEM;
3368
3369         first_key.objectid = (u64)-1;
3370
3371         ins_sizes = (u32 *)ins_data;
3372         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
3373
3374         for (i = 0; i < nr; i++) {
3375                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
3376                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
3377         }
3378         ret = btrfs_insert_empty_items(trans, log, dst_path,
3379                                        ins_keys, ins_sizes, nr);
3380         if (ret) {
3381                 kfree(ins_data);
3382                 return ret;
3383         }
3384
3385         for (i = 0; i < nr; i++, dst_path->slots[0]++) {
3386                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
3387                                                    dst_path->slots[0]);
3388
3389                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
3390
3391                 if ((i == (nr - 1)))
3392                         last_key = ins_keys[i];
3393
3394                 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
3395                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
3396                                                     dst_path->slots[0],
3397                                                     struct btrfs_inode_item);
3398                         fill_inode_item(trans, dst_path->nodes[0], inode_item,
3399                                         inode, inode_only == LOG_INODE_EXISTS,
3400                                         logged_isize);
3401                 } else {
3402                         copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
3403                                            src_offset, ins_sizes[i]);
3404                 }
3405
3406                 /*
3407                  * We set need_find_last_extent here in case we know we were
3408                  * processing other items and then walk into the first extent in
3409                  * the inode.  If we don't hit an extent then nothing changes,
3410                  * we'll do the last search the next time around.
3411                  */
3412                 if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY) {
3413                         has_extents = true;
3414                         if (first_key.objectid == (u64)-1)
3415                                 first_key = ins_keys[i];
3416                 } else {
3417                         need_find_last_extent = false;
3418                 }
3419
3420                 /* take a reference on file data extents so that truncates
3421                  * or deletes of this inode don't have to relog the inode
3422                  * again
3423                  */
3424                 if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY &&
3425                     !skip_csum) {
3426                         int found_type;
3427                         extent = btrfs_item_ptr(src, start_slot + i,
3428                                                 struct btrfs_file_extent_item);
3429
3430                         if (btrfs_file_extent_generation(src, extent) < trans->transid)
3431                                 continue;
3432
3433                         found_type = btrfs_file_extent_type(src, extent);
3434                         if (found_type == BTRFS_FILE_EXTENT_REG) {
3435                                 u64 ds, dl, cs, cl;
3436                                 ds = btrfs_file_extent_disk_bytenr(src,
3437                                                                 extent);
3438                                 /* ds == 0 is a hole */
3439                                 if (ds == 0)
3440                                         continue;
3441
3442                                 dl = btrfs_file_extent_disk_num_bytes(src,
3443                                                                 extent);
3444                                 cs = btrfs_file_extent_offset(src, extent);
3445                                 cl = btrfs_file_extent_num_bytes(src,
3446                                                                 extent);
3447                                 if (btrfs_file_extent_compression(src,
3448                                                                   extent)) {
3449                                         cs = 0;
3450                                         cl = dl;
3451                                 }
3452
3453                                 ret = btrfs_lookup_csums_range(
3454                                                 log->fs_info->csum_root,
3455                                                 ds + cs, ds + cs + cl - 1,
3456                                                 &ordered_sums, 0);
3457                                 if (ret) {
3458                                         btrfs_release_path(dst_path);
3459                                         kfree(ins_data);
3460                                         return ret;
3461                                 }
3462                         }
3463                 }
3464         }
3465
3466         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
3467         btrfs_release_path(dst_path);
3468         kfree(ins_data);
3469
3470         /*
3471          * we have to do this after the loop above to avoid changing the
3472          * log tree while trying to change the log tree.
3473          */
3474         ret = 0;
3475         while (!list_empty(&ordered_sums)) {
3476                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3477                                                    struct btrfs_ordered_sum,
3478                                                    list);
3479                 if (!ret)
3480                         ret = btrfs_csum_file_blocks(trans, log, sums);
3481                 list_del(&sums->list);
3482                 kfree(sums);
3483         }
3484
3485         if (!has_extents)
3486                 return ret;
3487
3488         if (need_find_last_extent && *last_extent == first_key.offset) {
3489                 /*
3490                  * We don't have any leafs between our current one and the one
3491                  * we processed before that can have file extent items for our
3492                  * inode (and have a generation number smaller than our current
3493                  * transaction id).
3494                  */
3495                 need_find_last_extent = false;
3496         }
3497
3498         /*
3499          * Because we use btrfs_search_forward we could skip leaves that were
3500          * not modified and then assume *last_extent is valid when it really
3501          * isn't.  So back up to the previous leaf and read the end of the last
3502          * extent before we go and fill in holes.
3503          */
3504         if (need_find_last_extent) {
3505                 u64 len;
3506
3507                 ret = btrfs_prev_leaf(BTRFS_I(inode)->root, src_path);
3508                 if (ret < 0)
3509                         return ret;
3510                 if (ret)
3511                         goto fill_holes;
3512                 if (src_path->slots[0])
3513                         src_path->slots[0]--;
3514                 src = src_path->nodes[0];
3515                 btrfs_item_key_to_cpu(src, &key, src_path->slots[0]);
3516                 if (key.objectid != btrfs_ino(inode) ||
3517                     key.type != BTRFS_EXTENT_DATA_KEY)
3518                         goto fill_holes;
3519                 extent = btrfs_item_ptr(src, src_path->slots[0],
3520                                         struct btrfs_file_extent_item);
3521                 if (btrfs_file_extent_type(src, extent) ==
3522                     BTRFS_FILE_EXTENT_INLINE) {
3523                         len = btrfs_file_extent_inline_len(src,
3524                                                            src_path->slots[0],
3525                                                            extent);
3526                         *last_extent = ALIGN(key.offset + len,
3527                                              log->sectorsize);
3528                 } else {
3529                         len = btrfs_file_extent_num_bytes(src, extent);
3530                         *last_extent = key.offset + len;
3531                 }
3532         }
3533 fill_holes:
3534         /* So we did prev_leaf, now we need to move to the next leaf, but a few
3535          * things could have happened
3536          *
3537          * 1) A merge could have happened, so we could currently be on a leaf
3538          * that holds what we were copying in the first place.
