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 "inode-map.h"
21#include "block-group.h"
22#include "space-info.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 const loff_t start,
100 const loff_t end,
101 struct btrfs_log_ctx *ctx);
102static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
103 struct btrfs_root *root,
104 struct btrfs_path *path, u64 objectid);
105static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
106 struct btrfs_root *root,
107 struct btrfs_root *log,
108 struct btrfs_path *path,
109 u64 dirid, int del_all);
110
111/*
112 * tree logging is a special write ahead log used to make sure that
113 * fsyncs and O_SYNCs can happen without doing full tree commits.
114 *
115 * Full tree commits are expensive because they require commonly
116 * modified blocks to be recowed, creating many dirty pages in the
117 * extent tree an 4x-6x higher write load than ext3.
118 *
119 * Instead of doing a tree commit on every fsync, we use the
120 * key ranges and transaction ids to find items for a given file or directory
121 * that have changed in this transaction. Those items are copied into
122 * a special tree (one per subvolume root), that tree is written to disk
123 * and then the fsync is considered complete.
124 *
125 * After a crash, items are copied out of the log-tree back into the
126 * subvolume tree. Any file data extents found are recorded in the extent
127 * allocation tree, and the log-tree freed.
128 *
129 * The log tree is read three times, once to pin down all the extents it is
130 * using in ram and once, once to create all the inodes logged in the tree
131 * and once to do all the other items.
132 */
133
134/*
135 * start a sub transaction and setup the log tree
136 * this increments the log tree writer count to make the people
137 * syncing the tree wait for us to finish
138 */
139static int start_log_trans(struct btrfs_trans_handle *trans,
140 struct btrfs_root *root,
141 struct btrfs_log_ctx *ctx)
142{
143 struct btrfs_fs_info *fs_info = root->fs_info;
144 int ret = 0;
145
146 mutex_lock(&root->log_mutex);
147
148 if (root->log_root) {
149 if (btrfs_need_log_full_commit(trans)) {
150 ret = -EAGAIN;
151 goto out;
152 }
153
154 if (!root->log_start_pid) {
155 clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
156 root->log_start_pid = current->pid;
157 } else if (root->log_start_pid != current->pid) {
158 set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
159 }
160 } else {
161 mutex_lock(&fs_info->tree_log_mutex);
162 if (!fs_info->log_root_tree)
163 ret = btrfs_init_log_root_tree(trans, fs_info);
164 mutex_unlock(&fs_info->tree_log_mutex);
165 if (ret)
166 goto out;
167
168 ret = btrfs_add_log_tree(trans, root);
169 if (ret)
170 goto out;
171
172 set_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
173 clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
174 root->log_start_pid = current->pid;
175 }
176
177 atomic_inc(&root->log_batch);
178 atomic_inc(&root->log_writers);
179 if (ctx) {
180 int index = root->log_transid % 2;
181 list_add_tail(&ctx->list, &root->log_ctxs[index]);
182 ctx->log_transid = root->log_transid;
183 }
184
185out:
186 mutex_unlock(&root->log_mutex);
187 return ret;
188}
189
190/*
191 * returns 0 if there was a log transaction running and we were able
192 * to join, or returns -ENOENT if there were not transactions
193 * in progress
194 */
195static int join_running_log_trans(struct btrfs_root *root)
196{
197 int ret = -ENOENT;
198
199 if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state))
200 return ret;
201
202 mutex_lock(&root->log_mutex);
203 if (root->log_root) {
204 ret = 0;
205 atomic_inc(&root->log_writers);
206 }
207 mutex_unlock(&root->log_mutex);
208 return ret;
209}
210
211/*
212 * This either makes the current running log transaction wait
213 * until you call btrfs_end_log_trans() or it makes any future
214 * log transactions wait until you call btrfs_end_log_trans()
215 */
216void btrfs_pin_log_trans(struct btrfs_root *root)
217{
218 mutex_lock(&root->log_mutex);
219 atomic_inc(&root->log_writers);
220 mutex_unlock(&root->log_mutex);
221}
222
223/*
224 * indicate we're done making changes to the log tree
225 * and wake up anyone waiting to do a sync
226 */
227void btrfs_end_log_trans(struct btrfs_root *root)
228{
229 if (atomic_dec_and_test(&root->log_writers)) {
230 /* atomic_dec_and_test implies a barrier */
231 cond_wake_up_nomb(&root->log_writer_wait);
232 }
233}
234
235static int btrfs_write_tree_block(struct extent_buffer *buf)
236{
237 return filemap_fdatawrite_range(buf->pages[0]->mapping, buf->start,
238 buf->start + buf->len - 1);
239}
240
241static void btrfs_wait_tree_block_writeback(struct extent_buffer *buf)
242{
243 filemap_fdatawait_range(buf->pages[0]->mapping,
244 buf->start, buf->start + buf->len - 1);
245}
246
247/*
248 * the walk control struct is used to pass state down the chain when
249 * processing the log tree. The stage field tells us which part
250 * of the log tree processing we are currently doing. The others
251 * are state fields used for that specific part
252 */
253struct walk_control {
254 /* should we free the extent on disk when done? This is used
255 * at transaction commit time while freeing a log tree
256 */
257 int free;
258
259 /* should we write out the extent buffer? This is used
260 * while flushing the log tree to disk during a sync
261 */
262 int write;
263
264 /* should we wait for the extent buffer io to finish? Also used
265 * while flushing the log tree to disk for a sync
266 */
267 int wait;
268
269 /* pin only walk, we record which extents on disk belong to the
270 * log trees
271 */
272 int pin;
273
274 /* what stage of the replay code we're currently in */
275 int stage;
276
277 /*
278 * Ignore any items from the inode currently being processed. Needs
279 * to be set every time we find a BTRFS_INODE_ITEM_KEY and we are in
280 * the LOG_WALK_REPLAY_INODES stage.
281 */
282 bool ignore_cur_inode;
283
284 /* the root we are currently replaying */
285 struct btrfs_root *replay_dest;
286
287 /* the trans handle for the current replay */
288 struct btrfs_trans_handle *trans;
289
290 /* the function that gets used to process blocks we find in the
291 * tree. Note the extent_buffer might not be up to date when it is
292 * passed in, and it must be checked or read if you need the data
293 * inside it
294 */
295 int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
296 struct walk_control *wc, u64 gen, int level);
297};
298
299/*
300 * process_func used to pin down extents, write them or wait on them
301 */
302static int process_one_buffer(struct btrfs_root *log,
303 struct extent_buffer *eb,
304 struct walk_control *wc, u64 gen, int level)
305{
306 struct btrfs_fs_info *fs_info = log->fs_info;
307 int ret = 0;
308
309 /*
310 * If this fs is mixed then we need to be able to process the leaves to
311 * pin down any logged extents, so we have to read the block.
312 */
313 if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
314 ret = btrfs_read_buffer(eb, gen, level, NULL);
315 if (ret)
316 return ret;
317 }
318
319 if (wc->pin)
320 ret = btrfs_pin_extent_for_log_replay(wc->trans, eb->start,
321 eb->len);
322
323 if (!ret && btrfs_buffer_uptodate(eb, gen, 0)) {
324 if (wc->pin && btrfs_header_level(eb) == 0)
325 ret = btrfs_exclude_logged_extents(eb);
326 if (wc->write)
327 btrfs_write_tree_block(eb);
328 if (wc->wait)
329 btrfs_wait_tree_block_writeback(eb);
330 }
331 return ret;
332}
333
334/*
335 * Item overwrite used by replay and tree logging. eb, slot and key all refer
336 * to the src data we are copying out.
337 *
338 * root is the tree we are copying into, and path is a scratch
339 * path for use in this function (it should be released on entry and
340 * will be released on exit).
341 *
342 * If the key is already in the destination tree the existing item is
343 * overwritten. If the existing item isn't big enough, it is extended.
344 * If it is too large, it is truncated.
345 *
346 * If the key isn't in the destination yet, a new item is inserted.
347 */
348static noinline int overwrite_item(struct btrfs_trans_handle *trans,
349 struct btrfs_root *root,
350 struct btrfs_path *path,
351 struct extent_buffer *eb, int slot,
352 struct btrfs_key *key)
353{
354 int ret;
355 u32 item_size;
356 u64 saved_i_size = 0;
357 int save_old_i_size = 0;
358 unsigned long src_ptr;
359 unsigned long dst_ptr;
360 int overwrite_root = 0;
361 bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
362
363 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
364 overwrite_root = 1;
365
366 item_size = btrfs_item_size_nr(eb, slot);
367 src_ptr = btrfs_item_ptr_offset(eb, slot);
368
369 /* look for the key in the destination tree */
370 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
371 if (ret < 0)
372 return ret;
373
374 if (ret == 0) {
375 char *src_copy;
376 char *dst_copy;
377 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
378 path->slots[0]);
379 if (dst_size != item_size)
380 goto insert;
381
382 if (item_size == 0) {
383 btrfs_release_path(path);
384 return 0;
385 }
386 dst_copy = kmalloc(item_size, GFP_NOFS);
387 src_copy = kmalloc(item_size, GFP_NOFS);
388 if (!dst_copy || !src_copy) {
389 btrfs_release_path(path);
390 kfree(dst_copy);
391 kfree(src_copy);
392 return -ENOMEM;
393 }
394
395 read_extent_buffer(eb, src_copy, src_ptr, item_size);
396
397 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
398 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
399 item_size);
400 ret = memcmp(dst_copy, src_copy, item_size);
401
402 kfree(dst_copy);
403 kfree(src_copy);
404 /*
405 * they have the same contents, just return, this saves
406 * us from cowing blocks in the destination tree and doing
407 * extra writes that may not have been done by a previous
408 * sync
409 */
410 if (ret == 0) {
411 btrfs_release_path(path);
412 return 0;
413 }
414
415 /*
416 * We need to load the old nbytes into the inode so when we
417 * replay the extents we've logged we get the right nbytes.
418 */
419 if (inode_item) {
420 struct btrfs_inode_item *item;
421 u64 nbytes;
422 u32 mode;
423
424 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
425 struct btrfs_inode_item);
426 nbytes = btrfs_inode_nbytes(path->nodes[0], item);
427 item = btrfs_item_ptr(eb, slot,
428 struct btrfs_inode_item);
429 btrfs_set_inode_nbytes(eb, item, nbytes);
430
431 /*
432 * If this is a directory we need to reset the i_size to
433 * 0 so that we can set it up properly when replaying
434 * the rest of the items in this log.
435 */
436 mode = btrfs_inode_mode(eb, item);
437 if (S_ISDIR(mode))
438 btrfs_set_inode_size(eb, item, 0);
439 }
440 } else if (inode_item) {
441 struct btrfs_inode_item *item;
442 u32 mode;
443
444 /*
445 * New inode, set nbytes to 0 so that the nbytes comes out
446 * properly when we replay the extents.
447 */
448 item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
449 btrfs_set_inode_nbytes(eb, item, 0);
450
451 /*
452 * If this is a directory we need to reset the i_size to 0 so
453 * that we can set it up properly when replaying the rest of
454 * the items in this log.
455 */
456 mode = btrfs_inode_mode(eb, item);
457 if (S_ISDIR(mode))
458 btrfs_set_inode_size(eb, item, 0);
459 }
460insert:
461 btrfs_release_path(path);
462 /* try to insert the key into the destination tree */
463 path->skip_release_on_error = 1;
464 ret = btrfs_insert_empty_item(trans, root, path,
465 key, item_size);
466 path->skip_release_on_error = 0;
467
468 /* make sure any existing item is the correct size */
469 if (ret == -EEXIST || ret == -EOVERFLOW) {
470 u32 found_size;
471 found_size = btrfs_item_size_nr(path->nodes[0],
472 path->slots[0]);
473 if (found_size > item_size)
474 btrfs_truncate_item(path, item_size, 1);
475 else if (found_size < item_size)
476 btrfs_extend_item(path, item_size - found_size);
477 } else if (ret) {
478 return ret;
479 }
480 dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
481 path->slots[0]);
482
483 /* don't overwrite an existing inode if the generation number
484 * was logged as zero. This is done when the tree logging code
485 * is just logging an inode to make sure it exists after recovery.
486 *
487 * Also, don't overwrite i_size on directories during replay.
488 * log replay inserts and removes directory items based on the
489 * state of the tree found in the subvolume, and i_size is modified
490 * as it goes
491 */
492 if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
493 struct btrfs_inode_item *src_item;
494 struct btrfs_inode_item *dst_item;
495
496 src_item = (struct btrfs_inode_item *)src_ptr;
497 dst_item = (struct btrfs_inode_item *)dst_ptr;
498
499 if (btrfs_inode_generation(eb, src_item) == 0) {
500 struct extent_buffer *dst_eb = path->nodes[0];
501 const u64 ino_size = btrfs_inode_size(eb, src_item);
502
503 /*
504 * For regular files an ino_size == 0 is used only when
505 * logging that an inode exists, as part of a directory
506 * fsync, and the inode wasn't fsynced before. In this
507 * case don't set the size of the inode in the fs/subvol
508 * tree, otherwise we would be throwing valid data away.
509 */
510 if (S_ISREG(btrfs_inode_mode(eb, src_item)) &&
511 S_ISREG(btrfs_inode_mode(dst_eb, dst_item)) &&
512 ino_size != 0)
513 btrfs_set_inode_size(dst_eb, dst_item, ino_size);
514 goto no_copy;
515 }
516
517 if (overwrite_root &&
518 S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
519 S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
520 save_old_i_size = 1;
521 saved_i_size = btrfs_inode_size(path->nodes[0],
522 dst_item);
523 }
524 }
525
526 copy_extent_buffer(path->nodes[0], eb, dst_ptr,
527 src_ptr, item_size);
528
529 if (save_old_i_size) {
530 struct btrfs_inode_item *dst_item;
531 dst_item = (struct btrfs_inode_item *)dst_ptr;
532 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
533 }
534
535 /* make sure the generation is filled in */
536 if (key->type == BTRFS_INODE_ITEM_KEY) {
537 struct btrfs_inode_item *dst_item;
538 dst_item = (struct btrfs_inode_item *)dst_ptr;
539 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
540 btrfs_set_inode_generation(path->nodes[0], dst_item,
541 trans->transid);
542 }
543 }
544no_copy:
545 btrfs_mark_buffer_dirty(path->nodes[0]);
546 btrfs_release_path(path);
547 return 0;
548}
549
550/*
551 * simple helper to read an inode off the disk from a given root
552 * This can only be called for subvolume roots and not for the log
553 */
554static noinline struct inode *read_one_inode(struct btrfs_root *root,
555 u64 objectid)
556{
557 struct inode *inode;
558
559 inode = btrfs_iget(root->fs_info->sb, objectid, root);
560 if (IS_ERR(inode))
561 inode = NULL;
562 return inode;
563}
564
565/* replays a single extent in 'eb' at 'slot' with 'key' into the
566 * subvolume 'root'. path is released on entry and should be released
567 * on exit.
568 *
569 * extents in the log tree have not been allocated out of the extent
570 * tree yet. So, this completes the allocation, taking a reference
571 * as required if the extent already exists or creating a new extent
572 * if it isn't in the extent allocation tree yet.
573 *
574 * The extent is inserted into the file, dropping any existing extents
575 * from the file that overlap the new one.
576 */
577static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
578 struct btrfs_root *root,
579 struct btrfs_path *path,
580 struct extent_buffer *eb, int slot,
581 struct btrfs_key *key)
582{
583 struct btrfs_fs_info *fs_info = root->fs_info;
584 int found_type;
585 u64 extent_end;
586 u64 start = key->offset;
587 u64 nbytes = 0;
588 struct btrfs_file_extent_item *item;
589 struct inode *inode = NULL;
590 unsigned long size;
591 int ret = 0;
592
593 item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
594 found_type = btrfs_file_extent_type(eb, item);
595
596 if (found_type == BTRFS_FILE_EXTENT_REG ||
597 found_type == BTRFS_FILE_EXTENT_PREALLOC) {
598 nbytes = btrfs_file_extent_num_bytes(eb, item);
599 extent_end = start + nbytes;
600
601 /*
602 * We don't add to the inodes nbytes if we are prealloc or a
603 * hole.
604 */
605 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
606 nbytes = 0;
607 } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
608 size = btrfs_file_extent_ram_bytes(eb, item);
609 nbytes = btrfs_file_extent_ram_bytes(eb, item);
610 extent_end = ALIGN(start + size,
611 fs_info->sectorsize);
612 } else {
613 ret = 0;
614 goto out;
615 }
616
617 inode = read_one_inode(root, key->objectid);
618 if (!inode) {
619 ret = -EIO;
620 goto out;
621 }
622
623 /*
624 * first check to see if we already have this extent in the
625 * file. This must be done before the btrfs_drop_extents run
626 * so we don't try to drop this extent.
