Loading...
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 "ctree.h"
22#include "transaction.h"
23#include "disk-io.h"
24#include "locking.h"
25#include "print-tree.h"
26#include "compat.h"
27#include "tree-log.h"
28
29/* magic values for the inode_only field in btrfs_log_inode:
30 *
31 * LOG_INODE_ALL means to log everything
32 * LOG_INODE_EXISTS means to log just enough to recreate the inode
33 * during log replay
34 */
35#define LOG_INODE_ALL 0
36#define LOG_INODE_EXISTS 1
37
38/*
39 * directory trouble cases
40 *
41 * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
42 * log, we must force a full commit before doing an fsync of the directory
43 * where the unlink was done.
44 * ---> record transid of last unlink/rename per directory
45 *
46 * mkdir foo/some_dir
47 * normal commit
48 * rename foo/some_dir foo2/some_dir
49 * mkdir foo/some_dir
50 * fsync foo/some_dir/some_file
51 *
52 * The fsync above will unlink the original some_dir without recording
53 * it in its new location (foo2). After a crash, some_dir will be gone
54 * unless the fsync of some_file forces a full commit
55 *
56 * 2) we must log any new names for any file or dir that is in the fsync
57 * log. ---> check inode while renaming/linking.
58 *
59 * 2a) we must log any new names for any file or dir during rename
60 * when the directory they are being removed from was logged.
61 * ---> check inode and old parent dir during rename
62 *
63 * 2a is actually the more important variant. With the extra logging
64 * a crash might unlink the old name without recreating the new one
65 *
66 * 3) after a crash, we must go through any directories with a link count
67 * of zero and redo the rm -rf
68 *
69 * mkdir f1/foo
70 * normal commit
71 * rm -rf f1/foo
72 * fsync(f1)
73 *
74 * The directory f1 was fully removed from the FS, but fsync was never
75 * called on f1, only its parent dir. After a crash the rm -rf must
76 * be replayed. This must be able to recurse down the entire
77 * directory tree. The inode link count fixup code takes care of the
78 * ugly details.
79 */
80
81/*
82 * stages for the tree walking. The first
83 * stage (0) is to only pin down the blocks we find
84 * the second stage (1) is to make sure that all the inodes
85 * we find in the log are created in the subvolume.
86 *
87 * The last stage is to deal with directories and links and extents
88 * and all the other fun semantics
89 */
90#define LOG_WALK_PIN_ONLY 0
91#define LOG_WALK_REPLAY_INODES 1
92#define LOG_WALK_REPLAY_ALL 2
93
94static int btrfs_log_inode(struct btrfs_trans_handle *trans,
95 struct btrfs_root *root, struct inode *inode,
96 int inode_only);
97static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
98 struct btrfs_root *root,
99 struct btrfs_path *path, u64 objectid);
100static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
101 struct btrfs_root *root,
102 struct btrfs_root *log,
103 struct btrfs_path *path,
104 u64 dirid, int del_all);
105
106/*
107 * tree logging is a special write ahead log used to make sure that
108 * fsyncs and O_SYNCs can happen without doing full tree commits.
109 *
110 * Full tree commits are expensive because they require commonly
111 * modified blocks to be recowed, creating many dirty pages in the
112 * extent tree an 4x-6x higher write load than ext3.
113 *
114 * Instead of doing a tree commit on every fsync, we use the
115 * key ranges and transaction ids to find items for a given file or directory
116 * that have changed in this transaction. Those items are copied into
117 * a special tree (one per subvolume root), that tree is written to disk
118 * and then the fsync is considered complete.
119 *
120 * After a crash, items are copied out of the log-tree back into the
121 * subvolume tree. Any file data extents found are recorded in the extent
122 * allocation tree, and the log-tree freed.
123 *
124 * The log tree is read three times, once to pin down all the extents it is
125 * using in ram and once, once to create all the inodes logged in the tree
126 * and once to do all the other items.
127 */
128
129/*
130 * start a sub transaction and setup the log tree
131 * this increments the log tree writer count to make the people
132 * syncing the tree wait for us to finish
133 */
134static int start_log_trans(struct btrfs_trans_handle *trans,
135 struct btrfs_root *root)
136{
137 int ret;
138 int err = 0;
139
140 mutex_lock(&root->log_mutex);
141 if (root->log_root) {
142 if (!root->log_start_pid) {
143 root->log_start_pid = current->pid;
144 root->log_multiple_pids = false;
145 } else if (root->log_start_pid != current->pid) {
146 root->log_multiple_pids = true;
147 }
148
149 root->log_batch++;
150 atomic_inc(&root->log_writers);
151 mutex_unlock(&root->log_mutex);
152 return 0;
153 }
154 root->log_multiple_pids = false;
155 root->log_start_pid = current->pid;
156 mutex_lock(&root->fs_info->tree_log_mutex);
157 if (!root->fs_info->log_root_tree) {
158 ret = btrfs_init_log_root_tree(trans, root->fs_info);
159 if (ret)
160 err = ret;
161 }
162 if (err == 0 && !root->log_root) {
163 ret = btrfs_add_log_tree(trans, root);
164 if (ret)
165 err = ret;
166 }
167 mutex_unlock(&root->fs_info->tree_log_mutex);
168 root->log_batch++;
169 atomic_inc(&root->log_writers);
170 mutex_unlock(&root->log_mutex);
171 return err;
172}
173
174/*
175 * returns 0 if there was a log transaction running and we were able
176 * to join, or returns -ENOENT if there were not transactions
177 * in progress
178 */
179static int join_running_log_trans(struct btrfs_root *root)
180{
181 int ret = -ENOENT;
182
183 smp_mb();
184 if (!root->log_root)
185 return -ENOENT;
186
187 mutex_lock(&root->log_mutex);
188 if (root->log_root) {
189 ret = 0;
190 atomic_inc(&root->log_writers);
191 }
192 mutex_unlock(&root->log_mutex);
193 return ret;
194}
195
196/*
197 * This either makes the current running log transaction wait
198 * until you call btrfs_end_log_trans() or it makes any future
199 * log transactions wait until you call btrfs_end_log_trans()
200 */
201int btrfs_pin_log_trans(struct btrfs_root *root)
202{
203 int ret = -ENOENT;
204
205 mutex_lock(&root->log_mutex);
206 atomic_inc(&root->log_writers);
207 mutex_unlock(&root->log_mutex);
208 return ret;
209}
210
211/*
212 * indicate we're done making changes to the log tree
213 * and wake up anyone waiting to do a sync
214 */
215int btrfs_end_log_trans(struct btrfs_root *root)
216{
217 if (atomic_dec_and_test(&root->log_writers)) {
218 smp_mb();
219 if (waitqueue_active(&root->log_writer_wait))
220 wake_up(&root->log_writer_wait);
221 }
222 return 0;
223}
224
225
226/*
227 * the walk control struct is used to pass state down the chain when
228 * processing the log tree. The stage field tells us which part
229 * of the log tree processing we are currently doing. The others
230 * are state fields used for that specific part
231 */
232struct walk_control {
233 /* should we free the extent on disk when done? This is used
234 * at transaction commit time while freeing a log tree
235 */
236 int free;
237
238 /* should we write out the extent buffer? This is used
239 * while flushing the log tree to disk during a sync
240 */
241 int write;
242
243 /* should we wait for the extent buffer io to finish? Also used
244 * while flushing the log tree to disk for a sync
245 */
246 int wait;
247
248 /* pin only walk, we record which extents on disk belong to the
249 * log trees
250 */
251 int pin;
252
253 /* what stage of the replay code we're currently in */
254 int stage;
255
256 /* the root we are currently replaying */
257 struct btrfs_root *replay_dest;
258
259 /* the trans handle for the current replay */
260 struct btrfs_trans_handle *trans;
261
262 /* the function that gets used to process blocks we find in the
263 * tree. Note the extent_buffer might not be up to date when it is
264 * passed in, and it must be checked or read if you need the data
265 * inside it
266 */
267 int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
268 struct walk_control *wc, u64 gen);
269};
270
271/*
272 * process_func used to pin down extents, write them or wait on them
273 */
274static int process_one_buffer(struct btrfs_root *log,
275 struct extent_buffer *eb,
276 struct walk_control *wc, u64 gen)
277{
278 if (wc->pin)
279 btrfs_pin_extent(log->fs_info->extent_root,
280 eb->start, eb->len, 0);
281
282 if (btrfs_buffer_uptodate(eb, gen)) {
283 if (wc->write)
284 btrfs_write_tree_block(eb);
285 if (wc->wait)
286 btrfs_wait_tree_block_writeback(eb);
287 }
288 return 0;
289}
290
291/*
292 * Item overwrite used by replay and tree logging. eb, slot and key all refer
293 * to the src data we are copying out.
294 *
295 * root is the tree we are copying into, and path is a scratch
296 * path for use in this function (it should be released on entry and
297 * will be released on exit).
298 *
299 * If the key is already in the destination tree the existing item is
300 * overwritten. If the existing item isn't big enough, it is extended.
301 * If it is too large, it is truncated.
302 *
303 * If the key isn't in the destination yet, a new item is inserted.
304 */
305static noinline int overwrite_item(struct btrfs_trans_handle *trans,
306 struct btrfs_root *root,
307 struct btrfs_path *path,
308 struct extent_buffer *eb, int slot,
309 struct btrfs_key *key)
310{
311 int ret;
312 u32 item_size;
313 u64 saved_i_size = 0;
314 int save_old_i_size = 0;
315 unsigned long src_ptr;
316 unsigned long dst_ptr;
317 int overwrite_root = 0;
318
319 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
320 overwrite_root = 1;
321
322 item_size = btrfs_item_size_nr(eb, slot);
323 src_ptr = btrfs_item_ptr_offset(eb, slot);
324
325 /* look for the key in the destination tree */
326 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
327 if (ret == 0) {
328 char *src_copy;
329 char *dst_copy;
330 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
331 path->slots[0]);
332 if (dst_size != item_size)
333 goto insert;
334
335 if (item_size == 0) {
336 btrfs_release_path(path);
337 return 0;
338 }
339 dst_copy = kmalloc(item_size, GFP_NOFS);
340 src_copy = kmalloc(item_size, GFP_NOFS);
341 if (!dst_copy || !src_copy) {
342 btrfs_release_path(path);
343 kfree(dst_copy);
344 kfree(src_copy);
345 return -ENOMEM;
346 }
347
348 read_extent_buffer(eb, src_copy, src_ptr, item_size);
349
350 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
351 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
352 item_size);
353 ret = memcmp(dst_copy, src_copy, item_size);
354
355 kfree(dst_copy);
356 kfree(src_copy);
357 /*
358 * they have the same contents, just return, this saves
359 * us from cowing blocks in the destination tree and doing
360 * extra writes that may not have been done by a previous
361 * sync
362 */
363 if (ret == 0) {
364 btrfs_release_path(path);
365 return 0;
366 }
367
368 }
369insert:
370 btrfs_release_path(path);
371 /* try to insert the key into the destination tree */
372 ret = btrfs_insert_empty_item(trans, root, path,
373 key, item_size);
374
375 /* make sure any existing item is the correct size */
376 if (ret == -EEXIST) {
377 u32 found_size;
378 found_size = btrfs_item_size_nr(path->nodes[0],
379 path->slots[0]);
380 if (found_size > item_size) {
381 btrfs_truncate_item(trans, root, path, item_size, 1);
382 } else if (found_size < item_size) {
383 ret = btrfs_extend_item(trans, root, path,
384 item_size - found_size);
385 }
386 } else if (ret) {
387 return ret;
388 }
389 dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
390 path->slots[0]);
391
392 /* don't overwrite an existing inode if the generation number
393 * was logged as zero. This is done when the tree logging code
394 * is just logging an inode to make sure it exists after recovery.
395 *
396 * Also, don't overwrite i_size on directories during replay.
397 * log replay inserts and removes directory items based on the
398 * state of the tree found in the subvolume, and i_size is modified
399 * as it goes
400 */
401 if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
402 struct btrfs_inode_item *src_item;
403 struct btrfs_inode_item *dst_item;
404
405 src_item = (struct btrfs_inode_item *)src_ptr;
406 dst_item = (struct btrfs_inode_item *)dst_ptr;
407
408 if (btrfs_inode_generation(eb, src_item) == 0)
409 goto no_copy;
410
411 if (overwrite_root &&
412 S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
413 S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
414 save_old_i_size = 1;
415 saved_i_size = btrfs_inode_size(path->nodes[0],
416 dst_item);
417 }
418 }
419
420 copy_extent_buffer(path->nodes[0], eb, dst_ptr,
421 src_ptr, item_size);
422
423 if (save_old_i_size) {
424 struct btrfs_inode_item *dst_item;
425 dst_item = (struct btrfs_inode_item *)dst_ptr;
426 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
427 }
428
429 /* make sure the generation is filled in */
430 if (key->type == BTRFS_INODE_ITEM_KEY) {
431 struct btrfs_inode_item *dst_item;
432 dst_item = (struct btrfs_inode_item *)dst_ptr;
433 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
434 btrfs_set_inode_generation(path->nodes[0], dst_item,
435 trans->transid);
436 }
437 }
438no_copy:
439 btrfs_mark_buffer_dirty(path->nodes[0]);
440 btrfs_release_path(path);
441 return 0;
442}
443
444/*
445 * simple helper to read an inode off the disk from a given root
446 * This can only be called for subvolume roots and not for the log
447 */
448static noinline struct inode *read_one_inode(struct btrfs_root *root,
449 u64 objectid)
450{
451 struct btrfs_key key;
452 struct inode *inode;
453
454 key.objectid = objectid;
455 key.type = BTRFS_INODE_ITEM_KEY;
456 key.offset = 0;
457 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
458 if (IS_ERR(inode)) {
459 inode = NULL;
460 } else if (is_bad_inode(inode)) {
461 iput(inode);
462 inode = NULL;
463 }
464 return inode;
465}
466
467/* replays a single extent in 'eb' at 'slot' with 'key' into the
468 * subvolume 'root'. path is released on entry and should be released
469 * on exit.
470 *
471 * extents in the log tree have not been allocated out of the extent
472 * tree yet. So, this completes the allocation, taking a reference
473 * as required if the extent already exists or creating a new extent
474 * if it isn't in the extent allocation tree yet.
475 *
476 * The extent is inserted into the file, dropping any existing extents
477 * from the file that overlap the new one.
478 */
479static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
480 struct btrfs_root *root,
481 struct btrfs_path *path,
482 struct extent_buffer *eb, int slot,
483 struct btrfs_key *key)
484{
485 int found_type;
486 u64 mask = root->sectorsize - 1;
487 u64 extent_end;
488 u64 alloc_hint;
489 u64 start = key->offset;
490 u64 saved_nbytes;
491 struct btrfs_file_extent_item *item;
492 struct inode *inode = NULL;
493 unsigned long size;
494 int ret = 0;
495
496 item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
497 found_type = btrfs_file_extent_type(eb, item);
498
499 if (found_type == BTRFS_FILE_EXTENT_REG ||
500 found_type == BTRFS_FILE_EXTENT_PREALLOC)
501 extent_end = start + btrfs_file_extent_num_bytes(eb, item);
502 else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
503 size = btrfs_file_extent_inline_len(eb, item);
504 extent_end = (start + size + mask) & ~mask;
505 } else {
506 ret = 0;
507 goto out;
508 }
509
510 inode = read_one_inode(root, key->objectid);
511 if (!inode) {
512 ret = -EIO;
513 goto out;
514 }
515
516 /*
517 * first check to see if we already have this extent in the
518 * file. This must be done before the btrfs_drop_extents run
519 * so we don't try to drop this extent.
520 */
521 ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
522 start, 0);
523
524 if (ret == 0 &&
525 (found_type == BTRFS_FILE_EXTENT_REG ||
526 found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
527 struct btrfs_file_extent_item cmp1;
528 struct btrfs_file_extent_item cmp2;
529 struct btrfs_file_extent_item *existing;
530 struct extent_buffer *leaf;
531
532 leaf = path->nodes[0];
533 existing = btrfs_item_ptr(leaf, path->slots[0],
534 struct btrfs_file_extent_item);
535
536 read_extent_buffer(eb, &cmp1, (unsigned long)item,
537 sizeof(cmp1));
538 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
539 sizeof(cmp2));
540
541 /*
542 * we already have a pointer to this exact extent,
543 * we don't have to do anything
544 */
545 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
546 btrfs_release_path(path);
547 goto out;
548 }
549 }
550 btrfs_release_path(path);
551
552 saved_nbytes = inode_get_bytes(inode);
553 /* drop any overlapping extents */
554 ret = btrfs_drop_extents(trans, inode, start, extent_end,
555 &alloc_hint, 1);
556 BUG_ON(ret);
557
558 if (found_type == BTRFS_FILE_EXTENT_REG ||
559 found_type == BTRFS_FILE_EXTENT_PREALLOC) {
560 u64 offset;
561 unsigned long dest_offset;
562 struct btrfs_key ins;
563
564 ret = btrfs_insert_empty_item(trans, root, path, key,
565 sizeof(*item));
566 BUG_ON(ret);
567 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
568 path->slots[0]);
569 copy_extent_buffer(path->nodes[0], eb, dest_offset,
570 (unsigned long)item, sizeof(*item));
571
572 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
573 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
574 ins.type = BTRFS_EXTENT_ITEM_KEY;
575 offset = key->offset - btrfs_file_extent_offset(eb, item);
576
577 if (ins.objectid > 0) {
578 u64 csum_start;
579 u64 csum_end;
580 LIST_HEAD(ordered_sums);
581 /*
582 * is this extent already allocated in the extent
583 * allocation tree? If so, just add a reference
584 */
585 ret = btrfs_lookup_extent(root, ins.objectid,
586 ins.offset);
587 if (ret == 0) {
588 ret = btrfs_inc_extent_ref(trans, root,
589 ins.objectid, ins.offset,
590 0, root->root_key.objectid,
591 key->objectid, offset);
592 BUG_ON(ret);
593 } else {
594 /*
595 * insert the extent pointer in the extent
596 * allocation tree
597 */
598 ret = btrfs_alloc_logged_file_extent(trans,
599 root, root->root_key.objectid,
600 key->objectid, offset, &ins);
601 BUG_ON(ret);
602 }
603 btrfs_release_path(path);
604
605 if (btrfs_file_extent_compression(eb, item)) {
606 csum_start = ins.objectid;
607 csum_end = csum_start + ins.offset;
608 } else {
609 csum_start = ins.objectid +
610 btrfs_file_extent_offset(eb, item);
611 csum_end = csum_start +
612 btrfs_file_extent_num_bytes(eb, item);
613 }
614
615 ret = btrfs_lookup_csums_range(root->log_root,
616 csum_start, csum_end - 1,
617 &ordered_sums, 0);
618 BUG_ON(ret);
619 while (!list_empty(&ordered_sums)) {
620 struct btrfs_ordered_sum *sums;
621 sums = list_entry(ordered_sums.next,
622 struct btrfs_ordered_sum,
623 list);
624 ret = btrfs_csum_file_blocks(trans,
625 root->fs_info->csum_root,
626 sums);
627 BUG_ON(ret);
628 list_del(&sums->list);
629 kfree(sums);
630 }
631 } else {
632 btrfs_release_path(path);
633 }
634 } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
635 /* inline extents are easy, we just overwrite them */
636 ret = overwrite_item(trans, root, path, eb, slot, key);
637 BUG_ON(ret);
638 }
639
640 inode_set_bytes(inode, saved_nbytes);
641 btrfs_update_inode(trans, root, inode);
642out:
643 if (inode)
644 iput(inode);
645 return ret;
646}
647
648/*
649 * when cleaning up conflicts between the directory names in the
650 * subvolume, directory names in the log and directory names in the
651 * inode back references, we may have to unlink inodes from directories.
652 *
653 * This is a helper function to do the unlink of a specific directory
654 * item
655 */
656static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
657 struct btrfs_root *root,
658 struct btrfs_path *path,
659 struct inode *dir,
660 struct btrfs_dir_item *di)
661{
662 struct inode *inode;
663 char *name;
664 int name_len;
665 struct extent_buffer *leaf;
666 struct btrfs_key location;
667 int ret;
668
669 leaf = path->nodes[0];
670
671 btrfs_dir_item_key_to_cpu(leaf, di, &location);
672 name_len = btrfs_dir_name_len(leaf, di);
673 name = kmalloc(name_len, GFP_NOFS);
674 if (!name)
675 return -ENOMEM;
676
677 read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
678 btrfs_release_path(path);
679
680 inode = read_one_inode(root, location.objectid);
681 if (!inode) {
682 kfree(name);
683 return -EIO;
684 }
685
686 ret = link_to_fixup_dir(trans, root, path, location.objectid);
687 BUG_ON(ret);
688
689 ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
690 BUG_ON(ret);
691 kfree(name);
692
693 iput(inode);
694 return ret;
695}
696
697/*
698 * helper function to see if a given name and sequence number found
699 * in an inode back reference are already in a directory and correctly
700 * point to this inode
701 */
702static noinline int inode_in_dir(struct btrfs_root *root,
703 struct btrfs_path *path,
704 u64 dirid, u64 objectid, u64 index,
705 const char *name, int name_len)
706{
707 struct btrfs_dir_item *di;
708 struct btrfs_key location;
709 int match = 0;
710
711 di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
712 index, name, name_len, 0);
713 if (di && !IS_ERR(di)) {
714 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
715 if (location.objectid != objectid)
716 goto out;
717 } else
718 goto out;
719 btrfs_release_path(path);
720
721 di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
722 if (di && !IS_ERR(di)) {
723 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
724 if (location.objectid != objectid)
725 goto out;
726 } else
727 goto out;
728 match = 1;
729out:
730 btrfs_release_path(path);
731 return match;
732}
733
734/*
735 * helper function to check a log tree for a named back reference in
736 * an inode. This is used to decide if a back reference that is
737 * found in the subvolume conflicts with what we find in the log.
738 *
739 * inode backreferences may have multiple refs in a single item,
740 * during replay we process one reference at a time, and we don't
741 * want to delete valid links to a file from the subvolume if that
742 * link is also in the log.
743 */
744static noinline int backref_in_log(struct btrfs_root *log,
745 struct btrfs_key *key,
746 char *name, int namelen)
747{
748 struct btrfs_path *path;
749 struct btrfs_inode_ref *ref;
750 unsigned long ptr;
751 unsigned long ptr_end;
752 unsigned long name_ptr;
753 int found_name_len;
754 int item_size;
755 int ret;
756 int match = 0;
757
758 path = btrfs_alloc_path();
759 if (!path)
760 return -ENOMEM;
761
762 ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
763 if (ret != 0)
764 goto out;
765
766 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
767 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
768 ptr_end = ptr + item_size;
769 while (ptr < ptr_end) {
770 ref = (struct btrfs_inode_ref *)ptr;
771 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
772 if (found_name_len == namelen) {
773 name_ptr = (unsigned long)(ref + 1);
774 ret = memcmp_extent_buffer(path->nodes[0], name,
775 name_ptr, namelen);
776 if (ret == 0) {
777 match = 1;
778 goto out;
779 }
780 }
781 ptr = (unsigned long)(ref + 1) + found_name_len;
782 }
783out:
784 btrfs_free_path(path);
785 return match;
786}
787
788
789/*
790 * replay one inode back reference item found in the log tree.
791 * eb, slot and key refer to the buffer and key found in the log tree.
792 * root is the destination we are replaying into, and path is for temp
793 * use by this function. (it should be released on return).
794 */
795static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
796 struct btrfs_root *root,
797 struct btrfs_root *log,
798 struct btrfs_path *path,
799 struct extent_buffer *eb, int slot,
800 struct btrfs_key *key)
801{
802 struct btrfs_inode_ref *ref;
803 struct btrfs_dir_item *di;
804 struct inode *dir;
805 struct inode *inode;
806 unsigned long ref_ptr;
807 unsigned long ref_end;
808 char *name;
809 int namelen;
810 int ret;
811 int search_done = 0;
812
813 /*
814 * it is possible that we didn't log all the parent directories
815 * for a given inode. If we don't find the dir, just don't
816 * copy the back ref in. The link count fixup code will take
817 * care of the rest
818 */
819 dir = read_one_inode(root, key->offset);
820 if (!dir)
821 return -ENOENT;
822
823 inode = read_one_inode(root, key->objectid);
824 if (!inode) {
825 iput(dir);
826 return -EIO;
827 }
828
829 ref_ptr = btrfs_item_ptr_offset(eb, slot);
830 ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
831
832again:
833 ref = (struct btrfs_inode_ref *)ref_ptr;
834
835 namelen = btrfs_inode_ref_name_len(eb, ref);
836 name = kmalloc(namelen, GFP_NOFS);
837 BUG_ON(!name);
838
839 read_extent_buffer(eb, name, (unsigned long)(ref + 1), namelen);
840
841 /* if we already have a perfect match, we're done */
842 if (inode_in_dir(root, path, btrfs_ino(dir), btrfs_ino(inode),
843 btrfs_inode_ref_index(eb, ref),
844 name, namelen)) {
845 goto out;
846 }
847
848 /*
849 * look for a conflicting back reference in the metadata.
850 * if we find one we have to unlink that name of the file
851 * before we add our new link. Later on, we overwrite any
852 * existing back reference, and we don't want to create
853 * dangling pointers in the directory.
854 */
855
856 if (search_done)
857 goto insert;
858
859 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
860 if (ret == 0) {
861 char *victim_name;
862 int victim_name_len;
863 struct btrfs_inode_ref *victim_ref;
864 unsigned long ptr;
865 unsigned long ptr_end;
866 struct extent_buffer *leaf = path->nodes[0];
867
868 /* are we trying to overwrite a back ref for the root directory
869 * if so, just jump out, we're done
870 */
871 if (key->objectid == key->offset)
872 goto out_nowrite;
873
874 /* check all the names in this back reference to see
875 * if they are in the log. if so, we allow them to stay
876 * otherwise they must be unlinked as a conflict
877 */
878 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
879 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
880 while (ptr < ptr_end) {
881 victim_ref = (struct btrfs_inode_ref *)ptr;
882 victim_name_len = btrfs_inode_ref_name_len(leaf,
883 victim_ref);
884 victim_name = kmalloc(victim_name_len, GFP_NOFS);
885 BUG_ON(!victim_name);
886
887 read_extent_buffer(leaf, victim_name,
888 (unsigned long)(victim_ref + 1),
889 victim_name_len);
890
891 if (!backref_in_log(log, key, victim_name,
892 victim_name_len)) {
893 btrfs_inc_nlink(inode);
894 btrfs_release_path(path);
895
896 ret = btrfs_unlink_inode(trans, root, dir,
897 inode, victim_name,
898 victim_name_len);
899 }
900 kfree(victim_name);
901 ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
902 }
903 BUG_ON(ret);
904
905 /*
906 * NOTE: we have searched root tree and checked the
907 * coresponding ref, it does not need to check again.
908 */
909 search_done = 1;
910 }
911 btrfs_release_path(path);
912
913 /* look for a conflicting sequence number */
914 di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
915 btrfs_inode_ref_index(eb, ref),
916 name, namelen, 0);
917 if (di && !IS_ERR(di)) {
918 ret = drop_one_dir_item(trans, root, path, dir, di);
919 BUG_ON(ret);
920 }
921 btrfs_release_path(path);
922
923 /* look for a conflicing name */
924 di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
925 name, namelen, 0);
926 if (di && !IS_ERR(di)) {
927 ret = drop_one_dir_item(trans, root, path, dir, di);
928 BUG_ON(ret);
929 }
930 btrfs_release_path(path);
931
932insert:
933 /* insert our name */
934 ret = btrfs_add_link(trans, dir, inode, name, namelen, 0,
935 btrfs_inode_ref_index(eb, ref));
936 BUG_ON(ret);
937
938 btrfs_update_inode(trans, root, inode);
939
940out:
941 ref_ptr = (unsigned long)(ref + 1) + namelen;
942 kfree(name);
943 if (ref_ptr < ref_end)
944 goto again;
945
946 /* finally write the back reference in the inode */
947 ret = overwrite_item(trans, root, path, eb, slot, key);
948 BUG_ON(ret);
949
950out_nowrite:
951 btrfs_release_path(path);
952 iput(dir);
953 iput(inode);
954 return 0;
955}
956
957static int insert_orphan_item(struct btrfs_trans_handle *trans,
958 struct btrfs_root *root, u64 offset)
959{
960 int ret;
961 ret = btrfs_find_orphan_item(root, offset);
962 if (ret > 0)
963 ret = btrfs_insert_orphan_item(trans, root, offset);
964 return ret;
965}
966
967
968/*
969 * There are a few corners where the link count of the file can't
970 * be properly maintained during replay. So, instead of adding
971 * lots of complexity to the log code, we just scan the backrefs
972 * for any file that has been through replay.
973 *
974 * The scan will update the link count on the inode to reflect the
975 * number of back refs found. If it goes down to zero, the iput
976 * will free the inode.
977 */
978static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
979 struct btrfs_root *root,
980 struct inode *inode)
981{
982 struct btrfs_path *path;
983 int ret;
984 struct btrfs_key key;
985 u64 nlink = 0;
986 unsigned long ptr;
987 unsigned long ptr_end;
988 int name_len;
989 u64 ino = btrfs_ino(inode);
990
991 key.objectid = ino;
992 key.type = BTRFS_INODE_REF_KEY;
993 key.offset = (u64)-1;
994
995 path = btrfs_alloc_path();
996 if (!path)
997 return -ENOMEM;
998
999 while (1) {
1000 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1001 if (ret < 0)
1002 break;
1003 if (ret > 0) {
1004 if (path->slots[0] == 0)
1005 break;
1006 path->slots[0]--;
1007 }
1008 btrfs_item_key_to_cpu(path->nodes[0], &key,
1009 path->slots[0]);
1010 if (key.objectid != ino ||
1011 key.type != BTRFS_INODE_REF_KEY)
1012 break;
1013 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1014 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1015 path->slots[0]);
1016 while (ptr < ptr_end) {
1017 struct btrfs_inode_ref *ref;
1018
1019 ref = (struct btrfs_inode_ref *)ptr;
1020 name_len = btrfs_inode_ref_name_len(path->nodes[0],
1021 ref);
1022 ptr = (unsigned long)(ref + 1) + name_len;
1023 nlink++;
1024 }
1025
1026 if (key.offset == 0)
1027 break;
1028 key.offset--;
1029 btrfs_release_path(path);
1030 }
1031 btrfs_release_path(path);
1032 if (nlink != inode->i_nlink) {
1033 inode->i_nlink = nlink;
1034 btrfs_update_inode(trans, root, inode);
1035 }
1036 BTRFS_I(inode)->index_cnt = (u64)-1;
1037
1038 if (inode->i_nlink == 0) {
1039 if (S_ISDIR(inode->i_mode)) {
1040 ret = replay_dir_deletes(trans, root, NULL, path,
1041 ino, 1);
1042 BUG_ON(ret);
1043 }
1044 ret = insert_orphan_item(trans, root, ino);
1045 BUG_ON(ret);
1046 }
1047 btrfs_free_path(path);
1048
1049 return 0;
1050}
1051
1052static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1053 struct btrfs_root *root,
1054 struct btrfs_path *path)
1055{
1056 int ret;
1057 struct btrfs_key key;
1058 struct inode *inode;
1059
1060 key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1061 key.type = BTRFS_ORPHAN_ITEM_KEY;
1062 key.offset = (u64)-1;
1063 while (1) {
1064 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1065 if (ret < 0)
1066 break;
1067
1068 if (ret == 1) {
1069 if (path->slots[0] == 0)
1070 break;
1071 path->slots[0]--;
1072 }
1073
1074 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1075 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1076 key.type != BTRFS_ORPHAN_ITEM_KEY)
1077 break;
1078
1079 ret = btrfs_del_item(trans, root, path);
1080 if (ret)
1081 goto out;
1082
1083 btrfs_release_path(path);
1084 inode = read_one_inode(root, key.offset);
1085 if (!inode)
1086 return -EIO;
1087
1088 ret = fixup_inode_link_count(trans, root, inode);
1089 BUG_ON(ret);
1090
1091 iput(inode);
1092
1093 /*
1094 * fixup on a directory may create new entries,
1095 * make sure we always look for the highset possible
1096 * offset
1097 */
1098 key.offset = (u64)-1;
1099 }
1100 ret = 0;
1101out:
1102 btrfs_release_path(path);
1103 return ret;
1104}
1105
1106
1107/*
1108 * record a given inode in the fixup dir so we can check its link
1109 * count when replay is done. The link count is incremented here
1110 * so the inode won't go away until we check it
1111 */
1112static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1113 struct btrfs_root *root,
1114 struct btrfs_path *path,
1115 u64 objectid)
1116{
1117 struct btrfs_key key;
1118 int ret = 0;
1119 struct inode *inode;
1120
1121 inode = read_one_inode(root, objectid);
1122 if (!inode)
1123 return -EIO;
1124
1125 key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1126 btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1127 key.offset = objectid;
1128
1129 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1130
1131 btrfs_release_path(path);
1132 if (ret == 0) {
1133 btrfs_inc_nlink(inode);
1134 btrfs_update_inode(trans, root, inode);
1135 } else if (ret == -EEXIST) {
1136 ret = 0;
1137 } else {
1138 BUG();
1139 }
1140 iput(inode);
1141
1142 return ret;
1143}
1144
1145/*
1146 * when replaying the log for a directory, we only insert names
1147 * for inodes that actually exist. This means an fsync on a directory
1148 * does not implicitly fsync all the new files in it
1149 */
1150static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1151 struct btrfs_root *root,
1152 struct btrfs_path *path,
1153 u64 dirid, u64 index,
1154 char *name, int name_len, u8 type,
1155 struct btrfs_key *location)
1156{
1157 struct inode *inode;
1158 struct inode *dir;
1159 int ret;
1160
1161 inode = read_one_inode(root, location->objectid);
1162 if (!inode)
1163 return -ENOENT;
1164
1165 dir = read_one_inode(root, dirid);
1166 if (!dir) {
1167 iput(inode);
1168 return -EIO;
1169 }
1170 ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1171
1172 /* FIXME, put inode into FIXUP list */
1173
1174 iput(inode);
1175 iput(dir);
1176 return ret;
1177}
1178
1179/*
1180 * take a single entry in a log directory item and replay it into
1181 * the subvolume.
1182 *
1183 * if a conflicting item exists in the subdirectory already,
1184 * the inode it points to is unlinked and put into the link count
1185 * fix up tree.
1186 *
1187 * If a name from the log points to a file or directory that does
1188 * not exist in the FS, it is skipped. fsyncs on directories
1189 * do not force down inodes inside that directory, just changes to the
1190 * names or unlinks in a directory.
1191 */
1192static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1193 struct btrfs_root *root,
1194 struct btrfs_path *path,
1195 struct extent_buffer *eb,
1196 struct btrfs_dir_item *di,
1197 struct btrfs_key *key)
1198{
1199 char *name;
1200 int name_len;
1201 struct btrfs_dir_item *dst_di;
1202 struct btrfs_key found_key;
1203 struct btrfs_key log_key;
1204 struct inode *dir;
1205 u8 log_type;
1206 int exists;
1207 int ret;
1208
1209 dir = read_one_inode(root, key->objectid);
1210 if (!dir)
1211 return -EIO;
1212
1213 name_len = btrfs_dir_name_len(eb, di);
1214 name = kmalloc(name_len, GFP_NOFS);
1215 if (!name)
1216 return -ENOMEM;
1217
1218 log_type = btrfs_dir_type(eb, di);
1219 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1220 name_len);
1221
1222 btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1223 exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1224 if (exists == 0)
1225 exists = 1;
1226 else
1227 exists = 0;
1228 btrfs_release_path(path);
1229
1230 if (key->type == BTRFS_DIR_ITEM_KEY) {
1231 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1232 name, name_len, 1);
1233 } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1234 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1235 key->objectid,
1236 key->offset, name,
1237 name_len, 1);
1238 } else {
1239 BUG();
1240 }
1241 if (IS_ERR_OR_NULL(dst_di)) {
1242 /* we need a sequence number to insert, so we only
1243 * do inserts for the BTRFS_DIR_INDEX_KEY types
1244 */
1245 if (key->type != BTRFS_DIR_INDEX_KEY)
1246 goto out;
1247 goto insert;
1248 }
1249
1250 btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1251 /* the existing item matches the logged item */
1252 if (found_key.objectid == log_key.objectid &&
1253 found_key.type == log_key.type &&
1254 found_key.offset == log_key.offset &&
1255 btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1256 goto out;
1257 }
1258
1259 /*
1260 * don't drop the conflicting directory entry if the inode
1261 * for the new entry doesn't exist
1262 */
1263 if (!exists)
1264 goto out;
1265
1266 ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1267 BUG_ON(ret);
1268
1269 if (key->type == BTRFS_DIR_INDEX_KEY)
1270 goto insert;
1271out:
1272 btrfs_release_path(path);
1273 kfree(name);
1274 iput(dir);
1275 return 0;
1276
1277insert:
1278 btrfs_release_path(path);
1279 ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1280 name, name_len, log_type, &log_key);
1281
1282 BUG_ON(ret && ret != -ENOENT);
1283 goto out;
1284}
1285
1286/*
1287 * find all the names in a directory item and reconcile them into
1288 * the subvolume. Only BTRFS_DIR_ITEM_KEY types will have more than
1289 * one name in a directory item, but the same code gets used for
1290 * both directory index types
1291 */
1292static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1293 struct btrfs_root *root,
1294 struct btrfs_path *path,
1295 struct extent_buffer *eb, int slot,
1296 struct btrfs_key *key)
1297{
1298 int ret;
1299 u32 item_size = btrfs_item_size_nr(eb, slot);
1300 struct btrfs_dir_item *di;
1301 int name_len;
1302 unsigned long ptr;
1303 unsigned long ptr_end;
1304
1305 ptr = btrfs_item_ptr_offset(eb, slot);
1306 ptr_end = ptr + item_size;
1307 while (ptr < ptr_end) {
1308 di = (struct btrfs_dir_item *)ptr;
1309 if (verify_dir_item(root, eb, di))
1310 return -EIO;
1311 name_len = btrfs_dir_name_len(eb, di);
1312 ret = replay_one_name(trans, root, path, eb, di, key);
1313 BUG_ON(ret);
1314 ptr = (unsigned long)(di + 1);
1315 ptr += name_len;
1316 }
1317 return 0;
1318}
1319
1320/*
1321 * directory replay has two parts. There are the standard directory
1322 * items in the log copied from the subvolume, and range items
1323 * created in the log while the subvolume was logged.
1324 *
1325 * The range items tell us which parts of the key space the log
1326 * is authoritative for. During replay, if a key in the subvolume
1327 * directory is in a logged range item, but not actually in the log
1328 * that means it was deleted from the directory before the fsync
1329 * and should be removed.
