Btrfs: fix fsync xattr loss in the fast fsync path
[sfrench/cifs-2.6.git] / fs / btrfs / tree-log.c
1 /*
2  * Copyright (C) 2008 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/sched.h>
20 #include <linux/slab.h>
21 #include <linux/blkdev.h>
22 #include <linux/list_sort.h>
23 #include "tree-log.h"
24 #include "disk-io.h"
25 #include "locking.h"
26 #include "print-tree.h"
27 #include "backref.h"
28 #include "hash.h"
29
30 /* magic values for the inode_only field in btrfs_log_inode:
31  *
32  * LOG_INODE_ALL means to log everything
33  * LOG_INODE_EXISTS means to log just enough to recreate the inode
34  * during log replay
35  */
36 #define LOG_INODE_ALL 0
37 #define LOG_INODE_EXISTS 1
38
39 /*
40  * directory trouble cases
41  *
42  * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
43  * log, we must force a full commit before doing an fsync of the directory
44  * where the unlink was done.
45  * ---> record transid of last unlink/rename per directory
46  *
47  * mkdir foo/some_dir
48  * normal commit
49  * rename foo/some_dir foo2/some_dir
50  * mkdir foo/some_dir
51  * fsync foo/some_dir/some_file
52  *
53  * The fsync above will unlink the original some_dir without recording
54  * it in its new location (foo2).  After a crash, some_dir will be gone
55  * unless the fsync of some_file forces a full commit
56  *
57  * 2) we must log any new names for any file or dir that is in the fsync
58  * log. ---> check inode while renaming/linking.
59  *
60  * 2a) we must log any new names for any file or dir during rename
61  * when the directory they are being removed from was logged.
62  * ---> check inode and old parent dir during rename
63  *
64  *  2a is actually the more important variant.  With the extra logging
65  *  a crash might unlink the old name without recreating the new one
66  *
67  * 3) after a crash, we must go through any directories with a link count
68  * of zero and redo the rm -rf
69  *
70  * mkdir f1/foo
71  * normal commit
72  * rm -rf f1/foo
73  * fsync(f1)
74  *
75  * The directory f1 was fully removed from the FS, but fsync was never
76  * called on f1, only its parent dir.  After a crash the rm -rf must
77  * be replayed.  This must be able to recurse down the entire
78  * directory tree.  The inode link count fixup code takes care of the
79  * ugly details.
80  */
81
82 /*
83  * stages for the tree walking.  The first
84  * stage (0) is to only pin down the blocks we find
85  * the second stage (1) is to make sure that all the inodes
86  * we find in the log are created in the subvolume.
87  *
88  * The last stage is to deal with directories and links and extents
89  * and all the other fun semantics
90  */
91 #define LOG_WALK_PIN_ONLY 0
92 #define LOG_WALK_REPLAY_INODES 1
93 #define LOG_WALK_REPLAY_DIR_INDEX 2
94 #define LOG_WALK_REPLAY_ALL 3
95
96 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
97                            struct btrfs_root *root, struct inode *inode,
98                            int inode_only,
99                            const loff_t start,
100                            const loff_t end,
101                            struct btrfs_log_ctx *ctx);
102 static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
103                              struct btrfs_root *root,
104                              struct btrfs_path *path, u64 objectid);
105 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
106                                        struct btrfs_root *root,
107                                        struct btrfs_root *log,
108                                        struct btrfs_path *path,
109                                        u64 dirid, int del_all);
110
111 /*
112  * tree logging is a special write ahead log used to make sure that
113  * fsyncs and O_SYNCs can happen without doing full tree commits.
114  *
115  * Full tree commits are expensive because they require commonly
116  * modified blocks to be recowed, creating many dirty pages in the
117  * extent tree an 4x-6x higher write load than ext3.
118  *
119  * Instead of doing a tree commit on every fsync, we use the
120  * key ranges and transaction ids to find items for a given file or directory
121  * that have changed in this transaction.  Those items are copied into
122  * a special tree (one per subvolume root), that tree is written to disk
123  * and then the fsync is considered complete.
124  *
125  * After a crash, items are copied out of the log-tree back into the
126  * subvolume tree.  Any file data extents found are recorded in the extent
127  * allocation tree, and the log-tree freed.
128  *
129  * The log tree is read three times, once to pin down all the extents it is
130  * using in ram and once, once to create all the inodes logged in the tree
131  * and once to do all the other items.
132  */
133
134 /*
135  * start a sub transaction and setup the log tree
136  * this increments the log tree writer count to make the people
137  * syncing the tree wait for us to finish
138  */
139 static int start_log_trans(struct btrfs_trans_handle *trans,
140                            struct btrfs_root *root,
141                            struct btrfs_log_ctx *ctx)
142 {
143         int index;
144         int ret;
145
146         mutex_lock(&root->log_mutex);
147         if (root->log_root) {
148                 if (btrfs_need_log_full_commit(root->fs_info, trans)) {
149                         ret = -EAGAIN;
150                         goto out;
151                 }
152                 if (!root->log_start_pid) {
153                         root->log_start_pid = current->pid;
154                         clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
155                 } else if (root->log_start_pid != current->pid) {
156                         set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
157                 }
158
159                 atomic_inc(&root->log_batch);
160                 atomic_inc(&root->log_writers);
161                 if (ctx) {
162                         index = root->log_transid % 2;
163                         list_add_tail(&ctx->list, &root->log_ctxs[index]);
164                         ctx->log_transid = root->log_transid;
165                 }
166                 mutex_unlock(&root->log_mutex);
167                 return 0;
168         }
169
170         ret = 0;
171         mutex_lock(&root->fs_info->tree_log_mutex);
172         if (!root->fs_info->log_root_tree)
173                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
174         mutex_unlock(&root->fs_info->tree_log_mutex);
175         if (ret)
176                 goto out;
177
178         if (!root->log_root) {
179                 ret = btrfs_add_log_tree(trans, root);
180                 if (ret)
181                         goto out;
182         }
183         clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
184         root->log_start_pid = current->pid;
185         atomic_inc(&root->log_batch);
186         atomic_inc(&root->log_writers);
187         if (ctx) {
188                 index = root->log_transid % 2;
189                 list_add_tail(&ctx->list, &root->log_ctxs[index]);
190                 ctx->log_transid = root->log_transid;
191         }
192 out:
193         mutex_unlock(&root->log_mutex);
194         return ret;
195 }
196
197 /*
198  * returns 0 if there was a log transaction running and we were able
199  * to join, or returns -ENOENT if there were not transactions
200  * in progress
201  */
202 static int join_running_log_trans(struct btrfs_root *root)
203 {
204         int ret = -ENOENT;
205
206         smp_mb();
207         if (!root->log_root)
208                 return -ENOENT;
209
210         mutex_lock(&root->log_mutex);
211         if (root->log_root) {
212                 ret = 0;
213                 atomic_inc(&root->log_writers);
214         }
215         mutex_unlock(&root->log_mutex);
216         return ret;
217 }
218
219 /*
220  * This either makes the current running log transaction wait
221  * until you call btrfs_end_log_trans() or it makes any future
222  * log transactions wait until you call btrfs_end_log_trans()
223  */
224 int btrfs_pin_log_trans(struct btrfs_root *root)
225 {
226         int ret = -ENOENT;
227
228         mutex_lock(&root->log_mutex);
229         atomic_inc(&root->log_writers);
230         mutex_unlock(&root->log_mutex);
231         return ret;
232 }
233
234 /*
235  * indicate we're done making changes to the log tree
236  * and wake up anyone waiting to do a sync
237  */
238 void btrfs_end_log_trans(struct btrfs_root *root)
239 {
240         if (atomic_dec_and_test(&root->log_writers)) {
241                 smp_mb();
242                 if (waitqueue_active(&root->log_writer_wait))
243                         wake_up(&root->log_writer_wait);
244         }
245 }
246
247
248 /*
249  * the walk control struct is used to pass state down the chain when
250  * processing the log tree.  The stage field tells us which part
251  * of the log tree processing we are currently doing.  The others
252  * are state fields used for that specific part
253  */
254 struct walk_control {
255         /* should we free the extent on disk when done?  This is used
256          * at transaction commit time while freeing a log tree
257          */
258         int free;
259
260         /* should we write out the extent buffer?  This is used
261          * while flushing the log tree to disk during a sync
262          */
263         int write;
264
265         /* should we wait for the extent buffer io to finish?  Also used
266          * while flushing the log tree to disk for a sync
267          */
268         int wait;
269
270         /* pin only walk, we record which extents on disk belong to the
271          * log trees
272          */
273         int pin;
274
275         /* what stage of the replay code we're currently in */
276         int stage;
277
278         /* the root we are currently replaying */
279         struct btrfs_root *replay_dest;
280
281         /* the trans handle for the current replay */
282         struct btrfs_trans_handle *trans;
283
284         /* the function that gets used to process blocks we find in the
285          * tree.  Note the extent_buffer might not be up to date when it is
286          * passed in, and it must be checked or read if you need the data
287          * inside it
288          */
289         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
290                             struct walk_control *wc, u64 gen);
291 };
292
293 /*
294  * process_func used to pin down extents, write them or wait on them
295  */
296 static int process_one_buffer(struct btrfs_root *log,
297                               struct extent_buffer *eb,
298                               struct walk_control *wc, u64 gen)
299 {
300         int ret = 0;
301
302         /*
303          * If this fs is mixed then we need to be able to process the leaves to
304          * pin down any logged extents, so we have to read the block.
305          */
306         if (btrfs_fs_incompat(log->fs_info, MIXED_GROUPS)) {
307                 ret = btrfs_read_buffer(eb, gen);
308                 if (ret)
309                         return ret;
310         }
311
312         if (wc->pin)
313                 ret = btrfs_pin_extent_for_log_replay(log->fs_info->extent_root,
314                                                       eb->start, eb->len);
315
316         if (!ret && btrfs_buffer_uptodate(eb, gen, 0)) {
317                 if (wc->pin && btrfs_header_level(eb) == 0)
318                         ret = btrfs_exclude_logged_extents(log, eb);
319                 if (wc->write)
320                         btrfs_write_tree_block(eb);
321                 if (wc->wait)
322                         btrfs_wait_tree_block_writeback(eb);
323         }
324         return ret;
325 }
326
327 /*
328  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
329  * to the src data we are copying out.
330  *
331  * root is the tree we are copying into, and path is a scratch
332  * path for use in this function (it should be released on entry and
333  * will be released on exit).
334  *
335  * If the key is already in the destination tree the existing item is
336  * overwritten.  If the existing item isn't big enough, it is extended.
337  * If it is too large, it is truncated.
338  *
339  * If the key isn't in the destination yet, a new item is inserted.
340  */
341 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
342                                    struct btrfs_root *root,
343                                    struct btrfs_path *path,
344                                    struct extent_buffer *eb, int slot,
345                                    struct btrfs_key *key)
346 {
347         int ret;
348         u32 item_size;
349         u64 saved_i_size = 0;
350         int save_old_i_size = 0;
351         unsigned long src_ptr;
352         unsigned long dst_ptr;
353         int overwrite_root = 0;
354         bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
355
356         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
357                 overwrite_root = 1;
358
359         item_size = btrfs_item_size_nr(eb, slot);
360         src_ptr = btrfs_item_ptr_offset(eb, slot);
361
362         /* look for the key in the destination tree */
363         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
364         if (ret < 0)
365                 return ret;
366
367         if (ret == 0) {
368                 char *src_copy;
369                 char *dst_copy;
370                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
371                                                   path->slots[0]);
372                 if (dst_size != item_size)
373                         goto insert;
374
375                 if (item_size == 0) {
376                         btrfs_release_path(path);
377                         return 0;
378                 }
379                 dst_copy = kmalloc(item_size, GFP_NOFS);
380                 src_copy = kmalloc(item_size, GFP_NOFS);
381                 if (!dst_copy || !src_copy) {
382                         btrfs_release_path(path);
383                         kfree(dst_copy);
384                         kfree(src_copy);
385                         return -ENOMEM;
386                 }
387
388                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
389
390                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
391                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
392                                    item_size);
393                 ret = memcmp(dst_copy, src_copy, item_size);
394
395                 kfree(dst_copy);
396                 kfree(src_copy);
397                 /*
398                  * they have the same contents, just return, this saves
399                  * us from cowing blocks in the destination tree and doing
400                  * extra writes that may not have been done by a previous
401                  * sync
402                  */
403                 if (ret == 0) {
404                         btrfs_release_path(path);
405                         return 0;
406                 }
407
408                 /*
409                  * We need to load the old nbytes into the inode so when we
410                  * replay the extents we've logged we get the right nbytes.
411                  */
412                 if (inode_item) {
413                         struct btrfs_inode_item *item;
414                         u64 nbytes;
415                         u32 mode;
416
417                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
418                                               struct btrfs_inode_item);
419                         nbytes = btrfs_inode_nbytes(path->nodes[0], item);
420                         item = btrfs_item_ptr(eb, slot,
421                                               struct btrfs_inode_item);
422                         btrfs_set_inode_nbytes(eb, item, nbytes);
423
424                         /*
425                          * If this is a directory we need to reset the i_size to
426                          * 0 so that we can set it up properly when replaying
427                          * the rest of the items in this log.
