c766649ad4538cf613f44d106c98041774da18f9
[linux-3.10.git] / 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 "ctree.h"
21 #include "transaction.h"
22 #include "disk-io.h"
23 #include "locking.h"
24 #include "print-tree.h"
25 #include "compat.h"
26 #include "tree-log.h"
27
28 /* magic values for the inode_only field in btrfs_log_inode:
29  *
30  * LOG_INODE_ALL means to log everything
31  * LOG_INODE_EXISTS means to log just enough to recreate the inode
32  * during log replay
33  */
34 #define LOG_INODE_ALL 0
35 #define LOG_INODE_EXISTS 1
36
37 /*
38  * stages for the tree walking.  The first
39  * stage (0) is to only pin down the blocks we find
40  * the second stage (1) is to make sure that all the inodes
41  * we find in the log are created in the subvolume.
42  *
43  * The last stage is to deal with directories and links and extents
44  * and all the other fun semantics
45  */
46 #define LOG_WALK_PIN_ONLY 0
47 #define LOG_WALK_REPLAY_INODES 1
48 #define LOG_WALK_REPLAY_ALL 2
49
50 static int __btrfs_log_inode(struct btrfs_trans_handle *trans,
51                              struct btrfs_root *root, struct inode *inode,
52                              int inode_only);
53
54 /*
55  * tree logging is a special write ahead log used to make sure that
56  * fsyncs and O_SYNCs can happen without doing full tree commits.
57  *
58  * Full tree commits are expensive because they require commonly
59  * modified blocks to be recowed, creating many dirty pages in the
60  * extent tree an 4x-6x higher write load than ext3.
61  *
62  * Instead of doing a tree commit on every fsync, we use the
63  * key ranges and transaction ids to find items for a given file or directory
64  * that have changed in this transaction.  Those items are copied into
65  * a special tree (one per subvolume root), that tree is written to disk
66  * and then the fsync is considered complete.
67  *
68  * After a crash, items are copied out of the log-tree back into the
69  * subvolume tree.  Any file data extents found are recorded in the extent
70  * allocation tree, and the log-tree freed.
71  *
72  * The log tree is read three times, once to pin down all the extents it is
73  * using in ram and once, once to create all the inodes logged in the tree
74  * and once to do all the other items.
75  */
76
77 /*
78  * btrfs_add_log_tree adds a new per-subvolume log tree into the
79  * tree of log tree roots.  This must be called with a tree log transaction
80  * running (see start_log_trans).
81  */
82 static int btrfs_add_log_tree(struct btrfs_trans_handle *trans,
83                       struct btrfs_root *root)
84 {
85         struct btrfs_key key;
86         struct btrfs_root_item root_item;
87         struct btrfs_inode_item *inode_item;
88         struct extent_buffer *leaf;
89         struct btrfs_root *new_root = root;
90         int ret;
91         u64 objectid = root->root_key.objectid;
92
93         leaf = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
94                                       BTRFS_TREE_LOG_OBJECTID,
95                                       trans->transid, 0, 0, 0);
96         if (IS_ERR(leaf)) {
97                 ret = PTR_ERR(leaf);
98                 return ret;
99         }
100
101         btrfs_set_header_nritems(leaf, 0);
102         btrfs_set_header_level(leaf, 0);
103         btrfs_set_header_bytenr(leaf, leaf->start);
104         btrfs_set_header_generation(leaf, trans->transid);
105         btrfs_set_header_owner(leaf, BTRFS_TREE_LOG_OBJECTID);
106
107         write_extent_buffer(leaf, root->fs_info->fsid,
108                             (unsigned long)btrfs_header_fsid(leaf),
109                             BTRFS_FSID_SIZE);
110         btrfs_mark_buffer_dirty(leaf);
111
112         inode_item = &root_item.inode;
113         memset(inode_item, 0, sizeof(*inode_item));
114         inode_item->generation = cpu_to_le64(1);
115         inode_item->size = cpu_to_le64(3);
116         inode_item->nlink = cpu_to_le32(1);
117         inode_item->nbytes = cpu_to_le64(root->leafsize);
118         inode_item->mode = cpu_to_le32(S_IFDIR | 0755);
119
120         btrfs_set_root_bytenr(&root_item, leaf->start);
121         btrfs_set_root_generation(&root_item, trans->transid);
122         btrfs_set_root_level(&root_item, 0);
123         btrfs_set_root_refs(&root_item, 0);
124         btrfs_set_root_used(&root_item, 0);
125
126         memset(&root_item.drop_progress, 0, sizeof(root_item.drop_progress));
127         root_item.drop_level = 0;
128
129         btrfs_tree_unlock(leaf);
130         free_extent_buffer(leaf);
131         leaf = NULL;
132
133         btrfs_set_root_dirid(&root_item, 0);
134
135         key.objectid = BTRFS_TREE_LOG_OBJECTID;
136         key.offset = objectid;
137         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
138         ret = btrfs_insert_root(trans, root->fs_info->log_root_tree, &key,
139                                 &root_item);
140         if (ret)
141                 goto fail;
142
143         new_root = btrfs_read_fs_root_no_radix(root->fs_info->log_root_tree,
144                                                &key);
145         BUG_ON(!new_root);
146
147         WARN_ON(root->log_root);
148         root->log_root = new_root;
149
150         /*
151          * log trees do not get reference counted because they go away
152          * before a real commit is actually done.  They do store pointers
153          * to file data extents, and those reference counts still get
154          * updated (along with back refs to the log tree).
155          */
156         new_root->ref_cows = 0;
157         new_root->last_trans = trans->transid;
158 fail:
159         return ret;
160 }
161
162 /*
163  * start a sub transaction and setup the log tree
164  * this increments the log tree writer count to make the people
165  * syncing the tree wait for us to finish
166  */
167 static int start_log_trans(struct btrfs_trans_handle *trans,
168                            struct btrfs_root *root)
169 {
170         int ret;
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                 BUG_ON(ret);
175         }
176         if (!root->log_root) {
177                 ret = btrfs_add_log_tree(trans, root);
178                 BUG_ON(ret);
179         }
180         atomic_inc(&root->fs_info->tree_log_writers);
181         root->fs_info->tree_log_batch++;
182         mutex_unlock(&root->fs_info->tree_log_mutex);
183         return 0;
184 }
185
186 /*
187  * returns 0 if there was a log transaction running and we were able
188  * to join, or returns -ENOENT if there were not transactions
189  * in progress
190  */
191 static int join_running_log_trans(struct btrfs_root *root)
192 {
193         int ret = -ENOENT;
194
195         smp_mb();
196         if (!root->log_root)
197                 return -ENOENT;
198
199         mutex_lock(&root->fs_info->tree_log_mutex);
200         if (root->log_root) {
201                 ret = 0;
202                 atomic_inc(&root->fs_info->tree_log_writers);
203                 root->fs_info->tree_log_batch++;
204         }
205         mutex_unlock(&root->fs_info->tree_log_mutex);
206         return ret;
207 }
208
209 /*
210  * indicate we're done making changes to the log tree
211  * and wake up anyone waiting to do a sync
212  */
213 static int end_log_trans(struct btrfs_root *root)
214 {
215         atomic_dec(&root->fs_info->tree_log_writers);
216         smp_mb();
217         if (waitqueue_active(&root->fs_info->tree_log_wait))
218                 wake_up(&root->fs_info->tree_log_wait);
219         return 0;
220 }
221
222
223 /*
224  * the walk control struct is used to pass state down the chain when
225  * processing the log tree.  The stage field tells us which part
226  * of the log tree processing we are currently doing.  The others
227  * are state fields used for that specific part
228  */
229 struct walk_control {
230         /* should we free the extent on disk when done?  This is used
231          * at transaction commit time while freeing a log tree
232          */
233         int free;
234
235         /* should we write out the extent buffer?  This is used
236          * while flushing the log tree to disk during a sync
237          */
238         int write;
239
240         /* should we wait for the extent buffer io to finish?  Also used
241          * while flushing the log tree to disk for a sync
242          */
243         int wait;
244
245         /* pin only walk, we record which extents on disk belong to the
246          * log trees
247          */
248         int pin;
249
250         /* what stage of the replay code we're currently in */
251         int stage;
252
253         /* the root we are currently replaying */
254         struct btrfs_root *replay_dest;
255
256         /* the trans handle for the current replay */
257         struct btrfs_trans_handle *trans;
258
259         /* the function that gets used to process blocks we find in the
260          * tree.  Note the extent_buffer might not be up to date when it is
261          * passed in, and it must be checked or read if you need the data
262          * inside it
263          */
264         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
265                             struct walk_control *wc, u64 gen);
266 };
267
268 /*
269  * process_func used to pin down extents, write them or wait on them
270  */
271 static int process_one_buffer(struct btrfs_root *log,
272                               struct extent_buffer *eb,
273                               struct walk_control *wc, u64 gen)
274 {
275         if (wc->pin) {
276                 mutex_lock(&log->fs_info->pinned_mutex);
277                 btrfs_update_pinned_extents(log->fs_info->extent_root,
278                                             eb->start, eb->len, 1);
279                 mutex_unlock(&log->fs_info->pinned_mutex);
280         }
281
282         if (btrfs_buffer_uptodate(eb, gen)) {
283                 if (wc->write)
284                         btrfs_write_tree_block(eb);
285                 if (wc->wait)
286                         btrfs_wait_tree_block_writeback(eb);
287         }
288         return 0;
289 }
290
291 /*
292  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
293  * to the src data we are copying out.
294  *
295  * root is the tree we are copying into, and path is a scratch
296  * path for use in this function (it should be released on entry and
297  * will be released on exit).
298  *
299  * If the key is already in the destination tree the existing item is
300  * overwritten.  If the existing item isn't big enough, it is extended.
301  * If it is too large, it is truncated.
302  *
303  * If the key isn't in the destination yet, a new item is inserted.
304  */
305 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
306                                    struct btrfs_root *root,
307                                    struct btrfs_path *path,
308                                    struct extent_buffer *eb, int slot,
309                                    struct btrfs_key *key)
310 {
311         int ret;
312         u32 item_size;
313         u64 saved_i_size = 0;
314         int save_old_i_size = 0;
315         unsigned long src_ptr;
316         unsigned long dst_ptr;
317         int overwrite_root = 0;
318
319         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
320                 overwrite_root = 1;
321
322         item_size = btrfs_item_size_nr(eb, slot);
323         src_ptr = btrfs_item_ptr_offset(eb, slot);
324
325         /* look for the key in the destination tree */
326         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
327         if (ret == 0) {
328                 char *src_copy;
329                 char *dst_copy;
330                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
331                                                   path->slots[0]);
332                 if (dst_size != item_size)
333                         goto insert;
334
335                 if (item_size == 0) {
336                         btrfs_release_path(root, path);
337                         return 0;
338                 }
339                 dst_copy = kmalloc(item_size, GFP_NOFS);
340                 src_copy = kmalloc(item_size, GFP_NOFS);
341
342                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
343
344                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
345                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
346                                    item_size);
347                 ret = memcmp(dst_copy, src_copy, item_size);
348
349                 kfree(dst_copy);
350                 kfree(src_copy);
351                 /*
352                  * they have the same contents, just return, this saves
353                  * us from cowing blocks in the destination tree and doing
354                  * extra writes that may not have been done by a previous
355                  * sync
356                  */
357                 if (ret == 0) {
358                         btrfs_release_path(root, path);
359                         return 0;
360                 }
361
362         }
363 insert:
364         btrfs_release_path(root, path);
365         /* try to insert the key into the destination tree */
366         ret = btrfs_insert_empty_item(trans, root, path,
367                                       key, item_size);
368
369         /* make sure any existing item is the correct size */
370         if (ret == -EEXIST) {
371                 u32 found_size;
372                 found_size = btrfs_item_size_nr(path->nodes[0],
373                                                 path->slots[0]);
374                 if (found_size > item_size) {
375                         btrfs_truncate_item(trans, root, path, item_size, 1);
376                 } else if (found_size < item_size) {
377                         ret = btrfs_del_item(trans, root,
378                                              path);
379                         BUG_ON(ret);
380
381                         btrfs_release_path(root, path);
382                         ret = btrfs_insert_empty_item(trans,
383                                   root, path, key, item_size);
384                         BUG_ON(ret);
385                 }
386         } else if (ret) {
387                 BUG();
388         }
389         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
390                                         path->slots[0]);
391
392         /* don't overwrite an existing inode if the generation number
393          * was logged as zero.  This is done when the tree logging code
394          * is just logging an inode to make sure it exists after recovery.