1330 */
1331static noinline int find_dir_range(struct btrfs_root *root,
1332 struct btrfs_path *path,
1333 u64 dirid, int key_type,
1334 u64 *start_ret, u64 *end_ret)
1335{
1336 struct btrfs_key key;
1337 u64 found_end;
1338 struct btrfs_dir_log_item *item;
1339 int ret;
1340 int nritems;
1341
1342 if (*start_ret == (u64)-1)
1343 return 1;
1344
1345 key.objectid = dirid;
1346 key.type = key_type;
1347 key.offset = *start_ret;
1348
1349 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1350 if (ret < 0)
1351 goto out;
1352 if (ret > 0) {
1353 if (path->slots[0] == 0)
1354 goto out;
1355 path->slots[0]--;
1356 }
1357 if (ret != 0)
1358 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1359
1360 if (key.type != key_type || key.objectid != dirid) {
1361 ret = 1;
1362 goto next;
1363 }
1364 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1365 struct btrfs_dir_log_item);
1366 found_end = btrfs_dir_log_end(path->nodes[0], item);
1367
1368 if (*start_ret >= key.offset && *start_ret <= found_end) {
1369 ret = 0;
1370 *start_ret = key.offset;
1371 *end_ret = found_end;
1372 goto out;
1373 }
1374 ret = 1;
1375next:
1376 /* check the next slot in the tree to see if it is a valid item */
1377 nritems = btrfs_header_nritems(path->nodes[0]);
1378 if (path->slots[0] >= nritems) {
1379 ret = btrfs_next_leaf(root, path);
1380 if (ret)
1381 goto out;
1382 } else {
1383 path->slots[0]++;
1384 }
1385
1386 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1387
1388 if (key.type != key_type || key.objectid != dirid) {
1389 ret = 1;
1390 goto out;
1391 }
1392 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1393 struct btrfs_dir_log_item);
1394 found_end = btrfs_dir_log_end(path->nodes[0], item);
1395 *start_ret = key.offset;
1396 *end_ret = found_end;
1397 ret = 0;
1398out:
1399 btrfs_release_path(path);
1400 return ret;
1401}
1402
1403/*
1404 * this looks for a given directory item in the log. If the directory
1405 * item is not in the log, the item is removed and the inode it points
1406 * to is unlinked
1407 */
1408static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1409 struct btrfs_root *root,
1410 struct btrfs_root *log,
1411 struct btrfs_path *path,
1412 struct btrfs_path *log_path,
1413 struct inode *dir,
1414 struct btrfs_key *dir_key)
1415{
1416 int ret;
1417 struct extent_buffer *eb;
1418 int slot;
1419 u32 item_size;
1420 struct btrfs_dir_item *di;
1421 struct btrfs_dir_item *log_di;
1422 int name_len;
1423 unsigned long ptr;
1424 unsigned long ptr_end;
1425 char *name;
1426 struct inode *inode;
1427 struct btrfs_key location;
1428
1429again:
1430 eb = path->nodes[0];
1431 slot = path->slots[0];
1432 item_size = btrfs_item_size_nr(eb, slot);
1433 ptr = btrfs_item_ptr_offset(eb, slot);
1434 ptr_end = ptr + item_size;
1435 while (ptr < ptr_end) {
1436 di = (struct btrfs_dir_item *)ptr;
1437 if (verify_dir_item(root, eb, di)) {
1438 ret = -EIO;
1439 goto out;
1440 }
1441
1442 name_len = btrfs_dir_name_len(eb, di);
1443 name = kmalloc(name_len, GFP_NOFS);
1444 if (!name) {
1445 ret = -ENOMEM;
1446 goto out;
1447 }
1448 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1449 name_len);
1450 log_di = NULL;
1451 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
1452 log_di = btrfs_lookup_dir_item(trans, log, log_path,
1453 dir_key->objectid,
1454 name, name_len, 0);
1455 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
1456 log_di = btrfs_lookup_dir_index_item(trans, log,
1457 log_path,
1458 dir_key->objectid,
1459 dir_key->offset,
1460 name, name_len, 0);
1461 }
1462 if (IS_ERR_OR_NULL(log_di)) {
1463 btrfs_dir_item_key_to_cpu(eb, di, &location);
1464 btrfs_release_path(path);
1465 btrfs_release_path(log_path);
1466 inode = read_one_inode(root, location.objectid);
1467 if (!inode) {
1468 kfree(name);
1469 return -EIO;
1470 }
1471
1472 ret = link_to_fixup_dir(trans, root,
1473 path, location.objectid);
1474 BUG_ON(ret);
1475 btrfs_inc_nlink(inode);
1476 ret = btrfs_unlink_inode(trans, root, dir, inode,
1477 name, name_len);
1478 BUG_ON(ret);
1479 kfree(name);
1480 iput(inode);
1481
1482 /* there might still be more names under this key
1483 * check and repeat if required
1484 */
1485 ret = btrfs_search_slot(NULL, root, dir_key, path,
1486 0, 0);
1487 if (ret == 0)
1488 goto again;
1489 ret = 0;
1490 goto out;
1491 }
1492 btrfs_release_path(log_path);
1493 kfree(name);
1494
1495 ptr = (unsigned long)(di + 1);
1496 ptr += name_len;
1497 }
1498 ret = 0;
1499out:
1500 btrfs_release_path(path);
1501 btrfs_release_path(log_path);
1502 return ret;
1503}
1504
1505/*
1506 * deletion replay happens before we copy any new directory items
1507 * out of the log or out of backreferences from inodes. It
1508 * scans the log to find ranges of keys that log is authoritative for,
1509 * and then scans the directory to find items in those ranges that are
1510 * not present in the log.
1511 *
1512 * Anything we don't find in the log is unlinked and removed from the
1513 * directory.
1514 */
1515static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1516 struct btrfs_root *root,
1517 struct btrfs_root *log,
1518 struct btrfs_path *path,
1519 u64 dirid, int del_all)
1520{
1521 u64 range_start;
1522 u64 range_end;
1523 int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1524 int ret = 0;
1525 struct btrfs_key dir_key;
1526 struct btrfs_key found_key;
1527 struct btrfs_path *log_path;
1528 struct inode *dir;
1529
1530 dir_key.objectid = dirid;
1531 dir_key.type = BTRFS_DIR_ITEM_KEY;
1532 log_path = btrfs_alloc_path();
1533 if (!log_path)
1534 return -ENOMEM;
1535
1536 dir = read_one_inode(root, dirid);
1537 /* it isn't an error if the inode isn't there, that can happen
1538 * because we replay the deletes before we copy in the inode item
1539 * from the log
1540 */
1541 if (!dir) {
1542 btrfs_free_path(log_path);
1543 return 0;
1544 }
1545again:
1546 range_start = 0;
1547 range_end = 0;
1548 while (1) {
1549 if (del_all)
1550 range_end = (u64)-1;
1551 else {
1552 ret = find_dir_range(log, path, dirid, key_type,
1553 &range_start, &range_end);
1554 if (ret != 0)
1555 break;
1556 }
1557
1558 dir_key.offset = range_start;
1559 while (1) {
1560 int nritems;
1561 ret = btrfs_search_slot(NULL, root, &dir_key, path,
1562 0, 0);
1563 if (ret < 0)
1564 goto out;
1565
1566 nritems = btrfs_header_nritems(path->nodes[0]);
1567 if (path->slots[0] >= nritems) {
1568 ret = btrfs_next_leaf(root, path);
1569 if (ret)
1570 break;
1571 }
1572 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1573 path->slots[0]);
1574 if (found_key.objectid != dirid ||
1575 found_key.type != dir_key.type)
1576 goto next_type;
1577
1578 if (found_key.offset > range_end)
1579 break;
1580
1581 ret = check_item_in_log(trans, root, log, path,
1582 log_path, dir,
1583 &found_key);
1584 BUG_ON(ret);
1585 if (found_key.offset == (u64)-1)
1586 break;
1587 dir_key.offset = found_key.offset + 1;
1588 }
1589 btrfs_release_path(path);
1590 if (range_end == (u64)-1)
1591 break;
1592 range_start = range_end + 1;
1593 }
1594
1595next_type:
1596 ret = 0;
1597 if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1598 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1599 dir_key.type = BTRFS_DIR_INDEX_KEY;
1600 btrfs_release_path(path);
1601 goto again;
1602 }
1603out:
1604 btrfs_release_path(path);
1605 btrfs_free_path(log_path);
1606 iput(dir);
1607 return ret;
1608}
1609
1610/*
1611 * the process_func used to replay items from the log tree. This
1612 * gets called in two different stages. The first stage just looks
1613 * for inodes and makes sure they are all copied into the subvolume.
1614 *
1615 * The second stage copies all the other item types from the log into
1616 * the subvolume. The two stage approach is slower, but gets rid of
1617 * lots of complexity around inodes referencing other inodes that exist
1618 * only in the log (references come from either directory items or inode
1619 * back refs).
1620 */
1621static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
1622 struct walk_control *wc, u64 gen)
1623{
1624 int nritems;
1625 struct btrfs_path *path;
1626 struct btrfs_root *root = wc->replay_dest;
1627 struct btrfs_key key;
1628 int level;
1629 int i;
1630 int ret;
1631
1632 btrfs_read_buffer(eb, gen);
1633
1634 level = btrfs_header_level(eb);
1635
1636 if (level != 0)
1637 return 0;
1638
1639 path = btrfs_alloc_path();
1640 if (!path)
1641 return -ENOMEM;
1642
1643 nritems = btrfs_header_nritems(eb);
1644 for (i = 0; i < nritems; i++) {
1645 btrfs_item_key_to_cpu(eb, &key, i);
1646
1647 /* inode keys are done during the first stage */
1648 if (key.type == BTRFS_INODE_ITEM_KEY &&
1649 wc->stage == LOG_WALK_REPLAY_INODES) {
1650 struct btrfs_inode_item *inode_item;
1651 u32 mode;
1652
1653 inode_item = btrfs_item_ptr(eb, i,
1654 struct btrfs_inode_item);
1655 mode = btrfs_inode_mode(eb, inode_item);
1656 if (S_ISDIR(mode)) {
1657 ret = replay_dir_deletes(wc->trans,
1658 root, log, path, key.objectid, 0);
1659 BUG_ON(ret);
1660 }
1661 ret = overwrite_item(wc->trans, root, path,
1662 eb, i, &key);
1663 BUG_ON(ret);
1664
1665 /* for regular files, make sure corresponding
1666 * orhpan item exist. extents past the new EOF
1667 * will be truncated later by orphan cleanup.
1668 */
1669 if (S_ISREG(mode)) {
1670 ret = insert_orphan_item(wc->trans, root,
1671 key.objectid);
1672 BUG_ON(ret);
1673 }
1674
1675 ret = link_to_fixup_dir(wc->trans, root,
1676 path, key.objectid);
1677 BUG_ON(ret);
1678 }
1679 if (wc->stage < LOG_WALK_REPLAY_ALL)
1680 continue;
1681
1682 /* these keys are simply copied */
1683 if (key.type == BTRFS_XATTR_ITEM_KEY) {
1684 ret = overwrite_item(wc->trans, root, path,
1685 eb, i, &key);
1686 BUG_ON(ret);
1687 } else if (key.type == BTRFS_INODE_REF_KEY) {
1688 ret = add_inode_ref(wc->trans, root, log, path,
1689 eb, i, &key);
1690 BUG_ON(ret && ret != -ENOENT);
1691 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
1692 ret = replay_one_extent(wc->trans, root, path,
1693 eb, i, &key);
1694 BUG_ON(ret);
1695 } else if (key.type == BTRFS_DIR_ITEM_KEY ||
1696 key.type == BTRFS_DIR_INDEX_KEY) {
1697 ret = replay_one_dir_item(wc->trans, root, path,
1698 eb, i, &key);
1699 BUG_ON(ret);
1700 }
1701 }
1702 btrfs_free_path(path);
1703 return 0;
1704}
1705
1706static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
1707 struct btrfs_root *root,
1708 struct btrfs_path *path, int *level,
1709 struct walk_control *wc)
1710{
1711 u64 root_owner;
1712 u64 bytenr;
1713 u64 ptr_gen;
1714 struct extent_buffer *next;
1715 struct extent_buffer *cur;
1716 struct extent_buffer *parent;
1717 u32 blocksize;
1718 int ret = 0;
1719
1720 WARN_ON(*level < 0);
1721 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1722
1723 while (*level > 0) {
1724 WARN_ON(*level < 0);
1725 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1726 cur = path->nodes[*level];
1727
1728 if (btrfs_header_level(cur) != *level)
1729 WARN_ON(1);
1730
1731 if (path->slots[*level] >=
1732 btrfs_header_nritems(cur))
1733 break;
1734
1735 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1736 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1737 blocksize = btrfs_level_size(root, *level - 1);
1738
1739 parent = path->nodes[*level];
1740 root_owner = btrfs_header_owner(parent);
1741
1742 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
1743 if (!next)
1744 return -ENOMEM;
1745
1746 if (*level == 1) {
1747 ret = wc->process_func(root, next, wc, ptr_gen);
1748 if (ret)
1749 return ret;
1750
1751 path->slots[*level]++;
1752 if (wc->free) {
1753 btrfs_read_buffer(next, ptr_gen);
1754
1755 btrfs_tree_lock(next);
1756 btrfs_set_lock_blocking(next);
1757 clean_tree_block(trans, root, next);
1758 btrfs_wait_tree_block_writeback(next);
1759 btrfs_tree_unlock(next);
1760
1761 WARN_ON(root_owner !=
1762 BTRFS_TREE_LOG_OBJECTID);
1763 ret = btrfs_free_reserved_extent(root,
1764 bytenr, blocksize);
1765 BUG_ON(ret);
1766 }
1767 free_extent_buffer(next);
1768 continue;
1769 }
1770 btrfs_read_buffer(next, ptr_gen);
1771
1772 WARN_ON(*level <= 0);
1773 if (path->nodes[*level-1])
1774 free_extent_buffer(path->nodes[*level-1]);
1775 path->nodes[*level-1] = next;
1776 *level = btrfs_header_level(next);
1777 path->slots[*level] = 0;
1778 cond_resched();
1779 }
1780 WARN_ON(*level < 0);
1781 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1782
1783 path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
1784
1785 cond_resched();
1786 return 0;
1787}
1788
1789static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
1790 struct btrfs_root *root,
1791 struct btrfs_path *path, int *level,
1792 struct walk_control *wc)
1793{
1794 u64 root_owner;
1795 int i;
1796 int slot;
1797 int ret;
1798
1799 for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1800 slot = path->slots[i];
1801 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
1802 path->slots[i]++;
1803 *level = i;
1804 WARN_ON(*level == 0);
1805 return 0;
1806 } else {
1807 struct extent_buffer *parent;
1808 if (path->nodes[*level] == root->node)
1809 parent = path->nodes[*level];
1810 else
1811 parent = path->nodes[*level + 1];
1812
1813 root_owner = btrfs_header_owner(parent);
1814 ret = wc->process_func(root, path->nodes[*level], wc,
1815 btrfs_header_generation(path->nodes[*level]));
1816 if (ret)
1817 return ret;
1818
1819 if (wc->free) {
1820 struct extent_buffer *next;
1821
1822 next = path->nodes[*level];
1823
1824 btrfs_tree_lock(next);
1825 btrfs_set_lock_blocking(next);
1826 clean_tree_block(trans, root, next);
1827 btrfs_wait_tree_block_writeback(next);
1828 btrfs_tree_unlock(next);
1829
1830 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
1831 ret = btrfs_free_reserved_extent(root,
1832 path->nodes[*level]->start,
1833 path->nodes[*level]->len);
1834 BUG_ON(ret);
1835 }
1836 free_extent_buffer(path->nodes[*level]);
1837 path->nodes[*level] = NULL;
1838 *level = i + 1;
1839 }
1840 }
1841 return 1;
1842}
1843
1844/*
1845 * drop the reference count on the tree rooted at 'snap'. This traverses
1846 * the tree freeing any blocks that have a ref count of zero after being
1847 * decremented.
1848 */
1849static int walk_log_tree(struct btrfs_trans_handle *trans,
1850 struct btrfs_root *log, struct walk_control *wc)
1851{
1852 int ret = 0;
1853 int wret;
1854 int level;
1855 struct btrfs_path *path;
1856 int i;
1857 int orig_level;
1858
1859 path = btrfs_alloc_path();
1860 if (!path)
1861 return -ENOMEM;
1862
1863 level = btrfs_header_level(log->node);
1864 orig_level = level;
1865 path->nodes[level] = log->node;
1866 extent_buffer_get(log->node);
1867 path->slots[level] = 0;
1868
1869 while (1) {
1870 wret = walk_down_log_tree(trans, log, path, &level, wc);
1871 if (wret > 0)
1872 break;
1873 if (wret < 0)
1874 ret = wret;
1875
1876 wret = walk_up_log_tree(trans, log, path, &level, wc);
1877 if (wret > 0)
1878 break;
1879 if (wret < 0)
1880 ret = wret;
1881 }
1882
1883 /* was the root node processed? if not, catch it here */
1884 if (path->nodes[orig_level]) {
1885 wc->process_func(log, path->nodes[orig_level], wc,
1886 btrfs_header_generation(path->nodes[orig_level]));
1887 if (wc->free) {
1888 struct extent_buffer *next;
1889
1890 next = path->nodes[orig_level];
1891
1892 btrfs_tree_lock(next);
1893 btrfs_set_lock_blocking(next);
1894 clean_tree_block(trans, log, next);
1895 btrfs_wait_tree_block_writeback(next);
1896 btrfs_tree_unlock(next);
1897
1898 WARN_ON(log->root_key.objectid !=
1899 BTRFS_TREE_LOG_OBJECTID);
1900 ret = btrfs_free_reserved_extent(log, next->start,
1901 next->len);
1902 BUG_ON(ret);
1903 }
1904 }
1905
1906 for (i = 0; i <= orig_level; i++) {
1907 if (path->nodes[i]) {
1908 free_extent_buffer(path->nodes[i]);
1909 path->nodes[i] = NULL;
1910 }
1911 }
1912 btrfs_free_path(path);
1913 return ret;
1914}
1915
1916/*
1917 * helper function to update the item for a given subvolumes log root
1918 * in the tree of log roots
1919 */
1920static int update_log_root(struct btrfs_trans_handle *trans,
1921 struct btrfs_root *log)
1922{
1923 int ret;
1924
1925 if (log->log_transid == 1) {
1926 /* insert root item on the first sync */
1927 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
1928 &log->root_key, &log->root_item);
1929 } else {
1930 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
1931 &log->root_key, &log->root_item);
1932 }
1933 return ret;
1934}
1935
1936static int wait_log_commit(struct btrfs_trans_handle *trans,
1937 struct btrfs_root *root, unsigned long transid)
1938{
1939 DEFINE_WAIT(wait);
1940 int index = transid % 2;
1941
1942 /*
1943 * we only allow two pending log transactions at a time,
1944 * so we know that if ours is more than 2 older than the
1945 * current transaction, we're done
1946 */
1947 do {
1948 prepare_to_wait(&root->log_commit_wait[index],
1949 &wait, TASK_UNINTERRUPTIBLE);
1950 mutex_unlock(&root->log_mutex);
1951
1952 if (root->fs_info->last_trans_log_full_commit !=
1953 trans->transid && root->log_transid < transid + 2 &&
1954 atomic_read(&root->log_commit[index]))
1955 schedule();
1956
1957 finish_wait(&root->log_commit_wait[index], &wait);
1958 mutex_lock(&root->log_mutex);
1959 } while (root->log_transid < transid + 2 &&
1960 atomic_read(&root->log_commit[index]));
1961 return 0;
1962}
1963
1964static int wait_for_writer(struct btrfs_trans_handle *trans,
1965 struct btrfs_root *root)
1966{
1967 DEFINE_WAIT(wait);
1968 while (atomic_read(&root->log_writers)) {
1969 prepare_to_wait(&root->log_writer_wait,
1970 &wait, TASK_UNINTERRUPTIBLE);
1971 mutex_unlock(&root->log_mutex);
1972 if (root->fs_info->last_trans_log_full_commit !=
1973 trans->transid && atomic_read(&root->log_writers))
1974 schedule();
1975 mutex_lock(&root->log_mutex);
1976 finish_wait(&root->log_writer_wait, &wait);
1977 }
1978 return 0;
1979}
1980
1981/*
1982 * btrfs_sync_log does sends a given tree log down to the disk and
1983 * updates the super blocks to record it. When this call is done,
1984 * you know that any inodes previously logged are safely on disk only
1985 * if it returns 0.
1986 *
1987 * Any other return value means you need to call btrfs_commit_transaction.
1988 * Some of the edge cases for fsyncing directories that have had unlinks
1989 * or renames done in the past mean that sometimes the only safe
1990 * fsync is to commit the whole FS. When btrfs_sync_log returns -EAGAIN,
1991 * that has happened.
1992 */
1993int btrfs_sync_log(struct btrfs_trans_handle *trans,
1994 struct btrfs_root *root)
1995{
1996 int index1;
1997 int index2;
1998 int mark;
1999 int ret;
2000 struct btrfs_root *log = root->log_root;
2001 struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
2002 unsigned long log_transid = 0;
2003
2004 mutex_lock(&root->log_mutex);
2005 index1 = root->log_transid % 2;
2006 if (atomic_read(&root->log_commit[index1])) {
2007 wait_log_commit(trans, root, root->log_transid);
2008 mutex_unlock(&root->log_mutex);
2009 return 0;
2010 }
2011 atomic_set(&root->log_commit[index1], 1);
2012
2013 /* wait for previous tree log sync to complete */
2014 if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2015 wait_log_commit(trans, root, root->log_transid - 1);
2016
2017 while (1) {
2018 unsigned long batch = root->log_batch;
2019 if (root->log_multiple_pids) {
2020 mutex_unlock(&root->log_mutex);
2021 schedule_timeout_uninterruptible(1);
2022 mutex_lock(&root->log_mutex);
2023 }
2024 wait_for_writer(trans, root);
2025 if (batch == root->log_batch)
2026 break;
2027 }
2028
2029 /* bail out if we need to do a full commit */
2030 if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2031 ret = -EAGAIN;
2032 mutex_unlock(&root->log_mutex);
2033 goto out;
2034 }
2035
2036 log_transid = root->log_transid;
2037 if (log_transid % 2 == 0)
2038 mark = EXTENT_DIRTY;
2039 else
2040 mark = EXTENT_NEW;
2041
2042 /* we start IO on all the marked extents here, but we don't actually
2043 * wait for them until later.
2044 */
2045 ret = btrfs_write_marked_extents(log, &log->dirty_log_pages, mark);
2046 BUG_ON(ret);
2047
2048 btrfs_set_root_node(&log->root_item, log->node);
2049
2050 root->log_batch = 0;
2051 root->log_transid++;
2052 log->log_transid = root->log_transid;
2053 root->log_start_pid = 0;
2054 smp_mb();
2055 /*
2056 * IO has been started, blocks of the log tree have WRITTEN flag set
2057 * in their headers. new modifications of the log will be written to
2058 * new positions. so it's safe to allow log writers to go in.
2059 */
2060 mutex_unlock(&root->log_mutex);
2061
2062 mutex_lock(&log_root_tree->log_mutex);
2063 log_root_tree->log_batch++;
2064 atomic_inc(&log_root_tree->log_writers);
2065 mutex_unlock(&log_root_tree->log_mutex);
2066
2067 ret = update_log_root(trans, log);
2068
2069 mutex_lock(&log_root_tree->log_mutex);
2070 if (atomic_dec_and_test(&log_root_tree->log_writers)) {
2071 smp_mb();
2072 if (waitqueue_active(&log_root_tree->log_writer_wait))
2073 wake_up(&log_root_tree->log_writer_wait);
2074 }
2075
2076 if (ret) {
2077 BUG_ON(ret != -ENOSPC);
2078 root->fs_info->last_trans_log_full_commit = trans->transid;
2079 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2080 mutex_unlock(&log_root_tree->log_mutex);
2081 ret = -EAGAIN;
2082 goto out;
2083 }
2084
2085 index2 = log_root_tree->log_transid % 2;
2086 if (atomic_read(&log_root_tree->log_commit[index2])) {
2087 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2088 wait_log_commit(trans, log_root_tree,
2089 log_root_tree->log_transid);
2090 mutex_unlock(&log_root_tree->log_mutex);
2091 ret = 0;
2092 goto out;
2093 }
2094 atomic_set(&log_root_tree->log_commit[index2], 1);
2095
2096 if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2097 wait_log_commit(trans, log_root_tree,
2098 log_root_tree->log_transid - 1);
2099 }
2100
2101 wait_for_writer(trans, log_root_tree);
2102
2103 /*
2104 * now that we've moved on to the tree of log tree roots,
2105 * check the full commit flag again
2106 */
2107 if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2108 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2109 mutex_unlock(&log_root_tree->log_mutex);
2110 ret = -EAGAIN;
2111 goto out_wake_log_root;
2112 }
2113
2114 ret = btrfs_write_and_wait_marked_extents(log_root_tree,
2115 &log_root_tree->dirty_log_pages,
2116 EXTENT_DIRTY | EXTENT_NEW);
2117 BUG_ON(ret);
2118 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2119
2120 btrfs_set_super_log_root(&root->fs_info->super_for_commit,
2121 log_root_tree->node->start);
2122 btrfs_set_super_log_root_level(&root->fs_info->super_for_commit,
2123 btrfs_header_level(log_root_tree->node));
2124
2125 log_root_tree->log_batch = 0;
2126 log_root_tree->log_transid++;
2127 smp_mb();
2128
2129 mutex_unlock(&log_root_tree->log_mutex);
2130
2131 /*
2132 * nobody else is going to jump in and write the the ctree
2133 * super here because the log_commit atomic below is protecting
2134 * us. We must be called with a transaction handle pinning
2135 * the running transaction open, so a full commit can't hop
2136 * in and cause problems either.
2137 */
2138 btrfs_scrub_pause_super(root);
2139 write_ctree_super(trans, root->fs_info->tree_root, 1);
2140 btrfs_scrub_continue_super(root);
2141 ret = 0;
2142
2143 mutex_lock(&root->log_mutex);
2144 if (root->last_log_commit < log_transid)
2145 root->last_log_commit = log_transid;
2146 mutex_unlock(&root->log_mutex);
2147
2148out_wake_log_root:
2149 atomic_set(&log_root_tree->log_commit[index2], 0);
2150 smp_mb();
2151 if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2152 wake_up(&log_root_tree->log_commit_wait[index2]);
2153out:
2154 atomic_set(&root->log_commit[index1], 0);
2155 smp_mb();
2156 if (waitqueue_active(&root->log_commit_wait[index1]))
2157 wake_up(&root->log_commit_wait[index1]);
2158 return ret;
2159}
2160
2161static void free_log_tree(struct btrfs_trans_handle *trans,
2162 struct btrfs_root *log)
2163{
2164 int ret;
2165 u64 start;
2166 u64 end;
2167 struct walk_control wc = {
2168 .free = 1,
2169 .process_func = process_one_buffer
2170 };
2171
2172 ret = walk_log_tree(trans, log, &wc);
2173 BUG_ON(ret);
2174
2175 while (1) {
2176 ret = find_first_extent_bit(&log->dirty_log_pages,
2177 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW);
2178 if (ret)
2179 break;
2180
2181 clear_extent_bits(&log->dirty_log_pages, start, end,
2182 EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2183 }
2184
2185 free_extent_buffer(log->node);
2186 kfree(log);
2187}
2188
2189/*
2190 * free all the extents used by the tree log. This should be called
2191 * at commit time of the full transaction
2192 */
2193int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2194{
2195 if (root->log_root) {
2196 free_log_tree(trans, root->log_root);
2197 root->log_root = NULL;
2198 }
2199 return 0;
2200}
2201
2202int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2203 struct btrfs_fs_info *fs_info)
2204{
2205 if (fs_info->log_root_tree) {
2206 free_log_tree(trans, fs_info->log_root_tree);
2207 fs_info->log_root_tree = NULL;
2208 }
2209 return 0;
2210}
2211
2212/*
2213 * If both a file and directory are logged, and unlinks or renames are
2214 * mixed in, we have a few interesting corners:
2215 *
2216 * create file X in dir Y
2217 * link file X to X.link in dir Y
2218 * fsync file X
2219 * unlink file X but leave X.link
2220 * fsync dir Y
2221 *
2222 * After a crash we would expect only X.link to exist. But file X
2223 * didn't get fsync'd again so the log has back refs for X and X.link.
2224 *
2225 * We solve this by removing directory entries and inode backrefs from the
2226 * log when a file that was logged in the current transaction is
2227 * unlinked. Any later fsync will include the updated log entries, and
2228 * we'll be able to reconstruct the proper directory items from backrefs.
2229 *
2230 * This optimizations allows us to avoid relogging the entire inode
2231 * or the entire directory.
2232 */
2233int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2234 struct btrfs_root *root,
2235 const char *name, int name_len,
2236 struct inode *dir, u64 index)
2237{
2238 struct btrfs_root *log;
2239 struct btrfs_dir_item *di;
2240 struct btrfs_path *path;
2241 int ret;
2242 int err = 0;
2243 int bytes_del = 0;
2244 u64 dir_ino = btrfs_ino(dir);
2245
2246 if (BTRFS_I(dir)->logged_trans < trans->transid)
2247 return 0;
2248
2249 ret = join_running_log_trans(root);
2250 if (ret)
2251 return 0;
2252
2253 mutex_lock(&BTRFS_I(dir)->log_mutex);
2254
2255 log = root->log_root;
2256 path = btrfs_alloc_path();
2257 if (!path) {
2258 err = -ENOMEM;
2259 goto out_unlock;
2260 }
2261
2262 di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2263 name, name_len, -1);
2264 if (IS_ERR(di)) {
2265 err = PTR_ERR(di);
2266 goto fail;
2267 }
2268 if (di) {
2269 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2270 bytes_del += name_len;
2271 BUG_ON(ret);
2272 }
2273 btrfs_release_path(path);
2274 di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2275 index, name, name_len, -1);
2276 if (IS_ERR(di)) {
2277 err = PTR_ERR(di);
2278 goto fail;
2279 }
2280 if (di) {
2281 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2282 bytes_del += name_len;
2283 BUG_ON(ret);
2284 }
2285
2286 /* update the directory size in the log to reflect the names
2287 * we have removed
2288 */
2289 if (bytes_del) {
2290 struct btrfs_key key;
2291
2292 key.objectid = dir_ino;
2293 key.offset = 0;
2294 key.type = BTRFS_INODE_ITEM_KEY;
2295 btrfs_release_path(path);
2296
2297 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2298 if (ret < 0) {
2299 err = ret;
2300 goto fail;
2301 }
2302 if (ret == 0) {
2303 struct btrfs_inode_item *item;
2304 u64 i_size;
2305
2306 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2307 struct btrfs_inode_item);
2308 i_size = btrfs_inode_size(path->nodes[0], item);
2309 if (i_size > bytes_del)
2310 i_size -= bytes_del;
2311 else
2312 i_size = 0;
2313 btrfs_set_inode_size(path->nodes[0], item, i_size);
2314 btrfs_mark_buffer_dirty(path->nodes[0]);
2315 } else
2316 ret = 0;
2317 btrfs_release_path(path);
2318 }
2319fail:
2320 btrfs_free_path(path);
2321out_unlock:
2322 mutex_unlock(&BTRFS_I(dir)->log_mutex);
2323 if (ret == -ENOSPC) {
2324 root->fs_info->last_trans_log_full_commit = trans->transid;
2325 ret = 0;
2326 }
2327 btrfs_end_log_trans(root);
2328
2329 return err;
2330}
2331
2332/* see comments for btrfs_del_dir_entries_in_log */
2333int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2334 struct btrfs_root *root,
2335 const char *name, int name_len,
2336 struct inode *inode, u64 dirid)
2337{
2338 struct btrfs_root *log;
2339 u64 index;
2340 int ret;
2341
2342 if (BTRFS_I(inode)->logged_trans < trans->transid)
2343 return 0;
2344
2345 ret = join_running_log_trans(root);
2346 if (ret)
2347 return 0;
2348 log = root->log_root;
2349 mutex_lock(&BTRFS_I(inode)->log_mutex);
2350
2351 ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
2352 dirid, &index);
2353 mutex_unlock(&BTRFS_I(inode)->log_mutex);
2354 if (ret == -ENOSPC) {
2355 root->fs_info->last_trans_log_full_commit = trans->transid;
2356 ret = 0;
2357 }
2358 btrfs_end_log_trans(root);
2359
2360 return ret;
2361}
2362
2363/*
2364 * creates a range item in the log for 'dirid'. first_offset and
2365 * last_offset tell us which parts of the key space the log should
2366 * be considered authoritative for.
2367 */
2368static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2369 struct btrfs_root *log,
2370 struct btrfs_path *path,
2371 int key_type, u64 dirid,
2372 u64 first_offset, u64 last_offset)
2373{
2374 int ret;
2375 struct btrfs_key key;
2376 struct btrfs_dir_log_item *item;
2377
2378 key.objectid = dirid;
2379 key.offset = first_offset;
2380 if (key_type == BTRFS_DIR_ITEM_KEY)
2381 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2382 else
2383 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2384 ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2385 if (ret)
2386 return ret;
2387
2388 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2389 struct btrfs_dir_log_item);
2390 btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2391 btrfs_mark_buffer_dirty(path->nodes[0]);
2392 btrfs_release_path(path);
2393 return 0;
2394}
2395
2396/*
2397 * log all the items included in the current transaction for a given
2398 * directory. This also creates the range items in the log tree required
2399 * to replay anything deleted before the fsync
2400 */
2401static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2402 struct btrfs_root *root, struct inode *inode,
2403 struct btrfs_path *path,
2404 struct btrfs_path *dst_path, int key_type,
2405 u64 min_offset, u64 *last_offset_ret)
2406{
2407 struct btrfs_key min_key;
2408 struct btrfs_key max_key;
2409 struct btrfs_root *log = root->log_root;
2410 struct extent_buffer *src;
2411 int err = 0;
2412 int ret;
2413 int i;
2414 int nritems;
2415 u64 first_offset = min_offset;
2416 u64 last_offset = (u64)-1;
2417 u64 ino = btrfs_ino(inode);
2418
2419 log = root->log_root;
2420 max_key.objectid = ino;
2421 max_key.offset = (u64)-1;
2422 max_key.type = key_type;
2423
2424 min_key.objectid = ino;
2425 min_key.type = key_type;
2426 min_key.offset = min_offset;
2427
2428 path->keep_locks = 1;
2429
2430 ret = btrfs_search_forward(root, &min_key, &max_key,
2431 path, 0, trans->transid);
2432
2433 /*
2434 * we didn't find anything from this transaction, see if there
2435 * is anything at all
2436 */
2437 if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
2438 min_key.objectid = ino;
2439 min_key.type = key_type;
2440 min_key.offset = (u64)-1;
2441 btrfs_release_path(path);
2442 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2443 if (ret < 0) {
2444 btrfs_release_path(path);
2445 return ret;
2446 }
2447 ret = btrfs_previous_item(root, path, ino, key_type);
2448
2449 /* if ret == 0 there are items for this type,
2450 * create a range to tell us the last key of this type.
2451 * otherwise, there are no items in this directory after
2452 * *min_offset, and we create a range to indicate that.
2453 */
2454 if (ret == 0) {
2455 struct btrfs_key tmp;
2456 btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2457 path->slots[0]);
2458 if (key_type == tmp.type)
2459 first_offset = max(min_offset, tmp.offset) + 1;
2460 }
2461 goto done;
2462 }
2463
2464 /* go backward to find any previous key */
2465 ret = btrfs_previous_item(root, path, ino, key_type);
2466 if (ret == 0) {
2467 struct btrfs_key tmp;
2468 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2469 if (key_type == tmp.type) {
2470 first_offset = tmp.offset;
2471 ret = overwrite_item(trans, log, dst_path,
2472 path->nodes[0], path->slots[0],
2473 &tmp);
2474 if (ret) {
2475 err = ret;
2476 goto done;
2477 }
2478 }
2479 }
2480 btrfs_release_path(path);
2481
2482 /* find the first key from this transaction again */
2483 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2484 if (ret != 0) {
2485 WARN_ON(1);
2486 goto done;
2487 }
2488
2489 /*
2490 * we have a block from this transaction, log every item in it
2491 * from our directory
2492 */
2493 while (1) {
2494 struct btrfs_key tmp;
2495 src = path->nodes[0];
2496 nritems = btrfs_header_nritems(src);
2497 for (i = path->slots[0]; i < nritems; i++) {
2498 btrfs_item_key_to_cpu(src, &min_key, i);
2499
2500 if (min_key.objectid != ino || min_key.type != key_type)
2501 goto done;
2502 ret = overwrite_item(trans, log, dst_path, src, i,
2503 &min_key);
2504 if (ret) {
2505 err = ret;
2506 goto done;
2507 }
2508 }
2509 path->slots[0] = nritems;
2510
2511 /*
2512 * look ahead to the next item and see if it is also
2513 * from this directory and from this transaction
2514 */
2515 ret = btrfs_next_leaf(root, path);
2516 if (ret == 1) {
2517 last_offset = (u64)-1;
2518 goto done;
2519 }
2520 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2521 if (tmp.objectid != ino || tmp.type != key_type) {
2522 last_offset = (u64)-1;
2523 goto done;
2524 }
2525 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2526 ret = overwrite_item(trans, log, dst_path,
2527 path->nodes[0], path->slots[0],
2528 &tmp);
2529 if (ret)
2530 err = ret;
2531 else
2532 last_offset = tmp.offset;
2533 goto done;
2534 }
2535 }
2536done:
2537 btrfs_release_path(path);
2538 btrfs_release_path(dst_path);
2539
2540 if (err == 0) {
2541 *last_offset_ret = last_offset;
2542 /*
2543 * insert the log range keys to indicate where the log
2544 * is valid
2545 */
2546 ret = insert_dir_log_key(trans, log, path, key_type,
2547 ino, first_offset, last_offset);
2548 if (ret)
2549 err = ret;
2550 }
2551 return err;
2552}
2553
2554/*
2555 * logging directories is very similar to logging inodes, We find all the items
2556 * from the current transaction and write them to the log.
2557 *
2558 * The recovery code scans the directory in the subvolume, and if it finds a
2559 * key in the range logged that is not present in the log tree, then it means
2560 * that dir entry was unlinked during the transaction.
2561 *
2562 * In order for that scan to work, we must include one key smaller than
2563 * the smallest logged by this transaction and one key larger than the largest
2564 * key logged by this transaction.
2565 */
2566static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
2567 struct btrfs_root *root, struct inode *inode,
2568 struct btrfs_path *path,
2569 struct btrfs_path *dst_path)
2570{
2571 u64 min_key;
2572 u64 max_key;
2573 int ret;
2574 int key_type = BTRFS_DIR_ITEM_KEY;
2575
2576again:
2577 min_key = 0;
2578 max_key = 0;
2579 while (1) {
2580 ret = log_dir_items(trans, root, inode, path,
2581 dst_path, key_type, min_key,
2582 &max_key);
2583 if (ret)
2584 return ret;
2585 if (max_key == (u64)-1)
2586 break;
2587 min_key = max_key + 1;
2588 }
2589
2590 if (key_type == BTRFS_DIR_ITEM_KEY) {
2591 key_type = BTRFS_DIR_INDEX_KEY;
2592 goto again;
2593 }
2594 return 0;
2595}
2596
2597/*
2598 * a helper function to drop items from the log before we relog an
2599 * inode. max_key_type indicates the highest item type to remove.
2600 * This cannot be run for file data extents because it does not
2601 * free the extents they point to.