428                          */
429                         mode = btrfs_inode_mode(eb, item);
430                         if (S_ISDIR(mode))
431                                 btrfs_set_inode_size(eb, item, 0);
432                 }
433         } else if (inode_item) {
434                 struct btrfs_inode_item *item;
435                 u32 mode;
436
437                 /*
438                  * New inode, set nbytes to 0 so that the nbytes comes out
439                  * properly when we replay the extents.
440                  */
441                 item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
442                 btrfs_set_inode_nbytes(eb, item, 0);
443
444                 /*
445                  * If this is a directory we need to reset the i_size to 0 so
446                  * that we can set it up properly when replaying the rest of
447                  * the items in this log.
448                  */
449                 mode = btrfs_inode_mode(eb, item);
450                 if (S_ISDIR(mode))
451                         btrfs_set_inode_size(eb, item, 0);
452         }
453 insert:
454         btrfs_release_path(path);
455         /* try to insert the key into the destination tree */
456         path->skip_release_on_error = 1;
457         ret = btrfs_insert_empty_item(trans, root, path,
458                                       key, item_size);
459         path->skip_release_on_error = 0;
460
461         /* make sure any existing item is the correct size */
462         if (ret == -EEXIST || ret == -EOVERFLOW) {
463                 u32 found_size;
464                 found_size = btrfs_item_size_nr(path->nodes[0],
465                                                 path->slots[0]);
466                 if (found_size > item_size)
467                         btrfs_truncate_item(root, path, item_size, 1);
468                 else if (found_size < item_size)
469                         btrfs_extend_item(root, path,
470                                           item_size - found_size);
471         } else if (ret) {
472                 return ret;
473         }
474         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
475                                         path->slots[0]);
476
477         /* don't overwrite an existing inode if the generation number
478          * was logged as zero.  This is done when the tree logging code
479          * is just logging an inode to make sure it exists after recovery.
480          *
481          * Also, don't overwrite i_size on directories during replay.
482          * log replay inserts and removes directory items based on the
483          * state of the tree found in the subvolume, and i_size is modified
484          * as it goes
485          */
486         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
487                 struct btrfs_inode_item *src_item;
488                 struct btrfs_inode_item *dst_item;
489
490                 src_item = (struct btrfs_inode_item *)src_ptr;
491                 dst_item = (struct btrfs_inode_item *)dst_ptr;
492
493                 if (btrfs_inode_generation(eb, src_item) == 0) {
494                         struct extent_buffer *dst_eb = path->nodes[0];
495                         const u64 ino_size = btrfs_inode_size(eb, src_item);
496
497                         /*
498                          * For regular files an ino_size == 0 is used only when
499                          * logging that an inode exists, as part of a directory
500                          * fsync, and the inode wasn't fsynced before. In this
501                          * case don't set the size of the inode in the fs/subvol
502                          * tree, otherwise we would be throwing valid data away.
503                          */
504                         if (S_ISREG(btrfs_inode_mode(eb, src_item)) &&
505                             S_ISREG(btrfs_inode_mode(dst_eb, dst_item)) &&
506                             ino_size != 0) {
507                                 struct btrfs_map_token token;
508
509                                 btrfs_init_map_token(&token);
510                                 btrfs_set_token_inode_size(dst_eb, dst_item,
511                                                            ino_size, &token);
512                         }
513                         goto no_copy;
514                 }
515
516                 if (overwrite_root &&
517                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
518                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
519                         save_old_i_size = 1;
520                         saved_i_size = btrfs_inode_size(path->nodes[0],
521                                                         dst_item);
522                 }
523         }
524
525         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
526                            src_ptr, item_size);
527
528         if (save_old_i_size) {
529                 struct btrfs_inode_item *dst_item;
530                 dst_item = (struct btrfs_inode_item *)dst_ptr;
531                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
532         }
533
534         /* make sure the generation is filled in */
535         if (key->type == BTRFS_INODE_ITEM_KEY) {
536                 struct btrfs_inode_item *dst_item;
537                 dst_item = (struct btrfs_inode_item *)dst_ptr;
538                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
539                         btrfs_set_inode_generation(path->nodes[0], dst_item,
540                                                    trans->transid);
541                 }
542         }
543 no_copy:
544         btrfs_mark_buffer_dirty(path->nodes[0]);
545         btrfs_release_path(path);
546         return 0;
547 }
548
549 /*
550  * simple helper to read an inode off the disk from a given root
551  * This can only be called for subvolume roots and not for the log
552  */
553 static noinline struct inode *read_one_inode(struct btrfs_root *root,
554                                              u64 objectid)
555 {
556         struct btrfs_key key;
557         struct inode *inode;
558
559         key.objectid = objectid;
560         key.type = BTRFS_INODE_ITEM_KEY;
561         key.offset = 0;
562         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
563         if (IS_ERR(inode)) {
564                 inode = NULL;
565         } else if (is_bad_inode(inode)) {
566                 iput(inode);
567                 inode = NULL;
568         }
569         return inode;
570 }
571
572 /* replays a single extent in 'eb' at 'slot' with 'key' into the
573  * subvolume 'root'.  path is released on entry and should be released
574  * on exit.
575  *
576  * extents in the log tree have not been allocated out of the extent
577  * tree yet.  So, this completes the allocation, taking a reference
578  * as required if the extent already exists or creating a new extent
579  * if it isn't in the extent allocation tree yet.
580  *
581  * The extent is inserted into the file, dropping any existing extents
582  * from the file that overlap the new one.
583  */
584 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
585                                       struct btrfs_root *root,
586                                       struct btrfs_path *path,
587                                       struct extent_buffer *eb, int slot,
588                                       struct btrfs_key *key)
589 {
590         int found_type;
591         u64 extent_end;
592         u64 start = key->offset;
593         u64 nbytes = 0;
594         struct btrfs_file_extent_item *item;
595         struct inode *inode = NULL;
596         unsigned long size;
597         int ret = 0;
598
599         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
600         found_type = btrfs_file_extent_type(eb, item);
601
602         if (found_type == BTRFS_FILE_EXTENT_REG ||
603             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
604                 nbytes = btrfs_file_extent_num_bytes(eb, item);
605                 extent_end = start + nbytes;
606
607                 /*
608                  * We don't add to the inodes nbytes if we are prealloc or a
609                  * hole.
610                  */
611                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
612                         nbytes = 0;
613         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
614                 size = btrfs_file_extent_inline_len(eb, slot, item);
615                 nbytes = btrfs_file_extent_ram_bytes(eb, item);
616                 extent_end = ALIGN(start + size, root->sectorsize);
617         } else {
618                 ret = 0;
619                 goto out;
620         }
621
622         inode = read_one_inode(root, key->objectid);
623         if (!inode) {
624                 ret = -EIO;
625                 goto out;
626         }
627
628         /*
629          * first check to see if we already have this extent in the
630          * file.  This must be done before the btrfs_drop_extents run
631          * so we don't try to drop this extent.
632          */
633         ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
634                                        start, 0);
635
636         if (ret == 0 &&
637             (found_type == BTRFS_FILE_EXTENT_REG ||
638              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
639                 struct btrfs_file_extent_item cmp1;
640                 struct btrfs_file_extent_item cmp2;
641                 struct btrfs_file_extent_item *existing;
642                 struct extent_buffer *leaf;
643
644                 leaf = path->nodes[0];
645                 existing = btrfs_item_ptr(leaf, path->slots[0],
646                                           struct btrfs_file_extent_item);
647
648                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
649                                    sizeof(cmp1));
650                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
651                                    sizeof(cmp2));
652
653                 /*
654                  * we already have a pointer to this exact extent,
655                  * we don't have to do anything
656                  */
657                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
658                         btrfs_release_path(path);
659                         goto out;
660                 }
661         }
662         btrfs_release_path(path);
663
664         /* drop any overlapping extents */
665         ret = btrfs_drop_extents(trans, root, inode, start, extent_end, 1);
666         if (ret)
667                 goto out;
668
669         if (found_type == BTRFS_FILE_EXTENT_REG ||
670             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
671                 u64 offset;
672                 unsigned long dest_offset;
673                 struct btrfs_key ins;
674
675                 ret = btrfs_insert_empty_item(trans, root, path, key,
676                                               sizeof(*item));
677                 if (ret)
678                         goto out;
679                 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
680                                                     path->slots[0]);
681                 copy_extent_buffer(path->nodes[0], eb, dest_offset,
682                                 (unsigned long)item,  sizeof(*item));
683
684                 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
685                 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
686                 ins.type = BTRFS_EXTENT_ITEM_KEY;
687                 offset = key->offset - btrfs_file_extent_offset(eb, item);
688
689                 if (ins.objectid > 0) {
690                         u64 csum_start;
691                         u64 csum_end;
692                         LIST_HEAD(ordered_sums);
693                         /*
694                          * is this extent already allocated in the extent
695                          * allocation tree?  If so, just add a reference
696                          */
697                         ret = btrfs_lookup_data_extent(root, ins.objectid,
698                                                 ins.offset);
699                         if (ret == 0) {
700                                 ret = btrfs_inc_extent_ref(trans, root,
701                                                 ins.objectid, ins.offset,
702                                                 0, root->root_key.objectid,
703                                                 key->objectid, offset, 0);
704                                 if (ret)
705                                         goto out;
706                         } else {
707                                 /*
708                                  * insert the extent pointer in the extent
709                                  * allocation tree
710                                  */
711                                 ret = btrfs_alloc_logged_file_extent(trans,
712                                                 root, root->root_key.objectid,
713                                                 key->objectid, offset, &ins);
714                                 if (ret)
715                                         goto out;
716                         }
717                         btrfs_release_path(path);
718
719                         if (btrfs_file_extent_compression(eb, item)) {
720                                 csum_start = ins.objectid;
721                                 csum_end = csum_start + ins.offset;
722                         } else {
723                                 csum_start = ins.objectid +
724                                         btrfs_file_extent_offset(eb, item);
725                                 csum_end = csum_start +
726                                         btrfs_file_extent_num_bytes(eb, item);
727                         }
728
729                         ret = btrfs_lookup_csums_range(root->log_root,
730                                                 csum_start, csum_end - 1,
731                                                 &ordered_sums, 0);
732                         if (ret)
733                                 goto out;
734                         while (!list_empty(&ordered_sums)) {
735                                 struct btrfs_ordered_sum *sums;
736                                 sums = list_entry(ordered_sums.next,
737                                                 struct btrfs_ordered_sum,
738                                                 list);
739                                 if (!ret)
740                                         ret = btrfs_csum_file_blocks(trans,
741                                                 root->fs_info->csum_root,
742                                                 sums);
743                                 list_del(&sums->list);
744                                 kfree(sums);
745                         }
746                         if (ret)
747                                 goto out;
748                 } else {
749                         btrfs_release_path(path);
750                 }
751         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
752                 /* inline extents are easy, we just overwrite them */
753                 ret = overwrite_item(trans, root, path, eb, slot, key);
754                 if (ret)
755                         goto out;
756         }
757
758         inode_add_bytes(inode, nbytes);
759         ret = btrfs_update_inode(trans, root, inode);
760 out:
761         if (inode)
762                 iput(inode);
763         return ret;
764 }
765
766 /*
767  * when cleaning up conflicts between the directory names in the
768  * subvolume, directory names in the log and directory names in the
769  * inode back references, we may have to unlink inodes from directories.
770  *
771  * This is a helper function to do the unlink of a specific directory
772  * item
773  */
774 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
775                                       struct btrfs_root *root,
776                                       struct btrfs_path *path,
777                                       struct inode *dir,
778                                       struct btrfs_dir_item *di)
779 {
780         struct inode *inode;
781         char *name;
782         int name_len;
783         struct extent_buffer *leaf;
784         struct btrfs_key location;
785         int ret;
786
787         leaf = path->nodes[0];
788
789         btrfs_dir_item_key_to_cpu(leaf, di, &location);
790         name_len = btrfs_dir_name_len(leaf, di);
791         name = kmalloc(name_len, GFP_NOFS);
792         if (!name)
793                 return -ENOMEM;
794
795         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
796         btrfs_release_path(path);
797
798         inode = read_one_inode(root, location.objectid);
799         if (!inode) {
800                 ret = -EIO;
801                 goto out;
802         }
803
804         ret = link_to_fixup_dir(trans, root, path, location.objectid);
805         if (ret)
806                 goto out;
807
808         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
809         if (ret)
810                 goto out;
811         else
812                 ret = btrfs_run_delayed_items(trans, root);
813 out:
814         kfree(name);
815         iput(inode);
816         return ret;
817 }
818
819 /*
820  * helper function to see if a given name and sequence number found
821  * in an inode back reference are already in a directory and correctly
822  * point to this inode
823  */
824 static noinline int inode_in_dir(struct btrfs_root *root,
825                                  struct btrfs_path *path,
826                                  u64 dirid, u64 objectid, u64 index,
827                                  const char *name, int name_len)
828 {
829         struct btrfs_dir_item *di;
830         struct btrfs_key location;
831         int match = 0;
832
833         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
834                                          index, name, name_len, 0);
835         if (di && !IS_ERR(di)) {
836                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
837                 if (location.objectid != objectid)
838                         goto out;
839         } else
840                 goto out;
841         btrfs_release_path(path);
842
843         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
844         if (di && !IS_ERR(di)) {
845                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
846                 if (location.objectid != objectid)
847                         goto out;
848         } else
849                 goto out;
850         match = 1;
851 out:
852         btrfs_release_path(path);
853         return match;
854 }
855
856 /*
857  * helper function to check a log tree for a named back reference in
858  * an inode.  This is used to decide if a back reference that is
859  * found in the subvolume conflicts with what we find in the log.
