Btrfs: Change btree locking to use explicit blocking points
[linux-3.10.git] / fs / btrfs / tree-log.c
1 /*
2  * Copyright (C) 2008 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/sched.h>
20 #include "ctree.h"
21 #include "transaction.h"
22 #include "disk-io.h"
23 #include "locking.h"
24 #include "print-tree.h"
25 #include "compat.h"
26 #include "tree-log.h"
27
28 /* magic values for the inode_only field in btrfs_log_inode:
29  *
30  * LOG_INODE_ALL means to log everything
31  * LOG_INODE_EXISTS means to log just enough to recreate the inode
32  * during log replay
33  */
34 #define LOG_INODE_ALL 0
35 #define LOG_INODE_EXISTS 1
36
37 /*
38  * stages for the tree walking.  The first
39  * stage (0) is to only pin down the blocks we find
40  * the second stage (1) is to make sure that all the inodes
41  * we find in the log are created in the subvolume.
42  *
43  * The last stage is to deal with directories and links and extents
44  * and all the other fun semantics
45  */
46 #define LOG_WALK_PIN_ONLY 0
47 #define LOG_WALK_REPLAY_INODES 1
48 #define LOG_WALK_REPLAY_ALL 2
49
50 static int __btrfs_log_inode(struct btrfs_trans_handle *trans,
51                              struct btrfs_root *root, struct inode *inode,
52                              int inode_only);
53 static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
54                              struct btrfs_root *root,
55                              struct btrfs_path *path, u64 objectid);
56
57 /*
58  * tree logging is a special write ahead log used to make sure that
59  * fsyncs and O_SYNCs can happen without doing full tree commits.
60  *
61  * Full tree commits are expensive because they require commonly
62  * modified blocks to be recowed, creating many dirty pages in the
63  * extent tree an 4x-6x higher write load than ext3.
64  *
65  * Instead of doing a tree commit on every fsync, we use the
66  * key ranges and transaction ids to find items for a given file or directory
67  * that have changed in this transaction.  Those items are copied into
68  * a special tree (one per subvolume root), that tree is written to disk
69  * and then the fsync is considered complete.
70  *
71  * After a crash, items are copied out of the log-tree back into the
72  * subvolume tree.  Any file data extents found are recorded in the extent
73  * allocation tree, and the log-tree freed.
74  *
75  * The log tree is read three times, once to pin down all the extents it is
76  * using in ram and once, once to create all the inodes logged in the tree
77  * and once to do all the other items.
78  */
79
80 /*
81  * start a sub transaction and setup the log tree
82  * this increments the log tree writer count to make the people
83  * syncing the tree wait for us to finish
84  */
85 static int start_log_trans(struct btrfs_trans_handle *trans,
86                            struct btrfs_root *root)
87 {
88         int ret;
89
90         mutex_lock(&root->log_mutex);
91         if (root->log_root) {
92                 root->log_batch++;
93                 atomic_inc(&root->log_writers);
94                 mutex_unlock(&root->log_mutex);
95                 return 0;
96         }
97         mutex_lock(&root->fs_info->tree_log_mutex);
98         if (!root->fs_info->log_root_tree) {
99                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
100                 BUG_ON(ret);
101         }
102         if (!root->log_root) {
103                 ret = btrfs_add_log_tree(trans, root);
104                 BUG_ON(ret);
105         }
106         mutex_unlock(&root->fs_info->tree_log_mutex);
107         root->log_batch++;
108         atomic_inc(&root->log_writers);
109         mutex_unlock(&root->log_mutex);
110         return 0;
111 }
112
113 /*
114  * returns 0 if there was a log transaction running and we were able
115  * to join, or returns -ENOENT if there were not transactions
116  * in progress
117  */
118 static int join_running_log_trans(struct btrfs_root *root)
119 {
120         int ret = -ENOENT;
121
122         smp_mb();
123         if (!root->log_root)
124                 return -ENOENT;
125
126         mutex_lock(&root->log_mutex);
127         if (root->log_root) {
128                 ret = 0;
129                 atomic_inc(&root->log_writers);
130         }
131         mutex_unlock(&root->log_mutex);
132         return ret;
133 }
134
135 /*
136  * indicate we're done making changes to the log tree
137  * and wake up anyone waiting to do a sync
138  */
139 static int end_log_trans(struct btrfs_root *root)
140 {
141         if (atomic_dec_and_test(&root->log_writers)) {
142                 smp_mb();
143                 if (waitqueue_active(&root->log_writer_wait))
144                         wake_up(&root->log_writer_wait);
145         }
146         return 0;
147 }
148
149
150 /*
151  * the walk control struct is used to pass state down the chain when
152  * processing the log tree.  The stage field tells us which part
153  * of the log tree processing we are currently doing.  The others
154  * are state fields used for that specific part
155  */
156 struct walk_control {
157         /* should we free the extent on disk when done?  This is used
158          * at transaction commit time while freeing a log tree
159          */
160         int free;
161
162         /* should we write out the extent buffer?  This is used
163          * while flushing the log tree to disk during a sync
164          */
165         int write;
166
167         /* should we wait for the extent buffer io to finish?  Also used
168          * while flushing the log tree to disk for a sync
169          */
170         int wait;
171
172         /* pin only walk, we record which extents on disk belong to the
173          * log trees
174          */
175         int pin;
176
177         /* what stage of the replay code we're currently in */
178         int stage;
179
180         /* the root we are currently replaying */
181         struct btrfs_root *replay_dest;
182
183         /* the trans handle for the current replay */
184         struct btrfs_trans_handle *trans;
185
186         /* the function that gets used to process blocks we find in the
187          * tree.  Note the extent_buffer might not be up to date when it is
188          * passed in, and it must be checked or read if you need the data
189          * inside it
190          */
191         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
192                             struct walk_control *wc, u64 gen);
193 };
194
195 /*
196  * process_func used to pin down extents, write them or wait on them
197  */
198 static int process_one_buffer(struct btrfs_root *log,
199                               struct extent_buffer *eb,
200                               struct walk_control *wc, u64 gen)
201 {
202         if (wc->pin) {
203                 mutex_lock(&log->fs_info->pinned_mutex);
204                 btrfs_update_pinned_extents(log->fs_info->extent_root,
205                                             eb->start, eb->len, 1);
206                 mutex_unlock(&log->fs_info->pinned_mutex);
207         }
208
209         if (btrfs_buffer_uptodate(eb, gen)) {
210                 if (wc->write)
211                         btrfs_write_tree_block(eb);
212                 if (wc->wait)
213                         btrfs_wait_tree_block_writeback(eb);
214         }
215         return 0;
216 }
217
218 /*
219  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
220  * to the src data we are copying out.
221  *
222  * root is the tree we are copying into, and path is a scratch
223  * path for use in this function (it should be released on entry and
224  * will be released on exit).
225  *
226  * If the key is already in the destination tree the existing item is
227  * overwritten.  If the existing item isn't big enough, it is extended.
228  * If it is too large, it is truncated.
229  *
230  * If the key isn't in the destination yet, a new item is inserted.
231  */
232 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
233                                    struct btrfs_root *root,
234                                    struct btrfs_path *path,
235                                    struct extent_buffer *eb, int slot,
236                                    struct btrfs_key *key)
237 {
238         int ret;
239         u32 item_size;
240         u64 saved_i_size = 0;
241         int save_old_i_size = 0;
242         unsigned long src_ptr;
243         unsigned long dst_ptr;
244         int overwrite_root = 0;
245
246         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
247                 overwrite_root = 1;
248
249         item_size = btrfs_item_size_nr(eb, slot);
250         src_ptr = btrfs_item_ptr_offset(eb, slot);
251
252         /* look for the key in the destination tree */
253         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
254         if (ret == 0) {
255                 char *src_copy;
256                 char *dst_copy;
257                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
258                                                   path->slots[0]);
259                 if (dst_size != item_size)
260                         goto insert;
261
262                 if (item_size == 0) {
263                         btrfs_release_path(root, path);
264                         return 0;
265                 }
266                 dst_copy = kmalloc(item_size, GFP_NOFS);
267                 src_copy = kmalloc(item_size, GFP_NOFS);
268
269                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
270
271                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
272                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
273                                    item_size);
274                 ret = memcmp(dst_copy, src_copy, item_size);
275
276                 kfree(dst_copy);
277                 kfree(src_copy);
278                 /*
279                  * they have the same contents, just return, this saves
280                  * us from cowing blocks in the destination tree and doing
281                  * extra writes that may not have been done by a previous
282                  * sync
283                  */
284                 if (ret == 0) {
285                         btrfs_release_path(root, path);
286                         return 0;
287                 }
288
289         }
290 insert:
291         btrfs_release_path(root, path);
292         /* try to insert the key into the destination tree */
293         ret = btrfs_insert_empty_item(trans, root, path,
294                                       key, item_size);
295
296         /* make sure any existing item is the correct size */
297         if (ret == -EEXIST) {
298                 u32 found_size;
299                 found_size = btrfs_item_size_nr(path->nodes[0],
300                                                 path->slots[0]);
301                 if (found_size > item_size) {
302                         btrfs_truncate_item(trans, root, path, item_size, 1);
303                 } else if (found_size < item_size) {
304                         ret = btrfs_extend_item(trans, root, path,
305                                                 item_size - found_size);
306                         BUG_ON(ret);
307                 }
308         } else if (ret) {
309                 BUG();
310         }
311         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
312                                         path->slots[0]);
313
314         /* don't overwrite an existing inode if the generation number
315          * was logged as zero.  This is done when the tree logging code
316          * is just logging an inode to make sure it exists after recovery.
317          *
318          * Also, don't overwrite i_size on directories during replay.
319          * log replay inserts and removes directory items based on the
320          * state of the tree found in the subvolume, and i_size is modified
321          * as it goes
322          */
323         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
324                 struct btrfs_inode_item *src_item;
325                 struct btrfs_inode_item *dst_item;
326
327                 src_item = (struct btrfs_inode_item *)src_ptr;
328                 dst_item = (struct btrfs_inode_item *)dst_ptr;
329
330                 if (btrfs_inode_generation(eb, src_item) == 0)
331                         goto no_copy;
332
333                 if (overwrite_root &&
334                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
335                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
336                         save_old_i_size = 1;
337                         saved_i_size = btrfs_inode_size(path->nodes[0],
338                                                         dst_item);
339                 }
340         }
341
342         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
343                            src_ptr, item_size);
344
345         if (save_old_i_size) {
346                 struct btrfs_inode_item *dst_item;
347                 dst_item = (struct btrfs_inode_item *)dst_ptr;
348                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
349         }
350
351         /* make sure the generation is filled in */
352         if (key->type == BTRFS_INODE_ITEM_KEY) {
353                 struct btrfs_inode_item *dst_item;
354                 dst_item = (struct btrfs_inode_item *)dst_ptr;
355                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
356                         btrfs_set_inode_generation(path->nodes[0], dst_item,
357                                                    trans->transid);
358                 }
359         }
360 no_copy:
361         btrfs_mark_buffer_dirty(path->nodes[0]);
362         btrfs_release_path(root, path);
363         return 0;
364 }
365
366 /*
367  * simple helper to read an inode off the disk from a given root
368  * This can only be called for subvolume roots and not for the log
369  */
370 static noinline struct inode *read_one_inode(struct btrfs_root *root,
371                                              u64 objectid)
372 {
373         struct inode *inode;
374         inode = btrfs_iget_locked(root->fs_info->sb, objectid, root);
375         if (inode->i_state & I_NEW) {
376                 BTRFS_I(inode)->root = root;
377                 BTRFS_I(inode)->location.objectid = objectid;
378                 BTRFS_I(inode)->location.type = BTRFS_INODE_ITEM_KEY;
379                 BTRFS_I(inode)->location.offset = 0;
380                 btrfs_read_locked_inode(inode);
381                 unlock_new_inode(inode);
382
383         }
384         if (is_bad_inode(inode)) {
385                 iput(inode);
386                 inode = NULL;
387         }
388         return inode;
389 }
390
391 /* replays a single extent in 'eb' at 'slot' with 'key' into the
392  * subvolume 'root'.  path is released on entry and should be released
393  * on exit.