395          *
396          * Also, don't overwrite i_size on directories during replay.
397          * log replay inserts and removes directory items based on the
398          * state of the tree found in the subvolume, and i_size is modified
399          * as it goes
400          */
401         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
402                 struct btrfs_inode_item *src_item;
403                 struct btrfs_inode_item *dst_item;
404
405                 src_item = (struct btrfs_inode_item *)src_ptr;
406                 dst_item = (struct btrfs_inode_item *)dst_ptr;
407
408                 if (btrfs_inode_generation(eb, src_item) == 0)
409                         goto no_copy;
410
411                 if (overwrite_root &&
412                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
413                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
414                         save_old_i_size = 1;
415                         saved_i_size = btrfs_inode_size(path->nodes[0],
416                                                         dst_item);
417                 }
418         }
419
420         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
421                            src_ptr, item_size);
422
423         if (save_old_i_size) {
424                 struct btrfs_inode_item *dst_item;
425                 dst_item = (struct btrfs_inode_item *)dst_ptr;
426                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
427         }
428
429         /* make sure the generation is filled in */
430         if (key->type == BTRFS_INODE_ITEM_KEY) {
431                 struct btrfs_inode_item *dst_item;
432                 dst_item = (struct btrfs_inode_item *)dst_ptr;
433                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
434                         btrfs_set_inode_generation(path->nodes[0], dst_item,
435                                                    trans->transid);
436                 }
437         }
438
439         if (overwrite_root &&
440             key->type == BTRFS_EXTENT_DATA_KEY) {
441                 int extent_type;
442                 struct btrfs_file_extent_item *fi;
443
444                 fi = (struct btrfs_file_extent_item *)dst_ptr;
445                 extent_type = btrfs_file_extent_type(path->nodes[0], fi);
446                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
447                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
448                         struct btrfs_key ins;
449                         ins.objectid = btrfs_file_extent_disk_bytenr(
450                                                         path->nodes[0], fi);
451                         ins.offset = btrfs_file_extent_disk_num_bytes(
452                                                         path->nodes[0], fi);
453                         ins.type = BTRFS_EXTENT_ITEM_KEY;
454
455                         /*
456                          * is this extent already allocated in the extent
457                          * allocation tree?  If so, just add a reference
458                          */
459                         ret = btrfs_lookup_extent(root, ins.objectid,
460                                                   ins.offset);
461                         if (ret == 0) {
462                                 ret = btrfs_inc_extent_ref(trans, root,
463                                                 ins.objectid, ins.offset,
464                                                 path->nodes[0]->start,
465                                                 root->root_key.objectid,
466                                                 trans->transid, key->objectid);
467                         } else {
468                                 /*
469                                  * insert the extent pointer in the extent
470                                  * allocation tree
471                                  */
472                                 ret = btrfs_alloc_logged_extent(trans, root,
473                                                 path->nodes[0]->start,
474                                                 root->root_key.objectid,
475                                                 trans->transid, key->objectid,
476                                                 &ins);
477                                 BUG_ON(ret);
478                         }
479                 }
480         }
481 no_copy:
482         btrfs_mark_buffer_dirty(path->nodes[0]);
483         btrfs_release_path(root, path);
484         return 0;
485 }
486
487 /*
488  * simple helper to read an inode off the disk from a given root
489  * This can only be called for subvolume roots and not for the log
490  */
491 static noinline struct inode *read_one_inode(struct btrfs_root *root,
492                                              u64 objectid)
493 {
494         struct inode *inode;
495         inode = btrfs_iget_locked(root->fs_info->sb, objectid, root);
496         if (inode->i_state & I_NEW) {
497                 BTRFS_I(inode)->root = root;
498                 BTRFS_I(inode)->location.objectid = objectid;
499                 BTRFS_I(inode)->location.type = BTRFS_INODE_ITEM_KEY;
500                 BTRFS_I(inode)->location.offset = 0;
501                 btrfs_read_locked_inode(inode);
502                 unlock_new_inode(inode);
503
504         }
505         if (is_bad_inode(inode)) {
506                 iput(inode);
507                 inode = NULL;
508         }
509         return inode;
510 }
511
512 /* replays a single extent in 'eb' at 'slot' with 'key' into the
513  * subvolume 'root'.  path is released on entry and should be released
514  * on exit.
515  *
516  * extents in the log tree have not been allocated out of the extent
517  * tree yet.  So, this completes the allocation, taking a reference
518  * as required if the extent already exists or creating a new extent
519  * if it isn't in the extent allocation tree yet.
520  *
521  * The extent is inserted into the file, dropping any existing extents
522  * from the file that overlap the new one.
523  */
524 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
525                                       struct btrfs_root *root,
526                                       struct btrfs_path *path,
527                                       struct extent_buffer *eb, int slot,
528                                       struct btrfs_key *key)
529 {
530         int found_type;
531         u64 mask = root->sectorsize - 1;
532         u64 extent_end;
533         u64 alloc_hint;
534         u64 start = key->offset;
535         struct btrfs_file_extent_item *item;
536         struct inode *inode = NULL;
537         unsigned long size;
538         int ret = 0;
539
540         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
541         found_type = btrfs_file_extent_type(eb, item);
542
543         if (found_type == BTRFS_FILE_EXTENT_REG ||
544             found_type == BTRFS_FILE_EXTENT_PREALLOC)
545                 extent_end = start + btrfs_file_extent_num_bytes(eb, item);
546         else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
547                 size = btrfs_file_extent_inline_len(eb, item);
548                 extent_end = (start + size + mask) & ~mask;
549         } else {
550                 ret = 0;
551                 goto out;
552         }
553
554         inode = read_one_inode(root, key->objectid);
555         if (!inode) {
556                 ret = -EIO;
557                 goto out;
558         }
559
560         /*
561          * first check to see if we already have this extent in the
562          * file.  This must be done before the btrfs_drop_extents run
563          * so we don't try to drop this extent.
564          */
565         ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
566                                        start, 0);
567
568         if (ret == 0 &&
569             (found_type == BTRFS_FILE_EXTENT_REG ||
570              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
571                 struct btrfs_file_extent_item cmp1;
572                 struct btrfs_file_extent_item cmp2;
573                 struct btrfs_file_extent_item *existing;
574                 struct extent_buffer *leaf;
575
576                 leaf = path->nodes[0];
577                 existing = btrfs_item_ptr(leaf, path->slots[0],
578                                           struct btrfs_file_extent_item);
579
580                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
581                                    sizeof(cmp1));
582                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
583                                    sizeof(cmp2));
584
585                 /*
586                  * we already have a pointer to this exact extent,
587                  * we don't have to do anything
588                  */
589                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
590                         btrfs_release_path(root, path);
591                         goto out;
592                 }
593         }
594         btrfs_release_path(root, path);
595
596         /* drop any overlapping extents */
597         ret = btrfs_drop_extents(trans, root, inode,
598                          start, extent_end, start, &alloc_hint);
599         BUG_ON(ret);
600
601         /* insert the extent */
602         ret = overwrite_item(trans, root, path, eb, slot, key);
603         BUG_ON(ret);
604
605         /* btrfs_drop_extents changes i_bytes & i_blocks, update it here */
606         inode_add_bytes(inode, extent_end - start);
607         btrfs_update_inode(trans, root, inode);
608 out:
609         if (inode)
610                 iput(inode);
611         return ret;
612 }
613
614 /*
615  * when cleaning up conflicts between the directory names in the
616  * subvolume, directory names in the log and directory names in the
617  * inode back references, we may have to unlink inodes from directories.
618  *
619  * This is a helper function to do the unlink of a specific directory
620  * item
621  */
622 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
623                                       struct btrfs_root *root,
624                                       struct btrfs_path *path,
625                                       struct inode *dir,
626                                       struct btrfs_dir_item *di)
627 {
628         struct inode *inode;
629         char *name;
630         int name_len;
631         struct extent_buffer *leaf;
632         struct btrfs_key location;
633         int ret;
634
635         leaf = path->nodes[0];
636
637         btrfs_dir_item_key_to_cpu(leaf, di, &location);
638         name_len = btrfs_dir_name_len(leaf, di);
639         name = kmalloc(name_len, GFP_NOFS);
640         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
641         btrfs_release_path(root, path);
642
643         inode = read_one_inode(root, location.objectid);
644         BUG_ON(!inode);
645
646         btrfs_inc_nlink(inode);
647         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
648         kfree(name);
649
650         iput(inode);
651         return ret;
652 }
653
654 /*
655  * helper function to see if a given name and sequence number found
656  * in an inode back reference are already in a directory and correctly
657  * point to this inode
658  */
659 static noinline int inode_in_dir(struct btrfs_root *root,
660                                  struct btrfs_path *path,
661                                  u64 dirid, u64 objectid, u64 index,
662                                  const char *name, int name_len)
663 {
664         struct btrfs_dir_item *di;
665         struct btrfs_key location;
666         int match = 0;
667
668         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
669                                          index, name, name_len, 0);
670         if (di && !IS_ERR(di)) {
671                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
672                 if (location.objectid != objectid)
673                         goto out;
674         } else
675                 goto out;
676         btrfs_release_path(root, path);
677
678         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
679         if (di && !IS_ERR(di)) {
680                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
681                 if (location.objectid != objectid)
682                         goto out;
683         } else
684                 goto out;
685         match = 1;
686 out:
687         btrfs_release_path(root, path);
688         return match;
689 }
690
691 /*
692  * helper function to check a log tree for a named back reference in
693  * an inode.  This is used to decide if a back reference that is
694  * found in the subvolume conflicts with what we find in the log.
695  *
696  * inode backreferences may have multiple refs in a single item,
697  * during replay we process one reference at a time, and we don't
698  * want to delete valid links to a file from the subvolume if that
699  * link is also in the log.
700  */
701 static noinline int backref_in_log(struct btrfs_root *log,
702                                    struct btrfs_key *key,
703                                    char *name, int namelen)
704 {
705         struct btrfs_path *path;
706         struct btrfs_inode_ref *ref;
707         unsigned long ptr;
708         unsigned long ptr_end;
709         unsigned long name_ptr;
710         int found_name_len;
711         int item_size;
712         int ret;
713         int match = 0;
714
715         path = btrfs_alloc_path();
716         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
717         if (ret != 0)
718                 goto out;
719
720         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
721         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
722         ptr_end = ptr + item_size;
723         while (ptr < ptr_end) {
724                 ref = (struct btrfs_inode_ref *)ptr;
725                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
726                 if (found_name_len == namelen) {
727                         name_ptr = (unsigned long)(ref + 1);
728                         ret = memcmp_extent_buffer(path->nodes[0], name,
729                                                    name_ptr, namelen);
730                         if (ret == 0) {
731                                 match = 1;
732                                 goto out;
733                         }
734                 }
735                 ptr = (unsigned long)(ref + 1) + found_name_len;
736         }
737 out:
738         btrfs_free_path(path);
739         return match;
740 }
741
742
743 /*
744  * replay one inode back reference item found in the log tree.
745  * eb, slot and key refer to the buffer and key found in the log tree.
746  * root is the destination we are replaying into, and path is for temp
747  * use by this function.  (it should be released on return).