860  *
861  * inode backreferences may have multiple refs in a single item,
862  * during replay we process one reference at a time, and we don't
863  * want to delete valid links to a file from the subvolume if that
864  * link is also in the log.
865  */
866 static noinline int backref_in_log(struct btrfs_root *log,
867                                    struct btrfs_key *key,
868                                    u64 ref_objectid,
869                                    const char *name, int namelen)
870 {
871         struct btrfs_path *path;
872         struct btrfs_inode_ref *ref;
873         unsigned long ptr;
874         unsigned long ptr_end;
875         unsigned long name_ptr;
876         int found_name_len;
877         int item_size;
878         int ret;
879         int match = 0;
880
881         path = btrfs_alloc_path();
882         if (!path)
883                 return -ENOMEM;
884
885         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
886         if (ret != 0)
887                 goto out;
888
889         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
890
891         if (key->type == BTRFS_INODE_EXTREF_KEY) {
892                 if (btrfs_find_name_in_ext_backref(path, ref_objectid,
893                                                    name, namelen, NULL))
894                         match = 1;
895
896                 goto out;
897         }
898
899         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
900         ptr_end = ptr + item_size;
901         while (ptr < ptr_end) {
902                 ref = (struct btrfs_inode_ref *)ptr;
903                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
904                 if (found_name_len == namelen) {
905                         name_ptr = (unsigned long)(ref + 1);
906                         ret = memcmp_extent_buffer(path->nodes[0], name,
907                                                    name_ptr, namelen);
908                         if (ret == 0) {
909                                 match = 1;
910                                 goto out;
911                         }
912                 }
913                 ptr = (unsigned long)(ref + 1) + found_name_len;
914         }
915 out:
916         btrfs_free_path(path);
917         return match;
918 }
919
920 static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
921                                   struct btrfs_root *root,
922                                   struct btrfs_path *path,
923                                   struct btrfs_root *log_root,
924                                   struct inode *dir, struct inode *inode,
925                                   struct extent_buffer *eb,
926                                   u64 inode_objectid, u64 parent_objectid,
927                                   u64 ref_index, char *name, int namelen,
928                                   int *search_done)
929 {
930         int ret;
931         char *victim_name;
932         int victim_name_len;
933         struct extent_buffer *leaf;
934         struct btrfs_dir_item *di;
935         struct btrfs_key search_key;
936         struct btrfs_inode_extref *extref;
937
938 again:
939         /* Search old style refs */
940         search_key.objectid = inode_objectid;
941         search_key.type = BTRFS_INODE_REF_KEY;
942         search_key.offset = parent_objectid;
943         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
944         if (ret == 0) {
945                 struct btrfs_inode_ref *victim_ref;
946                 unsigned long ptr;
947                 unsigned long ptr_end;
948
949                 leaf = path->nodes[0];
950
951                 /* are we trying to overwrite a back ref for the root directory
952                  * if so, just jump out, we're done
953                  */
954                 if (search_key.objectid == search_key.offset)
955                         return 1;
956
957                 /* check all the names in this back reference to see
958                  * if they are in the log.  if so, we allow them to stay
959                  * otherwise they must be unlinked as a conflict
960                  */
961                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
962                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
963                 while (ptr < ptr_end) {
964                         victim_ref = (struct btrfs_inode_ref *)ptr;
965                         victim_name_len = btrfs_inode_ref_name_len(leaf,
966                                                                    victim_ref);
967                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
968                         if (!victim_name)
969                                 return -ENOMEM;
970
971                         read_extent_buffer(leaf, victim_name,
972                                            (unsigned long)(victim_ref + 1),
973                                            victim_name_len);
974
975                         if (!backref_in_log(log_root, &search_key,
976                                             parent_objectid,
977                                             victim_name,
978                                             victim_name_len)) {
979                                 inc_nlink(inode);
980                                 btrfs_release_path(path);
981
982                                 ret = btrfs_unlink_inode(trans, root, dir,
983                                                          inode, victim_name,
984                                                          victim_name_len);
985                                 kfree(victim_name);
986                                 if (ret)
987                                         return ret;
988                                 ret = btrfs_run_delayed_items(trans, root);
989                                 if (ret)
990                                         return ret;
991                                 *search_done = 1;
992                                 goto again;
993                         }
994                         kfree(victim_name);
995
996                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
997                 }
998
999                 /*
1000                  * NOTE: we have searched root tree and checked the
1001                  * coresponding ref, it does not need to check again.
1002                  */
1003                 *search_done = 1;
1004         }
1005         btrfs_release_path(path);
1006
1007         /* Same search but for extended refs */
1008         extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
1009                                            inode_objectid, parent_objectid, 0,
1010                                            0);
1011         if (!IS_ERR_OR_NULL(extref)) {
1012                 u32 item_size;
1013                 u32 cur_offset = 0;
1014                 unsigned long base;
1015                 struct inode *victim_parent;
1016
1017                 leaf = path->nodes[0];
1018
1019                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1020                 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
1021
1022                 while (cur_offset < item_size) {
1023                         extref = (struct btrfs_inode_extref *)(base + cur_offset);
1024
1025                         victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
1026
1027                         if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
1028                                 goto next;
1029
1030                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
1031                         if (!victim_name)
1032                                 return -ENOMEM;
1033                         read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
1034                                            victim_name_len);
1035
1036                         search_key.objectid = inode_objectid;
1037                         search_key.type = BTRFS_INODE_EXTREF_KEY;
1038                         search_key.offset = btrfs_extref_hash(parent_objectid,
1039                                                               victim_name,
1040                                                               victim_name_len);
1041                         ret = 0;
1042                         if (!backref_in_log(log_root, &search_key,
1043                                             parent_objectid, victim_name,
1044                                             victim_name_len)) {
1045                                 ret = -ENOENT;
1046                                 victim_parent = read_one_inode(root,
1047                                                                parent_objectid);
1048                                 if (victim_parent) {
1049                                         inc_nlink(inode);
1050                                         btrfs_release_path(path);
1051
1052                                         ret = btrfs_unlink_inode(trans, root,
1053                                                                  victim_parent,
1054                                                                  inode,
1055                                                                  victim_name,
1056                                                                  victim_name_len);
1057                                         if (!ret)
1058                                                 ret = btrfs_run_delayed_items(
1059                                                                   trans, root);
1060                                 }
1061                                 iput(victim_parent);
1062                                 kfree(victim_name);
1063                                 if (ret)
1064                                         return ret;
1065                                 *search_done = 1;
1066                                 goto again;
1067                         }
1068                         kfree(victim_name);
1069                         if (ret)
1070                                 return ret;
1071 next:
1072                         cur_offset += victim_name_len + sizeof(*extref);
1073                 }
1074                 *search_done = 1;
1075         }
1076         btrfs_release_path(path);
1077
1078         /* look for a conflicting sequence number */
1079         di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
1080                                          ref_index, name, namelen, 0);
1081         if (di && !IS_ERR(di)) {
1082                 ret = drop_one_dir_item(trans, root, path, dir, di);
1083                 if (ret)
1084                         return ret;
1085         }
1086         btrfs_release_path(path);
1087
1088         /* look for a conflicing name */
1089         di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
1090                                    name, namelen, 0);
1091         if (di && !IS_ERR(di)) {
1092                 ret = drop_one_dir_item(trans, root, path, dir, di);
1093                 if (ret)
1094                         return ret;
1095         }
1096         btrfs_release_path(path);
1097
1098         return 0;
1099 }
1100
1101 static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1102                              u32 *namelen, char **name, u64 *index,
1103                              u64 *parent_objectid)
1104 {
1105         struct btrfs_inode_extref *extref;
1106
1107         extref = (struct btrfs_inode_extref *)ref_ptr;
1108
1109         *namelen = btrfs_inode_extref_name_len(eb, extref);
1110         *name = kmalloc(*namelen, GFP_NOFS);
1111         if (*name == NULL)
1112                 return -ENOMEM;
1113
1114         read_extent_buffer(eb, *name, (unsigned long)&extref->name,
1115                            *namelen);
1116
1117         *index = btrfs_inode_extref_index(eb, extref);
1118         if (parent_objectid)
1119                 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
1120
1121         return 0;
1122 }
1123
1124 static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1125                           u32 *namelen, char **name, u64 *index)
1126 {
1127         struct btrfs_inode_ref *ref;
1128
1129         ref = (struct btrfs_inode_ref *)ref_ptr;
1130
1131         *namelen = btrfs_inode_ref_name_len(eb, ref);
1132         *name = kmalloc(*namelen, GFP_NOFS);
1133         if (*name == NULL)
1134                 return -ENOMEM;
1135
1136         read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1137
1138         *index = btrfs_inode_ref_index(eb, ref);
1139
1140         return 0;
1141 }
1142
1143 /*
1144  * replay one inode back reference item found in the log tree.
1145  * eb, slot and key refer to the buffer and key found in the log tree.
1146  * root is the destination we are replaying into, and path is for temp
1147  * use by this function.  (it should be released on return).
1148  */
1149 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1150                                   struct btrfs_root *root,
1151                                   struct btrfs_root *log,
1152                                   struct btrfs_path *path,
1153                                   struct extent_buffer *eb, int slot,
1154                                   struct btrfs_key *key)
1155 {
1156         struct inode *dir = NULL;
1157         struct inode *inode = NULL;
1158         unsigned long ref_ptr;
1159         unsigned long ref_end;
1160         char *name = NULL;
1161         int namelen;
1162         int ret;
1163         int search_done = 0;
1164         int log_ref_ver = 0;
1165         u64 parent_objectid;
1166         u64 inode_objectid;
1167         u64 ref_index = 0;
1168         int ref_struct_size;
1169
1170         ref_ptr = btrfs_item_ptr_offset(eb, slot);
1171         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1172
1173         if (key->type == BTRFS_INODE_EXTREF_KEY) {
1174                 struct btrfs_inode_extref *r;
1175
1176                 ref_struct_size = sizeof(struct btrfs_inode_extref);
1177                 log_ref_ver = 1;
1178                 r = (struct btrfs_inode_extref *)ref_ptr;
1179                 parent_objectid = btrfs_inode_extref_parent(eb, r);
1180         } else {
1181                 ref_struct_size = sizeof(struct btrfs_inode_ref);
1182                 parent_objectid = key->offset;
1183         }
1184         inode_objectid = key->objectid;
1185
1186         /*
1187          * it is possible that we didn't log all the parent directories
1188          * for a given inode.  If we don't find the dir, just don't
1189          * copy the back ref in.  The link count fixup code will take
1190          * care of the rest
1191          */
1192         dir = read_one_inode(root, parent_objectid);
1193         if (!dir) {
1194                 ret = -ENOENT;
1195                 goto out;
1196         }
1197
1198         inode = read_one_inode(root, inode_objectid);
1199         if (!inode) {
1200                 ret = -EIO;
1201                 goto out;
1202         }
1203
1204         while (ref_ptr < ref_end) {
1205                 if (log_ref_ver) {
1206                         ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1207                                                 &ref_index, &parent_objectid);
1208                         /*
1209                          * parent object can change from one array
1210                          * item to another.
1211                          */
1212                         if (!dir)
1213                                 dir = read_one_inode(root, parent_objectid);
1214                         if (!dir) {
1215                                 ret = -ENOENT;
1216                                 goto out;
1217                         }
1218                 } else {
1219                         ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1220                                              &ref_index);
1221                 }
1222                 if (ret)
1223                         goto out;
1224
1225                 /* if we already have a perfect match, we're done */
1226                 if (!inode_in_dir(root, path, btrfs_ino(dir), btrfs_ino(inode),
1227                                   ref_index, name, namelen)) {
1228                         /*
1229                          * look for a conflicting back reference in the
1230                          * metadata. if we find one we have to unlink that name
1231                          * of the file before we add our new link.  Later on, we
1232                          * overwrite any existing back reference, and we don't
1233                          * want to create dangling pointers in the directory.