2602 */
2603static int drop_objectid_items(struct btrfs_trans_handle *trans,
2604 struct btrfs_root *log,
2605 struct btrfs_path *path,
2606 u64 objectid, int max_key_type)
2607{
2608 int ret;
2609 struct btrfs_key key;
2610 struct btrfs_key found_key;
2611
2612 key.objectid = objectid;
2613 key.type = max_key_type;
2614 key.offset = (u64)-1;
2615
2616 while (1) {
2617 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
2618 BUG_ON(ret == 0);
2619 if (ret < 0)
2620 break;
2621
2622 if (path->slots[0] == 0)
2623 break;
2624
2625 path->slots[0]--;
2626 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2627 path->slots[0]);
2628
2629 if (found_key.objectid != objectid)
2630 break;
2631
2632 ret = btrfs_del_item(trans, log, path);
2633 if (ret)
2634 break;
2635 btrfs_release_path(path);
2636 }
2637 btrfs_release_path(path);
2638 return ret;
2639}
2640
2641static noinline int copy_items(struct btrfs_trans_handle *trans,
2642 struct btrfs_root *log,
2643 struct btrfs_path *dst_path,
2644 struct extent_buffer *src,
2645 int start_slot, int nr, int inode_only)
2646{
2647 unsigned long src_offset;
2648 unsigned long dst_offset;
2649 struct btrfs_file_extent_item *extent;
2650 struct btrfs_inode_item *inode_item;
2651 int ret;
2652 struct btrfs_key *ins_keys;
2653 u32 *ins_sizes;
2654 char *ins_data;
2655 int i;
2656 struct list_head ordered_sums;
2657
2658 INIT_LIST_HEAD(&ordered_sums);
2659
2660 ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
2661 nr * sizeof(u32), GFP_NOFS);
2662 if (!ins_data)
2663 return -ENOMEM;
2664
2665 ins_sizes = (u32 *)ins_data;
2666 ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
2667
2668 for (i = 0; i < nr; i++) {
2669 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
2670 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
2671 }
2672 ret = btrfs_insert_empty_items(trans, log, dst_path,
2673 ins_keys, ins_sizes, nr);
2674 if (ret) {
2675 kfree(ins_data);
2676 return ret;
2677 }
2678
2679 for (i = 0; i < nr; i++, dst_path->slots[0]++) {
2680 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
2681 dst_path->slots[0]);
2682
2683 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
2684
2685 copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
2686 src_offset, ins_sizes[i]);
2687
2688 if (inode_only == LOG_INODE_EXISTS &&
2689 ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
2690 inode_item = btrfs_item_ptr(dst_path->nodes[0],
2691 dst_path->slots[0],
2692 struct btrfs_inode_item);
2693 btrfs_set_inode_size(dst_path->nodes[0], inode_item, 0);
2694
2695 /* set the generation to zero so the recover code
2696 * can tell the difference between an logging
2697 * just to say 'this inode exists' and a logging
2698 * to say 'update this inode with these values'
2699 */
2700 btrfs_set_inode_generation(dst_path->nodes[0],
2701 inode_item, 0);
2702 }
2703 /* take a reference on file data extents so that truncates
2704 * or deletes of this inode don't have to relog the inode
2705 * again
2706 */
2707 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY) {
2708 int found_type;
2709 extent = btrfs_item_ptr(src, start_slot + i,
2710 struct btrfs_file_extent_item);
2711
2712 if (btrfs_file_extent_generation(src, extent) < trans->transid)
2713 continue;
2714
2715 found_type = btrfs_file_extent_type(src, extent);
2716 if (found_type == BTRFS_FILE_EXTENT_REG ||
2717 found_type == BTRFS_FILE_EXTENT_PREALLOC) {
2718 u64 ds, dl, cs, cl;
2719 ds = btrfs_file_extent_disk_bytenr(src,
2720 extent);
2721 /* ds == 0 is a hole */
2722 if (ds == 0)
2723 continue;
2724
2725 dl = btrfs_file_extent_disk_num_bytes(src,
2726 extent);
2727 cs = btrfs_file_extent_offset(src, extent);
2728 cl = btrfs_file_extent_num_bytes(src,
2729 extent);
2730 if (btrfs_file_extent_compression(src,
2731 extent)) {
2732 cs = 0;
2733 cl = dl;
2734 }
2735
2736 ret = btrfs_lookup_csums_range(
2737 log->fs_info->csum_root,
2738 ds + cs, ds + cs + cl - 1,
2739 &ordered_sums, 0);
2740 BUG_ON(ret);
2741 }
2742 }
2743 }
2744
2745 btrfs_mark_buffer_dirty(dst_path->nodes[0]);
2746 btrfs_release_path(dst_path);
2747 kfree(ins_data);
2748
2749 /*
2750 * we have to do this after the loop above to avoid changing the
2751 * log tree while trying to change the log tree.
2752 */
2753 ret = 0;
2754 while (!list_empty(&ordered_sums)) {
2755 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
2756 struct btrfs_ordered_sum,
2757 list);
2758 if (!ret)
2759 ret = btrfs_csum_file_blocks(trans, log, sums);
2760 list_del(&sums->list);
2761 kfree(sums);
2762 }
2763 return ret;
2764}
2765
2766/* log a single inode in the tree log.
2767 * At least one parent directory for this inode must exist in the tree
2768 * or be logged already.
2769 *
2770 * Any items from this inode changed by the current transaction are copied
2771 * to the log tree. An extra reference is taken on any extents in this
2772 * file, allowing us to avoid a whole pile of corner cases around logging
2773 * blocks that have been removed from the tree.
2774 *
2775 * See LOG_INODE_ALL and related defines for a description of what inode_only
2776 * does.
2777 *
2778 * This handles both files and directories.
2779 */
2780static int btrfs_log_inode(struct btrfs_trans_handle *trans,
2781 struct btrfs_root *root, struct inode *inode,
2782 int inode_only)
2783{
2784 struct btrfs_path *path;
2785 struct btrfs_path *dst_path;
2786 struct btrfs_key min_key;
2787 struct btrfs_key max_key;
2788 struct btrfs_root *log = root->log_root;
2789 struct extent_buffer *src = NULL;
2790 int err = 0;
2791 int ret;
2792 int nritems;
2793 int ins_start_slot = 0;
2794 int ins_nr;
2795 u64 ino = btrfs_ino(inode);
2796
2797 log = root->log_root;
2798
2799 path = btrfs_alloc_path();
2800 if (!path)
2801 return -ENOMEM;
2802 dst_path = btrfs_alloc_path();
2803 if (!dst_path) {
2804 btrfs_free_path(path);
2805 return -ENOMEM;
2806 }
2807
2808 min_key.objectid = ino;
2809 min_key.type = BTRFS_INODE_ITEM_KEY;
2810 min_key.offset = 0;
2811
2812 max_key.objectid = ino;
2813
2814 /* today the code can only do partial logging of directories */
2815 if (!S_ISDIR(inode->i_mode))
2816 inode_only = LOG_INODE_ALL;
2817
2818 if (inode_only == LOG_INODE_EXISTS || S_ISDIR(inode->i_mode))
2819 max_key.type = BTRFS_XATTR_ITEM_KEY;
2820 else
2821 max_key.type = (u8)-1;
2822 max_key.offset = (u64)-1;
2823
2824 ret = btrfs_commit_inode_delayed_items(trans, inode);
2825 if (ret) {
2826 btrfs_free_path(path);
2827 btrfs_free_path(dst_path);
2828 return ret;
2829 }
2830
2831 mutex_lock(&BTRFS_I(inode)->log_mutex);
2832
2833 /*
2834 * a brute force approach to making sure we get the most uptodate
2835 * copies of everything.
2836 */
2837 if (S_ISDIR(inode->i_mode)) {
2838 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
2839
2840 if (inode_only == LOG_INODE_EXISTS)
2841 max_key_type = BTRFS_XATTR_ITEM_KEY;
2842 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
2843 } else {
2844 ret = btrfs_truncate_inode_items(trans, log, inode, 0, 0);
2845 }
2846 if (ret) {
2847 err = ret;
2848 goto out_unlock;
2849 }
2850 path->keep_locks = 1;
2851
2852 while (1) {
2853 ins_nr = 0;
2854 ret = btrfs_search_forward(root, &min_key, &max_key,
2855 path, 0, trans->transid);
2856 if (ret != 0)
2857 break;
2858again:
2859 /* note, ins_nr might be > 0 here, cleanup outside the loop */
2860 if (min_key.objectid != ino)
2861 break;
2862 if (min_key.type > max_key.type)
2863 break;
2864
2865 src = path->nodes[0];
2866 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
2867 ins_nr++;
2868 goto next_slot;
2869 } else if (!ins_nr) {
2870 ins_start_slot = path->slots[0];
2871 ins_nr = 1;
2872 goto next_slot;
2873 }
2874
2875 ret = copy_items(trans, log, dst_path, src, ins_start_slot,
2876 ins_nr, inode_only);
2877 if (ret) {
2878 err = ret;
2879 goto out_unlock;
2880 }
2881 ins_nr = 1;
2882 ins_start_slot = path->slots[0];
2883next_slot:
2884
2885 nritems = btrfs_header_nritems(path->nodes[0]);
2886 path->slots[0]++;
2887 if (path->slots[0] < nritems) {
2888 btrfs_item_key_to_cpu(path->nodes[0], &min_key,
2889 path->slots[0]);
2890 goto again;
2891 }
2892 if (ins_nr) {
2893 ret = copy_items(trans, log, dst_path, src,
2894 ins_start_slot,
2895 ins_nr, inode_only);
2896 if (ret) {
2897 err = ret;
2898 goto out_unlock;
2899 }
2900 ins_nr = 0;
2901 }
2902 btrfs_release_path(path);
2903
2904 if (min_key.offset < (u64)-1)
2905 min_key.offset++;
2906 else if (min_key.type < (u8)-1)
2907 min_key.type++;
2908 else if (min_key.objectid < (u64)-1)
2909 min_key.objectid++;
2910 else
2911 break;
2912 }
2913 if (ins_nr) {
2914 ret = copy_items(trans, log, dst_path, src,
2915 ins_start_slot,
2916 ins_nr, inode_only);
2917 if (ret) {
2918 err = ret;
2919 goto out_unlock;
2920 }
2921 ins_nr = 0;
2922 }
2923 WARN_ON(ins_nr);
2924 if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
2925 btrfs_release_path(path);
2926 btrfs_release_path(dst_path);
2927 ret = log_directory_changes(trans, root, inode, path, dst_path);
2928 if (ret) {
2929 err = ret;
2930 goto out_unlock;
2931 }
2932 }
2933 BTRFS_I(inode)->logged_trans = trans->transid;
2934out_unlock:
2935 mutex_unlock(&BTRFS_I(inode)->log_mutex);
2936
2937 btrfs_free_path(path);
2938 btrfs_free_path(dst_path);
2939 return err;
2940}
2941
2942/*
2943 * follow the dentry parent pointers up the chain and see if any
2944 * of the directories in it require a full commit before they can
2945 * be logged. Returns zero if nothing special needs to be done or 1 if
2946 * a full commit is required.
2947 */
2948static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
2949 struct inode *inode,
2950 struct dentry *parent,
2951 struct super_block *sb,
2952 u64 last_committed)
2953{
2954 int ret = 0;
2955 struct btrfs_root *root;
2956 struct dentry *old_parent = NULL;
2957
2958 /*
2959 * for regular files, if its inode is already on disk, we don't
2960 * have to worry about the parents at all. This is because
2961 * we can use the last_unlink_trans field to record renames
2962 * and other fun in this file.
2963 */
2964 if (S_ISREG(inode->i_mode) &&
2965 BTRFS_I(inode)->generation <= last_committed &&
2966 BTRFS_I(inode)->last_unlink_trans <= last_committed)
2967 goto out;
2968
2969 if (!S_ISDIR(inode->i_mode)) {
2970 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
2971 goto out;
2972 inode = parent->d_inode;
2973 }
2974
2975 while (1) {
2976 BTRFS_I(inode)->logged_trans = trans->transid;
2977 smp_mb();
2978
2979 if (BTRFS_I(inode)->last_unlink_trans > last_committed) {
2980 root = BTRFS_I(inode)->root;
2981
2982 /*
2983 * make sure any commits to the log are forced
2984 * to be full commits
2985 */
2986 root->fs_info->last_trans_log_full_commit =
2987 trans->transid;
2988 ret = 1;
2989 break;
2990 }
2991
2992 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
2993 break;
2994
2995 if (IS_ROOT(parent))
2996 break;
2997
2998 parent = dget_parent(parent);
2999 dput(old_parent);
3000 old_parent = parent;
3001 inode = parent->d_inode;
3002
3003 }
3004 dput(old_parent);
3005out:
3006 return ret;
3007}
3008
3009static int inode_in_log(struct btrfs_trans_handle *trans,
3010 struct inode *inode)
3011{
3012 struct btrfs_root *root = BTRFS_I(inode)->root;
3013 int ret = 0;
3014
3015 mutex_lock(&root->log_mutex);
3016 if (BTRFS_I(inode)->logged_trans == trans->transid &&
3017 BTRFS_I(inode)->last_sub_trans <= root->last_log_commit)
3018 ret = 1;
3019 mutex_unlock(&root->log_mutex);
3020 return ret;
3021}
3022
3023
3024/*
3025 * helper function around btrfs_log_inode to make sure newly created
3026 * parent directories also end up in the log. A minimal inode and backref
3027 * only logging is done of any parent directories that are older than
3028 * the last committed transaction
3029 */
3030int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
3031 struct btrfs_root *root, struct inode *inode,
3032 struct dentry *parent, int exists_only)
3033{
3034 int inode_only = exists_only ? LOG_INODE_EXISTS : LOG_INODE_ALL;
3035 struct super_block *sb;
3036 struct dentry *old_parent = NULL;
3037 int ret = 0;
3038 u64 last_committed = root->fs_info->last_trans_committed;
3039
3040 sb = inode->i_sb;
3041
3042 if (btrfs_test_opt(root, NOTREELOG)) {
3043 ret = 1;
3044 goto end_no_trans;
3045 }
3046
3047 if (root->fs_info->last_trans_log_full_commit >
3048 root->fs_info->last_trans_committed) {
3049 ret = 1;
3050 goto end_no_trans;
3051 }
3052
3053 if (root != BTRFS_I(inode)->root ||
3054 btrfs_root_refs(&root->root_item) == 0) {
3055 ret = 1;
3056 goto end_no_trans;
3057 }
3058
3059 ret = check_parent_dirs_for_sync(trans, inode, parent,
3060 sb, last_committed);
3061 if (ret)
3062 goto end_no_trans;
3063
3064 if (inode_in_log(trans, inode)) {
3065 ret = BTRFS_NO_LOG_SYNC;
3066 goto end_no_trans;
3067 }
3068
3069 ret = start_log_trans(trans, root);
3070 if (ret)
3071 goto end_trans;
3072
3073 ret = btrfs_log_inode(trans, root, inode, inode_only);
3074 if (ret)
3075 goto end_trans;
3076
3077 /*
3078 * for regular files, if its inode is already on disk, we don't
3079 * have to worry about the parents at all. This is because
3080 * we can use the last_unlink_trans field to record renames
3081 * and other fun in this file.
3082 */
3083 if (S_ISREG(inode->i_mode) &&
3084 BTRFS_I(inode)->generation <= last_committed &&
3085 BTRFS_I(inode)->last_unlink_trans <= last_committed) {
3086 ret = 0;
3087 goto end_trans;
3088 }
3089
3090 inode_only = LOG_INODE_EXISTS;
3091 while (1) {
3092 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3093 break;
3094
3095 inode = parent->d_inode;
3096 if (root != BTRFS_I(inode)->root)
3097 break;
3098
3099 if (BTRFS_I(inode)->generation >
3100 root->fs_info->last_trans_committed) {
3101 ret = btrfs_log_inode(trans, root, inode, inode_only);
3102 if (ret)
3103 goto end_trans;
3104 }
3105 if (IS_ROOT(parent))
3106 break;
3107
3108 parent = dget_parent(parent);
3109 dput(old_parent);
3110 old_parent = parent;
3111 }
3112 ret = 0;
3113end_trans:
3114 dput(old_parent);
3115 if (ret < 0) {
3116 BUG_ON(ret != -ENOSPC);
3117 root->fs_info->last_trans_log_full_commit = trans->transid;
3118 ret = 1;
3119 }
3120 btrfs_end_log_trans(root);
3121end_no_trans:
3122 return ret;
3123}
3124
3125/*
3126 * it is not safe to log dentry if the chunk root has added new
3127 * chunks. This returns 0 if the dentry was logged, and 1 otherwise.
3128 * If this returns 1, you must commit the transaction to safely get your
3129 * data on disk.
3130 */
3131int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
3132 struct btrfs_root *root, struct dentry *dentry)
3133{
3134 struct dentry *parent = dget_parent(dentry);
3135 int ret;
3136
3137 ret = btrfs_log_inode_parent(trans, root, dentry->d_inode, parent, 0);
3138 dput(parent);
3139
3140 return ret;
3141}
3142
3143/*
3144 * should be called during mount to recover any replay any log trees
3145 * from the FS
3146 */
3147int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
3148{
3149 int ret;
3150 struct btrfs_path *path;
3151 struct btrfs_trans_handle *trans;
3152 struct btrfs_key key;
3153 struct btrfs_key found_key;
3154 struct btrfs_key tmp_key;
3155 struct btrfs_root *log;
3156 struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
3157 struct walk_control wc = {
3158 .process_func = process_one_buffer,
3159 .stage = 0,
3160 };
3161
3162 path = btrfs_alloc_path();
3163 if (!path)
3164 return -ENOMEM;
3165
3166 fs_info->log_root_recovering = 1;
3167
3168 trans = btrfs_start_transaction(fs_info->tree_root, 0);
3169 BUG_ON(IS_ERR(trans));
3170
3171 wc.trans = trans;
3172 wc.pin = 1;
3173
3174 ret = walk_log_tree(trans, log_root_tree, &wc);
3175 BUG_ON(ret);
3176
3177again:
3178 key.objectid = BTRFS_TREE_LOG_OBJECTID;
3179 key.offset = (u64)-1;
3180 btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
3181
3182 while (1) {
3183 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
3184 if (ret < 0)
3185 break;
3186 if (ret > 0) {
3187 if (path->slots[0] == 0)
3188 break;
3189 path->slots[0]--;
3190 }
3191 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
3192 path->slots[0]);
3193 btrfs_release_path(path);
3194 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
3195 break;
3196
3197 log = btrfs_read_fs_root_no_radix(log_root_tree,
3198 &found_key);
3199 BUG_ON(IS_ERR(log));
3200
3201 tmp_key.objectid = found_key.offset;
3202 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
3203 tmp_key.offset = (u64)-1;
3204
3205 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
3206 BUG_ON(IS_ERR_OR_NULL(wc.replay_dest));
3207
3208 wc.replay_dest->log_root = log;
3209 btrfs_record_root_in_trans(trans, wc.replay_dest);
3210 ret = walk_log_tree(trans, log, &wc);
3211 BUG_ON(ret);
3212
3213 if (wc.stage == LOG_WALK_REPLAY_ALL) {
3214 ret = fixup_inode_link_counts(trans, wc.replay_dest,
3215 path);
3216 BUG_ON(ret);
3217 }
3218
3219 key.offset = found_key.offset - 1;
3220 wc.replay_dest->log_root = NULL;
3221 free_extent_buffer(log->node);
3222 free_extent_buffer(log->commit_root);
3223 kfree(log);
3224
3225 if (found_key.offset == 0)
3226 break;
3227 }
3228 btrfs_release_path(path);
3229
3230 /* step one is to pin it all, step two is to replay just inodes */
3231 if (wc.pin) {
3232 wc.pin = 0;
3233 wc.process_func = replay_one_buffer;
3234 wc.stage = LOG_WALK_REPLAY_INODES;
3235 goto again;
3236 }
3237 /* step three is to replay everything */
3238 if (wc.stage < LOG_WALK_REPLAY_ALL) {
3239 wc.stage++;
3240 goto again;
3241 }
3242
3243 btrfs_free_path(path);
3244
3245 free_extent_buffer(log_root_tree->node);
3246 log_root_tree->log_root = NULL;
3247 fs_info->log_root_recovering = 0;
3248
3249 /* step 4: commit the transaction, which also unpins the blocks */
3250 btrfs_commit_transaction(trans, fs_info->tree_root);
3251
3252 kfree(log_root_tree);
3253 return 0;
3254}
3255
3256/*
3257 * there are some corner cases where we want to force a full
3258 * commit instead of allowing a directory to be logged.
3259 *
3260 * They revolve around files there were unlinked from the directory, and
3261 * this function updates the parent directory so that a full commit is
3262 * properly done if it is fsync'd later after the unlinks are done.
3263 */
3264void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
3265 struct inode *dir, struct inode *inode,
3266 int for_rename)
3267{
3268 /*
3269 * when we're logging a file, if it hasn't been renamed
3270 * or unlinked, and its inode is fully committed on disk,
3271 * we don't have to worry about walking up the directory chain
3272 * to log its parents.
3273 *
3274 * So, we use the last_unlink_trans field to put this transid
3275 * into the file. When the file is logged we check it and
3276 * don't log the parents if the file is fully on disk.
3277 */
3278 if (S_ISREG(inode->i_mode))
3279 BTRFS_I(inode)->last_unlink_trans = trans->transid;
3280
3281 /*
3282 * if this directory was already logged any new
3283 * names for this file/dir will get recorded
3284 */
3285 smp_mb();
3286 if (BTRFS_I(dir)->logged_trans == trans->transid)
3287 return;
3288
3289 /*
3290 * if the inode we're about to unlink was logged,
3291 * the log will be properly updated for any new names
3292 */
3293 if (BTRFS_I(inode)->logged_trans == trans->transid)
3294 return;
3295
3296 /*
3297 * when renaming files across directories, if the directory
3298 * there we're unlinking from gets fsync'd later on, there's
3299 * no way to find the destination directory later and fsync it
3300 * properly. So, we have to be conservative and force commits
3301 * so the new name gets discovered.
3302 */
3303 if (for_rename)
3304 goto record;
3305
3306 /* we can safely do the unlink without any special recording */
3307 return;
3308
3309record:
3310 BTRFS_I(dir)->last_unlink_trans = trans->transid;
3311}
3312
3313/*
3314 * Call this after adding a new name for a file and it will properly
3315 * update the log to reflect the new name.
3316 *
3317 * It will return zero if all goes well, and it will return 1 if a
3318 * full transaction commit is required.
3319 */
3320int btrfs_log_new_name(struct btrfs_trans_handle *trans,
3321 struct inode *inode, struct inode *old_dir,
3322 struct dentry *parent)
3323{
3324 struct btrfs_root * root = BTRFS_I(inode)->root;
3325
3326 /*
3327 * this will force the logging code to walk the dentry chain
3328 * up for the file
3329 */
3330 if (S_ISREG(inode->i_mode))
3331 BTRFS_I(inode)->last_unlink_trans = trans->transid;
3332
3333 /*
3334 * if this inode hasn't been logged and directory we're renaming it
3335 * from hasn't been logged, we don't need to log it
3336 */
3337 if (BTRFS_I(inode)->logged_trans <=
3338 root->fs_info->last_trans_committed &&
3339 (!old_dir || BTRFS_I(old_dir)->logged_trans <=
3340 root->fs_info->last_trans_committed))
3341 return 0;
3342
3343 return btrfs_log_inode_parent(trans, root, inode, parent, 1);
3344}
3345
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2008 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/blkdev.h>
9#include <linux/list_sort.h>
10#include <linux/iversion.h>
11#include "misc.h"
12#include "ctree.h"
13#include "tree-log.h"
14#include "disk-io.h"
15#include "locking.h"
16#include "print-tree.h"
17#include "backref.h"
18#include "compression.h"
19#include "qgroup.h"
20#include "block-group.h"
21#include "space-info.h"
22#include "zoned.h"
23
24/* magic values for the inode_only field in btrfs_log_inode:
25 *
26 * LOG_INODE_ALL means to log everything
27 * LOG_INODE_EXISTS means to log just enough to recreate the inode
28 * during log replay
29 */
30enum {
31 LOG_INODE_ALL,
32 LOG_INODE_EXISTS,
33 LOG_OTHER_INODE,
34 LOG_OTHER_INODE_ALL,
35};
36
37/*
38 * directory trouble cases
39 *
40 * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
41 * log, we must force a full commit before doing an fsync of the directory
42 * where the unlink was done.
43 * ---> record transid of last unlink/rename per directory
44 *
45 * mkdir foo/some_dir
46 * normal commit
47 * rename foo/some_dir foo2/some_dir
48 * mkdir foo/some_dir
49 * fsync foo/some_dir/some_file
50 *
51 * The fsync above will unlink the original some_dir without recording
52 * it in its new location (foo2). After a crash, some_dir will be gone
53 * unless the fsync of some_file forces a full commit
54 *
55 * 2) we must log any new names for any file or dir that is in the fsync
56 * log. ---> check inode while renaming/linking.
57 *
58 * 2a) we must log any new names for any file or dir during rename
59 * when the directory they are being removed from was logged.
60 * ---> check inode and old parent dir during rename
61 *
62 * 2a is actually the more important variant. With the extra logging
63 * a crash might unlink the old name without recreating the new one
64 *
65 * 3) after a crash, we must go through any directories with a link count
66 * of zero and redo the rm -rf
67 *
68 * mkdir f1/foo
69 * normal commit
70 * rm -rf f1/foo
71 * fsync(f1)
72 *
73 * The directory f1 was fully removed from the FS, but fsync was never
74 * called on f1, only its parent dir. After a crash the rm -rf must
75 * be replayed. This must be able to recurse down the entire
76 * directory tree. The inode link count fixup code takes care of the
77 * ugly details.
78 */
79
80/*
81 * stages for the tree walking. The first
82 * stage (0) is to only pin down the blocks we find
83 * the second stage (1) is to make sure that all the inodes
84 * we find in the log are created in the subvolume.
85 *
86 * The last stage is to deal with directories and links and extents
87 * and all the other fun semantics
88 */
89enum {
90 LOG_WALK_PIN_ONLY,
91 LOG_WALK_REPLAY_INODES,
92 LOG_WALK_REPLAY_DIR_INDEX,
93 LOG_WALK_REPLAY_ALL,
94};
95
96static int btrfs_log_inode(struct btrfs_trans_handle *trans,
97 struct btrfs_root *root, struct btrfs_inode *inode,
98 int inode_only,
99 struct btrfs_log_ctx *ctx);
100static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
101 struct btrfs_root *root,
102 struct btrfs_path *path, u64 objectid);
103static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
104 struct btrfs_root *root,
105 struct btrfs_root *log,
106 struct btrfs_path *path,
107 u64 dirid, int del_all);
108static void wait_log_commit(struct btrfs_root *root, int transid);
109
110/*
111 * tree logging is a special write ahead log used to make sure that
112 * fsyncs and O_SYNCs can happen without doing full tree commits.
113 *
114 * Full tree commits are expensive because they require commonly
115 * modified blocks to be recowed, creating many dirty pages in the
116 * extent tree an 4x-6x higher write load than ext3.
117 *
118 * Instead of doing a tree commit on every fsync, we use the
119 * key ranges and transaction ids to find items for a given file or directory
120 * that have changed in this transaction. Those items are copied into
121 * a special tree (one per subvolume root), that tree is written to disk
122 * and then the fsync is considered complete.
123 *
124 * After a crash, items are copied out of the log-tree back into the
125 * subvolume tree. Any file data extents found are recorded in the extent
126 * allocation tree, and the log-tree freed.
127 *
128 * The log tree is read three times, once to pin down all the extents it is
129 * using in ram and once, once to create all the inodes logged in the tree
130 * and once to do all the other items.
131 */
132
133/*
134 * start a sub transaction and setup the log tree
135 * this increments the log tree writer count to make the people
136 * syncing the tree wait for us to finish
137 */
138static int start_log_trans(struct btrfs_trans_handle *trans,
139 struct btrfs_root *root,
140 struct btrfs_log_ctx *ctx)
141{
142 struct btrfs_fs_info *fs_info = root->fs_info;
143 struct btrfs_root *tree_root = fs_info->tree_root;
144 const bool zoned = btrfs_is_zoned(fs_info);
145 int ret = 0;
146 bool created = false;
147
148 /*
149 * First check if the log root tree was already created. If not, create
150 * it before locking the root's log_mutex, just to keep lockdep happy.
151 */
152 if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state)) {
153 mutex_lock(&tree_root->log_mutex);
154 if (!fs_info->log_root_tree) {
155 ret = btrfs_init_log_root_tree(trans, fs_info);
156 if (!ret) {
157 set_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state);
158 created = true;
159 }
160 }
161 mutex_unlock(&tree_root->log_mutex);
162 if (ret)
163 return ret;
164 }
165
166 mutex_lock(&root->log_mutex);
167
168again:
169 if (root->log_root) {
170 int index = (root->log_transid + 1) % 2;
171
172 if (btrfs_need_log_full_commit(trans)) {
173 ret = -EAGAIN;
174 goto out;
175 }
176
177 if (zoned && atomic_read(&root->log_commit[index])) {
178 wait_log_commit(root, root->log_transid - 1);
179 goto again;
180 }
181
182 if (!root->log_start_pid) {
183 clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
184 root->log_start_pid = current->pid;
185 } else if (root->log_start_pid != current->pid) {
186 set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
187 }
188 } else {
189 /*
190 * This means fs_info->log_root_tree was already created
191 * for some other FS trees. Do the full commit not to mix
192 * nodes from multiple log transactions to do sequential
193 * writing.
194 */
195 if (zoned && !created) {
196 ret = -EAGAIN;
197 goto out;
198 }
199
200 ret = btrfs_add_log_tree(trans, root);
201 if (ret)
202 goto out;
203
204 set_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
205 clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
206 root->log_start_pid = current->pid;
207 }
208
209 atomic_inc(&root->log_writers);
210 if (ctx && !ctx->logging_new_name) {
211 int index = root->log_transid % 2;
212 list_add_tail(&ctx->list, &root->log_ctxs[index]);
213 ctx->log_transid = root->log_transid;
214 }
215
216out:
217 mutex_unlock(&root->log_mutex);
218 return ret;
219}
220
221/*
222 * returns 0 if there was a log transaction running and we were able
223 * to join, or returns -ENOENT if there were not transactions
224 * in progress
225 */
226static int join_running_log_trans(struct btrfs_root *root)
227{
228 const bool zoned = btrfs_is_zoned(root->fs_info);
229 int ret = -ENOENT;
230
231 if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state))
232 return ret;
233
234 mutex_lock(&root->log_mutex);
235again:
236 if (root->log_root) {
237 int index = (root->log_transid + 1) % 2;
238
239 ret = 0;
240 if (zoned && atomic_read(&root->log_commit[index])) {
241 wait_log_commit(root, root->log_transid - 1);
242 goto again;
243 }
244 atomic_inc(&root->log_writers);
245 }
246 mutex_unlock(&root->log_mutex);
247 return ret;
248}
249
250/*
251 * This either makes the current running log transaction wait
252 * until you call btrfs_end_log_trans() or it makes any future
253 * log transactions wait until you call btrfs_end_log_trans()
254 */
255void btrfs_pin_log_trans(struct btrfs_root *root)
256{
257 atomic_inc(&root->log_writers);
258}
259
260/*
261 * indicate we're done making changes to the log tree
262 * and wake up anyone waiting to do a sync
263 */
264void btrfs_end_log_trans(struct btrfs_root *root)
265{
266 if (atomic_dec_and_test(&root->log_writers)) {
267 /* atomic_dec_and_test implies a barrier */
268 cond_wake_up_nomb(&root->log_writer_wait);
269 }
270}
271
272static int btrfs_write_tree_block(struct extent_buffer *buf)
273{
274 return filemap_fdatawrite_range(buf->pages[0]->mapping, buf->start,
275 buf->start + buf->len - 1);
276}
277
278static void btrfs_wait_tree_block_writeback(struct extent_buffer *buf)
279{
280 filemap_fdatawait_range(buf->pages[0]->mapping,
281 buf->start, buf->start + buf->len - 1);
282}
283
284/*
285 * the walk control struct is used to pass state down the chain when
286 * processing the log tree. The stage field tells us which part
287 * of the log tree processing we are currently doing. The others
288 * are state fields used for that specific part
289 */
290struct walk_control {
291 /* should we free the extent on disk when done? This is used
292 * at transaction commit time while freeing a log tree
293 */
294 int free;
295
296 /* should we write out the extent buffer? This is used
297 * while flushing the log tree to disk during a sync
298 */
299 int write;
300
301 /* should we wait for the extent buffer io to finish? Also used
302 * while flushing the log tree to disk for a sync
303 */
304 int wait;
305
306 /* pin only walk, we record which extents on disk belong to the
307 * log trees
308 */
309 int pin;
310
311 /* what stage of the replay code we're currently in */
312 int stage;
313
314 /*
315 * Ignore any items from the inode currently being processed. Needs
316 * to be set every time we find a BTRFS_INODE_ITEM_KEY and we are in
317 * the LOG_WALK_REPLAY_INODES stage.
318 */
319 bool ignore_cur_inode;
320
321 /* the root we are currently replaying */
322 struct btrfs_root *replay_dest;
323
324 /* the trans handle for the current replay */
325 struct btrfs_trans_handle *trans;
326
327 /* the function that gets used to process blocks we find in the
328 * tree. Note the extent_buffer might not be up to date when it is
329 * passed in, and it must be checked or read if you need the data
330 * inside it
331 */
332 int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
333 struct walk_control *wc, u64 gen, int level);
334};
335
336/*
337 * process_func used to pin down extents, write them or wait on them
338 */
339static int process_one_buffer(struct btrfs_root *log,
340 struct extent_buffer *eb,
341 struct walk_control *wc, u64 gen, int level)
342{
343 struct btrfs_fs_info *fs_info = log->fs_info;
344 int ret = 0;
345
346 /*
347 * If this fs is mixed then we need to be able to process the leaves to
348 * pin down any logged extents, so we have to read the block.
349 */
350 if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
351 ret = btrfs_read_buffer(eb, gen, level, NULL);
352 if (ret)
353 return ret;
354 }
355
356 if (wc->pin)
357 ret = btrfs_pin_extent_for_log_replay(wc->trans, eb->start,
358 eb->len);
359
360 if (!ret && btrfs_buffer_uptodate(eb, gen, 0)) {
361 if (wc->pin && btrfs_header_level(eb) == 0)
362 ret = btrfs_exclude_logged_extents(eb);
363 if (wc->write)
364 btrfs_write_tree_block(eb);
365 if (wc->wait)
366 btrfs_wait_tree_block_writeback(eb);
367 }
368 return ret;
369}
370
371/*
372 * Item overwrite used by replay and tree logging. eb, slot and key all refer
373 * to the src data we are copying out.
374 *
375 * root is the tree we are copying into, and path is a scratch
376 * path for use in this function (it should be released on entry and
377 * will be released on exit).
378 *
379 * If the key is already in the destination tree the existing item is
380 * overwritten. If the existing item isn't big enough, it is extended.
381 * If it is too large, it is truncated.
382 *
383 * If the key isn't in the destination yet, a new item is inserted.
384 */
385static noinline int overwrite_item(struct btrfs_trans_handle *trans,
386 struct btrfs_root *root,
387 struct btrfs_path *path,
388 struct extent_buffer *eb, int slot,
389 struct btrfs_key *key)
390{
391 int ret;
392 u32 item_size;
393 u64 saved_i_size = 0;
394 int save_old_i_size = 0;
395 unsigned long src_ptr;
396 unsigned long dst_ptr;
397 int overwrite_root = 0;
398 bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
399
400 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
401 overwrite_root = 1;
402
403 item_size = btrfs_item_size_nr(eb, slot);
404 src_ptr = btrfs_item_ptr_offset(eb, slot);
405
406 /* look for the key in the destination tree */
407 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
408 if (ret < 0)
409 return ret;
410
411 if (ret == 0) {
412 char *src_copy;
413 char *dst_copy;
414 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
415 path->slots[0]);
416 if (dst_size != item_size)
417 goto insert;
418
419 if (item_size == 0) {
420 btrfs_release_path(path);
421 return 0;
422 }
423 dst_copy = kmalloc(item_size, GFP_NOFS);
424 src_copy = kmalloc(item_size, GFP_NOFS);
425 if (!dst_copy || !src_copy) {
426 btrfs_release_path(path);
427 kfree(dst_copy);
428 kfree(src_copy);
429 return -ENOMEM;
430 }
431
432 read_extent_buffer(eb, src_copy, src_ptr, item_size);
433
434 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
435 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
436 item_size);
437 ret = memcmp(dst_copy, src_copy, item_size);
438
439 kfree(dst_copy);
440 kfree(src_copy);
441 /*
442 * they have the same contents, just return, this saves
443 * us from cowing blocks in the destination tree and doing
444 * extra writes that may not have been done by a previous
445 * sync
446 */
447 if (ret == 0) {
448 btrfs_release_path(path);
449 return 0;
450 }
451
452 /*
453 * We need to load the old nbytes into the inode so when we
454 * replay the extents we've logged we get the right nbytes.
455 */
456 if (inode_item) {
457 struct btrfs_inode_item *item;
458 u64 nbytes;
459 u32 mode;
460
461 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
462 struct btrfs_inode_item);
463 nbytes = btrfs_inode_nbytes(path->nodes[0], item);
464 item = btrfs_item_ptr(eb, slot,
465 struct btrfs_inode_item);
466 btrfs_set_inode_nbytes(eb, item, nbytes);
467
468 /*
469 * If this is a directory we need to reset the i_size to
470 * 0 so that we can set it up properly when replaying
471 * the rest of the items in this log.
472 */
473 mode = btrfs_inode_mode(eb, item);
474 if (S_ISDIR(mode))
475 btrfs_set_inode_size(eb, item, 0);
476 }
477 } else if (inode_item) {
478 struct btrfs_inode_item *item;
479 u32 mode;
480
481 /*
482 * New inode, set nbytes to 0 so that the nbytes comes out
483 * properly when we replay the extents.
484 */
485 item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
486 btrfs_set_inode_nbytes(eb, item, 0);
487
488 /*
489 * If this is a directory we need to reset the i_size to 0 so
490 * that we can set it up properly when replaying the rest of
491 * the items in this log.
492 */
493 mode = btrfs_inode_mode(eb, item);
494 if (S_ISDIR(mode))
495 btrfs_set_inode_size(eb, item, 0);
496 }
497insert:
498 btrfs_release_path(path);
499 /* try to insert the key into the destination tree */
500 path->skip_release_on_error = 1;
501 ret = btrfs_insert_empty_item(trans, root, path,
502 key, item_size);
503 path->skip_release_on_error = 0;
504
505 /* make sure any existing item is the correct size */
506 if (ret == -EEXIST || ret == -EOVERFLOW) {
507 u32 found_size;
508 found_size = btrfs_item_size_nr(path->nodes[0],
509 path->slots[0]);
510 if (found_size > item_size)
511 btrfs_truncate_item(path, item_size, 1);
512 else if (found_size < item_size)
513 btrfs_extend_item(path, item_size - found_size);
514 } else if (ret) {
515 return ret;
516 }
517 dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
518 path->slots[0]);
519
520 /* don't overwrite an existing inode if the generation number
521 * was logged as zero. This is done when the tree logging code
522 * is just logging an inode to make sure it exists after recovery.
523 *
524 * Also, don't overwrite i_size on directories during replay.
525 * log replay inserts and removes directory items based on the
526 * state of the tree found in the subvolume, and i_size is modified
527 * as it goes
528 */
529 if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
530 struct btrfs_inode_item *src_item;
531 struct btrfs_inode_item *dst_item;
532
533 src_item = (struct btrfs_inode_item *)src_ptr;
534 dst_item = (struct btrfs_inode_item *)dst_ptr;
535
536 if (btrfs_inode_generation(eb, src_item) == 0) {
537 struct extent_buffer *dst_eb = path->nodes[0];
538 const u64 ino_size = btrfs_inode_size(eb, src_item);
539
540 /*
541 * For regular files an ino_size == 0 is used only when
542 * logging that an inode exists, as part of a directory
543 * fsync, and the inode wasn't fsynced before. In this
544 * case don't set the size of the inode in the fs/subvol
545 * tree, otherwise we would be throwing valid data away.
546 */
547 if (S_ISREG(btrfs_inode_mode(eb, src_item)) &&
548 S_ISREG(btrfs_inode_mode(dst_eb, dst_item)) &&
549 ino_size != 0)
550 btrfs_set_inode_size(dst_eb, dst_item, ino_size);
551 goto no_copy;
552 }
553
554 if (overwrite_root &&
555 S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
556 S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
557 save_old_i_size = 1;
558 saved_i_size = btrfs_inode_size(path->nodes[0],
559 dst_item);
560 }
561 }
562
563 copy_extent_buffer(path->nodes[0], eb, dst_ptr,
564 src_ptr, item_size);
565
566 if (save_old_i_size) {
567 struct btrfs_inode_item *dst_item;
568 dst_item = (struct btrfs_inode_item *)dst_ptr;
569 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
570 }
571
572 /* make sure the generation is filled in */
573 if (key->type == BTRFS_INODE_ITEM_KEY) {
574 struct btrfs_inode_item *dst_item;
575 dst_item = (struct btrfs_inode_item *)dst_ptr;
576 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
577 btrfs_set_inode_generation(path->nodes[0], dst_item,
578 trans->transid);
579 }
580 }
581no_copy:
582 btrfs_mark_buffer_dirty(path->nodes[0]);
583 btrfs_release_path(path);
584 return 0;
585}
586
587/*
588 * simple helper to read an inode off the disk from a given root
589 * This can only be called for subvolume roots and not for the log
590 */
591static noinline struct inode *read_one_inode(struct btrfs_root *root,
592 u64 objectid)
593{
594 struct inode *inode;
595
596 inode = btrfs_iget(root->fs_info->sb, objectid, root);
597 if (IS_ERR(inode))
598 inode = NULL;
599 return inode;
600}
601
602/* replays a single extent in 'eb' at 'slot' with 'key' into the
603 * subvolume 'root'. path is released on entry and should be released
604 * on exit.