3539          * 2) A split could have happened, and now not all of the items we want
3540          * are on the same leaf.
3541          *
3542          * So we need to adjust how we search for holes, we need to drop the
3543          * path and re-search for the first extent key we found, and then walk
3544          * forward until we hit the last one we copied.
3545          */
3546         if (need_find_last_extent) {
3547                 /* btrfs_prev_leaf could return 1 without releasing the path */
3548                 btrfs_release_path(src_path);
3549                 ret = btrfs_search_slot(NULL, BTRFS_I(inode)->root, &first_key,
3550                                         src_path, 0, 0);
3551                 if (ret < 0)
3552                         return ret;
3553                 ASSERT(ret == 0);
3554                 src = src_path->nodes[0];
3555                 i = src_path->slots[0];
3556         } else {
3557                 i = start_slot;
3558         }
3559
3560         /*
3561          * Ok so here we need to go through and fill in any holes we may have
3562          * to make sure that holes are punched for those areas in case they had
3563          * extents previously.
3564          */
3565         while (!done) {
3566                 u64 offset, len;
3567                 u64 extent_end;
3568
3569                 if (i >= btrfs_header_nritems(src_path->nodes[0])) {
3570                         ret = btrfs_next_leaf(BTRFS_I(inode)->root, src_path);
3571                         if (ret < 0)
3572                                 return ret;
3573                         ASSERT(ret == 0);
3574                         src = src_path->nodes[0];
3575                         i = 0;
3576                 }
3577
3578                 btrfs_item_key_to_cpu(src, &key, i);
3579                 if (!btrfs_comp_cpu_keys(&key, &last_key))
3580                         done = true;
3581                 if (key.objectid != btrfs_ino(inode) ||
3582                     key.type != BTRFS_EXTENT_DATA_KEY) {
3583                         i++;
3584                         continue;
3585                 }
3586                 extent = btrfs_item_ptr(src, i, struct btrfs_file_extent_item);
3587                 if (btrfs_file_extent_type(src, extent) ==
3588                     BTRFS_FILE_EXTENT_INLINE) {
3589                         len = btrfs_file_extent_inline_len(src, i, extent);
3590                         extent_end = ALIGN(key.offset + len, log->sectorsize);
3591                 } else {
3592                         len = btrfs_file_extent_num_bytes(src, extent);
3593                         extent_end = key.offset + len;
3594                 }
3595                 i++;
3596
3597                 if (*last_extent == key.offset) {
3598                         *last_extent = extent_end;
3599                         continue;
3600                 }
3601                 offset = *last_extent;
3602                 len = key.offset - *last_extent;
3603                 ret = btrfs_insert_file_extent(trans, log, btrfs_ino(inode),
3604                                                offset, 0, 0, len, 0, len, 0,
3605                                                0, 0);
3606                 if (ret)
3607                         break;
3608                 *last_extent = extent_end;
3609         }
3610         /*
3611          * Need to let the callers know we dropped the path so they should
3612          * re-search.
3613          */
3614         if (!ret && need_find_last_extent)
3615                 ret = 1;
3616         return ret;
3617 }
3618
3619 static int extent_cmp(void *priv, struct list_head *a, struct list_head *b)
3620 {
3621         struct extent_map *em1, *em2;
3622
3623         em1 = list_entry(a, struct extent_map, list);
3624         em2 = list_entry(b, struct extent_map, list);
3625
3626         if (em1->start < em2->start)
3627                 return -1;
3628         else if (em1->start > em2->start)
3629                 return 1;
3630         return 0;
3631 }
3632
3633 static int wait_ordered_extents(struct btrfs_trans_handle *trans,
3634                                 struct inode *inode,
3635                                 struct btrfs_root *root,
3636                                 const struct extent_map *em,
3637                                 const struct list_head *logged_list,
3638                                 bool *ordered_io_error)
3639 {
3640         struct btrfs_ordered_extent *ordered;
3641         struct btrfs_root *log = root->log_root;
3642         u64 mod_start = em->mod_start;
3643         u64 mod_len = em->mod_len;
3644         const bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3645         u64 csum_offset;
3646         u64 csum_len;
3647         LIST_HEAD(ordered_sums);
3648         int ret = 0;
3649
3650         *ordered_io_error = false;
3651
3652         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
3653             em->block_start == EXTENT_MAP_HOLE)
3654                 return 0;
3655
3656         /*
3657          * Wait far any ordered extent that covers our extent map. If it
3658          * finishes without an error, first check and see if our csums are on
3659          * our outstanding ordered extents.
3660          */
3661         list_for_each_entry(ordered, logged_list, log_list) {
3662                 struct btrfs_ordered_sum *sum;
3663
3664                 if (!mod_len)
3665                         break;
3666
3667                 if (ordered->file_offset + ordered->len <= mod_start ||
3668                     mod_start + mod_len <= ordered->file_offset)
3669                         continue;
3670
3671                 if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
3672                     !test_bit(BTRFS_ORDERED_IOERR, &ordered->flags) &&
3673                     !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
3674                         const u64 start = ordered->file_offset;
3675                         const u64 end = ordered->file_offset + ordered->len - 1;
3676
3677                         WARN_ON(ordered->inode != inode);
3678                         filemap_fdatawrite_range(inode->i_mapping, start, end);
3679                 }
3680
3681                 wait_event(ordered->wait,
3682                            (test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) ||
3683                             test_bit(BTRFS_ORDERED_IOERR, &ordered->flags)));
3684
3685                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags)) {
3686                         /*
3687                          * Clear the AS_EIO/AS_ENOSPC flags from the inode's
3688                          * i_mapping flags, so that the next fsync won't get
3689                          * an outdated io error too.
3690                          */
3691                         btrfs_inode_check_errors(inode);
3692                         *ordered_io_error = true;
3693                         break;
3694                 }
3695                 /*
3696                  * We are going to copy all the csums on this ordered extent, so
3697                  * go ahead and adjust mod_start and mod_len in case this
3698                  * ordered extent has already been logged.
