Btrfs: Add zlib compression support
[linux-2.6.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
27 /* magic values for the inode_only field in btrfs_log_inode:
28  *
29  * LOG_INODE_ALL means to log everything
30  * LOG_INODE_EXISTS means to log just enough to recreate the inode
31  * during log replay
32  */
33 #define LOG_INODE_ALL 0
34 #define LOG_INODE_EXISTS 1
35
36 /*
37  * stages for the tree walking.  The first
38  * stage (0) is to only pin down the blocks we find
39  * the second stage (1) is to make sure that all the inodes
40  * we find in the log are created in the subvolume.
41  *
42  * The last stage is to deal with directories and links and extents
43  * and all the other fun semantics
44  */
45 #define LOG_WALK_PIN_ONLY 0
46 #define LOG_WALK_REPLAY_INODES 1
47 #define LOG_WALK_REPLAY_ALL 2
48
49 static int __btrfs_log_inode(struct btrfs_trans_handle *trans,
50                              struct btrfs_root *root, struct inode *inode,
51                              int inode_only);
52
53 /*
54  * tree logging is a special write ahead log used to make sure that
55  * fsyncs and O_SYNCs can happen without doing full tree commits.
56  *
57  * Full tree commits are expensive because they require commonly
58  * modified blocks to be recowed, creating many dirty pages in the
59  * extent tree an 4x-6x higher write load than ext3.
60  *
61  * Instead of doing a tree commit on every fsync, we use the
62  * key ranges and transaction ids to find items for a given file or directory
63  * that have changed in this transaction.  Those items are copied into
64  * a special tree (one per subvolume root), that tree is written to disk
65  * and then the fsync is considered complete.
66  *
67  * After a crash, items are copied out of the log-tree back into the
68  * subvolume tree.  Any file data extents found are recorded in the extent
69  * allocation tree, and the log-tree freed.
70  *
71  * The log tree is read three times, once to pin down all the extents it is
72  * using in ram and once, once to create all the inodes logged in the tree
73  * and once to do all the other items.
74  */
75
76 /*
77  * btrfs_add_log_tree adds a new per-subvolume log tree into the
78  * tree of log tree roots.  This must be called with a tree log transaction
79  * running (see start_log_trans).
80  */
81 int btrfs_add_log_tree(struct btrfs_trans_handle *trans,
82                       struct btrfs_root *root)
83 {
84         struct btrfs_key key;
85         struct btrfs_root_item root_item;
86         struct btrfs_inode_item *inode_item;
87         struct extent_buffer *leaf;
88         struct btrfs_root *new_root = root;
89         int ret;
90         u64 objectid = root->root_key.objectid;
91
92         leaf = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
93                                       BTRFS_TREE_LOG_OBJECTID,
94                                       trans->transid, 0, 0, 0);
95         if (IS_ERR(leaf)) {
96                 ret = PTR_ERR(leaf);
97                 return ret;
98         }
99
100         btrfs_set_header_nritems(leaf, 0);
101         btrfs_set_header_level(leaf, 0);
102         btrfs_set_header_bytenr(leaf, leaf->start);
103         btrfs_set_header_generation(leaf, trans->transid);
104         btrfs_set_header_owner(leaf, BTRFS_TREE_LOG_OBJECTID);
105
106         write_extent_buffer(leaf, root->fs_info->fsid,
107                             (unsigned long)btrfs_header_fsid(leaf),
108                             BTRFS_FSID_SIZE);
109         btrfs_mark_buffer_dirty(leaf);
110
111         inode_item = &root_item.inode;
112         memset(inode_item, 0, sizeof(*inode_item));
113         inode_item->generation = cpu_to_le64(1);
114         inode_item->size = cpu_to_le64(3);
115         inode_item->nlink = cpu_to_le32(1);
116         inode_item->nbytes = cpu_to_le64(root->leafsize);
117         inode_item->mode = cpu_to_le32(S_IFDIR | 0755);
118
119         btrfs_set_root_bytenr(&root_item, leaf->start);
120         btrfs_set_root_level(&root_item, 0);
121         btrfs_set_root_refs(&root_item, 0);
122         btrfs_set_root_used(&root_item, 0);
123
124         memset(&root_item.drop_progress, 0, sizeof(root_item.drop_progress));
125         root_item.drop_level = 0;
126
127         btrfs_tree_unlock(leaf);
128         free_extent_buffer(leaf);
129         leaf = NULL;
130
131         btrfs_set_root_dirid(&root_item, 0);
132
133         key.objectid = BTRFS_TREE_LOG_OBJECTID;
134         key.offset = objectid;
135         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
136         ret = btrfs_insert_root(trans, root->fs_info->log_root_tree, &key,
137                                 &root_item);
138         if (ret)
139                 goto fail;
140
141         new_root = btrfs_read_fs_root_no_radix(root->fs_info->log_root_tree,
142                                                &key);
143         BUG_ON(!new_root);
144
145         WARN_ON(root->log_root);
146         root->log_root = new_root;
147
148         /*
149          * log trees do not get reference counted because they go away
150          * before a real commit is actually done.  They do store pointers
151          * to file data extents, and those reference counts still get
152          * updated (along with back refs to the log tree).
153          */
154         new_root->ref_cows = 0;
155         new_root->last_trans = trans->transid;
156 fail:
157         return ret;
158 }
159
160 /*
161  * start a sub transaction and setup the log tree
162  * this increments the log tree writer count to make the people
163  * syncing the tree wait for us to finish
164  */
165 static int start_log_trans(struct btrfs_trans_handle *trans,
166                            struct btrfs_root *root)
167 {
168         int ret;
169         mutex_lock(&root->fs_info->tree_log_mutex);
170         if (!root->fs_info->log_root_tree) {
171                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
172                 BUG_ON(ret);
173         }
174         if (!root->log_root) {
175                 ret = btrfs_add_log_tree(trans, root);
176                 BUG_ON(ret);
177         }
178         atomic_inc(&root->fs_info->tree_log_writers);
179         root->fs_info->tree_log_batch++;
180         mutex_unlock(&root->fs_info->tree_log_mutex);
181         return 0;
182 }
183
184 /*
185  * returns 0 if there was a log transaction running and we were able
186  * to join, or returns -ENOENT if there were not transactions
187  * in progress
188  */
189 static int join_running_log_trans(struct btrfs_root *root)
190 {
191         int ret = -ENOENT;
192
193         smp_mb();
194         if (!root->log_root)
195                 return -ENOENT;
196
197         mutex_lock(&root->fs_info->tree_log_mutex);
198         if (root->log_root) {
199                 ret = 0;
200                 atomic_inc(&root->fs_info->tree_log_writers);
201                 root->fs_info->tree_log_batch++;
202         }
203         mutex_unlock(&root->fs_info->tree_log_mutex);
204         return ret;
205 }
206
207 /*
208  * indicate we're done making changes to the log tree
209  * and wake up anyone waiting to do a sync
210  */
211 static int end_log_trans(struct btrfs_root *root)
212 {
213         atomic_dec(&root->fs_info->tree_log_writers);
214         smp_mb();
215         if (waitqueue_active(&root->fs_info->tree_log_wait))
216                 wake_up(&root->fs_info->tree_log_wait);
217         return 0;
218 }
219
220
221 /*
222  * the walk control struct is used to pass state down the chain when
223  * processing the log tree.  The stage field tells us which part
224  * of the log tree processing we are currently doing.  The others
225  * are state fields used for that specific part
226  */
227 struct walk_control {
228         /* should we free the extent on disk when done?  This is used
229          * at transaction commit time while freeing a log tree
230          */
231         int free;
232
233         /* should we write out the extent buffer?  This is used
234          * while flushing the log tree to disk during a sync
235          */
236         int write;
237
238         /* should we wait for the extent buffer io to finish?  Also used
239          * while flushing the log tree to disk for a sync
240          */
241         int wait;
242
243         /* pin only walk, we record which extents on disk belong to the
244          * log trees
245          */
246         int pin;
247
248         /* what stage of the replay code we're currently in */
249         int stage;
250
251         /* the root we are currently replaying */
252         struct btrfs_root *replay_dest;
253
254         /* the trans handle for the current replay */
255         struct btrfs_trans_handle *trans;
256
257         /* the function that gets used to process blocks we find in the
258          * tree.  Note the extent_buffer might not be up to date when it is
259          * passed in, and it must be checked or read if you need the data
260          * inside it
261          */
262         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
263                             struct walk_control *wc, u64 gen);
264 };
265
266 /*
267  * process_func used to pin down extents, write them or wait on them
268  */
269 static int process_one_buffer(struct btrfs_root *log,
270                               struct extent_buffer *eb,
271                               struct walk_control *wc, u64 gen)
272 {
273         if (wc->pin) {
274                 mutex_lock(&log->fs_info->alloc_mutex);
275                 btrfs_update_pinned_extents(log->fs_info->extent_root,
276                                             eb->start, eb->len, 1);
277                 mutex_unlock(&log->fs_info->alloc_mutex);
278         }
279
280         if (btrfs_buffer_uptodate(eb, gen)) {
281                 if (wc->write)
282                         btrfs_write_tree_block(eb);
283                 if (wc->wait)
284                         btrfs_wait_tree_block_writeback(eb);
285         }
286         return 0;
287 }
288
289 /*
290  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
291  * to the src data we are copying out.
292  *
293  * root is the tree we are copying into, and path is a scratch
294  * path for use in this function (it should be released on entry and
295  * will be released on exit).
296  *
297  * If the key is already in the destination tree the existing item is
298  * overwritten.  If the existing item isn't big enough, it is extended.
299  * If it is too large, it is truncated.
300  *
301  * If the key isn't in the destination yet, a new item is inserted.
302  */
303 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
304                                    struct btrfs_root *root,
305                                    struct btrfs_path *path,
306                                    struct extent_buffer *eb, int slot,
307                                    struct btrfs_key *key)
308 {
309         int ret;
310         u32 item_size;
311         u64 saved_i_size = 0;
312         int save_old_i_size = 0;
313         unsigned long src_ptr;
314         unsigned long dst_ptr;
315         int overwrite_root = 0;
316
317         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
318                 overwrite_root = 1;
319
320         item_size = btrfs_item_size_nr(eb, slot);
321         src_ptr = btrfs_item_ptr_offset(eb, slot);
322
323         /* look for the key in the destination tree */
324         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
325         if (ret == 0) {
326                 char *src_copy;
327                 char *dst_copy;
328                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
329                                                   path->slots[0]);
330                 if (dst_size != item_size)
331                         goto insert;
332
333                 if (item_size == 0) {
334                         btrfs_release_path(root, path);
335                         return 0;
336                 }
337                 dst_copy = kmalloc(item_size, GFP_NOFS);
338                 src_copy = kmalloc(item_size, GFP_NOFS);
339
340                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
341
342                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
343                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
344                                    item_size);
345                 ret = memcmp(dst_copy, src_copy, item_size);
346
347                 kfree(dst_copy);
348                 kfree(src_copy);
349                 /*
350                  * they have the same contents, just return, this saves
351                  * us from cowing blocks in the destination tree and doing
352                  * extra writes that may not have been done by a previous
353                  * sync
354                  */
355                 if (ret == 0) {
356                         btrfs_release_path(root, path);
357                         return 0;
358                 }
359
360         }
361 insert:
362         btrfs_release_path(root, path);
363         /* try to insert the key into the destination tree */
364         ret = btrfs_insert_empty_item(trans, root, path,
365                                       key, item_size);
366
367         /* make sure any existing item is the correct size */
368         if (ret == -EEXIST) {
369                 u32 found_size;
370                 found_size = btrfs_item_size_nr(path->nodes[0],
371                                                 path->slots[0]);
372                 if (found_size > item_size) {
373                         btrfs_truncate_item(trans, root, path, item_size, 1);
374                 } else if (found_size < item_size) {
375                         ret = btrfs_del_item(trans, root,
376                                              path);
377                         BUG_ON(ret);
378
379                         btrfs_release_path(root, path);
380                         ret = btrfs_insert_empty_item(trans,
381                                   root, path, key, item_size);
382                         BUG_ON(ret);
383                 }
384         } else if (ret) {
385                 BUG();
386         }
387         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
388                                         path->slots[0]);
389
390         /* don't overwrite an existing inode if the generation number
391          * was logged as zero.  This is done when the tree logging code
392          * is just logging an inode to make sure it exists after recovery.