748  */
749 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
750                                   struct btrfs_root *root,
751                                   struct btrfs_root *log,
752                                   struct btrfs_path *path,
753                                   struct extent_buffer *eb, int slot,
754                                   struct btrfs_key *key)
755 {
756         struct inode *dir;
757         int ret;
758         struct btrfs_key location;
759         struct btrfs_inode_ref *ref;
760         struct btrfs_dir_item *di;
761         struct inode *inode;
762         char *name;
763         int namelen;
764         unsigned long ref_ptr;
765         unsigned long ref_end;
766
767         location.objectid = key->objectid;
768         location.type = BTRFS_INODE_ITEM_KEY;
769         location.offset = 0;
770
771         /*
772          * it is possible that we didn't log all the parent directories
773          * for a given inode.  If we don't find the dir, just don't
774          * copy the back ref in.  The link count fixup code will take
775          * care of the rest
776          */
777         dir = read_one_inode(root, key->offset);
778         if (!dir)
779                 return -ENOENT;
780
781         inode = read_one_inode(root, key->objectid);
782         BUG_ON(!dir);
783
784         ref_ptr = btrfs_item_ptr_offset(eb, slot);
785         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
786
787 again:
788         ref = (struct btrfs_inode_ref *)ref_ptr;
789
790         namelen = btrfs_inode_ref_name_len(eb, ref);
791         name = kmalloc(namelen, GFP_NOFS);
792         BUG_ON(!name);
793
794         read_extent_buffer(eb, name, (unsigned long)(ref + 1), namelen);
795
796         /* if we already have a perfect match, we're done */
797         if (inode_in_dir(root, path, dir->i_ino, inode->i_ino,
798                          btrfs_inode_ref_index(eb, ref),
799                          name, namelen)) {
800                 goto out;
801         }
802
803         /*
804          * look for a conflicting back reference in the metadata.
805          * if we find one we have to unlink that name of the file
806          * before we add our new link.  Later on, we overwrite any
807          * existing back reference, and we don't want to create
808          * dangling pointers in the directory.
809          */
810 conflict_again:
811         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
812         if (ret == 0) {
813                 char *victim_name;
814                 int victim_name_len;
815                 struct btrfs_inode_ref *victim_ref;
816                 unsigned long ptr;
817                 unsigned long ptr_end;
818                 struct extent_buffer *leaf = path->nodes[0];
819
820                 /* are we trying to overwrite a back ref for the root directory
821                  * if so, just jump out, we're done
822                  */
823                 if (key->objectid == key->offset)
824                         goto out_nowrite;
825
826                 /* check all the names in this back reference to see
827                  * if they are in the log.  if so, we allow them to stay
828                  * otherwise they must be unlinked as a conflict
829                  */
830                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
831                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
832                 while(ptr < ptr_end) {
833                         victim_ref = (struct btrfs_inode_ref *)ptr;
834                         victim_name_len = btrfs_inode_ref_name_len(leaf,
835                                                                    victim_ref);
836                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
837                         BUG_ON(!victim_name);
838
839                         read_extent_buffer(leaf, victim_name,
840                                            (unsigned long)(victim_ref + 1),
841                                            victim_name_len);
842
843                         if (!backref_in_log(log, key, victim_name,
844                                             victim_name_len)) {
845                                 btrfs_inc_nlink(inode);
846                                 btrfs_release_path(root, path);
847                                 ret = btrfs_unlink_inode(trans, root, dir,
848                                                          inode, victim_name,
849                                                          victim_name_len);
850                                 kfree(victim_name);
851                                 btrfs_release_path(root, path);
852                                 goto conflict_again;
853                         }
854                         kfree(victim_name);
855                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
856                 }
857                 BUG_ON(ret);
858         }
859         btrfs_release_path(root, path);
860
861         /* look for a conflicting sequence number */
862         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
863                                          btrfs_inode_ref_index(eb, ref),
864                                          name, namelen, 0);
865         if (di && !IS_ERR(di)) {
866                 ret = drop_one_dir_item(trans, root, path, dir, di);
867                 BUG_ON(ret);
868         }
869         btrfs_release_path(root, path);
870
871
872         /* look for a conflicting name */
873         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
874                                    name, namelen, 0);
875         if (di && !IS_ERR(di)) {
876                 ret = drop_one_dir_item(trans, root, path, dir, di);
877                 BUG_ON(ret);
878         }
879         btrfs_release_path(root, path);
880
881         /* insert our name */
882         ret = btrfs_add_link(trans, dir, inode, name, namelen, 0,
883                              btrfs_inode_ref_index(eb, ref));
884         BUG_ON(ret);
885
886         btrfs_update_inode(trans, root, inode);
887
888 out:
889         ref_ptr = (unsigned long)(ref + 1) + namelen;
890         kfree(name);
891         if (ref_ptr < ref_end)
892                 goto again;
893
894         /* finally write the back reference in the inode */
895         ret = overwrite_item(trans, root, path, eb, slot, key);
896         BUG_ON(ret);
897
898 out_nowrite:
899         btrfs_release_path(root, path);
900         iput(dir);
901         iput(inode);
902         return 0;
903 }
904
905 /*
906  * replay one csum item from the log tree into the subvolume 'root'
907  * eb, slot and key all refer to the log tree
908  * path is for temp use by this function and should be released on return
909  *
910  * This copies the checksums out of the log tree and inserts them into
911  * the subvolume.  Any existing checksums for this range in the file
912  * are overwritten, and new items are added where required.
913  *
914  * We keep this simple by reusing the btrfs_ordered_sum code from
915  * the data=ordered mode.  This basically means making a copy
916  * of all the checksums in ram, which we have to do anyway for kmap
917  * rules.
918  *
919  * The copy is then sent down to btrfs_csum_file_blocks, which
920  * does all the hard work of finding existing items in the file
921  * or adding new ones.
922  */
923 static noinline int replay_one_csum(struct btrfs_trans_handle *trans,
924                                       struct btrfs_root *root,
925                                       struct btrfs_path *path,
926                                       struct extent_buffer *eb, int slot,
927                                       struct btrfs_key *key)
928 {
929         int ret;
930         u32 item_size = btrfs_item_size_nr(eb, slot);
931         u64 cur_offset;
932         u16 csum_size =
933                 btrfs_super_csum_size(&root->fs_info->super_copy);
934         unsigned long file_bytes;
935         struct btrfs_ordered_sum *sums;
936         struct btrfs_sector_sum *sector_sum;
937         struct inode *inode;
938         unsigned long ptr;
939
940         file_bytes = (item_size / csum_size) * root->sectorsize;
941         inode = read_one_inode(root, key->objectid);
942         if (!inode) {
943                 return -EIO;
944         }
945
946         sums = kzalloc(btrfs_ordered_sum_size(root, file_bytes), GFP_NOFS);
947         if (!sums) {
948                 iput(inode);
949                 return -ENOMEM;
950         }
951
952         INIT_LIST_HEAD(&sums->list);
953         sums->len = file_bytes;
954         sums->file_offset = key->offset;
955
956         /*
957          * copy all the sums into the ordered sum struct
958          */
959         sector_sum = sums->sums;
960         cur_offset = key->offset;
961         ptr = btrfs_item_ptr_offset(eb, slot);
962         while(item_size > 0) {
963                 sector_sum->offset = cur_offset;
964                 read_extent_buffer(eb, &sector_sum->sum, ptr, csum_size);
965                 sector_sum++;
966                 item_size -= csum_size;
967                 ptr += csum_size;
968                 cur_offset += root->sectorsize;
969         }
970
971         /* let btrfs_csum_file_blocks add them into the file */
972         ret = btrfs_csum_file_blocks(trans, root, inode, sums);
973         BUG_ON(ret);
974         kfree(sums);
975         iput(inode);
976
977         return 0;
978 }
979 /*
980  * There are a few corners where the link count of the file can't
981  * be properly maintained during replay.  So, instead of adding
982  * lots of complexity to the log code, we just scan the backrefs
983  * for any file that has been through replay.
984  *
985  * The scan will update the link count on the inode to reflect the
986  * number of back refs found.  If it goes down to zero, the iput
987  * will free the inode.
988  */
989 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
990                                            struct btrfs_root *root,
991                                            struct inode *inode)
992 {
993         struct btrfs_path *path;
994         int ret;
995         struct btrfs_key key;
996         u64 nlink = 0;
997         unsigned long ptr;
998         unsigned long ptr_end;
999         int name_len;
1000
1001         key.objectid = inode->i_ino;
1002         key.type = BTRFS_INODE_REF_KEY;
1003         key.offset = (u64)-1;
1004
1005         path = btrfs_alloc_path();
1006
1007         while(1) {
1008                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1009                 if (ret < 0)
1010                         break;
1011                 if (ret > 0) {
1012                         if (path->slots[0] == 0)
1013                                 break;
1014                         path->slots[0]--;
1015                 }
1016                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1017                                       path->slots[0]);
1018                 if (key.objectid != inode->i_ino ||
1019                     key.type != BTRFS_INODE_REF_KEY)
1020                         break;
1021                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1022                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1023                                                    path->slots[0]);
1024                 while(ptr < ptr_end) {
1025                         struct btrfs_inode_ref *ref;
1026
1027                         ref = (struct btrfs_inode_ref *)ptr;
1028                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1029                                                             ref);
1030                         ptr = (unsigned long)(ref + 1) + name_len;
1031                         nlink++;
1032                 }
1033
1034                 if (key.offset == 0)
1035                         break;
1036                 key.offset--;
1037                 btrfs_release_path(root, path);
1038         }
1039         btrfs_free_path(path);
1040         if (nlink != inode->i_nlink) {
1041                 inode->i_nlink = nlink;
1042                 btrfs_update_inode(trans, root, inode);
1043         }
1044         BTRFS_I(inode)->index_cnt = (u64)-1;
1045
1046         return 0;
1047 }
1048
1049 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1050                                             struct btrfs_root *root,
1051                                             struct btrfs_path *path)
1052 {
1053         int ret;
1054         struct btrfs_key key;
1055         struct inode *inode;
1056
1057         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1058         key.type = BTRFS_ORPHAN_ITEM_KEY;
1059         key.offset = (u64)-1;
1060         while(1) {
1061                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1062                 if (ret < 0)
1063                         break;
1064
1065                 if (ret == 1) {
1066                         if (path->slots[0] == 0)
1067                                 break;
1068                         path->slots[0]--;
1069                 }
1070
1071                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1072                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1073                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1074                         break;
1075
1076                 ret = btrfs_del_item(trans, root, path);
1077                 BUG_ON(ret);
1078
1079                 btrfs_release_path(root, path);
1080                 inode = read_one_inode(root, key.offset);
1081                 BUG_ON(!inode);
1082
1083                 ret = fixup_inode_link_count(trans, root, inode);
1084                 BUG_ON(ret);
1085
1086                 iput(inode);
1087
1088                 if (key.offset == 0)
1089                         break;
1090                 key.offset--;
1091         }
1092         btrfs_release_path(root, path);
1093         return 0;
1094 }
1095
1096
1097 /*
1098  * record a given inode in the fixup dir so we can check its link
1099  * count when replay is done.  The link count is incremented here
1100  * so the inode won't go away until we check it
1101  */
1102 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1103                                       struct btrfs_root *root,
1104                                       struct btrfs_path *path,
1105                                       u64 objectid)
1106 {
1107         struct btrfs_key key;
1108         int ret = 0;
1109         struct inode *inode;
1110
1111         inode = read_one_inode(root, objectid);
1112         BUG_ON(!inode);
1113
1114         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1115         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1116         key.offset = objectid;
1117
1118         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1119
1120         btrfs_release_path(root, path);
1121         if (ret == 0) {
1122                 btrfs_inc_nlink(inode);
1123                 btrfs_update_inode(trans, root, inode);
1124         } else if (ret == -EEXIST) {
1125                 ret = 0;
1126         } else {
1127                 BUG();
1128         }
1129         iput(inode);
1130
1131         return ret;
1132 }
1133
1134 /*
1135  * when replaying the log for a directory, we only insert names
1136  * for inodes that actually exist.  This means an fsync on a directory
1137  * does not implicitly fsync all the new files in it
1138  */
1139 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1140                                     struct btrfs_root *root,
1141                                     struct btrfs_path *path,
1142                                     u64 dirid, u64 index,
1143                                     char *name, int name_len, u8 type,
1144                                     struct btrfs_key *location)
1145 {
1146         struct inode *inode;
1147         struct inode *dir;
1148         int ret;
1149
1150         inode = read_one_inode(root, location->objectid);
1151         if (!inode)
1152                 return -ENOENT;
1153
1154         dir = read_one_inode(root, dirid);
1155         if (!dir) {
1156                 iput(inode);
1157                 return -EIO;
1158         }
1159         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1160
1161         /* FIXME, put inode into FIXUP list */
1162
1163         iput(inode);
1164         iput(dir);
1165         return ret;
1166 }
1167
1168 /*
1169  * take a single entry in a log directory item and replay it into
1170  * the subvolume.