3699                  */
3700                 if (ordered->file_offset > mod_start) {
3701                         if (ordered->file_offset + ordered->len >=
3702                             mod_start + mod_len)
3703                                 mod_len = ordered->file_offset - mod_start;
3704                         /*
3705                          * If we have this case
3706                          *
3707                          * |--------- logged extent ---------|
3708                          *       |----- ordered extent ----|
3709                          *
3710                          * Just don't mess with mod_start and mod_len, we'll
3711                          * just end up logging more csums than we need and it
3712                          * will be ok.
3713                          */
3714                 } else {
3715                         if (ordered->file_offset + ordered->len <
3716                             mod_start + mod_len) {
3717                                 mod_len = (mod_start + mod_len) -
3718                                         (ordered->file_offset + ordered->len);
3719                                 mod_start = ordered->file_offset +
3720                                         ordered->len;
3721                         } else {
3722                                 mod_len = 0;
3723                         }
3724                 }
3725
3726                 if (skip_csum)
3727                         continue;
3728
3729                 /*
3730                  * To keep us from looping for the above case of an ordered
3731                  * extent that falls inside of the logged extent.
3732                  */
3733                 if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM,
3734                                      &ordered->flags))
3735                         continue;
3736
3737                 if (ordered->csum_bytes_left) {
3738                         btrfs_start_ordered_extent(inode, ordered, 0);
3739                         wait_event(ordered->wait,
3740                                    ordered->csum_bytes_left == 0);
3741                 }
3742
3743                 list_for_each_entry(sum, &ordered->list, list) {
3744                         ret = btrfs_csum_file_blocks(trans, log, sum);
3745                         if (ret)
3746                                 break;
3747                 }
3748         }
3749
3750         if (*ordered_io_error || !mod_len || ret || skip_csum)
3751                 return ret;
3752
3753         if (em->compress_type) {
3754                 csum_offset = 0;
3755                 csum_len = max(em->block_len, em->orig_block_len);
3756         } else {
3757                 csum_offset = mod_start - em->start;
3758                 csum_len = mod_len;
3759         }
3760
3761         /* block start is already adjusted for the file extent offset. */
3762         ret = btrfs_lookup_csums_range(log->fs_info->csum_root,
3763                                        em->block_start + csum_offset,
3764                                        em->block_start + csum_offset +
3765                                        csum_len - 1, &ordered_sums, 0);
3766         if (ret)
3767                 return ret;
3768
3769         while (!list_empty(&ordered_sums)) {
3770                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3771                                                    struct btrfs_ordered_sum,
3772                                                    list);
3773                 if (!ret)
3774                         ret = btrfs_csum_file_blocks(trans, log, sums);
3775                 list_del(&sums->list);
3776                 kfree(sums);
3777         }
3778
3779         return ret;
3780 }
3781
3782 static int log_one_extent(struct btrfs_trans_handle *trans,
3783                           struct inode *inode, struct btrfs_root *root,
3784                           const struct extent_map *em,
3785                           struct btrfs_path *path,
3786                           const struct list_head *logged_list,
3787                           struct btrfs_log_ctx *ctx)
3788 {
3789         struct btrfs_root *log = root->log_root;
3790         struct btrfs_file_extent_item *fi;
3791         struct extent_buffer *leaf;
3792         struct btrfs_map_token token;
3793         struct btrfs_key key;
3794         u64 extent_offset = em->start - em->orig_start;
3795         u64 block_len;
3796         int ret;
3797         int extent_inserted = 0;
3798         bool ordered_io_err = false;
3799
3800         ret = wait_ordered_extents(trans, inode, root, em, logged_list,
3801                                    &ordered_io_err);
3802         if (ret)
3803                 return ret;
3804
3805         if (ordered_io_err) {
3806                 ctx->io_err = -EIO;
3807                 return 0;
3808         }
3809
3810         btrfs_init_map_token(&token);
3811
3812         ret = __btrfs_drop_extents(trans, log, inode, path, em->start,
3813                                    em->start + em->len, NULL, 0, 1,
3814                                    sizeof(*fi), &extent_inserted);
3815         if (ret)
3816                 return ret;
3817
3818         if (!extent_inserted) {
3819                 key.objectid = btrfs_ino(inode);
3820                 key.type = BTRFS_EXTENT_DATA_KEY;
3821                 key.offset = em->start;
3822
3823                 ret = btrfs_insert_empty_item(trans, log, path, &key,
3824                                               sizeof(*fi));
3825                 if (ret)
3826                         return ret;
3827         }
3828         leaf = path->nodes[0];
3829         fi = btrfs_item_ptr(leaf, path->slots[0],
3830                             struct btrfs_file_extent_item);
3831
3832         btrfs_set_token_file_extent_generation(leaf, fi, trans->transid,
3833                                                &token);
3834         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3835                 btrfs_set_token_file_extent_type(leaf, fi,
3836                                                  BTRFS_FILE_EXTENT_PREALLOC,
3837                                                  &token);
3838         else
3839                 btrfs_set_token_file_extent_type(leaf, fi,
3840                                                  BTRFS_FILE_EXTENT_REG,
3841                                                  &token);
3842
3843         block_len = max(em->block_len, em->orig_block_len);
3844         if (em->compress_type != BTRFS_COMPRESS_NONE) {
3845                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3846                                                         em->block_start,
3847                                                         &token);
3848                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3849                                                            &token);
3850         } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
3851                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3852                                                         em->block_start -
3853                                                         extent_offset, &token);
3854                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3855                                                            &token);
3856         } else {
3857                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi, 0, &token);
3858                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, 0,
3859                                                            &token);
3860         }
3861
3862         btrfs_set_token_file_extent_offset(leaf, fi, extent_offset, &token);
3863         btrfs_set_token_file_extent_num_bytes(leaf, fi, em->len, &token);
3864         btrfs_set_token_file_extent_ram_bytes(leaf, fi, em->ram_bytes, &token);
3865         btrfs_set_token_file_extent_compression(leaf, fi, em->compress_type,
3866                                                 &token);
3867         btrfs_set_token_file_extent_encryption(leaf, fi, 0, &token);
3868         btrfs_set_token_file_extent_other_encoding(leaf, fi, 0, &token);
3869         btrfs_mark_buffer_dirty(leaf);
3870
3871         btrfs_release_path(path);
3872
3873         return ret;
3874 }
3875
3876 static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
3877                                      struct btrfs_root *root,
3878                                      struct inode *inode,
3879                                      struct btrfs_path *path,
3880                                      struct list_head *logged_list,
3881                                      struct btrfs_log_ctx *ctx)
3882 {
3883         struct extent_map *em, *n;
3884         struct list_head extents;
3885         struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3886         u64 test_gen;
3887         int ret = 0;
3888         int num = 0;
3889
3890         INIT_LIST_HEAD(&extents);
3891
3892         write_lock(&tree->lock);
3893         test_gen = root->fs_info->last_trans_committed;
3894
3895         list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
3896                 list_del_init(&em->list);
3897
3898                 /*
3899                  * Just an arbitrary number, this can be really CPU intensive
3900                  * once we start getting a lot of extents, and really once we
3901                  * have a bunch of extents we just want to commit since it will
3902                  * be faster.