393          *
394          * Also, don't overwrite i_size on directories during replay.
395          * log replay inserts and removes directory items based on the
396          * state of the tree found in the subvolume, and i_size is modified
397          * as it goes
398          */
399         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
400                 struct btrfs_inode_item *src_item;
401                 struct btrfs_inode_item *dst_item;
402
403                 src_item = (struct btrfs_inode_item *)src_ptr;
404                 dst_item = (struct btrfs_inode_item *)dst_ptr;
405
406                 if (btrfs_inode_generation(eb, src_item) == 0)
407                         goto no_copy;
408
409                 if (overwrite_root &&
410                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
411                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
412                         save_old_i_size = 1;
413                         saved_i_size = btrfs_inode_size(path->nodes[0],
414                                                         dst_item);
415                 }
416         }
417
418         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
419                            src_ptr, item_size);
420
421         if (save_old_i_size) {
422                 struct btrfs_inode_item *dst_item;
423                 dst_item = (struct btrfs_inode_item *)dst_ptr;
424                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
425         }
426
427         /* make sure the generation is filled in */
428         if (key->type == BTRFS_INODE_ITEM_KEY) {
429                 struct btrfs_inode_item *dst_item;
430                 dst_item = (struct btrfs_inode_item *)dst_ptr;
431                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
432                         btrfs_set_inode_generation(path->nodes[0], dst_item,
433                                                    trans->transid);
434                 }
435         }
436
437         if (overwrite_root &&
438             key->type == BTRFS_EXTENT_DATA_KEY) {
439                 int extent_type;
440                 struct btrfs_file_extent_item *fi;
441
442                 fi = (struct btrfs_file_extent_item *)dst_ptr;
443                 extent_type = btrfs_file_extent_type(path->nodes[0], fi);
444                 if (extent_type == BTRFS_FILE_EXTENT_REG) {
445                         struct btrfs_key ins;
446                         ins.objectid = btrfs_file_extent_disk_bytenr(
447                                                         path->nodes[0], fi);
448                         ins.offset = btrfs_file_extent_disk_num_bytes(
449                                                         path->nodes[0], fi);
450                         ins.type = BTRFS_EXTENT_ITEM_KEY;
451
452                         /*
453                          * is this extent already allocated in the extent
454                          * allocation tree?  If so, just add a reference
455                          */
456                         ret = btrfs_lookup_extent(root, ins.objectid,
457                                                   ins.offset);
458                         if (ret == 0) {
459                                 ret = btrfs_inc_extent_ref(trans, root,
460                                                 ins.objectid, ins.offset,
461                                                 path->nodes[0]->start,
462                                                 root->root_key.objectid,
463                                                 trans->transid, key->objectid);
464                         } else {
465                                 /*
466                                  * insert the extent pointer in the extent
467                                  * allocation tree
468                                  */
469                                 ret = btrfs_alloc_logged_extent(trans, root,
470                                                 path->nodes[0]->start,
471                                                 root->root_key.objectid,
472                                                 trans->transid, key->objectid,
473                                                 &ins);
474                                 BUG_ON(ret);
475                         }
476                 }
477         }
478 no_copy:
479         btrfs_mark_buffer_dirty(path->nodes[0]);
480         btrfs_release_path(root, path);
481         return 0;
482 }
483
484 /*
485  * simple helper to read an inode off the disk from a given root
486  * This can only be called for subvolume roots and not for the log
487  */
488 static noinline struct inode *read_one_inode(struct btrfs_root *root,
489                                              u64 objectid)
490 {
491         struct inode *inode;
492         inode = btrfs_iget_locked(root->fs_info->sb, objectid, root);
493         if (inode->i_state & I_NEW) {
494                 BTRFS_I(inode)->root = root;
495                 BTRFS_I(inode)->location.objectid = objectid;
496                 BTRFS_I(inode)->location.type = BTRFS_INODE_ITEM_KEY;
497                 BTRFS_I(inode)->location.offset = 0;
498                 btrfs_read_locked_inode(inode);
499                 unlock_new_inode(inode);
500
501         }
502         if (is_bad_inode(inode)) {
503                 iput(inode);
504                 inode = NULL;
505         }
506         return inode;
507 }
508
509 /* replays a single extent in 'eb' at 'slot' with 'key' into the
510  * subvolume 'root'.  path is released on entry and should be released
511  * on exit.
512  *
513  * extents in the log tree have not been allocated out of the extent
514  * tree yet.  So, this completes the allocation, taking a reference
515  * as required if the extent already exists or creating a new extent
516  * if it isn't in the extent allocation tree yet.
517  *
518  * The extent is inserted into the file, dropping any existing extents
519  * from the file that overlap the new one.
520  */
521 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
522                                       struct btrfs_root *root,
523                                       struct btrfs_path *path,
524                                       struct extent_buffer *eb, int slot,
525                                       struct btrfs_key *key)
526 {
527         int found_type;
528         u64 mask = root->sectorsize - 1;
529         u64 extent_end;
530         u64 alloc_hint;
531         u64 start = key->offset;
532         struct btrfs_file_extent_item *item;
533         struct inode *inode = NULL;
534         unsigned long size;
535         int ret = 0;
536
537         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
538         found_type = btrfs_file_extent_type(eb, item);
539
540         if (found_type == BTRFS_FILE_EXTENT_REG)
541                 extent_end = start + btrfs_file_extent_num_bytes(eb, item);
542         else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
543                 size = btrfs_file_extent_inline_len(eb, item);
544                 extent_end = (start + size + mask) & ~mask;
545         } else {
546                 ret = 0;
547                 goto out;
548         }
549
550         inode = read_one_inode(root, key->objectid);
551         if (!inode) {
552                 ret = -EIO;
553                 goto out;
554         }
555
556         /*
557          * first check to see if we already have this extent in the
558          * file.  This must be done before the btrfs_drop_extents run
559          * so we don't try to drop this extent.
560          */
561         ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
562                                        start, 0);
563
564         if (ret == 0 && found_type == BTRFS_FILE_EXTENT_REG) {
565                 struct btrfs_file_extent_item cmp1;
566                 struct btrfs_file_extent_item cmp2;
567                 struct btrfs_file_extent_item *existing;
568                 struct extent_buffer *leaf;
569
570                 leaf = path->nodes[0];
571                 existing = btrfs_item_ptr(leaf, path->slots[0],
572                                           struct btrfs_file_extent_item);
573
574                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
575                                    sizeof(cmp1));
576                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
577                                    sizeof(cmp2));
578
579                 /*
580                  * we already have a pointer to this exact extent,
581                  * we don't have to do anything
582                  */
583                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
584                         btrfs_release_path(root, path);
585                         goto out;
586                 }
587         }
588         btrfs_release_path(root, path);
589
590         /* drop any overlapping extents */
591         ret = btrfs_drop_extents(trans, root, inode,
592                          start, extent_end, start, &alloc_hint);
593         BUG_ON(ret);
594
595         /* insert the extent */
596         ret = overwrite_item(trans, root, path, eb, slot, key);
597         BUG_ON(ret);
598
599         /* btrfs_drop_extents changes i_bytes & i_blocks, update it here */
600         inode_add_bytes(inode, extent_end - start);
601         btrfs_update_inode(trans, root, inode);
602 out:
603         if (inode)
604                 iput(inode);
605         return ret;
606 }
607
608 /*
609  * when cleaning up conflicts between the directory names in the
610  * subvolume, directory names in the log and directory names in the
611  * inode back references, we may have to unlink inodes from directories.
612  *
613  * This is a helper function to do the unlink of a specific directory
614  * item
615  */
616 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
617                                       struct btrfs_root *root,
618                                       struct btrfs_path *path,
619                                       struct inode *dir,
620                                       struct btrfs_dir_item *di)
621 {
622         struct inode *inode;
623         char *name;
624         int name_len;
625         struct extent_buffer *leaf;
626         struct btrfs_key location;
627         int ret;
628
629         leaf = path->nodes[0];
630
631         btrfs_dir_item_key_to_cpu(leaf, di, &location);
632         name_len = btrfs_dir_name_len(leaf, di);
633         name = kmalloc(name_len, GFP_NOFS);
634         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
635         btrfs_release_path(root, path);
636
637         inode = read_one_inode(root, location.objectid);
638         BUG_ON(!inode);
639
640         btrfs_inc_nlink(inode);
641         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
642         kfree(name);
643
644         iput(inode);
645         return ret;
646 }
647
648 /*
649  * helper function to see if a given name and sequence number found
650  * in an inode back reference are already in a directory and correctly
651  * point to this inode
652  */
653 static noinline int inode_in_dir(struct btrfs_root *root,
654                                  struct btrfs_path *path,
655                                  u64 dirid, u64 objectid, u64 index,
656                                  const char *name, int name_len)
657 {
658         struct btrfs_dir_item *di;
659         struct btrfs_key location;
660         int match = 0;
661
662         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
663                                          index, name, name_len, 0);
664         if (di && !IS_ERR(di)) {
665                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
666                 if (location.objectid != objectid)
667                         goto out;
668         } else
669                 goto out;
670         btrfs_release_path(root, path);
671
672         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
673         if (di && !IS_ERR(di)) {
674                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
675                 if (location.objectid != objectid)
676                         goto out;
677         } else
678                 goto out;
679         match = 1;
680 out:
681         btrfs_release_path(root, path);
682         return match;
683 }
684
685 /*
686  * helper function to check a log tree for a named back reference in
687  * an inode.  This is used to decide if a back reference that is
688  * found in the subvolume conflicts with what we find in the log.
689  *
690  * inode backreferences may have multiple refs in a single item,
691  * during replay we process one reference at a time, and we don't
692  * want to delete valid links to a file from the subvolume if that
693  * link is also in the log.
694  */
695 static noinline int backref_in_log(struct btrfs_root *log,
696                                    struct btrfs_key *key,
697                                    char *name, int namelen)
698 {
699         struct btrfs_path *path;
700         struct btrfs_inode_ref *ref;
701         unsigned long ptr;
702         unsigned long ptr_end;
703         unsigned long name_ptr;
704         int found_name_len;
705         int item_size;
706         int ret;
707         int match = 0;
708
709         path = btrfs_alloc_path();
710         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
711         if (ret != 0)
712                 goto out;
713
714         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
715         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
716         ptr_end = ptr + item_size;
717         while (ptr < ptr_end) {
718                 ref = (struct btrfs_inode_ref *)ptr;
719                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
720                 if (found_name_len == namelen) {
721                         name_ptr = (unsigned long)(ref + 1);
722                         ret = memcmp_extent_buffer(path->nodes[0], name,
723                                                    name_ptr, namelen);
724                         if (ret == 0) {
725                                 match = 1;
726                                 goto out;
727                         }
728                 }
729                 ptr = (unsigned long)(ref + 1) + found_name_len;
730         }
731 out:
732         btrfs_free_path(path);
733         return match;
734 }
735
736
737 /*
738  * replay one inode back reference item found in the log tree.
739  * eb, slot and key refer to the buffer and key found in the log tree.
740  * root is the destination we are replaying into, and path is for temp
741  * use by this function.  (it should be released on return).