1171  *
1172  * if a conflicting item exists in the subdirectory already,
1173  * the inode it points to is unlinked and put into the link count
1174  * fix up tree.
1175  *
1176  * If a name from the log points to a file or directory that does
1177  * not exist in the FS, it is skipped.  fsyncs on directories
1178  * do not force down inodes inside that directory, just changes to the
1179  * names or unlinks in a directory.
1180  */
1181 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1182                                     struct btrfs_root *root,
1183                                     struct btrfs_path *path,
1184                                     struct extent_buffer *eb,
1185                                     struct btrfs_dir_item *di,
1186                                     struct btrfs_key *key)
1187 {
1188         char *name;
1189         int name_len;
1190         struct btrfs_dir_item *dst_di;
1191         struct btrfs_key found_key;
1192         struct btrfs_key log_key;
1193         struct inode *dir;
1194         u8 log_type;
1195         int exists;
1196         int ret;
1197
1198         dir = read_one_inode(root, key->objectid);
1199         BUG_ON(!dir);
1200
1201         name_len = btrfs_dir_name_len(eb, di);
1202         name = kmalloc(name_len, GFP_NOFS);
1203         log_type = btrfs_dir_type(eb, di);
1204         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1205                    name_len);
1206
1207         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1208         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1209         if (exists == 0)
1210                 exists = 1;
1211         else
1212                 exists = 0;
1213         btrfs_release_path(root, path);
1214
1215         if (key->type == BTRFS_DIR_ITEM_KEY) {
1216                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1217                                        name, name_len, 1);
1218         }
1219         else if (key->type == BTRFS_DIR_INDEX_KEY) {
1220                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1221                                                      key->objectid,
1222                                                      key->offset, name,
1223                                                      name_len, 1);
1224         } else {
1225                 BUG();
1226         }
1227         if (!dst_di || IS_ERR(dst_di)) {
1228                 /* we need a sequence number to insert, so we only
1229                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1230                  */
1231                 if (key->type != BTRFS_DIR_INDEX_KEY)
1232                         goto out;
1233                 goto insert;
1234         }
1235
1236         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1237         /* the existing item matches the logged item */
1238         if (found_key.objectid == log_key.objectid &&
1239             found_key.type == log_key.type &&
1240             found_key.offset == log_key.offset &&
1241             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1242                 goto out;
1243         }
1244
1245         /*
1246          * don't drop the conflicting directory entry if the inode
1247          * for the new entry doesn't exist
1248          */
1249         if (!exists)
1250                 goto out;
1251
1252         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1253         BUG_ON(ret);
1254
1255         if (key->type == BTRFS_DIR_INDEX_KEY)
1256                 goto insert;
1257 out:
1258         btrfs_release_path(root, path);
1259         kfree(name);
1260         iput(dir);
1261         return 0;
1262
1263 insert:
1264         btrfs_release_path(root, path);
1265         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1266                               name, name_len, log_type, &log_key);
1267
1268         if (ret && ret != -ENOENT)
1269                 BUG();
1270         goto out;
1271 }
1272
1273 /*
1274  * find all the names in a directory item and reconcile them into
1275  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1276  * one name in a directory item, but the same code gets used for
1277  * both directory index types
1278  */
1279 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1280                                         struct btrfs_root *root,
1281                                         struct btrfs_path *path,
1282                                         struct extent_buffer *eb, int slot,
1283                                         struct btrfs_key *key)
1284 {
1285         int ret;
1286         u32 item_size = btrfs_item_size_nr(eb, slot);
1287         struct btrfs_dir_item *di;
1288         int name_len;
1289         unsigned long ptr;
1290         unsigned long ptr_end;
1291
1292         ptr = btrfs_item_ptr_offset(eb, slot);
1293         ptr_end = ptr + item_size;
1294         while(ptr < ptr_end) {
1295                 di = (struct btrfs_dir_item *)ptr;
1296                 name_len = btrfs_dir_name_len(eb, di);
1297                 ret = replay_one_name(trans, root, path, eb, di, key);
1298                 BUG_ON(ret);
1299                 ptr = (unsigned long)(di + 1);
1300                 ptr += name_len;
1301         }
1302         return 0;
1303 }
1304
1305 /*
1306  * directory replay has two parts.  There are the standard directory
1307  * items in the log copied from the subvolume, and range items
1308  * created in the log while the subvolume was logged.
1309  *
1310  * The range items tell us which parts of the key space the log
1311  * is authoritative for.  During replay, if a key in the subvolume
1312  * directory is in a logged range item, but not actually in the log
1313  * that means it was deleted from the directory before the fsync
1314  * and should be removed.
1315  */
1316 static noinline int find_dir_range(struct btrfs_root *root,
1317                                    struct btrfs_path *path,
1318                                    u64 dirid, int key_type,
1319                                    u64 *start_ret, u64 *end_ret)
1320 {
1321         struct btrfs_key key;
1322         u64 found_end;
1323         struct btrfs_dir_log_item *item;
1324         int ret;
1325         int nritems;
1326
1327         if (*start_ret == (u64)-1)
1328                 return 1;
1329
1330         key.objectid = dirid;
1331         key.type = key_type;
1332         key.offset = *start_ret;
1333
1334         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1335         if (ret < 0)
1336                 goto out;
1337         if (ret > 0) {
1338                 if (path->slots[0] == 0)
1339                         goto out;
1340                 path->slots[0]--;
1341         }
1342         if (ret != 0)
1343                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1344
1345         if (key.type != key_type || key.objectid != dirid) {
1346                 ret = 1;
1347                 goto next;
1348         }
1349         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1350                               struct btrfs_dir_log_item);
1351         found_end = btrfs_dir_log_end(path->nodes[0], item);
1352
1353         if (*start_ret >= key.offset && *start_ret <= found_end) {
1354                 ret = 0;
1355                 *start_ret = key.offset;
1356                 *end_ret = found_end;
1357                 goto out;
1358         }
1359         ret = 1;
1360 next:
1361         /* check the next slot in the tree to see if it is a valid item */
1362         nritems = btrfs_header_nritems(path->nodes[0]);
1363         if (path->slots[0] >= nritems) {
1364                 ret = btrfs_next_leaf(root, path);
1365                 if (ret)
1366                         goto out;
1367         } else {
1368                 path->slots[0]++;
1369         }
1370
1371         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1372
1373         if (key.type != key_type || key.objectid != dirid) {
1374                 ret = 1;
1375                 goto out;
1376         }
1377         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1378                               struct btrfs_dir_log_item);
1379         found_end = btrfs_dir_log_end(path->nodes[0], item);
1380         *start_ret = key.offset;
1381         *end_ret = found_end;
1382         ret = 0;
1383 out:
1384         btrfs_release_path(root, path);
1385         return ret;
1386 }
1387
1388 /*
1389  * this looks for a given directory item in the log.  If the directory
1390  * item is not in the log, the item is removed and the inode it points
1391  * to is unlinked
1392  */
1393 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1394                                       struct btrfs_root *root,
1395                                       struct btrfs_root *log,
1396                                       struct btrfs_path *path,
1397                                       struct btrfs_path *log_path,
1398                                       struct inode *dir,
1399                                       struct btrfs_key *dir_key)
1400 {
1401         int ret;
1402         struct extent_buffer *eb;
1403         int slot;
1404         u32 item_size;
1405         struct btrfs_dir_item *di;
1406         struct btrfs_dir_item *log_di;
1407         int name_len;
1408         unsigned long ptr;
1409         unsigned long ptr_end;
1410         char *name;
1411         struct inode *inode;
1412         struct btrfs_key location;
1413
1414 again:
1415         eb = path->nodes[0];
1416         slot = path->slots[0];
1417         item_size = btrfs_item_size_nr(eb, slot);
1418         ptr = btrfs_item_ptr_offset(eb, slot);
1419         ptr_end = ptr + item_size;
1420         while(ptr < ptr_end) {
1421                 di = (struct btrfs_dir_item *)ptr;
1422                 name_len = btrfs_dir_name_len(eb, di);
1423                 name = kmalloc(name_len, GFP_NOFS);
1424                 if (!name) {
1425                         ret = -ENOMEM;
1426                         goto out;
1427                 }
1428                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1429                                   name_len);
1430                 log_di = NULL;
1431                 if (dir_key->type == BTRFS_DIR_ITEM_KEY) {
1432                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1433                                                        dir_key->objectid,
1434                                                        name, name_len, 0);
1435                 } else if (dir_key->type == BTRFS_DIR_INDEX_KEY) {
1436                         log_di = btrfs_lookup_dir_index_item(trans, log,
1437                                                      log_path,
1438                                                      dir_key->objectid,
1439                                                      dir_key->offset,
1440                                                      name, name_len, 0);
1441                 }
1442                 if (!log_di || IS_ERR(log_di)) {
1443                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1444                         btrfs_release_path(root, path);
1445                         btrfs_release_path(log, log_path);
1446                         inode = read_one_inode(root, location.objectid);
1447                         BUG_ON(!inode);
1448
1449                         ret = link_to_fixup_dir(trans, root,
1450                                                 path, location.objectid);
1451                         BUG_ON(ret);
1452                         btrfs_inc_nlink(inode);
1453                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1454                                                  name, name_len);
1455                         BUG_ON(ret);
1456                         kfree(name);
1457                         iput(inode);
1458
1459                         /* there might still be more names under this key
1460                          * check and repeat if required
1461                          */
1462                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1463                                                 0, 0);
1464                         if (ret == 0)
1465                                 goto again;
1466                         ret = 0;
1467                         goto out;
1468                 }
1469                 btrfs_release_path(log, log_path);
1470                 kfree(name);
1471
1472                 ptr = (unsigned long)(di + 1);
1473                 ptr += name_len;
1474         }
1475         ret = 0;
1476 out:
1477         btrfs_release_path(root, path);
1478         btrfs_release_path(log, log_path);
1479         return ret;
1480 }
1481
1482 /*
1483  * deletion replay happens before we copy any new directory items
1484  * out of the log or out of backreferences from inodes.  It
1485  * scans the log to find ranges of keys that log is authoritative for,
1486  * and then scans the directory to find items in those ranges that are
1487  * not present in the log.
1488  *
1489  * Anything we don't find in the log is unlinked and removed from the
1490  * directory.