605 *
606 * extents in the log tree have not been allocated out of the extent
607 * tree yet. So, this completes the allocation, taking a reference
608 * as required if the extent already exists or creating a new extent
609 * if it isn't in the extent allocation tree yet.
610 *
611 * The extent is inserted into the file, dropping any existing extents
612 * from the file that overlap the new one.
613 */
614static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
615 struct btrfs_root *root,
616 struct btrfs_path *path,
617 struct extent_buffer *eb, int slot,
618 struct btrfs_key *key)
619{
620 struct btrfs_drop_extents_args drop_args = { 0 };
621 struct btrfs_fs_info *fs_info = root->fs_info;
622 int found_type;
623 u64 extent_end;
624 u64 start = key->offset;
625 u64 nbytes = 0;
626 struct btrfs_file_extent_item *item;
627 struct inode *inode = NULL;
628 unsigned long size;
629 int ret = 0;
630
631 item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
632 found_type = btrfs_file_extent_type(eb, item);
633
634 if (found_type == BTRFS_FILE_EXTENT_REG ||
635 found_type == BTRFS_FILE_EXTENT_PREALLOC) {
636 nbytes = btrfs_file_extent_num_bytes(eb, item);
637 extent_end = start + nbytes;
638
639 /*
640 * We don't add to the inodes nbytes if we are prealloc or a
641 * hole.
642 */
643 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
644 nbytes = 0;
645 } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
646 size = btrfs_file_extent_ram_bytes(eb, item);
647 nbytes = btrfs_file_extent_ram_bytes(eb, item);
648 extent_end = ALIGN(start + size,
649 fs_info->sectorsize);
650 } else {
651 ret = 0;
652 goto out;
653 }
654
655 inode = read_one_inode(root, key->objectid);
656 if (!inode) {
657 ret = -EIO;
658 goto out;
659 }
660
661 /*
662 * first check to see if we already have this extent in the
663 * file. This must be done before the btrfs_drop_extents run
664 * so we don't try to drop this extent.
665 */
666 ret = btrfs_lookup_file_extent(trans, root, path,
667 btrfs_ino(BTRFS_I(inode)), start, 0);
668
669 if (ret == 0 &&
670 (found_type == BTRFS_FILE_EXTENT_REG ||
671 found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
672 struct btrfs_file_extent_item cmp1;
673 struct btrfs_file_extent_item cmp2;
674 struct btrfs_file_extent_item *existing;
675 struct extent_buffer *leaf;
676
677 leaf = path->nodes[0];
678 existing = btrfs_item_ptr(leaf, path->slots[0],
679 struct btrfs_file_extent_item);
680
681 read_extent_buffer(eb, &cmp1, (unsigned long)item,
682 sizeof(cmp1));
683 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
684 sizeof(cmp2));
685
686 /*
687 * we already have a pointer to this exact extent,
688 * we don't have to do anything
689 */
690 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
691 btrfs_release_path(path);
692 goto out;
693 }
694 }
695 btrfs_release_path(path);
696
697 /* drop any overlapping extents */
698 drop_args.start = start;
699 drop_args.end = extent_end;
700 drop_args.drop_cache = true;
701 ret = btrfs_drop_extents(trans, root, BTRFS_I(inode), &drop_args);
702 if (ret)
703 goto out;
704
705 if (found_type == BTRFS_FILE_EXTENT_REG ||
706 found_type == BTRFS_FILE_EXTENT_PREALLOC) {
707 u64 offset;
708 unsigned long dest_offset;
709 struct btrfs_key ins;
710
711 if (btrfs_file_extent_disk_bytenr(eb, item) == 0 &&
712 btrfs_fs_incompat(fs_info, NO_HOLES))
713 goto update_inode;
714
715 ret = btrfs_insert_empty_item(trans, root, path, key,
716 sizeof(*item));
717 if (ret)
718 goto out;
719 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
720 path->slots[0]);
721 copy_extent_buffer(path->nodes[0], eb, dest_offset,
722 (unsigned long)item, sizeof(*item));
723
724 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
725 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
726 ins.type = BTRFS_EXTENT_ITEM_KEY;
727 offset = key->offset - btrfs_file_extent_offset(eb, item);
728
729 /*
730 * Manually record dirty extent, as here we did a shallow
731 * file extent item copy and skip normal backref update,
732 * but modifying extent tree all by ourselves.
733 * So need to manually record dirty extent for qgroup,
734 * as the owner of the file extent changed from log tree
735 * (doesn't affect qgroup) to fs/file tree(affects qgroup)
736 */
737 ret = btrfs_qgroup_trace_extent(trans,
738 btrfs_file_extent_disk_bytenr(eb, item),
739 btrfs_file_extent_disk_num_bytes(eb, item),
740 GFP_NOFS);
741 if (ret < 0)
742 goto out;
743
744 if (ins.objectid > 0) {
745 struct btrfs_ref ref = { 0 };
746 u64 csum_start;
747 u64 csum_end;
748 LIST_HEAD(ordered_sums);
749
750 /*
751 * is this extent already allocated in the extent
752 * allocation tree? If so, just add a reference
753 */
754 ret = btrfs_lookup_data_extent(fs_info, ins.objectid,
755 ins.offset);
756 if (ret < 0) {
757 goto out;
758 } else if (ret == 0) {
759 btrfs_init_generic_ref(&ref,
760 BTRFS_ADD_DELAYED_REF,
761 ins.objectid, ins.offset, 0);
762 btrfs_init_data_ref(&ref,
763 root->root_key.objectid,
764 key->objectid, offset);
765 ret = btrfs_inc_extent_ref(trans, &ref);
766 if (ret)
767 goto out;
768 } else {
769 /*
770 * insert the extent pointer in the extent
771 * allocation tree
772 */
773 ret = btrfs_alloc_logged_file_extent(trans,
774 root->root_key.objectid,
775 key->objectid, offset, &ins);
776 if (ret)
777 goto out;
778 }
779 btrfs_release_path(path);
780
781 if (btrfs_file_extent_compression(eb, item)) {
782 csum_start = ins.objectid;
783 csum_end = csum_start + ins.offset;
784 } else {
785 csum_start = ins.objectid +
786 btrfs_file_extent_offset(eb, item);
787 csum_end = csum_start +
788 btrfs_file_extent_num_bytes(eb, item);
789 }
790
791 ret = btrfs_lookup_csums_range(root->log_root,
792 csum_start, csum_end - 1,
793 &ordered_sums, 0);
794 if (ret)
795 goto out;
796 /*
797 * Now delete all existing cums in the csum root that
798 * cover our range. We do this because we can have an
799 * extent that is completely referenced by one file
800 * extent item and partially referenced by another
801 * file extent item (like after using the clone or
802 * extent_same ioctls). In this case if we end up doing
803 * the replay of the one that partially references the
804 * extent first, and we do not do the csum deletion
805 * below, we can get 2 csum items in the csum tree that
806 * overlap each other. For example, imagine our log has
807 * the two following file extent items:
808 *
809 * key (257 EXTENT_DATA 409600)
810 * extent data disk byte 12845056 nr 102400
811 * extent data offset 20480 nr 20480 ram 102400
812 *
813 * key (257 EXTENT_DATA 819200)
814 * extent data disk byte 12845056 nr 102400
815 * extent data offset 0 nr 102400 ram 102400
816 *
817 * Where the second one fully references the 100K extent
818 * that starts at disk byte 12845056, and the log tree
819 * has a single csum item that covers the entire range
820 * of the extent:
821 *
822 * key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
823 *
824 * After the first file extent item is replayed, the
825 * csum tree gets the following csum item:
826 *
827 * key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
828 *
829 * Which covers the 20K sub-range starting at offset 20K
830 * of our extent. Now when we replay the second file
831 * extent item, if we do not delete existing csum items
832 * that cover any of its blocks, we end up getting two
833 * csum items in our csum tree that overlap each other:
834 *
835 * key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
836 * key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
837 *
838 * Which is a problem, because after this anyone trying
839 * to lookup up for the checksum of any block of our
840 * extent starting at an offset of 40K or higher, will
841 * end up looking at the second csum item only, which
842 * does not contain the checksum for any block starting
843 * at offset 40K or higher of our extent.
844 */
845 while (!list_empty(&ordered_sums)) {
846 struct btrfs_ordered_sum *sums;
847 sums = list_entry(ordered_sums.next,
848 struct btrfs_ordered_sum,
849 list);
850 if (!ret)
851 ret = btrfs_del_csums(trans,
852 fs_info->csum_root,
853 sums->bytenr,
854 sums->len);
855 if (!ret)
856 ret = btrfs_csum_file_blocks(trans,
857 fs_info->csum_root, sums);
858 list_del(&sums->list);
859 kfree(sums);
860 }
861 if (ret)
862 goto out;
863 } else {
864 btrfs_release_path(path);
865 }
866 } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
867 /* inline extents are easy, we just overwrite them */
868 ret = overwrite_item(trans, root, path, eb, slot, key);
869 if (ret)
870 goto out;
871 }
872
873 ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start,
874 extent_end - start);
875 if (ret)
876 goto out;
877
878update_inode:
879 btrfs_update_inode_bytes(BTRFS_I(inode), nbytes, drop_args.bytes_found);
880 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
881out:
882 if (inode)
883 iput(inode);
884 return ret;
885}
886
887/*
888 * when cleaning up conflicts between the directory names in the
889 * subvolume, directory names in the log and directory names in the
890 * inode back references, we may have to unlink inodes from directories.
891 *
892 * This is a helper function to do the unlink of a specific directory
893 * item
894 */
895static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
896 struct btrfs_root *root,
897 struct btrfs_path *path,
898 struct btrfs_inode *dir,
899 struct btrfs_dir_item *di)
900{
901 struct inode *inode;
902 char *name;
903 int name_len;
904 struct extent_buffer *leaf;
905 struct btrfs_key location;
906 int ret;
907
908 leaf = path->nodes[0];
909
910 btrfs_dir_item_key_to_cpu(leaf, di, &location);
911 name_len = btrfs_dir_name_len(leaf, di);
912 name = kmalloc(name_len, GFP_NOFS);
913 if (!name)
914 return -ENOMEM;
915
916 read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
917 btrfs_release_path(path);
918
919 inode = read_one_inode(root, location.objectid);
920 if (!inode) {
921 ret = -EIO;
922 goto out;
923 }
924
925 ret = link_to_fixup_dir(trans, root, path, location.objectid);
926 if (ret)
927 goto out;
928
929 ret = btrfs_unlink_inode(trans, root, dir, BTRFS_I(inode), name,
930 name_len);
931 if (ret)
932 goto out;
933 else
934 ret = btrfs_run_delayed_items(trans);
935out:
936 kfree(name);
937 iput(inode);
938 return ret;
939}
940
941/*
942 * See if a given name and sequence number found in an inode back reference are
943 * already in a directory and correctly point to this inode.
944 *
945 * Returns: < 0 on error, 0 if the directory entry does not exists and 1 if it
946 * exists.
947 */
948static noinline int inode_in_dir(struct btrfs_root *root,
949 struct btrfs_path *path,
950 u64 dirid, u64 objectid, u64 index,
951 const char *name, int name_len)
952{
953 struct btrfs_dir_item *di;
954 struct btrfs_key location;
955 int ret = 0;
956
957 di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
958 index, name, name_len, 0);
959 if (IS_ERR(di)) {
960 if (PTR_ERR(di) != -ENOENT)
961 ret = PTR_ERR(di);
962 goto out;
963 } else if (di) {
964 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
965 if (location.objectid != objectid)
966 goto out;
967 } else {
968 goto out;
969 }
970
971 btrfs_release_path(path);
972 di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
973 if (IS_ERR(di)) {
974 ret = PTR_ERR(di);
975 goto out;
976 } else if (di) {
977 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
978 if (location.objectid == objectid)
979 ret = 1;
980 }
981out:
982 btrfs_release_path(path);
983 return ret;
984}
985
986/*
987 * helper function to check a log tree for a named back reference in
988 * an inode. This is used to decide if a back reference that is
989 * found in the subvolume conflicts with what we find in the log.
990 *
991 * inode backreferences may have multiple refs in a single item,
992 * during replay we process one reference at a time, and we don't
993 * want to delete valid links to a file from the subvolume if that
994 * link is also in the log.
995 */
996static noinline int backref_in_log(struct btrfs_root *log,
997 struct btrfs_key *key,
998 u64 ref_objectid,
999 const char *name, int namelen)
1000{
1001 struct btrfs_path *path;
1002 int ret;
1003
1004 path = btrfs_alloc_path();
1005 if (!path)
1006 return -ENOMEM;
1007
1008 ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
1009 if (ret < 0) {
1010 goto out;
1011 } else if (ret == 1) {
1012 ret = 0;
1013 goto out;
1014 }
1015
1016 if (key->type == BTRFS_INODE_EXTREF_KEY)
1017 ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
1018 path->slots[0],
1019 ref_objectid,
1020 name, namelen);
1021 else
1022 ret = !!btrfs_find_name_in_backref(path->nodes[0],
1023 path->slots[0],
1024 name, namelen);
1025out:
1026 btrfs_free_path(path);
1027 return ret;
1028}
1029
1030static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
1031 struct btrfs_root *root,
1032 struct btrfs_path *path,
1033 struct btrfs_root *log_root,
1034 struct btrfs_inode *dir,
1035 struct btrfs_inode *inode,
1036 u64 inode_objectid, u64 parent_objectid,
1037 u64 ref_index, char *name, int namelen,
1038 int *search_done)
1039{
1040 int ret;
1041 char *victim_name;
1042 int victim_name_len;
1043 struct extent_buffer *leaf;
1044 struct btrfs_dir_item *di;
1045 struct btrfs_key search_key;
1046 struct btrfs_inode_extref *extref;
1047
1048again:
1049 /* Search old style refs */
1050 search_key.objectid = inode_objectid;
1051 search_key.type = BTRFS_INODE_REF_KEY;
1052 search_key.offset = parent_objectid;
1053 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
1054 if (ret == 0) {
1055 struct btrfs_inode_ref *victim_ref;
1056 unsigned long ptr;
1057 unsigned long ptr_end;
1058
1059 leaf = path->nodes[0];
1060
1061 /* are we trying to overwrite a back ref for the root directory
1062 * if so, just jump out, we're done
1063 */
1064 if (search_key.objectid == search_key.offset)
1065 return 1;
1066
1067 /* check all the names in this back reference to see
1068 * if they are in the log. if so, we allow them to stay
1069 * otherwise they must be unlinked as a conflict
1070 */
1071 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1072 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
1073 while (ptr < ptr_end) {
1074 victim_ref = (struct btrfs_inode_ref *)ptr;
1075 victim_name_len = btrfs_inode_ref_name_len(leaf,
1076 victim_ref);
1077 victim_name = kmalloc(victim_name_len, GFP_NOFS);
1078 if (!victim_name)
1079 return -ENOMEM;
1080
1081 read_extent_buffer(leaf, victim_name,
1082 (unsigned long)(victim_ref + 1),
1083 victim_name_len);
1084
1085 ret = backref_in_log(log_root, &search_key,
1086 parent_objectid, victim_name,
1087 victim_name_len);
1088 if (ret < 0) {
1089 kfree(victim_name);
1090 return ret;
1091 } else if (!ret) {
1092 inc_nlink(&inode->vfs_inode);
1093 btrfs_release_path(path);
1094
1095 ret = btrfs_unlink_inode(trans, root, dir, inode,
1096 victim_name, victim_name_len);
1097 kfree(victim_name);
1098 if (ret)
1099 return ret;
1100 ret = btrfs_run_delayed_items(trans);
1101 if (ret)
1102 return ret;
1103 *search_done = 1;
1104 goto again;
1105 }
1106 kfree(victim_name);
1107
1108 ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
1109 }
1110
1111 /*
1112 * NOTE: we have searched root tree and checked the
1113 * corresponding ref, it does not need to check again.
1114 */
1115 *search_done = 1;
1116 }
1117 btrfs_release_path(path);
1118
1119 /* Same search but for extended refs */
1120 extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
1121 inode_objectid, parent_objectid, 0,
1122 0);
1123 if (!IS_ERR_OR_NULL(extref)) {
1124 u32 item_size;
1125 u32 cur_offset = 0;
1126 unsigned long base;
1127 struct inode *victim_parent;
1128
1129 leaf = path->nodes[0];
1130
1131 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1132 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
1133
1134 while (cur_offset < item_size) {
1135 extref = (struct btrfs_inode_extref *)(base + cur_offset);
1136
1137 victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
1138
1139 if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
1140 goto next;
1141
1142 victim_name = kmalloc(victim_name_len, GFP_NOFS);
1143 if (!victim_name)
1144 return -ENOMEM;
1145 read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
1146 victim_name_len);
1147
1148 search_key.objectid = inode_objectid;
1149 search_key.type = BTRFS_INODE_EXTREF_KEY;
1150 search_key.offset = btrfs_extref_hash(parent_objectid,
1151 victim_name,
1152 victim_name_len);
1153 ret = backref_in_log(log_root, &search_key,
1154 parent_objectid, victim_name,
1155 victim_name_len);
1156 if (ret < 0) {
1157 return ret;
1158 } else if (!ret) {
1159 ret = -ENOENT;
1160 victim_parent = read_one_inode(root,
1161 parent_objectid);
1162 if (victim_parent) {
1163 inc_nlink(&inode->vfs_inode);
1164 btrfs_release_path(path);
1165
1166 ret = btrfs_unlink_inode(trans, root,
1167 BTRFS_I(victim_parent),
1168 inode,
1169 victim_name,
1170 victim_name_len);
1171 if (!ret)
1172 ret = btrfs_run_delayed_items(
1173 trans);
1174 }
1175 iput(victim_parent);
1176 kfree(victim_name);
1177 if (ret)
1178 return ret;
1179 *search_done = 1;
1180 goto again;
1181 }
1182 kfree(victim_name);
1183next:
1184 cur_offset += victim_name_len + sizeof(*extref);
1185 }
1186 *search_done = 1;
1187 }
1188 btrfs_release_path(path);
1189
1190 /* look for a conflicting sequence number */
1191 di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
1192 ref_index, name, namelen, 0);
1193 if (IS_ERR(di)) {
1194 if (PTR_ERR(di) != -ENOENT)
1195 return PTR_ERR(di);
1196 } else if (di) {
1197 ret = drop_one_dir_item(trans, root, path, dir, di);
1198 if (ret)
1199 return ret;
1200 }
1201 btrfs_release_path(path);
1202
1203 /* look for a conflicting name */
1204 di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
1205 name, namelen, 0);
1206 if (IS_ERR(di)) {
1207 return PTR_ERR(di);
1208 } else if (di) {
1209 ret = drop_one_dir_item(trans, root, path, dir, di);
1210 if (ret)
1211 return ret;
1212 }
1213 btrfs_release_path(path);
1214
1215 return 0;
1216}
1217
1218static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1219 u32 *namelen, char **name, u64 *index,
1220 u64 *parent_objectid)
1221{
1222 struct btrfs_inode_extref *extref;
1223
1224 extref = (struct btrfs_inode_extref *)ref_ptr;
1225
1226 *namelen = btrfs_inode_extref_name_len(eb, extref);
1227 *name = kmalloc(*namelen, GFP_NOFS);
1228 if (*name == NULL)
1229 return -ENOMEM;
1230
1231 read_extent_buffer(eb, *name, (unsigned long)&extref->name,
1232 *namelen);
1233
1234 if (index)
1235 *index = btrfs_inode_extref_index(eb, extref);
1236 if (parent_objectid)
1237 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
1238
1239 return 0;
1240}
1241
1242static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1243 u32 *namelen, char **name, u64 *index)
1244{
1245 struct btrfs_inode_ref *ref;
1246
1247 ref = (struct btrfs_inode_ref *)ref_ptr;
1248
1249 *namelen = btrfs_inode_ref_name_len(eb, ref);
1250 *name = kmalloc(*namelen, GFP_NOFS);
1251 if (*name == NULL)
1252 return -ENOMEM;
1253
1254 read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1255
1256 if (index)
1257 *index = btrfs_inode_ref_index(eb, ref);
1258
1259 return 0;
1260}
1261
1262/*
1263 * Take an inode reference item from the log tree and iterate all names from the
1264 * inode reference item in the subvolume tree with the same key (if it exists).
1265 * For any name that is not in the inode reference item from the log tree, do a
1266 * proper unlink of that name (that is, remove its entry from the inode
1267 * reference item and both dir index keys).
1268 */
1269static int unlink_old_inode_refs(struct btrfs_trans_handle *trans,
1270 struct btrfs_root *root,
1271 struct btrfs_path *path,
1272 struct btrfs_inode *inode,
1273 struct extent_buffer *log_eb,
1274 int log_slot,
1275 struct btrfs_key *key)
1276{
1277 int ret;
1278 unsigned long ref_ptr;
1279 unsigned long ref_end;
1280 struct extent_buffer *eb;
1281
1282again:
1283 btrfs_release_path(path);
1284 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1285 if (ret > 0) {
1286 ret = 0;
1287 goto out;
1288 }
1289 if (ret < 0)
1290 goto out;
1291
1292 eb = path->nodes[0];
1293 ref_ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
1294 ref_end = ref_ptr + btrfs_item_size_nr(eb, path->slots[0]);
1295 while (ref_ptr < ref_end) {
1296 char *name = NULL;
1297 int namelen;
1298 u64 parent_id;
1299
1300 if (key->type == BTRFS_INODE_EXTREF_KEY) {
1301 ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1302 NULL, &parent_id);
1303 } else {
1304 parent_id = key->offset;
1305 ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1306 NULL);
1307 }
1308 if (ret)
1309 goto out;
1310
1311 if (key->type == BTRFS_INODE_EXTREF_KEY)
1312 ret = !!btrfs_find_name_in_ext_backref(log_eb, log_slot,
1313 parent_id, name,
1314 namelen);
1315 else
1316 ret = !!btrfs_find_name_in_backref(log_eb, log_slot,
1317 name, namelen);
1318
1319 if (!ret) {
1320 struct inode *dir;
1321
1322 btrfs_release_path(path);
1323 dir = read_one_inode(root, parent_id);
1324 if (!dir) {
1325 ret = -ENOENT;
1326 kfree(name);
1327 goto out;
1328 }
1329 ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
1330 inode, name, namelen);
1331 kfree(name);
1332 iput(dir);
1333 if (ret)
1334 goto out;
1335 goto again;
1336 }
1337
1338 kfree(name);
1339 ref_ptr += namelen;
1340 if (key->type == BTRFS_INODE_EXTREF_KEY)
1341 ref_ptr += sizeof(struct btrfs_inode_extref);
1342 else
1343 ref_ptr += sizeof(struct btrfs_inode_ref);
1344 }
1345 ret = 0;
1346 out:
1347 btrfs_release_path(path);
1348 return ret;
1349}
1350
1351static int btrfs_inode_ref_exists(struct inode *inode, struct inode *dir,
1352 const u8 ref_type, const char *name,
1353 const int namelen)
1354{
1355 struct btrfs_key key;
1356 struct btrfs_path *path;
1357 const u64 parent_id = btrfs_ino(BTRFS_I(dir));
1358 int ret;
1359
1360 path = btrfs_alloc_path();
1361 if (!path)
1362 return -ENOMEM;
1363
1364 key.objectid = btrfs_ino(BTRFS_I(inode));
1365 key.type = ref_type;
1366 if (key.type == BTRFS_INODE_REF_KEY)
1367 key.offset = parent_id;
1368 else
1369 key.offset = btrfs_extref_hash(parent_id, name, namelen);
1370
1371 ret = btrfs_search_slot(NULL, BTRFS_I(inode)->root, &key, path, 0, 0);
1372 if (ret < 0)
1373 goto out;
1374 if (ret > 0) {
1375 ret = 0;
1376 goto out;
1377 }
1378 if (key.type == BTRFS_INODE_EXTREF_KEY)
1379 ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
1380 path->slots[0], parent_id, name, namelen);
1381 else
1382 ret = !!btrfs_find_name_in_backref(path->nodes[0], path->slots[0],
1383 name, namelen);
1384
1385out:
1386 btrfs_free_path(path);
1387 return ret;
1388}
1389
1390static int add_link(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1391 struct inode *dir, struct inode *inode, const char *name,
1392 int namelen, u64 ref_index)
1393{
1394 struct btrfs_dir_item *dir_item;
1395 struct btrfs_key key;
1396 struct btrfs_path *path;
1397 struct inode *other_inode = NULL;
1398 int ret;
1399
1400 path = btrfs_alloc_path();
1401 if (!path)
1402 return -ENOMEM;
1403
1404 dir_item = btrfs_lookup_dir_item(NULL, root, path,
1405 btrfs_ino(BTRFS_I(dir)),
1406 name, namelen, 0);
1407 if (!dir_item) {
1408 btrfs_release_path(path);
1409 goto add_link;
1410 } else if (IS_ERR(dir_item)) {
1411 ret = PTR_ERR(dir_item);
1412 goto out;
1413 }
1414
1415 /*
1416 * Our inode's dentry collides with the dentry of another inode which is
1417 * in the log but not yet processed since it has a higher inode number.
1418 * So delete that other dentry.
1419 */
1420 btrfs_dir_item_key_to_cpu(path->nodes[0], dir_item, &key);
1421 btrfs_release_path(path);
1422 other_inode = read_one_inode(root, key.objectid);
1423 if (!other_inode) {
1424 ret = -ENOENT;
1425 goto out;
1426 }
1427 ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir), BTRFS_I(other_inode),
1428 name, namelen);
1429 if (ret)
1430 goto out;
1431 /*
1432 * If we dropped the link count to 0, bump it so that later the iput()
1433 * on the inode will not free it. We will fixup the link count later.
1434 */
1435 if (other_inode->i_nlink == 0)
1436 inc_nlink(other_inode);
1437
1438 ret = btrfs_run_delayed_items(trans);
1439 if (ret)
1440 goto out;
1441add_link:
1442 ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
1443 name, namelen, 0, ref_index);
1444out:
1445 iput(other_inode);
1446 btrfs_free_path(path);
1447
1448 return ret;
1449}
1450
1451/*
1452 * replay one inode back reference item found in the log tree.
1453 * eb, slot and key refer to the buffer and key found in the log tree.
1454 * root is the destination we are replaying into, and path is for temp
1455 * use by this function. (it should be released on return).
1456 */
1457static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1458 struct btrfs_root *root,
1459 struct btrfs_root *log,
1460 struct btrfs_path *path,
1461 struct extent_buffer *eb, int slot,
1462 struct btrfs_key *key)
1463{
1464 struct inode *dir = NULL;
1465 struct inode *inode = NULL;
1466 unsigned long ref_ptr;
1467 unsigned long ref_end;
1468 char *name = NULL;
1469 int namelen;
1470 int ret;
1471 int search_done = 0;
1472 int log_ref_ver = 0;
1473 u64 parent_objectid;
1474 u64 inode_objectid;
1475 u64 ref_index = 0;
1476 int ref_struct_size;
1477
1478 ref_ptr = btrfs_item_ptr_offset(eb, slot);
1479 ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1480
1481 if (key->type == BTRFS_INODE_EXTREF_KEY) {
1482 struct btrfs_inode_extref *r;
1483
1484 ref_struct_size = sizeof(struct btrfs_inode_extref);
1485 log_ref_ver = 1;
1486 r = (struct btrfs_inode_extref *)ref_ptr;
1487 parent_objectid = btrfs_inode_extref_parent(eb, r);
1488 } else {
1489 ref_struct_size = sizeof(struct btrfs_inode_ref);
1490 parent_objectid = key->offset;
1491 }
1492 inode_objectid = key->objectid;
1493
1494 /*
1495 * it is possible that we didn't log all the parent directories
1496 * for a given inode. If we don't find the dir, just don't
1497 * copy the back ref in. The link count fixup code will take
1498 * care of the rest
1499 */
1500 dir = read_one_inode(root, parent_objectid);
1501 if (!dir) {
1502 ret = -ENOENT;
1503 goto out;
1504 }
1505
1506 inode = read_one_inode(root, inode_objectid);
1507 if (!inode) {
1508 ret = -EIO;
1509 goto out;
1510 }
1511
1512 while (ref_ptr < ref_end) {
1513 if (log_ref_ver) {
1514 ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1515 &ref_index, &parent_objectid);
1516 /*
1517 * parent object can change from one array
1518 * item to another.
1519 */
1520 if (!dir)
1521 dir = read_one_inode(root, parent_objectid);
1522 if (!dir) {
1523 ret = -ENOENT;
1524 goto out;
1525 }
1526 } else {
1527 ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1528 &ref_index);
1529 }
1530 if (ret)
1531 goto out;
1532
1533 ret = inode_in_dir(root, path, btrfs_ino(BTRFS_I(dir)),
1534 btrfs_ino(BTRFS_I(inode)), ref_index,
1535 name, namelen);
1536 if (ret < 0) {
1537 goto out;
1538 } else if (ret == 0) {
1539 /*
1540 * look for a conflicting back reference in the
1541 * metadata. if we find one we have to unlink that name
1542 * of the file before we add our new link. Later on, we
1543 * overwrite any existing back reference, and we don't
1544 * want to create dangling pointers in the directory.
1545 */
1546
1547 if (!search_done) {
1548 ret = __add_inode_ref(trans, root, path, log,
1549 BTRFS_I(dir),
1550 BTRFS_I(inode),
1551 inode_objectid,
1552 parent_objectid,
1553 ref_index, name, namelen,
1554 &search_done);
1555 if (ret) {
1556 if (ret == 1)
1557 ret = 0;
1558 goto out;
1559 }
1560 }
1561
1562 /*
1563 * If a reference item already exists for this inode
1564 * with the same parent and name, but different index,
1565 * drop it and the corresponding directory index entries
1566 * from the parent before adding the new reference item
1567 * and dir index entries, otherwise we would fail with
1568 * -EEXIST returned from btrfs_add_link() below.
1569 */
1570 ret = btrfs_inode_ref_exists(inode, dir, key->type,
1571 name, namelen);
1572 if (ret > 0) {
1573 ret = btrfs_unlink_inode(trans, root,
1574 BTRFS_I(dir),
1575 BTRFS_I(inode),
1576 name, namelen);
1577 /*
1578 * If we dropped the link count to 0, bump it so
1579 * that later the iput() on the inode will not
1580 * free it. We will fixup the link count later.
1581 */
1582 if (!ret && inode->i_nlink == 0)
1583 inc_nlink(inode);
1584 }
1585 if (ret < 0)
1586 goto out;
1587
1588 /* insert our name */
1589 ret = add_link(trans, root, dir, inode, name, namelen,
1590 ref_index);
1591 if (ret)
1592 goto out;
1593
1594 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
1595 if (ret)
1596 goto out;
1597 }
1598 /* Else, ret == 1, we already have a perfect match, we're done. */
1599
1600 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1601 kfree(name);
1602 name = NULL;
1603 if (log_ref_ver) {
1604 iput(dir);
1605 dir = NULL;
1606 }
1607 }
1608
1609 /*
1610 * Before we overwrite the inode reference item in the subvolume tree
1611 * with the item from the log tree, we must unlink all names from the
1612 * parent directory that are in the subvolume's tree inode reference
1613 * item, otherwise we end up with an inconsistent subvolume tree where
1614 * dir index entries exist for a name but there is no inode reference
1615 * item with the same name.
1616 */
1617 ret = unlink_old_inode_refs(trans, root, path, BTRFS_I(inode), eb, slot,
1618 key);
1619 if (ret)
1620 goto out;
1621
1622 /* finally write the back reference in the inode */
1623 ret = overwrite_item(trans, root, path, eb, slot, key);
1624out:
1625 btrfs_release_path(path);
1626 kfree(name);
1627 iput(dir);
1628 iput(inode);
1629 return ret;
1630}
1631
1632static int count_inode_extrefs(struct btrfs_root *root,
1633 struct btrfs_inode *inode, struct btrfs_path *path)
1634{
1635 int ret = 0;
1636 int name_len;
1637 unsigned int nlink = 0;
1638 u32 item_size;
1639 u32 cur_offset = 0;
1640 u64 inode_objectid = btrfs_ino(inode);
1641 u64 offset = 0;
1642 unsigned long ptr;
1643 struct btrfs_inode_extref *extref;
1644 struct extent_buffer *leaf;
1645
1646 while (1) {
1647 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1648 &extref, &offset);
1649 if (ret)
1650 break;
1651
1652 leaf = path->nodes[0];
1653 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1654 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1655 cur_offset = 0;
1656
1657 while (cur_offset < item_size) {
1658 extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1659 name_len = btrfs_inode_extref_name_len(leaf, extref);
1660
1661 nlink++;
1662
1663 cur_offset += name_len + sizeof(*extref);
1664 }
1665
1666 offset++;
1667 btrfs_release_path(path);
1668 }
1669 btrfs_release_path(path);
1670
1671 if (ret < 0 && ret != -ENOENT)
1672 return ret;
1673 return nlink;
1674}
1675
1676static int count_inode_refs(struct btrfs_root *root,
1677 struct btrfs_inode *inode, struct btrfs_path *path)
1678{
1679 int ret;
1680 struct btrfs_key key;
1681 unsigned int nlink = 0;
1682 unsigned long ptr;
1683 unsigned long ptr_end;
1684 int name_len;
1685 u64 ino = btrfs_ino(inode);
1686
1687 key.objectid = ino;
1688 key.type = BTRFS_INODE_REF_KEY;
1689 key.offset = (u64)-1;
1690
1691 while (1) {
1692 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1693 if (ret < 0)
1694 break;
1695 if (ret > 0) {
1696 if (path->slots[0] == 0)
1697 break;
1698 path->slots[0]--;
1699 }
1700process_slot:
1701 btrfs_item_key_to_cpu(path->nodes[0], &key,
1702 path->slots[0]);
1703 if (key.objectid != ino ||
1704 key.type != BTRFS_INODE_REF_KEY)
1705 break;
1706 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1707 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1708 path->slots[0]);
1709 while (ptr < ptr_end) {
1710 struct btrfs_inode_ref *ref;
1711
1712 ref = (struct btrfs_inode_ref *)ptr;
1713 name_len = btrfs_inode_ref_name_len(path->nodes[0],
1714 ref);
1715 ptr = (unsigned long)(ref + 1) + name_len;
1716 nlink++;
1717 }
1718
1719 if (key.offset == 0)
1720 break;
1721 if (path->slots[0] > 0) {
1722 path->slots[0]--;
1723 goto process_slot;
1724 }
1725 key.offset--;
1726 btrfs_release_path(path);
1727 }
1728 btrfs_release_path(path);
1729
1730 return nlink;
1731}
1732
1733/*
1734 * There are a few corners where the link count of the file can't
1735 * be properly maintained during replay. So, instead of adding
1736 * lots of complexity to the log code, we just scan the backrefs
1737 * for any file that has been through replay.
1738 *
1739 * The scan will update the link count on the inode to reflect the
1740 * number of back refs found. If it goes down to zero, the iput
1741 * will free the inode.
1742 */
1743static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1744 struct btrfs_root *root,
1745 struct inode *inode)
1746{
1747 struct btrfs_path *path;
1748 int ret;
1749 u64 nlink = 0;
1750 u64 ino = btrfs_ino(BTRFS_I(inode));
1751
1752 path = btrfs_alloc_path();
1753 if (!path)
1754 return -ENOMEM;
1755
1756 ret = count_inode_refs(root, BTRFS_I(inode), path);
1757 if (ret < 0)
1758 goto out;
1759
1760 nlink = ret;
1761
1762 ret = count_inode_extrefs(root, BTRFS_I(inode), path);
1763 if (ret < 0)
1764 goto out;
1765
1766 nlink += ret;
1767
1768 ret = 0;
1769
1770 if (nlink != inode->i_nlink) {
1771 set_nlink(inode, nlink);
1772 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
1773 if (ret)
1774 goto out;
1775 }
1776 BTRFS_I(inode)->index_cnt = (u64)-1;
1777
1778 if (inode->i_nlink == 0) {
1779 if (S_ISDIR(inode->i_mode)) {
1780 ret = replay_dir_deletes(trans, root, NULL, path,
1781 ino, 1);
1782 if (ret)
1783 goto out;
1784 }
1785 ret = btrfs_insert_orphan_item(trans, root, ino);
1786 if (ret == -EEXIST)
1787 ret = 0;
1788 }
1789
1790out:
1791 btrfs_free_path(path);
1792 return ret;
1793}
1794
1795static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1796 struct btrfs_root *root,
1797 struct btrfs_path *path)
1798{
1799 int ret;
1800 struct btrfs_key key;
1801 struct inode *inode;
1802
1803 key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1804 key.type = BTRFS_ORPHAN_ITEM_KEY;
1805 key.offset = (u64)-1;
1806 while (1) {
1807 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1808 if (ret < 0)
1809 break;
1810
1811 if (ret == 1) {
1812 ret = 0;
1813 if (path->slots[0] == 0)
1814 break;
1815 path->slots[0]--;
1816 }
1817
1818 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1819 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1820 key.type != BTRFS_ORPHAN_ITEM_KEY)
1821 break;
1822
1823 ret = btrfs_del_item(trans, root, path);
1824 if (ret)
1825 break;
1826
1827 btrfs_release_path(path);
1828 inode = read_one_inode(root, key.offset);
1829 if (!inode) {
1830 ret = -EIO;
1831 break;
1832 }
1833
1834 ret = fixup_inode_link_count(trans, root, inode);
1835 iput(inode);
1836 if (ret)
1837 break;
1838
1839 /*
1840 * fixup on a directory may create new entries,
1841 * make sure we always look for the highset possible
1842 * offset
1843 */
1844 key.offset = (u64)-1;
1845 }
1846 btrfs_release_path(path);
1847 return ret;
1848}
1849
1850
1851/*
1852 * record a given inode in the fixup dir so we can check its link
1853 * count when replay is done. The link count is incremented here
1854 * so the inode won't go away until we check it
1855 */
1856static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1857 struct btrfs_root *root,
1858 struct btrfs_path *path,
1859 u64 objectid)
1860{
1861 struct btrfs_key key;
1862 int ret = 0;
1863 struct inode *inode;
1864
1865 inode = read_one_inode(root, objectid);
1866 if (!inode)
1867 return -EIO;
1868
1869 key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1870 key.type = BTRFS_ORPHAN_ITEM_KEY;
1871 key.offset = objectid;
1872
1873 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1874
1875 btrfs_release_path(path);
1876 if (ret == 0) {
1877 if (!inode->i_nlink)
1878 set_nlink(inode, 1);
1879 else
1880 inc_nlink(inode);
1881 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
1882 } else if (ret == -EEXIST) {
1883 ret = 0;
1884 }
1885 iput(inode);
1886
1887 return ret;
1888}
1889
1890/*
1891 * when replaying the log for a directory, we only insert names
1892 * for inodes that actually exist. This means an fsync on a directory
1893 * does not implicitly fsync all the new files in it
1894 */
1895static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1896 struct btrfs_root *root,
1897 u64 dirid, u64 index,
1898 char *name, int name_len,
1899 struct btrfs_key *location)
1900{
1901 struct inode *inode;
1902 struct inode *dir;
1903 int ret;
1904
1905 inode = read_one_inode(root, location->objectid);
1906 if (!inode)
1907 return -ENOENT;
1908
1909 dir = read_one_inode(root, dirid);
1910 if (!dir) {
1911 iput(inode);
1912 return -EIO;
1913 }
1914
1915 ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode), name,
1916 name_len, 1, index);
1917
1918 /* FIXME, put inode into FIXUP list */
1919
1920 iput(inode);
1921 iput(dir);
1922 return ret;
1923}
1924
1925/*
1926 * take a single entry in a log directory item and replay it into
1927 * the subvolume.