742  */
743 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
744                                   struct btrfs_root *root,
745                                   struct btrfs_root *log,
746                                   struct btrfs_path *path,
747                                   struct extent_buffer *eb, int slot,
748                                   struct btrfs_key *key)
749 {
750         struct inode *dir;
751         int ret;
752         struct btrfs_key location;
753         struct btrfs_inode_ref *ref;
754         struct btrfs_dir_item *di;
755         struct inode *inode;
756         char *name;
757         int namelen;
758         unsigned long ref_ptr;
759         unsigned long ref_end;
760
761         location.objectid = key->objectid;
762         location.type = BTRFS_INODE_ITEM_KEY;
763         location.offset = 0;
764
765         /*
766          * it is possible that we didn't log all the parent directories
767          * for a given inode.  If we don't find the dir, just don't
768          * copy the back ref in.  The link count fixup code will take
769          * care of the rest
770          */
771         dir = read_one_inode(root, key->offset);
772         if (!dir)
773                 return -ENOENT;
774
775         inode = read_one_inode(root, key->objectid);
776         BUG_ON(!dir);
777
778         ref_ptr = btrfs_item_ptr_offset(eb, slot);
779         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
780
781 again:
782         ref = (struct btrfs_inode_ref *)ref_ptr;
783
784         namelen = btrfs_inode_ref_name_len(eb, ref);
785         name = kmalloc(namelen, GFP_NOFS);
786         BUG_ON(!name);
787
788         read_extent_buffer(eb, name, (unsigned long)(ref + 1), namelen);
789
790         /* if we already have a perfect match, we're done */
791         if (inode_in_dir(root, path, dir->i_ino, inode->i_ino,
792                          btrfs_inode_ref_index(eb, ref),
793                          name, namelen)) {
794                 goto out;
795         }
796
797         /*
798          * look for a conflicting back reference in the metadata.
799          * if we find one we have to unlink that name of the file
800          * before we add our new link.  Later on, we overwrite any
801          * existing back reference, and we don't want to create
802          * dangling pointers in the directory.
803          */
804 conflict_again:
805         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
806         if (ret == 0) {
807                 char *victim_name;
808                 int victim_name_len;
809                 struct btrfs_inode_ref *victim_ref;
810                 unsigned long ptr;
811                 unsigned long ptr_end;
812                 struct extent_buffer *leaf = path->nodes[0];
813
814                 /* are we trying to overwrite a back ref for the root directory
815                  * if so, just jump out, we're done
816                  */
817                 if (key->objectid == key->offset)
818                         goto out_nowrite;
819
820                 /* check all the names in this back reference to see
821                  * if they are in the log.  if so, we allow them to stay
822                  * otherwise they must be unlinked as a conflict
823                  */
824                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
825                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
826                 while(ptr < ptr_end) {
827                         victim_ref = (struct btrfs_inode_ref *)ptr;
828                         victim_name_len = btrfs_inode_ref_name_len(leaf,
829                                                                    victim_ref);
830                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
831                         BUG_ON(!victim_name);
832
833                         read_extent_buffer(leaf, victim_name,
834                                            (unsigned long)(victim_ref + 1),
835                                            victim_name_len);
836
837                         if (!backref_in_log(log, key, victim_name,
838                                             victim_name_len)) {
839                                 btrfs_inc_nlink(inode);
840                                 btrfs_release_path(root, path);
841                                 ret = btrfs_unlink_inode(trans, root, dir,
842                                                          inode, victim_name,
843                                                          victim_name_len);
844                                 kfree(victim_name);
845                                 btrfs_release_path(root, path);
846                                 goto conflict_again;
847                         }
848                         kfree(victim_name);
849                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
850                 }
851                 BUG_ON(ret);
852         }
853         btrfs_release_path(root, path);
854
855         /* look for a conflicting sequence number */
856         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
857                                          btrfs_inode_ref_index(eb, ref),
858                                          name, namelen, 0);
859         if (di && !IS_ERR(di)) {
860                 ret = drop_one_dir_item(trans, root, path, dir, di);
861                 BUG_ON(ret);
862         }
863         btrfs_release_path(root, path);
864
865
866         /* look for a conflicting name */
867         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
868                                    name, namelen, 0);
869         if (di && !IS_ERR(di)) {
870                 ret = drop_one_dir_item(trans, root, path, dir, di);
871                 BUG_ON(ret);
872         }
873         btrfs_release_path(root, path);
874
875         /* insert our name */
876         ret = btrfs_add_link(trans, dir, inode, name, namelen, 0,
877                              btrfs_inode_ref_index(eb, ref));
878         BUG_ON(ret);
879
880         btrfs_update_inode(trans, root, inode);
881
882 out:
883         ref_ptr = (unsigned long)(ref + 1) + namelen;
884         kfree(name);
885         if (ref_ptr < ref_end)
886                 goto again;
887
888         /* finally write the back reference in the inode */
889         ret = overwrite_item(trans, root, path, eb, slot, key);
890         BUG_ON(ret);
891
892 out_nowrite:
893         btrfs_release_path(root, path);
894         iput(dir);
895         iput(inode);
896         return 0;
897 }
898
899 /*
900  * replay one csum item from the log tree into the subvolume 'root'
901  * eb, slot and key all refer to the log tree
902  * path is for temp use by this function and should be released on return
903  *
904  * This copies the checksums out of the log tree and inserts them into
905  * the subvolume.  Any existing checksums for this range in the file
906  * are overwritten, and new items are added where required.
907  *
908  * We keep this simple by reusing the btrfs_ordered_sum code from
909  * the data=ordered mode.  This basically means making a copy
910  * of all the checksums in ram, which we have to do anyway for kmap
911  * rules.
912  *
913  * The copy is then sent down to btrfs_csum_file_blocks, which
914  * does all the hard work of finding existing items in the file
915  * or adding new ones.
916  */
917 static noinline int replay_one_csum(struct btrfs_trans_handle *trans,
918                                       struct btrfs_root *root,
919                                       struct btrfs_path *path,
920                                       struct extent_buffer *eb, int slot,
921                                       struct btrfs_key *key)
922 {
923         int ret;
924         u32 item_size = btrfs_item_size_nr(eb, slot);
925         u64 cur_offset;
926         unsigned long file_bytes;
927         struct btrfs_ordered_sum *sums;
928         struct btrfs_sector_sum *sector_sum;
929         struct inode *inode;
930         unsigned long ptr;
931
932         file_bytes = (item_size / BTRFS_CRC32_SIZE) * root->sectorsize;
933         inode = read_one_inode(root, key->objectid);
934         if (!inode) {
935                 return -EIO;
936         }
937
938         sums = kzalloc(btrfs_ordered_sum_size(root, file_bytes), GFP_NOFS);
939         if (!sums) {
940                 iput(inode);
941                 return -ENOMEM;
942         }
943
944         INIT_LIST_HEAD(&sums->list);
945         sums->len = file_bytes;
946         sums->file_offset = key->offset;
947
948         /*
949          * copy all the sums into the ordered sum struct
950          */
951         sector_sum = sums->sums;
952         cur_offset = key->offset;
953         ptr = btrfs_item_ptr_offset(eb, slot);
954         while(item_size > 0) {
955                 sector_sum->offset = cur_offset;
956                 read_extent_buffer(eb, &sector_sum->sum, ptr, BTRFS_CRC32_SIZE);
957                 sector_sum++;
958                 item_size -= BTRFS_CRC32_SIZE;
959                 ptr += BTRFS_CRC32_SIZE;
960                 cur_offset += root->sectorsize;
961         }
962
963         /* let btrfs_csum_file_blocks add them into the file */
964         ret = btrfs_csum_file_blocks(trans, root, inode, sums);
965         BUG_ON(ret);
966         kfree(sums);
967         iput(inode);
968
969         return 0;
970 }
971 /*
972  * There are a few corners where the link count of the file can't
973  * be properly maintained during replay.  So, instead of adding
974  * lots of complexity to the log code, we just scan the backrefs
975  * for any file that has been through replay.
976  *
977  * The scan will update the link count on the inode to reflect the
978  * number of back refs found.  If it goes down to zero, the iput
979  * will free the inode.
980  */
981 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
982                                            struct btrfs_root *root,
983                                            struct inode *inode)
984 {
985         struct btrfs_path *path;
986         int ret;
987         struct btrfs_key key;
988         u64 nlink = 0;
989         unsigned long ptr;
990         unsigned long ptr_end;
991         int name_len;
992
993         key.objectid = inode->i_ino;
994         key.type = BTRFS_INODE_REF_KEY;
995         key.offset = (u64)-1;
996
997         path = btrfs_alloc_path();
998
999         while(1) {
1000                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1001                 if (ret < 0)
1002                         break;
1003                 if (ret > 0) {
1004                         if (path->slots[0] == 0)
1005                                 break;
1006                         path->slots[0]--;
1007                 }
1008                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1009                                       path->slots[0]);
1010                 if (key.objectid != inode->i_ino ||
1011                     key.type != BTRFS_INODE_REF_KEY)
1012                         break;
1013                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1014                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1015                                                    path->slots[0]);
1016                 while(ptr < ptr_end) {
1017                         struct btrfs_inode_ref *ref;
1018
1019                         ref = (struct btrfs_inode_ref *)ptr;
1020                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1021                                                             ref);
1022                         ptr = (unsigned long)(ref + 1) + name_len;
1023                         nlink++;
1024                 }
1025
1026                 if (key.offset == 0)
1027                         break;
1028                 key.offset--;
1029                 btrfs_release_path(root, path);
1030         }
1031         btrfs_free_path(path);
1032         if (nlink != inode->i_nlink) {
1033                 inode->i_nlink = nlink;
1034                 btrfs_update_inode(trans, root, inode);
1035         }
1036         BTRFS_I(inode)->index_cnt = (u64)-1;
1037
1038         return 0;
1039 }
1040
1041 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1042                                             struct btrfs_root *root,
1043                                             struct btrfs_path *path)
1044 {
1045         int ret;
1046         struct btrfs_key key;
1047         struct inode *inode;
1048
1049         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1050         key.type = BTRFS_ORPHAN_ITEM_KEY;
1051         key.offset = (u64)-1;
1052         while(1) {
1053                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1054                 if (ret < 0)
1055                         break;
1056
1057                 if (ret == 1) {
1058                         if (path->slots[0] == 0)
1059                                 break;
1060                         path->slots[0]--;
1061                 }
1062
1063                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1064                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1065                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1066                         break;
1067
1068                 ret = btrfs_del_item(trans, root, path);
1069                 BUG_ON(ret);
1070
1071                 btrfs_release_path(root, path);
1072                 inode = read_one_inode(root, key.offset);
1073                 BUG_ON(!inode);
1074
1075                 ret = fixup_inode_link_count(trans, root, inode);
1076                 BUG_ON(ret);
1077
1078                 iput(inode);
1079
1080                 if (key.offset == 0)
1081                         break;
1082                 key.offset--;
1083         }
1084         btrfs_release_path(root, path);
1085         return 0;
1086 }
1087
1088
1089 /*
1090  * record a given inode in the fixup dir so we can check its link
1091  * count when replay is done.  The link count is incremented here
1092  * so the inode won't go away until we check it
1093  */
1094 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1095                                       struct btrfs_root *root,
1096                                       struct btrfs_path *path,
1097                                       u64 objectid)
1098 {
1099         struct btrfs_key key;
1100         int ret = 0;
1101         struct inode *inode;
1102
1103         inode = read_one_inode(root, objectid);
1104         BUG_ON(!inode);
1105
1106         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1107         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1108         key.offset = objectid;
1109
1110         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1111
1112         btrfs_release_path(root, path);
1113         if (ret == 0) {
1114                 btrfs_inc_nlink(inode);
1115                 btrfs_update_inode(trans, root, inode);
1116         } else if (ret == -EEXIST) {
1117                 ret = 0;
1118         } else {
1119                 BUG();
1120         }
1121         iput(inode);
1122
1123         return ret;
1124 }
1125
1126 /*
1127  * when replaying the log for a directory, we only insert names
1128  * for inodes that actually exist.  This means an fsync on a directory
1129  * does not implicitly fsync all the new files in it
1130  */
1131 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1132                                     struct btrfs_root *root,
1133                                     struct btrfs_path *path,
1134                                     u64 dirid, u64 index,
1135                                     char *name, int name_len, u8 type,
1136                                     struct btrfs_key *location)
1137 {
1138         struct inode *inode;
1139         struct inode *dir;
1140         int ret;
1141
1142         inode = read_one_inode(root, location->objectid);
1143         if (!inode)
1144                 return -ENOENT;
1145
1146         dir = read_one_inode(root, dirid);
1147         if (!dir) {
1148                 iput(inode);
1149                 return -EIO;
1150         }
1151         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1152
1153         /* FIXME, put inode into FIXUP list */
1154
1155         iput(inode);
1156         iput(dir);
1157         return ret;
1158 }
1159
1160 /*
1161  * take a single entry in a log directory item and replay it into
1162  * the subvolume.