1491  */
1492 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1493                                        struct btrfs_root *root,
1494                                        struct btrfs_root *log,
1495                                        struct btrfs_path *path,
1496                                        u64 dirid)
1497 {
1498         u64 range_start;
1499         u64 range_end;
1500         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1501         int ret = 0;
1502         struct btrfs_key dir_key;
1503         struct btrfs_key found_key;
1504         struct btrfs_path *log_path;
1505         struct inode *dir;
1506
1507         dir_key.objectid = dirid;
1508         dir_key.type = BTRFS_DIR_ITEM_KEY;
1509         log_path = btrfs_alloc_path();
1510         if (!log_path)
1511                 return -ENOMEM;
1512
1513         dir = read_one_inode(root, dirid);
1514         /* it isn't an error if the inode isn't there, that can happen
1515          * because we replay the deletes before we copy in the inode item
1516          * from the log
1517          */
1518         if (!dir) {
1519                 btrfs_free_path(log_path);
1520                 return 0;
1521         }
1522 again:
1523         range_start = 0;
1524         range_end = 0;
1525         while(1) {
1526                 ret = find_dir_range(log, path, dirid, key_type,
1527                                      &range_start, &range_end);
1528                 if (ret != 0)
1529                         break;
1530
1531                 dir_key.offset = range_start;
1532                 while(1) {
1533                         int nritems;
1534                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
1535                                                 0, 0);
1536                         if (ret < 0)
1537                                 goto out;
1538
1539                         nritems = btrfs_header_nritems(path->nodes[0]);
1540                         if (path->slots[0] >= nritems) {
1541                                 ret = btrfs_next_leaf(root, path);
1542                                 if (ret)
1543                                         break;
1544                         }
1545                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1546                                               path->slots[0]);
1547                         if (found_key.objectid != dirid ||
1548                             found_key.type != dir_key.type)
1549                                 goto next_type;
1550
1551                         if (found_key.offset > range_end)
1552                                 break;
1553
1554                         ret = check_item_in_log(trans, root, log, path,
1555                                                 log_path, dir, &found_key);
1556                         BUG_ON(ret);
1557                         if (found_key.offset == (u64)-1)
1558                                 break;
1559                         dir_key.offset = found_key.offset + 1;
1560                 }
1561                 btrfs_release_path(root, path);
1562                 if (range_end == (u64)-1)
1563                         break;
1564                 range_start = range_end + 1;
1565         }
1566
1567 next_type:
1568         ret = 0;
1569         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1570                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1571                 dir_key.type = BTRFS_DIR_INDEX_KEY;
1572                 btrfs_release_path(root, path);
1573                 goto again;
1574         }
1575 out:
1576         btrfs_release_path(root, path);
1577         btrfs_free_path(log_path);
1578         iput(dir);
1579         return ret;
1580 }
1581
1582 /*
1583  * the process_func used to replay items from the log tree.  This
1584  * gets called in two different stages.  The first stage just looks
1585  * for inodes and makes sure they are all copied into the subvolume.
1586  *
1587  * The second stage copies all the other item types from the log into
1588  * the subvolume.  The two stage approach is slower, but gets rid of
1589  * lots of complexity around inodes referencing other inodes that exist
1590  * only in the log (references come from either directory items or inode
1591  * back refs).
1592  */
1593 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
1594                              struct walk_control *wc, u64 gen)
1595 {
1596         int nritems;
1597         struct btrfs_path *path;
1598         struct btrfs_root *root = wc->replay_dest;
1599         struct btrfs_key key;
1600         u32 item_size;
1601         int level;
1602         int i;
1603         int ret;
1604
1605         btrfs_read_buffer(eb, gen);
1606
1607         level = btrfs_header_level(eb);
1608
1609         if (level != 0)
1610                 return 0;
1611
1612         path = btrfs_alloc_path();
1613         BUG_ON(!path);
1614
1615         nritems = btrfs_header_nritems(eb);
1616         for (i = 0; i < nritems; i++) {
1617                 btrfs_item_key_to_cpu(eb, &key, i);
1618                 item_size = btrfs_item_size_nr(eb, i);
1619
1620                 /* inode keys are done during the first stage */
1621                 if (key.type == BTRFS_INODE_ITEM_KEY &&
1622                     wc->stage == LOG_WALK_REPLAY_INODES) {
1623                         struct inode *inode;
1624                         struct btrfs_inode_item *inode_item;
1625                         u32 mode;
1626
1627                         inode_item = btrfs_item_ptr(eb, i,
1628                                             struct btrfs_inode_item);
1629                         mode = btrfs_inode_mode(eb, inode_item);
1630                         if (S_ISDIR(mode)) {
1631                                 ret = replay_dir_deletes(wc->trans,
1632                                          root, log, path, key.objectid);
1633                                 BUG_ON(ret);
1634                         }
1635                         ret = overwrite_item(wc->trans, root, path,
1636                                              eb, i, &key);
1637                         BUG_ON(ret);
1638
1639                         /* for regular files, truncate away
1640                          * extents past the new EOF
1641                          */
1642                         if (S_ISREG(mode)) {
1643                                 inode = read_one_inode(root,
1644                                                        key.objectid);
1645                                 BUG_ON(!inode);
1646
1647                                 ret = btrfs_truncate_inode_items(wc->trans,
1648                                         root, inode, inode->i_size,
1649                                         BTRFS_EXTENT_DATA_KEY);
1650                                 BUG_ON(ret);
1651                                 iput(inode);
1652                         }
1653                         ret = link_to_fixup_dir(wc->trans, root,
1654                                                 path, key.objectid);
1655                         BUG_ON(ret);
1656                 }
1657                 if (wc->stage < LOG_WALK_REPLAY_ALL)
1658                         continue;
1659
1660                 /* these keys are simply copied */
1661                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
1662                         ret = overwrite_item(wc->trans, root, path,
1663                                              eb, i, &key);
1664                         BUG_ON(ret);
1665                 } else if (key.type == BTRFS_INODE_REF_KEY) {
1666                         ret = add_inode_ref(wc->trans, root, log, path,
1667                                             eb, i, &key);
1668                         BUG_ON(ret && ret != -ENOENT);
1669                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
1670                         ret = replay_one_extent(wc->trans, root, path,
1671                                                 eb, i, &key);
1672                         BUG_ON(ret);
1673                 } else if (key.type == BTRFS_CSUM_ITEM_KEY) {
1674                         ret = replay_one_csum(wc->trans, root, path,
1675                                               eb, i, &key);
1676                         BUG_ON(ret);
1677                 } else if (key.type == BTRFS_DIR_ITEM_KEY ||
1678                            key.type == BTRFS_DIR_INDEX_KEY) {
1679                         ret = replay_one_dir_item(wc->trans, root, path,
1680                                                   eb, i, &key);
1681                         BUG_ON(ret);
1682                 }
1683         }
1684         btrfs_free_path(path);
1685         return 0;
1686 }
1687
1688 static int noinline walk_down_log_tree(struct btrfs_trans_handle *trans,
1689                                    struct btrfs_root *root,
1690                                    struct btrfs_path *path, int *level,
1691                                    struct walk_control *wc)
1692 {
1693         u64 root_owner;
1694         u64 root_gen;
1695         u64 bytenr;
1696         u64 ptr_gen;
1697         struct extent_buffer *next;
1698         struct extent_buffer *cur;
1699         struct extent_buffer *parent;
1700         u32 blocksize;
1701         int ret = 0;
1702
1703         WARN_ON(*level < 0);
1704         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1705
1706         while(*level > 0) {
1707                 WARN_ON(*level < 0);
1708                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1709                 cur = path->nodes[*level];
1710
1711                 if (btrfs_header_level(cur) != *level)
1712                         WARN_ON(1);
1713
1714                 if (path->slots[*level] >=
1715                     btrfs_header_nritems(cur))
1716                         break;
1717
1718                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1719                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1720                 blocksize = btrfs_level_size(root, *level - 1);
1721
1722                 parent = path->nodes[*level];
1723                 root_owner = btrfs_header_owner(parent);
1724                 root_gen = btrfs_header_generation(parent);
1725
1726                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
1727
1728                 wc->process_func(root, next, wc, ptr_gen);
1729
1730                 if (*level == 1) {
1731                         path->slots[*level]++;
1732                         if (wc->free) {
1733                                 btrfs_read_buffer(next, ptr_gen);
1734
1735                                 btrfs_tree_lock(next);
1736                                 clean_tree_block(trans, root, next);
1737                                 btrfs_wait_tree_block_writeback(next);
1738                                 btrfs_tree_unlock(next);
1739
1740                                 ret = btrfs_drop_leaf_ref(trans, root, next);
1741                                 BUG_ON(ret);
1742
1743                                 WARN_ON(root_owner !=
1744                                         BTRFS_TREE_LOG_OBJECTID);
1745                                 ret = btrfs_free_reserved_extent(root,
1746                                                          bytenr, blocksize);
1747                                 BUG_ON(ret);
1748                         }
1749                         free_extent_buffer(next);
1750                         continue;
1751                 }
1752                 btrfs_read_buffer(next, ptr_gen);
1753
1754                 WARN_ON(*level <= 0);
1755                 if (path->nodes[*level-1])
1756                         free_extent_buffer(path->nodes[*level-1]);
1757                 path->nodes[*level-1] = next;
1758                 *level = btrfs_header_level(next);
1759                 path->slots[*level] = 0;
1760                 cond_resched();
1761         }
1762         WARN_ON(*level < 0);
1763         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1764
1765         if (path->nodes[*level] == root->node) {
1766                 parent = path->nodes[*level];
1767         } else {
1768                 parent = path->nodes[*level + 1];
1769         }
1770         bytenr = path->nodes[*level]->start;
1771
1772         blocksize = btrfs_level_size(root, *level);
1773         root_owner = btrfs_header_owner(parent);
1774         root_gen = btrfs_header_generation(parent);
1775
1776         wc->process_func(root, path->nodes[*level], wc,
1777                          btrfs_header_generation(path->nodes[*level]));
1778
1779         if (wc->free) {
1780                 next = path->nodes[*level];
1781                 btrfs_tree_lock(next);
1782                 clean_tree_block(trans, root, next);
1783                 btrfs_wait_tree_block_writeback(next);
1784                 btrfs_tree_unlock(next);
1785
1786                 if (*level == 0) {
1787                         ret = btrfs_drop_leaf_ref(trans, root, next);
1788                         BUG_ON(ret);
1789                 }
1790                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1791                 ret = btrfs_free_reserved_extent(root, bytenr, blocksize);
1792                 BUG_ON(ret);
1793         }
1794         free_extent_buffer(path->nodes[*level]);
1795         path->nodes[*level] = NULL;
1796         *level += 1;
1797
1798         cond_resched();
1799         return 0;
1800 }
1801
1802 static int noinline walk_up_log_tree(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_root *root,
1804                                  struct btrfs_path *path, int *level,
1805                                  struct walk_control *wc)
1806 {
1807         u64 root_owner;
1808         u64 root_gen;
1809         int i;
1810         int slot;
1811         int ret;
1812
1813         for(i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1814                 slot = path->slots[i];
1815                 if (slot < btrfs_header_nritems(path->nodes[i]) - 1) {
1816                         struct extent_buffer *node;
1817                         node = path->nodes[i];
1818                         path->slots[i]++;
1819                         *level = i;
1820                         WARN_ON(*level == 0);
1821                         return 0;
1822                 } else {
1823                         struct extent_buffer *parent;
1824                         if (path->nodes[*level] == root->node)
1825                                 parent = path->nodes[*level];
1826                         else
1827                                 parent = path->nodes[*level + 1];
1828
1829                         root_owner = btrfs_header_owner(parent);
1830                         root_gen = btrfs_header_generation(parent);
1831                         wc->process_func(root, path->nodes[*level], wc,
1832                                  btrfs_header_generation(path->nodes[*level]));
1833                         if (wc->free) {
1834                                 struct extent_buffer *next;
1835
1836                                 next = path->nodes[*level];
1837
1838                                 btrfs_tree_lock(next);
1839                                 clean_tree_block(trans, root, next);
1840                                 btrfs_wait_tree_block_writeback(next);
1841                                 btrfs_tree_unlock(next);
1842
1843                                 if (*level == 0) {
1844                                         ret = btrfs_drop_leaf_ref(trans, root,
1845                                                                   next);
1846                                         BUG_ON(ret);
1847                                 }
1848
1849                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1850                                 ret = btrfs_free_reserved_extent(root,
1851                                                 path->nodes[*level]->start,
1852                                                 path->nodes[*level]->len);
1853                                 BUG_ON(ret);
1854                         }
1855                         free_extent_buffer(path->nodes[*level]);
1856                         path->nodes[*level] = NULL;
1857                         *level = i + 1;
1858                 }
1859         }
1860         return 1;
1861 }
1862
1863 /*
1864  * drop the reference count on the tree rooted at 'snap'.  This traverses
1865  * the tree freeing any blocks that have a ref count of zero after being
1866  * decremented.