394  *
395  * extents in the log tree have not been allocated out of the extent
396  * tree yet.  So, this completes the allocation, taking a reference
397  * as required if the extent already exists or creating a new extent
398  * if it isn't in the extent allocation tree yet.
399  *
400  * The extent is inserted into the file, dropping any existing extents
401  * from the file that overlap the new one.
402  */
403 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
404                                       struct btrfs_root *root,
405                                       struct btrfs_path *path,
406                                       struct extent_buffer *eb, int slot,
407                                       struct btrfs_key *key)
408 {
409         int found_type;
410         u64 mask = root->sectorsize - 1;
411         u64 extent_end;
412         u64 alloc_hint;
413         u64 start = key->offset;
414         u64 saved_nbytes;
415         struct btrfs_file_extent_item *item;
416         struct inode *inode = NULL;
417         unsigned long size;
418         int ret = 0;
419
420         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
421         found_type = btrfs_file_extent_type(eb, item);
422
423         if (found_type == BTRFS_FILE_EXTENT_REG ||
424             found_type == BTRFS_FILE_EXTENT_PREALLOC)
425                 extent_end = start + btrfs_file_extent_num_bytes(eb, item);
426         else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
427                 size = btrfs_file_extent_inline_len(eb, item);
428                 extent_end = (start + size + mask) & ~mask;
429         } else {
430                 ret = 0;
431                 goto out;
432         }
433
434         inode = read_one_inode(root, key->objectid);
435         if (!inode) {
436                 ret = -EIO;
437                 goto out;
438         }
439
440         /*
441          * first check to see if we already have this extent in the
442          * file.  This must be done before the btrfs_drop_extents run
443          * so we don't try to drop this extent.
444          */
445         ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
446                                        start, 0);
447
448         if (ret == 0 &&
449             (found_type == BTRFS_FILE_EXTENT_REG ||
450              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
451                 struct btrfs_file_extent_item cmp1;
452                 struct btrfs_file_extent_item cmp2;
453                 struct btrfs_file_extent_item *existing;
454                 struct extent_buffer *leaf;
455
456                 leaf = path->nodes[0];
457                 existing = btrfs_item_ptr(leaf, path->slots[0],
458                                           struct btrfs_file_extent_item);
459
460                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
461                                    sizeof(cmp1));
462                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
463                                    sizeof(cmp2));
464
465                 /*
466                  * we already have a pointer to this exact extent,
467                  * we don't have to do anything
468                  */
469                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
470                         btrfs_release_path(root, path);
471                         goto out;
472                 }
473         }
474         btrfs_release_path(root, path);
475
476         saved_nbytes = inode_get_bytes(inode);
477         /* drop any overlapping extents */
478         ret = btrfs_drop_extents(trans, root, inode,
479                          start, extent_end, start, &alloc_hint);
480         BUG_ON(ret);
481
482         if (found_type == BTRFS_FILE_EXTENT_REG ||
483             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
484                 unsigned long dest_offset;
485                 struct btrfs_key ins;
486
487                 ret = btrfs_insert_empty_item(trans, root, path, key,
488                                               sizeof(*item));
489                 BUG_ON(ret);
490                 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
491                                                     path->slots[0]);
492                 copy_extent_buffer(path->nodes[0], eb, dest_offset,
493                                 (unsigned long)item,  sizeof(*item));
494
495                 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
496                 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
497                 ins.type = BTRFS_EXTENT_ITEM_KEY;
498
499                 if (ins.objectid > 0) {
500                         u64 csum_start;
501                         u64 csum_end;
502                         LIST_HEAD(ordered_sums);
503                         /*
504                          * is this extent already allocated in the extent
505                          * allocation tree?  If so, just add a reference
506                          */
507                         ret = btrfs_lookup_extent(root, ins.objectid,
508                                                 ins.offset);
509                         if (ret == 0) {
510                                 ret = btrfs_inc_extent_ref(trans, root,
511                                                 ins.objectid, ins.offset,
512                                                 path->nodes[0]->start,
513                                                 root->root_key.objectid,
514                                                 trans->transid, key->objectid);
515                         } else {
516                                 /*
517                                  * insert the extent pointer in the extent
518                                  * allocation tree
519                                  */
520                                 ret = btrfs_alloc_logged_extent(trans, root,
521                                                 path->nodes[0]->start,
522                                                 root->root_key.objectid,
523                                                 trans->transid, key->objectid,
524                                                 &ins);
525                                 BUG_ON(ret);
526                         }
527                         btrfs_release_path(root, path);
528
529                         if (btrfs_file_extent_compression(eb, item)) {
530                                 csum_start = ins.objectid;
531                                 csum_end = csum_start + ins.offset;
532                         } else {
533                                 csum_start = ins.objectid +
534                                         btrfs_file_extent_offset(eb, item);
535                                 csum_end = csum_start +
536                                         btrfs_file_extent_num_bytes(eb, item);
537                         }
538
539                         ret = btrfs_lookup_csums_range(root->log_root,
540                                                 csum_start, csum_end - 1,
541                                                 &ordered_sums);
542                         BUG_ON(ret);
543                         while (!list_empty(&ordered_sums)) {
544                                 struct btrfs_ordered_sum *sums;
545                                 sums = list_entry(ordered_sums.next,
546                                                 struct btrfs_ordered_sum,
547                                                 list);
548                                 ret = btrfs_csum_file_blocks(trans,
549                                                 root->fs_info->csum_root,
550                                                 sums);
551                                 BUG_ON(ret);
552                                 list_del(&sums->list);
553                                 kfree(sums);
554                         }
555                 } else {
556                         btrfs_release_path(root, path);
557                 }
558         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
559                 /* inline extents are easy, we just overwrite them */
560                 ret = overwrite_item(trans, root, path, eb, slot, key);
561                 BUG_ON(ret);
562         }
563
564         inode_set_bytes(inode, saved_nbytes);
565         btrfs_update_inode(trans, root, inode);
566 out:
567         if (inode)
568                 iput(inode);
569         return ret;
570 }
571
572 /*
573  * when cleaning up conflicts between the directory names in the
574  * subvolume, directory names in the log and directory names in the
575  * inode back references, we may have to unlink inodes from directories.
576  *
577  * This is a helper function to do the unlink of a specific directory
578  * item
579  */
580 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
581                                       struct btrfs_root *root,
582                                       struct btrfs_path *path,
583                                       struct inode *dir,
584                                       struct btrfs_dir_item *di)
585 {
586         struct inode *inode;
587         char *name;
588         int name_len;
589         struct extent_buffer *leaf;
590         struct btrfs_key location;
591         int ret;
592
593         leaf = path->nodes[0];
594
595         btrfs_dir_item_key_to_cpu(leaf, di, &location);
596         name_len = btrfs_dir_name_len(leaf, di);
597         name = kmalloc(name_len, GFP_NOFS);
598         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
599         btrfs_release_path(root, path);
600
601         inode = read_one_inode(root, location.objectid);
602         BUG_ON(!inode);
603
604         ret = link_to_fixup_dir(trans, root, path, location.objectid);
605         BUG_ON(ret);
606         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
607         BUG_ON(ret);
608         kfree(name);
609
610         iput(inode);
611         return ret;
612 }
613
614 /*
615  * helper function to see if a given name and sequence number found
616  * in an inode back reference are already in a directory and correctly
617  * point to this inode
618  */
619 static noinline int inode_in_dir(struct btrfs_root *root,
620                                  struct btrfs_path *path,
621                                  u64 dirid, u64 objectid, u64 index,
622                                  const char *name, int name_len)
623 {
624         struct btrfs_dir_item *di;
625         struct btrfs_key location;
626         int match = 0;
627
628         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
629                                          index, name, name_len, 0);
630         if (di && !IS_ERR(di)) {
631                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
632                 if (location.objectid != objectid)
633                         goto out;
634         } else
635                 goto out;
636         btrfs_release_path(root, path);
637
638         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
639         if (di && !IS_ERR(di)) {
640                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
641                 if (location.objectid != objectid)
642                         goto out;
643         } else
644                 goto out;
645         match = 1;
646 out:
647         btrfs_release_path(root, path);
648         return match;
649 }
650
651 /*
652  * helper function to check a log tree for a named back reference in
653  * an inode.  This is used to decide if a back reference that is
654  * found in the subvolume conflicts with what we find in the log.
655  *
656  * inode backreferences may have multiple refs in a single item,
657  * during replay we process one reference at a time, and we don't
658  * want to delete valid links to a file from the subvolume if that
659  * link is also in the log.
660  */
661 static noinline int backref_in_log(struct btrfs_root *log,
662                                    struct btrfs_key *key,
663                                    char *name, int namelen)
664 {
665         struct btrfs_path *path;
666         struct btrfs_inode_ref *ref;
667         unsigned long ptr;
668         unsigned long ptr_end;
669         unsigned long name_ptr;
670         int found_name_len;
671         int item_size;
672         int ret;
673         int match = 0;
674
675         path = btrfs_alloc_path();
676         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
677         if (ret != 0)
678                 goto out;
679
680         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
681         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
682         ptr_end = ptr + item_size;
683         while (ptr < ptr_end) {
684                 ref = (struct btrfs_inode_ref *)ptr;
685                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
686                 if (found_name_len == namelen) {
687                         name_ptr = (unsigned long)(ref + 1);
688                         ret = memcmp_extent_buffer(path->nodes[0], name,
689                                                    name_ptr, namelen);
690                         if (ret == 0) {
691                                 match = 1;
692                                 goto out;
693                         }
694                 }
695                 ptr = (unsigned long)(ref + 1) + found_name_len;
696         }
697 out:
698         btrfs_free_path(path);
699         return match;
700 }
701
702
703 /*
704  * replay one inode back reference item found in the log tree.
705  * eb, slot and key refer to the buffer and key found in the log tree.
706  * root is the destination we are replaying into, and path is for temp
707  * use by this function.  (it should be released on return).
708  */
709 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
710                                   struct btrfs_root *root,
711                                   struct btrfs_root *log,
712                                   struct btrfs_path *path,
713                                   struct extent_buffer *eb, int slot,
714                                   struct btrfs_key *key)
715 {
716         struct inode *dir;
717         int ret;
718         struct btrfs_key location;
719         struct btrfs_inode_ref *ref;
720         struct btrfs_dir_item *di;
721         struct inode *inode;
722         char *name;
723         int namelen;
724         unsigned long ref_ptr;
725         unsigned long ref_end;
726
727         location.objectid = key->objectid;
728         location.type = BTRFS_INODE_ITEM_KEY;
729         location.offset = 0;
730
731         /*
732          * it is possible that we didn't log all the parent directories
733          * for a given inode.  If we don't find the dir, just don't
734          * copy the back ref in.  The link count fixup code will take
735          * care of the rest
736          */
737         dir = read_one_inode(root, key->offset);
738         if (!dir)
739                 return -ENOENT;
740
741         inode = read_one_inode(root, key->objectid);
742         BUG_ON(!dir);
743
744         ref_ptr = btrfs_item_ptr_offset(eb, slot);
745         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
746
747 again:
748         ref = (struct btrfs_inode_ref *)ref_ptr;
749
750         namelen = btrfs_inode_ref_name_len(eb, ref);
751         name = kmalloc(namelen, GFP_NOFS);
752         BUG_ON(!name);
753
754         read_extent_buffer(eb, name, (unsigned long)(ref + 1), namelen);
755
756         /* if we already have a perfect match, we're done */
757         if (inode_in_dir(root, path, dir->i_ino, inode->i_ino,
758                          btrfs_inode_ref_index(eb, ref),
759                          name, namelen)) {
760                 goto out;
761         }
762
763         /*
764          * look for a conflicting back reference in the metadata.