1928 *
1929 * if a conflicting item exists in the subdirectory already,
1930 * the inode it points to is unlinked and put into the link count
1931 * fix up tree.
1932 *
1933 * If a name from the log points to a file or directory that does
1934 * not exist in the FS, it is skipped. fsyncs on directories
1935 * do not force down inodes inside that directory, just changes to the
1936 * names or unlinks in a directory.
1937 *
1938 * Returns < 0 on error, 0 if the name wasn't replayed (dentry points to a
1939 * non-existing inode) and 1 if the name was replayed.
1940 */
1941static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1942 struct btrfs_root *root,
1943 struct btrfs_path *path,
1944 struct extent_buffer *eb,
1945 struct btrfs_dir_item *di,
1946 struct btrfs_key *key)
1947{
1948 char *name;
1949 int name_len;
1950 struct btrfs_dir_item *dst_di;
1951 struct btrfs_key found_key;
1952 struct btrfs_key log_key;
1953 struct inode *dir;
1954 u8 log_type;
1955 bool exists;
1956 int ret;
1957 bool update_size = (key->type == BTRFS_DIR_INDEX_KEY);
1958 bool name_added = false;
1959
1960 dir = read_one_inode(root, key->objectid);
1961 if (!dir)
1962 return -EIO;
1963
1964 name_len = btrfs_dir_name_len(eb, di);
1965 name = kmalloc(name_len, GFP_NOFS);
1966 if (!name) {
1967 ret = -ENOMEM;
1968 goto out;
1969 }
1970
1971 log_type = btrfs_dir_type(eb, di);
1972 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1973 name_len);
1974
1975 btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1976 ret = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1977 btrfs_release_path(path);
1978 if (ret < 0)
1979 goto out;
1980 exists = (ret == 0);
1981 ret = 0;
1982
1983 if (key->type == BTRFS_DIR_ITEM_KEY) {
1984 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1985 name, name_len, 1);
1986 } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1987 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1988 key->objectid,
1989 key->offset, name,
1990 name_len, 1);
1991 } else {
1992 /* Corruption */
1993 ret = -EINVAL;
1994 goto out;
1995 }
1996
1997 if (dst_di == ERR_PTR(-ENOENT))
1998 dst_di = NULL;
1999
2000 if (IS_ERR(dst_di)) {
2001 ret = PTR_ERR(dst_di);
2002 goto out;
2003 } else if (!dst_di) {
2004 /* we need a sequence number to insert, so we only
2005 * do inserts for the BTRFS_DIR_INDEX_KEY types
2006 */
2007 if (key->type != BTRFS_DIR_INDEX_KEY)
2008 goto out;
2009 goto insert;
2010 }
2011
2012 btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
2013 /* the existing item matches the logged item */
2014 if (found_key.objectid == log_key.objectid &&
2015 found_key.type == log_key.type &&
2016 found_key.offset == log_key.offset &&
2017 btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
2018 update_size = false;
2019 goto out;
2020 }
2021
2022 /*
2023 * don't drop the conflicting directory entry if the inode
2024 * for the new entry doesn't exist
2025 */
2026 if (!exists)
2027 goto out;
2028
2029 ret = drop_one_dir_item(trans, root, path, BTRFS_I(dir), dst_di);
2030 if (ret)
2031 goto out;
2032
2033 if (key->type == BTRFS_DIR_INDEX_KEY)
2034 goto insert;
2035out:
2036 btrfs_release_path(path);
2037 if (!ret && update_size) {
2038 btrfs_i_size_write(BTRFS_I(dir), dir->i_size + name_len * 2);
2039 ret = btrfs_update_inode(trans, root, BTRFS_I(dir));
2040 }
2041 kfree(name);
2042 iput(dir);
2043 if (!ret && name_added)
2044 ret = 1;
2045 return ret;
2046
2047insert:
2048 /*
2049 * Check if the inode reference exists in the log for the given name,
2050 * inode and parent inode
2051 */
2052 found_key.objectid = log_key.objectid;
2053 found_key.type = BTRFS_INODE_REF_KEY;
2054 found_key.offset = key->objectid;
2055 ret = backref_in_log(root->log_root, &found_key, 0, name, name_len);
2056 if (ret < 0) {
2057 goto out;
2058 } else if (ret) {
2059 /* The dentry will be added later. */
2060 ret = 0;
2061 update_size = false;
2062 goto out;
2063 }
2064
2065 found_key.objectid = log_key.objectid;
2066 found_key.type = BTRFS_INODE_EXTREF_KEY;
2067 found_key.offset = key->objectid;
2068 ret = backref_in_log(root->log_root, &found_key, key->objectid, name,
2069 name_len);
2070 if (ret < 0) {
2071 goto out;
2072 } else if (ret) {
2073 /* The dentry will be added later. */
2074 ret = 0;
2075 update_size = false;
2076 goto out;
2077 }
2078 btrfs_release_path(path);
2079 ret = insert_one_name(trans, root, key->objectid, key->offset,
2080 name, name_len, &log_key);
2081 if (ret && ret != -ENOENT && ret != -EEXIST)
2082 goto out;
2083 if (!ret)
2084 name_added = true;
2085 update_size = false;
2086 ret = 0;
2087 goto out;
2088}
2089
2090/*
2091 * find all the names in a directory item and reconcile them into
2092 * the subvolume. Only BTRFS_DIR_ITEM_KEY types will have more than
2093 * one name in a directory item, but the same code gets used for
2094 * both directory index types
2095 */
2096static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
2097 struct btrfs_root *root,
2098 struct btrfs_path *path,
2099 struct extent_buffer *eb, int slot,
2100 struct btrfs_key *key)
2101{
2102 int ret = 0;
2103 u32 item_size = btrfs_item_size_nr(eb, slot);
2104 struct btrfs_dir_item *di;
2105 int name_len;
2106 unsigned long ptr;
2107 unsigned long ptr_end;
2108 struct btrfs_path *fixup_path = NULL;
2109
2110 ptr = btrfs_item_ptr_offset(eb, slot);
2111 ptr_end = ptr + item_size;
2112 while (ptr < ptr_end) {
2113 di = (struct btrfs_dir_item *)ptr;
2114 name_len = btrfs_dir_name_len(eb, di);
2115 ret = replay_one_name(trans, root, path, eb, di, key);
2116 if (ret < 0)
2117 break;
2118 ptr = (unsigned long)(di + 1);
2119 ptr += name_len;
2120
2121 /*
2122 * If this entry refers to a non-directory (directories can not
2123 * have a link count > 1) and it was added in the transaction
2124 * that was not committed, make sure we fixup the link count of
2125 * the inode it the entry points to. Otherwise something like
2126 * the following would result in a directory pointing to an
2127 * inode with a wrong link that does not account for this dir
2128 * entry:
2129 *
2130 * mkdir testdir
2131 * touch testdir/foo
2132 * touch testdir/bar
2133 * sync
2134 *
2135 * ln testdir/bar testdir/bar_link
2136 * ln testdir/foo testdir/foo_link
2137 * xfs_io -c "fsync" testdir/bar
2138 *
2139 * <power failure>
2140 *
2141 * mount fs, log replay happens
2142 *
2143 * File foo would remain with a link count of 1 when it has two
2144 * entries pointing to it in the directory testdir. This would
2145 * make it impossible to ever delete the parent directory has
2146 * it would result in stale dentries that can never be deleted.
2147 */
2148 if (ret == 1 && btrfs_dir_type(eb, di) != BTRFS_FT_DIR) {
2149 struct btrfs_key di_key;
2150
2151 if (!fixup_path) {
2152 fixup_path = btrfs_alloc_path();
2153 if (!fixup_path) {
2154 ret = -ENOMEM;
2155 break;
2156 }
2157 }
2158
2159 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2160 ret = link_to_fixup_dir(trans, root, fixup_path,
2161 di_key.objectid);
2162 if (ret)
2163 break;
2164 }
2165 ret = 0;
2166 }
2167 btrfs_free_path(fixup_path);
2168 return ret;
2169}
2170
2171/*
2172 * directory replay has two parts. There are the standard directory
2173 * items in the log copied from the subvolume, and range items
2174 * created in the log while the subvolume was logged.
2175 *
2176 * The range items tell us which parts of the key space the log
2177 * is authoritative for. During replay, if a key in the subvolume
2178 * directory is in a logged range item, but not actually in the log
2179 * that means it was deleted from the directory before the fsync
2180 * and should be removed.
2181 */
2182static noinline int find_dir_range(struct btrfs_root *root,
2183 struct btrfs_path *path,
2184 u64 dirid, int key_type,
2185 u64 *start_ret, u64 *end_ret)
2186{
2187 struct btrfs_key key;
2188 u64 found_end;
2189 struct btrfs_dir_log_item *item;
2190 int ret;
2191 int nritems;
2192
2193 if (*start_ret == (u64)-1)
2194 return 1;
2195
2196 key.objectid = dirid;
2197 key.type = key_type;
2198 key.offset = *start_ret;
2199
2200 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2201 if (ret < 0)
2202 goto out;
2203 if (ret > 0) {
2204 if (path->slots[0] == 0)
2205 goto out;
2206 path->slots[0]--;
2207 }
2208 if (ret != 0)
2209 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2210
2211 if (key.type != key_type || key.objectid != dirid) {
2212 ret = 1;
2213 goto next;
2214 }
2215 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2216 struct btrfs_dir_log_item);
2217 found_end = btrfs_dir_log_end(path->nodes[0], item);
2218
2219 if (*start_ret >= key.offset && *start_ret <= found_end) {
2220 ret = 0;
2221 *start_ret = key.offset;
2222 *end_ret = found_end;
2223 goto out;
2224 }
2225 ret = 1;
2226next:
2227 /* check the next slot in the tree to see if it is a valid item */
2228 nritems = btrfs_header_nritems(path->nodes[0]);
2229 path->slots[0]++;
2230 if (path->slots[0] >= nritems) {
2231 ret = btrfs_next_leaf(root, path);
2232 if (ret)
2233 goto out;
2234 }
2235
2236 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2237
2238 if (key.type != key_type || key.objectid != dirid) {
2239 ret = 1;
2240 goto out;
2241 }
2242 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2243 struct btrfs_dir_log_item);
2244 found_end = btrfs_dir_log_end(path->nodes[0], item);
2245 *start_ret = key.offset;
2246 *end_ret = found_end;
2247 ret = 0;
2248out:
2249 btrfs_release_path(path);
2250 return ret;
2251}
2252
2253/*
2254 * this looks for a given directory item in the log. If the directory
2255 * item is not in the log, the item is removed and the inode it points
2256 * to is unlinked
2257 */
2258static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
2259 struct btrfs_root *root,
2260 struct btrfs_root *log,
2261 struct btrfs_path *path,
2262 struct btrfs_path *log_path,
2263 struct inode *dir,
2264 struct btrfs_key *dir_key)
2265{
2266 int ret;
2267 struct extent_buffer *eb;
2268 int slot;
2269 u32 item_size;
2270 struct btrfs_dir_item *di;
2271 struct btrfs_dir_item *log_di;
2272 int name_len;
2273 unsigned long ptr;
2274 unsigned long ptr_end;
2275 char *name;
2276 struct inode *inode;
2277 struct btrfs_key location;
2278
2279again:
2280 eb = path->nodes[0];
2281 slot = path->slots[0];
2282 item_size = btrfs_item_size_nr(eb, slot);
2283 ptr = btrfs_item_ptr_offset(eb, slot);
2284 ptr_end = ptr + item_size;
2285 while (ptr < ptr_end) {
2286 di = (struct btrfs_dir_item *)ptr;
2287 name_len = btrfs_dir_name_len(eb, di);
2288 name = kmalloc(name_len, GFP_NOFS);
2289 if (!name) {
2290 ret = -ENOMEM;
2291 goto out;
2292 }
2293 read_extent_buffer(eb, name, (unsigned long)(di + 1),
2294 name_len);
2295 log_di = NULL;
2296 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
2297 log_di = btrfs_lookup_dir_item(trans, log, log_path,
2298 dir_key->objectid,
2299 name, name_len, 0);
2300 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
2301 log_di = btrfs_lookup_dir_index_item(trans, log,
2302 log_path,
2303 dir_key->objectid,
2304 dir_key->offset,
2305 name, name_len, 0);
2306 }
2307 if (!log_di || log_di == ERR_PTR(-ENOENT)) {
2308 btrfs_dir_item_key_to_cpu(eb, di, &location);
2309 btrfs_release_path(path);
2310 btrfs_release_path(log_path);
2311 inode = read_one_inode(root, location.objectid);
2312 if (!inode) {
2313 kfree(name);
2314 return -EIO;
2315 }
2316
2317 ret = link_to_fixup_dir(trans, root,
2318 path, location.objectid);
2319 if (ret) {
2320 kfree(name);
2321 iput(inode);
2322 goto out;
2323 }
2324
2325 inc_nlink(inode);
2326 ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
2327 BTRFS_I(inode), name, name_len);
2328 if (!ret)
2329 ret = btrfs_run_delayed_items(trans);
2330 kfree(name);
2331 iput(inode);
2332 if (ret)
2333 goto out;
2334
2335 /* there might still be more names under this key
2336 * check and repeat if required
2337 */
2338 ret = btrfs_search_slot(NULL, root, dir_key, path,
2339 0, 0);
2340 if (ret == 0)
2341 goto again;
2342 ret = 0;
2343 goto out;
2344 } else if (IS_ERR(log_di)) {
2345 kfree(name);
2346 return PTR_ERR(log_di);
2347 }
2348 btrfs_release_path(log_path);
2349 kfree(name);
2350
2351 ptr = (unsigned long)(di + 1);
2352 ptr += name_len;
2353 }
2354 ret = 0;
2355out:
2356 btrfs_release_path(path);
2357 btrfs_release_path(log_path);
2358 return ret;
2359}
2360
2361static int replay_xattr_deletes(struct btrfs_trans_handle *trans,
2362 struct btrfs_root *root,
2363 struct btrfs_root *log,
2364 struct btrfs_path *path,
2365 const u64 ino)
2366{
2367 struct btrfs_key search_key;
2368 struct btrfs_path *log_path;
2369 int i;
2370 int nritems;
2371 int ret;
2372
2373 log_path = btrfs_alloc_path();
2374 if (!log_path)
2375 return -ENOMEM;
2376
2377 search_key.objectid = ino;
2378 search_key.type = BTRFS_XATTR_ITEM_KEY;
2379 search_key.offset = 0;
2380again:
2381 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
2382 if (ret < 0)
2383 goto out;
2384process_leaf:
2385 nritems = btrfs_header_nritems(path->nodes[0]);
2386 for (i = path->slots[0]; i < nritems; i++) {
2387 struct btrfs_key key;
2388 struct btrfs_dir_item *di;
2389 struct btrfs_dir_item *log_di;
2390 u32 total_size;
2391 u32 cur;
2392
2393 btrfs_item_key_to_cpu(path->nodes[0], &key, i);
2394 if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY) {
2395 ret = 0;
2396 goto out;
2397 }
2398
2399 di = btrfs_item_ptr(path->nodes[0], i, struct btrfs_dir_item);
2400 total_size = btrfs_item_size_nr(path->nodes[0], i);
2401 cur = 0;
2402 while (cur < total_size) {
2403 u16 name_len = btrfs_dir_name_len(path->nodes[0], di);
2404 u16 data_len = btrfs_dir_data_len(path->nodes[0], di);
2405 u32 this_len = sizeof(*di) + name_len + data_len;
2406 char *name;
2407
2408 name = kmalloc(name_len, GFP_NOFS);
2409 if (!name) {
2410 ret = -ENOMEM;
2411 goto out;
2412 }
2413 read_extent_buffer(path->nodes[0], name,
2414 (unsigned long)(di + 1), name_len);
2415
2416 log_di = btrfs_lookup_xattr(NULL, log, log_path, ino,
2417 name, name_len, 0);
2418 btrfs_release_path(log_path);
2419 if (!log_di) {
2420 /* Doesn't exist in log tree, so delete it. */
2421 btrfs_release_path(path);
2422 di = btrfs_lookup_xattr(trans, root, path, ino,
2423 name, name_len, -1);
2424 kfree(name);
2425 if (IS_ERR(di)) {
2426 ret = PTR_ERR(di);
2427 goto out;
2428 }
2429 ASSERT(di);
2430 ret = btrfs_delete_one_dir_name(trans, root,
2431 path, di);
2432 if (ret)
2433 goto out;
2434 btrfs_release_path(path);
2435 search_key = key;
2436 goto again;
2437 }
2438 kfree(name);
2439 if (IS_ERR(log_di)) {
2440 ret = PTR_ERR(log_di);
2441 goto out;
2442 }
2443 cur += this_len;
2444 di = (struct btrfs_dir_item *)((char *)di + this_len);
2445 }
2446 }
2447 ret = btrfs_next_leaf(root, path);
2448 if (ret > 0)
2449 ret = 0;
2450 else if (ret == 0)
2451 goto process_leaf;
2452out:
2453 btrfs_free_path(log_path);
2454 btrfs_release_path(path);
2455 return ret;
2456}
2457
2458
2459/*
2460 * deletion replay happens before we copy any new directory items
2461 * out of the log or out of backreferences from inodes. It
2462 * scans the log to find ranges of keys that log is authoritative for,
2463 * and then scans the directory to find items in those ranges that are
2464 * not present in the log.
2465 *
2466 * Anything we don't find in the log is unlinked and removed from the
2467 * directory.
2468 */
2469static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
2470 struct btrfs_root *root,
2471 struct btrfs_root *log,
2472 struct btrfs_path *path,
2473 u64 dirid, int del_all)
2474{
2475 u64 range_start;
2476 u64 range_end;
2477 int key_type = BTRFS_DIR_LOG_ITEM_KEY;
2478 int ret = 0;
2479 struct btrfs_key dir_key;
2480 struct btrfs_key found_key;
2481 struct btrfs_path *log_path;
2482 struct inode *dir;
2483
2484 dir_key.objectid = dirid;
2485 dir_key.type = BTRFS_DIR_ITEM_KEY;
2486 log_path = btrfs_alloc_path();
2487 if (!log_path)
2488 return -ENOMEM;
2489
2490 dir = read_one_inode(root, dirid);
2491 /* it isn't an error if the inode isn't there, that can happen
2492 * because we replay the deletes before we copy in the inode item
2493 * from the log
2494 */
2495 if (!dir) {
2496 btrfs_free_path(log_path);
2497 return 0;
2498 }
2499again:
2500 range_start = 0;
2501 range_end = 0;
2502 while (1) {
2503 if (del_all)
2504 range_end = (u64)-1;
2505 else {
2506 ret = find_dir_range(log, path, dirid, key_type,
2507 &range_start, &range_end);
2508 if (ret != 0)
2509 break;
2510 }
2511
2512 dir_key.offset = range_start;
2513 while (1) {
2514 int nritems;
2515 ret = btrfs_search_slot(NULL, root, &dir_key, path,
2516 0, 0);
2517 if (ret < 0)
2518 goto out;
2519
2520 nritems = btrfs_header_nritems(path->nodes[0]);
2521 if (path->slots[0] >= nritems) {
2522 ret = btrfs_next_leaf(root, path);
2523 if (ret == 1)
2524 break;
2525 else if (ret < 0)
2526 goto out;
2527 }
2528 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2529 path->slots[0]);
2530 if (found_key.objectid != dirid ||
2531 found_key.type != dir_key.type)
2532 goto next_type;
2533
2534 if (found_key.offset > range_end)
2535 break;
2536
2537 ret = check_item_in_log(trans, root, log, path,
2538 log_path, dir,
2539 &found_key);
2540 if (ret)
2541 goto out;
2542 if (found_key.offset == (u64)-1)
2543 break;
2544 dir_key.offset = found_key.offset + 1;
2545 }
2546 btrfs_release_path(path);
2547 if (range_end == (u64)-1)
2548 break;
2549 range_start = range_end + 1;
2550 }
2551
2552next_type:
2553 ret = 0;
2554 if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
2555 key_type = BTRFS_DIR_LOG_INDEX_KEY;
2556 dir_key.type = BTRFS_DIR_INDEX_KEY;
2557 btrfs_release_path(path);
2558 goto again;
2559 }
2560out:
2561 btrfs_release_path(path);
2562 btrfs_free_path(log_path);
2563 iput(dir);
2564 return ret;
2565}
2566
2567/*
2568 * the process_func used to replay items from the log tree. This
2569 * gets called in two different stages. The first stage just looks
2570 * for inodes and makes sure they are all copied into the subvolume.
2571 *
2572 * The second stage copies all the other item types from the log into
2573 * the subvolume. The two stage approach is slower, but gets rid of
2574 * lots of complexity around inodes referencing other inodes that exist
2575 * only in the log (references come from either directory items or inode
2576 * back refs).
2577 */
2578static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
2579 struct walk_control *wc, u64 gen, int level)
2580{
2581 int nritems;
2582 struct btrfs_path *path;
2583 struct btrfs_root *root = wc->replay_dest;
2584 struct btrfs_key key;
2585 int i;
2586 int ret;
2587
2588 ret = btrfs_read_buffer(eb, gen, level, NULL);
2589 if (ret)
2590 return ret;
2591
2592 level = btrfs_header_level(eb);
2593
2594 if (level != 0)
2595 return 0;
2596
2597 path = btrfs_alloc_path();
2598 if (!path)
2599 return -ENOMEM;
2600
2601 nritems = btrfs_header_nritems(eb);
2602 for (i = 0; i < nritems; i++) {
2603 btrfs_item_key_to_cpu(eb, &key, i);
2604
2605 /* inode keys are done during the first stage */
2606 if (key.type == BTRFS_INODE_ITEM_KEY &&
2607 wc->stage == LOG_WALK_REPLAY_INODES) {
2608 struct btrfs_inode_item *inode_item;
2609 u32 mode;
2610
2611 inode_item = btrfs_item_ptr(eb, i,
2612 struct btrfs_inode_item);
2613 /*
2614 * If we have a tmpfile (O_TMPFILE) that got fsync'ed
2615 * and never got linked before the fsync, skip it, as
2616 * replaying it is pointless since it would be deleted
2617 * later. We skip logging tmpfiles, but it's always
2618 * possible we are replaying a log created with a kernel
2619 * that used to log tmpfiles.
2620 */
2621 if (btrfs_inode_nlink(eb, inode_item) == 0) {
2622 wc->ignore_cur_inode = true;
2623 continue;
2624 } else {
2625 wc->ignore_cur_inode = false;
2626 }
2627 ret = replay_xattr_deletes(wc->trans, root, log,
2628 path, key.objectid);
2629 if (ret)
2630 break;
2631 mode = btrfs_inode_mode(eb, inode_item);
2632 if (S_ISDIR(mode)) {
2633 ret = replay_dir_deletes(wc->trans,
2634 root, log, path, key.objectid, 0);
2635 if (ret)
2636 break;
2637 }
2638 ret = overwrite_item(wc->trans, root, path,
2639 eb, i, &key);
2640 if (ret)
2641 break;
2642
2643 /*
2644 * Before replaying extents, truncate the inode to its
2645 * size. We need to do it now and not after log replay
2646 * because before an fsync we can have prealloc extents
2647 * added beyond the inode's i_size. If we did it after,
2648 * through orphan cleanup for example, we would drop
2649 * those prealloc extents just after replaying them.
2650 */
2651 if (S_ISREG(mode)) {
2652 struct btrfs_drop_extents_args drop_args = { 0 };
2653 struct inode *inode;
2654 u64 from;
2655
2656 inode = read_one_inode(root, key.objectid);
2657 if (!inode) {
2658 ret = -EIO;
2659 break;
2660 }
2661 from = ALIGN(i_size_read(inode),
2662 root->fs_info->sectorsize);
2663 drop_args.start = from;
2664 drop_args.end = (u64)-1;
2665 drop_args.drop_cache = true;
2666 ret = btrfs_drop_extents(wc->trans, root,
2667 BTRFS_I(inode),
2668 &drop_args);
2669 if (!ret) {
2670 inode_sub_bytes(inode,
2671 drop_args.bytes_found);
2672 /* Update the inode's nbytes. */
2673 ret = btrfs_update_inode(wc->trans,
2674 root, BTRFS_I(inode));
2675 }
2676 iput(inode);
2677 if (ret)
2678 break;
2679 }
2680
2681 ret = link_to_fixup_dir(wc->trans, root,
2682 path, key.objectid);
2683 if (ret)
2684 break;
2685 }
2686
2687 if (wc->ignore_cur_inode)
2688 continue;
2689
2690 if (key.type == BTRFS_DIR_INDEX_KEY &&
2691 wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
2692 ret = replay_one_dir_item(wc->trans, root, path,
2693 eb, i, &key);
2694 if (ret)
2695 break;
2696 }
2697
2698 if (wc->stage < LOG_WALK_REPLAY_ALL)
2699 continue;
2700
2701 /* these keys are simply copied */
2702 if (key.type == BTRFS_XATTR_ITEM_KEY) {
2703 ret = overwrite_item(wc->trans, root, path,
2704 eb, i, &key);
2705 if (ret)
2706 break;
2707 } else if (key.type == BTRFS_INODE_REF_KEY ||
2708 key.type == BTRFS_INODE_EXTREF_KEY) {
2709 ret = add_inode_ref(wc->trans, root, log, path,
2710 eb, i, &key);
2711 if (ret && ret != -ENOENT)
2712 break;
2713 ret = 0;
2714 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
2715 ret = replay_one_extent(wc->trans, root, path,
2716 eb, i, &key);
2717 if (ret)
2718 break;
2719 } else if (key.type == BTRFS_DIR_ITEM_KEY) {
2720 ret = replay_one_dir_item(wc->trans, root, path,
2721 eb, i, &key);
2722 if (ret)
2723 break;
2724 }
2725 }
2726 btrfs_free_path(path);
2727 return ret;
2728}
2729
2730/*
2731 * Correctly adjust the reserved bytes occupied by a log tree extent buffer
2732 */
2733static void unaccount_log_buffer(struct btrfs_fs_info *fs_info, u64 start)
2734{
2735 struct btrfs_block_group *cache;
2736
2737 cache = btrfs_lookup_block_group(fs_info, start);
2738 if (!cache) {
2739 btrfs_err(fs_info, "unable to find block group for %llu", start);
2740 return;
2741 }
2742
2743 spin_lock(&cache->space_info->lock);
2744 spin_lock(&cache->lock);
2745 cache->reserved -= fs_info->nodesize;
2746 cache->space_info->bytes_reserved -= fs_info->nodesize;
2747 spin_unlock(&cache->lock);
2748 spin_unlock(&cache->space_info->lock);
2749
2750 btrfs_put_block_group(cache);
2751}
2752
2753static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
2754 struct btrfs_root *root,
2755 struct btrfs_path *path, int *level,
2756 struct walk_control *wc)
2757{
2758 struct btrfs_fs_info *fs_info = root->fs_info;
2759 u64 bytenr;
2760 u64 ptr_gen;
2761 struct extent_buffer *next;
2762 struct extent_buffer *cur;
2763 u32 blocksize;
2764 int ret = 0;
2765
2766 while (*level > 0) {
2767 struct btrfs_key first_key;
2768
2769 cur = path->nodes[*level];
2770
2771 WARN_ON(btrfs_header_level(cur) != *level);
2772
2773 if (path->slots[*level] >=
2774 btrfs_header_nritems(cur))
2775 break;
2776
2777 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
2778 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
2779 btrfs_node_key_to_cpu(cur, &first_key, path->slots[*level]);
2780 blocksize = fs_info->nodesize;
2781
2782 next = btrfs_find_create_tree_block(fs_info, bytenr,
2783 btrfs_header_owner(cur),
2784 *level - 1);
2785 if (IS_ERR(next))
2786 return PTR_ERR(next);
2787
2788 if (*level == 1) {
2789 ret = wc->process_func(root, next, wc, ptr_gen,
2790 *level - 1);
2791 if (ret) {
2792 free_extent_buffer(next);
2793 return ret;
2794 }
2795
2796 path->slots[*level]++;
2797 if (wc->free) {
2798 ret = btrfs_read_buffer(next, ptr_gen,
2799 *level - 1, &first_key);
2800 if (ret) {
2801 free_extent_buffer(next);
2802 return ret;
2803 }
2804
2805 if (trans) {
2806 btrfs_tree_lock(next);
2807 btrfs_clean_tree_block(next);
2808 btrfs_wait_tree_block_writeback(next);
2809 btrfs_tree_unlock(next);
2810 ret = btrfs_pin_reserved_extent(trans,
2811 bytenr, blocksize);
2812 if (ret) {
2813 free_extent_buffer(next);
2814 return ret;
2815 }
2816 btrfs_redirty_list_add(
2817 trans->transaction, next);
2818 } else {
2819 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
2820 clear_extent_buffer_dirty(next);
2821 unaccount_log_buffer(fs_info, bytenr);
2822 }
2823 }
2824 free_extent_buffer(next);
2825 continue;
2826 }
2827 ret = btrfs_read_buffer(next, ptr_gen, *level - 1, &first_key);
2828 if (ret) {
2829 free_extent_buffer(next);
2830 return ret;
2831 }
2832
2833 if (path->nodes[*level-1])
2834 free_extent_buffer(path->nodes[*level-1]);
2835 path->nodes[*level-1] = next;
2836 *level = btrfs_header_level(next);
2837 path->slots[*level] = 0;
2838 cond_resched();
2839 }
2840 path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2841
2842 cond_resched();
2843 return 0;
2844}
2845
2846static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2847 struct btrfs_root *root,
2848 struct btrfs_path *path, int *level,
2849 struct walk_control *wc)
2850{
2851 struct btrfs_fs_info *fs_info = root->fs_info;
2852 int i;
2853 int slot;
2854 int ret;
2855
2856 for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2857 slot = path->slots[i];
2858 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2859 path->slots[i]++;
2860 *level = i;
2861 WARN_ON(*level == 0);
2862 return 0;
2863 } else {
2864 ret = wc->process_func(root, path->nodes[*level], wc,
2865 btrfs_header_generation(path->nodes[*level]),
2866 *level);
2867 if (ret)
2868 return ret;
2869
2870 if (wc->free) {
2871 struct extent_buffer *next;
2872
2873 next = path->nodes[*level];
2874
2875 if (trans) {
2876 btrfs_tree_lock(next);
2877 btrfs_clean_tree_block(next);
2878 btrfs_wait_tree_block_writeback(next);
2879 btrfs_tree_unlock(next);
2880 ret = btrfs_pin_reserved_extent(trans,
2881 path->nodes[*level]->start,
2882 path->nodes[*level]->len);
2883 if (ret)
2884 return ret;
2885 } else {
2886 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
2887 clear_extent_buffer_dirty(next);
2888
2889 unaccount_log_buffer(fs_info,
2890 path->nodes[*level]->start);
2891 }
2892 }
2893 free_extent_buffer(path->nodes[*level]);
2894 path->nodes[*level] = NULL;
2895 *level = i + 1;
2896 }
2897 }
2898 return 1;
2899}
2900
2901/*
2902 * drop the reference count on the tree rooted at 'snap'. This traverses
2903 * the tree freeing any blocks that have a ref count of zero after being
2904 * decremented.
2905 */
2906static int walk_log_tree(struct btrfs_trans_handle *trans,
2907 struct btrfs_root *log, struct walk_control *wc)
2908{
2909 struct btrfs_fs_info *fs_info = log->fs_info;
2910 int ret = 0;
2911 int wret;
2912 int level;
2913 struct btrfs_path *path;
2914 int orig_level;
2915
2916 path = btrfs_alloc_path();
2917 if (!path)
2918 return -ENOMEM;
2919
2920 level = btrfs_header_level(log->node);
2921 orig_level = level;
2922 path->nodes[level] = log->node;
2923 atomic_inc(&log->node->refs);
2924 path->slots[level] = 0;
2925
2926 while (1) {
2927 wret = walk_down_log_tree(trans, log, path, &level, wc);
2928 if (wret > 0)
2929 break;
2930 if (wret < 0) {
2931 ret = wret;
2932 goto out;
2933 }
2934
2935 wret = walk_up_log_tree(trans, log, path, &level, wc);
2936 if (wret > 0)
2937 break;
2938 if (wret < 0) {
2939 ret = wret;
2940 goto out;
2941 }
2942 }
2943
2944 /* was the root node processed? if not, catch it here */
2945 if (path->nodes[orig_level]) {
2946 ret = wc->process_func(log, path->nodes[orig_level], wc,
2947 btrfs_header_generation(path->nodes[orig_level]),
2948 orig_level);
2949 if (ret)
2950 goto out;
2951 if (wc->free) {
2952 struct extent_buffer *next;
2953
2954 next = path->nodes[orig_level];
2955
2956 if (trans) {
2957 btrfs_tree_lock(next);
2958 btrfs_clean_tree_block(next);
2959 btrfs_wait_tree_block_writeback(next);
2960 btrfs_tree_unlock(next);
2961 ret = btrfs_pin_reserved_extent(trans,
2962 next->start, next->len);
2963 if (ret)
2964 goto out;
2965 } else {
2966 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
2967 clear_extent_buffer_dirty(next);
2968 unaccount_log_buffer(fs_info, next->start);
2969 }
2970 }
2971 }
2972
2973out:
2974 btrfs_free_path(path);
2975 return ret;
2976}
2977
2978/*
2979 * helper function to update the item for a given subvolumes log root
2980 * in the tree of log roots
2981 */
2982static int update_log_root(struct btrfs_trans_handle *trans,
2983 struct btrfs_root *log,
2984 struct btrfs_root_item *root_item)
2985{
2986 struct btrfs_fs_info *fs_info = log->fs_info;
2987 int ret;
2988
2989 if (log->log_transid == 1) {
2990 /* insert root item on the first sync */
2991 ret = btrfs_insert_root(trans, fs_info->log_root_tree,
2992 &log->root_key, root_item);
2993 } else {
2994 ret = btrfs_update_root(trans, fs_info->log_root_tree,
2995 &log->root_key, root_item);
2996 }
2997 return ret;
2998}
2999
3000static void wait_log_commit(struct btrfs_root *root, int transid)
3001{
3002 DEFINE_WAIT(wait);
3003 int index = transid % 2;
3004
3005 /*
3006 * we only allow two pending log transactions at a time,
3007 * so we know that if ours is more than 2 older than the
3008 * current transaction, we're done
3009 */
3010 for (;;) {
3011 prepare_to_wait(&root->log_commit_wait[index],
3012 &wait, TASK_UNINTERRUPTIBLE);
3013
3014 if (!(root->log_transid_committed < transid &&
3015 atomic_read(&root->log_commit[index])))
3016 break;
3017
3018 mutex_unlock(&root->log_mutex);
3019 schedule();
3020 mutex_lock(&root->log_mutex);
3021 }
3022 finish_wait(&root->log_commit_wait[index], &wait);
3023}
3024
3025static void wait_for_writer(struct btrfs_root *root)
3026{
3027 DEFINE_WAIT(wait);
3028
3029 for (;;) {
3030 prepare_to_wait(&root->log_writer_wait, &wait,
3031 TASK_UNINTERRUPTIBLE);
3032 if (!atomic_read(&root->log_writers))
3033 break;
3034
3035 mutex_unlock(&root->log_mutex);
3036 schedule();
3037 mutex_lock(&root->log_mutex);
3038 }
3039 finish_wait(&root->log_writer_wait, &wait);
3040}
3041
3042static inline void btrfs_remove_log_ctx(struct btrfs_root *root,
3043 struct btrfs_log_ctx *ctx)
3044{
3045 if (!ctx)
3046 return;
3047
3048 mutex_lock(&root->log_mutex);
3049 list_del_init(&ctx->list);
3050 mutex_unlock(&root->log_mutex);
3051}
3052
3053/*
3054 * Invoked in log mutex context, or be sure there is no other task which
3055 * can access the list.
3056 */
3057static inline void btrfs_remove_all_log_ctxs(struct btrfs_root *root,
3058 int index, int error)
3059{
3060 struct btrfs_log_ctx *ctx;
3061 struct btrfs_log_ctx *safe;
3062
3063 list_for_each_entry_safe(ctx, safe, &root->log_ctxs[index], list) {
3064 list_del_init(&ctx->list);
3065 ctx->log_ret = error;
3066 }
3067
3068 INIT_LIST_HEAD(&root->log_ctxs[index]);
3069}
3070
3071/*
3072 * btrfs_sync_log does sends a given tree log down to the disk and
3073 * updates the super blocks to record it. When this call is done,
3074 * you know that any inodes previously logged are safely on disk only
3075 * if it returns 0.
3076 *
3077 * Any other return value means you need to call btrfs_commit_transaction.
3078 * Some of the edge cases for fsyncing directories that have had unlinks
3079 * or renames done in the past mean that sometimes the only safe
3080 * fsync is to commit the whole FS. When btrfs_sync_log returns -EAGAIN,
3081 * that has happened.
3082 */
3083int btrfs_sync_log(struct btrfs_trans_handle *trans,
3084 struct btrfs_root *root, struct btrfs_log_ctx *ctx)
3085{
3086 int index1;
3087 int index2;
3088 int mark;
3089 int ret;
3090 struct btrfs_fs_info *fs_info = root->fs_info;
3091 struct btrfs_root *log = root->log_root;
3092 struct btrfs_root *log_root_tree = fs_info->log_root_tree;
3093 struct btrfs_root_item new_root_item;
3094 int log_transid = 0;
3095 struct btrfs_log_ctx root_log_ctx;
3096 struct blk_plug plug;
3097 u64 log_root_start;
3098 u64 log_root_level;
3099
3100 mutex_lock(&root->log_mutex);
3101 log_transid = ctx->log_transid;
3102 if (root->log_transid_committed >= log_transid) {
3103 mutex_unlock(&root->log_mutex);
3104 return ctx->log_ret;
3105 }
3106
3107 index1 = log_transid % 2;
3108 if (atomic_read(&root->log_commit[index1])) {
3109 wait_log_commit(root, log_transid);
3110 mutex_unlock(&root->log_mutex);
3111 return ctx->log_ret;
3112 }
3113 ASSERT(log_transid == root->log_transid);
3114 atomic_set(&root->log_commit[index1], 1);
3115
3116 /* wait for previous tree log sync to complete */
3117 if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
3118 wait_log_commit(root, log_transid - 1);
3119
3120 while (1) {
3121 int batch = atomic_read(&root->log_batch);
3122 /* when we're on an ssd, just kick the log commit out */
3123 if (!btrfs_test_opt(fs_info, SSD) &&
3124 test_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state)) {
3125 mutex_unlock(&root->log_mutex);
3126 schedule_timeout_uninterruptible(1);
3127 mutex_lock(&root->log_mutex);
3128 }
3129 wait_for_writer(root);
3130 if (batch == atomic_read(&root->log_batch))
3131 break;
3132 }
3133
3134 /* bail out if we need to do a full commit */
3135 if (btrfs_need_log_full_commit(trans)) {
3136 ret = -EAGAIN;
3137 mutex_unlock(&root->log_mutex);
3138 goto out;
3139 }
3140
3141 if (log_transid % 2 == 0)
3142 mark = EXTENT_DIRTY;
3143 else
3144 mark = EXTENT_NEW;
3145
3146 /* we start IO on all the marked extents here, but we don't actually
3147 * wait for them until later.
3148 */
3149 blk_start_plug(&plug);
3150 ret = btrfs_write_marked_extents(fs_info, &log->dirty_log_pages, mark);
3151 /*
3152 * -EAGAIN happens when someone, e.g., a concurrent transaction
3153 * commit, writes a dirty extent in this tree-log commit. This
3154 * concurrent write will create a hole writing out the extents,
3155 * and we cannot proceed on a zoned filesystem, requiring
3156 * sequential writing. While we can bail out to a full commit
3157 * here, but we can continue hoping the concurrent writing fills
3158 * the hole.