627 */
628 ret = btrfs_lookup_file_extent(trans, root, path,
629 btrfs_ino(BTRFS_I(inode)), start, 0);
630
631 if (ret == 0 &&
632 (found_type == BTRFS_FILE_EXTENT_REG ||
633 found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
634 struct btrfs_file_extent_item cmp1;
635 struct btrfs_file_extent_item cmp2;
636 struct btrfs_file_extent_item *existing;
637 struct extent_buffer *leaf;
638
639 leaf = path->nodes[0];
640 existing = btrfs_item_ptr(leaf, path->slots[0],
641 struct btrfs_file_extent_item);
642
643 read_extent_buffer(eb, &cmp1, (unsigned long)item,
644 sizeof(cmp1));
645 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
646 sizeof(cmp2));
647
648 /*
649 * we already have a pointer to this exact extent,
650 * we don't have to do anything
651 */
652 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
653 btrfs_release_path(path);
654 goto out;
655 }
656 }
657 btrfs_release_path(path);
658
659 /* drop any overlapping extents */
660 ret = btrfs_drop_extents(trans, root, inode, start, extent_end, 1);
661 if (ret)
662 goto out;
663
664 if (found_type == BTRFS_FILE_EXTENT_REG ||
665 found_type == BTRFS_FILE_EXTENT_PREALLOC) {
666 u64 offset;
667 unsigned long dest_offset;
668 struct btrfs_key ins;
669
670 if (btrfs_file_extent_disk_bytenr(eb, item) == 0 &&
671 btrfs_fs_incompat(fs_info, NO_HOLES))
672 goto update_inode;
673
674 ret = btrfs_insert_empty_item(trans, root, path, key,
675 sizeof(*item));
676 if (ret)
677 goto out;
678 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
679 path->slots[0]);
680 copy_extent_buffer(path->nodes[0], eb, dest_offset,
681 (unsigned long)item, sizeof(*item));
682
683 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
684 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
685 ins.type = BTRFS_EXTENT_ITEM_KEY;
686 offset = key->offset - btrfs_file_extent_offset(eb, item);
687
688 /*
689 * Manually record dirty extent, as here we did a shallow
690 * file extent item copy and skip normal backref update,
691 * but modifying extent tree all by ourselves.
692 * So need to manually record dirty extent for qgroup,
693 * as the owner of the file extent changed from log tree
694 * (doesn't affect qgroup) to fs/file tree(affects qgroup)
695 */
696 ret = btrfs_qgroup_trace_extent(trans,
697 btrfs_file_extent_disk_bytenr(eb, item),
698 btrfs_file_extent_disk_num_bytes(eb, item),
699 GFP_NOFS);
700 if (ret < 0)
701 goto out;
702
703 if (ins.objectid > 0) {
704 struct btrfs_ref ref = { 0 };
705 u64 csum_start;
706 u64 csum_end;
707 LIST_HEAD(ordered_sums);
708
709 /*
710 * is this extent already allocated in the extent
711 * allocation tree? If so, just add a reference
712 */
713 ret = btrfs_lookup_data_extent(fs_info, ins.objectid,
714 ins.offset);
715 if (ret == 0) {
716 btrfs_init_generic_ref(&ref,
717 BTRFS_ADD_DELAYED_REF,
718 ins.objectid, ins.offset, 0);
719 btrfs_init_data_ref(&ref,
720 root->root_key.objectid,
721 key->objectid, offset);
722 ret = btrfs_inc_extent_ref(trans, &ref);
723 if (ret)
724 goto out;
725 } else {
726 /*
727 * insert the extent pointer in the extent
728 * allocation tree
729 */
730 ret = btrfs_alloc_logged_file_extent(trans,
731 root->root_key.objectid,
732 key->objectid, offset, &ins);
733 if (ret)
734 goto out;
735 }
736 btrfs_release_path(path);
737
738 if (btrfs_file_extent_compression(eb, item)) {
739 csum_start = ins.objectid;
740 csum_end = csum_start + ins.offset;
741 } else {
742 csum_start = ins.objectid +
743 btrfs_file_extent_offset(eb, item);
744 csum_end = csum_start +
745 btrfs_file_extent_num_bytes(eb, item);
746 }
747
748 ret = btrfs_lookup_csums_range(root->log_root,
749 csum_start, csum_end - 1,
750 &ordered_sums, 0);
751 if (ret)
752 goto out;
753 /*
754 * Now delete all existing cums in the csum root that
755 * cover our range. We do this because we can have an
756 * extent that is completely referenced by one file
757 * extent item and partially referenced by another
758 * file extent item (like after using the clone or
759 * extent_same ioctls). In this case if we end up doing
760 * the replay of the one that partially references the
761 * extent first, and we do not do the csum deletion
762 * below, we can get 2 csum items in the csum tree that
763 * overlap each other. For example, imagine our log has
764 * the two following file extent items:
765 *
766 * key (257 EXTENT_DATA 409600)
767 * extent data disk byte 12845056 nr 102400
768 * extent data offset 20480 nr 20480 ram 102400
769 *
770 * key (257 EXTENT_DATA 819200)
771 * extent data disk byte 12845056 nr 102400
772 * extent data offset 0 nr 102400 ram 102400
773 *
774 * Where the second one fully references the 100K extent
775 * that starts at disk byte 12845056, and the log tree
776 * has a single csum item that covers the entire range
777 * of the extent:
778 *
779 * key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
780 *
781 * After the first file extent item is replayed, the
782 * csum tree gets the following csum item:
783 *
784 * key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
785 *
786 * Which covers the 20K sub-range starting at offset 20K
787 * of our extent. Now when we replay the second file
788 * extent item, if we do not delete existing csum items
789 * that cover any of its blocks, we end up getting two
790 * csum items in our csum tree that overlap each other:
791 *
792 * key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
793 * key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
794 *
795 * Which is a problem, because after this anyone trying
796 * to lookup up for the checksum of any block of our
797 * extent starting at an offset of 40K or higher, will
798 * end up looking at the second csum item only, which
799 * does not contain the checksum for any block starting
800 * at offset 40K or higher of our extent.
801 */
802 while (!list_empty(&ordered_sums)) {
803 struct btrfs_ordered_sum *sums;
804 sums = list_entry(ordered_sums.next,
805 struct btrfs_ordered_sum,
806 list);
807 if (!ret)
808 ret = btrfs_del_csums(trans,
809 fs_info->csum_root,
810 sums->bytenr,
811 sums->len);
812 if (!ret)
813 ret = btrfs_csum_file_blocks(trans,
814 fs_info->csum_root, sums);
815 list_del(&sums->list);
816 kfree(sums);
817 }
818 if (ret)
819 goto out;
820 } else {
821 btrfs_release_path(path);
822 }
823 } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
824 /* inline extents are easy, we just overwrite them */
825 ret = overwrite_item(trans, root, path, eb, slot, key);
826 if (ret)
827 goto out;
828 }
829
830 ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start,
831 extent_end - start);
832 if (ret)
833 goto out;
834
835 inode_add_bytes(inode, nbytes);
836update_inode:
837 ret = btrfs_update_inode(trans, root, inode);
838out:
839 if (inode)
840 iput(inode);
841 return ret;
842}
843
844/*
845 * when cleaning up conflicts between the directory names in the
846 * subvolume, directory names in the log and directory names in the
847 * inode back references, we may have to unlink inodes from directories.
848 *
849 * This is a helper function to do the unlink of a specific directory
850 * item
851 */
852static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
853 struct btrfs_root *root,
854 struct btrfs_path *path,
855 struct btrfs_inode *dir,
856 struct btrfs_dir_item *di)
857{
858 struct inode *inode;
859 char *name;
860 int name_len;
861 struct extent_buffer *leaf;
862 struct btrfs_key location;
863 int ret;
864
865 leaf = path->nodes[0];
866
867 btrfs_dir_item_key_to_cpu(leaf, di, &location);
868 name_len = btrfs_dir_name_len(leaf, di);
869 name = kmalloc(name_len, GFP_NOFS);
870 if (!name)
871 return -ENOMEM;
872
873 read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
874 btrfs_release_path(path);
875
876 inode = read_one_inode(root, location.objectid);
877 if (!inode) {
878 ret = -EIO;
879 goto out;
880 }
881
882 ret = link_to_fixup_dir(trans, root, path, location.objectid);
883 if (ret)
884 goto out;
885
886 ret = btrfs_unlink_inode(trans, root, dir, BTRFS_I(inode), name,
887 name_len);
888 if (ret)
889 goto out;
890 else
891 ret = btrfs_run_delayed_items(trans);
892out:
893 kfree(name);
894 iput(inode);
895 return ret;
896}
897
898/*
899 * helper function to see if a given name and sequence number found
900 * in an inode back reference are already in a directory and correctly
901 * point to this inode
902 */
903static noinline int inode_in_dir(struct btrfs_root *root,
904 struct btrfs_path *path,
905 u64 dirid, u64 objectid, u64 index,
906 const char *name, int name_len)
907{
908 struct btrfs_dir_item *di;
909 struct btrfs_key location;
910 int match = 0;
911
912 di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
913 index, name, name_len, 0);
914 if (di && !IS_ERR(di)) {
915 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
916 if (location.objectid != objectid)
917 goto out;
918 } else
919 goto out;
920 btrfs_release_path(path);
921
922 di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
923 if (di && !IS_ERR(di)) {
924 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
925 if (location.objectid != objectid)
926 goto out;
927 } else
928 goto out;
929 match = 1;
930out:
931 btrfs_release_path(path);
932 return match;
933}
934
935/*
936 * helper function to check a log tree for a named back reference in
937 * an inode. This is used to decide if a back reference that is
938 * found in the subvolume conflicts with what we find in the log.
939 *
940 * inode backreferences may have multiple refs in a single item,
941 * during replay we process one reference at a time, and we don't
942 * want to delete valid links to a file from the subvolume if that
943 * link is also in the log.
944 */
945static noinline int backref_in_log(struct btrfs_root *log,
946 struct btrfs_key *key,
947 u64 ref_objectid,
948 const char *name, int namelen)
949{
950 struct btrfs_path *path;
951 int ret;
952
953 path = btrfs_alloc_path();
954 if (!path)
955 return -ENOMEM;
956
957 ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
958 if (ret < 0) {
959 goto out;
960 } else if (ret == 1) {
961 ret = 0;
962 goto out;
963 }
964
965 if (key->type == BTRFS_INODE_EXTREF_KEY)
966 ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
967 path->slots[0],
968 ref_objectid,
969 name, namelen);
970 else
971 ret = !!btrfs_find_name_in_backref(path->nodes[0],
972 path->slots[0],
973 name, namelen);
974out:
975 btrfs_free_path(path);
976 return ret;
977}
978
979static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
980 struct btrfs_root *root,
981 struct btrfs_path *path,
982 struct btrfs_root *log_root,
983 struct btrfs_inode *dir,
984 struct btrfs_inode *inode,
985 u64 inode_objectid, u64 parent_objectid,
986 u64 ref_index, char *name, int namelen,
987 int *search_done)
988{
989 int ret;
990 char *victim_name;
991 int victim_name_len;
992 struct extent_buffer *leaf;
993 struct btrfs_dir_item *di;
994 struct btrfs_key search_key;
995 struct btrfs_inode_extref *extref;
996
997again:
998 /* Search old style refs */
999 search_key.objectid = inode_objectid;
1000 search_key.type = BTRFS_INODE_REF_KEY;
1001 search_key.offset = parent_objectid;
1002 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
1003 if (ret == 0) {
1004 struct btrfs_inode_ref *victim_ref;
1005 unsigned long ptr;
1006 unsigned long ptr_end;
1007
1008 leaf = path->nodes[0];
1009
1010 /* are we trying to overwrite a back ref for the root directory
1011 * if so, just jump out, we're done
1012 */
1013 if (search_key.objectid == search_key.offset)
1014 return 1;
1015
1016 /* check all the names in this back reference to see
1017 * if they are in the log. if so, we allow them to stay
1018 * otherwise they must be unlinked as a conflict
1019 */
1020 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1021 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
1022 while (ptr < ptr_end) {
1023 victim_ref = (struct btrfs_inode_ref *)ptr;
1024 victim_name_len = btrfs_inode_ref_name_len(leaf,
1025 victim_ref);
1026 victim_name = kmalloc(victim_name_len, GFP_NOFS);
1027 if (!victim_name)
1028 return -ENOMEM;
1029
1030 read_extent_buffer(leaf, victim_name,
1031 (unsigned long)(victim_ref + 1),
1032 victim_name_len);
1033
1034 ret = backref_in_log(log_root, &search_key,
1035 parent_objectid, victim_name,
1036 victim_name_len);
1037 if (ret < 0) {
1038 kfree(victim_name);
1039 return ret;
1040 } else if (!ret) {
1041 inc_nlink(&inode->vfs_inode);
1042 btrfs_release_path(path);
1043
1044 ret = btrfs_unlink_inode(trans, root, dir, inode,
1045 victim_name, victim_name_len);
1046 kfree(victim_name);
1047 if (ret)
1048 return ret;
1049 ret = btrfs_run_delayed_items(trans);
1050 if (ret)
1051 return ret;
1052 *search_done = 1;
1053 goto again;
1054 }
1055 kfree(victim_name);
1056
1057 ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
1058 }
1059
1060 /*
1061 * NOTE: we have searched root tree and checked the
1062 * corresponding ref, it does not need to check again.
1063 */
1064 *search_done = 1;
1065 }
1066 btrfs_release_path(path);
1067
1068 /* Same search but for extended refs */
1069 extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
1070 inode_objectid, parent_objectid, 0,
1071 0);
1072 if (!IS_ERR_OR_NULL(extref)) {
1073 u32 item_size;
1074 u32 cur_offset = 0;
1075 unsigned long base;
1076 struct inode *victim_parent;
1077
1078 leaf = path->nodes[0];
1079
1080 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1081 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
1082
1083 while (cur_offset < item_size) {
1084 extref = (struct btrfs_inode_extref *)(base + cur_offset);
1085
1086 victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
1087
1088 if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
1089 goto next;
1090
1091 victim_name = kmalloc(victim_name_len, GFP_NOFS);
1092 if (!victim_name)
1093 return -ENOMEM;
1094 read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
1095 victim_name_len);
1096
1097 search_key.objectid = inode_objectid;
1098 search_key.type = BTRFS_INODE_EXTREF_KEY;
1099 search_key.offset = btrfs_extref_hash(parent_objectid,
1100 victim_name,
1101 victim_name_len);
1102 ret = backref_in_log(log_root, &search_key,
1103 parent_objectid, victim_name,
1104 victim_name_len);
1105 if (ret < 0) {
1106 return ret;
1107 } else if (!ret) {
1108 ret = -ENOENT;
1109 victim_parent = read_one_inode(root,
1110 parent_objectid);
1111 if (victim_parent) {
1112 inc_nlink(&inode->vfs_inode);
1113 btrfs_release_path(path);
1114
1115 ret = btrfs_unlink_inode(trans, root,
1116 BTRFS_I(victim_parent),
1117 inode,
1118 victim_name,
1119 victim_name_len);
1120 if (!ret)
1121 ret = btrfs_run_delayed_items(
1122 trans);
1123 }
1124 iput(victim_parent);
1125 kfree(victim_name);
1126 if (ret)
1127 return ret;
1128 *search_done = 1;
1129 goto again;
1130 }
1131 kfree(victim_name);
1132next:
1133 cur_offset += victim_name_len + sizeof(*extref);
1134 }
1135 *search_done = 1;
1136 }
1137 btrfs_release_path(path);
1138
1139 /* look for a conflicting sequence number */
1140 di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
1141 ref_index, name, namelen, 0);
1142 if (di && !IS_ERR(di)) {
1143 ret = drop_one_dir_item(trans, root, path, dir, di);
1144 if (ret)
1145 return ret;
1146 }
1147 btrfs_release_path(path);
1148
1149 /* look for a conflicting name */
1150 di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
1151 name, namelen, 0);
1152 if (di && !IS_ERR(di)) {
1153 ret = drop_one_dir_item(trans, root, path, dir, di);
1154 if (ret)
1155 return ret;
1156 }
1157 btrfs_release_path(path);
1158
1159 return 0;
1160}
1161
1162static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1163 u32 *namelen, char **name, u64 *index,
1164 u64 *parent_objectid)
1165{
1166 struct btrfs_inode_extref *extref;
1167
1168 extref = (struct btrfs_inode_extref *)ref_ptr;
1169
1170 *namelen = btrfs_inode_extref_name_len(eb, extref);
1171 *name = kmalloc(*namelen, GFP_NOFS);
1172 if (*name == NULL)
1173 return -ENOMEM;
1174
1175 read_extent_buffer(eb, *name, (unsigned long)&extref->name,
1176 *namelen);
1177
1178 if (index)
1179 *index = btrfs_inode_extref_index(eb, extref);
1180 if (parent_objectid)
1181 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
1182
1183 return 0;
1184}
1185
1186static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
1187 u32 *namelen, char **name, u64 *index)
1188{
1189 struct btrfs_inode_ref *ref;
1190
1191 ref = (struct btrfs_inode_ref *)ref_ptr;
1192
1193 *namelen = btrfs_inode_ref_name_len(eb, ref);
1194 *name = kmalloc(*namelen, GFP_NOFS);
1195 if (*name == NULL)
1196 return -ENOMEM;
1197
1198 read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1199
1200 if (index)
1201 *index = btrfs_inode_ref_index(eb, ref);
1202
1203 return 0;
1204}
1205
1206/*
1207 * Take an inode reference item from the log tree and iterate all names from the
1208 * inode reference item in the subvolume tree with the same key (if it exists).
1209 * For any name that is not in the inode reference item from the log tree, do a
1210 * proper unlink of that name (that is, remove its entry from the inode
1211 * reference item and both dir index keys).