1163  *
1164  * if a conflicting item exists in the subdirectory already,
1165  * the inode it points to is unlinked and put into the link count
1166  * fix up tree.
1167  *
1168  * If a name from the log points to a file or directory that does
1169  * not exist in the FS, it is skipped.  fsyncs on directories
1170  * do not force down inodes inside that directory, just changes to the
1171  * names or unlinks in a directory.
1172  */
1173 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1174                                     struct btrfs_root *root,
1175                                     struct btrfs_path *path,
1176                                     struct extent_buffer *eb,
1177                                     struct btrfs_dir_item *di,
1178                                     struct btrfs_key *key)
1179 {
1180         char *name;
1181         int name_len;
1182         struct btrfs_dir_item *dst_di;
1183         struct btrfs_key found_key;
1184         struct btrfs_key log_key;
1185         struct inode *dir;
1186         u8 log_type;
1187         int exists;
1188         int ret;
1189
1190         dir = read_one_inode(root, key->objectid);
1191         BUG_ON(!dir);
1192
1193         name_len = btrfs_dir_name_len(eb, di);
1194         name = kmalloc(name_len, GFP_NOFS);
1195         log_type = btrfs_dir_type(eb, di);
1196         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1197                    name_len);
1198
1199         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1200         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1201         if (exists == 0)
1202                 exists = 1;
1203         else
1204                 exists = 0;
1205         btrfs_release_path(root, path);
1206
1207         if (key->type == BTRFS_DIR_ITEM_KEY) {
1208                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1209                                        name, name_len, 1);
1210         }
1211         else if (key->type == BTRFS_DIR_INDEX_KEY) {
1212                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1213                                                      key->objectid,
1214                                                      key->offset, name,
1215                                                      name_len, 1);
1216         } else {
1217                 BUG();
1218         }
1219         if (!dst_di || IS_ERR(dst_di)) {
1220                 /* we need a sequence number to insert, so we only
1221                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1222                  */
1223                 if (key->type != BTRFS_DIR_INDEX_KEY)
1224                         goto out;
1225                 goto insert;
1226         }
1227
1228         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1229         /* the existing item matches the logged item */
1230         if (found_key.objectid == log_key.objectid &&
1231             found_key.type == log_key.type &&
1232             found_key.offset == log_key.offset &&
1233             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1234                 goto out;
1235         }
1236
1237         /*
1238          * don't drop the conflicting directory entry if the inode
1239          * for the new entry doesn't exist
1240          */
1241         if (!exists)
1242                 goto out;
1243
1244         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1245         BUG_ON(ret);
1246
1247         if (key->type == BTRFS_DIR_INDEX_KEY)
1248                 goto insert;
1249 out:
1250         btrfs_release_path(root, path);
1251         kfree(name);
1252         iput(dir);
1253         return 0;
1254
1255 insert:
1256         btrfs_release_path(root, path);
1257         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1258                               name, name_len, log_type, &log_key);
1259
1260         if (ret && ret != -ENOENT)
1261                 BUG();
1262         goto out;
1263 }
1264
1265 /*
1266  * find all the names in a directory item and reconcile them into
1267  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1268  * one name in a directory item, but the same code gets used for
1269  * both directory index types
1270  */
1271 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1272                                         struct btrfs_root *root,
1273                                         struct btrfs_path *path,
1274                                         struct extent_buffer *eb, int slot,
1275                                         struct btrfs_key *key)
1276 {
1277         int ret;
1278         u32 item_size = btrfs_item_size_nr(eb, slot);
1279         struct btrfs_dir_item *di;
1280         int name_len;
1281         unsigned long ptr;
1282         unsigned long ptr_end;
1283
1284         ptr = btrfs_item_ptr_offset(eb, slot);
1285         ptr_end = ptr + item_size;
1286         while(ptr < ptr_end) {
1287                 di = (struct btrfs_dir_item *)ptr;
1288                 name_len = btrfs_dir_name_len(eb, di);
1289                 ret = replay_one_name(trans, root, path, eb, di, key);
1290                 BUG_ON(ret);
1291                 ptr = (unsigned long)(di + 1);
1292                 ptr += name_len;
1293         }
1294         return 0;
1295 }
1296
1297 /*
1298  * directory replay has two parts.  There are the standard directory
1299  * items in the log copied from the subvolume, and range items
1300  * created in the log while the subvolume was logged.
1301  *
1302  * The range items tell us which parts of the key space the log
1303  * is authoritative for.  During replay, if a key in the subvolume
1304  * directory is in a logged range item, but not actually in the log
1305  * that means it was deleted from the directory before the fsync
1306  * and should be removed.
1307  */
1308 static noinline int find_dir_range(struct btrfs_root *root,
1309                                    struct btrfs_path *path,
1310                                    u64 dirid, int key_type,
1311                                    u64 *start_ret, u64 *end_ret)
1312 {
1313         struct btrfs_key key;
1314         u64 found_end;
1315         struct btrfs_dir_log_item *item;
1316         int ret;
1317         int nritems;
1318
1319         if (*start_ret == (u64)-1)
1320                 return 1;
1321
1322         key.objectid = dirid;
1323         key.type = key_type;
1324         key.offset = *start_ret;
1325
1326         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1327         if (ret < 0)
1328                 goto out;
1329         if (ret > 0) {
1330                 if (path->slots[0] == 0)
1331                         goto out;
1332                 path->slots[0]--;
1333         }
1334         if (ret != 0)
1335                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1336
1337         if (key.type != key_type || key.objectid != dirid) {
1338                 ret = 1;
1339                 goto next;
1340         }
1341         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1342                               struct btrfs_dir_log_item);
1343         found_end = btrfs_dir_log_end(path->nodes[0], item);
1344
1345         if (*start_ret >= key.offset && *start_ret <= found_end) {
1346                 ret = 0;
1347                 *start_ret = key.offset;
1348                 *end_ret = found_end;
1349                 goto out;
1350         }
1351         ret = 1;
1352 next:
1353         /* check the next slot in the tree to see if it is a valid item */
1354         nritems = btrfs_header_nritems(path->nodes[0]);
1355         if (path->slots[0] >= nritems) {
1356                 ret = btrfs_next_leaf(root, path);
1357                 if (ret)
1358                         goto out;
1359         } else {
1360                 path->slots[0]++;
1361         }
1362
1363         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1364
1365         if (key.type != key_type || key.objectid != dirid) {
1366                 ret = 1;
1367                 goto out;
1368         }
1369         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1370                               struct btrfs_dir_log_item);
1371         found_end = btrfs_dir_log_end(path->nodes[0], item);
1372         *start_ret = key.offset;
1373         *end_ret = found_end;
1374         ret = 0;
1375 out:
1376         btrfs_release_path(root, path);
1377         return ret;
1378 }
1379
1380 /*
1381  * this looks for a given directory item in the log.  If the directory
1382  * item is not in the log, the item is removed and the inode it points
1383  * to is unlinked
1384  */
1385 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1386                                       struct btrfs_root *root,
1387                                       struct btrfs_root *log,
1388                                       struct btrfs_path *path,
1389                                       struct btrfs_path *log_path,
1390                                       struct inode *dir,
1391                                       struct btrfs_key *dir_key)
1392 {
1393         int ret;
1394         struct extent_buffer *eb;
1395         int slot;
1396         u32 item_size;
1397         struct btrfs_dir_item *di;
1398         struct btrfs_dir_item *log_di;
1399         int name_len;
1400         unsigned long ptr;
1401         unsigned long ptr_end;
1402         char *name;
1403         struct inode *inode;
1404         struct btrfs_key location;
1405
1406 again:
1407         eb = path->nodes[0];
1408         slot = path->slots[0];
1409         item_size = btrfs_item_size_nr(eb, slot);
1410         ptr = btrfs_item_ptr_offset(eb, slot);
1411         ptr_end = ptr + item_size;
1412         while(ptr < ptr_end) {
1413                 di = (struct btrfs_dir_item *)ptr;
1414                 name_len = btrfs_dir_name_len(eb, di);
1415                 name = kmalloc(name_len, GFP_NOFS);
1416                 if (!name) {
1417                         ret = -ENOMEM;
1418                         goto out;
1419                 }
1420                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1421                                   name_len);
1422                 log_di = NULL;
1423                 if (dir_key->type == BTRFS_DIR_ITEM_KEY) {
1424                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1425                                                        dir_key->objectid,
1426                                                        name, name_len, 0);
1427                 } else if (dir_key->type == BTRFS_DIR_INDEX_KEY) {
1428                         log_di = btrfs_lookup_dir_index_item(trans, log,
1429                                                      log_path,
1430                                                      dir_key->objectid,
1431                                                      dir_key->offset,
1432                                                      name, name_len, 0);
1433                 }
1434                 if (!log_di || IS_ERR(log_di)) {
1435                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1436                         btrfs_release_path(root, path);
1437                         btrfs_release_path(log, log_path);
1438                         inode = read_one_inode(root, location.objectid);
1439                         BUG_ON(!inode);
1440
1441                         ret = link_to_fixup_dir(trans, root,
1442                                                 path, location.objectid);
1443                         BUG_ON(ret);
1444                         btrfs_inc_nlink(inode);
1445                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1446                                                  name, name_len);
1447                         BUG_ON(ret);
1448                         kfree(name);
1449                         iput(inode);
1450
1451                         /* there might still be more names under this key
1452                          * check and repeat if required
1453                          */
1454                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1455                                                 0, 0);
1456                         if (ret == 0)
1457                                 goto again;
1458                         ret = 0;
1459                         goto out;
1460                 }
1461                 btrfs_release_path(log, log_path);
1462                 kfree(name);
1463
1464                 ptr = (unsigned long)(di + 1);
1465                 ptr += name_len;
1466         }
1467         ret = 0;
1468 out:
1469         btrfs_release_path(root, path);
1470         btrfs_release_path(log, log_path);
1471         return ret;
1472 }
1473
1474 /*
1475  * deletion replay happens before we copy any new directory items
1476  * out of the log or out of backreferences from inodes.  It
1477  * scans the log to find ranges of keys that log is authoritative for,
1478  * and then scans the directory to find items in those ranges that are
1479  * not present in the log.
1480  *
1481  * Anything we don't find in the log is unlinked and removed from the
1482  * directory.