3159 */
3160 if (ret == -EAGAIN && btrfs_is_zoned(fs_info))
3161 ret = 0;
3162 if (ret) {
3163 blk_finish_plug(&plug);
3164 btrfs_abort_transaction(trans, ret);
3165 btrfs_set_log_full_commit(trans);
3166 mutex_unlock(&root->log_mutex);
3167 goto out;
3168 }
3169
3170 /*
3171 * We _must_ update under the root->log_mutex in order to make sure we
3172 * have a consistent view of the log root we are trying to commit at
3173 * this moment.
3174 *
3175 * We _must_ copy this into a local copy, because we are not holding the
3176 * log_root_tree->log_mutex yet. This is important because when we
3177 * commit the log_root_tree we must have a consistent view of the
3178 * log_root_tree when we update the super block to point at the
3179 * log_root_tree bytenr. If we update the log_root_tree here we'll race
3180 * with the commit and possibly point at the new block which we may not
3181 * have written out.
3182 */
3183 btrfs_set_root_node(&log->root_item, log->node);
3184 memcpy(&new_root_item, &log->root_item, sizeof(new_root_item));
3185
3186 root->log_transid++;
3187 log->log_transid = root->log_transid;
3188 root->log_start_pid = 0;
3189 /*
3190 * IO has been started, blocks of the log tree have WRITTEN flag set
3191 * in their headers. new modifications of the log will be written to
3192 * new positions. so it's safe to allow log writers to go in.
3193 */
3194 mutex_unlock(&root->log_mutex);
3195
3196 if (btrfs_is_zoned(fs_info)) {
3197 mutex_lock(&fs_info->tree_root->log_mutex);
3198 if (!log_root_tree->node) {
3199 ret = btrfs_alloc_log_tree_node(trans, log_root_tree);
3200 if (ret) {
3201 mutex_unlock(&fs_info->tree_root->log_mutex);
3202 goto out;
3203 }
3204 }
3205 mutex_unlock(&fs_info->tree_root->log_mutex);
3206 }
3207
3208 btrfs_init_log_ctx(&root_log_ctx, NULL);
3209
3210 mutex_lock(&log_root_tree->log_mutex);
3211
3212 index2 = log_root_tree->log_transid % 2;
3213 list_add_tail(&root_log_ctx.list, &log_root_tree->log_ctxs[index2]);
3214 root_log_ctx.log_transid = log_root_tree->log_transid;
3215
3216 /*
3217 * Now we are safe to update the log_root_tree because we're under the
3218 * log_mutex, and we're a current writer so we're holding the commit
3219 * open until we drop the log_mutex.
3220 */
3221 ret = update_log_root(trans, log, &new_root_item);
3222 if (ret) {
3223 if (!list_empty(&root_log_ctx.list))
3224 list_del_init(&root_log_ctx.list);
3225
3226 blk_finish_plug(&plug);
3227 btrfs_set_log_full_commit(trans);
3228
3229 if (ret != -ENOSPC) {
3230 btrfs_abort_transaction(trans, ret);
3231 mutex_unlock(&log_root_tree->log_mutex);
3232 goto out;
3233 }
3234 btrfs_wait_tree_log_extents(log, mark);
3235 mutex_unlock(&log_root_tree->log_mutex);
3236 ret = -EAGAIN;
3237 goto out;
3238 }
3239
3240 if (log_root_tree->log_transid_committed >= root_log_ctx.log_transid) {
3241 blk_finish_plug(&plug);
3242 list_del_init(&root_log_ctx.list);
3243 mutex_unlock(&log_root_tree->log_mutex);
3244 ret = root_log_ctx.log_ret;
3245 goto out;
3246 }
3247
3248 index2 = root_log_ctx.log_transid % 2;
3249 if (atomic_read(&log_root_tree->log_commit[index2])) {
3250 blk_finish_plug(&plug);
3251 ret = btrfs_wait_tree_log_extents(log, mark);
3252 wait_log_commit(log_root_tree,
3253 root_log_ctx.log_transid);
3254 mutex_unlock(&log_root_tree->log_mutex);
3255 if (!ret)
3256 ret = root_log_ctx.log_ret;
3257 goto out;
3258 }
3259 ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
3260 atomic_set(&log_root_tree->log_commit[index2], 1);
3261
3262 if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
3263 wait_log_commit(log_root_tree,
3264 root_log_ctx.log_transid - 1);
3265 }
3266
3267 /*
3268 * now that we've moved on to the tree of log tree roots,
3269 * check the full commit flag again
3270 */
3271 if (btrfs_need_log_full_commit(trans)) {
3272 blk_finish_plug(&plug);
3273 btrfs_wait_tree_log_extents(log, mark);
3274 mutex_unlock(&log_root_tree->log_mutex);
3275 ret = -EAGAIN;
3276 goto out_wake_log_root;
3277 }
3278
3279 ret = btrfs_write_marked_extents(fs_info,
3280 &log_root_tree->dirty_log_pages,
3281 EXTENT_DIRTY | EXTENT_NEW);
3282 blk_finish_plug(&plug);
3283 /*
3284 * As described above, -EAGAIN indicates a hole in the extents. We
3285 * cannot wait for these write outs since the waiting cause a
3286 * deadlock. Bail out to the full commit instead.
3287 */
3288 if (ret == -EAGAIN && btrfs_is_zoned(fs_info)) {
3289 btrfs_set_log_full_commit(trans);
3290 btrfs_wait_tree_log_extents(log, mark);
3291 mutex_unlock(&log_root_tree->log_mutex);
3292 goto out_wake_log_root;
3293 } else if (ret) {
3294 btrfs_set_log_full_commit(trans);
3295 btrfs_abort_transaction(trans, ret);
3296 mutex_unlock(&log_root_tree->log_mutex);
3297 goto out_wake_log_root;
3298 }
3299 ret = btrfs_wait_tree_log_extents(log, mark);
3300 if (!ret)
3301 ret = btrfs_wait_tree_log_extents(log_root_tree,
3302 EXTENT_NEW | EXTENT_DIRTY);
3303 if (ret) {
3304 btrfs_set_log_full_commit(trans);
3305 mutex_unlock(&log_root_tree->log_mutex);
3306 goto out_wake_log_root;
3307 }
3308
3309 log_root_start = log_root_tree->node->start;
3310 log_root_level = btrfs_header_level(log_root_tree->node);
3311 log_root_tree->log_transid++;
3312 mutex_unlock(&log_root_tree->log_mutex);
3313
3314 /*
3315 * Here we are guaranteed that nobody is going to write the superblock
3316 * for the current transaction before us and that neither we do write
3317 * our superblock before the previous transaction finishes its commit
3318 * and writes its superblock, because:
3319 *
3320 * 1) We are holding a handle on the current transaction, so no body
3321 * can commit it until we release the handle;
3322 *
3323 * 2) Before writing our superblock we acquire the tree_log_mutex, so
3324 * if the previous transaction is still committing, and hasn't yet
3325 * written its superblock, we wait for it to do it, because a
3326 * transaction commit acquires the tree_log_mutex when the commit
3327 * begins and releases it only after writing its superblock.
3328 */
3329 mutex_lock(&fs_info->tree_log_mutex);
3330
3331 /*
3332 * The previous transaction writeout phase could have failed, and thus
3333 * marked the fs in an error state. We must not commit here, as we
3334 * could have updated our generation in the super_for_commit and
3335 * writing the super here would result in transid mismatches. If there
3336 * is an error here just bail.
3337 */
3338 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
3339 ret = -EIO;
3340 btrfs_set_log_full_commit(trans);
3341 btrfs_abort_transaction(trans, ret);
3342 mutex_unlock(&fs_info->tree_log_mutex);
3343 goto out_wake_log_root;
3344 }
3345
3346 btrfs_set_super_log_root(fs_info->super_for_commit, log_root_start);
3347 btrfs_set_super_log_root_level(fs_info->super_for_commit, log_root_level);
3348 ret = write_all_supers(fs_info, 1);
3349 mutex_unlock(&fs_info->tree_log_mutex);
3350 if (ret) {
3351 btrfs_set_log_full_commit(trans);
3352 btrfs_abort_transaction(trans, ret);
3353 goto out_wake_log_root;
3354 }
3355
3356 mutex_lock(&root->log_mutex);
3357 if (root->last_log_commit < log_transid)
3358 root->last_log_commit = log_transid;
3359 mutex_unlock(&root->log_mutex);
3360
3361out_wake_log_root:
3362 mutex_lock(&log_root_tree->log_mutex);
3363 btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
3364
3365 log_root_tree->log_transid_committed++;
3366 atomic_set(&log_root_tree->log_commit[index2], 0);
3367 mutex_unlock(&log_root_tree->log_mutex);
3368
3369 /*
3370 * The barrier before waitqueue_active (in cond_wake_up) is needed so
3371 * all the updates above are seen by the woken threads. It might not be
3372 * necessary, but proving that seems to be hard.
3373 */
3374 cond_wake_up(&log_root_tree->log_commit_wait[index2]);
3375out:
3376 mutex_lock(&root->log_mutex);
3377 btrfs_remove_all_log_ctxs(root, index1, ret);
3378 root->log_transid_committed++;
3379 atomic_set(&root->log_commit[index1], 0);
3380 mutex_unlock(&root->log_mutex);
3381
3382 /*
3383 * The barrier before waitqueue_active (in cond_wake_up) is needed so
3384 * all the updates above are seen by the woken threads. It might not be
3385 * necessary, but proving that seems to be hard.
3386 */
3387 cond_wake_up(&root->log_commit_wait[index1]);
3388 return ret;
3389}
3390
3391static void free_log_tree(struct btrfs_trans_handle *trans,
3392 struct btrfs_root *log)
3393{
3394 int ret;
3395 struct walk_control wc = {
3396 .free = 1,
3397 .process_func = process_one_buffer
3398 };
3399
3400 if (log->node) {
3401 ret = walk_log_tree(trans, log, &wc);
3402 if (ret) {
3403 if (trans)
3404 btrfs_abort_transaction(trans, ret);
3405 else
3406 btrfs_handle_fs_error(log->fs_info, ret, NULL);
3407 }
3408 }
3409
3410 clear_extent_bits(&log->dirty_log_pages, 0, (u64)-1,
3411 EXTENT_DIRTY | EXTENT_NEW | EXTENT_NEED_WAIT);
3412 extent_io_tree_release(&log->log_csum_range);
3413
3414 if (trans && log->node)
3415 btrfs_redirty_list_add(trans->transaction, log->node);
3416 btrfs_put_root(log);
3417}
3418
3419/*
3420 * free all the extents used by the tree log. This should be called
3421 * at commit time of the full transaction
3422 */
3423int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
3424{
3425 if (root->log_root) {
3426 free_log_tree(trans, root->log_root);
3427 root->log_root = NULL;
3428 clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
3429 }
3430 return 0;
3431}
3432
3433int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
3434 struct btrfs_fs_info *fs_info)
3435{
3436 if (fs_info->log_root_tree) {
3437 free_log_tree(trans, fs_info->log_root_tree);
3438 fs_info->log_root_tree = NULL;
3439 clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &fs_info->tree_root->state);
3440 }
3441 return 0;
3442}
3443
3444/*
3445 * Check if an inode was logged in the current transaction. We can't always rely
3446 * on an inode's logged_trans value, because it's an in-memory only field and
3447 * therefore not persisted. This means that its value is lost if the inode gets
3448 * evicted and loaded again from disk (in which case it has a value of 0, and
3449 * certainly it is smaller then any possible transaction ID), when that happens
3450 * the full_sync flag is set in the inode's runtime flags, so on that case we
3451 * assume eviction happened and ignore the logged_trans value, assuming the
3452 * worst case, that the inode was logged before in the current transaction.
3453 */
3454static bool inode_logged(struct btrfs_trans_handle *trans,
3455 struct btrfs_inode *inode)
3456{
3457 if (inode->logged_trans == trans->transid)
3458 return true;
3459
3460 if (inode->last_trans == trans->transid &&
3461 test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) &&
3462 !test_bit(BTRFS_FS_LOG_RECOVERING, &trans->fs_info->flags))
3463 return true;
3464
3465 return false;
3466}
3467
3468/*
3469 * If both a file and directory are logged, and unlinks or renames are
3470 * mixed in, we have a few interesting corners:
3471 *
3472 * create file X in dir Y
3473 * link file X to X.link in dir Y
3474 * fsync file X
3475 * unlink file X but leave X.link
3476 * fsync dir Y
3477 *
3478 * After a crash we would expect only X.link to exist. But file X
3479 * didn't get fsync'd again so the log has back refs for X and X.link.
3480 *
3481 * We solve this by removing directory entries and inode backrefs from the
3482 * log when a file that was logged in the current transaction is
3483 * unlinked. Any later fsync will include the updated log entries, and
3484 * we'll be able to reconstruct the proper directory items from backrefs.
3485 *
3486 * This optimizations allows us to avoid relogging the entire inode
3487 * or the entire directory.
3488 */
3489int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
3490 struct btrfs_root *root,
3491 const char *name, int name_len,
3492 struct btrfs_inode *dir, u64 index)
3493{
3494 struct btrfs_root *log;
3495 struct btrfs_dir_item *di;
3496 struct btrfs_path *path;
3497 int ret;
3498 int err = 0;
3499 u64 dir_ino = btrfs_ino(dir);
3500
3501 if (!inode_logged(trans, dir))
3502 return 0;
3503
3504 ret = join_running_log_trans(root);
3505 if (ret)
3506 return 0;
3507
3508 mutex_lock(&dir->log_mutex);
3509
3510 log = root->log_root;
3511 path = btrfs_alloc_path();
3512 if (!path) {
3513 err = -ENOMEM;
3514 goto out_unlock;
3515 }
3516
3517 di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
3518 name, name_len, -1);
3519 if (IS_ERR(di)) {
3520 err = PTR_ERR(di);
3521 goto fail;
3522 }
3523 if (di) {
3524 ret = btrfs_delete_one_dir_name(trans, log, path, di);
3525 if (ret) {
3526 err = ret;
3527 goto fail;
3528 }
3529 }
3530 btrfs_release_path(path);
3531 di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
3532 index, name, name_len, -1);
3533 if (IS_ERR(di)) {
3534 err = PTR_ERR(di);
3535 goto fail;
3536 }
3537 if (di) {
3538 ret = btrfs_delete_one_dir_name(trans, log, path, di);
3539 if (ret) {
3540 err = ret;
3541 goto fail;
3542 }
3543 }
3544
3545 /*
3546 * We do not need to update the size field of the directory's inode item
3547 * because on log replay we update the field to reflect all existing
3548 * entries in the directory (see overwrite_item()).
3549 */
3550fail:
3551 btrfs_free_path(path);
3552out_unlock:
3553 mutex_unlock(&dir->log_mutex);
3554 if (err == -ENOSPC) {
3555 btrfs_set_log_full_commit(trans);
3556 err = 0;
3557 } else if (err < 0 && err != -ENOENT) {
3558 /* ENOENT can be returned if the entry hasn't been fsynced yet */
3559 btrfs_abort_transaction(trans, err);
3560 }
3561
3562 btrfs_end_log_trans(root);
3563
3564 return err;
3565}
3566
3567/* see comments for btrfs_del_dir_entries_in_log */
3568int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
3569 struct btrfs_root *root,
3570 const char *name, int name_len,
3571 struct btrfs_inode *inode, u64 dirid)
3572{
3573 struct btrfs_root *log;
3574 u64 index;
3575 int ret;
3576
3577 if (!inode_logged(trans, inode))
3578 return 0;
3579
3580 ret = join_running_log_trans(root);
3581 if (ret)
3582 return 0;
3583 log = root->log_root;
3584 mutex_lock(&inode->log_mutex);
3585
3586 ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
3587 dirid, &index);
3588 mutex_unlock(&inode->log_mutex);
3589 if (ret == -ENOSPC) {
3590 btrfs_set_log_full_commit(trans);
3591 ret = 0;
3592 } else if (ret < 0 && ret != -ENOENT)
3593 btrfs_abort_transaction(trans, ret);
3594 btrfs_end_log_trans(root);
3595
3596 return ret;
3597}
3598
3599/*
3600 * creates a range item in the log for 'dirid'. first_offset and
3601 * last_offset tell us which parts of the key space the log should
3602 * be considered authoritative for.
3603 */
3604static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
3605 struct btrfs_root *log,
3606 struct btrfs_path *path,
3607 int key_type, u64 dirid,
3608 u64 first_offset, u64 last_offset)
3609{
3610 int ret;
3611 struct btrfs_key key;
3612 struct btrfs_dir_log_item *item;
3613
3614 key.objectid = dirid;
3615 key.offset = first_offset;
3616 if (key_type == BTRFS_DIR_ITEM_KEY)
3617 key.type = BTRFS_DIR_LOG_ITEM_KEY;
3618 else
3619 key.type = BTRFS_DIR_LOG_INDEX_KEY;
3620 ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
3621 if (ret)
3622 return ret;
3623
3624 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3625 struct btrfs_dir_log_item);
3626 btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
3627 btrfs_mark_buffer_dirty(path->nodes[0]);
3628 btrfs_release_path(path);
3629 return 0;
3630}
3631
3632/*
3633 * log all the items included in the current transaction for a given
3634 * directory. This also creates the range items in the log tree required
3635 * to replay anything deleted before the fsync
3636 */
3637static noinline int log_dir_items(struct btrfs_trans_handle *trans,
3638 struct btrfs_root *root, struct btrfs_inode *inode,
3639 struct btrfs_path *path,
3640 struct btrfs_path *dst_path, int key_type,
3641 struct btrfs_log_ctx *ctx,
3642 u64 min_offset, u64 *last_offset_ret)
3643{
3644 struct btrfs_key min_key;
3645 struct btrfs_root *log = root->log_root;
3646 struct extent_buffer *src;
3647 int err = 0;
3648 int ret;
3649 int i;
3650 int nritems;
3651 u64 first_offset = min_offset;
3652 u64 last_offset = (u64)-1;
3653 u64 ino = btrfs_ino(inode);
3654
3655 log = root->log_root;
3656
3657 min_key.objectid = ino;
3658 min_key.type = key_type;
3659 min_key.offset = min_offset;
3660
3661 ret = btrfs_search_forward(root, &min_key, path, trans->transid);
3662
3663 /*
3664 * we didn't find anything from this transaction, see if there
3665 * is anything at all
3666 */
3667 if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
3668 min_key.objectid = ino;
3669 min_key.type = key_type;
3670 min_key.offset = (u64)-1;
3671 btrfs_release_path(path);
3672 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3673 if (ret < 0) {
3674 btrfs_release_path(path);
3675 return ret;
3676 }
3677 ret = btrfs_previous_item(root, path, ino, key_type);
3678
3679 /* if ret == 0 there are items for this type,
3680 * create a range to tell us the last key of this type.
3681 * otherwise, there are no items in this directory after
3682 * *min_offset, and we create a range to indicate that.
3683 */
3684 if (ret == 0) {
3685 struct btrfs_key tmp;
3686 btrfs_item_key_to_cpu(path->nodes[0], &tmp,
3687 path->slots[0]);
3688 if (key_type == tmp.type)
3689 first_offset = max(min_offset, tmp.offset) + 1;
3690 }
3691 goto done;
3692 }
3693
3694 /* go backward to find any previous key */
3695 ret = btrfs_previous_item(root, path, ino, key_type);
3696 if (ret == 0) {
3697 struct btrfs_key tmp;
3698 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3699 if (key_type == tmp.type) {
3700 first_offset = tmp.offset;
3701 ret = overwrite_item(trans, log, dst_path,
3702 path->nodes[0], path->slots[0],
3703 &tmp);
3704 if (ret) {
3705 err = ret;
3706 goto done;
3707 }
3708 }
3709 }
3710 btrfs_release_path(path);
3711
3712 /*
3713 * Find the first key from this transaction again. See the note for
3714 * log_new_dir_dentries, if we're logging a directory recursively we
3715 * won't be holding its i_mutex, which means we can modify the directory
3716 * while we're logging it. If we remove an entry between our first
3717 * search and this search we'll not find the key again and can just
3718 * bail.
3719 */
3720search:
3721 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3722 if (ret != 0)
3723 goto done;
3724
3725 /*
3726 * we have a block from this transaction, log every item in it
3727 * from our directory
3728 */
3729 while (1) {
3730 struct btrfs_key tmp;
3731 src = path->nodes[0];
3732 nritems = btrfs_header_nritems(src);
3733 for (i = path->slots[0]; i < nritems; i++) {
3734 struct btrfs_dir_item *di;
3735
3736 btrfs_item_key_to_cpu(src, &min_key, i);
3737
3738 if (min_key.objectid != ino || min_key.type != key_type)
3739 goto done;
3740
3741 if (need_resched()) {
3742 btrfs_release_path(path);
3743 cond_resched();
3744 goto search;
3745 }
3746
3747 ret = overwrite_item(trans, log, dst_path, src, i,
3748 &min_key);
3749 if (ret) {
3750 err = ret;
3751 goto done;
3752 }
3753
3754 /*
3755 * We must make sure that when we log a directory entry,
3756 * the corresponding inode, after log replay, has a
3757 * matching link count. For example:
3758 *
3759 * touch foo
3760 * mkdir mydir
3761 * sync
3762 * ln foo mydir/bar
3763 * xfs_io -c "fsync" mydir
3764 * <crash>
3765 * <mount fs and log replay>
3766 *
3767 * Would result in a fsync log that when replayed, our
3768 * file inode would have a link count of 1, but we get
3769 * two directory entries pointing to the same inode.
3770 * After removing one of the names, it would not be
3771 * possible to remove the other name, which resulted
3772 * always in stale file handle errors, and would not
3773 * be possible to rmdir the parent directory, since
3774 * its i_size could never decrement to the value
3775 * BTRFS_EMPTY_DIR_SIZE, resulting in -ENOTEMPTY errors.
3776 */
3777 di = btrfs_item_ptr(src, i, struct btrfs_dir_item);
3778 btrfs_dir_item_key_to_cpu(src, di, &tmp);
3779 if (ctx &&
3780 (btrfs_dir_transid(src, di) == trans->transid ||
3781 btrfs_dir_type(src, di) == BTRFS_FT_DIR) &&
3782 tmp.type != BTRFS_ROOT_ITEM_KEY)
3783 ctx->log_new_dentries = true;
3784 }
3785 path->slots[0] = nritems;
3786
3787 /*
3788 * look ahead to the next item and see if it is also
3789 * from this directory and from this transaction
3790 */
3791 ret = btrfs_next_leaf(root, path);
3792 if (ret) {
3793 if (ret == 1)
3794 last_offset = (u64)-1;
3795 else
3796 err = ret;
3797 goto done;
3798 }
3799 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3800 if (tmp.objectid != ino || tmp.type != key_type) {
3801 last_offset = (u64)-1;
3802 goto done;
3803 }
3804 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
3805 ret = overwrite_item(trans, log, dst_path,
3806 path->nodes[0], path->slots[0],
3807 &tmp);
3808 if (ret)
3809 err = ret;
3810 else
3811 last_offset = tmp.offset;
3812 goto done;
3813 }
3814 }
3815done:
3816 btrfs_release_path(path);
3817 btrfs_release_path(dst_path);
3818
3819 if (err == 0) {
3820 *last_offset_ret = last_offset;
3821 /*
3822 * insert the log range keys to indicate where the log
3823 * is valid
3824 */
3825 ret = insert_dir_log_key(trans, log, path, key_type,
3826 ino, first_offset, last_offset);
3827 if (ret)
3828 err = ret;
3829 }
3830 return err;
3831}
3832
3833/*
3834 * logging directories is very similar to logging inodes, We find all the items
3835 * from the current transaction and write them to the log.
3836 *
3837 * The recovery code scans the directory in the subvolume, and if it finds a
3838 * key in the range logged that is not present in the log tree, then it means
3839 * that dir entry was unlinked during the transaction.
3840 *
3841 * In order for that scan to work, we must include one key smaller than
3842 * the smallest logged by this transaction and one key larger than the largest
3843 * key logged by this transaction.
3844 */
3845static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
3846 struct btrfs_root *root, struct btrfs_inode *inode,
3847 struct btrfs_path *path,
3848 struct btrfs_path *dst_path,
3849 struct btrfs_log_ctx *ctx)
3850{
3851 u64 min_key;
3852 u64 max_key;
3853 int ret;
3854 int key_type = BTRFS_DIR_ITEM_KEY;
3855
3856again:
3857 min_key = 0;
3858 max_key = 0;
3859 while (1) {
3860 ret = log_dir_items(trans, root, inode, path, dst_path, key_type,
3861 ctx, min_key, &max_key);
3862 if (ret)
3863 return ret;
3864 if (max_key == (u64)-1)
3865 break;
3866 min_key = max_key + 1;
3867 }
3868
3869 if (key_type == BTRFS_DIR_ITEM_KEY) {
3870 key_type = BTRFS_DIR_INDEX_KEY;
3871 goto again;
3872 }
3873 return 0;
3874}
3875
3876/*
3877 * a helper function to drop items from the log before we relog an
3878 * inode. max_key_type indicates the highest item type to remove.
3879 * This cannot be run for file data extents because it does not
3880 * free the extents they point to.
3881 */
3882static int drop_objectid_items(struct btrfs_trans_handle *trans,
3883 struct btrfs_root *log,
3884 struct btrfs_path *path,
3885 u64 objectid, int max_key_type)
3886{
3887 int ret;
3888 struct btrfs_key key;
3889 struct btrfs_key found_key;
3890 int start_slot;
3891
3892 key.objectid = objectid;
3893 key.type = max_key_type;
3894 key.offset = (u64)-1;
3895
3896 while (1) {
3897 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
3898 BUG_ON(ret == 0); /* Logic error */
3899 if (ret < 0)
3900 break;
3901
3902 if (path->slots[0] == 0)
3903 break;
3904
3905 path->slots[0]--;
3906 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
3907 path->slots[0]);
3908
3909 if (found_key.objectid != objectid)
3910 break;
3911
3912 found_key.offset = 0;
3913 found_key.type = 0;
3914 ret = btrfs_bin_search(path->nodes[0], &found_key, &start_slot);
3915 if (ret < 0)
3916 break;
3917
3918 ret = btrfs_del_items(trans, log, path, start_slot,
3919 path->slots[0] - start_slot + 1);
3920 /*
3921 * If start slot isn't 0 then we don't need to re-search, we've
3922 * found the last guy with the objectid in this tree.
3923 */
3924 if (ret || start_slot != 0)
3925 break;
3926 btrfs_release_path(path);
3927 }
3928 btrfs_release_path(path);
3929 if (ret > 0)
3930 ret = 0;
3931 return ret;
3932}
3933
3934static void fill_inode_item(struct btrfs_trans_handle *trans,
3935 struct extent_buffer *leaf,
3936 struct btrfs_inode_item *item,
3937 struct inode *inode, int log_inode_only,
3938 u64 logged_isize)
3939{
3940 struct btrfs_map_token token;
3941
3942 btrfs_init_map_token(&token, leaf);
3943
3944 if (log_inode_only) {
3945 /* set the generation to zero so the recover code
3946 * can tell the difference between an logging
3947 * just to say 'this inode exists' and a logging
3948 * to say 'update this inode with these values'
3949 */
3950 btrfs_set_token_inode_generation(&token, item, 0);
3951 btrfs_set_token_inode_size(&token, item, logged_isize);
3952 } else {
3953 btrfs_set_token_inode_generation(&token, item,
3954 BTRFS_I(inode)->generation);
3955 btrfs_set_token_inode_size(&token, item, inode->i_size);
3956 }
3957
3958 btrfs_set_token_inode_uid(&token, item, i_uid_read(inode));
3959 btrfs_set_token_inode_gid(&token, item, i_gid_read(inode));
3960 btrfs_set_token_inode_mode(&token, item, inode->i_mode);
3961 btrfs_set_token_inode_nlink(&token, item, inode->i_nlink);
3962
3963 btrfs_set_token_timespec_sec(&token, &item->atime,
3964 inode->i_atime.tv_sec);
3965 btrfs_set_token_timespec_nsec(&token, &item->atime,
3966 inode->i_atime.tv_nsec);
3967
3968 btrfs_set_token_timespec_sec(&token, &item->mtime,
3969 inode->i_mtime.tv_sec);
3970 btrfs_set_token_timespec_nsec(&token, &item->mtime,
3971 inode->i_mtime.tv_nsec);
3972
3973 btrfs_set_token_timespec_sec(&token, &item->ctime,
3974 inode->i_ctime.tv_sec);
3975 btrfs_set_token_timespec_nsec(&token, &item->ctime,
3976 inode->i_ctime.tv_nsec);
3977
3978 /*
3979 * We do not need to set the nbytes field, in fact during a fast fsync
3980 * its value may not even be correct, since a fast fsync does not wait
3981 * for ordered extent completion, which is where we update nbytes, it
3982 * only waits for writeback to complete. During log replay as we find
3983 * file extent items and replay them, we adjust the nbytes field of the
3984 * inode item in subvolume tree as needed (see overwrite_item()).
3985 */
3986
3987 btrfs_set_token_inode_sequence(&token, item, inode_peek_iversion(inode));
3988 btrfs_set_token_inode_transid(&token, item, trans->transid);
3989 btrfs_set_token_inode_rdev(&token, item, inode->i_rdev);
3990 btrfs_set_token_inode_flags(&token, item, BTRFS_I(inode)->flags);
3991 btrfs_set_token_inode_block_group(&token, item, 0);
3992}
3993
3994static int log_inode_item(struct btrfs_trans_handle *trans,
3995 struct btrfs_root *log, struct btrfs_path *path,
3996 struct btrfs_inode *inode)
3997{
3998 struct btrfs_inode_item *inode_item;
3999 int ret;
4000
4001 ret = btrfs_insert_empty_item(trans, log, path,
4002 &inode->location, sizeof(*inode_item));
4003 if (ret && ret != -EEXIST)
4004 return ret;
4005 inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4006 struct btrfs_inode_item);
4007 fill_inode_item(trans, path->nodes[0], inode_item, &inode->vfs_inode,
4008 0, 0);
4009 btrfs_release_path(path);
4010 return 0;
4011}
4012
4013static int log_csums(struct btrfs_trans_handle *trans,
4014 struct btrfs_inode *inode,
4015 struct btrfs_root *log_root,
4016 struct btrfs_ordered_sum *sums)
4017{
4018 const u64 lock_end = sums->bytenr + sums->len - 1;
4019 struct extent_state *cached_state = NULL;
4020 int ret;
4021
4022 /*
4023 * If this inode was not used for reflink operations in the current
4024 * transaction with new extents, then do the fast path, no need to
4025 * worry about logging checksum items with overlapping ranges.
4026 */
4027 if (inode->last_reflink_trans < trans->transid)
4028 return btrfs_csum_file_blocks(trans, log_root, sums);
4029
4030 /*
4031 * Serialize logging for checksums. This is to avoid racing with the
4032 * same checksum being logged by another task that is logging another
4033 * file which happens to refer to the same extent as well. Such races
4034 * can leave checksum items in the log with overlapping ranges.
4035 */
4036 ret = lock_extent_bits(&log_root->log_csum_range, sums->bytenr,
4037 lock_end, &cached_state);
4038 if (ret)
4039 return ret;
4040 /*
4041 * Due to extent cloning, we might have logged a csum item that covers a
4042 * subrange of a cloned extent, and later we can end up logging a csum
4043 * item for a larger subrange of the same extent or the entire range.
4044 * This would leave csum items in the log tree that cover the same range
4045 * and break the searches for checksums in the log tree, resulting in
4046 * some checksums missing in the fs/subvolume tree. So just delete (or
4047 * trim and adjust) any existing csum items in the log for this range.
4048 */
4049 ret = btrfs_del_csums(trans, log_root, sums->bytenr, sums->len);
4050 if (!ret)
4051 ret = btrfs_csum_file_blocks(trans, log_root, sums);
4052
4053 unlock_extent_cached(&log_root->log_csum_range, sums->bytenr, lock_end,
4054 &cached_state);
4055
4056 return ret;
4057}
4058
4059static noinline int copy_items(struct btrfs_trans_handle *trans,
4060 struct btrfs_inode *inode,
4061 struct btrfs_path *dst_path,
4062 struct btrfs_path *src_path,
4063 int start_slot, int nr, int inode_only,
4064 u64 logged_isize)
4065{
4066 struct btrfs_fs_info *fs_info = trans->fs_info;
4067 unsigned long src_offset;
4068 unsigned long dst_offset;
4069 struct btrfs_root *log = inode->root->log_root;
4070 struct btrfs_file_extent_item *extent;
4071 struct btrfs_inode_item *inode_item;
4072 struct extent_buffer *src = src_path->nodes[0];
4073 int ret;
4074 struct btrfs_key *ins_keys;
4075 u32 *ins_sizes;
4076 char *ins_data;
4077 int i;
4078 struct list_head ordered_sums;
4079 int skip_csum = inode->flags & BTRFS_INODE_NODATASUM;
4080
4081 INIT_LIST_HEAD(&ordered_sums);
4082
4083 ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
4084 nr * sizeof(u32), GFP_NOFS);
4085 if (!ins_data)
4086 return -ENOMEM;
4087
4088 ins_sizes = (u32 *)ins_data;
4089 ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
4090
4091 for (i = 0; i < nr; i++) {
4092 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
4093 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
4094 }
4095 ret = btrfs_insert_empty_items(trans, log, dst_path,
4096 ins_keys, ins_sizes, nr);
4097 if (ret) {
4098 kfree(ins_data);
4099 return ret;
4100 }
4101
4102 for (i = 0; i < nr; i++, dst_path->slots[0]++) {
4103 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
4104 dst_path->slots[0]);
4105
4106 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
4107
4108 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
4109 inode_item = btrfs_item_ptr(dst_path->nodes[0],
4110 dst_path->slots[0],
4111 struct btrfs_inode_item);
4112 fill_inode_item(trans, dst_path->nodes[0], inode_item,
4113 &inode->vfs_inode,
4114 inode_only == LOG_INODE_EXISTS,
4115 logged_isize);
4116 } else {
4117 copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
4118 src_offset, ins_sizes[i]);
4119 }
4120
4121 /* take a reference on file data extents so that truncates
4122 * or deletes of this inode don't have to relog the inode
4123 * again
4124 */
4125 if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY &&
4126 !skip_csum) {
4127 int found_type;
4128 extent = btrfs_item_ptr(src, start_slot + i,
4129 struct btrfs_file_extent_item);
4130
4131 if (btrfs_file_extent_generation(src, extent) < trans->transid)
4132 continue;
4133
4134 found_type = btrfs_file_extent_type(src, extent);
4135 if (found_type == BTRFS_FILE_EXTENT_REG) {
4136 u64 ds, dl, cs, cl;
4137 ds = btrfs_file_extent_disk_bytenr(src,
4138 extent);
4139 /* ds == 0 is a hole */
4140 if (ds == 0)
4141 continue;
4142
4143 dl = btrfs_file_extent_disk_num_bytes(src,
4144 extent);
4145 cs = btrfs_file_extent_offset(src, extent);
4146 cl = btrfs_file_extent_num_bytes(src,
4147 extent);
4148 if (btrfs_file_extent_compression(src,
4149 extent)) {
4150 cs = 0;
4151 cl = dl;
4152 }
4153
4154 ret = btrfs_lookup_csums_range(
4155 fs_info->csum_root,
4156 ds + cs, ds + cs + cl - 1,
4157 &ordered_sums, 0);
4158 if (ret)
4159 break;
4160 }
4161 }
4162 }
4163
4164 btrfs_mark_buffer_dirty(dst_path->nodes[0]);
4165 btrfs_release_path(dst_path);
4166 kfree(ins_data);
4167
4168 /*
4169 * we have to do this after the loop above to avoid changing the
4170 * log tree while trying to change the log tree.
4171 */
4172 while (!list_empty(&ordered_sums)) {
4173 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
4174 struct btrfs_ordered_sum,
4175 list);
4176 if (!ret)
4177 ret = log_csums(trans, inode, log, sums);
4178 list_del(&sums->list);
4179 kfree(sums);
4180 }
4181
4182 return ret;
4183}
4184
4185static int extent_cmp(void *priv, const struct list_head *a,
4186 const struct list_head *b)
4187{
4188 struct extent_map *em1, *em2;
4189
4190 em1 = list_entry(a, struct extent_map, list);
4191 em2 = list_entry(b, struct extent_map, list);
4192
4193 if (em1->start < em2->start)
4194 return -1;
4195 else if (em1->start > em2->start)
4196 return 1;
4197 return 0;
4198}
4199
4200static int log_extent_csums(struct btrfs_trans_handle *trans,
4201 struct btrfs_inode *inode,
4202 struct btrfs_root *log_root,
4203 const struct extent_map *em,
4204 struct btrfs_log_ctx *ctx)
4205{
4206 struct btrfs_ordered_extent *ordered;
4207 u64 csum_offset;
4208 u64 csum_len;
4209 u64 mod_start = em->mod_start;
4210 u64 mod_len = em->mod_len;
4211 LIST_HEAD(ordered_sums);
4212 int ret = 0;
4213
4214 if (inode->flags & BTRFS_INODE_NODATASUM ||
4215 test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
4216 em->block_start == EXTENT_MAP_HOLE)
4217 return 0;
4218
4219 list_for_each_entry(ordered, &ctx->ordered_extents, log_list) {
4220 const u64 ordered_end = ordered->file_offset + ordered->num_bytes;
4221 const u64 mod_end = mod_start + mod_len;
4222 struct btrfs_ordered_sum *sums;
4223
4224 if (mod_len == 0)
4225 break;
4226
4227 if (ordered_end <= mod_start)
4228 continue;
4229 if (mod_end <= ordered->file_offset)
4230 break;
4231
4232 /*
4233 * We are going to copy all the csums on this ordered extent, so
4234 * go ahead and adjust mod_start and mod_len in case this ordered
4235 * extent has already been logged.
4236 */
4237 if (ordered->file_offset > mod_start) {
4238 if (ordered_end >= mod_end)
4239 mod_len = ordered->file_offset - mod_start;
4240 /*
4241 * If we have this case
4242 *
4243 * |--------- logged extent ---------|
4244 * |----- ordered extent ----|
4245 *
4246 * Just don't mess with mod_start and mod_len, we'll
4247 * just end up logging more csums than we need and it
4248 * will be ok.
4249 */
4250 } else {
4251 if (ordered_end < mod_end) {
4252 mod_len = mod_end - ordered_end;
4253 mod_start = ordered_end;
4254 } else {
4255 mod_len = 0;
4256 }
4257 }
4258
4259 /*
4260 * To keep us from looping for the above case of an ordered
4261 * extent that falls inside of the logged extent.