1234                          */
1235
1236                         if (!search_done) {
1237                                 ret = __add_inode_ref(trans, root, path, log,
1238                                                       dir, inode, eb,
1239                                                       inode_objectid,
1240                                                       parent_objectid,
1241                                                       ref_index, name, namelen,
1242                                                       &search_done);
1243                                 if (ret) {
1244                                         if (ret == 1)
1245                                                 ret = 0;
1246                                         goto out;
1247                                 }
1248                         }
1249
1250                         /* insert our name */
1251                         ret = btrfs_add_link(trans, dir, inode, name, namelen,
1252                                              0, ref_index);
1253                         if (ret)
1254                                 goto out;
1255
1256                         btrfs_update_inode(trans, root, inode);
1257                 }
1258
1259                 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1260                 kfree(name);
1261                 name = NULL;
1262                 if (log_ref_ver) {
1263                         iput(dir);
1264                         dir = NULL;
1265                 }
1266         }
1267
1268         /* finally write the back reference in the inode */
1269         ret = overwrite_item(trans, root, path, eb, slot, key);
1270 out:
1271         btrfs_release_path(path);
1272         kfree(name);
1273         iput(dir);
1274         iput(inode);
1275         return ret;
1276 }
1277
1278 static int insert_orphan_item(struct btrfs_trans_handle *trans,
1279                               struct btrfs_root *root, u64 ino)
1280 {
1281         int ret;
1282
1283         ret = btrfs_insert_orphan_item(trans, root, ino);
1284         if (ret == -EEXIST)
1285                 ret = 0;
1286
1287         return ret;
1288 }
1289
1290 static int count_inode_extrefs(struct btrfs_root *root,
1291                                struct inode *inode, struct btrfs_path *path)
1292 {
1293         int ret = 0;
1294         int name_len;
1295         unsigned int nlink = 0;
1296         u32 item_size;
1297         u32 cur_offset = 0;
1298         u64 inode_objectid = btrfs_ino(inode);
1299         u64 offset = 0;
1300         unsigned long ptr;
1301         struct btrfs_inode_extref *extref;
1302         struct extent_buffer *leaf;
1303
1304         while (1) {
1305                 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1306                                             &extref, &offset);
1307                 if (ret)
1308                         break;
1309
1310                 leaf = path->nodes[0];
1311                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1312                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1313                 cur_offset = 0;
1314
1315                 while (cur_offset < item_size) {
1316                         extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1317                         name_len = btrfs_inode_extref_name_len(leaf, extref);
1318
1319                         nlink++;
1320
1321                         cur_offset += name_len + sizeof(*extref);
1322                 }
1323
1324                 offset++;
1325                 btrfs_release_path(path);
1326         }
1327         btrfs_release_path(path);
1328
1329         if (ret < 0 && ret != -ENOENT)
1330                 return ret;
1331         return nlink;
1332 }
1333
1334 static int count_inode_refs(struct btrfs_root *root,
1335                                struct inode *inode, struct btrfs_path *path)
1336 {
1337         int ret;
1338         struct btrfs_key key;
1339         unsigned int nlink = 0;
1340         unsigned long ptr;
1341         unsigned long ptr_end;
1342         int name_len;
1343         u64 ino = btrfs_ino(inode);
1344
1345         key.objectid = ino;
1346         key.type = BTRFS_INODE_REF_KEY;
1347         key.offset = (u64)-1;
1348
1349         while (1) {
1350                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1351                 if (ret < 0)
1352                         break;
1353                 if (ret > 0) {
1354                         if (path->slots[0] == 0)
1355                                 break;
1356                         path->slots[0]--;
1357                 }
1358 process_slot:
1359                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1360                                       path->slots[0]);
1361                 if (key.objectid != ino ||
1362                     key.type != BTRFS_INODE_REF_KEY)
1363                         break;
1364                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1365                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1366                                                    path->slots[0]);
1367                 while (ptr < ptr_end) {
1368                         struct btrfs_inode_ref *ref;
1369
1370                         ref = (struct btrfs_inode_ref *)ptr;
1371                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1372                                                             ref);
1373                         ptr = (unsigned long)(ref + 1) + name_len;
1374                         nlink++;
1375                 }
1376
1377                 if (key.offset == 0)
1378                         break;
1379                 if (path->slots[0] > 0) {
1380                         path->slots[0]--;
1381                         goto process_slot;
1382                 }
1383                 key.offset--;
1384                 btrfs_release_path(path);
1385         }
1386         btrfs_release_path(path);
1387
1388         return nlink;
1389 }
1390
1391 /*
1392  * There are a few corners where the link count of the file can't
1393  * be properly maintained during replay.  So, instead of adding
1394  * lots of complexity to the log code, we just scan the backrefs
1395  * for any file that has been through replay.
1396  *
1397  * The scan will update the link count on the inode to reflect the
1398  * number of back refs found.  If it goes down to zero, the iput
1399  * will free the inode.
1400  */
1401 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1402                                            struct btrfs_root *root,
1403                                            struct inode *inode)
1404 {
1405         struct btrfs_path *path;
1406         int ret;
1407         u64 nlink = 0;
1408         u64 ino = btrfs_ino(inode);
1409
1410         path = btrfs_alloc_path();
1411         if (!path)
1412                 return -ENOMEM;
1413
1414         ret = count_inode_refs(root, inode, path);
1415         if (ret < 0)
1416                 goto out;
1417
1418         nlink = ret;
1419
1420         ret = count_inode_extrefs(root, inode, path);
1421         if (ret < 0)
1422                 goto out;
1423
1424         nlink += ret;
1425
1426         ret = 0;
1427
1428         if (nlink != inode->i_nlink) {
1429                 set_nlink(inode, nlink);
1430                 btrfs_update_inode(trans, root, inode);
1431         }
1432         BTRFS_I(inode)->index_cnt = (u64)-1;
1433
1434         if (inode->i_nlink == 0) {
1435                 if (S_ISDIR(inode->i_mode)) {
1436                         ret = replay_dir_deletes(trans, root, NULL, path,
1437                                                  ino, 1);
1438                         if (ret)
1439                                 goto out;
1440                 }
1441                 ret = insert_orphan_item(trans, root, ino);
1442         }
1443
1444 out:
1445         btrfs_free_path(path);
1446         return ret;
1447 }
1448
1449 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1450                                             struct btrfs_root *root,
1451                                             struct btrfs_path *path)
1452 {
1453         int ret;
1454         struct btrfs_key key;
1455         struct inode *inode;
1456
1457         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1458         key.type = BTRFS_ORPHAN_ITEM_KEY;
1459         key.offset = (u64)-1;
1460         while (1) {
1461                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1462                 if (ret < 0)
1463                         break;
1464
1465                 if (ret == 1) {
1466                         if (path->slots[0] == 0)
1467                                 break;
1468                         path->slots[0]--;
1469                 }
1470
1471                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1472                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1473                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1474                         break;
1475
1476                 ret = btrfs_del_item(trans, root, path);
1477                 if (ret)
1478                         goto out;
1479
1480                 btrfs_release_path(path);
1481                 inode = read_one_inode(root, key.offset);
1482                 if (!inode)
1483                         return -EIO;
1484
1485                 ret = fixup_inode_link_count(trans, root, inode);
1486                 iput(inode);
1487                 if (ret)
1488                         goto out;
1489
1490                 /*
1491                  * fixup on a directory may create new entries,
1492                  * make sure we always look for the highset possible
1493                  * offset
1494                  */
1495                 key.offset = (u64)-1;
1496         }
1497         ret = 0;
1498 out:
1499         btrfs_release_path(path);
1500         return ret;
1501 }
1502
1503
1504 /*
1505  * record a given inode in the fixup dir so we can check its link
1506  * count when replay is done.  The link count is incremented here
1507  * so the inode won't go away until we check it
1508  */
1509 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1510                                       struct btrfs_root *root,
1511                                       struct btrfs_path *path,
1512                                       u64 objectid)
1513 {
1514         struct btrfs_key key;
1515         int ret = 0;
1516         struct inode *inode;
1517
1518         inode = read_one_inode(root, objectid);
1519         if (!inode)
1520                 return -EIO;
1521
1522         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1523         key.type = BTRFS_ORPHAN_ITEM_KEY;
1524         key.offset = objectid;
1525
1526         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1527
1528         btrfs_release_path(path);
1529         if (ret == 0) {
1530                 if (!inode->i_nlink)
1531                         set_nlink(inode, 1);
1532                 else
1533                         inc_nlink(inode);
1534                 ret = btrfs_update_inode(trans, root, inode);
1535         } else if (ret == -EEXIST) {
1536                 ret = 0;
1537         } else {
1538                 BUG(); /* Logic Error */
1539         }
1540         iput(inode);
1541
1542         return ret;
1543 }
1544
1545 /*
1546  * when replaying the log for a directory, we only insert names
1547  * for inodes that actually exist.  This means an fsync on a directory
1548  * does not implicitly fsync all the new files in it
1549  */
1550 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1551                                     struct btrfs_root *root,
1552                                     struct btrfs_path *path,
1553                                     u64 dirid, u64 index,
1554                                     char *name, int name_len, u8 type,
1555                                     struct btrfs_key *location)
1556 {
1557         struct inode *inode;
1558         struct inode *dir;
1559         int ret;
1560
1561         inode = read_one_inode(root, location->objectid);
1562         if (!inode)
1563                 return -ENOENT;
1564
1565         dir = read_one_inode(root, dirid);
1566         if (!dir) {
1567                 iput(inode);
1568                 return -EIO;
1569         }
1570
1571         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1572
1573         /* FIXME, put inode into FIXUP list */
1574
1575         iput(inode);
1576         iput(dir);
1577         return ret;
1578 }
1579
1580 /*
1581  * Return true if an inode reference exists in the log for the given name,
1582  * inode and parent inode.
1583  */
1584 static bool name_in_log_ref(struct btrfs_root *log_root,
1585                             const char *name, const int name_len,
1586                             const u64 dirid, const u64 ino)
1587 {
1588         struct btrfs_key search_key;
1589
1590         search_key.objectid = ino;
1591         search_key.type = BTRFS_INODE_REF_KEY;
1592         search_key.offset = dirid;
1593         if (backref_in_log(log_root, &search_key, dirid, name, name_len))
1594                 return true;
1595
1596         search_key.type = BTRFS_INODE_EXTREF_KEY;
1597         search_key.offset = btrfs_extref_hash(dirid, name, name_len);
1598         if (backref_in_log(log_root, &search_key, dirid, name, name_len))
1599                 return true;
1600
1601         return false;
1602 }
1603
1604 /*
1605  * take a single entry in a log directory item and replay it into
1606  * the subvolume.
1607  *
1608  * if a conflicting item exists in the subdirectory already,
1609  * the inode it points to is unlinked and put into the link count
1610  * fix up tree.
1611  *
1612  * If a name from the log points to a file or directory that does
1613  * not exist in the FS, it is skipped.  fsyncs on directories
1614  * do not force down inodes inside that directory, just changes to the
1615  * names or unlinks in a directory.
1616  */
1617 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1618                                     struct btrfs_root *root,
1619                                     struct btrfs_path *path,
1620                                     struct extent_buffer *eb,
1621                                     struct btrfs_dir_item *di,
1622                                     struct btrfs_key *key)
1623 {
1624         char *name;
1625         int name_len;
1626         struct btrfs_dir_item *dst_di;
1627         struct btrfs_key found_key;
1628         struct btrfs_key log_key;
1629         struct inode *dir;
1630         u8 log_type;
1631         int exists;
1632         int ret = 0;
1633         bool update_size = (key->type == BTRFS_DIR_INDEX_KEY);
1634
1635         dir = read_one_inode(root, key->objectid);
1636         if (!dir)
1637                 return -EIO;
1638
1639         name_len = btrfs_dir_name_len(eb, di);
1640         name = kmalloc(name_len, GFP_NOFS);
1641         if (!name) {
1642                 ret = -ENOMEM;
1643                 goto out;
1644         }
1645
1646         log_type = btrfs_dir_type(eb, di);
1647         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1648                    name_len);
1649
1650         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1651         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1652         if (exists == 0)
1653                 exists = 1;
1654         else
1655                 exists = 0;
1656         btrfs_release_path(path);
1657
1658         if (key->type == BTRFS_DIR_ITEM_KEY) {
1659                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1660                                        name, name_len, 1);
1661         } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1662                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1663                                                      key->objectid,
1664                                                      key->offset, name,
1665                                                      name_len, 1);
1666         } else {
1667                 /* Corruption */
1668                 ret = -EINVAL;
1669                 goto out;
1670         }
1671         if (IS_ERR_OR_NULL(dst_di)) {
1672                 /* we need a sequence number to insert, so we only
1673                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1674                  */
1675                 if (key->type != BTRFS_DIR_INDEX_KEY)
1676                         goto out;
1677                 goto insert;
1678         }
1679
1680         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1681         /* the existing item matches the logged item */
1682         if (found_key.objectid == log_key.objectid &&
1683             found_key.type == log_key.type &&
1684             found_key.offset == log_key.offset &&
1685             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1686                 update_size = false;
1687                 goto out;
1688         }
1689
1690         /*
1691          * don't drop the conflicting directory entry if the inode
1692          * for the new entry doesn't exist
1693          */
1694         if (!exists)
1695                 goto out;
1696
1697         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1698         if (ret)
1699                 goto out;
1700
1701         if (key->type == BTRFS_DIR_INDEX_KEY)
1702                 goto insert;
1703 out:
1704         btrfs_release_path(path);
1705         if (!ret && update_size) {
1706                 btrfs_i_size_write(dir, dir->i_size + name_len * 2);
1707                 ret = btrfs_update_inode(trans, root, dir);
1708         }
1709         kfree(name);
1710         iput(dir);
1711         return ret;
1712
1713 insert:
1714         if (name_in_log_ref(root->log_root, name, name_len,
1715                             key->objectid, log_key.objectid)) {
1716                 /* The dentry will be added later. */
1717                 ret = 0;
1718                 update_size = false;
1719                 goto out;
1720         }
1721         btrfs_release_path(path);
1722         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1723                               name, name_len, log_type, &log_key);
1724         if (ret && ret != -ENOENT && ret != -EEXIST)
1725                 goto out;
1726         update_size = false;
1727         ret = 0;
1728         goto out;
1729 }
1730
1731 /*
1732  * find all the names in a directory item and reconcile them into
1733  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1734  * one name in a directory item, but the same code gets used for
1735  * both directory index types
1736  */
1737 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1738                                         struct btrfs_root *root,
1739                                         struct btrfs_path *path,
1740                                         struct extent_buffer *eb, int slot,
1741                                         struct btrfs_key *key)
1742 {
1743         int ret;
1744         u32 item_size = btrfs_item_size_nr(eb, slot);
1745         struct btrfs_dir_item *di;
1746         int name_len;
1747         unsigned long ptr;
1748         unsigned long ptr_end;
1749
1750         ptr = btrfs_item_ptr_offset(eb, slot);
1751         ptr_end = ptr + item_size;
1752         while (ptr < ptr_end) {
1753                 di = (struct btrfs_dir_item *)ptr;
1754                 if (verify_dir_item(root, eb, di))
1755                         return -EIO;
1756                 name_len = btrfs_dir_name_len(eb, di);
1757                 ret = replay_one_name(trans, root, path, eb, di, key);
1758                 if (ret)
1759                         return ret;
1760                 ptr = (unsigned long)(di + 1);
1761                 ptr += name_len;
1762         }
1763         return 0;
1764 }
1765
1766 /*
1767  * directory replay has two parts.  There are the standard directory
1768  * items in the log copied from the subvolume, and range items
1769  * created in the log while the subvolume was logged.