1483  */
1484 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1485                                        struct btrfs_root *root,
1486                                        struct btrfs_root *log,
1487                                        struct btrfs_path *path,
1488                                        u64 dirid)
1489 {
1490         u64 range_start;
1491         u64 range_end;
1492         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1493         int ret = 0;
1494         struct btrfs_key dir_key;
1495         struct btrfs_key found_key;
1496         struct btrfs_path *log_path;
1497         struct inode *dir;
1498
1499         dir_key.objectid = dirid;
1500         dir_key.type = BTRFS_DIR_ITEM_KEY;
1501         log_path = btrfs_alloc_path();
1502         if (!log_path)
1503                 return -ENOMEM;
1504
1505         dir = read_one_inode(root, dirid);
1506         /* it isn't an error if the inode isn't there, that can happen
1507          * because we replay the deletes before we copy in the inode item
1508          * from the log
1509          */
1510         if (!dir) {
1511                 btrfs_free_path(log_path);
1512                 return 0;
1513         }
1514 again:
1515         range_start = 0;
1516         range_end = 0;
1517         while(1) {
1518                 ret = find_dir_range(log, path, dirid, key_type,
1519                                      &range_start, &range_end);
1520                 if (ret != 0)
1521                         break;
1522
1523                 dir_key.offset = range_start;
1524                 while(1) {
1525                         int nritems;
1526                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
1527                                                 0, 0);
1528                         if (ret < 0)
1529                                 goto out;
1530
1531                         nritems = btrfs_header_nritems(path->nodes[0]);
1532                         if (path->slots[0] >= nritems) {
1533                                 ret = btrfs_next_leaf(root, path);
1534                                 if (ret)
1535                                         break;
1536                         }
1537                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1538                                               path->slots[0]);
1539                         if (found_key.objectid != dirid ||
1540                             found_key.type != dir_key.type)
1541                                 goto next_type;
1542
1543                         if (found_key.offset > range_end)
1544                                 break;
1545
1546                         ret = check_item_in_log(trans, root, log, path,
1547                                                 log_path, dir, &found_key);
1548                         BUG_ON(ret);
1549                         if (found_key.offset == (u64)-1)
1550                                 break;
1551                         dir_key.offset = found_key.offset + 1;
1552                 }
1553                 btrfs_release_path(root, path);
1554                 if (range_end == (u64)-1)
1555                         break;
1556                 range_start = range_end + 1;
1557         }
1558
1559 next_type:
1560         ret = 0;
1561         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1562                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1563                 dir_key.type = BTRFS_DIR_INDEX_KEY;
1564                 btrfs_release_path(root, path);
1565                 goto again;
1566         }
1567 out:
1568         btrfs_release_path(root, path);
1569         btrfs_free_path(log_path);
1570         iput(dir);
1571         return ret;
1572 }
1573
1574 /*
1575  * the process_func used to replay items from the log tree.  This
1576  * gets called in two different stages.  The first stage just looks
1577  * for inodes and makes sure they are all copied into the subvolume.
1578  *
1579  * The second stage copies all the other item types from the log into
1580  * the subvolume.  The two stage approach is slower, but gets rid of
1581  * lots of complexity around inodes referencing other inodes that exist
1582  * only in the log (references come from either directory items or inode
1583  * back refs).
1584  */
1585 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
1586                              struct walk_control *wc, u64 gen)
1587 {
1588         int nritems;
1589         struct btrfs_path *path;
1590         struct btrfs_root *root = wc->replay_dest;
1591         struct btrfs_key key;
1592         u32 item_size;
1593         int level;
1594         int i;
1595         int ret;
1596
1597         btrfs_read_buffer(eb, gen);
1598
1599         level = btrfs_header_level(eb);
1600
1601         if (level != 0)
1602                 return 0;
1603
1604         path = btrfs_alloc_path();
1605         BUG_ON(!path);
1606
1607         nritems = btrfs_header_nritems(eb);
1608         for (i = 0; i < nritems; i++) {
1609                 btrfs_item_key_to_cpu(eb, &key, i);
1610                 item_size = btrfs_item_size_nr(eb, i);
1611
1612                 /* inode keys are done during the first stage */
1613                 if (key.type == BTRFS_INODE_ITEM_KEY &&
1614                     wc->stage == LOG_WALK_REPLAY_INODES) {
1615                         struct inode *inode;
1616                         struct btrfs_inode_item *inode_item;
1617                         u32 mode;
1618
1619                         inode_item = btrfs_item_ptr(eb, i,
1620                                             struct btrfs_inode_item);
1621                         mode = btrfs_inode_mode(eb, inode_item);
1622                         if (S_ISDIR(mode)) {
1623                                 ret = replay_dir_deletes(wc->trans,
1624                                          root, log, path, key.objectid);
1625                                 BUG_ON(ret);
1626                         }
1627                         ret = overwrite_item(wc->trans, root, path,
1628                                              eb, i, &key);
1629                         BUG_ON(ret);
1630
1631                         /* for regular files, truncate away
1632                          * extents past the new EOF
1633                          */
1634                         if (S_ISREG(mode)) {
1635                                 inode = read_one_inode(root,
1636                                                        key.objectid);
1637                                 BUG_ON(!inode);
1638
1639                                 ret = btrfs_truncate_inode_items(wc->trans,
1640                                         root, inode, inode->i_size,
1641                                         BTRFS_EXTENT_DATA_KEY);
1642                                 BUG_ON(ret);
1643                                 iput(inode);
1644                         }
1645                         ret = link_to_fixup_dir(wc->trans, root,
1646                                                 path, key.objectid);
1647                         BUG_ON(ret);
1648                 }
1649                 if (wc->stage < LOG_WALK_REPLAY_ALL)
1650                         continue;
1651
1652                 /* these keys are simply copied */
1653                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
1654                         ret = overwrite_item(wc->trans, root, path,
1655                                              eb, i, &key);
1656                         BUG_ON(ret);
1657                 } else if (key.type == BTRFS_INODE_REF_KEY) {
1658                         ret = add_inode_ref(wc->trans, root, log, path,
1659                                             eb, i, &key);
1660                         BUG_ON(ret && ret != -ENOENT);
1661                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
1662                         ret = replay_one_extent(wc->trans, root, path,
1663                                                 eb, i, &key);
1664                         BUG_ON(ret);
1665                 } else if (key.type == BTRFS_CSUM_ITEM_KEY) {
1666                         ret = replay_one_csum(wc->trans, root, path,
1667                                               eb, i, &key);
1668                         BUG_ON(ret);
1669                 } else if (key.type == BTRFS_DIR_ITEM_KEY ||
1670                            key.type == BTRFS_DIR_INDEX_KEY) {
1671                         ret = replay_one_dir_item(wc->trans, root, path,
1672                                                   eb, i, &key);
1673                         BUG_ON(ret);
1674                 }
1675         }
1676         btrfs_free_path(path);
1677         return 0;
1678 }
1679
1680 static int noinline walk_down_log_tree(struct btrfs_trans_handle *trans,
1681                                    struct btrfs_root *root,
1682                                    struct btrfs_path *path, int *level,
1683                                    struct walk_control *wc)
1684 {
1685         u64 root_owner;
1686         u64 root_gen;
1687         u64 bytenr;
1688         u64 ptr_gen;
1689         struct extent_buffer *next;
1690         struct extent_buffer *cur;
1691         struct extent_buffer *parent;
1692         u32 blocksize;
1693         int ret = 0;
1694
1695         WARN_ON(*level < 0);
1696         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1697
1698         while(*level > 0) {
1699                 WARN_ON(*level < 0);
1700                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1701                 cur = path->nodes[*level];
1702
1703                 if (btrfs_header_level(cur) != *level)
1704                         WARN_ON(1);
1705
1706                 if (path->slots[*level] >=
1707                     btrfs_header_nritems(cur))
1708                         break;
1709
1710                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1711                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1712                 blocksize = btrfs_level_size(root, *level - 1);
1713
1714                 parent = path->nodes[*level];
1715                 root_owner = btrfs_header_owner(parent);
1716                 root_gen = btrfs_header_generation(parent);
1717
1718                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
1719
1720                 wc->process_func(root, next, wc, ptr_gen);
1721
1722                 if (*level == 1) {
1723                         path->slots[*level]++;
1724                         if (wc->free) {
1725                                 btrfs_read_buffer(next, ptr_gen);
1726
1727                                 btrfs_tree_lock(next);
1728                                 clean_tree_block(trans, root, next);
1729                                 btrfs_wait_tree_block_writeback(next);
1730                                 btrfs_tree_unlock(next);
1731
1732                                 ret = btrfs_drop_leaf_ref(trans, root, next);
1733                                 BUG_ON(ret);
1734
1735                                 WARN_ON(root_owner !=
1736                                         BTRFS_TREE_LOG_OBJECTID);
1737                                 ret = btrfs_free_reserved_extent(root,
1738                                                          bytenr, blocksize);
1739                                 BUG_ON(ret);
1740                         }
1741                         free_extent_buffer(next);
1742                         continue;
1743                 }
1744                 btrfs_read_buffer(next, ptr_gen);
1745
1746                 WARN_ON(*level <= 0);
1747                 if (path->nodes[*level-1])
1748                         free_extent_buffer(path->nodes[*level-1]);
1749                 path->nodes[*level-1] = next;
1750                 *level = btrfs_header_level(next);
1751                 path->slots[*level] = 0;
1752                 cond_resched();
1753         }
1754         WARN_ON(*level < 0);
1755         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1756
1757         if (path->nodes[*level] == root->node) {
1758                 parent = path->nodes[*level];
1759         } else {
1760                 parent = path->nodes[*level + 1];
1761         }
1762         bytenr = path->nodes[*level]->start;
1763
1764         blocksize = btrfs_level_size(root, *level);
1765         root_owner = btrfs_header_owner(parent);
1766         root_gen = btrfs_header_generation(parent);
1767
1768         wc->process_func(root, path->nodes[*level], wc,
1769                          btrfs_header_generation(path->nodes[*level]));
1770
1771         if (wc->free) {
1772                 next = path->nodes[*level];
1773                 btrfs_tree_lock(next);
1774                 clean_tree_block(trans, root, next);
1775                 btrfs_wait_tree_block_writeback(next);
1776                 btrfs_tree_unlock(next);
1777
1778                 if (*level == 0) {
1779                         ret = btrfs_drop_leaf_ref(trans, root, next);
1780                         BUG_ON(ret);
1781                 }
1782                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1783                 ret = btrfs_free_reserved_extent(root, bytenr, blocksize);
1784                 BUG_ON(ret);
1785         }
1786         free_extent_buffer(path->nodes[*level]);
1787         path->nodes[*level] = NULL;
1788         *level += 1;
1789
1790         cond_resched();
1791         return 0;
1792 }
1793
1794 static int noinline walk_up_log_tree(struct btrfs_trans_handle *trans,
1795                                  struct btrfs_root *root,
1796                                  struct btrfs_path *path, int *level,
1797                                  struct walk_control *wc)
1798 {
1799         u64 root_owner;
1800         u64 root_gen;
1801         int i;
1802         int slot;
1803         int ret;
1804
1805         for(i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1806                 slot = path->slots[i];
1807                 if (slot < btrfs_header_nritems(path->nodes[i]) - 1) {
1808                         struct extent_buffer *node;
1809                         node = path->nodes[i];
1810                         path->slots[i]++;
1811                         *level = i;
1812                         WARN_ON(*level == 0);
1813                         return 0;
1814                 } else {
1815                         struct extent_buffer *parent;
1816                         if (path->nodes[*level] == root->node)
1817                                 parent = path->nodes[*level];
1818                         else
1819                                 parent = path->nodes[*level + 1];
1820
1821                         root_owner = btrfs_header_owner(parent);
1822                         root_gen = btrfs_header_generation(parent);
1823                         wc->process_func(root, path->nodes[*level], wc,
1824                                  btrfs_header_generation(path->nodes[*level]));
1825                         if (wc->free) {
1826                                 struct extent_buffer *next;
1827
1828                                 next = path->nodes[*level];
1829
1830                                 btrfs_tree_lock(next);
1831                                 clean_tree_block(trans, root, next);
1832                                 btrfs_wait_tree_block_writeback(next);
1833                                 btrfs_tree_unlock(next);
1834
1835                                 if (*level == 0) {
1836                                         ret = btrfs_drop_leaf_ref(trans, root,
1837                                                                   next);
1838                                         BUG_ON(ret);
1839                                 }
1840
1841                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1842                                 ret = btrfs_free_reserved_extent(root,
1843                                                 path->nodes[*level]->start,
1844                                                 path->nodes[*level]->len);
1845                                 BUG_ON(ret);
1846                         }
1847                         free_extent_buffer(path->nodes[*level]);
1848                         path->nodes[*level] = NULL;
1849                         *level = i + 1;
1850                 }
1851         }
1852         return 1;
1853 }
1854
1855 /*
1856  * drop the reference count on the tree rooted at 'snap'.  This traverses
1857  * the tree freeing any blocks that have a ref count of zero after being
1858  * decremented.