4262 */
4263 if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM, &ordered->flags))
4264 continue;
4265
4266 list_for_each_entry(sums, &ordered->list, list) {
4267 ret = log_csums(trans, inode, log_root, sums);
4268 if (ret)
4269 return ret;
4270 }
4271 }
4272
4273 /* We're done, found all csums in the ordered extents. */
4274 if (mod_len == 0)
4275 return 0;
4276
4277 /* If we're compressed we have to save the entire range of csums. */
4278 if (em->compress_type) {
4279 csum_offset = 0;
4280 csum_len = max(em->block_len, em->orig_block_len);
4281 } else {
4282 csum_offset = mod_start - em->start;
4283 csum_len = mod_len;
4284 }
4285
4286 /* block start is already adjusted for the file extent offset. */
4287 ret = btrfs_lookup_csums_range(trans->fs_info->csum_root,
4288 em->block_start + csum_offset,
4289 em->block_start + csum_offset +
4290 csum_len - 1, &ordered_sums, 0);
4291 if (ret)
4292 return ret;
4293
4294 while (!list_empty(&ordered_sums)) {
4295 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
4296 struct btrfs_ordered_sum,
4297 list);
4298 if (!ret)
4299 ret = log_csums(trans, inode, log_root, sums);
4300 list_del(&sums->list);
4301 kfree(sums);
4302 }
4303
4304 return ret;
4305}
4306
4307static int log_one_extent(struct btrfs_trans_handle *trans,
4308 struct btrfs_inode *inode, struct btrfs_root *root,
4309 const struct extent_map *em,
4310 struct btrfs_path *path,
4311 struct btrfs_log_ctx *ctx)
4312{
4313 struct btrfs_drop_extents_args drop_args = { 0 };
4314 struct btrfs_root *log = root->log_root;
4315 struct btrfs_file_extent_item *fi;
4316 struct extent_buffer *leaf;
4317 struct btrfs_map_token token;
4318 struct btrfs_key key;
4319 u64 extent_offset = em->start - em->orig_start;
4320 u64 block_len;
4321 int ret;
4322
4323 ret = log_extent_csums(trans, inode, log, em, ctx);
4324 if (ret)
4325 return ret;
4326
4327 drop_args.path = path;
4328 drop_args.start = em->start;
4329 drop_args.end = em->start + em->len;
4330 drop_args.replace_extent = true;
4331 drop_args.extent_item_size = sizeof(*fi);
4332 ret = btrfs_drop_extents(trans, log, inode, &drop_args);
4333 if (ret)
4334 return ret;
4335
4336 if (!drop_args.extent_inserted) {
4337 key.objectid = btrfs_ino(inode);
4338 key.type = BTRFS_EXTENT_DATA_KEY;
4339 key.offset = em->start;
4340
4341 ret = btrfs_insert_empty_item(trans, log, path, &key,
4342 sizeof(*fi));
4343 if (ret)
4344 return ret;
4345 }
4346 leaf = path->nodes[0];
4347 btrfs_init_map_token(&token, leaf);
4348 fi = btrfs_item_ptr(leaf, path->slots[0],
4349 struct btrfs_file_extent_item);
4350
4351 btrfs_set_token_file_extent_generation(&token, fi, trans->transid);
4352 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4353 btrfs_set_token_file_extent_type(&token, fi,
4354 BTRFS_FILE_EXTENT_PREALLOC);
4355 else
4356 btrfs_set_token_file_extent_type(&token, fi,
4357 BTRFS_FILE_EXTENT_REG);
4358
4359 block_len = max(em->block_len, em->orig_block_len);
4360 if (em->compress_type != BTRFS_COMPRESS_NONE) {
4361 btrfs_set_token_file_extent_disk_bytenr(&token, fi,
4362 em->block_start);
4363 btrfs_set_token_file_extent_disk_num_bytes(&token, fi, block_len);
4364 } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
4365 btrfs_set_token_file_extent_disk_bytenr(&token, fi,
4366 em->block_start -
4367 extent_offset);
4368 btrfs_set_token_file_extent_disk_num_bytes(&token, fi, block_len);
4369 } else {
4370 btrfs_set_token_file_extent_disk_bytenr(&token, fi, 0);
4371 btrfs_set_token_file_extent_disk_num_bytes(&token, fi, 0);
4372 }
4373
4374 btrfs_set_token_file_extent_offset(&token, fi, extent_offset);
4375 btrfs_set_token_file_extent_num_bytes(&token, fi, em->len);
4376 btrfs_set_token_file_extent_ram_bytes(&token, fi, em->ram_bytes);
4377 btrfs_set_token_file_extent_compression(&token, fi, em->compress_type);
4378 btrfs_set_token_file_extent_encryption(&token, fi, 0);
4379 btrfs_set_token_file_extent_other_encoding(&token, fi, 0);
4380 btrfs_mark_buffer_dirty(leaf);
4381
4382 btrfs_release_path(path);
4383
4384 return ret;
4385}
4386
4387/*
4388 * Log all prealloc extents beyond the inode's i_size to make sure we do not
4389 * lose them after doing a fast fsync and replaying the log. We scan the
4390 * subvolume's root instead of iterating the inode's extent map tree because
4391 * otherwise we can log incorrect extent items based on extent map conversion.
4392 * That can happen due to the fact that extent maps are merged when they
4393 * are not in the extent map tree's list of modified extents.
4394 */
4395static int btrfs_log_prealloc_extents(struct btrfs_trans_handle *trans,
4396 struct btrfs_inode *inode,
4397 struct btrfs_path *path)
4398{
4399 struct btrfs_root *root = inode->root;
4400 struct btrfs_key key;
4401 const u64 i_size = i_size_read(&inode->vfs_inode);
4402 const u64 ino = btrfs_ino(inode);
4403 struct btrfs_path *dst_path = NULL;
4404 bool dropped_extents = false;
4405 u64 truncate_offset = i_size;
4406 struct extent_buffer *leaf;
4407 int slot;
4408 int ins_nr = 0;
4409 int start_slot;
4410 int ret;
4411
4412 if (!(inode->flags & BTRFS_INODE_PREALLOC))
4413 return 0;
4414
4415 key.objectid = ino;
4416 key.type = BTRFS_EXTENT_DATA_KEY;
4417 key.offset = i_size;
4418 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4419 if (ret < 0)
4420 goto out;
4421
4422 /*
4423 * We must check if there is a prealloc extent that starts before the
4424 * i_size and crosses the i_size boundary. This is to ensure later we
4425 * truncate down to the end of that extent and not to the i_size, as
4426 * otherwise we end up losing part of the prealloc extent after a log
4427 * replay and with an implicit hole if there is another prealloc extent
4428 * that starts at an offset beyond i_size.
4429 */
4430 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
4431 if (ret < 0)
4432 goto out;
4433
4434 if (ret == 0) {
4435 struct btrfs_file_extent_item *ei;
4436
4437 leaf = path->nodes[0];
4438 slot = path->slots[0];
4439 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4440
4441 if (btrfs_file_extent_type(leaf, ei) ==
4442 BTRFS_FILE_EXTENT_PREALLOC) {
4443 u64 extent_end;
4444
4445 btrfs_item_key_to_cpu(leaf, &key, slot);
4446 extent_end = key.offset +
4447 btrfs_file_extent_num_bytes(leaf, ei);
4448
4449 if (extent_end > i_size)
4450 truncate_offset = extent_end;
4451 }
4452 } else {
4453 ret = 0;
4454 }
4455
4456 while (true) {
4457 leaf = path->nodes[0];
4458 slot = path->slots[0];
4459
4460 if (slot >= btrfs_header_nritems(leaf)) {
4461 if (ins_nr > 0) {
4462 ret = copy_items(trans, inode, dst_path, path,
4463 start_slot, ins_nr, 1, 0);
4464 if (ret < 0)
4465 goto out;
4466 ins_nr = 0;
4467 }
4468 ret = btrfs_next_leaf(root, path);
4469 if (ret < 0)
4470 goto out;
4471 if (ret > 0) {
4472 ret = 0;
4473 break;
4474 }
4475 continue;
4476 }
4477
4478 btrfs_item_key_to_cpu(leaf, &key, slot);
4479 if (key.objectid > ino)
4480 break;
4481 if (WARN_ON_ONCE(key.objectid < ino) ||
4482 key.type < BTRFS_EXTENT_DATA_KEY ||
4483 key.offset < i_size) {
4484 path->slots[0]++;
4485 continue;
4486 }
4487 if (!dropped_extents) {
4488 /*
4489 * Avoid logging extent items logged in past fsync calls
4490 * and leading to duplicate keys in the log tree.
4491 */
4492 do {
4493 ret = btrfs_truncate_inode_items(trans,
4494 root->log_root,
4495 inode, truncate_offset,
4496 BTRFS_EXTENT_DATA_KEY,
4497 NULL);
4498 } while (ret == -EAGAIN);
4499 if (ret)
4500 goto out;
4501 dropped_extents = true;
4502 }
4503 if (ins_nr == 0)
4504 start_slot = slot;
4505 ins_nr++;
4506 path->slots[0]++;
4507 if (!dst_path) {
4508 dst_path = btrfs_alloc_path();
4509 if (!dst_path) {
4510 ret = -ENOMEM;
4511 goto out;
4512 }
4513 }
4514 }
4515 if (ins_nr > 0)
4516 ret = copy_items(trans, inode, dst_path, path,
4517 start_slot, ins_nr, 1, 0);
4518out:
4519 btrfs_release_path(path);
4520 btrfs_free_path(dst_path);
4521 return ret;
4522}
4523
4524static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
4525 struct btrfs_root *root,
4526 struct btrfs_inode *inode,
4527 struct btrfs_path *path,
4528 struct btrfs_log_ctx *ctx)
4529{
4530 struct btrfs_ordered_extent *ordered;
4531 struct btrfs_ordered_extent *tmp;
4532 struct extent_map *em, *n;
4533 struct list_head extents;
4534 struct extent_map_tree *tree = &inode->extent_tree;
4535 int ret = 0;
4536 int num = 0;
4537
4538 INIT_LIST_HEAD(&extents);
4539
4540 write_lock(&tree->lock);
4541
4542 list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
4543 list_del_init(&em->list);
4544 /*
4545 * Just an arbitrary number, this can be really CPU intensive
4546 * once we start getting a lot of extents, and really once we
4547 * have a bunch of extents we just want to commit since it will
4548 * be faster.
4549 */
4550 if (++num > 32768) {
4551 list_del_init(&tree->modified_extents);
4552 ret = -EFBIG;
4553 goto process;
4554 }
4555
4556 if (em->generation < trans->transid)
4557 continue;
4558
4559 /* We log prealloc extents beyond eof later. */
4560 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) &&
4561 em->start >= i_size_read(&inode->vfs_inode))
4562 continue;
4563
4564 /* Need a ref to keep it from getting evicted from cache */
4565 refcount_inc(&em->refs);
4566 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
4567 list_add_tail(&em->list, &extents);
4568 num++;
4569 }
4570
4571 list_sort(NULL, &extents, extent_cmp);
4572process:
4573 while (!list_empty(&extents)) {
4574 em = list_entry(extents.next, struct extent_map, list);
4575
4576 list_del_init(&em->list);
4577
4578 /*
4579 * If we had an error we just need to delete everybody from our
4580 * private list.
4581 */
4582 if (ret) {
4583 clear_em_logging(tree, em);
4584 free_extent_map(em);
4585 continue;
4586 }
4587
4588 write_unlock(&tree->lock);
4589
4590 ret = log_one_extent(trans, inode, root, em, path, ctx);
4591 write_lock(&tree->lock);
4592 clear_em_logging(tree, em);
4593 free_extent_map(em);
4594 }
4595 WARN_ON(!list_empty(&extents));
4596 write_unlock(&tree->lock);
4597
4598 btrfs_release_path(path);
4599 if (!ret)
4600 ret = btrfs_log_prealloc_extents(trans, inode, path);
4601 if (ret)
4602 return ret;
4603
4604 /*
4605 * We have logged all extents successfully, now make sure the commit of
4606 * the current transaction waits for the ordered extents to complete
4607 * before it commits and wipes out the log trees, otherwise we would
4608 * lose data if an ordered extents completes after the transaction
4609 * commits and a power failure happens after the transaction commit.
4610 */
4611 list_for_each_entry_safe(ordered, tmp, &ctx->ordered_extents, log_list) {
4612 list_del_init(&ordered->log_list);
4613 set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags);
4614
4615 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
4616 spin_lock_irq(&inode->ordered_tree.lock);
4617 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
4618 set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
4619 atomic_inc(&trans->transaction->pending_ordered);
4620 }
4621 spin_unlock_irq(&inode->ordered_tree.lock);
4622 }
4623 btrfs_put_ordered_extent(ordered);
4624 }
4625
4626 return 0;
4627}
4628
4629static int logged_inode_size(struct btrfs_root *log, struct btrfs_inode *inode,
4630 struct btrfs_path *path, u64 *size_ret)
4631{
4632 struct btrfs_key key;
4633 int ret;
4634
4635 key.objectid = btrfs_ino(inode);
4636 key.type = BTRFS_INODE_ITEM_KEY;
4637 key.offset = 0;
4638
4639 ret = btrfs_search_slot(NULL, log, &key, path, 0, 0);
4640 if (ret < 0) {
4641 return ret;
4642 } else if (ret > 0) {
4643 *size_ret = 0;
4644 } else {
4645 struct btrfs_inode_item *item;
4646
4647 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4648 struct btrfs_inode_item);
4649 *size_ret = btrfs_inode_size(path->nodes[0], item);
4650 /*
4651 * If the in-memory inode's i_size is smaller then the inode
4652 * size stored in the btree, return the inode's i_size, so
4653 * that we get a correct inode size after replaying the log
4654 * when before a power failure we had a shrinking truncate
4655 * followed by addition of a new name (rename / new hard link).
4656 * Otherwise return the inode size from the btree, to avoid
4657 * data loss when replaying a log due to previously doing a
4658 * write that expands the inode's size and logging a new name
4659 * immediately after.
4660 */
4661 if (*size_ret > inode->vfs_inode.i_size)
4662 *size_ret = inode->vfs_inode.i_size;
4663 }
4664
4665 btrfs_release_path(path);
4666 return 0;
4667}
4668
4669/*
4670 * At the moment we always log all xattrs. This is to figure out at log replay
4671 * time which xattrs must have their deletion replayed. If a xattr is missing
4672 * in the log tree and exists in the fs/subvol tree, we delete it. This is
4673 * because if a xattr is deleted, the inode is fsynced and a power failure
4674 * happens, causing the log to be replayed the next time the fs is mounted,
4675 * we want the xattr to not exist anymore (same behaviour as other filesystems
4676 * with a journal, ext3/4, xfs, f2fs, etc).
4677 */
4678static int btrfs_log_all_xattrs(struct btrfs_trans_handle *trans,
4679 struct btrfs_root *root,
4680 struct btrfs_inode *inode,
4681 struct btrfs_path *path,
4682 struct btrfs_path *dst_path)
4683{
4684 int ret;
4685 struct btrfs_key key;
4686 const u64 ino = btrfs_ino(inode);
4687 int ins_nr = 0;
4688 int start_slot = 0;
4689 bool found_xattrs = false;
4690
4691 if (test_bit(BTRFS_INODE_NO_XATTRS, &inode->runtime_flags))
4692 return 0;
4693
4694 key.objectid = ino;
4695 key.type = BTRFS_XATTR_ITEM_KEY;
4696 key.offset = 0;
4697
4698 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4699 if (ret < 0)
4700 return ret;
4701
4702 while (true) {
4703 int slot = path->slots[0];
4704 struct extent_buffer *leaf = path->nodes[0];
4705 int nritems = btrfs_header_nritems(leaf);
4706
4707 if (slot >= nritems) {
4708 if (ins_nr > 0) {
4709 ret = copy_items(trans, inode, dst_path, path,
4710 start_slot, ins_nr, 1, 0);
4711 if (ret < 0)
4712 return ret;
4713 ins_nr = 0;
4714 }
4715 ret = btrfs_next_leaf(root, path);
4716 if (ret < 0)
4717 return ret;
4718 else if (ret > 0)
4719 break;
4720 continue;
4721 }
4722
4723 btrfs_item_key_to_cpu(leaf, &key, slot);
4724 if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY)
4725 break;
4726
4727 if (ins_nr == 0)
4728 start_slot = slot;
4729 ins_nr++;
4730 path->slots[0]++;
4731 found_xattrs = true;
4732 cond_resched();
4733 }
4734 if (ins_nr > 0) {
4735 ret = copy_items(trans, inode, dst_path, path,
4736 start_slot, ins_nr, 1, 0);
4737 if (ret < 0)
4738 return ret;
4739 }
4740
4741 if (!found_xattrs)
4742 set_bit(BTRFS_INODE_NO_XATTRS, &inode->runtime_flags);
4743
4744 return 0;
4745}
4746
4747/*
4748 * When using the NO_HOLES feature if we punched a hole that causes the
4749 * deletion of entire leafs or all the extent items of the first leaf (the one
4750 * that contains the inode item and references) we may end up not processing
4751 * any extents, because there are no leafs with a generation matching the
4752 * current transaction that have extent items for our inode. So we need to find
4753 * if any holes exist and then log them. We also need to log holes after any
4754 * truncate operation that changes the inode's size.
4755 */
4756static int btrfs_log_holes(struct btrfs_trans_handle *trans,
4757 struct btrfs_root *root,
4758 struct btrfs_inode *inode,
4759 struct btrfs_path *path)
4760{
4761 struct btrfs_fs_info *fs_info = root->fs_info;
4762 struct btrfs_key key;
4763 const u64 ino = btrfs_ino(inode);
4764 const u64 i_size = i_size_read(&inode->vfs_inode);
4765 u64 prev_extent_end = 0;
4766 int ret;
4767
4768 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || i_size == 0)
4769 return 0;
4770
4771 key.objectid = ino;
4772 key.type = BTRFS_EXTENT_DATA_KEY;
4773 key.offset = 0;
4774
4775 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4776 if (ret < 0)
4777 return ret;
4778
4779 while (true) {
4780 struct extent_buffer *leaf = path->nodes[0];
4781
4782 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
4783 ret = btrfs_next_leaf(root, path);
4784 if (ret < 0)
4785 return ret;
4786 if (ret > 0) {
4787 ret = 0;
4788 break;
4789 }
4790 leaf = path->nodes[0];
4791 }
4792
4793 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4794 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
4795 break;
4796
4797 /* We have a hole, log it. */
4798 if (prev_extent_end < key.offset) {
4799 const u64 hole_len = key.offset - prev_extent_end;
4800
4801 /*
4802 * Release the path to avoid deadlocks with other code
4803 * paths that search the root while holding locks on
4804 * leafs from the log root.
4805 */
4806 btrfs_release_path(path);
4807 ret = btrfs_insert_file_extent(trans, root->log_root,
4808 ino, prev_extent_end, 0,
4809 0, hole_len, 0, hole_len,
4810 0, 0, 0);
4811 if (ret < 0)
4812 return ret;
4813
4814 /*
4815 * Search for the same key again in the root. Since it's
4816 * an extent item and we are holding the inode lock, the
4817 * key must still exist. If it doesn't just emit warning
4818 * and return an error to fall back to a transaction
4819 * commit.
4820 */
4821 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4822 if (ret < 0)
4823 return ret;
4824 if (WARN_ON(ret > 0))
4825 return -ENOENT;
4826 leaf = path->nodes[0];
4827 }
4828
4829 prev_extent_end = btrfs_file_extent_end(path);
4830 path->slots[0]++;
4831 cond_resched();
4832 }
4833
4834 if (prev_extent_end < i_size) {
4835 u64 hole_len;
4836
4837 btrfs_release_path(path);
4838 hole_len = ALIGN(i_size - prev_extent_end, fs_info->sectorsize);
4839 ret = btrfs_insert_file_extent(trans, root->log_root,
4840 ino, prev_extent_end, 0, 0,
4841 hole_len, 0, hole_len,
4842 0, 0, 0);
4843 if (ret < 0)
4844 return ret;
4845 }
4846
4847 return 0;
4848}
4849
4850/*
4851 * When we are logging a new inode X, check if it doesn't have a reference that
4852 * matches the reference from some other inode Y created in a past transaction
4853 * and that was renamed in the current transaction. If we don't do this, then at
4854 * log replay time we can lose inode Y (and all its files if it's a directory):
4855 *
4856 * mkdir /mnt/x
4857 * echo "hello world" > /mnt/x/foobar
4858 * sync
4859 * mv /mnt/x /mnt/y
4860 * mkdir /mnt/x # or touch /mnt/x
4861 * xfs_io -c fsync /mnt/x
4862 * <power fail>
4863 * mount fs, trigger log replay
4864 *
4865 * After the log replay procedure, we would lose the first directory and all its
4866 * files (file foobar).
4867 * For the case where inode Y is not a directory we simply end up losing it:
4868 *
4869 * echo "123" > /mnt/foo
4870 * sync
4871 * mv /mnt/foo /mnt/bar
4872 * echo "abc" > /mnt/foo
4873 * xfs_io -c fsync /mnt/foo
4874 * <power fail>
4875 *
4876 * We also need this for cases where a snapshot entry is replaced by some other
4877 * entry (file or directory) otherwise we end up with an unreplayable log due to
4878 * attempts to delete the snapshot entry (entry of type BTRFS_ROOT_ITEM_KEY) as
4879 * if it were a regular entry:
4880 *
4881 * mkdir /mnt/x
4882 * btrfs subvolume snapshot /mnt /mnt/x/snap
4883 * btrfs subvolume delete /mnt/x/snap
4884 * rmdir /mnt/x
4885 * mkdir /mnt/x
4886 * fsync /mnt/x or fsync some new file inside it
4887 * <power fail>
4888 *
4889 * The snapshot delete, rmdir of x, mkdir of a new x and the fsync all happen in
4890 * the same transaction.
4891 */
4892static int btrfs_check_ref_name_override(struct extent_buffer *eb,
4893 const int slot,
4894 const struct btrfs_key *key,
4895 struct btrfs_inode *inode,
4896 u64 *other_ino, u64 *other_parent)
4897{
4898 int ret;
4899 struct btrfs_path *search_path;
4900 char *name = NULL;
4901 u32 name_len = 0;
4902 u32 item_size = btrfs_item_size_nr(eb, slot);
4903 u32 cur_offset = 0;
4904 unsigned long ptr = btrfs_item_ptr_offset(eb, slot);
4905
4906 search_path = btrfs_alloc_path();
4907 if (!search_path)
4908 return -ENOMEM;
4909 search_path->search_commit_root = 1;
4910 search_path->skip_locking = 1;
4911
4912 while (cur_offset < item_size) {
4913 u64 parent;
4914 u32 this_name_len;
4915 u32 this_len;
4916 unsigned long name_ptr;
4917 struct btrfs_dir_item *di;
4918
4919 if (key->type == BTRFS_INODE_REF_KEY) {
4920 struct btrfs_inode_ref *iref;
4921
4922 iref = (struct btrfs_inode_ref *)(ptr + cur_offset);
4923 parent = key->offset;
4924 this_name_len = btrfs_inode_ref_name_len(eb, iref);
4925 name_ptr = (unsigned long)(iref + 1);
4926 this_len = sizeof(*iref) + this_name_len;
4927 } else {
4928 struct btrfs_inode_extref *extref;
4929
4930 extref = (struct btrfs_inode_extref *)(ptr +
4931 cur_offset);
4932 parent = btrfs_inode_extref_parent(eb, extref);
4933 this_name_len = btrfs_inode_extref_name_len(eb, extref);
4934 name_ptr = (unsigned long)&extref->name;
4935 this_len = sizeof(*extref) + this_name_len;
4936 }
4937
4938 if (this_name_len > name_len) {
4939 char *new_name;
4940
4941 new_name = krealloc(name, this_name_len, GFP_NOFS);
4942 if (!new_name) {
4943 ret = -ENOMEM;
4944 goto out;
4945 }
4946 name_len = this_name_len;
4947 name = new_name;
4948 }
4949
4950 read_extent_buffer(eb, name, name_ptr, this_name_len);
4951 di = btrfs_lookup_dir_item(NULL, inode->root, search_path,
4952 parent, name, this_name_len, 0);
4953 if (di && !IS_ERR(di)) {
4954 struct btrfs_key di_key;
4955
4956 btrfs_dir_item_key_to_cpu(search_path->nodes[0],
4957 di, &di_key);
4958 if (di_key.type == BTRFS_INODE_ITEM_KEY) {
4959 if (di_key.objectid != key->objectid) {
4960 ret = 1;
4961 *other_ino = di_key.objectid;
4962 *other_parent = parent;
4963 } else {
4964 ret = 0;
4965 }
4966 } else {
4967 ret = -EAGAIN;
4968 }
4969 goto out;
4970 } else if (IS_ERR(di)) {
4971 ret = PTR_ERR(di);
4972 goto out;
4973 }
4974 btrfs_release_path(search_path);
4975
4976 cur_offset += this_len;
4977 }
4978 ret = 0;
4979out:
4980 btrfs_free_path(search_path);
4981 kfree(name);
4982 return ret;
4983}
4984
4985struct btrfs_ino_list {
4986 u64 ino;
4987 u64 parent;
4988 struct list_head list;
4989};
4990
4991static int log_conflicting_inodes(struct btrfs_trans_handle *trans,
4992 struct btrfs_root *root,
4993 struct btrfs_path *path,
4994 struct btrfs_log_ctx *ctx,
4995 u64 ino, u64 parent)
4996{
4997 struct btrfs_ino_list *ino_elem;
4998 LIST_HEAD(inode_list);
4999 int ret = 0;
5000
5001 ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
5002 if (!ino_elem)
5003 return -ENOMEM;
5004 ino_elem->ino = ino;
5005 ino_elem->parent = parent;
5006 list_add_tail(&ino_elem->list, &inode_list);
5007
5008 while (!list_empty(&inode_list)) {
5009 struct btrfs_fs_info *fs_info = root->fs_info;
5010 struct btrfs_key key;
5011 struct inode *inode;
5012
5013 ino_elem = list_first_entry(&inode_list, struct btrfs_ino_list,
5014 list);
5015 ino = ino_elem->ino;
5016 parent = ino_elem->parent;
5017 list_del(&ino_elem->list);
5018 kfree(ino_elem);
5019 if (ret)
5020 continue;
5021
5022 btrfs_release_path(path);
5023
5024 inode = btrfs_iget(fs_info->sb, ino, root);
5025 /*
5026 * If the other inode that had a conflicting dir entry was
5027 * deleted in the current transaction, we need to log its parent
5028 * directory.
5029 */
5030 if (IS_ERR(inode)) {
5031 ret = PTR_ERR(inode);
5032 if (ret == -ENOENT) {
5033 inode = btrfs_iget(fs_info->sb, parent, root);
5034 if (IS_ERR(inode)) {
5035 ret = PTR_ERR(inode);
5036 } else {
5037 ret = btrfs_log_inode(trans, root,
5038 BTRFS_I(inode),
5039 LOG_OTHER_INODE_ALL,
5040 ctx);
5041 btrfs_add_delayed_iput(inode);
5042 }
5043 }
5044 continue;
5045 }
5046 /*
5047 * If the inode was already logged skip it - otherwise we can
5048 * hit an infinite loop. Example:
5049 *
5050 * From the commit root (previous transaction) we have the
5051 * following inodes:
5052 *
5053 * inode 257 a directory
5054 * inode 258 with references "zz" and "zz_link" on inode 257
5055 * inode 259 with reference "a" on inode 257
5056 *
5057 * And in the current (uncommitted) transaction we have:
5058 *
5059 * inode 257 a directory, unchanged
5060 * inode 258 with references "a" and "a2" on inode 257
5061 * inode 259 with reference "zz_link" on inode 257
5062 * inode 261 with reference "zz" on inode 257
5063 *
5064 * When logging inode 261 the following infinite loop could
5065 * happen if we don't skip already logged inodes:
5066 *
5067 * - we detect inode 258 as a conflicting inode, with inode 261
5068 * on reference "zz", and log it;
5069 *
5070 * - we detect inode 259 as a conflicting inode, with inode 258
5071 * on reference "a", and log it;
5072 *
5073 * - we detect inode 258 as a conflicting inode, with inode 259
5074 * on reference "zz_link", and log it - again! After this we
5075 * repeat the above steps forever.
5076 */
5077 spin_lock(&BTRFS_I(inode)->lock);
5078 /*
5079 * Check the inode's logged_trans only instead of
5080 * btrfs_inode_in_log(). This is because the last_log_commit of
5081 * the inode is not updated when we only log that it exists and
5082 * it has the full sync bit set (see btrfs_log_inode()).
5083 */
5084 if (BTRFS_I(inode)->logged_trans == trans->transid) {
5085 spin_unlock(&BTRFS_I(inode)->lock);
5086 btrfs_add_delayed_iput(inode);
5087 continue;
5088 }
5089 spin_unlock(&BTRFS_I(inode)->lock);
5090 /*
5091 * We are safe logging the other inode without acquiring its
5092 * lock as long as we log with the LOG_INODE_EXISTS mode. We
5093 * are safe against concurrent renames of the other inode as
5094 * well because during a rename we pin the log and update the
5095 * log with the new name before we unpin it.
5096 */
5097 ret = btrfs_log_inode(trans, root, BTRFS_I(inode),
5098 LOG_OTHER_INODE, ctx);
5099 if (ret) {
5100 btrfs_add_delayed_iput(inode);
5101 continue;
5102 }
5103
5104 key.objectid = ino;
5105 key.type = BTRFS_INODE_REF_KEY;
5106 key.offset = 0;
5107 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5108 if (ret < 0) {
5109 btrfs_add_delayed_iput(inode);
5110 continue;
5111 }
5112
5113 while (true) {
5114 struct extent_buffer *leaf = path->nodes[0];
5115 int slot = path->slots[0];
5116 u64 other_ino = 0;
5117 u64 other_parent = 0;
5118
5119 if (slot >= btrfs_header_nritems(leaf)) {
5120 ret = btrfs_next_leaf(root, path);
5121 if (ret < 0) {
5122 break;
5123 } else if (ret > 0) {
5124 ret = 0;
5125 break;
5126 }
5127 continue;
5128 }
5129
5130 btrfs_item_key_to_cpu(leaf, &key, slot);
5131 if (key.objectid != ino ||
5132 (key.type != BTRFS_INODE_REF_KEY &&
5133 key.type != BTRFS_INODE_EXTREF_KEY)) {
5134 ret = 0;
5135 break;
5136 }
5137
5138 ret = btrfs_check_ref_name_override(leaf, slot, &key,
5139 BTRFS_I(inode), &other_ino,
5140 &other_parent);
5141 if (ret < 0)
5142 break;
5143 if (ret > 0) {
5144 ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
5145 if (!ino_elem) {
5146 ret = -ENOMEM;
5147 break;
5148 }
5149 ino_elem->ino = other_ino;
5150 ino_elem->parent = other_parent;
5151 list_add_tail(&ino_elem->list, &inode_list);
5152 ret = 0;
5153 }
5154 path->slots[0]++;
5155 }
5156 btrfs_add_delayed_iput(inode);
5157 }
5158
5159 return ret;
5160}
5161
5162static int copy_inode_items_to_log(struct btrfs_trans_handle *trans,
5163 struct btrfs_inode *inode,
5164 struct btrfs_key *min_key,
5165 const struct btrfs_key *max_key,
5166 struct btrfs_path *path,
5167 struct btrfs_path *dst_path,
5168 const u64 logged_isize,
5169 const bool recursive_logging,
5170 const int inode_only,
5171 struct btrfs_log_ctx *ctx,
5172 bool *need_log_inode_item)
5173{
5174 struct btrfs_root *root = inode->root;
5175 int ins_start_slot = 0;
5176 int ins_nr = 0;
5177 int ret;
5178
5179 while (1) {
5180 ret = btrfs_search_forward(root, min_key, path, trans->transid);
5181 if (ret < 0)
5182 return ret;
5183 if (ret > 0) {
5184 ret = 0;
5185 break;
5186 }
5187again:
5188 /* Note, ins_nr might be > 0 here, cleanup outside the loop */
5189 if (min_key->objectid != max_key->objectid)
5190 break;
5191 if (min_key->type > max_key->type)
5192 break;
5193
5194 if (min_key->type == BTRFS_INODE_ITEM_KEY)
5195 *need_log_inode_item = false;
5196
5197 if ((min_key->type == BTRFS_INODE_REF_KEY ||
5198 min_key->type == BTRFS_INODE_EXTREF_KEY) &&
5199 inode->generation == trans->transid &&
5200 !recursive_logging) {
5201 u64 other_ino = 0;
5202 u64 other_parent = 0;
5203
5204 ret = btrfs_check_ref_name_override(path->nodes[0],
5205 path->slots[0], min_key, inode,
5206 &other_ino, &other_parent);
5207 if (ret < 0) {
5208 return ret;
5209 } else if (ret > 0 && ctx &&
5210 other_ino != btrfs_ino(BTRFS_I(ctx->inode))) {
5211 if (ins_nr > 0) {
5212 ins_nr++;
5213 } else {
5214 ins_nr = 1;
5215 ins_start_slot = path->slots[0];
5216 }
5217 ret = copy_items(trans, inode, dst_path, path,
5218 ins_start_slot, ins_nr,
5219 inode_only, logged_isize);
5220 if (ret < 0)
5221 return ret;
5222 ins_nr = 0;
5223
5224 ret = log_conflicting_inodes(trans, root, path,
5225 ctx, other_ino, other_parent);
5226 if (ret)
5227 return ret;
5228 btrfs_release_path(path);
5229 goto next_key;
5230 }
5231 }
5232
5233 /* Skip xattrs, we log them later with btrfs_log_all_xattrs() */
5234 if (min_key->type == BTRFS_XATTR_ITEM_KEY) {
5235 if (ins_nr == 0)
5236 goto next_slot;
5237 ret = copy_items(trans, inode, dst_path, path,
5238 ins_start_slot,
5239 ins_nr, inode_only, logged_isize);
5240 if (ret < 0)
5241 return ret;
5242 ins_nr = 0;
5243 goto next_slot;
5244 }
5245
5246 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
5247 ins_nr++;
5248 goto next_slot;
5249 } else if (!ins_nr) {
5250 ins_start_slot = path->slots[0];
5251 ins_nr = 1;
5252 goto next_slot;
5253 }
5254
5255 ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
5256 ins_nr, inode_only, logged_isize);
5257 if (ret < 0)
5258 return ret;
5259 ins_nr = 1;
5260 ins_start_slot = path->slots[0];
5261next_slot:
5262 path->slots[0]++;
5263 if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
5264 btrfs_item_key_to_cpu(path->nodes[0], min_key,
5265 path->slots[0]);
5266 goto again;
5267 }
5268 if (ins_nr) {
5269 ret = copy_items(trans, inode, dst_path, path,
5270 ins_start_slot, ins_nr, inode_only,
5271 logged_isize);
5272 if (ret < 0)
5273 return ret;
5274 ins_nr = 0;
5275 }
5276 btrfs_release_path(path);
5277next_key:
5278 if (min_key->offset < (u64)-1) {
5279 min_key->offset++;
5280 } else if (min_key->type < max_key->type) {
5281 min_key->type++;
5282 min_key->offset = 0;
5283 } else {
5284 break;
5285 }
5286 }
5287 if (ins_nr)
5288 ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
5289 ins_nr, inode_only, logged_isize);
5290
5291 return ret;
5292}
5293
5294/* log a single inode in the tree log.
5295 * At least one parent directory for this inode must exist in the tree
5296 * or be logged already.
5297 *
5298 * Any items from this inode changed by the current transaction are copied
5299 * to the log tree. An extra reference is taken on any extents in this
5300 * file, allowing us to avoid a whole pile of corner cases around logging
5301 * blocks that have been removed from the tree.
5302 *
5303 * See LOG_INODE_ALL and related defines for a description of what inode_only
5304 * does.
5305 *
5306 * This handles both files and directories.
5307 */
5308static int btrfs_log_inode(struct btrfs_trans_handle *trans,
5309 struct btrfs_root *root, struct btrfs_inode *inode,
5310 int inode_only,
5311 struct btrfs_log_ctx *ctx)
5312{
5313 struct btrfs_path *path;
5314 struct btrfs_path *dst_path;
5315 struct btrfs_key min_key;
5316 struct btrfs_key max_key;
5317 struct btrfs_root *log = root->log_root;
5318 int err = 0;
5319 int ret = 0;
5320 bool fast_search = false;
5321 u64 ino = btrfs_ino(inode);
5322 struct extent_map_tree *em_tree = &inode->extent_tree;
5323 u64 logged_isize = 0;
5324 bool need_log_inode_item = true;
5325 bool xattrs_logged = false;
5326 bool recursive_logging = false;
5327
5328 path = btrfs_alloc_path();
5329 if (!path)
5330 return -ENOMEM;
5331 dst_path = btrfs_alloc_path();
5332 if (!dst_path) {
5333 btrfs_free_path(path);
5334 return -ENOMEM;
5335 }
5336
5337 min_key.objectid = ino;
5338 min_key.type = BTRFS_INODE_ITEM_KEY;
5339 min_key.offset = 0;
5340
5341 max_key.objectid = ino;
5342
5343
5344 /* today the code can only do partial logging of directories */
5345 if (S_ISDIR(inode->vfs_inode.i_mode) ||
5346 (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5347 &inode->runtime_flags) &&
5348 inode_only >= LOG_INODE_EXISTS))
5349 max_key.type = BTRFS_XATTR_ITEM_KEY;
5350 else
5351 max_key.type = (u8)-1;
5352 max_key.offset = (u64)-1;
5353
5354 /*
5355 * Only run delayed items if we are a directory. We want to make sure
5356 * all directory indexes hit the fs/subvolume tree so we can find them
5357 * and figure out which index ranges have to be logged.
5358 *
5359 * Otherwise commit the delayed inode only if the full sync flag is set,
5360 * as we want to make sure an up to date version is in the subvolume
5361 * tree so copy_inode_items_to_log() / copy_items() can find it and copy
5362 * it to the log tree. For a non full sync, we always log the inode item
5363 * based on the in-memory struct btrfs_inode which is always up to date.
5364 */
5365 if (S_ISDIR(inode->vfs_inode.i_mode))
5366 ret = btrfs_commit_inode_delayed_items(trans, inode);
5367 else if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags))
5368 ret = btrfs_commit_inode_delayed_inode(inode);
5369
5370 if (ret) {
5371 btrfs_free_path(path);
5372 btrfs_free_path(dst_path);
5373 return ret;
5374 }
5375
5376 if (inode_only == LOG_OTHER_INODE || inode_only == LOG_OTHER_INODE_ALL) {
5377 recursive_logging = true;
5378 if (inode_only == LOG_OTHER_INODE)
5379 inode_only = LOG_INODE_EXISTS;
5380 else
5381 inode_only = LOG_INODE_ALL;
5382 mutex_lock_nested(&inode->log_mutex, SINGLE_DEPTH_NESTING);
5383 } else {
5384 mutex_lock(&inode->log_mutex);
5385 }
5386
5387 /*
5388 * This is for cases where logging a directory could result in losing a
5389 * a file after replaying the log. For example, if we move a file from a
5390 * directory A to a directory B, then fsync directory A, we have no way
5391 * to known the file was moved from A to B, so logging just A would
5392 * result in losing the file after a log replay.
5393 */
5394 if (S_ISDIR(inode->vfs_inode.i_mode) &&
5395 inode_only == LOG_INODE_ALL &&
5396 inode->last_unlink_trans >= trans->transid) {
5397 btrfs_set_log_full_commit(trans);
5398 err = 1;
5399 goto out_unlock;
5400 }
5401
5402 /*
5403 * a brute force approach to making sure we get the most uptodate
5404 * copies of everything.
5405 */
5406 if (S_ISDIR(inode->vfs_inode.i_mode)) {
5407 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
5408
5409 clear_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags);
5410 if (inode_only == LOG_INODE_EXISTS)
5411 max_key_type = BTRFS_XATTR_ITEM_KEY;
5412 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
5413 } else {
5414 if (inode_only == LOG_INODE_EXISTS) {
5415 /*
5416 * Make sure the new inode item we write to the log has
5417 * the same isize as the current one (if it exists).
5418 * This is necessary to prevent data loss after log
5419 * replay, and also to prevent doing a wrong expanding
5420 * truncate - for e.g. create file, write 4K into offset
5421 * 0, fsync, write 4K into offset 4096, add hard link,
5422 * fsync some other file (to sync log), power fail - if
5423 * we use the inode's current i_size, after log replay
5424 * we get a 8Kb file, with the last 4Kb extent as a hole
5425 * (zeroes), as if an expanding truncate happened,
5426 * instead of getting a file of 4Kb only.