1770  *
1771  * The range items tell us which parts of the key space the log
1772  * is authoritative for.  During replay, if a key in the subvolume
1773  * directory is in a logged range item, but not actually in the log
1774  * that means it was deleted from the directory before the fsync
1775  * and should be removed.
1776  */
1777 static noinline int find_dir_range(struct btrfs_root *root,
1778                                    struct btrfs_path *path,
1779                                    u64 dirid, int key_type,
1780                                    u64 *start_ret, u64 *end_ret)
1781 {
1782         struct btrfs_key key;
1783         u64 found_end;
1784         struct btrfs_dir_log_item *item;
1785         int ret;
1786         int nritems;
1787
1788         if (*start_ret == (u64)-1)
1789                 return 1;
1790
1791         key.objectid = dirid;
1792         key.type = key_type;
1793         key.offset = *start_ret;
1794
1795         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1796         if (ret < 0)
1797                 goto out;
1798         if (ret > 0) {
1799                 if (path->slots[0] == 0)
1800                         goto out;
1801                 path->slots[0]--;
1802         }
1803         if (ret != 0)
1804                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1805
1806         if (key.type != key_type || key.objectid != dirid) {
1807                 ret = 1;
1808                 goto next;
1809         }
1810         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1811                               struct btrfs_dir_log_item);
1812         found_end = btrfs_dir_log_end(path->nodes[0], item);
1813
1814         if (*start_ret >= key.offset && *start_ret <= found_end) {
1815                 ret = 0;
1816                 *start_ret = key.offset;
1817                 *end_ret = found_end;
1818                 goto out;
1819         }
1820         ret = 1;
1821 next:
1822         /* check the next slot in the tree to see if it is a valid item */
1823         nritems = btrfs_header_nritems(path->nodes[0]);
1824         if (path->slots[0] >= nritems) {
1825                 ret = btrfs_next_leaf(root, path);
1826                 if (ret)
1827                         goto out;
1828         } else {
1829                 path->slots[0]++;
1830         }
1831
1832         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1833
1834         if (key.type != key_type || key.objectid != dirid) {
1835                 ret = 1;
1836                 goto out;
1837         }
1838         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1839                               struct btrfs_dir_log_item);
1840         found_end = btrfs_dir_log_end(path->nodes[0], item);
1841         *start_ret = key.offset;
1842         *end_ret = found_end;
1843         ret = 0;
1844 out:
1845         btrfs_release_path(path);
1846         return ret;
1847 }
1848
1849 /*
1850  * this looks for a given directory item in the log.  If the directory
1851  * item is not in the log, the item is removed and the inode it points
1852  * to is unlinked
1853  */
1854 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1855                                       struct btrfs_root *root,
1856                                       struct btrfs_root *log,
1857                                       struct btrfs_path *path,
1858                                       struct btrfs_path *log_path,
1859                                       struct inode *dir,
1860                                       struct btrfs_key *dir_key)
1861 {
1862         int ret;
1863         struct extent_buffer *eb;
1864         int slot;
1865         u32 item_size;
1866         struct btrfs_dir_item *di;
1867         struct btrfs_dir_item *log_di;
1868         int name_len;
1869         unsigned long ptr;
1870         unsigned long ptr_end;
1871         char *name;
1872         struct inode *inode;
1873         struct btrfs_key location;
1874
1875 again:
1876         eb = path->nodes[0];
1877         slot = path->slots[0];
1878         item_size = btrfs_item_size_nr(eb, slot);
1879         ptr = btrfs_item_ptr_offset(eb, slot);
1880         ptr_end = ptr + item_size;
1881         while (ptr < ptr_end) {
1882                 di = (struct btrfs_dir_item *)ptr;
1883                 if (verify_dir_item(root, eb, di)) {
1884                         ret = -EIO;
1885                         goto out;
1886                 }
1887
1888                 name_len = btrfs_dir_name_len(eb, di);
1889                 name = kmalloc(name_len, GFP_NOFS);
1890                 if (!name) {
1891                         ret = -ENOMEM;
1892                         goto out;
1893                 }
1894                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1895                                   name_len);
1896                 log_di = NULL;
1897                 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
1898                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1899                                                        dir_key->objectid,
1900                                                        name, name_len, 0);
1901                 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
1902                         log_di = btrfs_lookup_dir_index_item(trans, log,
1903                                                      log_path,
1904                                                      dir_key->objectid,
1905                                                      dir_key->offset,
1906                                                      name, name_len, 0);
1907                 }
1908                 if (!log_di || (IS_ERR(log_di) && PTR_ERR(log_di) == -ENOENT)) {
1909                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1910                         btrfs_release_path(path);
1911                         btrfs_release_path(log_path);
1912                         inode = read_one_inode(root, location.objectid);
1913                         if (!inode) {
1914                                 kfree(name);
1915                                 return -EIO;
1916                         }
1917
1918                         ret = link_to_fixup_dir(trans, root,
1919                                                 path, location.objectid);
1920                         if (ret) {
1921                                 kfree(name);
1922                                 iput(inode);
1923                                 goto out;
1924                         }
1925
1926                         inc_nlink(inode);
1927                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1928                                                  name, name_len);
1929                         if (!ret)
1930                                 ret = btrfs_run_delayed_items(trans, root);
1931                         kfree(name);
1932                         iput(inode);
1933                         if (ret)
1934                                 goto out;
1935
1936                         /* there might still be more names under this key
1937                          * check and repeat if required
1938                          */
1939                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1940                                                 0, 0);
1941                         if (ret == 0)
1942                                 goto again;
1943                         ret = 0;
1944                         goto out;
1945                 } else if (IS_ERR(log_di)) {
1946                         kfree(name);
1947                         return PTR_ERR(log_di);
1948                 }
1949                 btrfs_release_path(log_path);
1950                 kfree(name);
1951
1952                 ptr = (unsigned long)(di + 1);
1953                 ptr += name_len;
1954         }
1955         ret = 0;
1956 out:
1957         btrfs_release_path(path);
1958         btrfs_release_path(log_path);
1959         return ret;
1960 }
1961
1962 static int replay_xattr_deletes(struct btrfs_trans_handle *trans,
1963                               struct btrfs_root *root,
1964                               struct btrfs_root *log,
1965                               struct btrfs_path *path,
1966                               const u64 ino)
1967 {
1968         struct btrfs_key search_key;
1969         struct btrfs_path *log_path;
1970         int i;
1971         int nritems;
1972         int ret;
1973
1974         log_path = btrfs_alloc_path();
1975         if (!log_path)
1976                 return -ENOMEM;
1977
1978         search_key.objectid = ino;
1979         search_key.type = BTRFS_XATTR_ITEM_KEY;
1980         search_key.offset = 0;
1981 again:
1982         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
1983         if (ret < 0)
1984                 goto out;
1985 process_leaf:
1986         nritems = btrfs_header_nritems(path->nodes[0]);
1987         for (i = path->slots[0]; i < nritems; i++) {
1988                 struct btrfs_key key;
1989                 struct btrfs_dir_item *di;
1990                 struct btrfs_dir_item *log_di;
1991                 u32 total_size;
1992                 u32 cur;
1993
1994                 btrfs_item_key_to_cpu(path->nodes[0], &key, i);
1995                 if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY) {
1996                         ret = 0;
1997                         goto out;
1998                 }
1999
2000                 di = btrfs_item_ptr(path->nodes[0], i, struct btrfs_dir_item);
2001                 total_size = btrfs_item_size_nr(path->nodes[0], i);
2002                 cur = 0;
2003                 while (cur < total_size) {
2004                         u16 name_len = btrfs_dir_name_len(path->nodes[0], di);
2005                         u16 data_len = btrfs_dir_data_len(path->nodes[0], di);
2006                         u32 this_len = sizeof(*di) + name_len + data_len;
2007                         char *name;
2008
2009                         name = kmalloc(name_len, GFP_NOFS);
2010                         if (!name) {
2011                                 ret = -ENOMEM;
2012                                 goto out;
2013                         }
2014                         read_extent_buffer(path->nodes[0], name,
2015                                            (unsigned long)(di + 1), name_len);
2016
2017                         log_di = btrfs_lookup_xattr(NULL, log, log_path, ino,
2018                                                     name, name_len, 0);
2019                         btrfs_release_path(log_path);
2020                         if (!log_di) {
2021                                 /* Doesn't exist in log tree, so delete it. */
2022                                 btrfs_release_path(path);
2023                                 di = btrfs_lookup_xattr(trans, root, path, ino,
2024                                                         name, name_len, -1);
2025                                 kfree(name);
2026                                 if (IS_ERR(di)) {
2027                                         ret = PTR_ERR(di);
2028                                         goto out;
2029                                 }
2030                                 ASSERT(di);
2031                                 ret = btrfs_delete_one_dir_name(trans, root,
2032                                                                 path, di);
2033                                 if (ret)
2034                                         goto out;
2035                                 btrfs_release_path(path);
2036                                 search_key = key;
2037                                 goto again;
2038                         }
2039                         kfree(name);
2040                         if (IS_ERR(log_di)) {
2041                                 ret = PTR_ERR(log_di);
2042                                 goto out;
2043                         }
2044                         cur += this_len;
2045                         di = (struct btrfs_dir_item *)((char *)di + this_len);
2046                 }
2047         }
2048         ret = btrfs_next_leaf(root, path);
2049         if (ret > 0)
2050                 ret = 0;
2051         else if (ret == 0)
2052                 goto process_leaf;
2053 out:
2054         btrfs_free_path(log_path);
2055         btrfs_release_path(path);
2056         return ret;
2057 }
2058
2059
2060 /*
2061  * deletion replay happens before we copy any new directory items
2062  * out of the log or out of backreferences from inodes.  It
2063  * scans the log to find ranges of keys that log is authoritative for,
2064  * and then scans the directory to find items in those ranges that are
2065  * not present in the log.
2066  *
2067  * Anything we don't find in the log is unlinked and removed from the
2068  * directory.
2069  */
2070 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
2071                                        struct btrfs_root *root,
2072                                        struct btrfs_root *log,
2073                                        struct btrfs_path *path,
2074                                        u64 dirid, int del_all)
2075 {
2076         u64 range_start;
2077         u64 range_end;
2078         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
2079         int ret = 0;
2080         struct btrfs_key dir_key;
2081         struct btrfs_key found_key;
2082         struct btrfs_path *log_path;
2083         struct inode *dir;
2084
2085         dir_key.objectid = dirid;
2086         dir_key.type = BTRFS_DIR_ITEM_KEY;
2087         log_path = btrfs_alloc_path();
2088         if (!log_path)
2089                 return -ENOMEM;
2090
2091         dir = read_one_inode(root, dirid);
2092         /* it isn't an error if the inode isn't there, that can happen
2093          * because we replay the deletes before we copy in the inode item
2094          * from the log
2095          */
2096         if (!dir) {
2097                 btrfs_free_path(log_path);
2098                 return 0;
2099         }
2100 again:
2101         range_start = 0;
2102         range_end = 0;
2103         while (1) {
2104                 if (del_all)
2105                         range_end = (u64)-1;
2106                 else {
2107                         ret = find_dir_range(log, path, dirid, key_type,
2108                                              &range_start, &range_end);
2109                         if (ret != 0)
2110                                 break;
2111                 }
2112
2113                 dir_key.offset = range_start;
2114                 while (1) {
2115                         int nritems;
2116                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
2117                                                 0, 0);
2118                         if (ret < 0)
2119                                 goto out;
2120
2121                         nritems = btrfs_header_nritems(path->nodes[0]);
2122                         if (path->slots[0] >= nritems) {
2123                                 ret = btrfs_next_leaf(root, path);
2124                                 if (ret)
2125                                         break;
2126                         }
2127                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2128                                               path->slots[0]);
2129                         if (found_key.objectid != dirid ||
2130                             found_key.type != dir_key.type)
2131                                 goto next_type;
2132
2133                         if (found_key.offset > range_end)
2134                                 break;
2135
2136                         ret = check_item_in_log(trans, root, log, path,
2137                                                 log_path, dir,
2138                                                 &found_key);
2139                         if (ret)
2140                                 goto out;
2141                         if (found_key.offset == (u64)-1)
2142                                 break;
2143                         dir_key.offset = found_key.offset + 1;
2144                 }
2145                 btrfs_release_path(path);
2146                 if (range_end == (u64)-1)
2147                         break;
2148                 range_start = range_end + 1;
2149         }
2150
2151 next_type:
2152         ret = 0;
2153         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
2154                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
2155                 dir_key.type = BTRFS_DIR_INDEX_KEY;
2156                 btrfs_release_path(path);
2157                 goto again;
2158         }
2159 out:
2160         btrfs_release_path(path);
2161         btrfs_free_path(log_path);
2162         iput(dir);
2163         return ret;
2164 }
2165
2166 /*
2167  * the process_func used to replay items from the log tree.  This
2168  * gets called in two different stages.  The first stage just looks
2169  * for inodes and makes sure they are all copied into the subvolume.