1212 */
1213static int unlink_old_inode_refs(struct btrfs_trans_handle *trans,
1214 struct btrfs_root *root,
1215 struct btrfs_path *path,
1216 struct btrfs_inode *inode,
1217 struct extent_buffer *log_eb,
1218 int log_slot,
1219 struct btrfs_key *key)
1220{
1221 int ret;
1222 unsigned long ref_ptr;
1223 unsigned long ref_end;
1224 struct extent_buffer *eb;
1225
1226again:
1227 btrfs_release_path(path);
1228 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1229 if (ret > 0) {
1230 ret = 0;
1231 goto out;
1232 }
1233 if (ret < 0)
1234 goto out;
1235
1236 eb = path->nodes[0];
1237 ref_ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
1238 ref_end = ref_ptr + btrfs_item_size_nr(eb, path->slots[0]);
1239 while (ref_ptr < ref_end) {
1240 char *name = NULL;
1241 int namelen;
1242 u64 parent_id;
1243
1244 if (key->type == BTRFS_INODE_EXTREF_KEY) {
1245 ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1246 NULL, &parent_id);
1247 } else {
1248 parent_id = key->offset;
1249 ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1250 NULL);
1251 }
1252 if (ret)
1253 goto out;
1254
1255 if (key->type == BTRFS_INODE_EXTREF_KEY)
1256 ret = !!btrfs_find_name_in_ext_backref(log_eb, log_slot,
1257 parent_id, name,
1258 namelen);
1259 else
1260 ret = !!btrfs_find_name_in_backref(log_eb, log_slot,
1261 name, namelen);
1262
1263 if (!ret) {
1264 struct inode *dir;
1265
1266 btrfs_release_path(path);
1267 dir = read_one_inode(root, parent_id);
1268 if (!dir) {
1269 ret = -ENOENT;
1270 kfree(name);
1271 goto out;
1272 }
1273 ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
1274 inode, name, namelen);
1275 kfree(name);
1276 iput(dir);
1277 if (ret)
1278 goto out;
1279 goto again;
1280 }
1281
1282 kfree(name);
1283 ref_ptr += namelen;
1284 if (key->type == BTRFS_INODE_EXTREF_KEY)
1285 ref_ptr += sizeof(struct btrfs_inode_extref);
1286 else
1287 ref_ptr += sizeof(struct btrfs_inode_ref);
1288 }
1289 ret = 0;
1290 out:
1291 btrfs_release_path(path);
1292 return ret;
1293}
1294
1295static int btrfs_inode_ref_exists(struct inode *inode, struct inode *dir,
1296 const u8 ref_type, const char *name,
1297 const int namelen)
1298{
1299 struct btrfs_key key;
1300 struct btrfs_path *path;
1301 const u64 parent_id = btrfs_ino(BTRFS_I(dir));
1302 int ret;
1303
1304 path = btrfs_alloc_path();
1305 if (!path)
1306 return -ENOMEM;
1307
1308 key.objectid = btrfs_ino(BTRFS_I(inode));
1309 key.type = ref_type;
1310 if (key.type == BTRFS_INODE_REF_KEY)
1311 key.offset = parent_id;
1312 else
1313 key.offset = btrfs_extref_hash(parent_id, name, namelen);
1314
1315 ret = btrfs_search_slot(NULL, BTRFS_I(inode)->root, &key, path, 0, 0);
1316 if (ret < 0)
1317 goto out;
1318 if (ret > 0) {
1319 ret = 0;
1320 goto out;
1321 }
1322 if (key.type == BTRFS_INODE_EXTREF_KEY)
1323 ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
1324 path->slots[0], parent_id, name, namelen);
1325 else
1326 ret = !!btrfs_find_name_in_backref(path->nodes[0], path->slots[0],
1327 name, namelen);
1328
1329out:
1330 btrfs_free_path(path);
1331 return ret;
1332}
1333
1334static int add_link(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1335 struct inode *dir, struct inode *inode, const char *name,
1336 int namelen, u64 ref_index)
1337{
1338 struct btrfs_dir_item *dir_item;
1339 struct btrfs_key key;
1340 struct btrfs_path *path;
1341 struct inode *other_inode = NULL;
1342 int ret;
1343
1344 path = btrfs_alloc_path();
1345 if (!path)
1346 return -ENOMEM;
1347
1348 dir_item = btrfs_lookup_dir_item(NULL, root, path,
1349 btrfs_ino(BTRFS_I(dir)),
1350 name, namelen, 0);
1351 if (!dir_item) {
1352 btrfs_release_path(path);
1353 goto add_link;
1354 } else if (IS_ERR(dir_item)) {
1355 ret = PTR_ERR(dir_item);
1356 goto out;
1357 }
1358
1359 /*
1360 * Our inode's dentry collides with the dentry of another inode which is
1361 * in the log but not yet processed since it has a higher inode number.
1362 * So delete that other dentry.
1363 */
1364 btrfs_dir_item_key_to_cpu(path->nodes[0], dir_item, &key);
1365 btrfs_release_path(path);
1366 other_inode = read_one_inode(root, key.objectid);
1367 if (!other_inode) {
1368 ret = -ENOENT;
1369 goto out;
1370 }
1371 ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir), BTRFS_I(other_inode),
1372 name, namelen);
1373 if (ret)
1374 goto out;
1375 /*
1376 * If we dropped the link count to 0, bump it so that later the iput()
1377 * on the inode will not free it. We will fixup the link count later.
1378 */
1379 if (other_inode->i_nlink == 0)
1380 inc_nlink(other_inode);
1381
1382 ret = btrfs_run_delayed_items(trans);
1383 if (ret)
1384 goto out;
1385add_link:
1386 ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
1387 name, namelen, 0, ref_index);
1388out:
1389 iput(other_inode);
1390 btrfs_free_path(path);
1391
1392 return ret;
1393}
1394
1395/*
1396 * replay one inode back reference item found in the log tree.
1397 * eb, slot and key refer to the buffer and key found in the log tree.
1398 * root is the destination we are replaying into, and path is for temp
1399 * use by this function. (it should be released on return).
1400 */
1401static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1402 struct btrfs_root *root,
1403 struct btrfs_root *log,
1404 struct btrfs_path *path,
1405 struct extent_buffer *eb, int slot,
1406 struct btrfs_key *key)
1407{
1408 struct inode *dir = NULL;
1409 struct inode *inode = NULL;
1410 unsigned long ref_ptr;
1411 unsigned long ref_end;
1412 char *name = NULL;
1413 int namelen;
1414 int ret;
1415 int search_done = 0;
1416 int log_ref_ver = 0;
1417 u64 parent_objectid;
1418 u64 inode_objectid;
1419 u64 ref_index = 0;
1420 int ref_struct_size;
1421
1422 ref_ptr = btrfs_item_ptr_offset(eb, slot);
1423 ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1424
1425 if (key->type == BTRFS_INODE_EXTREF_KEY) {
1426 struct btrfs_inode_extref *r;
1427
1428 ref_struct_size = sizeof(struct btrfs_inode_extref);
1429 log_ref_ver = 1;
1430 r = (struct btrfs_inode_extref *)ref_ptr;
1431 parent_objectid = btrfs_inode_extref_parent(eb, r);
1432 } else {
1433 ref_struct_size = sizeof(struct btrfs_inode_ref);
1434 parent_objectid = key->offset;
1435 }
1436 inode_objectid = key->objectid;
1437
1438 /*
1439 * it is possible that we didn't log all the parent directories
1440 * for a given inode. If we don't find the dir, just don't
1441 * copy the back ref in. The link count fixup code will take
1442 * care of the rest
1443 */
1444 dir = read_one_inode(root, parent_objectid);
1445 if (!dir) {
1446 ret = -ENOENT;
1447 goto out;
1448 }
1449
1450 inode = read_one_inode(root, inode_objectid);
1451 if (!inode) {
1452 ret = -EIO;
1453 goto out;
1454 }
1455
1456 while (ref_ptr < ref_end) {
1457 if (log_ref_ver) {
1458 ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1459 &ref_index, &parent_objectid);
1460 /*
1461 * parent object can change from one array
1462 * item to another.
1463 */
1464 if (!dir)
1465 dir = read_one_inode(root, parent_objectid);
1466 if (!dir) {
1467 ret = -ENOENT;
1468 goto out;
1469 }
1470 } else {
1471 ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1472 &ref_index);
1473 }
1474 if (ret)
1475 goto out;
1476
1477 /* if we already have a perfect match, we're done */
1478 if (!inode_in_dir(root, path, btrfs_ino(BTRFS_I(dir)),
1479 btrfs_ino(BTRFS_I(inode)), ref_index,
1480 name, namelen)) {
1481 /*
1482 * look for a conflicting back reference in the
1483 * metadata. if we find one we have to unlink that name
1484 * of the file before we add our new link. Later on, we
1485 * overwrite any existing back reference, and we don't
1486 * want to create dangling pointers in the directory.
1487 */
1488
1489 if (!search_done) {
1490 ret = __add_inode_ref(trans, root, path, log,
1491 BTRFS_I(dir),
1492 BTRFS_I(inode),
1493 inode_objectid,
1494 parent_objectid,
1495 ref_index, name, namelen,
1496 &search_done);
1497 if (ret) {
1498 if (ret == 1)
1499 ret = 0;
1500 goto out;
1501 }
1502 }
1503
1504 /*
1505 * If a reference item already exists for this inode
1506 * with the same parent and name, but different index,
1507 * drop it and the corresponding directory index entries
1508 * from the parent before adding the new reference item
1509 * and dir index entries, otherwise we would fail with
1510 * -EEXIST returned from btrfs_add_link() below.
1511 */
1512 ret = btrfs_inode_ref_exists(inode, dir, key->type,
1513 name, namelen);
1514 if (ret > 0) {
1515 ret = btrfs_unlink_inode(trans, root,
1516 BTRFS_I(dir),
1517 BTRFS_I(inode),
1518 name, namelen);
1519 /*
1520 * If we dropped the link count to 0, bump it so
1521 * that later the iput() on the inode will not
1522 * free it. We will fixup the link count later.
1523 */
1524 if (!ret && inode->i_nlink == 0)
1525 inc_nlink(inode);
1526 }
1527 if (ret < 0)
1528 goto out;
1529
1530 /* insert our name */
1531 ret = add_link(trans, root, dir, inode, name, namelen,
1532 ref_index);
1533 if (ret)
1534 goto out;
1535
1536 btrfs_update_inode(trans, root, inode);
1537 }
1538
1539 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1540 kfree(name);
1541 name = NULL;
1542 if (log_ref_ver) {
1543 iput(dir);
1544 dir = NULL;
1545 }
1546 }
1547
1548 /*
1549 * Before we overwrite the inode reference item in the subvolume tree
1550 * with the item from the log tree, we must unlink all names from the
1551 * parent directory that are in the subvolume's tree inode reference
1552 * item, otherwise we end up with an inconsistent subvolume tree where
1553 * dir index entries exist for a name but there is no inode reference
1554 * item with the same name.
1555 */
1556 ret = unlink_old_inode_refs(trans, root, path, BTRFS_I(inode), eb, slot,
1557 key);
1558 if (ret)
1559 goto out;
1560
1561 /* finally write the back reference in the inode */
1562 ret = overwrite_item(trans, root, path, eb, slot, key);
1563out:
1564 btrfs_release_path(path);
1565 kfree(name);
1566 iput(dir);
1567 iput(inode);
1568 return ret;
1569}
1570
1571static int insert_orphan_item(struct btrfs_trans_handle *trans,
1572 struct btrfs_root *root, u64 ino)
1573{
1574 int ret;
1575
1576 ret = btrfs_insert_orphan_item(trans, root, ino);
1577 if (ret == -EEXIST)
1578 ret = 0;
1579
1580 return ret;
1581}
1582
1583static int count_inode_extrefs(struct btrfs_root *root,
1584 struct btrfs_inode *inode, struct btrfs_path *path)
1585{
1586 int ret = 0;
1587 int name_len;
1588 unsigned int nlink = 0;
1589 u32 item_size;
1590 u32 cur_offset = 0;
1591 u64 inode_objectid = btrfs_ino(inode);
1592 u64 offset = 0;
1593 unsigned long ptr;
1594 struct btrfs_inode_extref *extref;
1595 struct extent_buffer *leaf;
1596
1597 while (1) {
1598 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1599 &extref, &offset);
1600 if (ret)
1601 break;
1602
1603 leaf = path->nodes[0];
1604 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1605 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1606 cur_offset = 0;
1607
1608 while (cur_offset < item_size) {
1609 extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1610 name_len = btrfs_inode_extref_name_len(leaf, extref);
1611
1612 nlink++;
1613
1614 cur_offset += name_len + sizeof(*extref);
1615 }
1616
1617 offset++;
1618 btrfs_release_path(path);
1619 }
1620 btrfs_release_path(path);
1621
1622 if (ret < 0 && ret != -ENOENT)
1623 return ret;
1624 return nlink;
1625}
1626
1627static int count_inode_refs(struct btrfs_root *root,
1628 struct btrfs_inode *inode, struct btrfs_path *path)
1629{
1630 int ret;
1631 struct btrfs_key key;
1632 unsigned int nlink = 0;
1633 unsigned long ptr;
1634 unsigned long ptr_end;
1635 int name_len;
1636 u64 ino = btrfs_ino(inode);
1637
1638 key.objectid = ino;
1639 key.type = BTRFS_INODE_REF_KEY;
1640 key.offset = (u64)-1;
1641
1642 while (1) {
1643 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1644 if (ret < 0)
1645 break;
1646 if (ret > 0) {
1647 if (path->slots[0] == 0)
1648 break;
1649 path->slots[0]--;
1650 }
1651process_slot:
1652 btrfs_item_key_to_cpu(path->nodes[0], &key,
1653 path->slots[0]);
1654 if (key.objectid != ino ||
1655 key.type != BTRFS_INODE_REF_KEY)
1656 break;
1657 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1658 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1659 path->slots[0]);
1660 while (ptr < ptr_end) {
1661 struct btrfs_inode_ref *ref;
1662
1663 ref = (struct btrfs_inode_ref *)ptr;
1664 name_len = btrfs_inode_ref_name_len(path->nodes[0],
1665 ref);
1666 ptr = (unsigned long)(ref + 1) + name_len;
1667 nlink++;
1668 }
1669
1670 if (key.offset == 0)
1671 break;
1672 if (path->slots[0] > 0) {
1673 path->slots[0]--;
1674 goto process_slot;
1675 }
1676 key.offset--;
1677 btrfs_release_path(path);
1678 }
1679 btrfs_release_path(path);
1680
1681 return nlink;
1682}
1683
1684/*
1685 * There are a few corners where the link count of the file can't
1686 * be properly maintained during replay. So, instead of adding
1687 * lots of complexity to the log code, we just scan the backrefs
1688 * for any file that has been through replay.
1689 *
1690 * The scan will update the link count on the inode to reflect the
1691 * number of back refs found. If it goes down to zero, the iput
1692 * will free the inode.
1693 */
1694static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1695 struct btrfs_root *root,
1696 struct inode *inode)
1697{
1698 struct btrfs_path *path;
1699 int ret;
1700 u64 nlink = 0;
1701 u64 ino = btrfs_ino(BTRFS_I(inode));
1702
1703 path = btrfs_alloc_path();
1704 if (!path)
1705 return -ENOMEM;
1706
1707 ret = count_inode_refs(root, BTRFS_I(inode), path);
1708 if (ret < 0)
1709 goto out;
1710
1711 nlink = ret;
1712
1713 ret = count_inode_extrefs(root, BTRFS_I(inode), path);
1714 if (ret < 0)
1715 goto out;
1716
1717 nlink += ret;
1718
1719 ret = 0;
1720
1721 if (nlink != inode->i_nlink) {
1722 set_nlink(inode, nlink);
1723 btrfs_update_inode(trans, root, inode);
1724 }
1725 BTRFS_I(inode)->index_cnt = (u64)-1;
1726
1727 if (inode->i_nlink == 0) {
1728 if (S_ISDIR(inode->i_mode)) {
1729 ret = replay_dir_deletes(trans, root, NULL, path,
1730 ino, 1);
1731 if (ret)
1732 goto out;
1733 }
1734 ret = insert_orphan_item(trans, root, ino);
1735 }
1736
1737out:
1738 btrfs_free_path(path);
1739 return ret;
1740}
1741
1742static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1743 struct btrfs_root *root,
1744 struct btrfs_path *path)
1745{
1746 int ret;
1747 struct btrfs_key key;
1748 struct inode *inode;
1749
1750 key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1751 key.type = BTRFS_ORPHAN_ITEM_KEY;
1752 key.offset = (u64)-1;
1753 while (1) {
1754 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1755 if (ret < 0)
1756 break;
1757
1758 if (ret == 1) {
1759 if (path->slots[0] == 0)
1760 break;
1761 path->slots[0]--;
1762 }
1763
1764 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1765 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1766 key.type != BTRFS_ORPHAN_ITEM_KEY)
1767 break;
1768
1769 ret = btrfs_del_item(trans, root, path);
1770 if (ret)
1771 goto out;
1772
1773 btrfs_release_path(path);
1774 inode = read_one_inode(root, key.offset);
1775 if (!inode)
1776 return -EIO;
1777
1778 ret = fixup_inode_link_count(trans, root, inode);
1779 iput(inode);
1780 if (ret)
1781 goto out;
1782
1783 /*
1784 * fixup on a directory may create new entries,
1785 * make sure we always look for the highset possible
1786 * offset
1787 */
1788 key.offset = (u64)-1;
1789 }
1790 ret = 0;
1791out:
1792 btrfs_release_path(path);
1793 return ret;
1794}
1795
1796
1797/*
1798 * record a given inode in the fixup dir so we can check its link
1799 * count when replay is done. The link count is incremented here
1800 * so the inode won't go away until we check it
1801 */
1802static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1803 struct btrfs_root *root,
1804 struct btrfs_path *path,
1805 u64 objectid)
1806{
1807 struct btrfs_key key;
1808 int ret = 0;
1809 struct inode *inode;
1810
1811 inode = read_one_inode(root, objectid);
1812 if (!inode)
1813 return -EIO;
1814
1815 key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1816 key.type = BTRFS_ORPHAN_ITEM_KEY;
1817 key.offset = objectid;
1818
1819 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1820
1821 btrfs_release_path(path);
1822 if (ret == 0) {
1823 if (!inode->i_nlink)
1824 set_nlink(inode, 1);
1825 else
1826 inc_nlink(inode);
1827 ret = btrfs_update_inode(trans, root, inode);
1828 } else if (ret == -EEXIST) {
1829 ret = 0;
1830 } else {
1831 BUG(); /* Logic Error */
1832 }
1833 iput(inode);
1834
1835 return ret;
1836}
1837
1838/*
1839 * when replaying the log for a directory, we only insert names
1840 * for inodes that actually exist. This means an fsync on a directory
1841 * does not implicitly fsync all the new files in it
1842 */
1843static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1844 struct btrfs_root *root,
1845 u64 dirid, u64 index,
1846 char *name, int name_len,
1847 struct btrfs_key *location)
1848{
1849 struct inode *inode;
1850 struct inode *dir;
1851 int ret;
1852
1853 inode = read_one_inode(root, location->objectid);
1854 if (!inode)
1855 return -ENOENT;
1856
1857 dir = read_one_inode(root, dirid);
1858 if (!dir) {
1859 iput(inode);
1860 return -EIO;
1861 }
1862
1863 ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode), name,
1864 name_len, 1, index);
1865
1866 /* FIXME, put inode into FIXUP list */
1867
1868 iput(inode);
1869 iput(dir);
1870 return ret;
1871}
1872
1873/*
1874 * take a single entry in a log directory item and replay it into
1875 * the subvolume.