3903                  */
3904                 if (++num > 32768) {
3905                         list_del_init(&tree->modified_extents);
3906                         ret = -EFBIG;
3907                         goto process;
3908                 }
3909
3910                 if (em->generation <= test_gen)
3911                         continue;
3912                 /* Need a ref to keep it from getting evicted from cache */
3913                 atomic_inc(&em->refs);
3914                 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
3915                 list_add_tail(&em->list, &extents);
3916                 num++;
3917         }
3918
3919         list_sort(NULL, &extents, extent_cmp);
3920
3921 process:
3922         while (!list_empty(&extents)) {
3923                 em = list_entry(extents.next, struct extent_map, list);
3924
3925                 list_del_init(&em->list);
3926
3927                 /*
3928                  * If we had an error we just need to delete everybody from our
3929                  * private list.
3930                  */
3931                 if (ret) {
3932                         clear_em_logging(tree, em);
3933                         free_extent_map(em);
3934                         continue;
3935                 }
3936
3937                 write_unlock(&tree->lock);
3938
3939                 ret = log_one_extent(trans, inode, root, em, path, logged_list,
3940                                      ctx);
3941                 write_lock(&tree->lock);
3942                 clear_em_logging(tree, em);
3943                 free_extent_map(em);
3944         }
3945         WARN_ON(!list_empty(&extents));
3946         write_unlock(&tree->lock);
3947
3948         btrfs_release_path(path);
3949         return ret;
3950 }
3951
3952 static int logged_inode_size(struct btrfs_root *log, struct inode *inode,
3953                              struct btrfs_path *path, u64 *size_ret)
3954 {
3955         struct btrfs_key key;
3956         int ret;
3957
3958         key.objectid = btrfs_ino(inode);
3959         key.type = BTRFS_INODE_ITEM_KEY;
3960         key.offset = 0;
3961
3962         ret = btrfs_search_slot(NULL, log, &key, path, 0, 0);
3963         if (ret < 0) {
3964                 return ret;
3965         } else if (ret > 0) {
3966                 *size_ret = i_size_read(inode);
3967         } else {
3968                 struct btrfs_inode_item *item;
3969
3970                 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3971                                       struct btrfs_inode_item);
3972                 *size_ret = btrfs_inode_size(path->nodes[0], item);
3973         }
3974
3975         btrfs_release_path(path);
3976         return 0;
3977 }
3978
3979 /* log a single inode in the tree log.
3980  * At least one parent directory for this inode must exist in the tree
3981  * or be logged already.
3982  *
3983  * Any items from this inode changed by the current transaction are copied
3984  * to the log tree.  An extra reference is taken on any extents in this
3985  * file, allowing us to avoid a whole pile of corner cases around logging
3986  * blocks that have been removed from the tree.
3987  *
3988  * See LOG_INODE_ALL and related defines for a description of what inode_only
3989  * does.
3990  *
3991  * This handles both files and directories.
3992  */
3993 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
3994                            struct btrfs_root *root, struct inode *inode,
3995                            int inode_only,
3996                            const loff_t start,
3997                            const loff_t end,
3998                            struct btrfs_log_ctx *ctx)
3999 {
4000         struct btrfs_path *path;
4001         struct btrfs_path *dst_path;
4002         struct btrfs_key min_key;
4003         struct btrfs_key max_key;
4004         struct btrfs_root *log = root->log_root;
4005         struct extent_buffer *src = NULL;
4006         LIST_HEAD(logged_list);
4007         u64 last_extent = 0;
4008         int err = 0;
4009         int ret;
4010         int nritems;
4011         int ins_start_slot = 0;
4012         int ins_nr;
4013         bool fast_search = false;
4014         u64 ino = btrfs_ino(inode);
4015         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4016         u64 logged_isize = 0;
4017
4018         path = btrfs_alloc_path();
4019         if (!path)
4020                 return -ENOMEM;
4021         dst_path = btrfs_alloc_path();
4022         if (!dst_path) {
4023                 btrfs_free_path(path);
4024                 return -ENOMEM;
4025         }
4026
4027         min_key.objectid = ino;
4028         min_key.type = BTRFS_INODE_ITEM_KEY;
4029         min_key.offset = 0;
4030
4031         max_key.objectid = ino;
4032
4033
4034         /* today the code can only do partial logging of directories */
4035         if (S_ISDIR(inode->i_mode) ||
4036             (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4037                        &BTRFS_I(inode)->runtime_flags) &&
4038              inode_only == LOG_INODE_EXISTS))
4039                 max_key.type = BTRFS_XATTR_ITEM_KEY;
4040         else
4041                 max_key.type = (u8)-1;
4042         max_key.offset = (u64)-1;
4043
4044         /*
4045          * Only run delayed items if we are a dir or a new file.
4046          * Otherwise commit the delayed inode only, which is needed in
4047          * order for the log replay code to mark inodes for link count
4048          * fixup (create temporary BTRFS_TREE_LOG_FIXUP_OBJECTID items).