2170  *
2171  * The second stage copies all the other item types from the log into
2172  * the subvolume.  The two stage approach is slower, but gets rid of
2173  * lots of complexity around inodes referencing other inodes that exist
2174  * only in the log (references come from either directory items or inode
2175  * back refs).
2176  */
2177 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
2178                              struct walk_control *wc, u64 gen)
2179 {
2180         int nritems;
2181         struct btrfs_path *path;
2182         struct btrfs_root *root = wc->replay_dest;
2183         struct btrfs_key key;
2184         int level;
2185         int i;
2186         int ret;
2187
2188         ret = btrfs_read_buffer(eb, gen);
2189         if (ret)
2190                 return ret;
2191
2192         level = btrfs_header_level(eb);
2193
2194         if (level != 0)
2195                 return 0;
2196
2197         path = btrfs_alloc_path();
2198         if (!path)
2199                 return -ENOMEM;
2200
2201         nritems = btrfs_header_nritems(eb);
2202         for (i = 0; i < nritems; i++) {
2203                 btrfs_item_key_to_cpu(eb, &key, i);
2204
2205                 /* inode keys are done during the first stage */
2206                 if (key.type == BTRFS_INODE_ITEM_KEY &&
2207                     wc->stage == LOG_WALK_REPLAY_INODES) {
2208                         struct btrfs_inode_item *inode_item;
2209                         u32 mode;
2210
2211                         inode_item = btrfs_item_ptr(eb, i,
2212                                             struct btrfs_inode_item);
2213                         ret = replay_xattr_deletes(wc->trans, root, log,
2214                                                    path, key.objectid);
2215                         if (ret)
2216                                 break;
2217                         mode = btrfs_inode_mode(eb, inode_item);
2218                         if (S_ISDIR(mode)) {
2219                                 ret = replay_dir_deletes(wc->trans,
2220                                          root, log, path, key.objectid, 0);
2221                                 if (ret)
2222                                         break;
2223                         }
2224                         ret = overwrite_item(wc->trans, root, path,
2225                                              eb, i, &key);
2226                         if (ret)
2227                                 break;
2228
2229                         /* for regular files, make sure corresponding
2230                          * orhpan item exist. extents past the new EOF
2231                          * will be truncated later by orphan cleanup.
2232                          */
2233                         if (S_ISREG(mode)) {
2234                                 ret = insert_orphan_item(wc->trans, root,
2235                                                          key.objectid);
2236                                 if (ret)
2237                                         break;
2238                         }
2239
2240                         ret = link_to_fixup_dir(wc->trans, root,
2241                                                 path, key.objectid);
2242                         if (ret)
2243                                 break;
2244                 }
2245
2246                 if (key.type == BTRFS_DIR_INDEX_KEY &&
2247                     wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
2248                         ret = replay_one_dir_item(wc->trans, root, path,
2249                                                   eb, i, &key);
2250                         if (ret)
2251                                 break;
2252                 }
2253
2254                 if (wc->stage < LOG_WALK_REPLAY_ALL)
2255                         continue;
2256
2257                 /* these keys are simply copied */
2258                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
2259                         ret = overwrite_item(wc->trans, root, path,
2260                                              eb, i, &key);
2261                         if (ret)
2262                                 break;
2263                 } else if (key.type == BTRFS_INODE_REF_KEY ||
2264                            key.type == BTRFS_INODE_EXTREF_KEY) {
2265                         ret = add_inode_ref(wc->trans, root, log, path,
2266                                             eb, i, &key);
2267                         if (ret && ret != -ENOENT)
2268                                 break;
2269                         ret = 0;
2270                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
2271                         ret = replay_one_extent(wc->trans, root, path,
2272                                                 eb, i, &key);
2273                         if (ret)
2274                                 break;
2275                 } else if (key.type == BTRFS_DIR_ITEM_KEY) {
2276                         ret = replay_one_dir_item(wc->trans, root, path,
2277                                                   eb, i, &key);
2278                         if (ret)
2279                                 break;
2280                 }
2281         }
2282         btrfs_free_path(path);
2283         return ret;
2284 }
2285
2286 static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
2287                                    struct btrfs_root *root,
2288                                    struct btrfs_path *path, int *level,
2289                                    struct walk_control *wc)
2290 {
2291         u64 root_owner;
2292         u64 bytenr;
2293         u64 ptr_gen;
2294         struct extent_buffer *next;
2295         struct extent_buffer *cur;
2296         struct extent_buffer *parent;
2297         u32 blocksize;
2298         int ret = 0;
2299
2300         WARN_ON(*level < 0);
2301         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2302
2303         while (*level > 0) {
2304                 WARN_ON(*level < 0);
2305                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
2306                 cur = path->nodes[*level];
2307
2308                 WARN_ON(btrfs_header_level(cur) != *level);
2309
2310                 if (path->slots[*level] >=
2311                     btrfs_header_nritems(cur))
2312                         break;
2313
2314                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
2315                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
2316                 blocksize = root->nodesize;
2317
2318                 parent = path->nodes[*level];
2319                 root_owner = btrfs_header_owner(parent);
2320
2321                 next = btrfs_find_create_tree_block(root, bytenr);
2322                 if (!next)
2323                         return -ENOMEM;
2324
2325                 if (*level == 1) {
2326                         ret = wc->process_func(root, next, wc, ptr_gen);
2327                         if (ret) {
2328                                 free_extent_buffer(next);
2329                                 return ret;
2330                         }
2331
2332                         path->slots[*level]++;
2333                         if (wc->free) {
2334                                 ret = btrfs_read_buffer(next, ptr_gen);
2335                                 if (ret) {
2336                                         free_extent_buffer(next);
2337                                         return ret;
2338                                 }
2339
2340                                 if (trans) {
2341                                         btrfs_tree_lock(next);
2342                                         btrfs_set_lock_blocking(next);
2343                                         clean_tree_block(trans, root->fs_info,
2344                                                         next);
2345                                         btrfs_wait_tree_block_writeback(next);
2346                                         btrfs_tree_unlock(next);
2347                                 }
2348
2349                                 WARN_ON(root_owner !=
2350                                         BTRFS_TREE_LOG_OBJECTID);
2351                                 ret = btrfs_free_and_pin_reserved_extent(root,
2352                                                          bytenr, blocksize);
2353                                 if (ret) {
2354                                         free_extent_buffer(next);
2355                                         return ret;
2356                                 }
2357                         }
2358                         free_extent_buffer(next);
2359                         continue;
2360                 }
2361                 ret = btrfs_read_buffer(next, ptr_gen);
2362                 if (ret) {
2363                         free_extent_buffer(next);
2364                         return ret;
2365                 }
2366
2367                 WARN_ON(*level <= 0);
2368                 if (path->nodes[*level-1])
2369                         free_extent_buffer(path->nodes[*level-1]);
2370                 path->nodes[*level-1] = next;
2371                 *level = btrfs_header_level(next);
2372                 path->slots[*level] = 0;
2373                 cond_resched();
2374         }
2375         WARN_ON(*level < 0);
2376         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2377
2378         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2379
2380         cond_resched();
2381         return 0;
2382 }
2383
2384 static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2385                                  struct btrfs_root *root,
2386                                  struct btrfs_path *path, int *level,
2387                                  struct walk_control *wc)
2388 {
2389         u64 root_owner;
2390         int i;
2391         int slot;
2392         int ret;
2393
2394         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2395                 slot = path->slots[i];
2396                 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2397                         path->slots[i]++;
2398                         *level = i;
2399                         WARN_ON(*level == 0);
2400                         return 0;
2401                 } else {
2402                         struct extent_buffer *parent;
2403                         if (path->nodes[*level] == root->node)
2404                                 parent = path->nodes[*level];
2405                         else
2406                                 parent = path->nodes[*level + 1];
2407
2408                         root_owner = btrfs_header_owner(parent);
2409                         ret = wc->process_func(root, path->nodes[*level], wc,
2410                                  btrfs_header_generation(path->nodes[*level]));
2411                         if (ret)
2412                                 return ret;
2413
2414                         if (wc->free) {
2415                                 struct extent_buffer *next;
2416
2417                                 next = path->nodes[*level];
2418
2419                                 if (trans) {
2420                                         btrfs_tree_lock(next);
2421                                         btrfs_set_lock_blocking(next);
2422                                         clean_tree_block(trans, root->fs_info,
2423                                                         next);
2424                                         btrfs_wait_tree_block_writeback(next);
2425                                         btrfs_tree_unlock(next);
2426                                 }
2427
2428                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
2429                                 ret = btrfs_free_and_pin_reserved_extent(root,
2430                                                 path->nodes[*level]->start,
2431                                                 path->nodes[*level]->len);
2432                                 if (ret)
2433                                         return ret;
2434                         }
2435                         free_extent_buffer(path->nodes[*level]);
2436                         path->nodes[*level] = NULL;
2437                         *level = i + 1;
2438                 }
2439         }
2440         return 1;
2441 }
2442
2443 /*
2444  * drop the reference count on the tree rooted at 'snap'.  This traverses
2445  * the tree freeing any blocks that have a ref count of zero after being
2446  * decremented.
2447  */
2448 static int walk_log_tree(struct btrfs_trans_handle *trans,
2449                          struct btrfs_root *log, struct walk_control *wc)
2450 {
2451         int ret = 0;
2452         int wret;
2453         int level;
2454         struct btrfs_path *path;
2455         int orig_level;
2456
2457         path = btrfs_alloc_path();
2458         if (!path)
2459                 return -ENOMEM;
2460
2461         level = btrfs_header_level(log->node);
2462         orig_level = level;
2463         path->nodes[level] = log->node;
2464         extent_buffer_get(log->node);
2465         path->slots[level] = 0;
2466
2467         while (1) {
2468                 wret = walk_down_log_tree(trans, log, path, &level, wc);
2469                 if (wret > 0)
2470                         break;
2471                 if (wret < 0) {
2472                         ret = wret;
2473                         goto out;
2474                 }
2475
2476                 wret = walk_up_log_tree(trans, log, path, &level, wc);
2477                 if (wret > 0)
2478                         break;
2479                 if (wret < 0) {
2480                         ret = wret;
2481                         goto out;
2482                 }
2483         }
2484
2485         /* was the root node processed? if not, catch it here */
2486         if (path->nodes[orig_level]) {
2487                 ret = wc->process_func(log, path->nodes[orig_level], wc,
2488                          btrfs_header_generation(path->nodes[orig_level]));
2489                 if (ret)
2490                         goto out;
2491                 if (wc->free) {
2492                         struct extent_buffer *next;
2493
2494                         next = path->nodes[orig_level];
2495
2496                         if (trans) {
2497                                 btrfs_tree_lock(next);
2498                                 btrfs_set_lock_blocking(next);
2499                                 clean_tree_block(trans, log->fs_info, next);
2500                                 btrfs_wait_tree_block_writeback(next);
2501                                 btrfs_tree_unlock(next);
2502                         }
2503
2504                         WARN_ON(log->root_key.objectid !=
2505                                 BTRFS_TREE_LOG_OBJECTID);
2506                         ret = btrfs_free_and_pin_reserved_extent(log, next->start,
2507                                                          next->len);
2508                         if (ret)
2509                                 goto out;
2510                 }
2511         }
2512
2513 out:
2514         btrfs_free_path(path);
2515         return ret;
2516 }
2517
2518 /*
2519  * helper function to update the item for a given subvolumes log root
2520  * in the tree of log roots
2521  */
2522 static int update_log_root(struct btrfs_trans_handle *trans,
2523                            struct btrfs_root *log)
2524 {
2525         int ret;
2526
2527         if (log->log_transid == 1) {
2528                 /* insert root item on the first sync */
2529                 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
2530                                 &log->root_key, &log->root_item);
2531         } else {
2532                 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2533                                 &log->root_key, &log->root_item);
2534         }
2535         return ret;
2536 }
2537
2538 static void wait_log_commit(struct btrfs_trans_handle *trans,
2539                             struct btrfs_root *root, int transid)
2540 {
2541         DEFINE_WAIT(wait);
2542         int index = transid % 2;
2543
2544         /*
2545          * we only allow two pending log transactions at a time,
2546          * so we know that if ours is more than 2 older than the
2547          * current transaction, we're done
2548          */
2549         do {
2550                 prepare_to_wait(&root->log_commit_wait[index],
2551                                 &wait, TASK_UNINTERRUPTIBLE);
2552                 mutex_unlock(&root->log_mutex);
2553
2554                 if (root->log_transid_committed < transid &&
2555                     atomic_read(&root->log_commit[index]))
2556                         schedule();
2557
2558                 finish_wait(&root->log_commit_wait[index], &wait);
2559                 mutex_lock(&root->log_mutex);
2560         } while (root->log_transid_committed < transid &&
2561                  atomic_read(&root->log_commit[index]));
2562 }
2563
2564 static void wait_for_writer(struct btrfs_trans_handle *trans,
2565                             struct btrfs_root *root)
2566 {
2567         DEFINE_WAIT(wait);
2568
2569         while (atomic_read(&root->log_writers)) {
2570                 prepare_to_wait(&root->log_writer_wait,
2571                                 &wait, TASK_UNINTERRUPTIBLE);
2572                 mutex_unlock(&root->log_mutex);
2573                 if (atomic_read(&root->log_writers))
2574                         schedule();
2575                 finish_wait(&root->log_writer_wait, &wait);
2576                 mutex_lock(&root->log_mutex);
2577         }
2578 }
2579
2580 static inline void btrfs_remove_log_ctx(struct btrfs_root *root,
2581                                         struct btrfs_log_ctx *ctx)
2582 {
2583         if (!ctx)
2584                 return;
2585
2586         mutex_lock(&root->log_mutex);
2587         list_del_init(&ctx->list);
2588         mutex_unlock(&root->log_mutex);
2589 }
2590
2591 /* 
2592  * Invoked in log mutex context, or be sure there is no other task which
2593  * can access the list.