1876 *
1877 * if a conflicting item exists in the subdirectory already,
1878 * the inode it points to is unlinked and put into the link count
1879 * fix up tree.
1880 *
1881 * If a name from the log points to a file or directory that does
1882 * not exist in the FS, it is skipped. fsyncs on directories
1883 * do not force down inodes inside that directory, just changes to the
1884 * names or unlinks in a directory.
1885 *
1886 * Returns < 0 on error, 0 if the name wasn't replayed (dentry points to a
1887 * non-existing inode) and 1 if the name was replayed.
1888 */
1889static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1890 struct btrfs_root *root,
1891 struct btrfs_path *path,
1892 struct extent_buffer *eb,
1893 struct btrfs_dir_item *di,
1894 struct btrfs_key *key)
1895{
1896 char *name;
1897 int name_len;
1898 struct btrfs_dir_item *dst_di;
1899 struct btrfs_key found_key;
1900 struct btrfs_key log_key;
1901 struct inode *dir;
1902 u8 log_type;
1903 int exists;
1904 int ret = 0;
1905 bool update_size = (key->type == BTRFS_DIR_INDEX_KEY);
1906 bool name_added = false;
1907
1908 dir = read_one_inode(root, key->objectid);
1909 if (!dir)
1910 return -EIO;
1911
1912 name_len = btrfs_dir_name_len(eb, di);
1913 name = kmalloc(name_len, GFP_NOFS);
1914 if (!name) {
1915 ret = -ENOMEM;
1916 goto out;
1917 }
1918
1919 log_type = btrfs_dir_type(eb, di);
1920 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1921 name_len);
1922
1923 btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1924 exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1925 if (exists == 0)
1926 exists = 1;
1927 else
1928 exists = 0;
1929 btrfs_release_path(path);
1930
1931 if (key->type == BTRFS_DIR_ITEM_KEY) {
1932 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1933 name, name_len, 1);
1934 } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1935 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1936 key->objectid,
1937 key->offset, name,
1938 name_len, 1);
1939 } else {
1940 /* Corruption */
1941 ret = -EINVAL;
1942 goto out;
1943 }
1944 if (IS_ERR_OR_NULL(dst_di)) {
1945 /* we need a sequence number to insert, so we only
1946 * do inserts for the BTRFS_DIR_INDEX_KEY types
1947 */
1948 if (key->type != BTRFS_DIR_INDEX_KEY)
1949 goto out;
1950 goto insert;
1951 }
1952
1953 btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1954 /* the existing item matches the logged item */
1955 if (found_key.objectid == log_key.objectid &&
1956 found_key.type == log_key.type &&
1957 found_key.offset == log_key.offset &&
1958 btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1959 update_size = false;
1960 goto out;
1961 }
1962
1963 /*
1964 * don't drop the conflicting directory entry if the inode
1965 * for the new entry doesn't exist
1966 */
1967 if (!exists)
1968 goto out;
1969
1970 ret = drop_one_dir_item(trans, root, path, BTRFS_I(dir), dst_di);
1971 if (ret)
1972 goto out;
1973
1974 if (key->type == BTRFS_DIR_INDEX_KEY)
1975 goto insert;
1976out:
1977 btrfs_release_path(path);
1978 if (!ret && update_size) {
1979 btrfs_i_size_write(BTRFS_I(dir), dir->i_size + name_len * 2);
1980 ret = btrfs_update_inode(trans, root, dir);
1981 }
1982 kfree(name);
1983 iput(dir);
1984 if (!ret && name_added)
1985 ret = 1;
1986 return ret;
1987
1988insert:
1989 /*
1990 * Check if the inode reference exists in the log for the given name,
1991 * inode and parent inode
1992 */
1993 found_key.objectid = log_key.objectid;
1994 found_key.type = BTRFS_INODE_REF_KEY;
1995 found_key.offset = key->objectid;
1996 ret = backref_in_log(root->log_root, &found_key, 0, name, name_len);
1997 if (ret < 0) {
1998 goto out;
1999 } else if (ret) {
2000 /* The dentry will be added later. */
2001 ret = 0;
2002 update_size = false;
2003 goto out;
2004 }
2005
2006 found_key.objectid = log_key.objectid;
2007 found_key.type = BTRFS_INODE_EXTREF_KEY;
2008 found_key.offset = key->objectid;
2009 ret = backref_in_log(root->log_root, &found_key, key->objectid, name,
2010 name_len);
2011 if (ret < 0) {
2012 goto out;
2013 } else if (ret) {
2014 /* The dentry will be added later. */
2015 ret = 0;
2016 update_size = false;
2017 goto out;
2018 }
2019 btrfs_release_path(path);
2020 ret = insert_one_name(trans, root, key->objectid, key->offset,
2021 name, name_len, &log_key);
2022 if (ret && ret != -ENOENT && ret != -EEXIST)
2023 goto out;
2024 if (!ret)
2025 name_added = true;
2026 update_size = false;
2027 ret = 0;
2028 goto out;
2029}
2030
2031/*
2032 * find all the names in a directory item and reconcile them into
2033 * the subvolume. Only BTRFS_DIR_ITEM_KEY types will have more than
2034 * one name in a directory item, but the same code gets used for
2035 * both directory index types
2036 */
2037static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
2038 struct btrfs_root *root,
2039 struct btrfs_path *path,
2040 struct extent_buffer *eb, int slot,
2041 struct btrfs_key *key)
2042{
2043 int ret = 0;
2044 u32 item_size = btrfs_item_size_nr(eb, slot);
2045 struct btrfs_dir_item *di;
2046 int name_len;
2047 unsigned long ptr;
2048 unsigned long ptr_end;
2049 struct btrfs_path *fixup_path = NULL;
2050
2051 ptr = btrfs_item_ptr_offset(eb, slot);
2052 ptr_end = ptr + item_size;
2053 while (ptr < ptr_end) {
2054 di = (struct btrfs_dir_item *)ptr;
2055 name_len = btrfs_dir_name_len(eb, di);
2056 ret = replay_one_name(trans, root, path, eb, di, key);
2057 if (ret < 0)
2058 break;
2059 ptr = (unsigned long)(di + 1);
2060 ptr += name_len;
2061
2062 /*
2063 * If this entry refers to a non-directory (directories can not
2064 * have a link count > 1) and it was added in the transaction
2065 * that was not committed, make sure we fixup the link count of
2066 * the inode it the entry points to. Otherwise something like
2067 * the following would result in a directory pointing to an
2068 * inode with a wrong link that does not account for this dir
2069 * entry:
2070 *
2071 * mkdir testdir
2072 * touch testdir/foo
2073 * touch testdir/bar
2074 * sync
2075 *
2076 * ln testdir/bar testdir/bar_link
2077 * ln testdir/foo testdir/foo_link
2078 * xfs_io -c "fsync" testdir/bar
2079 *
2080 * <power failure>
2081 *
2082 * mount fs, log replay happens
2083 *
2084 * File foo would remain with a link count of 1 when it has two
2085 * entries pointing to it in the directory testdir. This would
2086 * make it impossible to ever delete the parent directory has
2087 * it would result in stale dentries that can never be deleted.
2088 */
2089 if (ret == 1 && btrfs_dir_type(eb, di) != BTRFS_FT_DIR) {
2090 struct btrfs_key di_key;
2091
2092 if (!fixup_path) {
2093 fixup_path = btrfs_alloc_path();
2094 if (!fixup_path) {
2095 ret = -ENOMEM;
2096 break;
2097 }
2098 }
2099
2100 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2101 ret = link_to_fixup_dir(trans, root, fixup_path,
2102 di_key.objectid);
2103 if (ret)
2104 break;
2105 }
2106 ret = 0;
2107 }
2108 btrfs_free_path(fixup_path);
2109 return ret;
2110}
2111
2112/*
2113 * directory replay has two parts. There are the standard directory
2114 * items in the log copied from the subvolume, and range items
2115 * created in the log while the subvolume was logged.
2116 *
2117 * The range items tell us which parts of the key space the log
2118 * is authoritative for. During replay, if a key in the subvolume
2119 * directory is in a logged range item, but not actually in the log
2120 * that means it was deleted from the directory before the fsync
2121 * and should be removed.
2122 */
2123static noinline int find_dir_range(struct btrfs_root *root,
2124 struct btrfs_path *path,
2125 u64 dirid, int key_type,
2126 u64 *start_ret, u64 *end_ret)
2127{
2128 struct btrfs_key key;
2129 u64 found_end;
2130 struct btrfs_dir_log_item *item;
2131 int ret;
2132 int nritems;
2133
2134 if (*start_ret == (u64)-1)
2135 return 1;
2136
2137 key.objectid = dirid;
2138 key.type = key_type;
2139 key.offset = *start_ret;
2140
2141 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2142 if (ret < 0)
2143 goto out;
2144 if (ret > 0) {
2145 if (path->slots[0] == 0)
2146 goto out;
2147 path->slots[0]--;
2148 }
2149 if (ret != 0)
2150 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2151
2152 if (key.type != key_type || key.objectid != dirid) {
2153 ret = 1;
2154 goto next;
2155 }
2156 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2157 struct btrfs_dir_log_item);
2158 found_end = btrfs_dir_log_end(path->nodes[0], item);
2159
2160 if (*start_ret >= key.offset && *start_ret <= found_end) {
2161 ret = 0;
2162 *start_ret = key.offset;
2163 *end_ret = found_end;
2164 goto out;
2165 }
2166 ret = 1;
2167next:
2168 /* check the next slot in the tree to see if it is a valid item */
2169 nritems = btrfs_header_nritems(path->nodes[0]);
2170 path->slots[0]++;
2171 if (path->slots[0] >= nritems) {
2172 ret = btrfs_next_leaf(root, path);
2173 if (ret)
2174 goto out;
2175 }
2176
2177 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2178
2179 if (key.type != key_type || key.objectid != dirid) {
2180 ret = 1;
2181 goto out;
2182 }
2183 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2184 struct btrfs_dir_log_item);
2185 found_end = btrfs_dir_log_end(path->nodes[0], item);
2186 *start_ret = key.offset;
2187 *end_ret = found_end;
2188 ret = 0;
2189out:
2190 btrfs_release_path(path);
2191 return ret;
2192}
2193
2194/*
2195 * this looks for a given directory item in the log. If the directory
2196 * item is not in the log, the item is removed and the inode it points
2197 * to is unlinked
2198 */
2199static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
2200 struct btrfs_root *root,
2201 struct btrfs_root *log,
2202 struct btrfs_path *path,
2203 struct btrfs_path *log_path,
2204 struct inode *dir,
2205 struct btrfs_key *dir_key)
2206{
2207 int ret;
2208 struct extent_buffer *eb;
2209 int slot;
2210 u32 item_size;
2211 struct btrfs_dir_item *di;
2212 struct btrfs_dir_item *log_di;
2213 int name_len;
2214 unsigned long ptr;
2215 unsigned long ptr_end;
2216 char *name;
2217 struct inode *inode;
2218 struct btrfs_key location;
2219
2220again:
2221 eb = path->nodes[0];
2222 slot = path->slots[0];
2223 item_size = btrfs_item_size_nr(eb, slot);
2224 ptr = btrfs_item_ptr_offset(eb, slot);
2225 ptr_end = ptr + item_size;
2226 while (ptr < ptr_end) {
2227 di = (struct btrfs_dir_item *)ptr;
2228 name_len = btrfs_dir_name_len(eb, di);
2229 name = kmalloc(name_len, GFP_NOFS);
2230 if (!name) {
2231 ret = -ENOMEM;
2232 goto out;
2233 }
2234 read_extent_buffer(eb, name, (unsigned long)(di + 1),
2235 name_len);
2236 log_di = NULL;
2237 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
2238 log_di = btrfs_lookup_dir_item(trans, log, log_path,
2239 dir_key->objectid,
2240 name, name_len, 0);
2241 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
2242 log_di = btrfs_lookup_dir_index_item(trans, log,
2243 log_path,
2244 dir_key->objectid,
2245 dir_key->offset,
2246 name, name_len, 0);
2247 }
2248 if (!log_di || log_di == ERR_PTR(-ENOENT)) {
2249 btrfs_dir_item_key_to_cpu(eb, di, &location);
2250 btrfs_release_path(path);
2251 btrfs_release_path(log_path);
2252 inode = read_one_inode(root, location.objectid);
2253 if (!inode) {
2254 kfree(name);
2255 return -EIO;
2256 }
2257
2258 ret = link_to_fixup_dir(trans, root,
2259 path, location.objectid);
2260 if (ret) {
2261 kfree(name);
2262 iput(inode);
2263 goto out;
2264 }
2265
2266 inc_nlink(inode);
2267 ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
2268 BTRFS_I(inode), name, name_len);
2269 if (!ret)
2270 ret = btrfs_run_delayed_items(trans);
2271 kfree(name);
2272 iput(inode);
2273 if (ret)
2274 goto out;
2275
2276 /* there might still be more names under this key
2277 * check and repeat if required
2278 */
2279 ret = btrfs_search_slot(NULL, root, dir_key, path,
2280 0, 0);
2281 if (ret == 0)
2282 goto again;
2283 ret = 0;
2284 goto out;
2285 } else if (IS_ERR(log_di)) {
2286 kfree(name);
2287 return PTR_ERR(log_di);
2288 }
2289 btrfs_release_path(log_path);
2290 kfree(name);
2291
2292 ptr = (unsigned long)(di + 1);
2293 ptr += name_len;
2294 }
2295 ret = 0;
2296out:
2297 btrfs_release_path(path);
2298 btrfs_release_path(log_path);
2299 return ret;
2300}
2301
2302static int replay_xattr_deletes(struct btrfs_trans_handle *trans,
2303 struct btrfs_root *root,
2304 struct btrfs_root *log,
2305 struct btrfs_path *path,
2306 const u64 ino)
2307{
2308 struct btrfs_key search_key;
2309 struct btrfs_path *log_path;
2310 int i;
2311 int nritems;
2312 int ret;
2313
2314 log_path = btrfs_alloc_path();
2315 if (!log_path)
2316 return -ENOMEM;
2317
2318 search_key.objectid = ino;
2319 search_key.type = BTRFS_XATTR_ITEM_KEY;
2320 search_key.offset = 0;
2321again:
2322 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
2323 if (ret < 0)
2324 goto out;
2325process_leaf:
2326 nritems = btrfs_header_nritems(path->nodes[0]);
2327 for (i = path->slots[0]; i < nritems; i++) {
2328 struct btrfs_key key;
2329 struct btrfs_dir_item *di;
2330 struct btrfs_dir_item *log_di;
2331 u32 total_size;
2332 u32 cur;
2333
2334 btrfs_item_key_to_cpu(path->nodes[0], &key, i);
2335 if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY) {
2336 ret = 0;
2337 goto out;
2338 }
2339
2340 di = btrfs_item_ptr(path->nodes[0], i, struct btrfs_dir_item);
2341 total_size = btrfs_item_size_nr(path->nodes[0], i);
2342 cur = 0;
2343 while (cur < total_size) {
2344 u16 name_len = btrfs_dir_name_len(path->nodes[0], di);
2345 u16 data_len = btrfs_dir_data_len(path->nodes[0], di);
2346 u32 this_len = sizeof(*di) + name_len + data_len;
2347 char *name;
2348
2349 name = kmalloc(name_len, GFP_NOFS);
2350 if (!name) {
2351 ret = -ENOMEM;
2352 goto out;
2353 }
2354 read_extent_buffer(path->nodes[0], name,
2355 (unsigned long)(di + 1), name_len);
2356
2357 log_di = btrfs_lookup_xattr(NULL, log, log_path, ino,
2358 name, name_len, 0);
2359 btrfs_release_path(log_path);
2360 if (!log_di) {
2361 /* Doesn't exist in log tree, so delete it. */
2362 btrfs_release_path(path);
2363 di = btrfs_lookup_xattr(trans, root, path, ino,
2364 name, name_len, -1);
2365 kfree(name);
2366 if (IS_ERR(di)) {
2367 ret = PTR_ERR(di);
2368 goto out;
2369 }
2370 ASSERT(di);
2371 ret = btrfs_delete_one_dir_name(trans, root,
2372 path, di);
2373 if (ret)
2374 goto out;
2375 btrfs_release_path(path);
2376 search_key = key;
2377 goto again;
2378 }
2379 kfree(name);
2380 if (IS_ERR(log_di)) {
2381 ret = PTR_ERR(log_di);
2382 goto out;
2383 }
2384 cur += this_len;
2385 di = (struct btrfs_dir_item *)((char *)di + this_len);
2386 }
2387 }
2388 ret = btrfs_next_leaf(root, path);
2389 if (ret > 0)
2390 ret = 0;
2391 else if (ret == 0)
2392 goto process_leaf;
2393out:
2394 btrfs_free_path(log_path);
2395 btrfs_release_path(path);
2396 return ret;
2397}
2398
2399
2400/*
2401 * deletion replay happens before we copy any new directory items
2402 * out of the log or out of backreferences from inodes. It
2403 * scans the log to find ranges of keys that log is authoritative for,
2404 * and then scans the directory to find items in those ranges that are
2405 * not present in the log.