1859  */
1860 static int walk_log_tree(struct btrfs_trans_handle *trans,
1861                          struct btrfs_root *log, struct walk_control *wc)
1862 {
1863         int ret = 0;
1864         int wret;
1865         int level;
1866         struct btrfs_path *path;
1867         int i;
1868         int orig_level;
1869
1870         path = btrfs_alloc_path();
1871         BUG_ON(!path);
1872
1873         level = btrfs_header_level(log->node);
1874         orig_level = level;
1875         path->nodes[level] = log->node;
1876         extent_buffer_get(log->node);
1877         path->slots[level] = 0;
1878
1879         while(1) {
1880                 wret = walk_down_log_tree(trans, log, path, &level, wc);
1881                 if (wret > 0)
1882                         break;
1883                 if (wret < 0)
1884                         ret = wret;
1885
1886                 wret = walk_up_log_tree(trans, log, path, &level, wc);
1887                 if (wret > 0)
1888                         break;
1889                 if (wret < 0)
1890                         ret = wret;
1891         }
1892
1893         /* was the root node processed? if not, catch it here */
1894         if (path->nodes[orig_level]) {
1895                 wc->process_func(log, path->nodes[orig_level], wc,
1896                          btrfs_header_generation(path->nodes[orig_level]));
1897                 if (wc->free) {
1898                         struct extent_buffer *next;
1899
1900                         next = path->nodes[orig_level];
1901
1902                         btrfs_tree_lock(next);
1903                         clean_tree_block(trans, log, next);
1904                         btrfs_wait_tree_block_writeback(next);
1905                         btrfs_tree_unlock(next);
1906
1907                         if (orig_level == 0) {
1908                                 ret = btrfs_drop_leaf_ref(trans, log,
1909                                                           next);
1910                                 BUG_ON(ret);
1911                         }
1912                         WARN_ON(log->root_key.objectid !=
1913                                 BTRFS_TREE_LOG_OBJECTID);
1914                         ret = btrfs_free_reserved_extent(log, next->start,
1915                                                          next->len);
1916                         BUG_ON(ret);
1917                 }
1918         }
1919
1920         for (i = 0; i <= orig_level; i++) {
1921                 if (path->nodes[i]) {
1922                         free_extent_buffer(path->nodes[i]);
1923                         path->nodes[i] = NULL;
1924                 }
1925         }
1926         btrfs_free_path(path);
1927         if (wc->free)
1928                 free_extent_buffer(log->node);
1929         return ret;
1930 }
1931
1932 int wait_log_commit(struct btrfs_root *log)
1933 {
1934         DEFINE_WAIT(wait);
1935         u64 transid = log->fs_info->tree_log_transid;
1936
1937         do {
1938                 prepare_to_wait(&log->fs_info->tree_log_wait, &wait,
1939                                 TASK_UNINTERRUPTIBLE);
1940                 mutex_unlock(&log->fs_info->tree_log_mutex);
1941                 if (atomic_read(&log->fs_info->tree_log_commit))
1942                         schedule();
1943                 finish_wait(&log->fs_info->tree_log_wait, &wait);
1944                 mutex_lock(&log->fs_info->tree_log_mutex);
1945         } while(transid == log->fs_info->tree_log_transid &&
1946                 atomic_read(&log->fs_info->tree_log_commit));
1947         return 0;
1948 }
1949
1950 /*
1951  * btrfs_sync_log does sends a given tree log down to the disk and
1952  * updates the super blocks to record it.  When this call is done,
1953  * you know that any inodes previously logged are safely on disk
1954  */
1955 int btrfs_sync_log(struct btrfs_trans_handle *trans,
1956                    struct btrfs_root *root)
1957 {
1958         int ret;
1959         unsigned long batch;
1960         struct btrfs_root *log = root->log_root;
1961
1962         mutex_lock(&log->fs_info->tree_log_mutex);
1963         if (atomic_read(&log->fs_info->tree_log_commit)) {
1964                 wait_log_commit(log);
1965                 goto out;
1966         }
1967         atomic_set(&log->fs_info->tree_log_commit, 1);
1968
1969         while(1) {
1970                 batch = log->fs_info->tree_log_batch;
1971                 mutex_unlock(&log->fs_info->tree_log_mutex);
1972                 schedule_timeout_uninterruptible(1);
1973                 mutex_lock(&log->fs_info->tree_log_mutex);
1974
1975                 while(atomic_read(&log->fs_info->tree_log_writers)) {
1976                         DEFINE_WAIT(wait);
1977                         prepare_to_wait(&log->fs_info->tree_log_wait, &wait,
1978                                         TASK_UNINTERRUPTIBLE);
1979                         mutex_unlock(&log->fs_info->tree_log_mutex);
1980                         if (atomic_read(&log->fs_info->tree_log_writers))
1981                                 schedule();
1982                         mutex_lock(&log->fs_info->tree_log_mutex);
1983                         finish_wait(&log->fs_info->tree_log_wait, &wait);
1984                 }
1985                 if (batch == log->fs_info->tree_log_batch)
1986                         break;
1987         }
1988
1989         ret = btrfs_write_and_wait_marked_extents(log, &log->dirty_log_pages);
1990         BUG_ON(ret);
1991         ret = btrfs_write_and_wait_marked_extents(root->fs_info->log_root_tree,
1992                                &root->fs_info->log_root_tree->dirty_log_pages);
1993         BUG_ON(ret);
1994
1995         btrfs_set_super_log_root(&root->fs_info->super_for_commit,
1996                                  log->fs_info->log_root_tree->node->start);
1997         btrfs_set_super_log_root_level(&root->fs_info->super_for_commit,
1998                        btrfs_header_level(log->fs_info->log_root_tree->node));
1999
2000         write_ctree_super(trans, log->fs_info->tree_root);
2001         log->fs_info->tree_log_transid++;
2002         log->fs_info->tree_log_batch = 0;
2003         atomic_set(&log->fs_info->tree_log_commit, 0);
2004         smp_mb();
2005         if (waitqueue_active(&log->fs_info->tree_log_wait))
2006                 wake_up(&log->fs_info->tree_log_wait);
2007 out:
2008         mutex_unlock(&log->fs_info->tree_log_mutex);
2009         return 0;
2010
2011 }
2012
2013 /* * free all the extents used by the tree log.  This should be called
2014  * at commit time of the full transaction
2015  */
2016 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2017 {
2018         int ret;
2019         struct btrfs_root *log;
2020         struct key;
2021         u64 start;
2022         u64 end;
2023         struct walk_control wc = {
2024                 .free = 1,
2025                 .process_func = process_one_buffer
2026         };
2027
2028         if (!root->log_root)
2029                 return 0;
2030
2031         log = root->log_root;
2032         ret = walk_log_tree(trans, log, &wc);
2033         BUG_ON(ret);
2034
2035         while(1) {
2036                 ret = find_first_extent_bit(&log->dirty_log_pages,
2037                                     0, &start, &end, EXTENT_DIRTY);
2038                 if (ret)
2039                         break;
2040
2041                 clear_extent_dirty(&log->dirty_log_pages,
2042                                    start, end, GFP_NOFS);
2043         }
2044
2045         log = root->log_root;
2046         ret = btrfs_del_root(trans, root->fs_info->log_root_tree,
2047                              &log->root_key);
2048         BUG_ON(ret);
2049         root->log_root = NULL;
2050         kfree(root->log_root);
2051         return 0;
2052 }
2053
2054 /*
2055  * helper function to update the item for a given subvolumes log root
2056  * in the tree of log roots
2057  */
2058 static int update_log_root(struct btrfs_trans_handle *trans,
2059                            struct btrfs_root *log)
2060 {
2061         u64 bytenr = btrfs_root_bytenr(&log->root_item);
2062         int ret;
2063
2064         if (log->node->start == bytenr)
2065                 return 0;
2066
2067         btrfs_set_root_bytenr(&log->root_item, log->node->start);
2068         btrfs_set_root_level(&log->root_item, btrfs_header_level(log->node));
2069         ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2070                                 &log->root_key, &log->root_item);
2071         BUG_ON(ret);
2072         return ret;
2073 }
2074
2075 /*
2076  * If both a file and directory are logged, and unlinks or renames are
2077  * mixed in, we have a few interesting corners:
2078  *
2079  * create file X in dir Y
2080  * link file X to X.link in dir Y
2081  * fsync file X
2082  * unlink file X but leave X.link
2083  * fsync dir Y
2084  *
2085  * After a crash we would expect only X.link to exist.  But file X
2086  * didn't get fsync'd again so the log has back refs for X and X.link.
2087  *
2088  * We solve this by removing directory entries and inode backrefs from the
2089  * log when a file that was logged in the current transaction is
2090  * unlinked.  Any later fsync will include the updated log entries, and
2091  * we'll be able to reconstruct the proper directory items from backrefs.
2092  *
2093  * This optimizations allows us to avoid relogging the entire inode
2094  * or the entire directory.
2095  */
2096 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2097                                  struct btrfs_root *root,
2098                                  const char *name, int name_len,
2099                                  struct inode *dir, u64 index)
2100 {
2101         struct btrfs_root *log;
2102         struct btrfs_dir_item *di;
2103         struct btrfs_path *path;
2104         int ret;
2105         int bytes_del = 0;
2106
2107         if (BTRFS_I(dir)->logged_trans < trans->transid)
2108                 return 0;
2109
2110         ret = join_running_log_trans(root);
2111         if (ret)
2112                 return 0;
2113
2114         mutex_lock(&BTRFS_I(dir)->log_mutex);
2115
2116         log = root->log_root;
2117         path = btrfs_alloc_path();
2118         di = btrfs_lookup_dir_item(trans, log, path, dir->i_ino,
2119                                    name, name_len, -1);
2120         if (di && !IS_ERR(di)) {
2121                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2122                 bytes_del += name_len;
2123                 BUG_ON(ret);
2124         }
2125         btrfs_release_path(log, path);
2126         di = btrfs_lookup_dir_index_item(trans, log, path, dir->i_ino,
2127                                          index, name, name_len, -1);
2128         if (di && !IS_ERR(di)) {
2129                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2130                 bytes_del += name_len;
2131                 BUG_ON(ret);
2132         }
2133
2134         /* update the directory size in the log to reflect the names
2135          * we have removed
2136          */
2137         if (bytes_del) {
2138                 struct btrfs_key key;
2139
2140                 key.objectid = dir->i_ino;
2141                 key.offset = 0;
2142                 key.type = BTRFS_INODE_ITEM_KEY;
2143                 btrfs_release_path(log, path);
2144
2145                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2146                 if (ret == 0) {
2147                         struct btrfs_inode_item *item;
2148                         u64 i_size;
2149
2150                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2151                                               struct btrfs_inode_item);
2152                         i_size = btrfs_inode_size(path->nodes[0], item);
2153                         if (i_size > bytes_del)
2154                                 i_size -= bytes_del;
2155                         else
2156                                 i_size = 0;
2157                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2158                         btrfs_mark_buffer_dirty(path->nodes[0]);
2159                 } else
2160                         ret = 0;
2161                 btrfs_release_path(log, path);
2162         }
2163
2164         btrfs_free_path(path);
2165         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2166         end_log_trans(root);
2167
2168         return 0;
2169 }
2170
2171 /* see comments for btrfs_del_dir_entries_in_log */
2172 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2173                                struct btrfs_root *root,
2174                                const char *name, int name_len,
2175                                struct inode *inode, u64 dirid)
2176 {
2177         struct btrfs_root *log;
2178         u64 index;
2179         int ret;
2180
2181         if (BTRFS_I(inode)->logged_trans < trans->transid)
2182                 return 0;
2183
2184         ret = join_running_log_trans(root);
2185         if (ret)
2186                 return 0;
2187         log = root->log_root;
2188         mutex_lock(&BTRFS_I(inode)->log_mutex);
2189
2190         ret = btrfs_del_inode_ref(trans, log, name, name_len, inode->i_ino,
2191                                   dirid, &index);
2192         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2193         end_log_trans(root);
2194
2195         return ret;
2196 }
2197
2198 /*
2199  * creates a range item in the log for 'dirid'.  first_offset and
2200  * last_offset tell us which parts of the key space the log should
2201  * be considered authoritative for.