4049          */
4050         if (S_ISDIR(inode->i_mode) ||
4051             BTRFS_I(inode)->generation > root->fs_info->last_trans_committed)
4052                 ret = btrfs_commit_inode_delayed_items(trans, inode);
4053         else
4054                 ret = btrfs_commit_inode_delayed_inode(inode);
4055
4056         if (ret) {
4057                 btrfs_free_path(path);
4058                 btrfs_free_path(dst_path);
4059                 return ret;
4060         }
4061
4062         mutex_lock(&BTRFS_I(inode)->log_mutex);
4063
4064         btrfs_get_logged_extents(inode, &logged_list, start, end);
4065
4066         /*
4067          * a brute force approach to making sure we get the most uptodate
4068          * copies of everything.
4069          */
4070         if (S_ISDIR(inode->i_mode)) {
4071                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
4072
4073                 if (inode_only == LOG_INODE_EXISTS) {
4074                         max_key_type = BTRFS_INODE_EXTREF_KEY;
4075                         max_key.type = max_key_type;
4076                 }
4077                 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
4078         } else {
4079                 if (inode_only == LOG_INODE_EXISTS) {
4080                         /*
4081                          * Make sure the new inode item we write to the log has
4082                          * the same isize as the current one (if it exists).
4083                          * This is necessary to prevent data loss after log
4084                          * replay, and also to prevent doing a wrong expanding
4085                          * truncate - for e.g. create file, write 4K into offset
4086                          * 0, fsync, write 4K into offset 4096, add hard link,
4087                          * fsync some other file (to sync log), power fail - if
4088                          * we use the inode's current i_size, after log replay
4089                          * we get a 8Kb file, with the last 4Kb extent as a hole
4090                          * (zeroes), as if an expanding truncate happened,
4091                          * instead of getting a file of 4Kb only.
4092                          */
4093                         err = logged_inode_size(log, inode, path,
4094                                                 &logged_isize);
4095                         if (err)
4096                                 goto out_unlock;
4097                 }
4098                 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4099                              &BTRFS_I(inode)->runtime_flags)) {
4100                         if (inode_only == LOG_INODE_EXISTS) {
4101                                 max_key.type = BTRFS_INODE_EXTREF_KEY;
4102                                 ret = drop_objectid_items(trans, log, path, ino,
4103                                                           max_key.type);
4104                         } else {
4105                                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4106                                           &BTRFS_I(inode)->runtime_flags);
4107                                 clear_bit(BTRFS_INODE_COPY_EVERYTHING,
4108                                           &BTRFS_I(inode)->runtime_flags);
4109                                 ret = btrfs_truncate_inode_items(trans, log,
4110                                                                  inode, 0, 0);
4111                         }
4112                 } else if (test_bit(BTRFS_INODE_COPY_EVERYTHING,
4113                                     &BTRFS_I(inode)->runtime_flags) ||
4114                            inode_only == LOG_INODE_EXISTS) {
4115                         if (inode_only == LOG_INODE_ALL) {
4116                                 clear_bit(BTRFS_INODE_COPY_EVERYTHING,
4117                                           &BTRFS_I(inode)->runtime_flags);
4118                                 fast_search = true;
4119                                 max_key.type = BTRFS_XATTR_ITEM_KEY;
4120                         } else {
4121                                 max_key.type = BTRFS_INODE_EXTREF_KEY;
4122                         }
4123                         ret = drop_objectid_items(trans, log, path, ino,
4124                                                   max_key.type);
4125                 } else {
4126                         if (inode_only == LOG_INODE_ALL)
4127                                 fast_search = true;
4128                         ret = log_inode_item(trans, log, dst_path, inode);
4129                         if (ret) {
4130                                 err = ret;
4131                                 goto out_unlock;
4132                         }
4133                         goto log_extents;
4134                 }
4135
4136         }
4137         if (ret) {
4138                 err = ret;
4139                 goto out_unlock;
4140         }
4141
4142         while (1) {
4143                 ins_nr = 0;
4144                 ret = btrfs_search_forward(root, &min_key,
4145                                            path, trans->transid);
4146                 if (ret != 0)
4147                         break;
4148 again:
4149                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
4150                 if (min_key.objectid != ino)
4151                         break;
4152                 if (min_key.type > max_key.type)
4153                         break;
4154
4155                 src = path->nodes[0];
4156                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
4157                         ins_nr++;
4158                         goto next_slot;
4159                 } else if (!ins_nr) {
4160                         ins_start_slot = path->slots[0];
4161                         ins_nr = 1;
4162                         goto next_slot;
4163                 }
4164
4165                 ret = copy_items(trans, inode, dst_path, path, &last_extent,
4166                                  ins_start_slot, ins_nr, inode_only,
4167                                  logged_isize);
4168                 if (ret < 0) {
4169                         err = ret;
4170                         goto out_unlock;
4171                 }
4172                 if (ret) {
4173                         ins_nr = 0;
4174                         btrfs_release_path(path);
4175                         continue;
4176                 }
4177                 ins_nr = 1;
4178                 ins_start_slot = path->slots[0];
4179 next_slot:
4180
4181                 nritems = btrfs_header_nritems(path->nodes[0]);
4182                 path->slots[0]++;
4183                 if (path->slots[0] < nritems) {
4184                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
4185                                               path->slots[0]);
4186                         goto again;
4187                 }
4188                 if (ins_nr) {
4189                         ret = copy_items(trans, inode, dst_path, path,
4190                                          &last_extent, ins_start_slot,
4191                                          ins_nr, inode_only, logged_isize);
4192                         if (ret < 0) {
4193                                 err = ret;
4194                                 goto out_unlock;
4195                         }
4196                         ret = 0;
4197                         ins_nr = 0;
4198                 }
4199                 btrfs_release_path(path);
4200
4201                 if (min_key.offset < (u64)-1) {
4202                         min_key.offset++;
4203                 } else if (min_key.type < max_key.type) {
4204                         min_key.type++;
4205                         min_key.offset = 0;
4206                 } else {
4207                         break;
4208                 }
4209         }
4210         if (ins_nr) {
4211                 ret = copy_items(trans, inode, dst_path, path, &last_extent,
4212                                  ins_start_slot, ins_nr, inode_only,
4213                                  logged_isize);
4214                 if (ret < 0) {
4215                         err = ret;
4216                         goto out_unlock;
4217                 }
4218                 ret = 0;
4219                 ins_nr = 0;
4220         }
4221
4222 log_extents:
4223         btrfs_release_path(path);
4224         btrfs_release_path(dst_path);
4225         if (fast_search) {
4226                 /*
4227                  * Some ordered extents started by fsync might have completed
4228                  * before we collected the ordered extents in logged_list, which
4229                  * means they're gone, not in our logged_list nor in the inode's
4230                  * ordered tree. We want the application/user space to know an
4231                  * error happened while attempting to persist file data so that
4232                  * it can take proper action. If such error happened, we leave
4233                  * without writing to the log tree and the fsync must report the
4234                  * file data write error and not commit the current transaction.