2406 *
2407 * Anything we don't find in the log is unlinked and removed from the
2408 * directory.
2409 */
2410static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
2411 struct btrfs_root *root,
2412 struct btrfs_root *log,
2413 struct btrfs_path *path,
2414 u64 dirid, int del_all)
2415{
2416 u64 range_start;
2417 u64 range_end;
2418 int key_type = BTRFS_DIR_LOG_ITEM_KEY;
2419 int ret = 0;
2420 struct btrfs_key dir_key;
2421 struct btrfs_key found_key;
2422 struct btrfs_path *log_path;
2423 struct inode *dir;
2424
2425 dir_key.objectid = dirid;
2426 dir_key.type = BTRFS_DIR_ITEM_KEY;
2427 log_path = btrfs_alloc_path();
2428 if (!log_path)
2429 return -ENOMEM;
2430
2431 dir = read_one_inode(root, dirid);
2432 /* it isn't an error if the inode isn't there, that can happen
2433 * because we replay the deletes before we copy in the inode item
2434 * from the log
2435 */
2436 if (!dir) {
2437 btrfs_free_path(log_path);
2438 return 0;
2439 }
2440again:
2441 range_start = 0;
2442 range_end = 0;
2443 while (1) {
2444 if (del_all)
2445 range_end = (u64)-1;
2446 else {
2447 ret = find_dir_range(log, path, dirid, key_type,
2448 &range_start, &range_end);
2449 if (ret != 0)
2450 break;
2451 }
2452
2453 dir_key.offset = range_start;
2454 while (1) {
2455 int nritems;
2456 ret = btrfs_search_slot(NULL, root, &dir_key, path,
2457 0, 0);
2458 if (ret < 0)
2459 goto out;
2460
2461 nritems = btrfs_header_nritems(path->nodes[0]);
2462 if (path->slots[0] >= nritems) {
2463 ret = btrfs_next_leaf(root, path);
2464 if (ret == 1)
2465 break;
2466 else if (ret < 0)
2467 goto out;
2468 }
2469 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2470 path->slots[0]);
2471 if (found_key.objectid != dirid ||
2472 found_key.type != dir_key.type)
2473 goto next_type;
2474
2475 if (found_key.offset > range_end)
2476 break;
2477
2478 ret = check_item_in_log(trans, root, log, path,
2479 log_path, dir,
2480 &found_key);
2481 if (ret)
2482 goto out;
2483 if (found_key.offset == (u64)-1)
2484 break;
2485 dir_key.offset = found_key.offset + 1;
2486 }
2487 btrfs_release_path(path);
2488 if (range_end == (u64)-1)
2489 break;
2490 range_start = range_end + 1;
2491 }
2492
2493next_type:
2494 ret = 0;
2495 if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
2496 key_type = BTRFS_DIR_LOG_INDEX_KEY;
2497 dir_key.type = BTRFS_DIR_INDEX_KEY;
2498 btrfs_release_path(path);
2499 goto again;
2500 }
2501out:
2502 btrfs_release_path(path);
2503 btrfs_free_path(log_path);
2504 iput(dir);
2505 return ret;
2506}
2507
2508/*
2509 * the process_func used to replay items from the log tree. This
2510 * gets called in two different stages. The first stage just looks
2511 * for inodes and makes sure they are all copied into the subvolume.
2512 *
2513 * The second stage copies all the other item types from the log into
2514 * the subvolume. The two stage approach is slower, but gets rid of
2515 * lots of complexity around inodes referencing other inodes that exist
2516 * only in the log (references come from either directory items or inode
2517 * back refs).
2518 */
2519static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
2520 struct walk_control *wc, u64 gen, int level)
2521{
2522 int nritems;
2523 struct btrfs_path *path;
2524 struct btrfs_root *root = wc->replay_dest;
2525 struct btrfs_key key;
2526 int i;
2527 int ret;
2528
2529 ret = btrfs_read_buffer(eb, gen, level, NULL);
2530 if (ret)
2531 return ret;
2532
2533 level = btrfs_header_level(eb);
2534
2535 if (level != 0)
2536 return 0;
2537
2538 path = btrfs_alloc_path();
2539 if (!path)
2540 return -ENOMEM;
2541
2542 nritems = btrfs_header_nritems(eb);
2543 for (i = 0; i < nritems; i++) {
2544 btrfs_item_key_to_cpu(eb, &key, i);
2545
2546 /* inode keys are done during the first stage */
2547 if (key.type == BTRFS_INODE_ITEM_KEY &&
2548 wc->stage == LOG_WALK_REPLAY_INODES) {
2549 struct btrfs_inode_item *inode_item;
2550 u32 mode;
2551
2552 inode_item = btrfs_item_ptr(eb, i,
2553 struct btrfs_inode_item);
2554 /*
2555 * If we have a tmpfile (O_TMPFILE) that got fsync'ed
2556 * and never got linked before the fsync, skip it, as
2557 * replaying it is pointless since it would be deleted
2558 * later. We skip logging tmpfiles, but it's always
2559 * possible we are replaying a log created with a kernel
2560 * that used to log tmpfiles.
2561 */
2562 if (btrfs_inode_nlink(eb, inode_item) == 0) {
2563 wc->ignore_cur_inode = true;
2564 continue;
2565 } else {
2566 wc->ignore_cur_inode = false;
2567 }
2568 ret = replay_xattr_deletes(wc->trans, root, log,
2569 path, key.objectid);
2570 if (ret)
2571 break;
2572 mode = btrfs_inode_mode(eb, inode_item);
2573 if (S_ISDIR(mode)) {
2574 ret = replay_dir_deletes(wc->trans,
2575 root, log, path, key.objectid, 0);
2576 if (ret)
2577 break;
2578 }
2579 ret = overwrite_item(wc->trans, root, path,
2580 eb, i, &key);
2581 if (ret)
2582 break;
2583
2584 /*
2585 * Before replaying extents, truncate the inode to its
2586 * size. We need to do it now and not after log replay
2587 * because before an fsync we can have prealloc extents
2588 * added beyond the inode's i_size. If we did it after,
2589 * through orphan cleanup for example, we would drop
2590 * those prealloc extents just after replaying them.
2591 */
2592 if (S_ISREG(mode)) {
2593 struct inode *inode;
2594 u64 from;
2595
2596 inode = read_one_inode(root, key.objectid);
2597 if (!inode) {
2598 ret = -EIO;
2599 break;
2600 }
2601 from = ALIGN(i_size_read(inode),
2602 root->fs_info->sectorsize);
2603 ret = btrfs_drop_extents(wc->trans, root, inode,
2604 from, (u64)-1, 1);
2605 if (!ret) {
2606 /* Update the inode's nbytes. */
2607 ret = btrfs_update_inode(wc->trans,
2608 root, inode);
2609 }
2610 iput(inode);
2611 if (ret)
2612 break;
2613 }
2614
2615 ret = link_to_fixup_dir(wc->trans, root,
2616 path, key.objectid);
2617 if (ret)
2618 break;
2619 }
2620
2621 if (wc->ignore_cur_inode)
2622 continue;
2623
2624 if (key.type == BTRFS_DIR_INDEX_KEY &&
2625 wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
2626 ret = replay_one_dir_item(wc->trans, root, path,
2627 eb, i, &key);
2628 if (ret)
2629 break;
2630 }
2631
2632 if (wc->stage < LOG_WALK_REPLAY_ALL)
2633 continue;
2634
2635 /* these keys are simply copied */
2636 if (key.type == BTRFS_XATTR_ITEM_KEY) {
2637 ret = overwrite_item(wc->trans, root, path,
2638 eb, i, &key);
2639 if (ret)
2640 break;
2641 } else if (key.type == BTRFS_INODE_REF_KEY ||
2642 key.type == BTRFS_INODE_EXTREF_KEY) {
2643 ret = add_inode_ref(wc->trans, root, log, path,
2644 eb, i, &key);
2645 if (ret && ret != -ENOENT)
2646 break;
2647 ret = 0;
2648 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
2649 ret = replay_one_extent(wc->trans, root, path,
2650 eb, i, &key);
2651 if (ret)
2652 break;
2653 } else if (key.type == BTRFS_DIR_ITEM_KEY) {
2654 ret = replay_one_dir_item(wc->trans, root, path,
2655 eb, i, &key);
2656 if (ret)
2657 break;
2658 }
2659 }
2660 btrfs_free_path(path);
2661 return ret;
2662}
2663
2664/*
2665 * Correctly adjust the reserved bytes occupied by a log tree extent buffer
2666 */
2667static void unaccount_log_buffer(struct btrfs_fs_info *fs_info, u64 start)
2668{
2669 struct btrfs_block_group *cache;
2670
2671 cache = btrfs_lookup_block_group(fs_info, start);
2672 if (!cache) {
2673 btrfs_err(fs_info, "unable to find block group for %llu", start);
2674 return;
2675 }
2676
2677 spin_lock(&cache->space_info->lock);
2678 spin_lock(&cache->lock);
2679 cache->reserved -= fs_info->nodesize;
2680 cache->space_info->bytes_reserved -= fs_info->nodesize;
2681 spin_unlock(&cache->lock);
2682 spin_unlock(&cache->space_info->lock);
2683
2684 btrfs_put_block_group(cache);
2685}
2686
2687static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
2688 struct btrfs_root *root,
2689 struct btrfs_path *path, int *level,
2690 struct walk_control *wc)
2691{
2692 struct btrfs_fs_info *fs_info = root->fs_info;
2693 u64 bytenr;
2694 u64 ptr_gen;
2695 struct extent_buffer *next;
2696 struct extent_buffer *cur;
2697 u32 blocksize;
2698 int ret = 0;
2699
2700 while (*level > 0) {
2701 struct btrfs_key first_key;
2702
2703 cur = path->nodes[*level];
2704
2705 WARN_ON(btrfs_header_level(cur) != *level);
2706
2707 if (path->slots[*level] >=
2708 btrfs_header_nritems(cur))
2709 break;
2710
2711 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
2712 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
2713 btrfs_node_key_to_cpu(cur, &first_key, path->slots[*level]);
2714 blocksize = fs_info->nodesize;
2715
2716 next = btrfs_find_create_tree_block(fs_info, bytenr);
2717 if (IS_ERR(next))
2718 return PTR_ERR(next);
2719
2720 if (*level == 1) {
2721 ret = wc->process_func(root, next, wc, ptr_gen,
2722 *level - 1);
2723 if (ret) {
2724 free_extent_buffer(next);
2725 return ret;
2726 }
2727
2728 path->slots[*level]++;
2729 if (wc->free) {
2730 ret = btrfs_read_buffer(next, ptr_gen,
2731 *level - 1, &first_key);
2732 if (ret) {
2733 free_extent_buffer(next);
2734 return ret;
2735 }
2736
2737 if (trans) {
2738 btrfs_tree_lock(next);
2739 btrfs_set_lock_blocking_write(next);
2740 btrfs_clean_tree_block(next);
2741 btrfs_wait_tree_block_writeback(next);
2742 btrfs_tree_unlock(next);
2743 ret = btrfs_pin_reserved_extent(trans,
2744 bytenr, blocksize);
2745 if (ret) {
2746 free_extent_buffer(next);
2747 return ret;
2748 }
2749 } else {
2750 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
2751 clear_extent_buffer_dirty(next);
2752 unaccount_log_buffer(fs_info, bytenr);
2753 }
2754 }
2755 free_extent_buffer(next);
2756 continue;
2757 }
2758 ret = btrfs_read_buffer(next, ptr_gen, *level - 1, &first_key);
2759 if (ret) {
2760 free_extent_buffer(next);
2761 return ret;
2762 }
2763
2764 if (path->nodes[*level-1])
2765 free_extent_buffer(path->nodes[*level-1]);
2766 path->nodes[*level-1] = next;
2767 *level = btrfs_header_level(next);
2768 path->slots[*level] = 0;
2769 cond_resched();
2770 }
2771 path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2772
2773 cond_resched();
2774 return 0;
2775}
2776
2777static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2778 struct btrfs_root *root,
2779 struct btrfs_path *path, int *level,
2780 struct walk_control *wc)
2781{
2782 struct btrfs_fs_info *fs_info = root->fs_info;
2783 int i;
2784 int slot;
2785 int ret;
2786
2787 for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2788 slot = path->slots[i];
2789 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2790 path->slots[i]++;
2791 *level = i;
2792 WARN_ON(*level == 0);
2793 return 0;
2794 } else {
2795 ret = wc->process_func(root, path->nodes[*level], wc,
2796 btrfs_header_generation(path->nodes[*level]),
2797 *level);
2798 if (ret)
2799 return ret;
2800
2801 if (wc->free) {
2802 struct extent_buffer *next;
2803
2804 next = path->nodes[*level];
2805
2806 if (trans) {
2807 btrfs_tree_lock(next);
2808 btrfs_set_lock_blocking_write(next);
2809 btrfs_clean_tree_block(next);
2810 btrfs_wait_tree_block_writeback(next);
2811 btrfs_tree_unlock(next);
2812 ret = btrfs_pin_reserved_extent(trans,
2813 path->nodes[*level]->start,
2814 path->nodes[*level]->len);
2815 if (ret)
2816 return ret;
2817 } else {
2818 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
2819 clear_extent_buffer_dirty(next);
2820
2821 unaccount_log_buffer(fs_info,
2822 path->nodes[*level]->start);
2823 }
2824 }
2825 free_extent_buffer(path->nodes[*level]);
2826 path->nodes[*level] = NULL;
2827 *level = i + 1;
2828 }
2829 }
2830 return 1;
2831}
2832
2833/*
2834 * drop the reference count on the tree rooted at 'snap'. This traverses
2835 * the tree freeing any blocks that have a ref count of zero after being
2836 * decremented.
2837 */
2838static int walk_log_tree(struct btrfs_trans_handle *trans,
2839 struct btrfs_root *log, struct walk_control *wc)
2840{
2841 struct btrfs_fs_info *fs_info = log->fs_info;
2842 int ret = 0;
2843 int wret;
2844 int level;
2845 struct btrfs_path *path;
2846 int orig_level;
2847
2848 path = btrfs_alloc_path();
2849 if (!path)
2850 return -ENOMEM;
2851
2852 level = btrfs_header_level(log->node);
2853 orig_level = level;
2854 path->nodes[level] = log->node;
2855 atomic_inc(&log->node->refs);
2856 path->slots[level] = 0;
2857
2858 while (1) {
2859 wret = walk_down_log_tree(trans, log, path, &level, wc);
2860 if (wret > 0)
2861 break;
2862 if (wret < 0) {
2863 ret = wret;
2864 goto out;
2865 }
2866
2867 wret = walk_up_log_tree(trans, log, path, &level, wc);
2868 if (wret > 0)
2869 break;
2870 if (wret < 0) {
2871 ret = wret;
2872 goto out;
2873 }
2874 }
2875
2876 /* was the root node processed? if not, catch it here */
2877 if (path->nodes[orig_level]) {
2878 ret = wc->process_func(log, path->nodes[orig_level], wc,
2879 btrfs_header_generation(path->nodes[orig_level]),
2880 orig_level);
2881 if (ret)
2882 goto out;
2883 if (wc->free) {
2884 struct extent_buffer *next;
2885
2886 next = path->nodes[orig_level];
2887
2888 if (trans) {
2889 btrfs_tree_lock(next);
2890 btrfs_set_lock_blocking_write(next);
2891 btrfs_clean_tree_block(next);
2892 btrfs_wait_tree_block_writeback(next);
2893 btrfs_tree_unlock(next);
2894 ret = btrfs_pin_reserved_extent(trans,
2895 next->start, next->len);
2896 if (ret)
2897 goto out;
2898 } else {
2899 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
2900 clear_extent_buffer_dirty(next);
2901 unaccount_log_buffer(fs_info, next->start);
2902 }
2903 }
2904 }
2905
2906out:
2907 btrfs_free_path(path);
2908 return ret;
2909}
2910
2911/*
2912 * helper function to update the item for a given subvolumes log root
2913 * in the tree of log roots
2914 */
2915static int update_log_root(struct btrfs_trans_handle *trans,
2916 struct btrfs_root *log,
2917 struct btrfs_root_item *root_item)
2918{
2919 struct btrfs_fs_info *fs_info = log->fs_info;
2920 int ret;
2921
2922 if (log->log_transid == 1) {
2923 /* insert root item on the first sync */
2924 ret = btrfs_insert_root(trans, fs_info->log_root_tree,
2925 &log->root_key, root_item);
2926 } else {
2927 ret = btrfs_update_root(trans, fs_info->log_root_tree,
2928 &log->root_key, root_item);
2929 }
2930 return ret;
2931}
2932
2933static void wait_log_commit(struct btrfs_root *root, int transid)
2934{
2935 DEFINE_WAIT(wait);
2936 int index = transid % 2;
2937
2938 /*
2939 * we only allow two pending log transactions at a time,
2940 * so we know that if ours is more than 2 older than the
2941 * current transaction, we're done
2942 */
2943 for (;;) {
2944 prepare_to_wait(&root->log_commit_wait[index],
2945 &wait, TASK_UNINTERRUPTIBLE);
2946
2947 if (!(root->log_transid_committed < transid &&
2948 atomic_read(&root->log_commit[index])))
2949 break;
2950
2951 mutex_unlock(&root->log_mutex);
2952 schedule();
2953 mutex_lock(&root->log_mutex);
2954 }
2955 finish_wait(&root->log_commit_wait[index], &wait);
2956}
2957
2958static void wait_for_writer(struct btrfs_root *root)
2959{
2960 DEFINE_WAIT(wait);
2961
2962 for (;;) {
2963 prepare_to_wait(&root->log_writer_wait, &wait,
2964 TASK_UNINTERRUPTIBLE);
2965 if (!atomic_read(&root->log_writers))
2966 break;
2967
2968 mutex_unlock(&root->log_mutex);
2969 schedule();
2970 mutex_lock(&root->log_mutex);
2971 }
2972 finish_wait(&root->log_writer_wait, &wait);
2973}
2974
2975static inline void btrfs_remove_log_ctx(struct btrfs_root *root,
2976 struct btrfs_log_ctx *ctx)
2977{
2978 if (!ctx)
2979 return;
2980
2981 mutex_lock(&root->log_mutex);
2982 list_del_init(&ctx->list);
2983 mutex_unlock(&root->log_mutex);
2984}
2985
2986/*
2987 * Invoked in log mutex context, or be sure there is no other task which
2988 * can access the list.