2202  */
2203 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2204                                        struct btrfs_root *log,
2205                                        struct btrfs_path *path,
2206                                        int key_type, u64 dirid,
2207                                        u64 first_offset, u64 last_offset)
2208 {
2209         int ret;
2210         struct btrfs_key key;
2211         struct btrfs_dir_log_item *item;
2212
2213         key.objectid = dirid;
2214         key.offset = first_offset;
2215         if (key_type == BTRFS_DIR_ITEM_KEY)
2216                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2217         else
2218                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2219         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2220         BUG_ON(ret);
2221
2222         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2223                               struct btrfs_dir_log_item);
2224         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2225         btrfs_mark_buffer_dirty(path->nodes[0]);
2226         btrfs_release_path(log, path);
2227         return 0;
2228 }
2229
2230 /*
2231  * log all the items included in the current transaction for a given
2232  * directory.  This also creates the range items in the log tree required
2233  * to replay anything deleted before the fsync
2234  */
2235 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2236                           struct btrfs_root *root, struct inode *inode,
2237                           struct btrfs_path *path,
2238                           struct btrfs_path *dst_path, int key_type,
2239                           u64 min_offset, u64 *last_offset_ret)
2240 {
2241         struct btrfs_key min_key;
2242         struct btrfs_key max_key;
2243         struct btrfs_root *log = root->log_root;
2244         struct extent_buffer *src;
2245         int ret;
2246         int i;
2247         int nritems;
2248         u64 first_offset = min_offset;
2249         u64 last_offset = (u64)-1;
2250
2251         log = root->log_root;
2252         max_key.objectid = inode->i_ino;
2253         max_key.offset = (u64)-1;
2254         max_key.type = key_type;
2255
2256         min_key.objectid = inode->i_ino;
2257         min_key.type = key_type;
2258         min_key.offset = min_offset;
2259
2260         path->keep_locks = 1;
2261
2262         ret = btrfs_search_forward(root, &min_key, &max_key,
2263                                    path, 0, trans->transid);
2264
2265         /*
2266          * we didn't find anything from this transaction, see if there
2267          * is anything at all
2268          */
2269         if (ret != 0 || min_key.objectid != inode->i_ino ||
2270             min_key.type != key_type) {
2271                 min_key.objectid = inode->i_ino;
2272                 min_key.type = key_type;
2273                 min_key.offset = (u64)-1;
2274                 btrfs_release_path(root, path);
2275                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2276                 if (ret < 0) {
2277                         btrfs_release_path(root, path);
2278                         return ret;
2279                 }
2280                 ret = btrfs_previous_item(root, path, inode->i_ino, key_type);
2281
2282                 /* if ret == 0 there are items for this type,
2283                  * create a range to tell us the last key of this type.
2284                  * otherwise, there are no items in this directory after
2285                  * *min_offset, and we create a range to indicate that.
2286                  */
2287                 if (ret == 0) {
2288                         struct btrfs_key tmp;
2289                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2290                                               path->slots[0]);
2291                         if (key_type == tmp.type) {
2292                                 first_offset = max(min_offset, tmp.offset) + 1;
2293                         }
2294                 }
2295                 goto done;
2296         }
2297
2298         /* go backward to find any previous key */
2299         ret = btrfs_previous_item(root, path, inode->i_ino, key_type);
2300         if (ret == 0) {
2301                 struct btrfs_key tmp;
2302                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2303                 if (key_type == tmp.type) {
2304                         first_offset = tmp.offset;
2305                         ret = overwrite_item(trans, log, dst_path,
2306                                              path->nodes[0], path->slots[0],
2307                                              &tmp);
2308                 }
2309         }
2310         btrfs_release_path(root, path);
2311
2312         /* find the first key from this transaction again */
2313         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2314         if (ret != 0) {
2315                 WARN_ON(1);
2316                 goto done;
2317         }
2318
2319         /*
2320          * we have a block from this transaction, log every item in it
2321          * from our directory
2322          */
2323         while(1) {
2324                 struct btrfs_key tmp;
2325                 src = path->nodes[0];
2326                 nritems = btrfs_header_nritems(src);
2327                 for (i = path->slots[0]; i < nritems; i++) {
2328                         btrfs_item_key_to_cpu(src, &min_key, i);
2329
2330                         if (min_key.objectid != inode->i_ino ||
2331                             min_key.type != key_type)
2332                                 goto done;
2333                         ret = overwrite_item(trans, log, dst_path, src, i,
2334                                              &min_key);
2335                         BUG_ON(ret);
2336                 }
2337                 path->slots[0] = nritems;
2338
2339                 /*
2340                  * look ahead to the next item and see if it is also
2341                  * from this directory and from this transaction
2342                  */
2343                 ret = btrfs_next_leaf(root, path);
2344                 if (ret == 1) {
2345                         last_offset = (u64)-1;
2346                         goto done;
2347                 }
2348                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2349                 if (tmp.objectid != inode->i_ino || tmp.type != key_type) {
2350                         last_offset = (u64)-1;
2351                         goto done;
2352                 }
2353                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2354                         ret = overwrite_item(trans, log, dst_path,
2355                                              path->nodes[0], path->slots[0],
2356                                              &tmp);
2357
2358                         BUG_ON(ret);
2359                         last_offset = tmp.offset;
2360                         goto done;
2361                 }
2362         }
2363 done:
2364         *last_offset_ret = last_offset;
2365         btrfs_release_path(root, path);
2366         btrfs_release_path(log, dst_path);
2367
2368         /* insert the log range keys to indicate where the log is valid */
2369         ret = insert_dir_log_key(trans, log, path, key_type, inode->i_ino,
2370                                  first_offset, last_offset);
2371         BUG_ON(ret);
2372         return 0;
2373 }
2374
2375 /*
2376  * logging directories is very similar to logging inodes, We find all the items
2377  * from the current transaction and write them to the log.
2378  *
2379  * The recovery code scans the directory in the subvolume, and if it finds a
2380  * key in the range logged that is not present in the log tree, then it means
2381  * that dir entry was unlinked during the transaction.
2382  *
2383  * In order for that scan to work, we must include one key smaller than
2384  * the smallest logged by this transaction and one key larger than the largest
2385  * key logged by this transaction.
2386  */
2387 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
2388                           struct btrfs_root *root, struct inode *inode,
2389                           struct btrfs_path *path,
2390                           struct btrfs_path *dst_path)
2391 {
2392         u64 min_key;
2393         u64 max_key;
2394         int ret;
2395         int key_type = BTRFS_DIR_ITEM_KEY;
2396
2397 again:
2398         min_key = 0;
2399         max_key = 0;
2400         while(1) {
2401                 ret = log_dir_items(trans, root, inode, path,
2402                                     dst_path, key_type, min_key,
2403                                     &max_key);
2404                 BUG_ON(ret);
2405                 if (max_key == (u64)-1)
2406                         break;
2407                 min_key = max_key + 1;
2408         }
2409
2410         if (key_type == BTRFS_DIR_ITEM_KEY) {
2411                 key_type = BTRFS_DIR_INDEX_KEY;
2412                 goto again;
2413         }
2414         return 0;
2415 }
2416
2417 /*
2418  * a helper function to drop items from the log before we relog an
2419  * inode.  max_key_type indicates the highest item type to remove.
2420  * This cannot be run for file data extents because it does not
2421  * free the extents they point to.
2422  */
2423 static int drop_objectid_items(struct btrfs_trans_handle *trans,
2424                                   struct btrfs_root *log,
2425                                   struct btrfs_path *path,
2426                                   u64 objectid, int max_key_type)
2427 {
2428         int ret;
2429         struct btrfs_key key;
2430         struct btrfs_key found_key;
2431
2432         key.objectid = objectid;
2433         key.type = max_key_type;
2434         key.offset = (u64)-1;
2435
2436         while(1) {
2437                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
2438
2439                 if (ret != 1)
2440                         break;
2441
2442                 if (path->slots[0] == 0)
2443                         break;
2444
2445                 path->slots[0]--;
2446                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2447                                       path->slots[0]);
2448
2449                 if (found_key.objectid != objectid)
2450                         break;
2451
2452                 ret = btrfs_del_item(trans, log, path);
2453                 BUG_ON(ret);
2454                 btrfs_release_path(log, path);
2455         }
2456         btrfs_release_path(log, path);
2457         return 0;
2458 }
2459
2460 static noinline int copy_items(struct btrfs_trans_handle *trans,
2461                                struct btrfs_root *log,
2462                                struct btrfs_path *dst_path,
2463                                struct extent_buffer *src,
2464                                int start_slot, int nr, int inode_only)
2465 {
2466         unsigned long src_offset;
2467         unsigned long dst_offset;
2468         struct btrfs_file_extent_item *extent;
2469         struct btrfs_inode_item *inode_item;
2470         int ret;
2471         struct btrfs_key *ins_keys;
2472         u32 *ins_sizes;
2473         char *ins_data;
2474         int i;
2475
2476         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
2477                            nr * sizeof(u32), GFP_NOFS);
2478         ins_sizes = (u32 *)ins_data;
2479         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
2480
2481         for (i = 0; i < nr; i++) {
2482                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
2483                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
2484         }
2485         ret = btrfs_insert_empty_items(trans, log, dst_path,
2486                                        ins_keys, ins_sizes, nr);
2487         BUG_ON(ret);
2488
2489         for (i = 0; i < nr; i++) {
2490                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
2491                                                    dst_path->slots[0]);
2492
2493                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
2494
2495                 copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
2496                                    src_offset, ins_sizes[i]);
2497
2498                 if (inode_only == LOG_INODE_EXISTS &&
2499                     ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
2500                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
2501                                                     dst_path->slots[0],
2502                                                     struct btrfs_inode_item);
2503                         btrfs_set_inode_size(dst_path->nodes[0], inode_item, 0);
2504
2505                         /* set the generation to zero so the recover code
2506                          * can tell the difference between an logging
2507                          * just to say 'this inode exists' and a logging
2508                          * to say 'update this inode with these values'
2509                          */
2510                         btrfs_set_inode_generation(dst_path->nodes[0],
2511                                                    inode_item, 0);
2512                 }
2513                 /* take a reference on file data extents so that truncates
2514                  * or deletes of this inode don't have to relog the inode
2515                  * again
2516                  */
2517                 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY) {
2518                         int found_type;
2519                         extent = btrfs_item_ptr(src, start_slot + i,
2520                                                 struct btrfs_file_extent_item);
2521
2522                         found_type = btrfs_file_extent_type(src, extent);
2523                         if (found_type == BTRFS_FILE_EXTENT_REG) {
2524                                 u64 ds = btrfs_file_extent_disk_bytenr(src,
2525                                                                    extent);
2526                                 u64 dl = btrfs_file_extent_disk_num_bytes(src,
2527                                                                       extent);
2528                                 /* ds == 0 is a hole */
2529                                 if (ds != 0) {
2530                                         ret = btrfs_inc_extent_ref(trans, log,
2531                                                    ds, dl,
2532                                                    dst_path->nodes[0]->start,
2533                                                    BTRFS_TREE_LOG_OBJECTID,
2534                                                    trans->transid,
2535                                                    ins_keys[i].objectid);
2536                                         BUG_ON(ret);
2537                                 }
2538                         }
2539                 }
2540                 dst_path->slots[0]++;
2541         }
2542
2543         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
2544         btrfs_release_path(log, dst_path);
2545         kfree(ins_data);
2546         return 0;
2547 }
2548
2549 /* log a single inode in the tree log.