765          * if we find one we have to unlink that name of the file
766          * before we add our new link.  Later on, we overwrite any
767          * existing back reference, and we don't want to create
768          * dangling pointers in the directory.
769          */
770 conflict_again:
771         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
772         if (ret == 0) {
773                 char *victim_name;
774                 int victim_name_len;
775                 struct btrfs_inode_ref *victim_ref;
776                 unsigned long ptr;
777                 unsigned long ptr_end;
778                 struct extent_buffer *leaf = path->nodes[0];
779
780                 /* are we trying to overwrite a back ref for the root directory
781                  * if so, just jump out, we're done
782                  */
783                 if (key->objectid == key->offset)
784                         goto out_nowrite;
785
786                 /* check all the names in this back reference to see
787                  * if they are in the log.  if so, we allow them to stay
788                  * otherwise they must be unlinked as a conflict
789                  */
790                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
791                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
792                 while (ptr < ptr_end) {
793                         victim_ref = (struct btrfs_inode_ref *)ptr;
794                         victim_name_len = btrfs_inode_ref_name_len(leaf,
795                                                                    victim_ref);
796                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
797                         BUG_ON(!victim_name);
798
799                         read_extent_buffer(leaf, victim_name,
800                                            (unsigned long)(victim_ref + 1),
801                                            victim_name_len);
802
803                         if (!backref_in_log(log, key, victim_name,
804                                             victim_name_len)) {
805                                 btrfs_inc_nlink(inode);
806                                 btrfs_release_path(root, path);
807                                 ret = btrfs_unlink_inode(trans, root, dir,
808                                                          inode, victim_name,
809                                                          victim_name_len);
810                                 kfree(victim_name);
811                                 btrfs_release_path(root, path);
812                                 goto conflict_again;
813                         }
814                         kfree(victim_name);
815                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
816                 }
817                 BUG_ON(ret);
818         }
819         btrfs_release_path(root, path);
820
821         /* look for a conflicting sequence number */
822         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
823                                          btrfs_inode_ref_index(eb, ref),
824                                          name, namelen, 0);
825         if (di && !IS_ERR(di)) {
826                 ret = drop_one_dir_item(trans, root, path, dir, di);
827                 BUG_ON(ret);
828         }
829         btrfs_release_path(root, path);
830
831
832         /* look for a conflicting name */
833         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
834                                    name, namelen, 0);
835         if (di && !IS_ERR(di)) {
836                 ret = drop_one_dir_item(trans, root, path, dir, di);
837                 BUG_ON(ret);
838         }
839         btrfs_release_path(root, path);
840
841         /* insert our name */
842         ret = btrfs_add_link(trans, dir, inode, name, namelen, 0,
843                              btrfs_inode_ref_index(eb, ref));
844         BUG_ON(ret);
845
846         btrfs_update_inode(trans, root, inode);
847
848 out:
849         ref_ptr = (unsigned long)(ref + 1) + namelen;
850         kfree(name);
851         if (ref_ptr < ref_end)
852                 goto again;
853
854         /* finally write the back reference in the inode */
855         ret = overwrite_item(trans, root, path, eb, slot, key);
856         BUG_ON(ret);
857
858 out_nowrite:
859         btrfs_release_path(root, path);
860         iput(dir);
861         iput(inode);
862         return 0;
863 }
864
865 /*
866  * There are a few corners where the link count of the file can't
867  * be properly maintained during replay.  So, instead of adding
868  * lots of complexity to the log code, we just scan the backrefs
869  * for any file that has been through replay.
870  *
871  * The scan will update the link count on the inode to reflect the
872  * number of back refs found.  If it goes down to zero, the iput
873  * will free the inode.
874  */
875 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
876                                            struct btrfs_root *root,
877                                            struct inode *inode)
878 {
879         struct btrfs_path *path;
880         int ret;
881         struct btrfs_key key;
882         u64 nlink = 0;
883         unsigned long ptr;
884         unsigned long ptr_end;
885         int name_len;
886
887         key.objectid = inode->i_ino;
888         key.type = BTRFS_INODE_REF_KEY;
889         key.offset = (u64)-1;
890
891         path = btrfs_alloc_path();
892
893         while (1) {
894                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
895                 if (ret < 0)
896                         break;
897                 if (ret > 0) {
898                         if (path->slots[0] == 0)
899                                 break;
900                         path->slots[0]--;
901                 }
902                 btrfs_item_key_to_cpu(path->nodes[0], &key,
903                                       path->slots[0]);
904                 if (key.objectid != inode->i_ino ||
905                     key.type != BTRFS_INODE_REF_KEY)
906                         break;
907                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
908                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
909                                                    path->slots[0]);
910                 while (ptr < ptr_end) {
911                         struct btrfs_inode_ref *ref;
912
913                         ref = (struct btrfs_inode_ref *)ptr;
914                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
915                                                             ref);
916                         ptr = (unsigned long)(ref + 1) + name_len;
917                         nlink++;
918                 }
919
920                 if (key.offset == 0)
921                         break;
922                 key.offset--;
923                 btrfs_release_path(root, path);
924         }
925         btrfs_free_path(path);
926         if (nlink != inode->i_nlink) {
927                 inode->i_nlink = nlink;
928                 btrfs_update_inode(trans, root, inode);
929         }
930         BTRFS_I(inode)->index_cnt = (u64)-1;
931
932         return 0;
933 }
934
935 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
936                                             struct btrfs_root *root,
937                                             struct btrfs_path *path)
938 {
939         int ret;
940         struct btrfs_key key;
941         struct inode *inode;
942
943         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
944         key.type = BTRFS_ORPHAN_ITEM_KEY;
945         key.offset = (u64)-1;
946         while (1) {
947                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
948                 if (ret < 0)
949                         break;
950
951                 if (ret == 1) {
952                         if (path->slots[0] == 0)
953                                 break;
954                         path->slots[0]--;
955                 }
956
957                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
958                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
959                     key.type != BTRFS_ORPHAN_ITEM_KEY)
960                         break;
961
962                 ret = btrfs_del_item(trans, root, path);
963                 BUG_ON(ret);
964
965                 btrfs_release_path(root, path);
966                 inode = read_one_inode(root, key.offset);
967                 BUG_ON(!inode);
968
969                 ret = fixup_inode_link_count(trans, root, inode);
970                 BUG_ON(ret);
971
972                 iput(inode);
973
974                 if (key.offset == 0)
975                         break;
976                 key.offset--;
977         }
978         btrfs_release_path(root, path);
979         return 0;
980 }
981
982
983 /*
984  * record a given inode in the fixup dir so we can check its link
985  * count when replay is done.  The link count is incremented here
986  * so the inode won't go away until we check it
987  */
988 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
989                                       struct btrfs_root *root,
990                                       struct btrfs_path *path,
991                                       u64 objectid)
992 {
993         struct btrfs_key key;
994         int ret = 0;
995         struct inode *inode;
996
997         inode = read_one_inode(root, objectid);
998         BUG_ON(!inode);
999
1000         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1001         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1002         key.offset = objectid;
1003
1004         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1005
1006         btrfs_release_path(root, path);
1007         if (ret == 0) {
1008                 btrfs_inc_nlink(inode);
1009                 btrfs_update_inode(trans, root, inode);
1010         } else if (ret == -EEXIST) {
1011                 ret = 0;
1012         } else {
1013                 BUG();
1014         }
1015         iput(inode);
1016
1017         return ret;
1018 }
1019
1020 /*
1021  * when replaying the log for a directory, we only insert names
1022  * for inodes that actually exist.  This means an fsync on a directory
1023  * does not implicitly fsync all the new files in it
1024  */
1025 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1026                                     struct btrfs_root *root,
1027                                     struct btrfs_path *path,
1028                                     u64 dirid, u64 index,
1029                                     char *name, int name_len, u8 type,
1030                                     struct btrfs_key *location)
1031 {
1032         struct inode *inode;
1033         struct inode *dir;
1034         int ret;
1035
1036         inode = read_one_inode(root, location->objectid);
1037         if (!inode)
1038                 return -ENOENT;
1039
1040         dir = read_one_inode(root, dirid);
1041         if (!dir) {
1042                 iput(inode);
1043                 return -EIO;
1044         }
1045         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1046
1047         /* FIXME, put inode into FIXUP list */
1048
1049         iput(inode);
1050         iput(dir);
1051         return ret;
1052 }
1053
1054 /*
1055  * take a single entry in a log directory item and replay it into
1056  * the subvolume.
1057  *
1058  * if a conflicting item exists in the subdirectory already,
1059  * the inode it points to is unlinked and put into the link count
1060  * fix up tree.
1061  *
1062  * If a name from the log points to a file or directory that does
1063  * not exist in the FS, it is skipped.  fsyncs on directories
1064  * do not force down inodes inside that directory, just changes to the
1065  * names or unlinks in a directory.
1066  */
1067 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1068                                     struct btrfs_root *root,
1069                                     struct btrfs_path *path,
1070                                     struct extent_buffer *eb,
1071                                     struct btrfs_dir_item *di,
1072                                     struct btrfs_key *key)
1073 {
1074         char *name;
1075         int name_len;
1076         struct btrfs_dir_item *dst_di;
1077         struct btrfs_key found_key;
1078         struct btrfs_key log_key;
1079         struct inode *dir;
1080         u8 log_type;
1081         int exists;
1082         int ret;
1083
1084         dir = read_one_inode(root, key->objectid);
1085         BUG_ON(!dir);
1086
1087         name_len = btrfs_dir_name_len(eb, di);
1088         name = kmalloc(name_len, GFP_NOFS);
1089         log_type = btrfs_dir_type(eb, di);
1090         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1091                    name_len);
1092
1093         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1094         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1095         if (exists == 0)
1096                 exists = 1;
1097         else
1098                 exists = 0;
1099         btrfs_release_path(root, path);
1100
1101         if (key->type == BTRFS_DIR_ITEM_KEY) {
1102                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1103                                        name, name_len, 1);
1104         } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1105                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1106                                                      key->objectid,
1107                                                      key->offset, name,
1108                                                      name_len, 1);
1109         } else {
1110                 BUG();
1111         }
1112         if (!dst_di || IS_ERR(dst_di)) {
1113                 /* we need a sequence number to insert, so we only
1114                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1115                  */
1116                 if (key->type != BTRFS_DIR_INDEX_KEY)
1117                         goto out;
1118                 goto insert;
1119         }
1120
1121         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1122         /* the existing item matches the logged item */
1123         if (found_key.objectid == log_key.objectid &&
1124             found_key.type == log_key.type &&
1125             found_key.offset == log_key.offset &&
1126             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1127                 goto out;
1128         }
1129
1130         /*
1131          * don't drop the conflicting directory entry if the inode
1132          * for the new entry doesn't exist
1133          */
1134         if (!exists)
1135                 goto out;
1136
1137         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1138         BUG_ON(ret);
1139
1140         if (key->type == BTRFS_DIR_INDEX_KEY)
1141                 goto insert;
1142 out:
1143         btrfs_release_path(root, path);
1144         kfree(name);
1145         iput(dir);
1146         return 0;
1147
1148 insert:
1149         btrfs_release_path(root, path);
1150         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1151                               name, name_len, log_type, &log_key);
1152
1153         if (ret && ret != -ENOENT)
1154                 BUG();
1155         goto out;
1156 }
1157
1158 /*
1159  * find all the names in a directory item and reconcile them into
1160  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1161  * one name in a directory item, but the same code gets used for
1162  * both directory index types
1163  */
1164 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1165                                         struct btrfs_root *root,
1166                                         struct btrfs_path *path,
1167                                         struct extent_buffer *eb, int slot,
1168                                         struct btrfs_key *key)
1169 {
1170         int ret;
1171         u32 item_size = btrfs_item_size_nr(eb, slot);
1172         struct btrfs_dir_item *di;
1173         int name_len;
1174         unsigned long ptr;
1175         unsigned long ptr_end;
1176
1177         ptr = btrfs_item_ptr_offset(eb, slot);
1178         ptr_end = ptr + item_size;
1179         while (ptr < ptr_end) {
1180                 di = (struct btrfs_dir_item *)ptr;
1181                 name_len = btrfs_dir_name_len(eb, di);
1182                 ret = replay_one_name(trans, root, path, eb, di, key);
1183                 BUG_ON(ret);
1184                 ptr = (unsigned long)(di + 1);
1185                 ptr += name_len;
1186         }
1187         return 0;
1188 }
1189
1190 /*
1191  * directory replay has two parts.  There are the standard directory
1192  * items in the log copied from the subvolume, and range items
1193  * created in the log while the subvolume was logged.