2594  */
2595 static inline void btrfs_remove_all_log_ctxs(struct btrfs_root *root,
2596                                              int index, int error)
2597 {
2598         struct btrfs_log_ctx *ctx;
2599
2600         if (!error) {
2601                 INIT_LIST_HEAD(&root->log_ctxs[index]);
2602                 return;
2603         }
2604
2605         list_for_each_entry(ctx, &root->log_ctxs[index], list)
2606                 ctx->log_ret = error;
2607
2608         INIT_LIST_HEAD(&root->log_ctxs[index]);
2609 }
2610
2611 /*
2612  * btrfs_sync_log does sends a given tree log down to the disk and
2613  * updates the super blocks to record it.  When this call is done,
2614  * you know that any inodes previously logged are safely on disk only
2615  * if it returns 0.
2616  *
2617  * Any other return value means you need to call btrfs_commit_transaction.
2618  * Some of the edge cases for fsyncing directories that have had unlinks
2619  * or renames done in the past mean that sometimes the only safe
2620  * fsync is to commit the whole FS.  When btrfs_sync_log returns -EAGAIN,
2621  * that has happened.
2622  */
2623 int btrfs_sync_log(struct btrfs_trans_handle *trans,
2624                    struct btrfs_root *root, struct btrfs_log_ctx *ctx)
2625 {
2626         int index1;
2627         int index2;
2628         int mark;
2629         int ret;
2630         struct btrfs_root *log = root->log_root;
2631         struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
2632         int log_transid = 0;
2633         struct btrfs_log_ctx root_log_ctx;
2634         struct blk_plug plug;
2635
2636         mutex_lock(&root->log_mutex);
2637         log_transid = ctx->log_transid;
2638         if (root->log_transid_committed >= log_transid) {
2639                 mutex_unlock(&root->log_mutex);
2640                 return ctx->log_ret;
2641         }
2642
2643         index1 = log_transid % 2;
2644         if (atomic_read(&root->log_commit[index1])) {
2645                 wait_log_commit(trans, root, log_transid);
2646                 mutex_unlock(&root->log_mutex);
2647                 return ctx->log_ret;
2648         }
2649         ASSERT(log_transid == root->log_transid);
2650         atomic_set(&root->log_commit[index1], 1);
2651
2652         /* wait for previous tree log sync to complete */
2653         if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2654                 wait_log_commit(trans, root, log_transid - 1);
2655
2656         while (1) {
2657                 int batch = atomic_read(&root->log_batch);
2658                 /* when we're on an ssd, just kick the log commit out */
2659                 if (!btrfs_test_opt(root, SSD) &&
2660                     test_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state)) {
2661                         mutex_unlock(&root->log_mutex);
2662                         schedule_timeout_uninterruptible(1);
2663                         mutex_lock(&root->log_mutex);
2664                 }
2665                 wait_for_writer(trans, root);
2666                 if (batch == atomic_read(&root->log_batch))
2667                         break;
2668         }
2669
2670         /* bail out if we need to do a full commit */
2671         if (btrfs_need_log_full_commit(root->fs_info, trans)) {
2672                 ret = -EAGAIN;
2673                 btrfs_free_logged_extents(log, log_transid);
2674                 mutex_unlock(&root->log_mutex);
2675                 goto out;
2676         }
2677
2678         if (log_transid % 2 == 0)
2679                 mark = EXTENT_DIRTY;
2680         else
2681                 mark = EXTENT_NEW;
2682
2683         /* we start IO on  all the marked extents here, but we don't actually
2684          * wait for them until later.
2685          */
2686         blk_start_plug(&plug);
2687         ret = btrfs_write_marked_extents(log, &log->dirty_log_pages, mark);
2688         if (ret) {
2689                 blk_finish_plug(&plug);
2690                 btrfs_abort_transaction(trans, root, ret);
2691                 btrfs_free_logged_extents(log, log_transid);
2692                 btrfs_set_log_full_commit(root->fs_info, trans);
2693                 mutex_unlock(&root->log_mutex);
2694                 goto out;
2695         }
2696
2697         btrfs_set_root_node(&log->root_item, log->node);
2698
2699         root->log_transid++;
2700         log->log_transid = root->log_transid;
2701         root->log_start_pid = 0;
2702         /*
2703          * IO has been started, blocks of the log tree have WRITTEN flag set
2704          * in their headers. new modifications of the log will be written to
2705          * new positions. so it's safe to allow log writers to go in.
2706          */
2707         mutex_unlock(&root->log_mutex);
2708
2709         btrfs_init_log_ctx(&root_log_ctx);
2710
2711         mutex_lock(&log_root_tree->log_mutex);
2712         atomic_inc(&log_root_tree->log_batch);
2713         atomic_inc(&log_root_tree->log_writers);
2714
2715         index2 = log_root_tree->log_transid % 2;
2716         list_add_tail(&root_log_ctx.list, &log_root_tree->log_ctxs[index2]);
2717         root_log_ctx.log_transid = log_root_tree->log_transid;
2718
2719         mutex_unlock(&log_root_tree->log_mutex);
2720
2721         ret = update_log_root(trans, log);
2722
2723         mutex_lock(&log_root_tree->log_mutex);
2724         if (atomic_dec_and_test(&log_root_tree->log_writers)) {
2725                 smp_mb();
2726                 if (waitqueue_active(&log_root_tree->log_writer_wait))
2727                         wake_up(&log_root_tree->log_writer_wait);
2728         }
2729
2730         if (ret) {
2731                 if (!list_empty(&root_log_ctx.list))
2732                         list_del_init(&root_log_ctx.list);
2733
2734                 blk_finish_plug(&plug);
2735                 btrfs_set_log_full_commit(root->fs_info, trans);
2736
2737                 if (ret != -ENOSPC) {
2738                         btrfs_abort_transaction(trans, root, ret);
2739                         mutex_unlock(&log_root_tree->log_mutex);
2740                         goto out;
2741                 }
2742                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2743                 btrfs_free_logged_extents(log, log_transid);
2744                 mutex_unlock(&log_root_tree->log_mutex);
2745                 ret = -EAGAIN;
2746                 goto out;
2747         }
2748
2749         if (log_root_tree->log_transid_committed >= root_log_ctx.log_transid) {
2750                 blk_finish_plug(&plug);
2751                 mutex_unlock(&log_root_tree->log_mutex);
2752                 ret = root_log_ctx.log_ret;
2753                 goto out;
2754         }
2755
2756         index2 = root_log_ctx.log_transid % 2;
2757         if (atomic_read(&log_root_tree->log_commit[index2])) {
2758                 blk_finish_plug(&plug);
2759                 ret = btrfs_wait_marked_extents(log, &log->dirty_log_pages,
2760                                                 mark);
2761                 btrfs_wait_logged_extents(trans, log, log_transid);
2762                 wait_log_commit(trans, log_root_tree,
2763                                 root_log_ctx.log_transid);
2764                 mutex_unlock(&log_root_tree->log_mutex);
2765                 if (!ret)
2766                         ret = root_log_ctx.log_ret;
2767                 goto out;
2768         }
2769         ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
2770         atomic_set(&log_root_tree->log_commit[index2], 1);
2771
2772         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2773                 wait_log_commit(trans, log_root_tree,
2774                                 root_log_ctx.log_transid - 1);
2775         }
2776
2777         wait_for_writer(trans, log_root_tree);
2778
2779         /*
2780          * now that we've moved on to the tree of log tree roots,
2781          * check the full commit flag again
2782          */
2783         if (btrfs_need_log_full_commit(root->fs_info, trans)) {
2784                 blk_finish_plug(&plug);
2785                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2786                 btrfs_free_logged_extents(log, log_transid);
2787                 mutex_unlock(&log_root_tree->log_mutex);
2788                 ret = -EAGAIN;
2789                 goto out_wake_log_root;
2790         }
2791
2792         ret = btrfs_write_marked_extents(log_root_tree,
2793                                          &log_root_tree->dirty_log_pages,
2794                                          EXTENT_DIRTY | EXTENT_NEW);
2795         blk_finish_plug(&plug);
2796         if (ret) {
2797                 btrfs_set_log_full_commit(root->fs_info, trans);
2798                 btrfs_abort_transaction(trans, root, ret);
2799                 btrfs_free_logged_extents(log, log_transid);
2800                 mutex_unlock(&log_root_tree->log_mutex);
2801                 goto out_wake_log_root;
2802         }
2803         ret = btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2804         if (!ret)
2805                 ret = btrfs_wait_marked_extents(log_root_tree,
2806                                                 &log_root_tree->dirty_log_pages,
2807                                                 EXTENT_NEW | EXTENT_DIRTY);
2808         if (ret) {
2809                 btrfs_set_log_full_commit(root->fs_info, trans);
2810                 btrfs_free_logged_extents(log, log_transid);
2811                 mutex_unlock(&log_root_tree->log_mutex);
2812                 goto out_wake_log_root;
2813         }
2814         btrfs_wait_logged_extents(trans, log, log_transid);
2815
2816         btrfs_set_super_log_root(root->fs_info->super_for_commit,
2817                                 log_root_tree->node->start);
2818         btrfs_set_super_log_root_level(root->fs_info->super_for_commit,
2819                                 btrfs_header_level(log_root_tree->node));
2820
2821         log_root_tree->log_transid++;
2822         mutex_unlock(&log_root_tree->log_mutex);
2823
2824         /*
2825          * nobody else is going to jump in and write the the ctree
2826          * super here because the log_commit atomic below is protecting
2827          * us.  We must be called with a transaction handle pinning
2828          * the running transaction open, so a full commit can't hop
2829          * in and cause problems either.
2830          */
2831         ret = write_ctree_super(trans, root->fs_info->tree_root, 1);
2832         if (ret) {
2833                 btrfs_set_log_full_commit(root->fs_info, trans);
2834                 btrfs_abort_transaction(trans, root, ret);
2835                 goto out_wake_log_root;
2836         }
2837
2838         mutex_lock(&root->log_mutex);
2839         if (root->last_log_commit < log_transid)
2840                 root->last_log_commit = log_transid;
2841         mutex_unlock(&root->log_mutex);
2842
2843 out_wake_log_root:
2844         /*
2845          * We needn't get log_mutex here because we are sure all
2846          * the other tasks are blocked.
2847          */
2848         btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
2849
2850         mutex_lock(&log_root_tree->log_mutex);
2851         log_root_tree->log_transid_committed++;
2852         atomic_set(&log_root_tree->log_commit[index2], 0);
2853         mutex_unlock(&log_root_tree->log_mutex);
2854
2855         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2856                 wake_up(&log_root_tree->log_commit_wait[index2]);
2857 out:
2858         /* See above. */
2859         btrfs_remove_all_log_ctxs(root, index1, ret);
2860
2861         mutex_lock(&root->log_mutex);
2862         root->log_transid_committed++;
2863         atomic_set(&root->log_commit[index1], 0);
2864         mutex_unlock(&root->log_mutex);
2865
2866         if (waitqueue_active(&root->log_commit_wait[index1]))
2867                 wake_up(&root->log_commit_wait[index1]);
2868         return ret;
2869 }
2870
2871 static void free_log_tree(struct btrfs_trans_handle *trans,
2872                           struct btrfs_root *log)
2873 {
2874         int ret;
2875         u64 start;
2876         u64 end;
2877         struct walk_control wc = {
2878                 .free = 1,
2879                 .process_func = process_one_buffer
2880         };
2881
2882         ret = walk_log_tree(trans, log, &wc);
2883         /* I don't think this can happen but just in case */
2884         if (ret)
2885                 btrfs_abort_transaction(trans, log, ret);
2886
2887         while (1) {
2888                 ret = find_first_extent_bit(&log->dirty_log_pages,
2889                                 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW,
2890                                 NULL);
2891                 if (ret)
2892                         break;
2893
2894                 clear_extent_bits(&log->dirty_log_pages, start, end,
2895                                   EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2896         }
2897
2898         /*
2899          * We may have short-circuited the log tree with the full commit logic
2900          * and left ordered extents on our list, so clear these out to keep us
2901          * from leaking inodes and memory.