2550  * At least one parent directory for this inode must exist in the tree
2551  * or be logged already.
2552  *
2553  * Any items from this inode changed by the current transaction are copied
2554  * to the log tree.  An extra reference is taken on any extents in this
2555  * file, allowing us to avoid a whole pile of corner cases around logging
2556  * blocks that have been removed from the tree.
2557  *
2558  * See LOG_INODE_ALL and related defines for a description of what inode_only
2559  * does.
2560  *
2561  * This handles both files and directories.
2562  */
2563 static int __btrfs_log_inode(struct btrfs_trans_handle *trans,
2564                              struct btrfs_root *root, struct inode *inode,
2565                              int inode_only)
2566 {
2567         struct btrfs_path *path;
2568         struct btrfs_path *dst_path;
2569         struct btrfs_key min_key;
2570         struct btrfs_key max_key;
2571         struct btrfs_root *log = root->log_root;
2572         struct extent_buffer *src = NULL;
2573         u32 size;
2574         int ret;
2575         int nritems;
2576         int ins_start_slot = 0;
2577         int ins_nr;
2578
2579         log = root->log_root;
2580
2581         path = btrfs_alloc_path();
2582         dst_path = btrfs_alloc_path();
2583
2584         min_key.objectid = inode->i_ino;
2585         min_key.type = BTRFS_INODE_ITEM_KEY;
2586         min_key.offset = 0;
2587
2588         max_key.objectid = inode->i_ino;
2589         if (inode_only == LOG_INODE_EXISTS || S_ISDIR(inode->i_mode))
2590                 max_key.type = BTRFS_XATTR_ITEM_KEY;
2591         else
2592                 max_key.type = (u8)-1;
2593         max_key.offset = (u64)-1;
2594
2595         /*
2596          * if this inode has already been logged and we're in inode_only
2597          * mode, we don't want to delete the things that have already
2598          * been written to the log.
2599          *
2600          * But, if the inode has been through an inode_only log,
2601          * the logged_trans field is not set.  This allows us to catch
2602          * any new names for this inode in the backrefs by logging it
2603          * again
2604          */
2605         if (inode_only == LOG_INODE_EXISTS &&
2606             BTRFS_I(inode)->logged_trans == trans->transid) {
2607                 btrfs_free_path(path);
2608                 btrfs_free_path(dst_path);
2609                 goto out;
2610         }
2611         mutex_lock(&BTRFS_I(inode)->log_mutex);
2612
2613         /*
2614          * a brute force approach to making sure we get the most uptodate
2615          * copies of everything.
2616          */
2617         if (S_ISDIR(inode->i_mode)) {
2618                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
2619
2620                 if (inode_only == LOG_INODE_EXISTS)
2621                         max_key_type = BTRFS_XATTR_ITEM_KEY;
2622                 ret = drop_objectid_items(trans, log, path,
2623                                           inode->i_ino, max_key_type);
2624         } else {
2625                 ret = btrfs_truncate_inode_items(trans, log, inode, 0, 0);
2626         }
2627         BUG_ON(ret);
2628         path->keep_locks = 1;
2629
2630         while(1) {
2631                 ins_nr = 0;
2632                 ret = btrfs_search_forward(root, &min_key, &max_key,
2633                                            path, 0, trans->transid);
2634                 if (ret != 0)
2635                         break;
2636 again:
2637                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
2638                 if (min_key.objectid != inode->i_ino)
2639                         break;
2640                 if (min_key.type > max_key.type)
2641                         break;
2642
2643                 src = path->nodes[0];
2644                 size = btrfs_item_size_nr(src, path->slots[0]);
2645                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
2646                         ins_nr++;
2647                         goto next_slot;
2648                 } else if (!ins_nr) {
2649                         ins_start_slot = path->slots[0];
2650                         ins_nr = 1;
2651                         goto next_slot;
2652                 }
2653
2654                 ret = copy_items(trans, log, dst_path, src, ins_start_slot,
2655                                  ins_nr, inode_only);
2656                 BUG_ON(ret);
2657                 ins_nr = 1;
2658                 ins_start_slot = path->slots[0];
2659 next_slot:
2660
2661                 nritems = btrfs_header_nritems(path->nodes[0]);
2662                 path->slots[0]++;
2663                 if (path->slots[0] < nritems) {
2664                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
2665                                               path->slots[0]);
2666                         goto again;
2667                 }
2668                 if (ins_nr) {
2669                         ret = copy_items(trans, log, dst_path, src,
2670                                          ins_start_slot,
2671                                          ins_nr, inode_only);
2672                         BUG_ON(ret);
2673                         ins_nr = 0;
2674                 }
2675                 btrfs_release_path(root, path);
2676
2677                 if (min_key.offset < (u64)-1)
2678                         min_key.offset++;
2679                 else if (min_key.type < (u8)-1)
2680                         min_key.type++;
2681                 else if (min_key.objectid < (u64)-1)
2682                         min_key.objectid++;
2683                 else
2684                         break;
2685         }
2686         if (ins_nr) {
2687                 ret = copy_items(trans, log, dst_path, src,
2688                                  ins_start_slot,
2689                                  ins_nr, inode_only);
2690                 BUG_ON(ret);
2691                 ins_nr = 0;
2692         }
2693         WARN_ON(ins_nr);
2694         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
2695                 btrfs_release_path(root, path);
2696                 btrfs_release_path(log, dst_path);
2697                 BTRFS_I(inode)->log_dirty_trans = 0;
2698                 ret = log_directory_changes(trans, root, inode, path, dst_path);
2699                 BUG_ON(ret);
2700         }
2701         BTRFS_I(inode)->logged_trans = trans->transid;
2702         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2703
2704         btrfs_free_path(path);
2705         btrfs_free_path(dst_path);
2706
2707         mutex_lock(&root->fs_info->tree_log_mutex);
2708         ret = update_log_root(trans, log);
2709         BUG_ON(ret);
2710         mutex_unlock(&root->fs_info->tree_log_mutex);
2711 out:
2712         return 0;
2713 }
2714
2715 int btrfs_log_inode(struct btrfs_trans_handle *trans,
2716                     struct btrfs_root *root, struct inode *inode,
2717                     int inode_only)
2718 {
2719         int ret;
2720
2721         start_log_trans(trans, root);
2722         ret = __btrfs_log_inode(trans, root, inode, inode_only);
2723         end_log_trans(root);
2724         return ret;
2725 }
2726
2727 /*
2728  * helper function around btrfs_log_inode to make sure newly created
2729  * parent directories also end up in the log.  A minimal inode and backref
2730  * only logging is done of any parent directories that are older than
2731  * the last committed transaction
2732  */
2733 int btrfs_log_dentry(struct btrfs_trans_handle *trans,
2734                     struct btrfs_root *root, struct dentry *dentry)
2735 {
2736         int inode_only = LOG_INODE_ALL;
2737         struct super_block *sb;
2738         int ret;
2739
2740         start_log_trans(trans, root);
2741         sb = dentry->d_inode->i_sb;
2742         while(1) {
2743                 ret = __btrfs_log_inode(trans, root, dentry->d_inode,
2744                                         inode_only);
2745                 BUG_ON(ret);
2746                 inode_only = LOG_INODE_EXISTS;
2747
2748                 dentry = dentry->d_parent;
2749                 if (!dentry || !dentry->d_inode || sb != dentry->d_inode->i_sb)
2750                         break;
2751
2752                 if (BTRFS_I(dentry->d_inode)->generation <=
2753                     root->fs_info->last_trans_committed)
2754                         break;
2755         }
2756         end_log_trans(root);
2757         return 0;
2758 }
2759
2760 /*
2761  * it is not safe to log dentry if the chunk root has added new
2762  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
2763  * If this returns 1, you must commit the transaction to safely get your
2764  * data on disk.
2765  */
2766 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
2767                           struct btrfs_root *root, struct dentry *dentry)
2768 {
2769         u64 gen;
2770         gen = root->fs_info->last_trans_new_blockgroup;
2771         if (gen > root->fs_info->last_trans_committed)
2772                 return 1;
2773         else
2774                 return btrfs_log_dentry(trans, root, dentry);
2775 }
2776
2777 /*
2778  * should be called during mount to recover any replay any log trees
2779  * from the FS
2780  */
2781 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
2782 {
2783         int ret;
2784         struct btrfs_path *path;
2785         struct btrfs_trans_handle *trans;
2786         struct btrfs_key key;
2787         struct btrfs_key found_key;
2788         struct btrfs_key tmp_key;
2789         struct btrfs_root *log;
2790         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
2791         u64 highest_inode;
2792         struct walk_control wc = {
2793                 .process_func = process_one_buffer,
2794                 .stage = 0,
2795         };
2796
2797         fs_info->log_root_recovering = 1;
2798         path = btrfs_alloc_path();
2799         BUG_ON(!path);
2800
2801         trans = btrfs_start_transaction(fs_info->tree_root, 1);
2802
2803         wc.trans = trans;
2804         wc.pin = 1;
2805
2806         walk_log_tree(trans, log_root_tree, &wc);
2807
2808 again:
2809         key.objectid = BTRFS_TREE_LOG_OBJECTID;
2810         key.offset = (u64)-1;
2811         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
2812
2813         while(1) {
2814                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
2815                 if (ret < 0)
2816                         break;
2817                 if (ret > 0) {
2818                         if (path->slots[0] == 0)
2819                                 break;
2820                         path->slots[0]--;
2821                 }
2822                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2823                                       path->slots[0]);
2824                 btrfs_release_path(log_root_tree, path);
2825                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
2826                         break;
2827
2828                 log = btrfs_read_fs_root_no_radix(log_root_tree,
2829                                                   &found_key);
2830                 BUG_ON(!log);
2831
2832
2833                 tmp_key.objectid = found_key.offset;
2834                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
2835                 tmp_key.offset = (u64)-1;
2836
2837                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
2838
2839                 BUG_ON(!wc.replay_dest);
2840
2841                 btrfs_record_root_in_trans(wc.replay_dest);
2842                 ret = walk_log_tree(trans, log, &wc);
2843                 BUG_ON(ret);
2844
2845                 if (wc.stage == LOG_WALK_REPLAY_ALL) {
2846                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
2847                                                       path);
2848                         BUG_ON(ret);
2849                 }
2850                 ret = btrfs_find_highest_inode(wc.replay_dest, &highest_inode);
2851                 if (ret == 0) {
2852                         wc.replay_dest->highest_inode = highest_inode;
2853                         wc.replay_dest->last_inode_alloc = highest_inode;
2854                 }
2855
2856                 key.offset = found_key.offset - 1;
2857                 free_extent_buffer(log->node);
2858                 kfree(log);
2859
2860                 if (found_key.offset == 0)
2861                         break;
2862         }
2863         btrfs_release_path(log_root_tree, path);
2864
2865         /* step one is to pin it all, step two is to replay just inodes */
2866         if (wc.pin) {
2867                 wc.pin = 0;
2868                 wc.process_func = replay_one_buffer;
2869                 wc.stage = LOG_WALK_REPLAY_INODES;
2870                 goto again;
2871         }
2872         /* step three is to replay everything */
2873         if (wc.stage < LOG_WALK_REPLAY_ALL) {
2874                 wc.stage++;
2875                 goto again;
2876         }
2877
2878         btrfs_free_path(path);
2879
2880         free_extent_buffer(log_root_tree->node);
2881         log_root_tree->log_root = NULL;
2882         fs_info->log_root_recovering = 0;
2883
2884         /* step 4: commit the transaction, which also unpins the blocks */
2885         btrfs_commit_transaction(trans, fs_info->tree_root);
2886
2887         kfree(log_root_tree);
2888         return 0;
2889 }