1194  *
1195  * The range items tell us which parts of the key space the log
1196  * is authoritative for.  During replay, if a key in the subvolume
1197  * directory is in a logged range item, but not actually in the log
1198  * that means it was deleted from the directory before the fsync
1199  * and should be removed.
1200  */
1201 static noinline int find_dir_range(struct btrfs_root *root,
1202                                    struct btrfs_path *path,
1203                                    u64 dirid, int key_type,
1204                                    u64 *start_ret, u64 *end_ret)
1205 {
1206         struct btrfs_key key;
1207         u64 found_end;
1208         struct btrfs_dir_log_item *item;
1209         int ret;
1210         int nritems;
1211
1212         if (*start_ret == (u64)-1)
1213                 return 1;
1214
1215         key.objectid = dirid;
1216         key.type = key_type;
1217         key.offset = *start_ret;
1218
1219         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1220         if (ret < 0)
1221                 goto out;
1222         if (ret > 0) {
1223                 if (path->slots[0] == 0)
1224                         goto out;
1225                 path->slots[0]--;
1226         }
1227         if (ret != 0)
1228                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1229
1230         if (key.type != key_type || key.objectid != dirid) {
1231                 ret = 1;
1232                 goto next;
1233         }
1234         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1235                               struct btrfs_dir_log_item);
1236         found_end = btrfs_dir_log_end(path->nodes[0], item);
1237
1238         if (*start_ret >= key.offset && *start_ret <= found_end) {
1239                 ret = 0;
1240                 *start_ret = key.offset;
1241                 *end_ret = found_end;
1242                 goto out;
1243         }
1244         ret = 1;
1245 next:
1246         /* check the next slot in the tree to see if it is a valid item */
1247         nritems = btrfs_header_nritems(path->nodes[0]);
1248         if (path->slots[0] >= nritems) {
1249                 ret = btrfs_next_leaf(root, path);
1250                 if (ret)
1251                         goto out;
1252         } else {
1253                 path->slots[0]++;
1254         }
1255
1256         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1257
1258         if (key.type != key_type || key.objectid != dirid) {
1259                 ret = 1;
1260                 goto out;
1261         }
1262         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1263                               struct btrfs_dir_log_item);
1264         found_end = btrfs_dir_log_end(path->nodes[0], item);
1265         *start_ret = key.offset;
1266         *end_ret = found_end;
1267         ret = 0;
1268 out:
1269         btrfs_release_path(root, path);
1270         return ret;
1271 }
1272
1273 /*
1274  * this looks for a given directory item in the log.  If the directory
1275  * item is not in the log, the item is removed and the inode it points
1276  * to is unlinked
1277  */
1278 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1279                                       struct btrfs_root *root,
1280                                       struct btrfs_root *log,
1281                                       struct btrfs_path *path,
1282                                       struct btrfs_path *log_path,
1283                                       struct inode *dir,
1284                                       struct btrfs_key *dir_key)
1285 {
1286         int ret;
1287         struct extent_buffer *eb;
1288         int slot;
1289         u32 item_size;
1290         struct btrfs_dir_item *di;
1291         struct btrfs_dir_item *log_di;
1292         int name_len;
1293         unsigned long ptr;
1294         unsigned long ptr_end;
1295         char *name;
1296         struct inode *inode;
1297         struct btrfs_key location;
1298
1299 again:
1300         eb = path->nodes[0];
1301         slot = path->slots[0];
1302         item_size = btrfs_item_size_nr(eb, slot);
1303         ptr = btrfs_item_ptr_offset(eb, slot);
1304         ptr_end = ptr + item_size;
1305         while (ptr < ptr_end) {
1306                 di = (struct btrfs_dir_item *)ptr;
1307                 name_len = btrfs_dir_name_len(eb, di);
1308                 name = kmalloc(name_len, GFP_NOFS);
1309                 if (!name) {
1310                         ret = -ENOMEM;
1311                         goto out;
1312                 }
1313                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1314                                   name_len);
1315                 log_di = NULL;
1316                 if (dir_key->type == BTRFS_DIR_ITEM_KEY) {
1317                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1318                                                        dir_key->objectid,
1319                                                        name, name_len, 0);
1320                 } else if (dir_key->type == BTRFS_DIR_INDEX_KEY) {
1321                         log_di = btrfs_lookup_dir_index_item(trans, log,
1322                                                      log_path,
1323                                                      dir_key->objectid,
1324                                                      dir_key->offset,
1325                                                      name, name_len, 0);
1326                 }
1327                 if (!log_di || IS_ERR(log_di)) {
1328                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1329                         btrfs_release_path(root, path);
1330                         btrfs_release_path(log, log_path);
1331                         inode = read_one_inode(root, location.objectid);
1332                         BUG_ON(!inode);
1333
1334                         ret = link_to_fixup_dir(trans, root,
1335                                                 path, location.objectid);
1336                         BUG_ON(ret);
1337                         btrfs_inc_nlink(inode);
1338                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1339                                                  name, name_len);
1340                         BUG_ON(ret);
1341                         kfree(name);
1342                         iput(inode);
1343
1344                         /* there might still be more names under this key
1345                          * check and repeat if required
1346                          */
1347                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1348                                                 0, 0);
1349                         if (ret == 0)
1350                                 goto again;
1351                         ret = 0;
1352                         goto out;
1353                 }
1354                 btrfs_release_path(log, log_path);
1355                 kfree(name);
1356
1357                 ptr = (unsigned long)(di + 1);
1358                 ptr += name_len;
1359         }
1360         ret = 0;
1361 out:
1362         btrfs_release_path(root, path);
1363         btrfs_release_path(log, log_path);
1364         return ret;
1365 }
1366
1367 /*
1368  * deletion replay happens before we copy any new directory items
1369  * out of the log or out of backreferences from inodes.  It
1370  * scans the log to find ranges of keys that log is authoritative for,
1371  * and then scans the directory to find items in those ranges that are
1372  * not present in the log.
1373  *
1374  * Anything we don't find in the log is unlinked and removed from the
1375  * directory.
1376  */
1377 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1378                                        struct btrfs_root *root,
1379                                        struct btrfs_root *log,
1380                                        struct btrfs_path *path,
1381                                        u64 dirid)
1382 {
1383         u64 range_start;
1384         u64 range_end;
1385         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1386         int ret = 0;
1387         struct btrfs_key dir_key;
1388         struct btrfs_key found_key;
1389         struct btrfs_path *log_path;
1390         struct inode *dir;
1391
1392         dir_key.objectid = dirid;
1393         dir_key.type = BTRFS_DIR_ITEM_KEY;
1394         log_path = btrfs_alloc_path();
1395         if (!log_path)
1396                 return -ENOMEM;
1397
1398         dir = read_one_inode(root, dirid);
1399         /* it isn't an error if the inode isn't there, that can happen
1400          * because we replay the deletes before we copy in the inode item
1401          * from the log
1402          */
1403         if (!dir) {
1404                 btrfs_free_path(log_path);
1405                 return 0;
1406         }
1407 again:
1408         range_start = 0;
1409         range_end = 0;
1410         while (1) {
1411                 ret = find_dir_range(log, path, dirid, key_type,
1412                                      &range_start, &range_end);
1413                 if (ret != 0)
1414                         break;
1415
1416                 dir_key.offset = range_start;
1417                 while (1) {
1418                         int nritems;
1419                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
1420                                                 0, 0);
1421                         if (ret < 0)
1422                                 goto out;
1423
1424                         nritems = btrfs_header_nritems(path->nodes[0]);
1425                         if (path->slots[0] >= nritems) {
1426                                 ret = btrfs_next_leaf(root, path);
1427                                 if (ret)
1428                                         break;
1429                         }
1430                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1431                                               path->slots[0]);
1432                         if (found_key.objectid != dirid ||
1433                             found_key.type != dir_key.type)
1434                                 goto next_type;
1435
1436                         if (found_key.offset > range_end)
1437                                 break;
1438
1439                         ret = check_item_in_log(trans, root, log, path,
1440                                                 log_path, dir, &found_key);
1441                         BUG_ON(ret);
1442                         if (found_key.offset == (u64)-1)
1443                                 break;
1444                         dir_key.offset = found_key.offset + 1;
1445                 }
1446                 btrfs_release_path(root, path);
1447                 if (range_end == (u64)-1)
1448                         break;
1449                 range_start = range_end + 1;
1450         }
1451
1452 next_type:
1453         ret = 0;
1454         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1455                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1456                 dir_key.type = BTRFS_DIR_INDEX_KEY;
1457                 btrfs_release_path(root, path);
1458                 goto again;
1459         }
1460 out:
1461         btrfs_release_path(root, path);
1462         btrfs_free_path(log_path);
1463         iput(dir);
1464         return ret;
1465 }
1466
1467 /*
1468  * the process_func used to replay items from the log tree.  This
1469  * gets called in two different stages.  The first stage just looks
1470  * for inodes and makes sure they are all copied into the subvolume.
1471  *
1472  * The second stage copies all the other item types from the log into
1473  * the subvolume.  The two stage approach is slower, but gets rid of
1474  * lots of complexity around inodes referencing other inodes that exist
1475  * only in the log (references come from either directory items or inode
1476  * back refs).