2989 */
2990static inline void btrfs_remove_all_log_ctxs(struct btrfs_root *root,
2991 int index, int error)
2992{
2993 struct btrfs_log_ctx *ctx;
2994 struct btrfs_log_ctx *safe;
2995
2996 list_for_each_entry_safe(ctx, safe, &root->log_ctxs[index], list) {
2997 list_del_init(&ctx->list);
2998 ctx->log_ret = error;
2999 }
3000
3001 INIT_LIST_HEAD(&root->log_ctxs[index]);
3002}
3003
3004/*
3005 * btrfs_sync_log does sends a given tree log down to the disk and
3006 * updates the super blocks to record it. When this call is done,
3007 * you know that any inodes previously logged are safely on disk only
3008 * if it returns 0.
3009 *
3010 * Any other return value means you need to call btrfs_commit_transaction.
3011 * Some of the edge cases for fsyncing directories that have had unlinks
3012 * or renames done in the past mean that sometimes the only safe
3013 * fsync is to commit the whole FS. When btrfs_sync_log returns -EAGAIN,
3014 * that has happened.
3015 */
3016int btrfs_sync_log(struct btrfs_trans_handle *trans,
3017 struct btrfs_root *root, struct btrfs_log_ctx *ctx)
3018{
3019 int index1;
3020 int index2;
3021 int mark;
3022 int ret;
3023 struct btrfs_fs_info *fs_info = root->fs_info;
3024 struct btrfs_root *log = root->log_root;
3025 struct btrfs_root *log_root_tree = fs_info->log_root_tree;
3026 struct btrfs_root_item new_root_item;
3027 int log_transid = 0;
3028 struct btrfs_log_ctx root_log_ctx;
3029 struct blk_plug plug;
3030
3031 mutex_lock(&root->log_mutex);
3032 log_transid = ctx->log_transid;
3033 if (root->log_transid_committed >= log_transid) {
3034 mutex_unlock(&root->log_mutex);
3035 return ctx->log_ret;
3036 }
3037
3038 index1 = log_transid % 2;
3039 if (atomic_read(&root->log_commit[index1])) {
3040 wait_log_commit(root, log_transid);
3041 mutex_unlock(&root->log_mutex);
3042 return ctx->log_ret;
3043 }
3044 ASSERT(log_transid == root->log_transid);
3045 atomic_set(&root->log_commit[index1], 1);
3046
3047 /* wait for previous tree log sync to complete */
3048 if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
3049 wait_log_commit(root, log_transid - 1);
3050
3051 while (1) {
3052 int batch = atomic_read(&root->log_batch);
3053 /* when we're on an ssd, just kick the log commit out */
3054 if (!btrfs_test_opt(fs_info, SSD) &&
3055 test_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state)) {
3056 mutex_unlock(&root->log_mutex);
3057 schedule_timeout_uninterruptible(1);
3058 mutex_lock(&root->log_mutex);
3059 }
3060 wait_for_writer(root);
3061 if (batch == atomic_read(&root->log_batch))
3062 break;
3063 }
3064
3065 /* bail out if we need to do a full commit */
3066 if (btrfs_need_log_full_commit(trans)) {
3067 ret = -EAGAIN;
3068 mutex_unlock(&root->log_mutex);
3069 goto out;
3070 }
3071
3072 if (log_transid % 2 == 0)
3073 mark = EXTENT_DIRTY;
3074 else
3075 mark = EXTENT_NEW;
3076
3077 /* we start IO on all the marked extents here, but we don't actually
3078 * wait for them until later.
3079 */
3080 blk_start_plug(&plug);
3081 ret = btrfs_write_marked_extents(fs_info, &log->dirty_log_pages, mark);
3082 if (ret) {
3083 blk_finish_plug(&plug);
3084 btrfs_abort_transaction(trans, ret);
3085 btrfs_set_log_full_commit(trans);
3086 mutex_unlock(&root->log_mutex);
3087 goto out;
3088 }
3089
3090 /*
3091 * We _must_ update under the root->log_mutex in order to make sure we
3092 * have a consistent view of the log root we are trying to commit at
3093 * this moment.
3094 *
3095 * We _must_ copy this into a local copy, because we are not holding the
3096 * log_root_tree->log_mutex yet. This is important because when we
3097 * commit the log_root_tree we must have a consistent view of the
3098 * log_root_tree when we update the super block to point at the
3099 * log_root_tree bytenr. If we update the log_root_tree here we'll race
3100 * with the commit and possibly point at the new block which we may not
3101 * have written out.
3102 */
3103 btrfs_set_root_node(&log->root_item, log->node);
3104 memcpy(&new_root_item, &log->root_item, sizeof(new_root_item));
3105
3106 root->log_transid++;
3107 log->log_transid = root->log_transid;
3108 root->log_start_pid = 0;
3109 /*
3110 * IO has been started, blocks of the log tree have WRITTEN flag set
3111 * in their headers. new modifications of the log will be written to
3112 * new positions. so it's safe to allow log writers to go in.
3113 */
3114 mutex_unlock(&root->log_mutex);
3115
3116 btrfs_init_log_ctx(&root_log_ctx, NULL);
3117
3118 mutex_lock(&log_root_tree->log_mutex);
3119
3120 index2 = log_root_tree->log_transid % 2;
3121 list_add_tail(&root_log_ctx.list, &log_root_tree->log_ctxs[index2]);
3122 root_log_ctx.log_transid = log_root_tree->log_transid;
3123
3124 /*
3125 * Now we are safe to update the log_root_tree because we're under the
3126 * log_mutex, and we're a current writer so we're holding the commit
3127 * open until we drop the log_mutex.
3128 */
3129 ret = update_log_root(trans, log, &new_root_item);
3130 if (ret) {
3131 if (!list_empty(&root_log_ctx.list))
3132 list_del_init(&root_log_ctx.list);
3133
3134 blk_finish_plug(&plug);
3135 btrfs_set_log_full_commit(trans);
3136
3137 if (ret != -ENOSPC) {
3138 btrfs_abort_transaction(trans, ret);
3139 mutex_unlock(&log_root_tree->log_mutex);
3140 goto out;
3141 }
3142 btrfs_wait_tree_log_extents(log, mark);
3143 mutex_unlock(&log_root_tree->log_mutex);
3144 ret = -EAGAIN;
3145 goto out;
3146 }
3147
3148 if (log_root_tree->log_transid_committed >= root_log_ctx.log_transid) {
3149 blk_finish_plug(&plug);
3150 list_del_init(&root_log_ctx.list);
3151 mutex_unlock(&log_root_tree->log_mutex);
3152 ret = root_log_ctx.log_ret;
3153 goto out;
3154 }
3155
3156 index2 = root_log_ctx.log_transid % 2;
3157 if (atomic_read(&log_root_tree->log_commit[index2])) {
3158 blk_finish_plug(&plug);
3159 ret = btrfs_wait_tree_log_extents(log, mark);
3160 wait_log_commit(log_root_tree,
3161 root_log_ctx.log_transid);
3162 mutex_unlock(&log_root_tree->log_mutex);
3163 if (!ret)
3164 ret = root_log_ctx.log_ret;
3165 goto out;
3166 }
3167 ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
3168 atomic_set(&log_root_tree->log_commit[index2], 1);
3169
3170 if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
3171 wait_log_commit(log_root_tree,
3172 root_log_ctx.log_transid - 1);
3173 }
3174
3175 /*
3176 * now that we've moved on to the tree of log tree roots,
3177 * check the full commit flag again
3178 */
3179 if (btrfs_need_log_full_commit(trans)) {
3180 blk_finish_plug(&plug);
3181 btrfs_wait_tree_log_extents(log, mark);
3182 mutex_unlock(&log_root_tree->log_mutex);
3183 ret = -EAGAIN;
3184 goto out_wake_log_root;
3185 }
3186
3187 ret = btrfs_write_marked_extents(fs_info,
3188 &log_root_tree->dirty_log_pages,
3189 EXTENT_DIRTY | EXTENT_NEW);
3190 blk_finish_plug(&plug);
3191 if (ret) {
3192 btrfs_set_log_full_commit(trans);
3193 btrfs_abort_transaction(trans, ret);
3194 mutex_unlock(&log_root_tree->log_mutex);
3195 goto out_wake_log_root;
3196 }
3197 ret = btrfs_wait_tree_log_extents(log, mark);
3198 if (!ret)
3199 ret = btrfs_wait_tree_log_extents(log_root_tree,
3200 EXTENT_NEW | EXTENT_DIRTY);
3201 if (ret) {
3202 btrfs_set_log_full_commit(trans);
3203 mutex_unlock(&log_root_tree->log_mutex);
3204 goto out_wake_log_root;
3205 }
3206
3207 btrfs_set_super_log_root(fs_info->super_for_commit,
3208 log_root_tree->node->start);
3209 btrfs_set_super_log_root_level(fs_info->super_for_commit,
3210 btrfs_header_level(log_root_tree->node));
3211
3212 log_root_tree->log_transid++;
3213 mutex_unlock(&log_root_tree->log_mutex);
3214
3215 /*
3216 * Nobody else is going to jump in and write the ctree
3217 * super here because the log_commit atomic below is protecting
3218 * us. We must be called with a transaction handle pinning
3219 * the running transaction open, so a full commit can't hop
3220 * in and cause problems either.
3221 */
3222 ret = write_all_supers(fs_info, 1);
3223 if (ret) {
3224 btrfs_set_log_full_commit(trans);
3225 btrfs_abort_transaction(trans, ret);
3226 goto out_wake_log_root;
3227 }
3228
3229 mutex_lock(&root->log_mutex);
3230 if (root->last_log_commit < log_transid)
3231 root->last_log_commit = log_transid;
3232 mutex_unlock(&root->log_mutex);
3233
3234out_wake_log_root:
3235 mutex_lock(&log_root_tree->log_mutex);
3236 btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
3237
3238 log_root_tree->log_transid_committed++;
3239 atomic_set(&log_root_tree->log_commit[index2], 0);
3240 mutex_unlock(&log_root_tree->log_mutex);
3241
3242 /*
3243 * The barrier before waitqueue_active (in cond_wake_up) is needed so
3244 * all the updates above are seen by the woken threads. It might not be
3245 * necessary, but proving that seems to be hard.
3246 */
3247 cond_wake_up(&log_root_tree->log_commit_wait[index2]);
3248out:
3249 mutex_lock(&root->log_mutex);
3250 btrfs_remove_all_log_ctxs(root, index1, ret);
3251 root->log_transid_committed++;
3252 atomic_set(&root->log_commit[index1], 0);
3253 mutex_unlock(&root->log_mutex);
3254
3255 /*
3256 * The barrier before waitqueue_active (in cond_wake_up) is needed so
3257 * all the updates above are seen by the woken threads. It might not be
3258 * necessary, but proving that seems to be hard.
3259 */
3260 cond_wake_up(&root->log_commit_wait[index1]);
3261 return ret;
3262}
3263
3264static void free_log_tree(struct btrfs_trans_handle *trans,
3265 struct btrfs_root *log)
3266{
3267 int ret;
3268 struct walk_control wc = {
3269 .free = 1,
3270 .process_func = process_one_buffer
3271 };
3272
3273 ret = walk_log_tree(trans, log, &wc);
3274 if (ret) {
3275 if (trans)
3276 btrfs_abort_transaction(trans, ret);
3277 else
3278 btrfs_handle_fs_error(log->fs_info, ret, NULL);
3279 }
3280
3281 clear_extent_bits(&log->dirty_log_pages, 0, (u64)-1,
3282 EXTENT_DIRTY | EXTENT_NEW | EXTENT_NEED_WAIT);
3283 extent_io_tree_release(&log->log_csum_range);
3284 btrfs_put_root(log);
3285}
3286
3287/*
3288 * free all the extents used by the tree log. This should be called
3289 * at commit time of the full transaction
3290 */
3291int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
3292{
3293 if (root->log_root) {
3294 free_log_tree(trans, root->log_root);
3295 root->log_root = NULL;
3296 clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
3297 }
3298 return 0;
3299}
3300
3301int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
3302 struct btrfs_fs_info *fs_info)
3303{
3304 if (fs_info->log_root_tree) {
3305 free_log_tree(trans, fs_info->log_root_tree);
3306 fs_info->log_root_tree = NULL;
3307 }
3308 return 0;
3309}
3310
3311/*
3312 * Check if an inode was logged in the current transaction. We can't always rely
3313 * on an inode's logged_trans value, because it's an in-memory only field and
3314 * therefore not persisted. This means that its value is lost if the inode gets
3315 * evicted and loaded again from disk (in which case it has a value of 0, and
3316 * certainly it is smaller then any possible transaction ID), when that happens
3317 * the full_sync flag is set in the inode's runtime flags, so on that case we
3318 * assume eviction happened and ignore the logged_trans value, assuming the
3319 * worst case, that the inode was logged before in the current transaction.
3320 */
3321static bool inode_logged(struct btrfs_trans_handle *trans,
3322 struct btrfs_inode *inode)
3323{
3324 if (inode->logged_trans == trans->transid)
3325 return true;
3326
3327 if (inode->last_trans == trans->transid &&
3328 test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) &&
3329 !test_bit(BTRFS_FS_LOG_RECOVERING, &trans->fs_info->flags))
3330 return true;
3331
3332 return false;
3333}
3334
3335/*
3336 * If both a file and directory are logged, and unlinks or renames are
3337 * mixed in, we have a few interesting corners:
3338 *
3339 * create file X in dir Y
3340 * link file X to X.link in dir Y
3341 * fsync file X
3342 * unlink file X but leave X.link
3343 * fsync dir Y
3344 *
3345 * After a crash we would expect only X.link to exist. But file X
3346 * didn't get fsync'd again so the log has back refs for X and X.link.
3347 *
3348 * We solve this by removing directory entries and inode backrefs from the
3349 * log when a file that was logged in the current transaction is
3350 * unlinked. Any later fsync will include the updated log entries, and
3351 * we'll be able to reconstruct the proper directory items from backrefs.
3352 *
3353 * This optimizations allows us to avoid relogging the entire inode
3354 * or the entire directory.