2902          */
2903         btrfs_free_logged_extents(log, 0);
2904         btrfs_free_logged_extents(log, 1);
2905
2906         free_extent_buffer(log->node);
2907         kfree(log);
2908 }
2909
2910 /*
2911  * free all the extents used by the tree log.  This should be called
2912  * at commit time of the full transaction
2913  */
2914 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2915 {
2916         if (root->log_root) {
2917                 free_log_tree(trans, root->log_root);
2918                 root->log_root = NULL;
2919         }
2920         return 0;
2921 }
2922
2923 int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2924                              struct btrfs_fs_info *fs_info)
2925 {
2926         if (fs_info->log_root_tree) {
2927                 free_log_tree(trans, fs_info->log_root_tree);
2928                 fs_info->log_root_tree = NULL;
2929         }
2930         return 0;
2931 }
2932
2933 /*
2934  * If both a file and directory are logged, and unlinks or renames are
2935  * mixed in, we have a few interesting corners:
2936  *
2937  * create file X in dir Y
2938  * link file X to X.link in dir Y
2939  * fsync file X
2940  * unlink file X but leave X.link
2941  * fsync dir Y
2942  *
2943  * After a crash we would expect only X.link to exist.  But file X
2944  * didn't get fsync'd again so the log has back refs for X and X.link.
2945  *
2946  * We solve this by removing directory entries and inode backrefs from the
2947  * log when a file that was logged in the current transaction is
2948  * unlinked.  Any later fsync will include the updated log entries, and
2949  * we'll be able to reconstruct the proper directory items from backrefs.
2950  *
2951  * This optimizations allows us to avoid relogging the entire inode
2952  * or the entire directory.
2953  */
2954 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2955                                  struct btrfs_root *root,
2956                                  const char *name, int name_len,
2957                                  struct inode *dir, u64 index)
2958 {
2959         struct btrfs_root *log;
2960         struct btrfs_dir_item *di;
2961         struct btrfs_path *path;
2962         int ret;
2963         int err = 0;
2964         int bytes_del = 0;
2965         u64 dir_ino = btrfs_ino(dir);
2966
2967         if (BTRFS_I(dir)->logged_trans < trans->transid)
2968                 return 0;
2969
2970         ret = join_running_log_trans(root);
2971         if (ret)
2972                 return 0;
2973
2974         mutex_lock(&BTRFS_I(dir)->log_mutex);
2975
2976         log = root->log_root;
2977         path = btrfs_alloc_path();
2978         if (!path) {
2979                 err = -ENOMEM;
2980                 goto out_unlock;
2981         }
2982
2983         di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2984                                    name, name_len, -1);
2985         if (IS_ERR(di)) {
2986                 err = PTR_ERR(di);
2987                 goto fail;
2988         }
2989         if (di) {
2990                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2991                 bytes_del += name_len;
2992                 if (ret) {
2993                         err = ret;
2994                         goto fail;
2995                 }
2996         }
2997         btrfs_release_path(path);
2998         di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2999                                          index, name, name_len, -1);
3000         if (IS_ERR(di)) {
3001                 err = PTR_ERR(di);
3002                 goto fail;
3003         }
3004         if (di) {
3005                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
3006                 bytes_del += name_len;
3007                 if (ret) {
3008                         err = ret;
3009                         goto fail;
3010                 }
3011         }
3012
3013         /* update the directory size in the log to reflect the names
3014          * we have removed
3015          */
3016         if (bytes_del) {
3017                 struct btrfs_key key;
3018
3019                 key.objectid = dir_ino;
3020                 key.offset = 0;
3021                 key.type = BTRFS_INODE_ITEM_KEY;
3022                 btrfs_release_path(path);
3023
3024                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
3025                 if (ret < 0) {
3026                         err = ret;
3027                         goto fail;
3028                 }
3029                 if (ret == 0) {
3030                         struct btrfs_inode_item *item;
3031                         u64 i_size;
3032
3033                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3034                                               struct btrfs_inode_item);
3035                         i_size = btrfs_inode_size(path->nodes[0], item);
3036                         if (i_size > bytes_del)
3037                                 i_size -= bytes_del;
3038                         else
3039                                 i_size = 0;
3040                         btrfs_set_inode_size(path->nodes[0], item, i_size);
3041                         btrfs_mark_buffer_dirty(path->nodes[0]);
3042                 } else
3043                         ret = 0;
3044                 btrfs_release_path(path);
3045         }
3046 fail:
3047         btrfs_free_path(path);
3048 out_unlock:
3049         mutex_unlock(&BTRFS_I(dir)->log_mutex);
3050         if (ret == -ENOSPC) {
3051                 btrfs_set_log_full_commit(root->fs_info, trans);
3052                 ret = 0;
3053         } else if (ret < 0)
3054                 btrfs_abort_transaction(trans, root, ret);
3055
3056         btrfs_end_log_trans(root);
3057
3058         return err;
3059 }
3060
3061 /* see comments for btrfs_del_dir_entries_in_log */
3062 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
3063                                struct btrfs_root *root,
3064                                const char *name, int name_len,
3065                                struct inode *inode, u64 dirid)
3066 {
3067         struct btrfs_root *log;
3068         u64 index;
3069         int ret;
3070
3071         if (BTRFS_I(inode)->logged_trans < trans->transid)
3072                 return 0;
3073
3074         ret = join_running_log_trans(root);
3075         if (ret)
3076                 return 0;
3077         log = root->log_root;
3078         mutex_lock(&BTRFS_I(inode)->log_mutex);
3079
3080         ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
3081                                   dirid, &index);
3082         mutex_unlock(&BTRFS_I(inode)->log_mutex);
3083         if (ret == -ENOSPC) {
3084                 btrfs_set_log_full_commit(root->fs_info, trans);
3085                 ret = 0;
3086         } else if (ret < 0 && ret != -ENOENT)
3087                 btrfs_abort_transaction(trans, root, ret);
3088         btrfs_end_log_trans(root);
3089
3090         return ret;
3091 }
3092
3093 /*
3094  * creates a range item in the log for 'dirid'.  first_offset and
3095  * last_offset tell us which parts of the key space the log should
3096  * be considered authoritative for.
3097  */
3098 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
3099                                        struct btrfs_root *log,
3100                                        struct btrfs_path *path,
3101                                        int key_type, u64 dirid,
3102                                        u64 first_offset, u64 last_offset)
3103 {
3104         int ret;
3105         struct btrfs_key key;
3106         struct btrfs_dir_log_item *item;
3107
3108         key.objectid = dirid;
3109         key.offset = first_offset;
3110         if (key_type == BTRFS_DIR_ITEM_KEY)
3111                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
3112         else
3113                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
3114         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
3115         if (ret)
3116                 return ret;
3117
3118         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3119                               struct btrfs_dir_log_item);
3120         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
3121         btrfs_mark_buffer_dirty(path->nodes[0]);
3122         btrfs_release_path(path);
3123         return 0;
3124 }
3125
3126 /*
3127  * log all the items included in the current transaction for a given
3128  * directory.  This also creates the range items in the log tree required
3129  * to replay anything deleted before the fsync
3130  */
3131 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
3132                           struct btrfs_root *root, struct inode *inode,
3133                           struct btrfs_path *path,
3134                           struct btrfs_path *dst_path, int key_type,
3135                           struct btrfs_log_ctx *ctx,
3136                           u64 min_offset, u64 *last_offset_ret)
3137 {
3138         struct btrfs_key min_key;
3139         struct btrfs_root *log = root->log_root;
3140         struct extent_buffer *src;
3141         int err = 0;
3142         int ret;
3143         int i;
3144         int nritems;
3145         u64 first_offset = min_offset;
3146         u64 last_offset = (u64)-1;
3147         u64 ino = btrfs_ino(inode);
3148
3149         log = root->log_root;
3150
3151         min_key.objectid = ino;
3152         min_key.type = key_type;
3153         min_key.offset = min_offset;
3154
3155         ret = btrfs_search_forward(root, &min_key, path, trans->transid);
3156
3157         /*
3158          * we didn't find anything from this transaction, see if there
3159          * is anything at all
3160          */
3161         if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
3162                 min_key.objectid = ino;
3163                 min_key.type = key_type;
3164                 min_key.offset = (u64)-1;
3165                 btrfs_release_path(path);
3166                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3167                 if (ret < 0) {
3168                         btrfs_release_path(path);
3169                         return ret;
3170                 }
3171                 ret = btrfs_previous_item(root, path, ino, key_type);
3172
3173                 /* if ret == 0 there are items for this type,
3174                  * create a range to tell us the last key of this type.
3175                  * otherwise, there are no items in this directory after
3176                  * *min_offset, and we create a range to indicate that.
3177                  */
3178                 if (ret == 0) {
3179                         struct btrfs_key tmp;
3180                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
3181                                               path->slots[0]);
3182                         if (key_type == tmp.type)
3183                                 first_offset = max(min_offset, tmp.offset) + 1;
3184                 }
3185                 goto done;
3186         }
3187
3188         /* go backward to find any previous key */
3189         ret = btrfs_previous_item(root, path, ino, key_type);
3190         if (ret == 0) {
3191                 struct btrfs_key tmp;
3192                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3193                 if (key_type == tmp.type) {
3194                         first_offset = tmp.offset;
3195                         ret = overwrite_item(trans, log, dst_path,
3196                                              path->nodes[0], path->slots[0],
3197                                              &tmp);
3198                         if (ret) {
3199                                 err = ret;
3200                                 goto done;
3201                         }
3202                 }
3203         }
3204         btrfs_release_path(path);
3205
3206         /* find the first key from this transaction again */
3207         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3208         if (WARN_ON(ret != 0))
3209                 goto done;
3210
3211         /*
3212          * we have a block from this transaction, log every item in it
3213          * from our directory
3214          */
3215         while (1) {
3216                 struct btrfs_key tmp;
3217                 src = path->nodes[0];
3218                 nritems = btrfs_header_nritems(src);
3219                 for (i = path->slots[0]; i < nritems; i++) {
3220                         struct btrfs_dir_item *di;
3221
3222                         btrfs_item_key_to_cpu(src, &min_key, i);
3223
3224                         if (min_key.objectid != ino || min_key.type != key_type)
3225                                 goto done;
3226                         ret = overwrite_item(trans, log, dst_path, src, i,
3227                                              &min_key);
3228                         if (ret) {
3229                                 err = ret;
3230                                 goto done;
3231                         }
3232
3233                         /*
3234                          * We must make sure that when we log a directory entry,
3235                          * the corresponding inode, after log replay, has a
3236                          * matching link count. For example:
3237                          *
3238                          * touch foo
3239                          * mkdir mydir
3240                          * sync
3241                          * ln foo mydir/bar
3242                          * xfs_io -c "fsync" mydir
3243                          * <crash>
3244                          * <mount fs and log replay>
3245                          *
3246                          * Would result in a fsync log that when replayed, our
3247                          * file inode would have a link count of 1, but we get
3248                          * two directory entries pointing to the same inode.
3249                          * After removing one of the names, it would not be
3250                          * possible to remove the other name, which resulted
3251                          * always in stale file handle errors, and would not
3252                          * be possible to rmdir the parent directory, since
3253                          * its i_size could never decrement to the value
3254                          * BTRFS_EMPTY_DIR_SIZE, resulting in -ENOTEMPTY errors.
3255                          */
3256                         di = btrfs_item_ptr(src, i, struct btrfs_dir_item);
3257                         btrfs_dir_item_key_to_cpu(src, di, &tmp);
3258                         if (ctx &&
3259                             (btrfs_dir_transid(src, di) == trans->transid ||
3260                              btrfs_dir_type(src, di) == BTRFS_FT_DIR) &&
3261                             tmp.type != BTRFS_ROOT_ITEM_KEY)
3262                                 ctx->log_new_dentries = true;
3263                 }
3264                 path->slots[0] = nritems;
3265
3266                 /*
3267                  * look ahead to the next item and see if it is also
3268                  * from this directory and from this transaction
3269                  */
3270                 ret = btrfs_next_leaf(root, path);
3271                 if (ret == 1) {
3272                         last_offset = (u64)-1;
3273                         goto done;
3274                 }
3275                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3276                 if (tmp.objectid != ino || tmp.type != key_type) {
3277                         last_offset = (u64)-1;
3278                         goto done;
3279                 }
3280                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
3281                         ret = overwrite_item(trans, log, dst_path,
3282                                              path->nodes[0], path->slots[0],
3283                                              &tmp);
3284                         if (ret)
3285                                 err = ret;
3286                         else
3287                                 last_offset = tmp.offset;
3288                         goto done;
3289                 }
3290         }
3291 done:
3292         btrfs_release_path(path);
3293         btrfs_release_path(dst_path);
3294
3295         if (err == 0) {
3296                 *last_offset_ret = last_offset;
3297                 /*
3298                  * insert the log range keys to indicate where the log
3299                  * is valid
3300                  */
3301                 ret = insert_dir_log_key(trans, log, path, key_type,
3302                                          ino, first_offset, last_offset);
3303                 if (ret)
3304                         err = ret;
3305         }
3306         return err;
3307 }
3308
3309 /*
3310  * logging directories is very similar to logging inodes, We find all the items
3311  * from the current transaction and write them to the log.
3312  *
3313  * The recovery code scans the directory in the subvolume, and if it finds a
3314  * key in the range logged that is not present in the log tree, then it means
3315  * that dir entry was unlinked during the transaction.
3316  *
3317  * In order for that scan to work, we must include one key smaller than
3318  * the smallest logged by this transaction and one key larger than the largest
3319  * key logged by this transaction.
3320  */