1477  */
1478 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
1479                              struct walk_control *wc, u64 gen)
1480 {
1481         int nritems;
1482         struct btrfs_path *path;
1483         struct btrfs_root *root = wc->replay_dest;
1484         struct btrfs_key key;
1485         u32 item_size;
1486         int level;
1487         int i;
1488         int ret;
1489
1490         btrfs_read_buffer(eb, gen);
1491
1492         level = btrfs_header_level(eb);
1493
1494         if (level != 0)
1495                 return 0;
1496
1497         path = btrfs_alloc_path();
1498         BUG_ON(!path);
1499
1500         nritems = btrfs_header_nritems(eb);
1501         for (i = 0; i < nritems; i++) {
1502                 btrfs_item_key_to_cpu(eb, &key, i);
1503                 item_size = btrfs_item_size_nr(eb, i);
1504
1505                 /* inode keys are done during the first stage */
1506                 if (key.type == BTRFS_INODE_ITEM_KEY &&
1507                     wc->stage == LOG_WALK_REPLAY_INODES) {
1508                         struct inode *inode;
1509                         struct btrfs_inode_item *inode_item;
1510                         u32 mode;
1511
1512                         inode_item = btrfs_item_ptr(eb, i,
1513                                             struct btrfs_inode_item);
1514                         mode = btrfs_inode_mode(eb, inode_item);
1515                         if (S_ISDIR(mode)) {
1516                                 ret = replay_dir_deletes(wc->trans,
1517                                          root, log, path, key.objectid);
1518                                 BUG_ON(ret);
1519                         }
1520                         ret = overwrite_item(wc->trans, root, path,
1521                                              eb, i, &key);
1522                         BUG_ON(ret);
1523
1524                         /* for regular files, truncate away
1525                          * extents past the new EOF
1526                          */
1527                         if (S_ISREG(mode)) {
1528                                 inode = read_one_inode(root,
1529                                                        key.objectid);
1530                                 BUG_ON(!inode);
1531
1532                                 ret = btrfs_truncate_inode_items(wc->trans,
1533                                         root, inode, inode->i_size,
1534                                         BTRFS_EXTENT_DATA_KEY);
1535                                 BUG_ON(ret);
1536                                 iput(inode);
1537                         }
1538                         ret = link_to_fixup_dir(wc->trans, root,
1539                                                 path, key.objectid);
1540                         BUG_ON(ret);
1541                 }
1542                 if (wc->stage < LOG_WALK_REPLAY_ALL)
1543                         continue;
1544
1545                 /* these keys are simply copied */
1546                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
1547                         ret = overwrite_item(wc->trans, root, path,
1548                                              eb, i, &key);
1549                         BUG_ON(ret);
1550                 } else if (key.type == BTRFS_INODE_REF_KEY) {
1551                         ret = add_inode_ref(wc->trans, root, log, path,
1552                                             eb, i, &key);
1553                         BUG_ON(ret && ret != -ENOENT);
1554                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
1555                         ret = replay_one_extent(wc->trans, root, path,
1556                                                 eb, i, &key);
1557                         BUG_ON(ret);
1558                 } else if (key.type == BTRFS_DIR_ITEM_KEY ||
1559                            key.type == BTRFS_DIR_INDEX_KEY) {
1560                         ret = replay_one_dir_item(wc->trans, root, path,
1561                                                   eb, i, &key);
1562                         BUG_ON(ret);
1563                 }
1564         }
1565         btrfs_free_path(path);
1566         return 0;
1567 }
1568
1569 static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
1570                                    struct btrfs_root *root,
1571                                    struct btrfs_path *path, int *level,
1572                                    struct walk_control *wc)
1573 {
1574         u64 root_owner;
1575         u64 root_gen;
1576         u64 bytenr;
1577         u64 ptr_gen;
1578         struct extent_buffer *next;
1579         struct extent_buffer *cur;
1580         struct extent_buffer *parent;
1581         u32 blocksize;
1582         int ret = 0;
1583
1584         WARN_ON(*level < 0);
1585         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1586
1587         while (*level > 0) {
1588                 WARN_ON(*level < 0);
1589                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1590                 cur = path->nodes[*level];
1591
1592                 if (btrfs_header_level(cur) != *level)
1593                         WARN_ON(1);
1594
1595                 if (path->slots[*level] >=
1596                     btrfs_header_nritems(cur))
1597                         break;
1598
1599                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1600                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1601                 blocksize = btrfs_level_size(root, *level - 1);
1602
1603                 parent = path->nodes[*level];
1604                 root_owner = btrfs_header_owner(parent);
1605                 root_gen = btrfs_header_generation(parent);
1606
1607                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
1608
1609                 wc->process_func(root, next, wc, ptr_gen);
1610
1611                 if (*level == 1) {
1612                         path->slots[*level]++;
1613                         if (wc->free) {
1614                                 btrfs_read_buffer(next, ptr_gen);
1615
1616                                 btrfs_tree_lock(next);
1617                                 clean_tree_block(trans, root, next);
1618                                 btrfs_set_lock_blocking(next);
1619                                 btrfs_wait_tree_block_writeback(next);
1620                                 btrfs_tree_unlock(next);
1621
1622                                 ret = btrfs_drop_leaf_ref(trans, root, next);
1623                                 BUG_ON(ret);
1624
1625                                 WARN_ON(root_owner !=
1626                                         BTRFS_TREE_LOG_OBJECTID);
1627                                 ret = btrfs_free_reserved_extent(root,
1628                                                          bytenr, blocksize);
1629                                 BUG_ON(ret);
1630                         }
1631                         free_extent_buffer(next);
1632                         continue;
1633                 }
1634                 btrfs_read_buffer(next, ptr_gen);
1635
1636                 WARN_ON(*level <= 0);
1637                 if (path->nodes[*level-1])
1638                         free_extent_buffer(path->nodes[*level-1]);
1639                 path->nodes[*level-1] = next;
1640                 *level = btrfs_header_level(next);
1641                 path->slots[*level] = 0;
1642                 cond_resched();
1643         }
1644         WARN_ON(*level < 0);
1645         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1646
1647         if (path->nodes[*level] == root->node)
1648                 parent = path->nodes[*level];
1649         else
1650                 parent = path->nodes[*level + 1];
1651
1652         bytenr = path->nodes[*level]->start;
1653
1654         blocksize = btrfs_level_size(root, *level);
1655         root_owner = btrfs_header_owner(parent);
1656         root_gen = btrfs_header_generation(parent);
1657
1658         wc->process_func(root, path->nodes[*level], wc,
1659                          btrfs_header_generation(path->nodes[*level]));
1660
1661         if (wc->free) {
1662                 next = path->nodes[*level];
1663                 btrfs_tree_lock(next);
1664                 clean_tree_block(trans, root, next);
1665                 btrfs_set_lock_blocking(next);
1666                 btrfs_wait_tree_block_writeback(next);
1667                 btrfs_tree_unlock(next);
1668
1669                 if (*level == 0) {
1670                         ret = btrfs_drop_leaf_ref(trans, root, next);
1671                         BUG_ON(ret);
1672                 }
1673                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1674                 ret = btrfs_free_reserved_extent(root, bytenr, blocksize);
1675                 BUG_ON(ret);
1676         }
1677         free_extent_buffer(path->nodes[*level]);
1678         path->nodes[*level] = NULL;
1679         *level += 1;
1680
1681         cond_resched();
1682         return 0;
1683 }
1684
1685 static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
1686                                  struct btrfs_root *root,
1687                                  struct btrfs_path *path, int *level,
1688                                  struct walk_control *wc)
1689 {
1690         u64 root_owner;
1691         u64 root_gen;
1692         int i;
1693         int slot;
1694         int ret;
1695
1696         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1697                 slot = path->slots[i];
1698                 if (slot < btrfs_header_nritems(path->nodes[i]) - 1) {
1699                         struct extent_buffer *node;
1700                         node = path->nodes[i];
1701                         path->slots[i]++;
1702                         *level = i;
1703                         WARN_ON(*level == 0);
1704                         return 0;
1705                 } else {
1706                         struct extent_buffer *parent;
1707                         if (path->nodes[*level] == root->node)
1708                                 parent = path->nodes[*level];
1709                         else
1710                                 parent = path->nodes[*level + 1];
1711
1712                         root_owner = btrfs_header_owner(parent);
1713                         root_gen = btrfs_header_generation(parent);
1714                         wc->process_func(root, path->nodes[*level], wc,
1715                                  btrfs_header_generation(path->nodes[*level]));
1716                         if (wc->free) {
1717                                 struct extent_buffer *next;
1718
1719                                 next = path->nodes[*level];
1720
1721                                 btrfs_tree_lock(next);
1722                                 clean_tree_block(trans, root, next);
1723                                 btrfs_set_lock_blocking(next);
1724                                 btrfs_wait_tree_block_writeback(next);
1725                                 btrfs_tree_unlock(next);
1726
1727                                 if (*level == 0) {
1728                                         ret = btrfs_drop_leaf_ref(trans, root,
1729                                                                   next);
1730                                         BUG_ON(ret);
1731                                 }
1732
1733                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1734                                 ret = btrfs_free_reserved_extent(root,
1735                                                 path->nodes[*level]->start,
1736                                                 path->nodes[*level]->len);
1737                                 BUG_ON(ret);
1738                         }
1739                         free_extent_buffer(path->nodes[*level]);
1740                         path->nodes[*level] = NULL;
1741                         *level = i + 1;
1742                 }
1743         }
1744         return 1;
1745 }
1746
1747 /*
1748  * drop the reference count on the tree rooted at 'snap'.  This traverses
1749  * the tree freeing any blocks that have a ref count of zero after being
1750  * decremented.
1751  */
1752 static int walk_log_tree(struct btrfs_trans_handle *trans,
1753                          struct btrfs_root *log, struct walk_control *wc)
1754 {
1755         int ret = 0;
1756         int wret;
1757         int level;
1758         struct btrfs_path *path;
1759         int i;
1760         int orig_level;
1761
1762         path = btrfs_alloc_path();
1763         BUG_ON(!path);
1764
1765         level = btrfs_header_level(log->node);
1766         orig_level = level;
1767         path->nodes[level] = log->node;
1768         extent_buffer_get(log->node);
1769         path->slots[level] = 0;
1770
1771         while (1) {
1772                 wret = walk_down_log_tree(trans, log, path, &level, wc);
1773                 if (wret > 0)
1774                         break;
1775                 if (wret < 0)
1776                         ret = wret;
1777
1778                 wret = walk_up_log_tree(trans, log, path, &level, wc);
1779                 if (wret > 0)
1780                         break;
1781                 if (wret < 0)
1782                         ret = wret;
1783         }
1784
1785         /* was the root node processed? if not, catch it here */
1786         if (path->nodes[orig_level]) {
1787                 wc->process_func(log, path->nodes[orig_level], wc,
1788                          btrfs_header_generation(path->nodes[orig_level]));
1789                 if (wc->free) {
1790                         struct extent_buffer *next;
1791
1792                         next = path->nodes[orig_level];
1793
1794                         btrfs_tree_lock(next);
1795                         clean_tree_block(trans, log, next);
1796                         btrfs_set_lock_blocking(next);
1797                         btrfs_wait_tree_block_writeback(next);
1798                         btrfs_tree_unlock(next);
1799
1800                         if (orig_level == 0) {
1801                                 ret = btrfs_drop_leaf_ref(trans, log,
1802                                                           next);
1803                                 BUG_ON(ret);
1804                         }
1805                         WARN_ON(log->root_key.objectid !=
1806                                 BTRFS_TREE_LOG_OBJECTID);
1807                         ret = btrfs_free_reserved_extent(log, next->start,
1808                                                          next->len);
1809                         BUG_ON(ret);
1810                 }
1811         }
1812
1813         for (i = 0; i <= orig_level; i++) {
1814                 if (path->nodes[i]) {
1815                         free_extent_buffer(path->nodes[i]);
1816                         path->nodes[i] = NULL;
1817                 }
1818         }
1819         btrfs_free_path(path);
1820         return ret;
1821 }
1822
1823 /*
1824  * helper function to update the item for a given subvolumes log root
1825  * in the tree of log roots
1826  */
1827 static int update_log_root(struct btrfs_trans_handle *trans,
1828                            struct btrfs_root *log)
1829 {
1830         int ret;
1831
1832         if (log->log_transid == 1) {
1833                 /* insert root item on the first sync */
1834                 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
1835                                 &log->root_key, &log->root_item);
1836         } else {
1837                 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
1838                                 &log->root_key, &log->root_item);
1839         }
1840         return ret;
1841 }
1842
1843 static int wait_log_commit(struct btrfs_root *root, unsigned long transid)
1844 {
1845         DEFINE_WAIT(wait);
1846         int index = transid % 2;
1847
1848         /*
1849          * we only allow two pending log transactions at a time,
1850          * so we know that if ours is more than 2 older than the
1851          * current transaction, we're done
1852          */
1853         do {
1854                 prepare_to_wait(&root->log_commit_wait[index],
1855                                 &wait, TASK_UNINTERRUPTIBLE);
1856                 mutex_unlock(&root->log_mutex);
1857                 if (root->log_transid < transid + 2 &&
1858                     atomic_read(&root->log_commit[index]))
1859                         schedule();
1860                 finish_wait(&root->log_commit_wait[index], &wait);
1861                 mutex_lock(&root->log_mutex);
1862         } while (root->log_transid < transid + 2 &&
1863                  atomic_read(&root->log_commit[index]));
1864         return 0;
1865 }
1866
1867 static int wait_for_writer(struct btrfs_root *root)
1868 {
1869         DEFINE_WAIT(wait);
1870         while (atomic_read(&root->log_writers)) {
1871                 prepare_to_wait(&root->log_writer_wait,
1872                                 &wait, TASK_UNINTERRUPTIBLE);
1873                 mutex_unlock(&root->log_mutex);
1874                 if (atomic_read(&root->log_writers))
1875                         schedule();
1876                 mutex_lock(&root->log_mutex);
1877                 finish_wait(&root->log_writer_wait, &wait);
1878         }
1879         return 0;
1880 }
1881
1882 /*
1883  * btrfs_sync_log does sends a given tree log down to the disk and
1884  * updates the super blocks to record it.  When this call is done,
1885  * you know that any inodes previously logged are safely on disk
1886  */
1887 int btrfs_sync_log(struct btrfs_trans_handle *trans,
1888                    struct btrfs_root *root)
1889 {
1890         int index1;
1891         int index2;
1892         int ret;
1893         struct btrfs_root *log = root->log_root;
1894         struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
1895
1896         mutex_lock(&root->log_mutex);
1897         index1 = root->log_transid % 2;
1898         if (atomic_read(&root->log_commit[index1])) {
1899                 wait_log_commit(root, root->log_transid);
1900                 mutex_unlock(&root->log_mutex);
1901                 return 0;
1902         }
1903         atomic_set(&root->log_commit[index1], 1);
1904
1905         /* wait for previous tree log sync to complete */
1906         if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
1907                 wait_log_commit(root, root->log_transid - 1);
1908
1909         while (1) {
1910                 unsigned long batch = root->log_batch;
1911                 mutex_unlock(&root->log_mutex);
1912                 schedule_timeout_uninterruptible(1);
1913                 mutex_lock(&root->log_mutex);
1914                 wait_for_writer(root);
1915                 if (batch == root->log_batch)
1916                         break;
1917         }
1918
1919         ret = btrfs_write_and_wait_marked_extents(log, &log->dirty_log_pages);
1920         BUG_ON(ret);
1921
1922         btrfs_set_root_bytenr(&log->root_item, log->node->start);
1923         btrfs_set_root_generation(&log->root_item, trans->transid);
1924         btrfs_set_root_level(&log->root_item, btrfs_header_level(log->node));
1925
1926         root->log_batch = 0;
1927         root->log_transid++;
1928         log->log_transid = root->log_transid;
1929         smp_mb();
1930         /*
1931          * log tree has been flushed to disk, new modifications of
1932          * the log will be written to new positions. so it's safe to
1933          * allow log writers to go in.