4235                  */
4236                 err = btrfs_inode_check_errors(inode);
4237                 if (err) {
4238                         ctx->io_err = err;
4239                         goto out_unlock;
4240                 }
4241                 ret = btrfs_log_changed_extents(trans, root, inode, dst_path,
4242                                                 &logged_list, ctx);
4243                 if (ret) {
4244                         err = ret;
4245                         goto out_unlock;
4246                 }
4247         } else if (inode_only == LOG_INODE_ALL) {
4248                 struct extent_map *em, *n;
4249
4250                 write_lock(&em_tree->lock);
4251                 /*
4252                  * We can't just remove every em if we're called for a ranged
4253                  * fsync - that is, one that doesn't cover the whole possible
4254                  * file range (0 to LLONG_MAX). This is because we can have
4255                  * em's that fall outside the range we're logging and therefore
4256                  * their ordered operations haven't completed yet
4257                  * (btrfs_finish_ordered_io() not invoked yet). This means we
4258                  * didn't get their respective file extent item in the fs/subvol
4259                  * tree yet, and need to let the next fast fsync (one which
4260                  * consults the list of modified extent maps) find the em so
4261                  * that it logs a matching file extent item and waits for the
4262                  * respective ordered operation to complete (if it's still
4263                  * running).
4264                  *
4265                  * Removing every em outside the range we're logging would make
4266                  * the next fast fsync not log their matching file extent items,
4267                  * therefore making us lose data after a log replay.
4268                  */
4269                 list_for_each_entry_safe(em, n, &em_tree->modified_extents,
4270                                          list) {
4271                         const u64 mod_end = em->mod_start + em->mod_len - 1;
4272
4273                         if (em->mod_start >= start && mod_end <= end)
4274                                 list_del_init(&em->list);
4275                 }
4276                 write_unlock(&em_tree->lock);
4277         }
4278
4279         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
4280                 ret = log_directory_changes(trans, root, inode, path, dst_path);
4281                 if (ret) {
4282                         err = ret;
4283                         goto out_unlock;
4284                 }
4285         }
4286
4287         BTRFS_I(inode)->logged_trans = trans->transid;
4288         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->last_sub_trans;
4289 out_unlock:
4290         if (unlikely(err))
4291                 btrfs_put_logged_extents(&logged_list);
4292         else
4293                 btrfs_submit_logged_extents(&logged_list, log);
4294         mutex_unlock(&BTRFS_I(inode)->log_mutex);
4295
4296         btrfs_free_path(path);
4297         btrfs_free_path(dst_path);
4298         return err;
4299 }
4300
4301 /*
4302  * follow the dentry parent pointers up the chain and see if any
4303  * of the directories in it require a full commit before they can
4304  * be logged.  Returns zero if nothing special needs to be done or 1 if
4305  * a full commit is required.
4306  */
4307 static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
4308                                                struct inode *inode,
4309                                                struct dentry *parent,
4310                                                struct super_block *sb,
4311                                                u64 last_committed)
4312 {
4313         int ret = 0;
4314         struct btrfs_root *root;
4315         struct dentry *old_parent = NULL;
4316         struct inode *orig_inode = inode;
4317
4318         /*
4319          * for regular files, if its inode is already on disk, we don't
4320          * have to worry about the parents at all.  This is because
4321          * we can use the last_unlink_trans field to record renames
4322          * and other fun in this file.
4323          */
4324         if (S_ISREG(inode->i_mode) &&
4325             BTRFS_I(inode)->generation <= last_committed &&
4326             BTRFS_I(inode)->last_unlink_trans <= last_committed)
4327                         goto out;
4328
4329         if (!S_ISDIR(inode->i_mode)) {
4330                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
4331                         goto out;
4332                 inode = parent->d_inode;
4333         }
4334
4335         while (1) {
4336                 /*
4337                  * If we are logging a directory then we start with our inode,
4338                  * not our parents inode, so we need to skipp setting the
4339                  * logged_trans so that further down in the log code we don't
4340                  * think this inode has already been logged.
4341                  */
4342                 if (inode != orig_inode)
4343                         BTRFS_I(inode)->logged_trans = trans->transid;
4344                 smp_mb();
4345
4346                 if (BTRFS_I(inode)->last_unlink_trans > last_committed) {
4347                         root = BTRFS_I(inode)->root;
4348
4349                         /*
4350                          * make sure any commits to the log are forced
4351                          * to be full commits
4352                          */
4353                         btrfs_set_log_full_commit(root->fs_info, trans);
4354                         ret = 1;
4355                         break;
4356                 }
4357
4358                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
4359                         break;
4360
4361                 if (IS_ROOT(parent))
4362                         break;
4363
4364                 parent = dget_parent(parent);
4365                 dput(old_parent);
4366                 old_parent = parent;
4367                 inode = parent->d_inode;
4368
4369         }
4370         dput(old_parent);
4371 out:
4372         return ret;
4373 }
4374
4375 /*
4376  * helper function around btrfs_log_inode to make sure newly created
4377  * parent directories also end up in the log.  A minimal inode and backref
4378  * only logging is done of any parent directories that are older than
4379  * the last committed transaction
4380  */
4381 static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
4382                                   struct btrfs_root *root, struct inode *inode,
4383                                   struct dentry *parent,
4384                                   const loff_t start,
4385                                   const loff_t end,
4386                                   int exists_only,
4387                                   struct btrfs_log_ctx *ctx)
4388 {
4389         int inode_only = exists_only ? LOG_INODE_EXISTS : LOG_INODE_ALL;
4390         struct super_block *sb;
4391         struct dentry *old_parent = NULL;
4392         int ret = 0;
4393         u64 last_committed = root->fs_info->last_trans_committed;
4394         const struct dentry * const first_parent = parent;
4395         const bool did_unlink = (BTRFS_I(inode)->last_unlink_trans >
4396                                  last_committed);
4397
4398         sb = inode->i_sb;
4399
4400         if (btrfs_test_opt(root, NOTREELOG)) {
4401                 ret = 1;
4402                 goto end_no_trans;
4403         }
4404
4405         /*
4406          * The prev transaction commit doesn't complete, we need do
4407          * full commit by ourselves.