1867  */
1868 static int walk_log_tree(struct btrfs_trans_handle *trans,
1869                          struct btrfs_root *log, struct walk_control *wc)
1870 {
1871         int ret = 0;
1872         int wret;
1873         int level;
1874         struct btrfs_path *path;
1875         int i;
1876         int orig_level;
1877
1878         path = btrfs_alloc_path();
1879         BUG_ON(!path);
1880
1881         level = btrfs_header_level(log->node);
1882         orig_level = level;
1883         path->nodes[level] = log->node;
1884         extent_buffer_get(log->node);
1885         path->slots[level] = 0;
1886
1887         while(1) {
1888                 wret = walk_down_log_tree(trans, log, path, &level, wc);
1889                 if (wret > 0)
1890                         break;
1891                 if (wret < 0)
1892                         ret = wret;
1893
1894                 wret = walk_up_log_tree(trans, log, path, &level, wc);
1895                 if (wret > 0)
1896                         break;
1897                 if (wret < 0)
1898                         ret = wret;
1899         }
1900
1901         /* was the root node processed? if not, catch it here */
1902         if (path->nodes[orig_level]) {
1903                 wc->process_func(log, path->nodes[orig_level], wc,
1904                          btrfs_header_generation(path->nodes[orig_level]));
1905                 if (wc->free) {
1906                         struct extent_buffer *next;
1907
1908                         next = path->nodes[orig_level];
1909
1910                         btrfs_tree_lock(next);
1911                         clean_tree_block(trans, log, next);
1912                         btrfs_wait_tree_block_writeback(next);
1913                         btrfs_tree_unlock(next);
1914
1915                         if (orig_level == 0) {
1916                                 ret = btrfs_drop_leaf_ref(trans, log,
1917                                                           next);
1918                                 BUG_ON(ret);
1919                         }
1920                         WARN_ON(log->root_key.objectid !=
1921                                 BTRFS_TREE_LOG_OBJECTID);
1922                         ret = btrfs_free_reserved_extent(log, next->start,
1923                                                          next->len);
1924                         BUG_ON(ret);
1925                 }
1926         }
1927
1928         for (i = 0; i <= orig_level; i++) {
1929                 if (path->nodes[i]) {
1930                         free_extent_buffer(path->nodes[i]);
1931                         path->nodes[i] = NULL;
1932                 }
1933         }
1934         btrfs_free_path(path);
1935         if (wc->free)
1936                 free_extent_buffer(log->node);
1937         return ret;
1938 }
1939
1940 static int wait_log_commit(struct btrfs_root *log)
1941 {
1942         DEFINE_WAIT(wait);
1943         u64 transid = log->fs_info->tree_log_transid;
1944
1945         do {
1946                 prepare_to_wait(&log->fs_info->tree_log_wait, &wait,
1947                                 TASK_UNINTERRUPTIBLE);
1948                 mutex_unlock(&log->fs_info->tree_log_mutex);
1949                 if (atomic_read(&log->fs_info->tree_log_commit))
1950                         schedule();
1951                 finish_wait(&log->fs_info->tree_log_wait, &wait);
1952                 mutex_lock(&log->fs_info->tree_log_mutex);
1953         } while(transid == log->fs_info->tree_log_transid &&
1954                 atomic_read(&log->fs_info->tree_log_commit));
1955         return 0;
1956 }
1957
1958 /*
1959  * btrfs_sync_log does sends a given tree log down to the disk and
1960  * updates the super blocks to record it.  When this call is done,
1961  * you know that any inodes previously logged are safely on disk
1962  */
1963 int btrfs_sync_log(struct btrfs_trans_handle *trans,
1964                    struct btrfs_root *root)
1965 {
1966         int ret;
1967         unsigned long batch;
1968         struct btrfs_root *log = root->log_root;
1969
1970         mutex_lock(&log->fs_info->tree_log_mutex);
1971         if (atomic_read(&log->fs_info->tree_log_commit)) {
1972                 wait_log_commit(log);
1973                 goto out;
1974         }
1975         atomic_set(&log->fs_info->tree_log_commit, 1);
1976
1977         while(1) {
1978                 batch = log->fs_info->tree_log_batch;
1979                 mutex_unlock(&log->fs_info->tree_log_mutex);
1980                 schedule_timeout_uninterruptible(1);
1981                 mutex_lock(&log->fs_info->tree_log_mutex);
1982
1983                 while(atomic_read(&log->fs_info->tree_log_writers)) {
1984                         DEFINE_WAIT(wait);
1985                         prepare_to_wait(&log->fs_info->tree_log_wait, &wait,
1986                                         TASK_UNINTERRUPTIBLE);
1987                         mutex_unlock(&log->fs_info->tree_log_mutex);
1988                         if (atomic_read(&log->fs_info->tree_log_writers))
1989                                 schedule();
1990                         mutex_lock(&log->fs_info->tree_log_mutex);
1991                         finish_wait(&log->fs_info->tree_log_wait, &wait);
1992                 }
1993                 if (batch == log->fs_info->tree_log_batch)
1994                         break;
1995         }
1996
1997         ret = btrfs_write_and_wait_marked_extents(log, &log->dirty_log_pages);
1998         BUG_ON(ret);
1999         ret = btrfs_write_and_wait_marked_extents(root->fs_info->log_root_tree,
2000                                &root->fs_info->log_root_tree->dirty_log_pages);
2001         BUG_ON(ret);
2002
2003         btrfs_set_super_log_root(&root->fs_info->super_for_commit,
2004                                  log->fs_info->log_root_tree->node->start);
2005         btrfs_set_super_log_root_level(&root->fs_info->super_for_commit,
2006                        btrfs_header_level(log->fs_info->log_root_tree->node));
2007
2008         write_ctree_super(trans, log->fs_info->tree_root);
2009         log->fs_info->tree_log_transid++;
2010         log->fs_info->tree_log_batch = 0;
2011         atomic_set(&log->fs_info->tree_log_commit, 0);
2012         smp_mb();
2013         if (waitqueue_active(&log->fs_info->tree_log_wait))
2014                 wake_up(&log->fs_info->tree_log_wait);
2015 out:
2016         mutex_unlock(&log->fs_info->tree_log_mutex);
2017         return 0;
2018
2019 }
2020
2021 /* * free all the extents used by the tree log.  This should be called
2022  * at commit time of the full transaction
2023  */
2024 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2025 {
2026         int ret;
2027         struct btrfs_root *log;
2028         struct key;
2029         u64 start;
2030         u64 end;
2031         struct walk_control wc = {
2032                 .free = 1,
2033                 .process_func = process_one_buffer
2034         };
2035
2036         if (!root->log_root)
2037                 return 0;
2038
2039         log = root->log_root;
2040         ret = walk_log_tree(trans, log, &wc);
2041         BUG_ON(ret);
2042
2043         while(1) {
2044                 ret = find_first_extent_bit(&log->dirty_log_pages,
2045                                     0, &start, &end, EXTENT_DIRTY);
2046                 if (ret)
2047                         break;
2048
2049                 clear_extent_dirty(&log->dirty_log_pages,
2050                                    start, end, GFP_NOFS);
2051         }
2052
2053         log = root->log_root;
2054         ret = btrfs_del_root(trans, root->fs_info->log_root_tree,
2055                              &log->root_key);
2056         BUG_ON(ret);
2057         root->log_root = NULL;
2058         kfree(root->log_root);
2059         return 0;
2060 }
2061
2062 /*
2063  * helper function to update the item for a given subvolumes log root
2064  * in the tree of log roots
2065  */
2066 static int update_log_root(struct btrfs_trans_handle *trans,
2067                            struct btrfs_root *log)
2068 {
2069         u64 bytenr = btrfs_root_bytenr(&log->root_item);
2070         int ret;
2071
2072         if (log->node->start == bytenr)
2073                 return 0;
2074
2075         btrfs_set_root_bytenr(&log->root_item, log->node->start);
2076         btrfs_set_root_generation(&log->root_item, trans->transid);
2077         btrfs_set_root_level(&log->root_item, btrfs_header_level(log->node));
2078         ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2079                                 &log->root_key, &log->root_item);
2080         BUG_ON(ret);
2081         return ret;
2082 }
2083
2084 /*
2085  * If both a file and directory are logged, and unlinks or renames are
2086  * mixed in, we have a few interesting corners:
2087  *
2088  * create file X in dir Y
2089  * link file X to X.link in dir Y
2090  * fsync file X
2091  * unlink file X but leave X.link
2092  * fsync dir Y
2093  *
2094  * After a crash we would expect only X.link to exist.  But file X
2095  * didn't get fsync'd again so the log has back refs for X and X.link.
2096  *
2097  * We solve this by removing directory entries and inode backrefs from the
2098  * log when a file that was logged in the current transaction is
2099  * unlinked.  Any later fsync will include the updated log entries, and
2100  * we'll be able to reconstruct the proper directory items from backrefs.
2101  *
2102  * This optimizations allows us to avoid relogging the entire inode
2103  * or the entire directory.
2104  */
2105 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2106                                  struct btrfs_root *root,
2107                                  const char *name, int name_len,
2108                                  struct inode *dir, u64 index)
2109 {
2110         struct btrfs_root *log;
2111         struct btrfs_dir_item *di;
2112         struct btrfs_path *path;
2113         int ret;
2114         int bytes_del = 0;
2115
2116         if (BTRFS_I(dir)->logged_trans < trans->transid)
2117                 return 0;
2118
2119         ret = join_running_log_trans(root);
2120         if (ret)
2121                 return 0;
2122
2123         mutex_lock(&BTRFS_I(dir)->log_mutex);
2124
2125         log = root->log_root;
2126         path = btrfs_alloc_path();
2127         di = btrfs_lookup_dir_item(trans, log, path, dir->i_ino,
2128                                    name, name_len, -1);
2129         if (di && !IS_ERR(di)) {
2130                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2131                 bytes_del += name_len;
2132                 BUG_ON(ret);
2133         }
2134         btrfs_release_path(log, path);
2135         di = btrfs_lookup_dir_index_item(trans, log, path, dir->i_ino,
2136                                          index, name, name_len, -1);
2137         if (di && !IS_ERR(di)) {
2138                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2139                 bytes_del += name_len;
2140                 BUG_ON(ret);
2141         }
2142
2143         /* update the directory size in the log to reflect the names
2144          * we have removed
2145          */
2146         if (bytes_del) {
2147                 struct btrfs_key key;
2148
2149                 key.objectid = dir->i_ino;
2150                 key.offset = 0;
2151                 key.type = BTRFS_INODE_ITEM_KEY;
2152                 btrfs_release_path(log, path);
2153
2154                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2155                 if (ret == 0) {
2156                         struct btrfs_inode_item *item;
2157                         u64 i_size;
2158
2159                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2160                                               struct btrfs_inode_item);
2161                         i_size = btrfs_inode_size(path->nodes[0], item);
2162                         if (i_size > bytes_del)
2163                                 i_size -= bytes_del;
2164                         else
2165                                 i_size = 0;
2166                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2167                         btrfs_mark_buffer_dirty(path->nodes[0]);
2168                 } else
2169                         ret = 0;
2170                 btrfs_release_path(log, path);
2171         }
2172
2173         btrfs_free_path(path);
2174         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2175         end_log_trans(root);
2176
2177         return 0;
2178 }
2179
2180 /* see comments for btrfs_del_dir_entries_in_log */
2181 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2182                                struct btrfs_root *root,
2183                                const char *name, int name_len,
2184                                struct inode *inode, u64 dirid)
2185 {
2186         struct btrfs_root *log;
2187         u64 index;
2188         int ret;
2189
2190         if (BTRFS_I(inode)->logged_trans < trans->transid)
2191                 return 0;
2192
2193         ret = join_running_log_trans(root);
2194         if (ret)
2195                 return 0;
2196         log = root->log_root;
2197         mutex_lock(&BTRFS_I(inode)->log_mutex);
2198
2199         ret = btrfs_del_inode_ref(trans, log, name, name_len, inode->i_ino,
2200                                   dirid, &index);
2201         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2202         end_log_trans(root);
2203
2204         return ret;
2205 }
2206
2207 /*
2208  * creates a range item in the log for 'dirid'.  first_offset and
2209  * last_offset tell us which parts of the key space the log should
2210  * be considered authoritative for.