1934          */
1935         mutex_unlock(&root->log_mutex);
1936
1937         mutex_lock(&log_root_tree->log_mutex);
1938         log_root_tree->log_batch++;
1939         atomic_inc(&log_root_tree->log_writers);
1940         mutex_unlock(&log_root_tree->log_mutex);
1941
1942         ret = update_log_root(trans, log);
1943         BUG_ON(ret);
1944
1945         mutex_lock(&log_root_tree->log_mutex);
1946         if (atomic_dec_and_test(&log_root_tree->log_writers)) {
1947                 smp_mb();
1948                 if (waitqueue_active(&log_root_tree->log_writer_wait))
1949                         wake_up(&log_root_tree->log_writer_wait);
1950         }
1951
1952         index2 = log_root_tree->log_transid % 2;
1953         if (atomic_read(&log_root_tree->log_commit[index2])) {
1954                 wait_log_commit(log_root_tree, log_root_tree->log_transid);
1955                 mutex_unlock(&log_root_tree->log_mutex);
1956                 goto out;
1957         }
1958         atomic_set(&log_root_tree->log_commit[index2], 1);
1959
1960         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2]))
1961                 wait_log_commit(log_root_tree, log_root_tree->log_transid - 1);
1962
1963         wait_for_writer(log_root_tree);
1964
1965         ret = btrfs_write_and_wait_marked_extents(log_root_tree,
1966                                 &log_root_tree->dirty_log_pages);
1967         BUG_ON(ret);
1968
1969         btrfs_set_super_log_root(&root->fs_info->super_for_commit,
1970                                 log_root_tree->node->start);
1971         btrfs_set_super_log_root_level(&root->fs_info->super_for_commit,
1972                                 btrfs_header_level(log_root_tree->node));
1973
1974         log_root_tree->log_batch = 0;
1975         log_root_tree->log_transid++;
1976         smp_mb();
1977
1978         mutex_unlock(&log_root_tree->log_mutex);
1979
1980         /*
1981          * nobody else is going to jump in and write the the ctree
1982          * super here because the log_commit atomic below is protecting
1983          * us.  We must be called with a transaction handle pinning
1984          * the running transaction open, so a full commit can't hop
1985          * in and cause problems either.
1986          */
1987         write_ctree_super(trans, root->fs_info->tree_root, 2);
1988
1989         atomic_set(&log_root_tree->log_commit[index2], 0);
1990         smp_mb();
1991         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
1992                 wake_up(&log_root_tree->log_commit_wait[index2]);
1993 out:
1994         atomic_set(&root->log_commit[index1], 0);
1995         smp_mb();
1996         if (waitqueue_active(&root->log_commit_wait[index1]))
1997                 wake_up(&root->log_commit_wait[index1]);
1998         return 0;
1999 }
2000
2001 /* * free all the extents used by the tree log.  This should be called
2002  * at commit time of the full transaction
2003  */
2004 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2005 {
2006         int ret;
2007         struct btrfs_root *log;
2008         struct key;
2009         u64 start;
2010         u64 end;
2011         struct walk_control wc = {
2012                 .free = 1,
2013                 .process_func = process_one_buffer
2014         };
2015
2016         if (!root->log_root || root->fs_info->log_root_recovering)
2017                 return 0;
2018
2019         log = root->log_root;
2020         ret = walk_log_tree(trans, log, &wc);
2021         BUG_ON(ret);
2022
2023         while (1) {
2024                 ret = find_first_extent_bit(&log->dirty_log_pages,
2025                                     0, &start, &end, EXTENT_DIRTY);
2026                 if (ret)
2027                         break;
2028
2029                 clear_extent_dirty(&log->dirty_log_pages,
2030                                    start, end, GFP_NOFS);
2031         }
2032
2033         if (log->log_transid > 0) {
2034                 ret = btrfs_del_root(trans, root->fs_info->log_root_tree,
2035                                      &log->root_key);
2036                 BUG_ON(ret);
2037         }
2038         root->log_root = NULL;
2039         free_extent_buffer(log->node);
2040         kfree(log);
2041         return 0;
2042 }
2043
2044 /*
2045  * If both a file and directory are logged, and unlinks or renames are
2046  * mixed in, we have a few interesting corners:
2047  *
2048  * create file X in dir Y
2049  * link file X to X.link in dir Y
2050  * fsync file X
2051  * unlink file X but leave X.link
2052  * fsync dir Y
2053  *
2054  * After a crash we would expect only X.link to exist.  But file X
2055  * didn't get fsync'd again so the log has back refs for X and X.link.
2056  *
2057  * We solve this by removing directory entries and inode backrefs from the
2058  * log when a file that was logged in the current transaction is
2059  * unlinked.  Any later fsync will include the updated log entries, and
2060  * we'll be able to reconstruct the proper directory items from backrefs.
2061  *
2062  * This optimizations allows us to avoid relogging the entire inode
2063  * or the entire directory.
2064  */
2065 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2066                                  struct btrfs_root *root,
2067                                  const char *name, int name_len,
2068                                  struct inode *dir, u64 index)
2069 {
2070         struct btrfs_root *log;
2071         struct btrfs_dir_item *di;
2072         struct btrfs_path *path;
2073         int ret;
2074         int bytes_del = 0;
2075
2076         if (BTRFS_I(dir)->logged_trans < trans->transid)
2077                 return 0;
2078
2079         ret = join_running_log_trans(root);
2080         if (ret)
2081                 return 0;
2082
2083         mutex_lock(&BTRFS_I(dir)->log_mutex);
2084
2085         log = root->log_root;
2086         path = btrfs_alloc_path();
2087         di = btrfs_lookup_dir_item(trans, log, path, dir->i_ino,
2088                                    name, name_len, -1);
2089         if (di && !IS_ERR(di)) {
2090                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2091                 bytes_del += name_len;
2092                 BUG_ON(ret);
2093         }
2094         btrfs_release_path(log, path);
2095         di = btrfs_lookup_dir_index_item(trans, log, path, dir->i_ino,
2096                                          index, name, name_len, -1);
2097         if (di && !IS_ERR(di)) {
2098                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2099                 bytes_del += name_len;
2100                 BUG_ON(ret);
2101         }
2102
2103         /* update the directory size in the log to reflect the names
2104          * we have removed
2105          */
2106         if (bytes_del) {
2107                 struct btrfs_key key;
2108
2109                 key.objectid = dir->i_ino;
2110                 key.offset = 0;
2111                 key.type = BTRFS_INODE_ITEM_KEY;
2112                 btrfs_release_path(log, path);
2113
2114                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2115                 if (ret == 0) {
2116                         struct btrfs_inode_item *item;
2117                         u64 i_size;
2118
2119                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2120                                               struct btrfs_inode_item);
2121                         i_size = btrfs_inode_size(path->nodes[0], item);
2122                         if (i_size > bytes_del)
2123                                 i_size -= bytes_del;
2124                         else
2125                                 i_size = 0;
2126                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2127                         btrfs_mark_buffer_dirty(path->nodes[0]);
2128                 } else
2129                         ret = 0;
2130                 btrfs_release_path(log, path);
2131         }
2132
2133         btrfs_free_path(path);
2134         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2135         end_log_trans(root);
2136
2137         return 0;
2138 }
2139
2140 /* see comments for btrfs_del_dir_entries_in_log */
2141 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2142                                struct btrfs_root *root,
2143                                const char *name, int name_len,
2144                                struct inode *inode, u64 dirid)
2145 {
2146         struct btrfs_root *log;
2147         u64 index;
2148         int ret;
2149
2150         if (BTRFS_I(inode)->logged_trans < trans->transid)
2151                 return 0;
2152
2153         ret = join_running_log_trans(root);
2154         if (ret)
2155                 return 0;
2156         log = root->log_root;
2157         mutex_lock(&BTRFS_I(inode)->log_mutex);
2158
2159         ret = btrfs_del_inode_ref(trans, log, name, name_len, inode->i_ino,
2160                                   dirid, &index);
2161         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2162         end_log_trans(root);
2163
2164         return ret;
2165 }
2166
2167 /*
2168  * creates a range item in the log for 'dirid'.  first_offset and
2169  * last_offset tell us which parts of the key space the log should
2170  * be considered authoritative for.