4408          */
4409         if (root->fs_info->last_trans_log_full_commit >
4410             root->fs_info->last_trans_committed) {
4411                 ret = 1;
4412                 goto end_no_trans;
4413         }
4414
4415         if (root != BTRFS_I(inode)->root ||
4416             btrfs_root_refs(&root->root_item) == 0) {
4417                 ret = 1;
4418                 goto end_no_trans;
4419         }
4420
4421         ret = check_parent_dirs_for_sync(trans, inode, parent,
4422                                          sb, last_committed);
4423         if (ret)
4424                 goto end_no_trans;
4425
4426         if (btrfs_inode_in_log(inode, trans->transid)) {
4427                 ret = BTRFS_NO_LOG_SYNC;
4428                 goto end_no_trans;
4429         }
4430
4431         ret = start_log_trans(trans, root, ctx);
4432         if (ret)
4433                 goto end_no_trans;
4434
4435         ret = btrfs_log_inode(trans, root, inode, inode_only, start, end, ctx);
4436         if (ret)
4437                 goto end_trans;
4438
4439         /*
4440          * for regular files, if its inode is already on disk, we don't
4441          * have to worry about the parents at all.  This is because
4442          * we can use the last_unlink_trans field to record renames
4443          * and other fun in this file.
4444          */
4445         if (S_ISREG(inode->i_mode) &&
4446             BTRFS_I(inode)->generation <= last_committed &&
4447             BTRFS_I(inode)->last_unlink_trans <= last_committed) {
4448                 ret = 0;
4449                 goto end_trans;
4450         }
4451
4452         while (1) {
4453                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
4454                         break;
4455
4456                 inode = parent->d_inode;
4457                 if (root != BTRFS_I(inode)->root)
4458                         break;
4459
4460                 /*
4461                  * On unlink we must make sure our immediate parent directory
4462                  * inode is fully logged. This is to prevent leaving dangling
4463                  * directory index entries and a wrong directory inode's i_size.
4464                  * Not doing so can result in a directory being impossible to
4465                  * delete after log replay (rmdir will always fail with error
4466                  * -ENOTEMPTY).
4467                  */
4468                 if (did_unlink && parent == first_parent)
4469                         inode_only = LOG_INODE_ALL;
4470                 else
4471                         inode_only = LOG_INODE_EXISTS;
4472
4473                 if (BTRFS_I(inode)->generation >
4474                     root->fs_info->last_trans_committed ||
4475                     inode_only == LOG_INODE_ALL) {
4476                         ret = btrfs_log_inode(trans, root, inode, inode_only,
4477                                               0, LLONG_MAX, ctx);
4478                         if (ret)
4479                                 goto end_trans;
4480                 }
4481                 if (IS_ROOT(parent))
4482                         break;
4483
4484                 parent = dget_parent(parent);
4485                 dput(old_parent);
4486                 old_parent = parent;
4487         }
4488         ret = 0;
4489 end_trans:
4490         dput(old_parent);
4491         if (ret < 0) {
4492                 btrfs_set_log_full_commit(root->fs_info, trans);
4493                 ret = 1;
4494         }
4495
4496         if (ret)
4497                 btrfs_remove_log_ctx(root, ctx);
4498         btrfs_end_log_trans(root);
4499 end_no_trans:
4500         return ret;
4501 }
4502
4503 /*
4504  * it is not safe to log dentry if the chunk root has added new
4505  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
4506  * If this returns 1, you must commit the transaction to safely get your
4507  * data on disk.
4508  */
4509 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
4510                           struct btrfs_root *root, struct dentry *dentry,
4511                           const loff_t start,
4512                           const loff_t end,
4513                           struct btrfs_log_ctx *ctx)
4514 {
4515         struct dentry *parent = dget_parent(dentry);
4516         int ret;
4517
4518         ret = btrfs_log_inode_parent(trans, root, dentry->d_inode, parent,
4519                                      start, end, 0, ctx);
4520         dput(parent);
4521
4522         return ret;
4523 }
4524
4525 /*
4526  * should be called during mount to recover any replay any log trees
4527  * from the FS
4528  */
4529 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
4530 {
4531         int ret;
4532         struct btrfs_path *path;
4533         struct btrfs_trans_handle *trans;
4534         struct btrfs_key key;
4535         struct btrfs_key found_key;
4536         struct btrfs_key tmp_key;
4537         struct btrfs_root *log;
4538         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
4539         struct walk_control wc = {
4540                 .process_func = process_one_buffer,
4541                 .stage = 0,
4542         };
4543
4544         path = btrfs_alloc_path();
4545         if (!path)
4546                 return -ENOMEM;
4547
4548         fs_info->log_root_recovering = 1;
4549
4550         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4551         if (IS_ERR(trans)) {
4552                 ret = PTR_ERR(trans);
4553                 goto error;
4554         }
4555
4556         wc.trans = trans;
4557         wc.pin = 1;
4558
4559         ret = walk_log_tree(trans, log_root_tree, &wc);
4560         if (ret) {
4561                 btrfs_error(fs_info, ret, "Failed to pin buffers while "
4562                             "recovering log root tree.");
4563                 goto error;
4564         }
4565
4566 again:
4567         key.objectid = BTRFS_TREE_LOG_OBJECTID;
4568         key.offset = (u64)-1;
4569         key.type = BTRFS_ROOT_ITEM_KEY;
4570
4571         while (1) {
4572                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
4573
4574                 if (ret < 0) {
4575                         btrfs_error(fs_info, ret,
4576                                     "Couldn't find tree log root.");
4577                         goto error;
4578                 }
4579                 if (ret > 0) {
4580                         if (path->slots[0] == 0)
4581                                 break;
4582                         path->slots[0]--;
4583                 }
4584                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
4585                                       path->slots[0]);
4586                 btrfs_release_path(path);
4587                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4588                         break;
4589