2211  */
2212 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2213                                        struct btrfs_root *log,
2214                                        struct btrfs_path *path,
2215                                        int key_type, u64 dirid,
2216                                        u64 first_offset, u64 last_offset)
2217 {
2218         int ret;
2219         struct btrfs_key key;
2220         struct btrfs_dir_log_item *item;
2221
2222         key.objectid = dirid;
2223         key.offset = first_offset;
2224         if (key_type == BTRFS_DIR_ITEM_KEY)
2225                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2226         else
2227                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2228         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2229         BUG_ON(ret);
2230
2231         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2232                               struct btrfs_dir_log_item);
2233         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2234         btrfs_mark_buffer_dirty(path->nodes[0]);
2235         btrfs_release_path(log, path);
2236         return 0;
2237 }
2238
2239 /*
2240  * log all the items included in the current transaction for a given
2241  * directory.  This also creates the range items in the log tree required
2242  * to replay anything deleted before the fsync
2243  */
2244 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2245                           struct btrfs_root *root, struct inode *inode,
2246                           struct btrfs_path *path,
2247                           struct btrfs_path *dst_path, int key_type,
2248                           u64 min_offset, u64 *last_offset_ret)
2249 {
2250         struct btrfs_key min_key;
2251         struct btrfs_key max_key;
2252         struct btrfs_root *log = root->log_root;
2253         struct extent_buffer *src;
2254         int ret;
2255         int i;
2256         int nritems;
2257         u64 first_offset = min_offset;
2258         u64 last_offset = (u64)-1;
2259
2260         log = root->log_root;
2261         max_key.objectid = inode->i_ino;
2262         max_key.offset = (u64)-1;
2263         max_key.type = key_type;
2264
2265         min_key.objectid = inode->i_ino;
2266         min_key.type = key_type;
2267         min_key.offset = min_offset;
2268
2269         path->keep_locks = 1;
2270
2271         ret = btrfs_search_forward(root, &min_key, &max_key,
2272                                    path, 0, trans->transid);
2273
2274         /*
2275          * we didn't find anything from this transaction, see if there
2276          * is anything at all
2277          */
2278         if (ret != 0 || min_key.objectid != inode->i_ino ||
2279             min_key.type != key_type) {
2280                 min_key.objectid = inode->i_ino;
2281                 min_key.type = key_type;
2282                 min_key.offset = (u64)-1;
2283                 btrfs_release_path(root, path);
2284                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2285                 if (ret < 0) {
2286                         btrfs_release_path(root, path);
2287                         return ret;
2288                 }
2289                 ret = btrfs_previous_item(root, path, inode->i_ino, key_type);
2290
2291                 /* if ret == 0 there are items for this type,
2292                  * create a range to tell us the last key of this type.
2293                  * otherwise, there are no items in this directory after
2294                  * *min_offset, and we create a range to indicate that.
2295                  */
2296                 if (ret == 0) {
2297                         struct btrfs_key tmp;
2298                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2299                                               path->slots[0]);
2300                         if (key_type == tmp.type) {
2301                                 first_offset = max(min_offset, tmp.offset) + 1;
2302                         }
2303                 }
2304                 goto done;
2305         }
2306
2307         /* go backward to find any previous key */
2308         ret = btrfs_previous_item(root, path, inode->i_ino, key_type);
2309         if (ret == 0) {
2310                 struct btrfs_key tmp;
2311                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2312                 if (key_type == tmp.type) {
2313                         first_offset = tmp.offset;
2314                         ret = overwrite_item(trans, log, dst_path,
2315                                              path->nodes[0], path->slots[0],
2316                                              &tmp);
2317                 }
2318         }
2319         btrfs_release_path(root, path);
2320
2321         /* find the first key from this transaction again */
2322         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2323         if (ret != 0) {
2324                 WARN_ON(1);
2325                 goto done;
2326         }
2327
2328         /*
2329          * we have a block from this transaction, log every item in it
2330          * from our directory
2331          */
2332         while(1) {
2333                 struct btrfs_key tmp;
2334                 src = path->nodes[0];
2335                 nritems = btrfs_header_nritems(src);
2336                 for (i = path->slots[0]; i < nritems; i++) {
2337                         btrfs_item_key_to_cpu(src, &min_key, i);
2338
2339                         if (min_key.objectid != inode->i_ino ||
2340                             min_key.type != key_type)
2341                                 goto done;
2342                         ret = overwrite_item(trans, log, dst_path, src, i,
2343                                              &min_key);
2344                         BUG_ON(ret);
2345                 }
2346                 path->slots[0] = nritems;
2347
2348                 /*
2349                  * look ahead to the next item and see if it is also
2350                  * from this directory and from this transaction
2351                  */
2352                 ret = btrfs_next_leaf(root, path);
2353                 if (ret == 1) {
2354                         last_offset = (u64)-1;
2355                         goto done;
2356                 }
2357                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2358                 if (tmp.objectid != inode->i_ino || tmp.type != key_type) {
2359                         last_offset = (u64)-1;
2360                         goto done;
2361                 }
2362                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2363                         ret = overwrite_item(trans, log, dst_path,
2364                                              path->nodes[0], path->slots[0],
2365                                              &tmp);
2366
2367                         BUG_ON(ret);
2368                         last_offset = tmp.offset;
2369                         goto done;
2370                 }
2371         }
2372 done:
2373         *last_offset_ret = last_offset;
2374         btrfs_release_path(root, path);
2375         btrfs_release_path(log, dst_path);
2376
2377         /* insert the log range keys to indicate where the log is valid */
2378         ret = insert_dir_log_key(trans, log, path, key_type, inode->i_ino,
2379                                  first_offset, last_offset);
2380         BUG_ON(ret);
2381         return 0;
2382 }
2383
2384 /*
2385  * logging directories is very similar to logging inodes, We find all the items
2386  * from the current transaction and write them to the log.
2387  *
2388  * The recovery code scans the directory in the subvolume, and if it finds a
2389  * key in the range logged that is not present in the log tree, then it means
2390  * that dir entry was unlinked during the transaction.
2391  *
2392  * In order for that scan to work, we must include one key smaller than
2393  * the smallest logged by this transaction and one key larger than the largest
2394  * key logged by this transaction.
2395  */
2396 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
2397                           struct btrfs_root *root, struct inode *inode,
2398                           struct btrfs_path *path,
2399                           struct btrfs_path *dst_path)
2400 {
2401         u64 min_key;
2402         u64 max_key;
2403         int ret;
2404         int key_type = BTRFS_DIR_ITEM_KEY;
2405
2406 again:
2407         min_key = 0;
2408         max_key = 0;
2409         while(1) {
2410                 ret = log_dir_items(trans, root, inode, path,
2411                                     dst_path, key_type, min_key,
2412                                     &max_key);
2413                 BUG_ON(ret);
2414                 if (max_key == (u64)-1)
2415                         break;
2416                 min_key = max_key + 1;
2417         }
2418
2419         if (key_type == BTRFS_DIR_ITEM_KEY) {
2420                 key_type = BTRFS_DIR_INDEX_KEY;
2421                 goto again;
2422         }
2423         return 0;
2424 }
2425
2426 /*
2427  * a helper function to drop items from the log before we relog an
2428  * inode.  max_key_type indicates the highest item type to remove.
2429  * This cannot be run for file data extents because it does not
2430  * free the extents they point to.
2431  */
2432 static int drop_objectid_items(struct btrfs_trans_handle *trans,
2433                                   struct btrfs_root *log,
2434                                   struct btrfs_path *path,
2435                                   u64 objectid, int max_key_type)
2436 {
2437         int ret;
2438         struct btrfs_key key;
2439         struct btrfs_key found_key;
2440
2441         key.objectid = objectid;
2442         key.type = max_key_type;
2443         key.offset = (u64)-1;
2444
2445         while(1) {
2446                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
2447
2448                 if (ret != 1)
2449                         break;
2450
2451                 if (path->slots[0] == 0)
2452                         break;
2453
2454                 path->slots[0]--;
2455                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2456                                       path->slots[0]);
2457
2458                 if (found_key.objectid != objectid)
2459                         break;
2460
2461                 ret = btrfs_del_item(trans, log, path);
2462                 BUG_ON(ret);
2463                 btrfs_release_path(log, path);
2464         }
2465         btrfs_release_path(log, path);
2466         return 0;
2467 }
2468
2469 static noinline int copy_items(struct btrfs_trans_handle *trans,
2470                                struct btrfs_root *log,
2471                                struct btrfs_path *dst_path,
2472                                struct extent_buffer *src,
2473                                int start_slot, int nr, int inode_only)
2474 {
2475         unsigned long src_offset;
2476         unsigned long dst_offset;
2477         struct btrfs_file_extent_item *extent;
2478         struct btrfs_inode_item *inode_item;
2479         int ret;
2480         struct btrfs_key *ins_keys;
2481         u32 *ins_sizes;
2482         char *ins_data;
2483         int i;
2484
2485         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
2486                            nr * sizeof(u32), GFP_NOFS);
2487         ins_sizes = (u32 *)ins_data;
2488         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
2489
2490         for (i = 0; i < nr; i++) {
2491                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
2492                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
2493         }
2494         ret = btrfs_insert_empty_items(trans, log, dst_path,
2495                                        ins_keys, ins_sizes, nr);
2496         BUG_ON(ret);
2497
2498         for (i = 0; i < nr; i++) {
2499                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
2500                                                    dst_path->slots[0]);
2501
2502                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
2503
2504                 copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
2505                                    src_offset, ins_sizes[i]);
2506
2507                 if (inode_only == LOG_INODE_EXISTS &&
2508                     ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
2509                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
2510                                                     dst_path->slots[0],
2511                                                     struct btrfs_inode_item);
2512                         btrfs_set_inode_size(dst_path->nodes[0], inode_item, 0);
2513
2514                         /* set the generation to zero so the recover code
2515                          * can tell the difference between an logging
2516                          * just to say 'this inode exists' and a logging
2517                          * to say 'update this inode with these values'
2518                          */
2519                         btrfs_set_inode_generation(dst_path->nodes[0],
2520                                                    inode_item, 0);
2521                 }
2522                 /* take a reference on file data extents so that truncates
2523                  * or deletes of this inode don't have to relog the inode
2524                  * again
2525                  */
2526                 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY) {
2527                         int found_type;
2528                         extent = btrfs_item_ptr(src, start_slot + i,
2529                                                 struct btrfs_file_extent_item);
2530
2531                         found_type = btrfs_file_extent_type(src, extent);
2532                         if (found_type == BTRFS_FILE_EXTENT_REG ||
2533                             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
2534                                 u64 ds = btrfs_file_extent_disk_bytenr(src,
2535                                                                    extent);
2536                                 u64 dl = btrfs_file_extent_disk_num_bytes(src,
2537                                                                       extent);
2538                                 /* ds == 0 is a hole */
2539                                 if (ds != 0) {
2540                                         ret = btrfs_inc_extent_ref(trans, log,
2541                                                    ds, dl,
2542                                                    dst_path->nodes[0]->start,
2543                                                    BTRFS_TREE_LOG_OBJECTID,
2544                                                    trans->transid,
2545                                                    ins_keys[i].objectid);
2546                                         BUG_ON(ret);
2547                                 }
2548                         }
2549                 }
2550                 dst_path->slots[0]++;
2551         }
2552
2553         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
2554         btrfs_release_path(log, dst_path);
2555         kfree(ins_data);
2556         return 0;
2557 }
2558
2559 /* log a single inode in the tree log.