2171  */
2172 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2173                                        struct btrfs_root *log,
2174                                        struct btrfs_path *path,
2175                                        int key_type, u64 dirid,
2176                                        u64 first_offset, u64 last_offset)
2177 {
2178         int ret;
2179         struct btrfs_key key;
2180         struct btrfs_dir_log_item *item;
2181
2182         key.objectid = dirid;
2183         key.offset = first_offset;
2184         if (key_type == BTRFS_DIR_ITEM_KEY)
2185                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2186         else
2187                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2188         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2189         BUG_ON(ret);
2190
2191         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2192                               struct btrfs_dir_log_item);
2193         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2194         btrfs_mark_buffer_dirty(path->nodes[0]);
2195         btrfs_release_path(log, path);
2196         return 0;
2197 }
2198
2199 /*
2200  * log all the items included in the current transaction for a given
2201  * directory.  This also creates the range items in the log tree required
2202  * to replay anything deleted before the fsync
2203  */
2204 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2205                           struct btrfs_root *root, struct inode *inode,
2206                           struct btrfs_path *path,
2207                           struct btrfs_path *dst_path, int key_type,
2208                           u64 min_offset, u64 *last_offset_ret)
2209 {
2210         struct btrfs_key min_key;
2211         struct btrfs_key max_key;
2212         struct btrfs_root *log = root->log_root;
2213         struct extent_buffer *src;
2214         int ret;
2215         int i;
2216         int nritems;
2217         u64 first_offset = min_offset;
2218         u64 last_offset = (u64)-1;
2219
2220         log = root->log_root;
2221         max_key.objectid = inode->i_ino;
2222         max_key.offset = (u64)-1;
2223         max_key.type = key_type;
2224
2225         min_key.objectid = inode->i_ino;
2226         min_key.type = key_type;
2227         min_key.offset = min_offset;
2228
2229         path->keep_locks = 1;
2230
2231         ret = btrfs_search_forward(root, &min_key, &max_key,
2232                                    path, 0, trans->transid);
2233
2234         /*
2235          * we didn't find anything from this transaction, see if there
2236          * is anything at all
2237          */
2238         if (ret != 0 || min_key.objectid != inode->i_ino ||
2239             min_key.type != key_type) {
2240                 min_key.objectid = inode->i_ino;
2241                 min_key.type = key_type;
2242                 min_key.offset = (u64)-1;
2243                 btrfs_release_path(root, path);
2244                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2245                 if (ret < 0) {
2246                         btrfs_release_path(root, path);
2247                         return ret;
2248                 }
2249                 ret = btrfs_previous_item(root, path, inode->i_ino, key_type);
2250
2251                 /* if ret == 0 there are items for this type,
2252                  * create a range to tell us the last key of this type.
2253                  * otherwise, there are no items in this directory after
2254                  * *min_offset, and we create a range to indicate that.
2255                  */
2256                 if (ret == 0) {
2257                         struct btrfs_key tmp;
2258                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2259                                               path->slots[0]);
2260                         if (key_type == tmp.type)
2261                                 first_offset = max(min_offset, tmp.offset) + 1;
2262                 }
2263                 goto done;
2264         }
2265
2266         /* go backward to find any previous key */
2267         ret = btrfs_previous_item(root, path, inode->i_ino, key_type);
2268         if (ret == 0) {
2269                 struct btrfs_key tmp;
2270                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2271                 if (key_type == tmp.type) {
2272                         first_offset = tmp.offset;
2273                         ret = overwrite_item(trans, log, dst_path,
2274                                              path->nodes[0], path->slots[0],
2275                                              &tmp);
2276                 }
2277         }
2278         btrfs_release_path(root, path);
2279
2280         /* find the first key from this transaction again */
2281         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2282         if (ret != 0) {
2283                 WARN_ON(1);
2284                 goto done;
2285         }
2286
2287         /*
2288          * we have a block from this transaction, log every item in it
2289          * from our directory
2290          */
2291         while (1) {
2292                 struct btrfs_key tmp;
2293                 src = path->nodes[0];
2294                 nritems = btrfs_header_nritems(src);
2295                 for (i = path->slots[0]; i < nritems; i++) {
2296                         btrfs_item_key_to_cpu(src, &min_key, i);
2297
2298                         if (min_key.objectid != inode->i_ino ||
2299                             min_key.type != key_type)
2300                                 goto done;
2301                         ret = overwrite_item(trans, log, dst_path, src, i,
2302                                              &min_key);
2303                         BUG_ON(ret);
2304                 }
2305                 path->slots[0] = nritems;
2306
2307                 /*
2308                  * look ahead to the next item and see if it is also
2309                  * from this directory and from this transaction
2310                  */
2311                 ret = btrfs_next_leaf(root, path);
2312                 if (ret == 1) {
2313                         last_offset = (u64)-1;
2314                         goto done;
2315                 }
2316                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2317                 if (tmp.objectid != inode->i_ino || tmp.type != key_type) {
2318                         last_offset = (u64)-1;
2319                         goto done;
2320                 }
2321                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2322                         ret = overwrite_item(trans, log, dst_path,
2323                                              path->nodes[0], path->slots[0],
2324                                              &tmp);
2325
2326                         BUG_ON(ret);
2327                         last_offset = tmp.offset;
2328                         goto done;
2329                 }
2330         }
2331 done:
2332         *last_offset_ret = last_offset;
2333         btrfs_release_path(root, path);
2334         btrfs_release_path(log, dst_path);
2335
2336         /* insert the log range keys to indicate where the log is valid */
2337         ret = insert_dir_log_key(trans, log, path, key_type, inode->i_ino,
2338                                  first_offset, last_offset);
2339         BUG_ON(ret);
2340         return 0;
2341 }
2342
2343 /*
2344  * logging directories is very similar to logging inodes, We find all the items
2345  * from the current transaction and write them to the log.
2346  *
2347  * The recovery code scans the directory in the subvolume, and if it finds a
2348  * key in the range logged that is not present in the log tree, then it means
2349  * that dir entry was unlinked during the transaction.
2350  *
2351  * In order for that scan to work, we must include one key smaller than
2352  * the smallest logged by this transaction and one key larger than the largest
2353  * key logged by this transaction.
2354  */
2355 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
2356                           struct btrfs_root *root, struct inode *inode,
2357                           struct btrfs_path *path,
2358                           struct btrfs_path *dst_path)
2359 {
2360         u64 min_key;
2361         u64 max_key;
2362         int ret;
2363         int key_type = BTRFS_DIR_ITEM_KEY;
2364
2365 again:
2366         min_key = 0;
2367         max_key = 0;
2368         while (1) {
2369                 ret = log_dir_items(trans, root, inode, path,
2370                                     dst_path, key_type, min_key,
2371                                     &max_key);
2372                 BUG_ON(ret);
2373                 if (max_key == (u64)-1)
2374                         break;
2375                 min_key = max_key + 1;
2376         }
2377
2378         if (key_type == BTRFS_DIR_ITEM_KEY) {
2379                 key_type = BTRFS_DIR_INDEX_KEY;
2380                 goto again;
2381         }
2382         return 0;
2383 }
2384
2385 /*
2386  * a helper function to drop items from the log before we relog an
2387  * inode.  max_key_type indicates the highest item type to remove.
2388  * This cannot be run for file data extents because it does not
2389  * free the extents they point to.
2390  */
2391 static int drop_objectid_items(struct btrfs_trans_handle *trans,
2392                                   struct btrfs_root *log,
2393                                   struct btrfs_path *path,
2394                                   u64 objectid, int max_key_type)
2395 {
2396         int ret;
2397         struct btrfs_key key;
2398         struct btrfs_key found_key;
2399
2400         key.objectid = objectid;
2401         key.type = max_key_type;
2402         key.offset = (u64)-1;
2403
2404         while (1) {
2405                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
2406
2407                 if (ret != 1)
2408                         break;
2409
2410                 if (path->slots[0] == 0)
2411                         break;
2412
2413                 path->slots[0]--;
2414                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2415                                       path->slots[0]);
2416
2417                 if (found_key.objectid != objectid)
2418                         break;
2419
2420                 ret = btrfs_del_item(trans, log, path);
2421                 BUG_ON(ret);
2422                 btrfs_release_path(log, path);
2423         }
2424         btrfs_release_path(log, path);
2425         return 0;
2426 }
2427
2428 static noinline int copy_items(struct btrfs_trans_handle *trans,
2429                                struct btrfs_root *log,
2430                                struct btrfs_path *dst_path,
2431                                struct extent_buffer *src,
2432                                int start_slot, int nr, int inode_only)
2433 {
2434         unsigned long src_offset;
2435         unsigned long dst_offset;
2436         struct btrfs_file_extent_item *extent;
2437         struct btrfs_inode_item *inode_item;
2438         int ret;
2439         struct btrfs_key *ins_keys;
2440         u32 *ins_sizes;
2441         char *ins_data;
2442         int i;
2443         struct list_head ordered_sums;
2444
2445         INIT_LIST_HEAD(&ordered_sums);
2446
2447         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
2448                            nr * sizeof(u32), GFP_NOFS);
2449         ins_sizes = (u32 *)ins_data;
2450         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
2451
2452         for (i = 0; i < nr; i++) {
2453                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
2454                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
2455         }
2456         ret = btrfs_insert_empty_items(trans, log, dst_path,
2457                                        ins_keys, ins_sizes, nr);
2458         BUG_ON(ret);
2459
2460         for (i = 0; i < nr; i++) {
2461                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
2462                                                    dst_path->slots[0]);
2463
2464                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
2465
2466                 copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
2467                                    src_offset, ins_sizes[i]);
2468
2469                 if (inode_only == LOG_INODE_EXISTS &&
2470                     ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
2471                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
2472                                                     dst_path->slots[0],
2473                                                     struct btrfs_inode_item);
2474                         btrfs_set_inode_size(dst_path->nodes[0], inode_item, 0);
2475
2476                         /* set the generation to zero so the recover code
2477                          * can tell the difference between an logging
2478                          * just to say 'this inode exists' and a logging
2479                          * to say 'update this inode with these values'
2480                          */
2481                         btrfs_set_inode_generation(dst_path->nodes[0],
2482                                                    inode_item, 0);
2483                 }
2484                 /* take a reference on file data extents so that truncates
2485                  * or deletes of this inode don't have to relog the inode
2486                  * again
2487                  */
2488                 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY) {
2489                         int found_type;
2490                         extent = btrfs_item_ptr(src, start_slot + i,
2491                                                 struct btrfs_file_extent_item);
2492
2493                         found_type = btrfs_file_extent_type(src, extent);
2494                         if (found_type == BTRFS_FILE_EXTENT_REG ||
2495                             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
2496                                 u64 ds = btrfs_file_extent_disk_bytenr(src,
2497                                                                    extent);
2498                                 u64 dl = btrfs_file_extent_disk_num_bytes(src,
2499                                                                       extent);
2500                                 u64 cs = btrfs_file_extent_offset(src, extent);
2501                                 u64 cl = btrfs_file_extent_num_bytes(src,
2502                                                                      extent);;
2503                                 if (btrfs_file_extent_compression(src,
2504                                                                   extent)) {
2505                                         cs = 0;
2506                                         cl = dl;
2507                                 }
2508                                 /* ds == 0 is a hole */
2509                                 if (ds != 0) {
2510                                         ret = btrfs_inc_extent_ref(trans, log,
2511                                                    ds, dl,
2512                                                    dst_path->nodes[0]->start,
2513                                                    BTRFS_TREE_LOG_OBJECTID,
2514                                                    trans->transid,
2515                                                    ins_keys[i].objectid);
2516                                         BUG_ON(ret);
2517                                         ret = btrfs_lookup_csums_range(
2518                                                    log->fs_info->csum_root,
2519                                                    ds + cs, ds + cs + cl - 1,
2520                                                    &ordered_sums);
2521                                         BUG_ON(ret);
2522                                 }
2523                         }
2524                 }
2525                 dst_path->slots[0]++;
2526         }
2527
2528         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
2529         btrfs_release_path(log, dst_path);
2530         kfree(ins_data);
2531
2532         /*
2533          * we have to do this after the loop above to avoid changing the
2534          * log tree while trying to change the log tree.