5427 */
5428 err = logged_inode_size(log, inode, path, &logged_isize);
5429 if (err)
5430 goto out_unlock;
5431 }
5432 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5433 &inode->runtime_flags)) {
5434 if (inode_only == LOG_INODE_EXISTS) {
5435 max_key.type = BTRFS_XATTR_ITEM_KEY;
5436 ret = drop_objectid_items(trans, log, path, ino,
5437 max_key.type);
5438 } else {
5439 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5440 &inode->runtime_flags);
5441 clear_bit(BTRFS_INODE_COPY_EVERYTHING,
5442 &inode->runtime_flags);
5443 while(1) {
5444 ret = btrfs_truncate_inode_items(trans,
5445 log, inode, 0, 0, NULL);
5446 if (ret != -EAGAIN)
5447 break;
5448 }
5449 }
5450 } else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
5451 &inode->runtime_flags) ||
5452 inode_only == LOG_INODE_EXISTS) {
5453 if (inode_only == LOG_INODE_ALL)
5454 fast_search = true;
5455 max_key.type = BTRFS_XATTR_ITEM_KEY;
5456 ret = drop_objectid_items(trans, log, path, ino,
5457 max_key.type);
5458 } else {
5459 if (inode_only == LOG_INODE_ALL)
5460 fast_search = true;
5461 goto log_extents;
5462 }
5463
5464 }
5465 if (ret) {
5466 err = ret;
5467 goto out_unlock;
5468 }
5469
5470 err = copy_inode_items_to_log(trans, inode, &min_key, &max_key,
5471 path, dst_path, logged_isize,
5472 recursive_logging, inode_only, ctx,
5473 &need_log_inode_item);
5474 if (err)
5475 goto out_unlock;
5476
5477 btrfs_release_path(path);
5478 btrfs_release_path(dst_path);
5479 err = btrfs_log_all_xattrs(trans, root, inode, path, dst_path);
5480 if (err)
5481 goto out_unlock;
5482 xattrs_logged = true;
5483 if (max_key.type >= BTRFS_EXTENT_DATA_KEY && !fast_search) {
5484 btrfs_release_path(path);
5485 btrfs_release_path(dst_path);
5486 err = btrfs_log_holes(trans, root, inode, path);
5487 if (err)
5488 goto out_unlock;
5489 }
5490log_extents:
5491 btrfs_release_path(path);
5492 btrfs_release_path(dst_path);
5493 if (need_log_inode_item) {
5494 err = log_inode_item(trans, log, dst_path, inode);
5495 if (err)
5496 goto out_unlock;
5497 /*
5498 * If we are doing a fast fsync and the inode was logged before
5499 * in this transaction, we don't need to log the xattrs because
5500 * they were logged before. If xattrs were added, changed or
5501 * deleted since the last time we logged the inode, then we have
5502 * already logged them because the inode had the runtime flag
5503 * BTRFS_INODE_COPY_EVERYTHING set.
5504 */
5505 if (!xattrs_logged && inode->logged_trans < trans->transid) {
5506 err = btrfs_log_all_xattrs(trans, root, inode, path,
5507 dst_path);
5508 if (err)
5509 goto out_unlock;
5510 btrfs_release_path(path);
5511 }
5512 }
5513 if (fast_search) {
5514 ret = btrfs_log_changed_extents(trans, root, inode, dst_path,
5515 ctx);
5516 if (ret) {
5517 err = ret;
5518 goto out_unlock;
5519 }
5520 } else if (inode_only == LOG_INODE_ALL) {
5521 struct extent_map *em, *n;
5522
5523 write_lock(&em_tree->lock);
5524 list_for_each_entry_safe(em, n, &em_tree->modified_extents, list)
5525 list_del_init(&em->list);
5526 write_unlock(&em_tree->lock);
5527 }
5528
5529 if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->vfs_inode.i_mode)) {
5530 ret = log_directory_changes(trans, root, inode, path, dst_path,
5531 ctx);
5532 if (ret) {
5533 err = ret;
5534 goto out_unlock;
5535 }
5536 }
5537
5538 /*
5539 * If we are logging that an ancestor inode exists as part of logging a
5540 * new name from a link or rename operation, don't mark the inode as
5541 * logged - otherwise if an explicit fsync is made against an ancestor,
5542 * the fsync considers the inode in the log and doesn't sync the log,
5543 * resulting in the ancestor missing after a power failure unless the
5544 * log was synced as part of an fsync against any other unrelated inode.
5545 * So keep it simple for this case and just don't flag the ancestors as
5546 * logged.
5547 */
5548 if (!ctx ||
5549 !(S_ISDIR(inode->vfs_inode.i_mode) && ctx->logging_new_name &&
5550 &inode->vfs_inode != ctx->inode)) {
5551 spin_lock(&inode->lock);
5552 inode->logged_trans = trans->transid;
5553 /*
5554 * Don't update last_log_commit if we logged that an inode exists.
5555 * We do this for two reasons:
5556 *
5557 * 1) We might have had buffered writes to this inode that were
5558 * flushed and had their ordered extents completed in this
5559 * transaction, but we did not previously log the inode with
5560 * LOG_INODE_ALL. Later the inode was evicted and after that
5561 * it was loaded again and this LOG_INODE_EXISTS log operation
5562 * happened. We must make sure that if an explicit fsync against
5563 * the inode is performed later, it logs the new extents, an
5564 * updated inode item, etc, and syncs the log. The same logic
5565 * applies to direct IO writes instead of buffered writes.
5566 *
5567 * 2) When we log the inode with LOG_INODE_EXISTS, its inode item
5568 * is logged with an i_size of 0 or whatever value was logged
5569 * before. If later the i_size of the inode is increased by a
5570 * truncate operation, the log is synced through an fsync of
5571 * some other inode and then finally an explicit fsync against
5572 * this inode is made, we must make sure this fsync logs the
5573 * inode with the new i_size, the hole between old i_size and
5574 * the new i_size, and syncs the log.
5575 */
5576 if (inode_only != LOG_INODE_EXISTS)
5577 inode->last_log_commit = inode->last_sub_trans;
5578 spin_unlock(&inode->lock);
5579 }
5580out_unlock:
5581 mutex_unlock(&inode->log_mutex);
5582
5583 btrfs_free_path(path);
5584 btrfs_free_path(dst_path);
5585 return err;
5586}
5587
5588/*
5589 * Check if we need to log an inode. This is used in contexts where while
5590 * logging an inode we need to log another inode (either that it exists or in
5591 * full mode). This is used instead of btrfs_inode_in_log() because the later
5592 * requires the inode to be in the log and have the log transaction committed,
5593 * while here we do not care if the log transaction was already committed - our
5594 * caller will commit the log later - and we want to avoid logging an inode
5595 * multiple times when multiple tasks have joined the same log transaction.
5596 */
5597static bool need_log_inode(struct btrfs_trans_handle *trans,
5598 struct btrfs_inode *inode)
5599{
5600 /*
5601 * If this inode does not have new/updated/deleted xattrs since the last
5602 * time it was logged and is flagged as logged in the current transaction,
5603 * we can skip logging it. As for new/deleted names, those are updated in
5604 * the log by link/unlink/rename operations.
5605 * In case the inode was logged and then evicted and reloaded, its
5606 * logged_trans will be 0, in which case we have to fully log it since
5607 * logged_trans is a transient field, not persisted.
5608 */
5609 if (inode->logged_trans == trans->transid &&
5610 !test_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags))
5611 return false;
5612
5613 return true;
5614}
5615
5616struct btrfs_dir_list {
5617 u64 ino;
5618 struct list_head list;
5619};
5620
5621/*
5622 * Log the inodes of the new dentries of a directory. See log_dir_items() for
5623 * details about the why it is needed.
5624 * This is a recursive operation - if an existing dentry corresponds to a
5625 * directory, that directory's new entries are logged too (same behaviour as
5626 * ext3/4, xfs, f2fs, reiserfs, nilfs2). Note that when logging the inodes
5627 * the dentries point to we do not lock their i_mutex, otherwise lockdep
5628 * complains about the following circular lock dependency / possible deadlock:
5629 *
5630 * CPU0 CPU1
5631 * ---- ----
5632 * lock(&type->i_mutex_dir_key#3/2);
5633 * lock(sb_internal#2);
5634 * lock(&type->i_mutex_dir_key#3/2);
5635 * lock(&sb->s_type->i_mutex_key#14);
5636 *
5637 * Where sb_internal is the lock (a counter that works as a lock) acquired by
5638 * sb_start_intwrite() in btrfs_start_transaction().
5639 * Not locking i_mutex of the inodes is still safe because:
5640 *
5641 * 1) For regular files we log with a mode of LOG_INODE_EXISTS. It's possible
5642 * that while logging the inode new references (names) are added or removed
5643 * from the inode, leaving the logged inode item with a link count that does
5644 * not match the number of logged inode reference items. This is fine because
5645 * at log replay time we compute the real number of links and correct the
5646 * link count in the inode item (see replay_one_buffer() and
5647 * link_to_fixup_dir());
5648 *
5649 * 2) For directories we log with a mode of LOG_INODE_ALL. It's possible that
5650 * while logging the inode's items new items with keys BTRFS_DIR_ITEM_KEY and
5651 * BTRFS_DIR_INDEX_KEY are added to fs/subvol tree and the logged inode item
5652 * has a size that doesn't match the sum of the lengths of all the logged
5653 * names. This does not result in a problem because if a dir_item key is
5654 * logged but its matching dir_index key is not logged, at log replay time we
5655 * don't use it to replay the respective name (see replay_one_name()). On the
5656 * other hand if only the dir_index key ends up being logged, the respective
5657 * name is added to the fs/subvol tree with both the dir_item and dir_index
5658 * keys created (see replay_one_name()).
5659 * The directory's inode item with a wrong i_size is not a problem as well,
5660 * since we don't use it at log replay time to set the i_size in the inode
5661 * item of the fs/subvol tree (see overwrite_item()).
5662 */
5663static int log_new_dir_dentries(struct btrfs_trans_handle *trans,
5664 struct btrfs_root *root,
5665 struct btrfs_inode *start_inode,
5666 struct btrfs_log_ctx *ctx)
5667{
5668 struct btrfs_fs_info *fs_info = root->fs_info;
5669 struct btrfs_root *log = root->log_root;
5670 struct btrfs_path *path;
5671 LIST_HEAD(dir_list);
5672 struct btrfs_dir_list *dir_elem;
5673 int ret = 0;
5674
5675 path = btrfs_alloc_path();
5676 if (!path)
5677 return -ENOMEM;
5678
5679 dir_elem = kmalloc(sizeof(*dir_elem), GFP_NOFS);
5680 if (!dir_elem) {
5681 btrfs_free_path(path);
5682 return -ENOMEM;
5683 }
5684 dir_elem->ino = btrfs_ino(start_inode);
5685 list_add_tail(&dir_elem->list, &dir_list);
5686
5687 while (!list_empty(&dir_list)) {
5688 struct extent_buffer *leaf;
5689 struct btrfs_key min_key;
5690 int nritems;
5691 int i;
5692
5693 dir_elem = list_first_entry(&dir_list, struct btrfs_dir_list,
5694 list);
5695 if (ret)
5696 goto next_dir_inode;
5697
5698 min_key.objectid = dir_elem->ino;
5699 min_key.type = BTRFS_DIR_ITEM_KEY;
5700 min_key.offset = 0;
5701again:
5702 btrfs_release_path(path);
5703 ret = btrfs_search_forward(log, &min_key, path, trans->transid);
5704 if (ret < 0) {
5705 goto next_dir_inode;
5706 } else if (ret > 0) {
5707 ret = 0;
5708 goto next_dir_inode;
5709 }
5710
5711process_leaf:
5712 leaf = path->nodes[0];
5713 nritems = btrfs_header_nritems(leaf);
5714 for (i = path->slots[0]; i < nritems; i++) {
5715 struct btrfs_dir_item *di;
5716 struct btrfs_key di_key;
5717 struct inode *di_inode;
5718 struct btrfs_dir_list *new_dir_elem;
5719 int log_mode = LOG_INODE_EXISTS;
5720 int type;
5721
5722 btrfs_item_key_to_cpu(leaf, &min_key, i);
5723 if (min_key.objectid != dir_elem->ino ||
5724 min_key.type != BTRFS_DIR_ITEM_KEY)
5725 goto next_dir_inode;
5726
5727 di = btrfs_item_ptr(leaf, i, struct btrfs_dir_item);
5728 type = btrfs_dir_type(leaf, di);
5729 if (btrfs_dir_transid(leaf, di) < trans->transid &&
5730 type != BTRFS_FT_DIR)
5731 continue;
5732 btrfs_dir_item_key_to_cpu(leaf, di, &di_key);
5733 if (di_key.type == BTRFS_ROOT_ITEM_KEY)
5734 continue;
5735
5736 btrfs_release_path(path);
5737 di_inode = btrfs_iget(fs_info->sb, di_key.objectid, root);
5738 if (IS_ERR(di_inode)) {
5739 ret = PTR_ERR(di_inode);
5740 goto next_dir_inode;
5741 }
5742
5743 if (!need_log_inode(trans, BTRFS_I(di_inode))) {
5744 btrfs_add_delayed_iput(di_inode);
5745 break;
5746 }
5747
5748 ctx->log_new_dentries = false;
5749 if (type == BTRFS_FT_DIR || type == BTRFS_FT_SYMLINK)
5750 log_mode = LOG_INODE_ALL;
5751 ret = btrfs_log_inode(trans, root, BTRFS_I(di_inode),
5752 log_mode, ctx);
5753 btrfs_add_delayed_iput(di_inode);
5754 if (ret)
5755 goto next_dir_inode;
5756 if (ctx->log_new_dentries) {
5757 new_dir_elem = kmalloc(sizeof(*new_dir_elem),
5758 GFP_NOFS);
5759 if (!new_dir_elem) {
5760 ret = -ENOMEM;
5761 goto next_dir_inode;
5762 }
5763 new_dir_elem->ino = di_key.objectid;
5764 list_add_tail(&new_dir_elem->list, &dir_list);
5765 }
5766 break;
5767 }
5768 if (i == nritems) {
5769 ret = btrfs_next_leaf(log, path);
5770 if (ret < 0) {
5771 goto next_dir_inode;
5772 } else if (ret > 0) {
5773 ret = 0;
5774 goto next_dir_inode;
5775 }
5776 goto process_leaf;
5777 }
5778 if (min_key.offset < (u64)-1) {
5779 min_key.offset++;
5780 goto again;
5781 }
5782next_dir_inode:
5783 list_del(&dir_elem->list);
5784 kfree(dir_elem);
5785 }
5786
5787 btrfs_free_path(path);
5788 return ret;
5789}
5790
5791static int btrfs_log_all_parents(struct btrfs_trans_handle *trans,
5792 struct btrfs_inode *inode,
5793 struct btrfs_log_ctx *ctx)
5794{
5795 struct btrfs_fs_info *fs_info = trans->fs_info;
5796 int ret;
5797 struct btrfs_path *path;
5798 struct btrfs_key key;
5799 struct btrfs_root *root = inode->root;
5800 const u64 ino = btrfs_ino(inode);
5801
5802 path = btrfs_alloc_path();
5803 if (!path)
5804 return -ENOMEM;
5805 path->skip_locking = 1;
5806 path->search_commit_root = 1;
5807
5808 key.objectid = ino;
5809 key.type = BTRFS_INODE_REF_KEY;
5810 key.offset = 0;
5811 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5812 if (ret < 0)
5813 goto out;
5814
5815 while (true) {
5816 struct extent_buffer *leaf = path->nodes[0];
5817 int slot = path->slots[0];
5818 u32 cur_offset = 0;
5819 u32 item_size;
5820 unsigned long ptr;
5821
5822 if (slot >= btrfs_header_nritems(leaf)) {
5823 ret = btrfs_next_leaf(root, path);
5824 if (ret < 0)
5825 goto out;
5826 else if (ret > 0)
5827 break;
5828 continue;
5829 }
5830
5831 btrfs_item_key_to_cpu(leaf, &key, slot);
5832 /* BTRFS_INODE_EXTREF_KEY is BTRFS_INODE_REF_KEY + 1 */
5833 if (key.objectid != ino || key.type > BTRFS_INODE_EXTREF_KEY)
5834 break;
5835
5836 item_size = btrfs_item_size_nr(leaf, slot);
5837 ptr = btrfs_item_ptr_offset(leaf, slot);
5838 while (cur_offset < item_size) {
5839 struct btrfs_key inode_key;
5840 struct inode *dir_inode;
5841
5842 inode_key.type = BTRFS_INODE_ITEM_KEY;
5843 inode_key.offset = 0;
5844
5845 if (key.type == BTRFS_INODE_EXTREF_KEY) {
5846 struct btrfs_inode_extref *extref;
5847
5848 extref = (struct btrfs_inode_extref *)
5849 (ptr + cur_offset);
5850 inode_key.objectid = btrfs_inode_extref_parent(
5851 leaf, extref);
5852 cur_offset += sizeof(*extref);
5853 cur_offset += btrfs_inode_extref_name_len(leaf,
5854 extref);
5855 } else {
5856 inode_key.objectid = key.offset;
5857 cur_offset = item_size;
5858 }
5859
5860 dir_inode = btrfs_iget(fs_info->sb, inode_key.objectid,
5861 root);
5862 /*
5863 * If the parent inode was deleted, return an error to
5864 * fallback to a transaction commit. This is to prevent
5865 * getting an inode that was moved from one parent A to
5866 * a parent B, got its former parent A deleted and then
5867 * it got fsync'ed, from existing at both parents after
5868 * a log replay (and the old parent still existing).
5869 * Example:
5870 *
5871 * mkdir /mnt/A
5872 * mkdir /mnt/B
5873 * touch /mnt/B/bar
5874 * sync
5875 * mv /mnt/B/bar /mnt/A/bar
5876 * mv -T /mnt/A /mnt/B
5877 * fsync /mnt/B/bar
5878 * <power fail>
5879 *
5880 * If we ignore the old parent B which got deleted,
5881 * after a log replay we would have file bar linked
5882 * at both parents and the old parent B would still
5883 * exist.
5884 */
5885 if (IS_ERR(dir_inode)) {
5886 ret = PTR_ERR(dir_inode);
5887 goto out;
5888 }
5889
5890 if (!need_log_inode(trans, BTRFS_I(dir_inode))) {
5891 btrfs_add_delayed_iput(dir_inode);
5892 continue;
5893 }
5894
5895 if (ctx)
5896 ctx->log_new_dentries = false;
5897 ret = btrfs_log_inode(trans, root, BTRFS_I(dir_inode),
5898 LOG_INODE_ALL, ctx);
5899 if (!ret && ctx && ctx->log_new_dentries)
5900 ret = log_new_dir_dentries(trans, root,
5901 BTRFS_I(dir_inode), ctx);
5902 btrfs_add_delayed_iput(dir_inode);
5903 if (ret)
5904 goto out;
5905 }
5906 path->slots[0]++;
5907 }
5908 ret = 0;
5909out:
5910 btrfs_free_path(path);
5911 return ret;
5912}
5913
5914static int log_new_ancestors(struct btrfs_trans_handle *trans,
5915 struct btrfs_root *root,
5916 struct btrfs_path *path,
5917 struct btrfs_log_ctx *ctx)
5918{
5919 struct btrfs_key found_key;
5920
5921 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
5922
5923 while (true) {
5924 struct btrfs_fs_info *fs_info = root->fs_info;
5925 struct extent_buffer *leaf = path->nodes[0];
5926 int slot = path->slots[0];
5927 struct btrfs_key search_key;
5928 struct inode *inode;
5929 u64 ino;
5930 int ret = 0;
5931
5932 btrfs_release_path(path);
5933
5934 ino = found_key.offset;
5935
5936 search_key.objectid = found_key.offset;
5937 search_key.type = BTRFS_INODE_ITEM_KEY;
5938 search_key.offset = 0;
5939 inode = btrfs_iget(fs_info->sb, ino, root);
5940 if (IS_ERR(inode))
5941 return PTR_ERR(inode);
5942
5943 if (BTRFS_I(inode)->generation >= trans->transid &&
5944 need_log_inode(trans, BTRFS_I(inode)))
5945 ret = btrfs_log_inode(trans, root, BTRFS_I(inode),
5946 LOG_INODE_EXISTS, ctx);
5947 btrfs_add_delayed_iput(inode);
5948 if (ret)
5949 return ret;
5950
5951 if (search_key.objectid == BTRFS_FIRST_FREE_OBJECTID)
5952 break;
5953
5954 search_key.type = BTRFS_INODE_REF_KEY;
5955 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
5956 if (ret < 0)
5957 return ret;
5958
5959 leaf = path->nodes[0];
5960 slot = path->slots[0];
5961 if (slot >= btrfs_header_nritems(leaf)) {
5962 ret = btrfs_next_leaf(root, path);
5963 if (ret < 0)
5964 return ret;
5965 else if (ret > 0)
5966 return -ENOENT;
5967 leaf = path->nodes[0];
5968 slot = path->slots[0];
5969 }
5970
5971 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5972 if (found_key.objectid != search_key.objectid ||
5973 found_key.type != BTRFS_INODE_REF_KEY)
5974 return -ENOENT;
5975 }
5976 return 0;
5977}
5978
5979static int log_new_ancestors_fast(struct btrfs_trans_handle *trans,
5980 struct btrfs_inode *inode,
5981 struct dentry *parent,
5982 struct btrfs_log_ctx *ctx)
5983{
5984 struct btrfs_root *root = inode->root;
5985 struct dentry *old_parent = NULL;
5986 struct super_block *sb = inode->vfs_inode.i_sb;
5987 int ret = 0;
5988
5989 while (true) {
5990 if (!parent || d_really_is_negative(parent) ||
5991 sb != parent->d_sb)
5992 break;
5993
5994 inode = BTRFS_I(d_inode(parent));
5995 if (root != inode->root)
5996 break;
5997
5998 if (inode->generation >= trans->transid &&
5999 need_log_inode(trans, inode)) {
6000 ret = btrfs_log_inode(trans, root, inode,
6001 LOG_INODE_EXISTS, ctx);
6002 if (ret)
6003 break;
6004 }
6005 if (IS_ROOT(parent))
6006 break;
6007
6008 parent = dget_parent(parent);
6009 dput(old_parent);
6010 old_parent = parent;
6011 }
6012 dput(old_parent);
6013
6014 return ret;
6015}
6016
6017static int log_all_new_ancestors(struct btrfs_trans_handle *trans,
6018 struct btrfs_inode *inode,
6019 struct dentry *parent,
6020 struct btrfs_log_ctx *ctx)
6021{
6022 struct btrfs_root *root = inode->root;
6023 const u64 ino = btrfs_ino(inode);
6024 struct btrfs_path *path;
6025 struct btrfs_key search_key;
6026 int ret;
6027
6028 /*
6029 * For a single hard link case, go through a fast path that does not
6030 * need to iterate the fs/subvolume tree.
6031 */
6032 if (inode->vfs_inode.i_nlink < 2)
6033 return log_new_ancestors_fast(trans, inode, parent, ctx);
6034
6035 path = btrfs_alloc_path();
6036 if (!path)
6037 return -ENOMEM;
6038
6039 search_key.objectid = ino;
6040 search_key.type = BTRFS_INODE_REF_KEY;
6041 search_key.offset = 0;
6042again:
6043 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
6044 if (ret < 0)
6045 goto out;
6046 if (ret == 0)
6047 path->slots[0]++;
6048
6049 while (true) {
6050 struct extent_buffer *leaf = path->nodes[0];
6051 int slot = path->slots[0];
6052 struct btrfs_key found_key;
6053
6054 if (slot >= btrfs_header_nritems(leaf)) {
6055 ret = btrfs_next_leaf(root, path);
6056 if (ret < 0)
6057 goto out;
6058 else if (ret > 0)
6059 break;
6060 continue;
6061 }
6062
6063 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6064 if (found_key.objectid != ino ||
6065 found_key.type > BTRFS_INODE_EXTREF_KEY)
6066 break;
6067
6068 /*
6069 * Don't deal with extended references because they are rare
6070 * cases and too complex to deal with (we would need to keep
6071 * track of which subitem we are processing for each item in
6072 * this loop, etc). So just return some error to fallback to
6073 * a transaction commit.
6074 */
6075 if (found_key.type == BTRFS_INODE_EXTREF_KEY) {
6076 ret = -EMLINK;
6077 goto out;
6078 }
6079
6080 /*
6081 * Logging ancestors needs to do more searches on the fs/subvol
6082 * tree, so it releases the path as needed to avoid deadlocks.
6083 * Keep track of the last inode ref key and resume from that key
6084 * after logging all new ancestors for the current hard link.
6085 */
6086 memcpy(&search_key, &found_key, sizeof(search_key));
6087
6088 ret = log_new_ancestors(trans, root, path, ctx);
6089 if (ret)
6090 goto out;
6091 btrfs_release_path(path);
6092 goto again;
6093 }
6094 ret = 0;
6095out:
6096 btrfs_free_path(path);
6097 return ret;
6098}
6099
6100/*
6101 * helper function around btrfs_log_inode to make sure newly created
6102 * parent directories also end up in the log. A minimal inode and backref
6103 * only logging is done of any parent directories that are older than
6104 * the last committed transaction
6105 */
6106static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
6107 struct btrfs_inode *inode,
6108 struct dentry *parent,
6109 int inode_only,
6110 struct btrfs_log_ctx *ctx)
6111{
6112 struct btrfs_root *root = inode->root;
6113 struct btrfs_fs_info *fs_info = root->fs_info;
6114 int ret = 0;
6115 bool log_dentries = false;
6116
6117 if (btrfs_test_opt(fs_info, NOTREELOG)) {
6118 ret = 1;
6119 goto end_no_trans;
6120 }
6121
6122 if (btrfs_root_refs(&root->root_item) == 0) {
6123 ret = 1;
6124 goto end_no_trans;
6125 }
6126
6127 /*
6128 * Skip already logged inodes or inodes corresponding to tmpfiles
6129 * (since logging them is pointless, a link count of 0 means they
6130 * will never be accessible).
6131 */
6132 if ((btrfs_inode_in_log(inode, trans->transid) &&
6133 list_empty(&ctx->ordered_extents)) ||
6134 inode->vfs_inode.i_nlink == 0) {
6135 ret = BTRFS_NO_LOG_SYNC;
6136 goto end_no_trans;
6137 }
6138
6139 ret = start_log_trans(trans, root, ctx);
6140 if (ret)
6141 goto end_no_trans;
6142
6143 ret = btrfs_log_inode(trans, root, inode, inode_only, ctx);
6144 if (ret)
6145 goto end_trans;
6146
6147 /*
6148 * for regular files, if its inode is already on disk, we don't
6149 * have to worry about the parents at all. This is because
6150 * we can use the last_unlink_trans field to record renames
6151 * and other fun in this file.
6152 */
6153 if (S_ISREG(inode->vfs_inode.i_mode) &&
6154 inode->generation < trans->transid &&
6155 inode->last_unlink_trans < trans->transid) {
6156 ret = 0;
6157 goto end_trans;
6158 }
6159
6160 if (S_ISDIR(inode->vfs_inode.i_mode) && ctx && ctx->log_new_dentries)
6161 log_dentries = true;
6162
6163 /*
6164 * On unlink we must make sure all our current and old parent directory
6165 * inodes are fully logged. This is to prevent leaving dangling
6166 * directory index entries in directories that were our parents but are
6167 * not anymore. Not doing this results in old parent directory being
6168 * impossible to delete after log replay (rmdir will always fail with
6169 * error -ENOTEMPTY).
6170 *
6171 * Example 1:
6172 *
6173 * mkdir testdir
6174 * touch testdir/foo
6175 * ln testdir/foo testdir/bar
6176 * sync
6177 * unlink testdir/bar
6178 * xfs_io -c fsync testdir/foo
6179 * <power failure>
6180 * mount fs, triggers log replay
6181 *
6182 * If we don't log the parent directory (testdir), after log replay the
6183 * directory still has an entry pointing to the file inode using the bar
6184 * name, but a matching BTRFS_INODE_[REF|EXTREF]_KEY does not exist and
6185 * the file inode has a link count of 1.
6186 *
6187 * Example 2:
6188 *
6189 * mkdir testdir
6190 * touch foo
6191 * ln foo testdir/foo2
6192 * ln foo testdir/foo3
6193 * sync
6194 * unlink testdir/foo3
6195 * xfs_io -c fsync foo
6196 * <power failure>
6197 * mount fs, triggers log replay
6198 *
6199 * Similar as the first example, after log replay the parent directory
6200 * testdir still has an entry pointing to the inode file with name foo3
6201 * but the file inode does not have a matching BTRFS_INODE_REF_KEY item
6202 * and has a link count of 2.
6203 */
6204 if (inode->last_unlink_trans >= trans->transid) {
6205 ret = btrfs_log_all_parents(trans, inode, ctx);
6206 if (ret)
6207 goto end_trans;
6208 }
6209
6210 ret = log_all_new_ancestors(trans, inode, parent, ctx);
6211 if (ret)
6212 goto end_trans;
6213
6214 if (log_dentries)
6215 ret = log_new_dir_dentries(trans, root, inode, ctx);
6216 else
6217 ret = 0;
6218end_trans:
6219 if (ret < 0) {
6220 btrfs_set_log_full_commit(trans);
6221 ret = 1;
6222 }
6223
6224 if (ret)
6225 btrfs_remove_log_ctx(root, ctx);
6226 btrfs_end_log_trans(root);
6227end_no_trans:
6228 return ret;
6229}
6230
6231/*
6232 * it is not safe to log dentry if the chunk root has added new
6233 * chunks. This returns 0 if the dentry was logged, and 1 otherwise.
6234 * If this returns 1, you must commit the transaction to safely get your
6235 * data on disk.
6236 */
6237int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
6238 struct dentry *dentry,
6239 struct btrfs_log_ctx *ctx)
6240{
6241 struct dentry *parent = dget_parent(dentry);
6242 int ret;
6243
6244 ret = btrfs_log_inode_parent(trans, BTRFS_I(d_inode(dentry)), parent,
6245 LOG_INODE_ALL, ctx);
6246 dput(parent);
6247
6248 return ret;
6249}
6250
6251/*
6252 * should be called during mount to recover any replay any log trees
6253 * from the FS
6254 */
6255int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
6256{
6257 int ret;
6258 struct btrfs_path *path;
6259 struct btrfs_trans_handle *trans;
6260 struct btrfs_key key;
6261 struct btrfs_key found_key;
6262 struct btrfs_root *log;
6263 struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
6264 struct walk_control wc = {
6265 .process_func = process_one_buffer,
6266 .stage = LOG_WALK_PIN_ONLY,
6267 };
6268
6269 path = btrfs_alloc_path();
6270 if (!path)
6271 return -ENOMEM;
6272
6273 set_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
6274
6275 trans = btrfs_start_transaction(fs_info->tree_root, 0);
6276 if (IS_ERR(trans)) {
6277 ret = PTR_ERR(trans);
6278 goto error;
6279 }
6280
6281 wc.trans = trans;
6282 wc.pin = 1;
6283
6284 ret = walk_log_tree(trans, log_root_tree, &wc);
6285 if (ret) {
6286 btrfs_handle_fs_error(fs_info, ret,
6287 "Failed to pin buffers while recovering log root tree.");
6288 goto error;
6289 }
6290
6291again:
6292 key.objectid = BTRFS_TREE_LOG_OBJECTID;
6293 key.offset = (u64)-1;
6294 key.type = BTRFS_ROOT_ITEM_KEY;
6295
6296 while (1) {
6297 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
6298
6299 if (ret < 0) {
6300 btrfs_handle_fs_error(fs_info, ret,
6301 "Couldn't find tree log root.");
6302 goto error;
6303 }
6304 if (ret > 0) {
6305 if (path->slots[0] == 0)
6306 break;
6307 path->slots[0]--;
6308 }
6309 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
6310 path->slots[0]);
6311 btrfs_release_path(path);
6312 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
6313 break;
6314
6315 log = btrfs_read_tree_root(log_root_tree, &found_key);
6316 if (IS_ERR(log)) {
6317 ret = PTR_ERR(log);
6318 btrfs_handle_fs_error(fs_info, ret,
6319 "Couldn't read tree log root.");
6320 goto error;
6321 }
6322
6323 wc.replay_dest = btrfs_get_fs_root(fs_info, found_key.offset,
6324 true);
6325 if (IS_ERR(wc.replay_dest)) {
6326 ret = PTR_ERR(wc.replay_dest);
6327
6328 /*
6329 * We didn't find the subvol, likely because it was
6330 * deleted. This is ok, simply skip this log and go to
6331 * the next one.
6332 *
6333 * We need to exclude the root because we can't have
6334 * other log replays overwriting this log as we'll read
6335 * it back in a few more times. This will keep our
6336 * block from being modified, and we'll just bail for
6337 * each subsequent pass.
6338 */
6339 if (ret == -ENOENT)
6340 ret = btrfs_pin_extent_for_log_replay(trans,
6341 log->node->start,
6342 log->node->len);
6343 btrfs_put_root(log);
6344
6345 if (!ret)
6346 goto next;
6347 btrfs_handle_fs_error(fs_info, ret,
6348 "Couldn't read target root for tree log recovery.");
6349 goto error;
6350 }
6351
6352 wc.replay_dest->log_root = log;
6353 ret = btrfs_record_root_in_trans(trans, wc.replay_dest);
6354 if (ret)
6355 /* The loop needs to continue due to the root refs */
6356 btrfs_handle_fs_error(fs_info, ret,
6357 "failed to record the log root in transaction");
6358 else
6359 ret = walk_log_tree(trans, log, &wc);
6360
6361 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
6362 ret = fixup_inode_link_counts(trans, wc.replay_dest,
6363 path);
6364 }
6365
6366 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
6367 struct btrfs_root *root = wc.replay_dest;
6368
6369 btrfs_release_path(path);
6370
6371 /*
6372 * We have just replayed everything, and the highest
6373 * objectid of fs roots probably has changed in case
6374 * some inode_item's got replayed.
6375 *
6376 * root->objectid_mutex is not acquired as log replay
6377 * could only happen during mount.
6378 */
6379 ret = btrfs_init_root_free_objectid(root);
6380 }
6381
6382 wc.replay_dest->log_root = NULL;
6383 btrfs_put_root(wc.replay_dest);
6384 btrfs_put_root(log);
6385
6386 if (ret)
6387 goto error;
6388next:
6389 if (found_key.offset == 0)
6390 break;
6391 key.offset = found_key.offset - 1;
6392 }
6393 btrfs_release_path(path);
6394
6395 /* step one is to pin it all, step two is to replay just inodes */
6396 if (wc.pin) {
6397 wc.pin = 0;
6398 wc.process_func = replay_one_buffer;
6399 wc.stage = LOG_WALK_REPLAY_INODES;
6400 goto again;
6401 }
6402 /* step three is to replay everything */
6403 if (wc.stage < LOG_WALK_REPLAY_ALL) {
6404 wc.stage++;
6405 goto again;
6406 }
6407
6408 btrfs_free_path(path);
6409
6410 /* step 4: commit the transaction, which also unpins the blocks */
6411 ret = btrfs_commit_transaction(trans);
6412 if (ret)
6413 return ret;
6414
6415 log_root_tree->log_root = NULL;
6416 clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
6417 btrfs_put_root(log_root_tree);
6418
6419 return 0;
6420error:
6421 if (wc.trans)
6422 btrfs_end_transaction(wc.trans);
6423 clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
6424 btrfs_free_path(path);
6425 return ret;
6426}
6427
6428/*
6429 * there are some corner cases where we want to force a full
6430 * commit instead of allowing a directory to be logged.
6431 *
6432 * They revolve around files there were unlinked from the directory, and
6433 * this function updates the parent directory so that a full commit is
6434 * properly done if it is fsync'd later after the unlinks are done.
6435 *
6436 * Must be called before the unlink operations (updates to the subvolume tree,
6437 * inodes, etc) are done.
6438 */
6439void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
6440 struct btrfs_inode *dir, struct btrfs_inode *inode,
6441 int for_rename)
6442{
6443 /*
6444 * when we're logging a file, if it hasn't been renamed
6445 * or unlinked, and its inode is fully committed on disk,
6446 * we don't have to worry about walking up the directory chain
6447 * to log its parents.
6448 *
6449 * So, we use the last_unlink_trans field to put this transid
6450 * into the file. When the file is logged we check it and
6451 * don't log the parents if the file is fully on disk.
6452 */
6453 mutex_lock(&inode->log_mutex);
6454 inode->last_unlink_trans = trans->transid;
6455 mutex_unlock(&inode->log_mutex);
6456
6457 /*
6458 * if this directory was already logged any new
6459 * names for this file/dir will get recorded
6460 */
6461 if (dir->logged_trans == trans->transid)
6462 return;
6463
6464 /*
6465 * if the inode we're about to unlink was logged,
6466 * the log will be properly updated for any new names
6467 */
6468 if (inode->logged_trans == trans->transid)
6469 return;
6470
6471 /*
6472 * when renaming files across directories, if the directory
6473 * there we're unlinking from gets fsync'd later on, there's
6474 * no way to find the destination directory later and fsync it
6475 * properly. So, we have to be conservative and force commits
6476 * so the new name gets discovered.
6477 */
6478 if (for_rename)
6479 goto record;
6480
6481 /* we can safely do the unlink without any special recording */
6482 return;
6483
6484record:
6485 mutex_lock(&dir->log_mutex);
6486 dir->last_unlink_trans = trans->transid;
6487 mutex_unlock(&dir->log_mutex);
6488}
6489
6490/*
6491 * Make sure that if someone attempts to fsync the parent directory of a deleted
6492 * snapshot, it ends up triggering a transaction commit. This is to guarantee
6493 * that after replaying the log tree of the parent directory's root we will not
6494 * see the snapshot anymore and at log replay time we will not see any log tree
6495 * corresponding to the deleted snapshot's root, which could lead to replaying
6496 * it after replaying the log tree of the parent directory (which would replay
6497 * the snapshot delete operation).
6498 *
6499 * Must be called before the actual snapshot destroy operation (updates to the
6500 * parent root and tree of tree roots trees, etc) are done.
6501 */
6502void btrfs_record_snapshot_destroy(struct btrfs_trans_handle *trans,
6503 struct btrfs_inode *dir)
6504{
6505 mutex_lock(&dir->log_mutex);
6506 dir->last_unlink_trans = trans->transid;
6507 mutex_unlock(&dir->log_mutex);
6508}
6509
6510/*
6511 * Call this after adding a new name for a file and it will properly
6512 * update the log to reflect the new name.
6513 */
6514void btrfs_log_new_name(struct btrfs_trans_handle *trans,
6515 struct btrfs_inode *inode, struct btrfs_inode *old_dir,
6516 struct dentry *parent)
6517{
6518 struct btrfs_log_ctx ctx;
6519
6520 /*
6521 * this will force the logging code to walk the dentry chain
6522 * up for the file
6523 */
6524 if (!S_ISDIR(inode->vfs_inode.i_mode))
6525 inode->last_unlink_trans = trans->transid;
6526
6527 /*
6528 * if this inode hasn't been logged and directory we're renaming it
6529 * from hasn't been logged, we don't need to log it
6530 */
6531 if (!inode_logged(trans, inode) &&
6532 (!old_dir || !inode_logged(trans, old_dir)))
6533 return;
6534
6535 /*
6536 * If we are doing a rename (old_dir is not NULL) from a directory that
6537 * was previously logged, make sure the next log attempt on the directory
6538 * is not skipped and logs the inode again. This is because the log may
6539 * not currently be authoritative for a range including the old
6540 * BTRFS_DIR_ITEM_KEY and BTRFS_DIR_INDEX_KEY keys, so we want to make
6541 * sure after a log replay we do not end up with both the new and old
6542 * dentries around (in case the inode is a directory we would have a
6543 * directory with two hard links and 2 inode references for different
6544 * parents). The next log attempt of old_dir will happen at
6545 * btrfs_log_all_parents(), called through btrfs_log_inode_parent()
6546 * below, because we have previously set inode->last_unlink_trans to the
6547 * current transaction ID, either here or at btrfs_record_unlink_dir() in
6548 * case inode is a directory.
6549 */
6550 if (old_dir)
6551 old_dir->logged_trans = 0;
6552
6553 btrfs_init_log_ctx(&ctx, &inode->vfs_inode);
6554 ctx.logging_new_name = true;
6555 /*
6556 * We don't care about the return value. If we fail to log the new name
6557 * then we know the next attempt to sync the log will fallback to a full
6558 * transaction commit (due to a call to btrfs_set_log_full_commit()), so
6559 * we don't need to worry about getting a log committed that has an
6560 * inconsistent state after a rename operation.
6561 */
6562 btrfs_log_inode_parent(trans, inode, parent, LOG_INODE_EXISTS, &ctx);
6563}
6564