3355 */
3356int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
3357 struct btrfs_root *root,
3358 const char *name, int name_len,
3359 struct btrfs_inode *dir, u64 index)
3360{
3361 struct btrfs_root *log;
3362 struct btrfs_dir_item *di;
3363 struct btrfs_path *path;
3364 int ret;
3365 int err = 0;
3366 int bytes_del = 0;
3367 u64 dir_ino = btrfs_ino(dir);
3368
3369 if (!inode_logged(trans, dir))
3370 return 0;
3371
3372 ret = join_running_log_trans(root);
3373 if (ret)
3374 return 0;
3375
3376 mutex_lock(&dir->log_mutex);
3377
3378 log = root->log_root;
3379 path = btrfs_alloc_path();
3380 if (!path) {
3381 err = -ENOMEM;
3382 goto out_unlock;
3383 }
3384
3385 di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
3386 name, name_len, -1);
3387 if (IS_ERR(di)) {
3388 err = PTR_ERR(di);
3389 goto fail;
3390 }
3391 if (di) {
3392 ret = btrfs_delete_one_dir_name(trans, log, path, di);
3393 bytes_del += name_len;
3394 if (ret) {
3395 err = ret;
3396 goto fail;
3397 }
3398 }
3399 btrfs_release_path(path);
3400 di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
3401 index, name, name_len, -1);
3402 if (IS_ERR(di)) {
3403 err = PTR_ERR(di);
3404 goto fail;
3405 }
3406 if (di) {
3407 ret = btrfs_delete_one_dir_name(trans, log, path, di);
3408 bytes_del += name_len;
3409 if (ret) {
3410 err = ret;
3411 goto fail;
3412 }
3413 }
3414
3415 /* update the directory size in the log to reflect the names
3416 * we have removed
3417 */
3418 if (bytes_del) {
3419 struct btrfs_key key;
3420
3421 key.objectid = dir_ino;
3422 key.offset = 0;
3423 key.type = BTRFS_INODE_ITEM_KEY;
3424 btrfs_release_path(path);
3425
3426 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
3427 if (ret < 0) {
3428 err = ret;
3429 goto fail;
3430 }
3431 if (ret == 0) {
3432 struct btrfs_inode_item *item;
3433 u64 i_size;
3434
3435 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3436 struct btrfs_inode_item);
3437 i_size = btrfs_inode_size(path->nodes[0], item);
3438 if (i_size > bytes_del)
3439 i_size -= bytes_del;
3440 else
3441 i_size = 0;
3442 btrfs_set_inode_size(path->nodes[0], item, i_size);
3443 btrfs_mark_buffer_dirty(path->nodes[0]);
3444 } else
3445 ret = 0;
3446 btrfs_release_path(path);
3447 }
3448fail:
3449 btrfs_free_path(path);
3450out_unlock:
3451 mutex_unlock(&dir->log_mutex);
3452 if (err == -ENOSPC) {
3453 btrfs_set_log_full_commit(trans);
3454 err = 0;
3455 } else if (err < 0 && err != -ENOENT) {
3456 /* ENOENT can be returned if the entry hasn't been fsynced yet */
3457 btrfs_abort_transaction(trans, err);
3458 }
3459
3460 btrfs_end_log_trans(root);
3461
3462 return err;
3463}
3464
3465/* see comments for btrfs_del_dir_entries_in_log */
3466int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
3467 struct btrfs_root *root,
3468 const char *name, int name_len,
3469 struct btrfs_inode *inode, u64 dirid)
3470{
3471 struct btrfs_root *log;
3472 u64 index;
3473 int ret;
3474
3475 if (!inode_logged(trans, inode))
3476 return 0;
3477
3478 ret = join_running_log_trans(root);
3479 if (ret)
3480 return 0;
3481 log = root->log_root;
3482 mutex_lock(&inode->log_mutex);
3483
3484 ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
3485 dirid, &index);
3486 mutex_unlock(&inode->log_mutex);
3487 if (ret == -ENOSPC) {
3488 btrfs_set_log_full_commit(trans);
3489 ret = 0;
3490 } else if (ret < 0 && ret != -ENOENT)
3491 btrfs_abort_transaction(trans, ret);
3492 btrfs_end_log_trans(root);
3493
3494 return ret;
3495}
3496
3497/*
3498 * creates a range item in the log for 'dirid'. first_offset and
3499 * last_offset tell us which parts of the key space the log should
3500 * be considered authoritative for.
3501 */
3502static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
3503 struct btrfs_root *log,
3504 struct btrfs_path *path,
3505 int key_type, u64 dirid,
3506 u64 first_offset, u64 last_offset)
3507{
3508 int ret;
3509 struct btrfs_key key;
3510 struct btrfs_dir_log_item *item;
3511
3512 key.objectid = dirid;
3513 key.offset = first_offset;
3514 if (key_type == BTRFS_DIR_ITEM_KEY)
3515 key.type = BTRFS_DIR_LOG_ITEM_KEY;
3516 else
3517 key.type = BTRFS_DIR_LOG_INDEX_KEY;
3518 ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
3519 if (ret)
3520 return ret;
3521
3522 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3523 struct btrfs_dir_log_item);
3524 btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
3525 btrfs_mark_buffer_dirty(path->nodes[0]);
3526 btrfs_release_path(path);
3527 return 0;
3528}
3529
3530/*
3531 * log all the items included in the current transaction for a given
3532 * directory. This also creates the range items in the log tree required
3533 * to replay anything deleted before the fsync
3534 */
3535static noinline int log_dir_items(struct btrfs_trans_handle *trans,
3536 struct btrfs_root *root, struct btrfs_inode *inode,
3537 struct btrfs_path *path,
3538 struct btrfs_path *dst_path, int key_type,
3539 struct btrfs_log_ctx *ctx,
3540 u64 min_offset, u64 *last_offset_ret)
3541{
3542 struct btrfs_key min_key;
3543 struct btrfs_root *log = root->log_root;
3544 struct extent_buffer *src;
3545 int err = 0;
3546 int ret;
3547 int i;
3548 int nritems;
3549 u64 first_offset = min_offset;
3550 u64 last_offset = (u64)-1;
3551 u64 ino = btrfs_ino(inode);
3552
3553 log = root->log_root;
3554
3555 min_key.objectid = ino;
3556 min_key.type = key_type;
3557 min_key.offset = min_offset;
3558
3559 ret = btrfs_search_forward(root, &min_key, path, trans->transid);
3560
3561 /*
3562 * we didn't find anything from this transaction, see if there
3563 * is anything at all
3564 */
3565 if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
3566 min_key.objectid = ino;
3567 min_key.type = key_type;
3568 min_key.offset = (u64)-1;
3569 btrfs_release_path(path);
3570 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3571 if (ret < 0) {
3572 btrfs_release_path(path);
3573 return ret;
3574 }
3575 ret = btrfs_previous_item(root, path, ino, key_type);
3576
3577 /* if ret == 0 there are items for this type,
3578 * create a range to tell us the last key of this type.
3579 * otherwise, there are no items in this directory after
3580 * *min_offset, and we create a range to indicate that.
3581 */
3582 if (ret == 0) {
3583 struct btrfs_key tmp;
3584 btrfs_item_key_to_cpu(path->nodes[0], &tmp,
3585 path->slots[0]);
3586 if (key_type == tmp.type)
3587 first_offset = max(min_offset, tmp.offset) + 1;
3588 }
3589 goto done;
3590 }
3591
3592 /* go backward to find any previous key */
3593 ret = btrfs_previous_item(root, path, ino, key_type);
3594 if (ret == 0) {
3595 struct btrfs_key tmp;
3596 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3597 if (key_type == tmp.type) {
3598 first_offset = tmp.offset;
3599 ret = overwrite_item(trans, log, dst_path,
3600 path->nodes[0], path->slots[0],
3601 &tmp);
3602 if (ret) {
3603 err = ret;
3604 goto done;
3605 }
3606 }
3607 }
3608 btrfs_release_path(path);
3609
3610 /*
3611 * Find the first key from this transaction again. See the note for
3612 * log_new_dir_dentries, if we're logging a directory recursively we
3613 * won't be holding its i_mutex, which means we can modify the directory
3614 * while we're logging it. If we remove an entry between our first
3615 * search and this search we'll not find the key again and can just
3616 * bail.
3617 */
3618 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
3619 if (ret != 0)
3620 goto done;
3621
3622 /*
3623 * we have a block from this transaction, log every item in it
3624 * from our directory
3625 */
3626 while (1) {
3627 struct btrfs_key tmp;
3628 src = path->nodes[0];
3629 nritems = btrfs_header_nritems(src);
3630 for (i = path->slots[0]; i < nritems; i++) {
3631 struct btrfs_dir_item *di;
3632
3633 btrfs_item_key_to_cpu(src, &min_key, i);
3634
3635 if (min_key.objectid != ino || min_key.type != key_type)
3636 goto done;
3637 ret = overwrite_item(trans, log, dst_path, src, i,
3638 &min_key);
3639 if (ret) {
3640 err = ret;
3641 goto done;
3642 }
3643
3644 /*
3645 * We must make sure that when we log a directory entry,
3646 * the corresponding inode, after log replay, has a
3647 * matching link count. For example:
3648 *
3649 * touch foo
3650 * mkdir mydir
3651 * sync
3652 * ln foo mydir/bar
3653 * xfs_io -c "fsync" mydir
3654 * <crash>
3655 * <mount fs and log replay>
3656 *
3657 * Would result in a fsync log that when replayed, our
3658 * file inode would have a link count of 1, but we get
3659 * two directory entries pointing to the same inode.
3660 * After removing one of the names, it would not be
3661 * possible to remove the other name, which resulted
3662 * always in stale file handle errors, and would not
3663 * be possible to rmdir the parent directory, since
3664 * its i_size could never decrement to the value
3665 * BTRFS_EMPTY_DIR_SIZE, resulting in -ENOTEMPTY errors.
3666 */
3667 di = btrfs_item_ptr(src, i, struct btrfs_dir_item);
3668 btrfs_dir_item_key_to_cpu(src, di, &tmp);
3669 if (ctx &&
3670 (btrfs_dir_transid(src, di) == trans->transid ||
3671 btrfs_dir_type(src, di) == BTRFS_FT_DIR) &&
3672 tmp.type != BTRFS_ROOT_ITEM_KEY)
3673 ctx->log_new_dentries = true;
3674 }
3675 path->slots[0] = nritems;
3676
3677 /*
3678 * look ahead to the next item and see if it is also
3679 * from this directory and from this transaction
3680 */
3681 ret = btrfs_next_leaf(root, path);
3682 if (ret) {
3683 if (ret == 1)
3684 last_offset = (u64)-1;
3685 else
3686 err = ret;
3687 goto done;
3688 }
3689 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
3690 if (tmp.objectid != ino || tmp.type != key_type) {
3691 last_offset = (u64)-1;
3692 goto done;
3693 }
3694 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
3695 ret = overwrite_item(trans, log, dst_path,
3696 path->nodes[0], path->slots[0],
3697 &tmp);
3698 if (ret)
3699 err = ret;
3700 else
3701 last_offset = tmp.offset;
3702 goto done;
3703 }
3704 }
3705done:
3706 btrfs_release_path(path);
3707 btrfs_release_path(dst_path);
3708
3709 if (err == 0) {
3710 *last_offset_ret = last_offset;
3711 /*
3712 * insert the log range keys to indicate where the log
3713 * is valid
3714 */
3715 ret = insert_dir_log_key(trans, log, path, key_type,
3716 ino, first_offset, last_offset);
3717 if (ret)
3718 err = ret;
3719 }
3720 return err;
3721}
3722
3723/*
3724 * logging directories is very similar to logging inodes, We find all the items
3725 * from the current transaction and write them to the log.
3726 *
3727 * The recovery code scans the directory in the subvolume, and if it finds a
3728 * key in the range logged that is not present in the log tree, then it means
3729 * that dir entry was unlinked during the transaction.
3730 *
3731 * In order for that scan to work, we must include one key smaller than
3732 * the smallest logged by this transaction and one key larger than the largest
3733 * key logged by this transaction.
3734 */
3735static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
3736 struct btrfs_root *root, struct btrfs_inode *inode,
3737 struct btrfs_path *path,
3738 struct btrfs_path *dst_path,
3739 struct btrfs_log_ctx *ctx)
3740{
3741 u64 min_key;
3742 u64 max_key;
3743 int ret;
3744 int key_type = BTRFS_DIR_ITEM_KEY;
3745
3746again:
3747 min_key = 0;
3748 max_key = 0;
3749 while (1) {
3750 ret = log_dir_items(trans, root, inode, path, dst_path, key_type,
3751 ctx, min_key, &max_key);
3752 if (ret)
3753 return ret;
3754 if (max_key == (u64)-1)
3755 break;
3756 min_key = max_key + 1;
3757 }
3758
3759 if (key_type == BTRFS_DIR_ITEM_KEY) {
3760 key_type = BTRFS_DIR_INDEX_KEY;
3761 goto again;
3762 }
3763 return 0;
3764}
3765
3766/*
3767 * a helper function to drop items from the log before we relog an
3768 * inode. max_key_type indicates the highest item type to remove.
3769 * This cannot be run for file data extents because it does not
3770 * free the extents they point to.
3771 */
3772static int drop_objectid_items(struct btrfs_trans_handle *trans,
3773 struct btrfs_root *log,
3774 struct btrfs_path *path,
3775 u64 objectid, int max_key_type)
3776{
3777 int ret;
3778 struct btrfs_key key;
3779 struct btrfs_key found_key;
3780 int start_slot;
3781
3782 key.objectid = objectid;
3783 key.type = max_key_type;
3784 key.offset = (u64)-1;
3785
3786 while (1) {
3787 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
3788 BUG_ON(ret == 0); /* Logic error */
3789 if (ret < 0)
3790 break;
3791
3792 if (path->slots[0] == 0)
3793 break;
3794
3795 path->slots[0]--;
3796 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
3797 path->slots[0]);
3798
3799 if (found_key.objectid != objectid)
3800 break;
3801
3802 found_key.offset = 0;
3803 found_key.type = 0;
3804 ret = btrfs_bin_search(path->nodes[0], &found_key, &start_slot);
3805 if (ret < 0)
3806 break;
3807
3808 ret = btrfs_del_items(trans, log, path, start_slot,
3809 path->slots[0] - start_slot + 1);
3810 /*
3811 * If start slot isn't 0 then we don't need to re-search, we've
3812 * found the last guy with the objectid in this tree.
3813 */
3814 if (ret || start_slot != 0)
3815 break;
3816 btrfs_release_path(path);
3817 }
3818 btrfs_release_path(path);
3819 if (ret > 0)
3820 ret = 0;
3821 return ret;
3822}
3823
3824static void fill_inode_item(struct btrfs_trans_handle *trans,
3825 struct extent_buffer *leaf,
3826 struct btrfs_inode_item *item,
3827 struct inode *inode, int log_inode_only,
3828 u64 logged_isize)
3829{
3830 struct btrfs_map_token token;
3831
3832 btrfs_init_map_token(&token, leaf);
3833
3834 if (log_inode_only) {
3835 /* set the generation to zero so the recover code
3836 * can tell the difference between an logging
3837 * just to say 'this inode exists' and a logging
3838 * to say 'update this inode with these values'
3839 */
3840 btrfs_set_token_inode_generation(&token, item, 0);
3841 btrfs_set_token_inode_size(&token, item, logged_isize);
3842 } else {
3843 btrfs_set_token_inode_generation(&token, item,
3844 BTRFS_I(inode)->generation);
3845 btrfs_set_token_inode_size(&token, item, inode->i_size);
3846 }
3847
3848 btrfs_set_token_inode_uid(&token, item, i_uid_read(inode));
3849 btrfs_set_token_inode_gid(&token, item, i_gid_read(inode));
3850 btrfs_set_token_inode_mode(&token, item, inode->i_mode);
3851 btrfs_set_token_inode_nlink(&token, item, inode->i_nlink);
3852
3853 btrfs_set_token_timespec_sec(&token, &item->atime,
3854 inode->i_atime.tv_sec);
3855 btrfs_set_token_timespec_nsec(&token, &item->atime,
3856 inode->i_atime.tv_nsec);
3857
3858 btrfs_set_token_timespec_sec(&token, &item->mtime,
3859 inode->i_mtime.tv_sec);
3860 btrfs_set_token_timespec_nsec(&token, &item->mtime,
3861 inode->i_mtime.tv_nsec);
3862
3863 btrfs_set_token_timespec_sec(&token, &item->ctime,
3864 inode->i_ctime.tv_sec);
3865 btrfs_set_token_timespec_nsec(&token, &item->ctime,
3866 inode->i_ctime.tv_nsec);
3867
3868 btrfs_set_token_inode_nbytes(&token, item, inode_get_bytes(inode));
3869
3870 btrfs_set_token_inode_sequence(&token, item, inode_peek_iversion(inode));
3871 btrfs_set_token_inode_transid(&token, item, trans->transid);
3872 btrfs_set_token_inode_rdev(&token, item, inode->i_rdev);
3873 btrfs_set_token_inode_flags(&token, item, BTRFS_I(inode)->flags);
3874 btrfs_set_token_inode_block_group(&token, item, 0);
3875}
3876
3877static int log_inode_item(struct btrfs_trans_handle *trans,
3878 struct btrfs_root *log, struct btrfs_path *path,
3879 struct btrfs_inode *inode)
3880{
3881 struct btrfs_inode_item *inode_item;
3882 int ret;
3883
3884 ret = btrfs_insert_empty_item(trans, log, path,
3885 &inode->location, sizeof(*inode_item));
3886 if (ret && ret != -EEXIST)
3887 return ret;
3888 inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3889 struct btrfs_inode_item);
3890 fill_inode_item(trans, path->nodes[0], inode_item, &inode->vfs_inode,
3891 0, 0);
3892 btrfs_release_path(path);
3893 return 0;
3894}
3895
3896static int log_csums(struct btrfs_trans_handle *trans,
3897 struct btrfs_inode *inode,
3898 struct btrfs_root *log_root,
3899 struct btrfs_ordered_sum *sums)
3900{
3901 const u64 lock_end = sums->bytenr + sums->len - 1;
3902 struct extent_state *cached_state = NULL;
3903 int ret;
3904
3905 /*
3906 * If this inode was not used for reflink operations in the current
3907 * transaction with new extents, then do the fast path, no need to
3908 * worry about logging checksum items with overlapping ranges.
3909 */
3910 if (inode->last_reflink_trans < trans->transid)
3911 return btrfs_csum_file_blocks(trans, log_root, sums);
3912
3913 /*
3914 * Serialize logging for checksums. This is to avoid racing with the
3915 * same checksum being logged by another task that is logging another
3916 * file which happens to refer to the same extent as well. Such races
3917 * can leave checksum items in the log with overlapping ranges.
3918 */
3919 ret = lock_extent_bits(&log_root->log_csum_range, sums->bytenr,
3920 lock_end, &cached_state);
3921 if (ret)
3922 return ret;
3923 /*
3924 * Due to extent cloning, we might have logged a csum item that covers a
3925 * subrange of a cloned extent, and later we can end up logging a csum
3926 * item for a larger subrange of the same extent or the entire range.