2535          */
2536         while (!list_empty(&ordered_sums)) {
2537                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
2538                                                    struct btrfs_ordered_sum,
2539                                                    list);
2540                 ret = btrfs_csum_file_blocks(trans, log, sums);
2541                 BUG_ON(ret);
2542                 list_del(&sums->list);
2543                 kfree(sums);
2544         }
2545         return 0;
2546 }
2547
2548 /* log a single inode in the tree log.
2549  * At least one parent directory for this inode must exist in the tree
2550  * or be logged already.
2551  *
2552  * Any items from this inode changed by the current transaction are copied
2553  * to the log tree.  An extra reference is taken on any extents in this
2554  * file, allowing us to avoid a whole pile of corner cases around logging
2555  * blocks that have been removed from the tree.
2556  *
2557  * See LOG_INODE_ALL and related defines for a description of what inode_only
2558  * does.
2559  *
2560  * This handles both files and directories.
2561  */
2562 static int __btrfs_log_inode(struct btrfs_trans_handle *trans,
2563                              struct btrfs_root *root, struct inode *inode,
2564                              int inode_only)
2565 {
2566         struct btrfs_path *path;
2567         struct btrfs_path *dst_path;
2568         struct btrfs_key min_key;
2569         struct btrfs_key max_key;
2570         struct btrfs_root *log = root->log_root;
2571         struct extent_buffer *src = NULL;
2572         u32 size;
2573         int ret;
2574         int nritems;
2575         int ins_start_slot = 0;
2576         int ins_nr;
2577
2578         log = root->log_root;
2579
2580         path = btrfs_alloc_path();
2581         dst_path = btrfs_alloc_path();
2582
2583         min_key.objectid = inode->i_ino;
2584         min_key.type = BTRFS_INODE_ITEM_KEY;
2585         min_key.offset = 0;
2586
2587         max_key.objectid = inode->i_ino;
2588         if (inode_only == LOG_INODE_EXISTS || S_ISDIR(inode->i_mode))
2589                 max_key.type = BTRFS_XATTR_ITEM_KEY;
2590         else
2591                 max_key.type = (u8)-1;
2592         max_key.offset = (u64)-1;
2593
2594         /*
2595          * if this inode has already been logged and we're in inode_only
2596          * mode, we don't want to delete the things that have already
2597          * been written to the log.
2598          *
2599          * But, if the inode has been through an inode_only log,
2600          * the logged_trans field is not set.  This allows us to catch
2601          * any new names for this inode in the backrefs by logging it
2602          * again
2603          */
2604         if (inode_only == LOG_INODE_EXISTS &&
2605             BTRFS_I(inode)->logged_trans == trans->transid) {
2606                 btrfs_free_path(path);
2607                 btrfs_free_path(dst_path);
2608                 goto out;
2609         }
2610         mutex_lock(&BTRFS_I(inode)->log_mutex);
2611
2612         /*
2613          * a brute force approach to making sure we get the most uptodate
2614          * copies of everything.
2615          */
2616         if (S_ISDIR(inode->i_mode)) {
2617                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
2618
2619                 if (inode_only == LOG_INODE_EXISTS)
2620                         max_key_type = BTRFS_XATTR_ITEM_KEY;
2621                 ret = drop_objectid_items(trans, log, path,
2622                                           inode->i_ino, max_key_type);
2623         } else {
2624                 ret = btrfs_truncate_inode_items(trans, log, inode, 0, 0);
2625         }
2626         BUG_ON(ret);
2627         path->keep_locks = 1;
2628
2629         while (1) {
2630                 ins_nr = 0;
2631                 ret = btrfs_search_forward(root, &min_key, &max_key,
2632                                            path, 0, trans->transid);
2633                 if (ret != 0)
2634                         break;
2635 again:
2636                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
2637                 if (min_key.objectid != inode->i_ino)
2638                         break;
2639                 if (min_key.type > max_key.type)
2640                         break;
2641
2642                 src = path->nodes[0];
2643                 size = btrfs_item_size_nr(src, path->slots[0]);
2644                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
2645                         ins_nr++;
2646                         goto next_slot;
2647                 } else if (!ins_nr) {
2648                         ins_start_slot = path->slots[0];
2649                         ins_nr = 1;
2650                         goto next_slot;
2651                 }
2652
2653                 ret = copy_items(trans, log, dst_path, src, ins_start_slot,
2654                                  ins_nr, inode_only);
2655                 BUG_ON(ret);
2656                 ins_nr = 1;
2657                 ins_start_slot = path->slots[0];
2658 next_slot:
2659
2660                 nritems = btrfs_header_nritems(path->nodes[0]);
2661                 path->slots[0]++;
2662                 if (path->slots[0] < nritems) {
2663                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
2664                                               path->slots[0]);
2665                         goto again;
2666                 }
2667                 if (ins_nr) {
2668                         ret = copy_items(trans, log, dst_path, src,
2669                                          ins_start_slot,
2670                                          ins_nr, inode_only);
2671                         BUG_ON(ret);
2672                         ins_nr = 0;
2673                 }
2674                 btrfs_release_path(root, path);
2675
2676                 if (min_key.offset < (u64)-1)
2677                         min_key.offset++;
2678                 else if (min_key.type < (u8)-1)
2679                         min_key.type++;
2680                 else if (min_key.objectid < (u64)-1)
2681                         min_key.objectid++;
2682                 else
2683                         break;
2684         }
2685         if (ins_nr) {
2686                 ret = copy_items(trans, log, dst_path, src,
2687                                  ins_start_slot,
2688                                  ins_nr, inode_only);
2689                 BUG_ON(ret);
2690                 ins_nr = 0;
2691         }
2692         WARN_ON(ins_nr);
2693         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
2694                 btrfs_release_path(root, path);
2695                 btrfs_release_path(log, dst_path);
2696                 BTRFS_I(inode)->log_dirty_trans = 0;
2697                 ret = log_directory_changes(trans, root, inode, path, dst_path);
2698                 BUG_ON(ret);
2699         }
2700         BTRFS_I(inode)->logged_trans = trans->transid;
2701         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2702
2703         btrfs_free_path(path);
2704         btrfs_free_path(dst_path);
2705 out:
2706         return 0;
2707 }
2708
2709 int btrfs_log_inode(struct btrfs_trans_handle *trans,
2710                     struct btrfs_root *root, struct inode *inode,
2711                     int inode_only)
2712 {
2713         int ret;
2714
2715         start_log_trans(trans, root);
2716         ret = __btrfs_log_inode(trans, root, inode, inode_only);
2717         end_log_trans(root);
2718         return ret;
2719 }
2720
2721 /*
2722  * helper function around btrfs_log_inode to make sure newly created
2723  * parent directories also end up in the log.  A minimal inode and backref
2724  * only logging is done of any parent directories that are older than
2725  * the last committed transaction
2726  */
2727 int btrfs_log_dentry(struct btrfs_trans_handle *trans,
2728                     struct btrfs_root *root, struct dentry *dentry)
2729 {
2730         int inode_only = LOG_INODE_ALL;
2731         struct super_block *sb;
2732         int ret;
2733
2734         start_log_trans(trans, root);
2735         sb = dentry->d_inode->i_sb;
2736         while (1) {
2737                 ret = __btrfs_log_inode(trans, root, dentry->d_inode,
2738                                         inode_only);
2739                 BUG_ON(ret);
2740                 inode_only = LOG_INODE_EXISTS;
2741
2742                 dentry = dentry->d_parent;
2743                 if (!dentry || !dentry->d_inode || sb != dentry->d_inode->i_sb)
2744                         break;
2745
2746                 if (BTRFS_I(dentry->d_inode)->generation <=
2747                     root->fs_info->last_trans_committed)
2748                         break;
2749         }
2750         end_log_trans(root);
2751         return 0;
2752 }
2753
2754 /*
2755  * it is not safe to log dentry if the chunk root has added new
2756  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
2757  * If this returns 1, you must commit the transaction to safely get your
2758  * data on disk.
2759  */
2760 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
2761                           struct btrfs_root *root, struct dentry *dentry)
2762 {
2763         u64 gen;
2764         gen = root->fs_info->last_trans_new_blockgroup;
2765         if (gen > root->fs_info->last_trans_committed)
2766                 return 1;
2767         else
2768                 return btrfs_log_dentry(trans, root, dentry);
2769 }
2770
2771 /*
2772  * should be called during mount to recover any replay any log trees
2773  * from the FS
2774  */
2775 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
2776 {
2777         int ret;
2778         struct btrfs_path *path;
2779         struct btrfs_trans_handle *trans;
2780         struct btrfs_key key;
2781         struct btrfs_key found_key;
2782         struct btrfs_key tmp_key;
2783         struct btrfs_root *log;
2784         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
2785         u64 highest_inode;
2786         struct walk_control wc = {
2787                 .process_func = process_one_buffer,
2788                 .stage = 0,
2789         };
2790
2791         fs_info->log_root_recovering = 1;
2792         path = btrfs_alloc_path();
2793         BUG_ON(!path);
2794
2795         trans = btrfs_start_transaction(fs_info->tree_root, 1);
2796
2797         wc.trans = trans;
2798         wc.pin = 1;
2799
2800         walk_log_tree(trans, log_root_tree, &wc);
2801
2802 again:
2803         key.objectid = BTRFS_TREE_LOG_OBJECTID;
2804         key.offset = (u64)-1;
2805         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
2806
2807         while (1) {
2808                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
2809                 if (ret < 0)
2810                         break;
2811                 if (ret > 0) {
2812                         if (path->slots[0] == 0)
2813                                 break;
2814                         path->slots[0]--;
2815                 }
2816                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2817                                       path->slots[0]);
2818                 btrfs_release_path(log_root_tree, path);
2819                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
2820                         break;
2821
2822                 log = btrfs_read_fs_root_no_radix(log_root_tree,
2823                                                   &found_key);
2824                 BUG_ON(!log);
2825
2826
2827                 tmp_key.objectid = found_key.offset;
2828                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
2829                 tmp_key.offset = (u64)-1;
2830
2831                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
2832                 BUG_ON(!wc.replay_dest);
2833
2834                 wc.replay_dest->log_root = log;
2835                 btrfs_record_root_in_trans(wc.replay_dest);
2836                 ret = walk_log_tree(trans, log, &wc);
2837                 BUG_ON(ret);
2838
2839                 if (wc.stage == LOG_WALK_REPLAY_ALL) {
2840                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
2841                                                       path);
2842                         BUG_ON(ret);
2843                 }
2844                 ret = btrfs_find_highest_inode(wc.replay_dest, &highest_inode);
2845                 if (ret == 0) {
2846                         wc.replay_dest->highest_inode = highest_inode;
2847                         wc.replay_dest->last_inode_alloc = highest_inode;
2848                 }
2849
2850                 key.offset = found_key.offset - 1;
2851                 wc.replay_dest->log_root = NULL;
2852                 free_extent_buffer(log->node);
2853                 kfree(log);
2854
2855                 if (found_key.offset == 0)
2856                         break;
2857         }
2858         btrfs_release_path(log_root_tree, path);
2859
2860         /* step one is to pin it all, step two is to replay just inodes */
2861         if (wc.pin) {
2862                 wc.pin = 0;
2863                 wc.process_func = replay_one_buffer;
2864                 wc.stage = LOG_WALK_REPLAY_INODES;
2865                 goto again;
2866         }
2867         /* step three is to replay everything */
2868         if (wc.stage < LOG_WALK_REPLAY_ALL) {
2869                 wc.stage++;
2870                 goto again;
2871         }
2872
2873         btrfs_free_path(path);
2874
2875         free_extent_buffer(log_root_tree->node);
2876         log_root_tree->log_root = NULL;
2877         fs_info->log_root_recovering = 0;
2878
2879         /* step 4: commit the transaction, which also unpins the blocks */
2880         btrfs_commit_transaction(trans, fs_info->tree_root);
2881
2882         kfree(log_root_tree);
2883         return 0;
2884 }