4590                 log = btrfs_read_fs_root(log_root_tree, &found_key);
4591                 if (IS_ERR(log)) {
4592                         ret = PTR_ERR(log);
4593                         btrfs_error(fs_info, ret,
4594                                     "Couldn't read tree log root.");
4595                         goto error;
4596                 }
4597
4598                 tmp_key.objectid = found_key.offset;
4599                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
4600                 tmp_key.offset = (u64)-1;
4601
4602                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
4603                 if (IS_ERR(wc.replay_dest)) {
4604                         ret = PTR_ERR(wc.replay_dest);
4605                         free_extent_buffer(log->node);
4606                         free_extent_buffer(log->commit_root);
4607                         kfree(log);
4608                         btrfs_error(fs_info, ret, "Couldn't read target root "
4609                                     "for tree log recovery.");
4610                         goto error;
4611                 }
4612
4613                 wc.replay_dest->log_root = log;
4614                 btrfs_record_root_in_trans(trans, wc.replay_dest);
4615                 ret = walk_log_tree(trans, log, &wc);
4616
4617                 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
4618                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
4619                                                       path);
4620                 }
4621
4622                 key.offset = found_key.offset - 1;
4623                 wc.replay_dest->log_root = NULL;
4624                 free_extent_buffer(log->node);
4625                 free_extent_buffer(log->commit_root);
4626                 kfree(log);
4627
4628                 if (ret)
4629                         goto error;
4630
4631                 if (found_key.offset == 0)
4632                         break;
4633         }
4634         btrfs_release_path(path);
4635
4636         /* step one is to pin it all, step two is to replay just inodes */
4637         if (wc.pin) {
4638                 wc.pin = 0;
4639                 wc.process_func = replay_one_buffer;
4640                 wc.stage = LOG_WALK_REPLAY_INODES;
4641                 goto again;
4642         }
4643         /* step three is to replay everything */
4644         if (wc.stage < LOG_WALK_REPLAY_ALL) {
4645                 wc.stage++;
4646                 goto again;
4647         }
4648
4649         btrfs_free_path(path);
4650
4651         /* step 4: commit the transaction, which also unpins the blocks */
4652         ret = btrfs_commit_transaction(trans, fs_info->tree_root);
4653         if (ret)
4654                 return ret;
4655
4656         free_extent_buffer(log_root_tree->node);
4657         log_root_tree->log_root = NULL;
4658         fs_info->log_root_recovering = 0;
4659         kfree(log_root_tree);
4660
4661         return 0;
4662 error:
4663         if (wc.trans)
4664                 btrfs_end_transaction(wc.trans, fs_info->tree_root);
4665         btrfs_free_path(path);
4666         return ret;
4667 }
4668
4669 /*
4670  * there are some corner cases where we want to force a full
4671  * commit instead of allowing a directory to be logged.
4672  *
4673  * They revolve around files there were unlinked from the directory, and
4674  * this function updates the parent directory so that a full commit is
4675  * properly done if it is fsync'd later after the unlinks are done.
4676  */
4677 void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
4678                              struct inode *dir, struct inode *inode,
4679                              int for_rename)
4680 {
4681         /*
4682          * when we're logging a file, if it hasn't been renamed
4683          * or unlinked, and its inode is fully committed on disk,
4684          * we don't have to worry about walking up the directory chain
4685          * to log its parents.
4686          *
4687          * So, we use the last_unlink_trans field to put this transid
4688          * into the file.  When the file is logged we check it and
4689          * don't log the parents if the file is fully on disk.
4690          */
4691         if (S_ISREG(inode->i_mode))
4692                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4693
4694         /*
4695          * if this directory was already logged any new
4696          * names for this file/dir will get recorded
4697          */
4698         smp_mb();
4699         if (BTRFS_I(dir)->logged_trans == trans->transid)
4700                 return;
4701
4702         /*
4703          * if the inode we're about to unlink was logged,
4704          * the log will be properly updated for any new names
4705          */
4706         if (BTRFS_I(inode)->logged_trans == trans->transid)
4707                 return;
4708
4709         /*
4710          * when renaming files across directories, if the directory
4711          * there we're unlinking from gets fsync'd later on, there's
4712          * no way to find the destination directory later and fsync it
4713          * properly.  So, we have to be conservative and force commits
4714          * so the new name gets discovered.
4715          */
4716         if (for_rename)
4717                 goto record;
4718
4719         /* we can safely do the unlink without any special recording */
4720         return;
4721
4722 record:
4723         BTRFS_I(dir)->last_unlink_trans = trans->transid;
4724 }
4725
4726 /*
4727  * Call this after adding a new name for a file and it will properly
4728  * update the log to reflect the new name.
4729  *
4730  * It will return zero if all goes well, and it will return 1 if a
4731  * full transaction commit is required.
4732  */
4733 int btrfs_log_new_name(struct btrfs_trans_handle *trans,
4734                         struct inode *inode, struct inode *old_dir,
4735                         struct dentry *parent)
4736 {
4737         struct btrfs_root * root = BTRFS_I(inode)->root;
4738
4739         /*
4740          * this will force the logging code to walk the dentry chain
4741          * up for the file
4742          */
4743         if (S_ISREG(inode->i_mode))
4744                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4745
4746         /*
4747          * if this inode hasn't been logged and directory we're renaming it
4748          * from hasn't been logged, we don't need to log it
4749          */
4750         if (BTRFS_I(inode)->logged_trans <=
4751             root->fs_info->last_trans_committed &&
4752             (!old_dir || BTRFS_I(old_dir)->logged_trans <=
4753                     root->fs_info->last_trans_committed))
4754                 return 0;
4755
4756         return btrfs_log_inode_parent(trans, root, inode, parent, 0,
4757                                       LLONG_MAX, 1, NULL);
4758 }
4759