2560  * At least one parent directory for this inode must exist in the tree
2561  * or be logged already.
2562  *
2563  * Any items from this inode changed by the current transaction are copied
2564  * to the log tree.  An extra reference is taken on any extents in this
2565  * file, allowing us to avoid a whole pile of corner cases around logging
2566  * blocks that have been removed from the tree.
2567  *
2568  * See LOG_INODE_ALL and related defines for a description of what inode_only
2569  * does.
2570  *
2571  * This handles both files and directories.
2572  */
2573 static int __btrfs_log_inode(struct btrfs_trans_handle *trans,
2574                              struct btrfs_root *root, struct inode *inode,
2575                              int inode_only)
2576 {
2577         struct btrfs_path *path;
2578         struct btrfs_path *dst_path;
2579         struct btrfs_key min_key;
2580         struct btrfs_key max_key;
2581         struct btrfs_root *log = root->log_root;
2582         struct extent_buffer *src = NULL;
2583         u32 size;
2584         int ret;
2585         int nritems;
2586         int ins_start_slot = 0;
2587         int ins_nr;
2588
2589         log = root->log_root;
2590
2591         path = btrfs_alloc_path();
2592         dst_path = btrfs_alloc_path();
2593
2594         min_key.objectid = inode->i_ino;
2595         min_key.type = BTRFS_INODE_ITEM_KEY;
2596         min_key.offset = 0;
2597
2598         max_key.objectid = inode->i_ino;
2599         if (inode_only == LOG_INODE_EXISTS || S_ISDIR(inode->i_mode))
2600                 max_key.type = BTRFS_XATTR_ITEM_KEY;
2601         else
2602                 max_key.type = (u8)-1;
2603         max_key.offset = (u64)-1;
2604
2605         /*
2606          * if this inode has already been logged and we're in inode_only
2607          * mode, we don't want to delete the things that have already
2608          * been written to the log.
2609          *
2610          * But, if the inode has been through an inode_only log,
2611          * the logged_trans field is not set.  This allows us to catch
2612          * any new names for this inode in the backrefs by logging it
2613          * again
2614          */
2615         if (inode_only == LOG_INODE_EXISTS &&
2616             BTRFS_I(inode)->logged_trans == trans->transid) {
2617                 btrfs_free_path(path);
2618                 btrfs_free_path(dst_path);
2619                 goto out;
2620         }
2621         mutex_lock(&BTRFS_I(inode)->log_mutex);
2622
2623         /*
2624          * a brute force approach to making sure we get the most uptodate
2625          * copies of everything.
2626          */
2627         if (S_ISDIR(inode->i_mode)) {
2628                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
2629
2630                 if (inode_only == LOG_INODE_EXISTS)
2631                         max_key_type = BTRFS_XATTR_ITEM_KEY;
2632                 ret = drop_objectid_items(trans, log, path,
2633                                           inode->i_ino, max_key_type);
2634         } else {
2635                 ret = btrfs_truncate_inode_items(trans, log, inode, 0, 0);
2636         }
2637         BUG_ON(ret);
2638         path->keep_locks = 1;
2639
2640         while(1) {
2641                 ins_nr = 0;
2642                 ret = btrfs_search_forward(root, &min_key, &max_key,
2643                                            path, 0, trans->transid);
2644                 if (ret != 0)
2645                         break;
2646 again:
2647                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
2648                 if (min_key.objectid != inode->i_ino)
2649                         break;
2650                 if (min_key.type > max_key.type)
2651                         break;
2652
2653                 src = path->nodes[0];
2654                 size = btrfs_item_size_nr(src, path->slots[0]);
2655                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
2656                         ins_nr++;
2657                         goto next_slot;
2658                 } else if (!ins_nr) {
2659                         ins_start_slot = path->slots[0];
2660                         ins_nr = 1;
2661                         goto next_slot;
2662                 }
2663
2664                 ret = copy_items(trans, log, dst_path, src, ins_start_slot,
2665                                  ins_nr, inode_only);
2666                 BUG_ON(ret);
2667                 ins_nr = 1;
2668                 ins_start_slot = path->slots[0];
2669 next_slot:
2670
2671                 nritems = btrfs_header_nritems(path->nodes[0]);
2672                 path->slots[0]++;
2673                 if (path->slots[0] < nritems) {
2674                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
2675                                               path->slots[0]);
2676                         goto again;
2677                 }
2678                 if (ins_nr) {
2679                         ret = copy_items(trans, log, dst_path, src,
2680                                          ins_start_slot,
2681                                          ins_nr, inode_only);
2682                         BUG_ON(ret);
2683                         ins_nr = 0;
2684                 }
2685                 btrfs_release_path(root, path);
2686
2687                 if (min_key.offset < (u64)-1)
2688                         min_key.offset++;
2689                 else if (min_key.type < (u8)-1)
2690                         min_key.type++;
2691                 else if (min_key.objectid < (u64)-1)
2692                         min_key.objectid++;
2693                 else
2694                         break;
2695         }
2696         if (ins_nr) {
2697                 ret = copy_items(trans, log, dst_path, src,
2698                                  ins_start_slot,
2699                                  ins_nr, inode_only);
2700                 BUG_ON(ret);
2701                 ins_nr = 0;
2702         }
2703         WARN_ON(ins_nr);
2704         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
2705                 btrfs_release_path(root, path);
2706                 btrfs_release_path(log, dst_path);
2707                 BTRFS_I(inode)->log_dirty_trans = 0;
2708                 ret = log_directory_changes(trans, root, inode, path, dst_path);
2709                 BUG_ON(ret);
2710         }
2711         BTRFS_I(inode)->logged_trans = trans->transid;
2712         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2713
2714         btrfs_free_path(path);
2715         btrfs_free_path(dst_path);
2716
2717         mutex_lock(&root->fs_info->tree_log_mutex);
2718         ret = update_log_root(trans, log);
2719         BUG_ON(ret);
2720         mutex_unlock(&root->fs_info->tree_log_mutex);
2721 out:
2722         return 0;
2723 }
2724
2725 int btrfs_log_inode(struct btrfs_trans_handle *trans,
2726                     struct btrfs_root *root, struct inode *inode,
2727                     int inode_only)
2728 {
2729         int ret;
2730
2731         start_log_trans(trans, root);
2732         ret = __btrfs_log_inode(trans, root, inode, inode_only);
2733         end_log_trans(root);
2734         return ret;
2735 }
2736
2737 /*
2738  * helper function around btrfs_log_inode to make sure newly created
2739  * parent directories also end up in the log.  A minimal inode and backref
2740  * only logging is done of any parent directories that are older than
2741  * the last committed transaction
2742  */
2743 int btrfs_log_dentry(struct btrfs_trans_handle *trans,
2744                     struct btrfs_root *root, struct dentry *dentry)
2745 {
2746         int inode_only = LOG_INODE_ALL;
2747         struct super_block *sb;
2748         int ret;
2749
2750         start_log_trans(trans, root);
2751         sb = dentry->d_inode->i_sb;
2752         while(1) {
2753                 ret = __btrfs_log_inode(trans, root, dentry->d_inode,
2754                                         inode_only);
2755                 BUG_ON(ret);
2756                 inode_only = LOG_INODE_EXISTS;
2757
2758                 dentry = dentry->d_parent;
2759                 if (!dentry || !dentry->d_inode || sb != dentry->d_inode->i_sb)
2760                         break;
2761
2762                 if (BTRFS_I(dentry->d_inode)->generation <=
2763                     root->fs_info->last_trans_committed)
2764                         break;
2765         }
2766         end_log_trans(root);
2767         return 0;
2768 }
2769
2770 /*
2771  * it is not safe to log dentry if the chunk root has added new
2772  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
2773  * If this returns 1, you must commit the transaction to safely get your
2774  * data on disk.
2775  */
2776 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
2777                           struct btrfs_root *root, struct dentry *dentry)
2778 {
2779         u64 gen;
2780         gen = root->fs_info->last_trans_new_blockgroup;
2781         if (gen > root->fs_info->last_trans_committed)
2782                 return 1;
2783         else
2784                 return btrfs_log_dentry(trans, root, dentry);
2785 }
2786
2787 /*
2788  * should be called during mount to recover any replay any log trees
2789  * from the FS
2790  */
2791 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
2792 {
2793         int ret;
2794         struct btrfs_path *path;
2795         struct btrfs_trans_handle *trans;
2796         struct btrfs_key key;
2797         struct btrfs_key found_key;
2798         struct btrfs_key tmp_key;
2799         struct btrfs_root *log;
2800         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
2801         u64 highest_inode;
2802         struct walk_control wc = {
2803                 .process_func = process_one_buffer,
2804                 .stage = 0,
2805         };
2806
2807         fs_info->log_root_recovering = 1;
2808         path = btrfs_alloc_path();
2809         BUG_ON(!path);
2810
2811         trans = btrfs_start_transaction(fs_info->tree_root, 1);
2812
2813         wc.trans = trans;
2814         wc.pin = 1;
2815
2816         walk_log_tree(trans, log_root_tree, &wc);
2817
2818 again:
2819         key.objectid = BTRFS_TREE_LOG_OBJECTID;
2820         key.offset = (u64)-1;
2821         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
2822
2823         while(1) {
2824                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
2825                 if (ret < 0)
2826                         break;
2827                 if (ret > 0) {
2828                         if (path->slots[0] == 0)
2829                                 break;
2830                         path->slots[0]--;
2831                 }
2832                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2833                                       path->slots[0]);
2834                 btrfs_release_path(log_root_tree, path);
2835                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
2836                         break;
2837
2838                 log = btrfs_read_fs_root_no_radix(log_root_tree,
2839                                                   &found_key);
2840                 BUG_ON(!log);
2841
2842
2843                 tmp_key.objectid = found_key.offset;
2844                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
2845                 tmp_key.offset = (u64)-1;
2846
2847                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
2848
2849                 BUG_ON(!wc.replay_dest);
2850
2851                 btrfs_record_root_in_trans(wc.replay_dest);
2852                 ret = walk_log_tree(trans, log, &wc);
2853                 BUG_ON(ret);
2854
2855                 if (wc.stage == LOG_WALK_REPLAY_ALL) {
2856                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
2857                                                       path);
2858                         BUG_ON(ret);
2859                 }
2860                 ret = btrfs_find_highest_inode(wc.replay_dest, &highest_inode);
2861                 if (ret == 0) {
2862                         wc.replay_dest->highest_inode = highest_inode;
2863                         wc.replay_dest->last_inode_alloc = highest_inode;
2864                 }
2865
2866                 key.offset = found_key.offset - 1;
2867                 free_extent_buffer(log->node);
2868                 kfree(log);
2869
2870                 if (found_key.offset == 0)
2871                         break;
2872         }
2873         btrfs_release_path(log_root_tree, path);
2874
2875         /* step one is to pin it all, step two is to replay just inodes */
2876         if (wc.pin) {
2877                 wc.pin = 0;
2878                 wc.process_func = replay_one_buffer;
2879                 wc.stage = LOG_WALK_REPLAY_INODES;
2880                 goto again;
2881         }
2882         /* step three is to replay everything */
2883         if (wc.stage < LOG_WALK_REPLAY_ALL) {
2884                 wc.stage++;
2885                 goto again;
2886         }
2887
2888         btrfs_free_path(path);
2889
2890         free_extent_buffer(log_root_tree->node);
2891         log_root_tree->log_root = NULL;
2892         fs_info->log_root_recovering = 0;
2893
2894         /* step 4: commit the transaction, which also unpins the blocks */
2895         btrfs_commit_transaction(trans, fs_info->tree_root);
2896
2897         kfree(log_root_tree);
2898         return 0;
2899 }