3927 * This would leave csum items in the log tree that cover the same range
3928 * and break the searches for checksums in the log tree, resulting in
3929 * some checksums missing in the fs/subvolume tree. So just delete (or
3930 * trim and adjust) any existing csum items in the log for this range.
3931 */
3932 ret = btrfs_del_csums(trans, log_root, sums->bytenr, sums->len);
3933 if (!ret)
3934 ret = btrfs_csum_file_blocks(trans, log_root, sums);
3935
3936 unlock_extent_cached(&log_root->log_csum_range, sums->bytenr, lock_end,
3937 &cached_state);
3938
3939 return ret;
3940}
3941
3942static noinline int copy_items(struct btrfs_trans_handle *trans,
3943 struct btrfs_inode *inode,
3944 struct btrfs_path *dst_path,
3945 struct btrfs_path *src_path,
3946 int start_slot, int nr, int inode_only,
3947 u64 logged_isize)
3948{
3949 struct btrfs_fs_info *fs_info = trans->fs_info;
3950 unsigned long src_offset;
3951 unsigned long dst_offset;
3952 struct btrfs_root *log = inode->root->log_root;
3953 struct btrfs_file_extent_item *extent;
3954 struct btrfs_inode_item *inode_item;
3955 struct extent_buffer *src = src_path->nodes[0];
3956 int ret;
3957 struct btrfs_key *ins_keys;
3958 u32 *ins_sizes;
3959 char *ins_data;
3960 int i;
3961 struct list_head ordered_sums;
3962 int skip_csum = inode->flags & BTRFS_INODE_NODATASUM;
3963
3964 INIT_LIST_HEAD(&ordered_sums);
3965
3966 ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
3967 nr * sizeof(u32), GFP_NOFS);
3968 if (!ins_data)
3969 return -ENOMEM;
3970
3971 ins_sizes = (u32 *)ins_data;
3972 ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
3973
3974 for (i = 0; i < nr; i++) {
3975 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
3976 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
3977 }
3978 ret = btrfs_insert_empty_items(trans, log, dst_path,
3979 ins_keys, ins_sizes, nr);
3980 if (ret) {
3981 kfree(ins_data);
3982 return ret;
3983 }
3984
3985 for (i = 0; i < nr; i++, dst_path->slots[0]++) {
3986 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
3987 dst_path->slots[0]);
3988
3989 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
3990
3991 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
3992 inode_item = btrfs_item_ptr(dst_path->nodes[0],
3993 dst_path->slots[0],
3994 struct btrfs_inode_item);
3995 fill_inode_item(trans, dst_path->nodes[0], inode_item,
3996 &inode->vfs_inode,
3997 inode_only == LOG_INODE_EXISTS,
3998 logged_isize);
3999 } else {
4000 copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
4001 src_offset, ins_sizes[i]);
4002 }
4003
4004 /* take a reference on file data extents so that truncates
4005 * or deletes of this inode don't have to relog the inode
4006 * again
4007 */
4008 if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY &&
4009 !skip_csum) {
4010 int found_type;
4011 extent = btrfs_item_ptr(src, start_slot + i,
4012 struct btrfs_file_extent_item);
4013
4014 if (btrfs_file_extent_generation(src, extent) < trans->transid)
4015 continue;
4016
4017 found_type = btrfs_file_extent_type(src, extent);
4018 if (found_type == BTRFS_FILE_EXTENT_REG) {
4019 u64 ds, dl, cs, cl;
4020 ds = btrfs_file_extent_disk_bytenr(src,
4021 extent);
4022 /* ds == 0 is a hole */
4023 if (ds == 0)
4024 continue;
4025
4026 dl = btrfs_file_extent_disk_num_bytes(src,
4027 extent);
4028 cs = btrfs_file_extent_offset(src, extent);
4029 cl = btrfs_file_extent_num_bytes(src,
4030 extent);
4031 if (btrfs_file_extent_compression(src,
4032 extent)) {
4033 cs = 0;
4034 cl = dl;
4035 }
4036
4037 ret = btrfs_lookup_csums_range(
4038 fs_info->csum_root,
4039 ds + cs, ds + cs + cl - 1,
4040 &ordered_sums, 0);
4041 if (ret)
4042 break;
4043 }
4044 }
4045 }
4046
4047 btrfs_mark_buffer_dirty(dst_path->nodes[0]);
4048 btrfs_release_path(dst_path);
4049 kfree(ins_data);
4050
4051 /*
4052 * we have to do this after the loop above to avoid changing the
4053 * log tree while trying to change the log tree.
4054 */
4055 while (!list_empty(&ordered_sums)) {
4056 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
4057 struct btrfs_ordered_sum,
4058 list);
4059 if (!ret)
4060 ret = log_csums(trans, inode, log, sums);
4061 list_del(&sums->list);
4062 kfree(sums);
4063 }
4064
4065 return ret;
4066}
4067
4068static int extent_cmp(void *priv, struct list_head *a, struct list_head *b)
4069{
4070 struct extent_map *em1, *em2;
4071
4072 em1 = list_entry(a, struct extent_map, list);
4073 em2 = list_entry(b, struct extent_map, list);
4074
4075 if (em1->start < em2->start)
4076 return -1;
4077 else if (em1->start > em2->start)
4078 return 1;
4079 return 0;
4080}
4081
4082static int log_extent_csums(struct btrfs_trans_handle *trans,
4083 struct btrfs_inode *inode,
4084 struct btrfs_root *log_root,
4085 const struct extent_map *em)
4086{
4087 u64 csum_offset;
4088 u64 csum_len;
4089 LIST_HEAD(ordered_sums);
4090 int ret = 0;
4091
4092 if (inode->flags & BTRFS_INODE_NODATASUM ||
4093 test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
4094 em->block_start == EXTENT_MAP_HOLE)
4095 return 0;
4096
4097 /* If we're compressed we have to save the entire range of csums. */
4098 if (em->compress_type) {
4099 csum_offset = 0;
4100 csum_len = max(em->block_len, em->orig_block_len);
4101 } else {
4102 csum_offset = em->mod_start - em->start;
4103 csum_len = em->mod_len;
4104 }
4105
4106 /* block start is already adjusted for the file extent offset. */
4107 ret = btrfs_lookup_csums_range(trans->fs_info->csum_root,
4108 em->block_start + csum_offset,
4109 em->block_start + csum_offset +
4110 csum_len - 1, &ordered_sums, 0);
4111 if (ret)
4112 return ret;
4113
4114 while (!list_empty(&ordered_sums)) {
4115 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
4116 struct btrfs_ordered_sum,
4117 list);
4118 if (!ret)
4119 ret = log_csums(trans, inode, log_root, sums);
4120 list_del(&sums->list);
4121 kfree(sums);
4122 }
4123
4124 return ret;
4125}
4126
4127static int log_one_extent(struct btrfs_trans_handle *trans,
4128 struct btrfs_inode *inode, struct btrfs_root *root,
4129 const struct extent_map *em,
4130 struct btrfs_path *path,
4131 struct btrfs_log_ctx *ctx)
4132{
4133 struct btrfs_root *log = root->log_root;
4134 struct btrfs_file_extent_item *fi;
4135 struct extent_buffer *leaf;
4136 struct btrfs_map_token token;
4137 struct btrfs_key key;
4138 u64 extent_offset = em->start - em->orig_start;
4139 u64 block_len;
4140 int ret;
4141 int extent_inserted = 0;
4142
4143 ret = log_extent_csums(trans, inode, log, em);
4144 if (ret)
4145 return ret;
4146
4147 ret = __btrfs_drop_extents(trans, log, inode, path, em->start,
4148 em->start + em->len, NULL, 0, 1,
4149 sizeof(*fi), &extent_inserted);
4150 if (ret)
4151 return ret;
4152
4153 if (!extent_inserted) {
4154 key.objectid = btrfs_ino(inode);
4155 key.type = BTRFS_EXTENT_DATA_KEY;
4156 key.offset = em->start;
4157
4158 ret = btrfs_insert_empty_item(trans, log, path, &key,
4159 sizeof(*fi));
4160 if (ret)
4161 return ret;
4162 }
4163 leaf = path->nodes[0];
4164 btrfs_init_map_token(&token, leaf);
4165 fi = btrfs_item_ptr(leaf, path->slots[0],
4166 struct btrfs_file_extent_item);
4167
4168 btrfs_set_token_file_extent_generation(&token, fi, trans->transid);
4169 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4170 btrfs_set_token_file_extent_type(&token, fi,
4171 BTRFS_FILE_EXTENT_PREALLOC);
4172 else
4173 btrfs_set_token_file_extent_type(&token, fi,
4174 BTRFS_FILE_EXTENT_REG);
4175
4176 block_len = max(em->block_len, em->orig_block_len);
4177 if (em->compress_type != BTRFS_COMPRESS_NONE) {
4178 btrfs_set_token_file_extent_disk_bytenr(&token, fi,
4179 em->block_start);
4180 btrfs_set_token_file_extent_disk_num_bytes(&token, fi, block_len);
4181 } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
4182 btrfs_set_token_file_extent_disk_bytenr(&token, fi,
4183 em->block_start -
4184 extent_offset);
4185 btrfs_set_token_file_extent_disk_num_bytes(&token, fi, block_len);
4186 } else {
4187 btrfs_set_token_file_extent_disk_bytenr(&token, fi, 0);
4188 btrfs_set_token_file_extent_disk_num_bytes(&token, fi, 0);
4189 }
4190
4191 btrfs_set_token_file_extent_offset(&token, fi, extent_offset);
4192 btrfs_set_token_file_extent_num_bytes(&token, fi, em->len);
4193 btrfs_set_token_file_extent_ram_bytes(&token, fi, em->ram_bytes);
4194 btrfs_set_token_file_extent_compression(&token, fi, em->compress_type);
4195 btrfs_set_token_file_extent_encryption(&token, fi, 0);
4196 btrfs_set_token_file_extent_other_encoding(&token, fi, 0);
4197 btrfs_mark_buffer_dirty(leaf);
4198
4199 btrfs_release_path(path);
4200
4201 return ret;
4202}
4203
4204/*
4205 * Log all prealloc extents beyond the inode's i_size to make sure we do not
4206 * lose them after doing a fast fsync and replaying the log. We scan the
4207 * subvolume's root instead of iterating the inode's extent map tree because
4208 * otherwise we can log incorrect extent items based on extent map conversion.
4209 * That can happen due to the fact that extent maps are merged when they
4210 * are not in the extent map tree's list of modified extents.
4211 */
4212static int btrfs_log_prealloc_extents(struct btrfs_trans_handle *trans,
4213 struct btrfs_inode *inode,
4214 struct btrfs_path *path)
4215{
4216 struct btrfs_root *root = inode->root;
4217 struct btrfs_key key;
4218 const u64 i_size = i_size_read(&inode->vfs_inode);
4219 const u64 ino = btrfs_ino(inode);
4220 struct btrfs_path *dst_path = NULL;
4221 bool dropped_extents = false;
4222 u64 truncate_offset = i_size;
4223 struct extent_buffer *leaf;
4224 int slot;
4225 int ins_nr = 0;
4226 int start_slot;
4227 int ret;
4228
4229 if (!(inode->flags & BTRFS_INODE_PREALLOC))
4230 return 0;
4231
4232 key.objectid = ino;
4233 key.type = BTRFS_EXTENT_DATA_KEY;
4234 key.offset = i_size;
4235 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4236 if (ret < 0)
4237 goto out;
4238
4239 /*
4240 * We must check if there is a prealloc extent that starts before the
4241 * i_size and crosses the i_size boundary. This is to ensure later we
4242 * truncate down to the end of that extent and not to the i_size, as
4243 * otherwise we end up losing part of the prealloc extent after a log
4244 * replay and with an implicit hole if there is another prealloc extent
4245 * that starts at an offset beyond i_size.
4246 */
4247 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
4248 if (ret < 0)
4249 goto out;
4250
4251 if (ret == 0) {
4252 struct btrfs_file_extent_item *ei;
4253
4254 leaf = path->nodes[0];
4255 slot = path->slots[0];
4256 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4257
4258 if (btrfs_file_extent_type(leaf, ei) ==
4259 BTRFS_FILE_EXTENT_PREALLOC) {
4260 u64 extent_end;
4261
4262 btrfs_item_key_to_cpu(leaf, &key, slot);
4263 extent_end = key.offset +
4264 btrfs_file_extent_num_bytes(leaf, ei);
4265
4266 if (extent_end > i_size)
4267 truncate_offset = extent_end;
4268 }
4269 } else {
4270 ret = 0;
4271 }
4272
4273 while (true) {
4274 leaf = path->nodes[0];
4275 slot = path->slots[0];
4276
4277 if (slot >= btrfs_header_nritems(leaf)) {
4278 if (ins_nr > 0) {
4279 ret = copy_items(trans, inode, dst_path, path,
4280 start_slot, ins_nr, 1, 0);
4281 if (ret < 0)
4282 goto out;
4283 ins_nr = 0;
4284 }
4285 ret = btrfs_next_leaf(root, path);
4286 if (ret < 0)
4287 goto out;
4288 if (ret > 0) {
4289 ret = 0;
4290 break;
4291 }
4292 continue;
4293 }
4294
4295 btrfs_item_key_to_cpu(leaf, &key, slot);
4296 if (key.objectid > ino)
4297 break;
4298 if (WARN_ON_ONCE(key.objectid < ino) ||
4299 key.type < BTRFS_EXTENT_DATA_KEY ||
4300 key.offset < i_size) {
4301 path->slots[0]++;
4302 continue;
4303 }
4304 if (!dropped_extents) {
4305 /*
4306 * Avoid logging extent items logged in past fsync calls
4307 * and leading to duplicate keys in the log tree.
4308 */
4309 do {
4310 ret = btrfs_truncate_inode_items(trans,
4311 root->log_root,
4312 &inode->vfs_inode,
4313 truncate_offset,
4314 BTRFS_EXTENT_DATA_KEY);
4315 } while (ret == -EAGAIN);
4316 if (ret)
4317 goto out;
4318 dropped_extents = true;
4319 }
4320 if (ins_nr == 0)
4321 start_slot = slot;
4322 ins_nr++;
4323 path->slots[0]++;
4324 if (!dst_path) {
4325 dst_path = btrfs_alloc_path();
4326 if (!dst_path) {
4327 ret = -ENOMEM;
4328 goto out;
4329 }
4330 }
4331 }
4332 if (ins_nr > 0)
4333 ret = copy_items(trans, inode, dst_path, path,
4334 start_slot, ins_nr, 1, 0);
4335out:
4336 btrfs_release_path(path);
4337 btrfs_free_path(dst_path);
4338 return ret;
4339}
4340
4341static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
4342 struct btrfs_root *root,
4343 struct btrfs_inode *inode,
4344 struct btrfs_path *path,
4345 struct btrfs_log_ctx *ctx,
4346 const u64 start,
4347 const u64 end)
4348{
4349 struct extent_map *em, *n;
4350 struct list_head extents;
4351 struct extent_map_tree *tree = &inode->extent_tree;
4352 u64 test_gen;
4353 int ret = 0;
4354 int num = 0;
4355
4356 INIT_LIST_HEAD(&extents);
4357
4358 write_lock(&tree->lock);
4359 test_gen = root->fs_info->last_trans_committed;
4360
4361 list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
4362 /*
4363 * Skip extents outside our logging range. It's important to do
4364 * it for correctness because if we don't ignore them, we may
4365 * log them before their ordered extent completes, and therefore
4366 * we could log them without logging their respective checksums
4367 * (the checksum items are added to the csum tree at the very
4368 * end of btrfs_finish_ordered_io()). Also leave such extents
4369 * outside of our range in the list, since we may have another
4370 * ranged fsync in the near future that needs them. If an extent
4371 * outside our range corresponds to a hole, log it to avoid
4372 * leaving gaps between extents (fsck will complain when we are
4373 * not using the NO_HOLES feature).
4374 */
4375 if ((em->start > end || em->start + em->len <= start) &&
4376 em->block_start != EXTENT_MAP_HOLE)
4377 continue;
4378
4379 list_del_init(&em->list);
4380 /*
4381 * Just an arbitrary number, this can be really CPU intensive
4382 * once we start getting a lot of extents, and really once we
4383 * have a bunch of extents we just want to commit since it will
4384 * be faster.
4385 */
4386 if (++num > 32768) {
4387 list_del_init(&tree->modified_extents);
4388 ret = -EFBIG;
4389 goto process;
4390 }
4391
4392 if (em->generation <= test_gen)
4393 continue;
4394
4395 /* We log prealloc extents beyond eof later. */
4396 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) &&
4397 em->start >= i_size_read(&inode->vfs_inode))
4398 continue;
4399
4400 /* Need a ref to keep it from getting evicted from cache */
4401 refcount_inc(&em->refs);
4402 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
4403 list_add_tail(&em->list, &extents);
4404 num++;
4405 }
4406
4407 list_sort(NULL, &extents, extent_cmp);
4408process:
4409 while (!list_empty(&extents)) {
4410 em = list_entry(extents.next, struct extent_map, list);
4411
4412 list_del_init(&em->list);
4413
4414 /*
4415 * If we had an error we just need to delete everybody from our
4416 * private list.
4417 */
4418 if (ret) {
4419 clear_em_logging(tree, em);
4420 free_extent_map(em);
4421 continue;
4422 }
4423
4424 write_unlock(&tree->lock);
4425
4426 ret = log_one_extent(trans, inode, root, em, path, ctx);
4427 write_lock(&tree->lock);
4428 clear_em_logging(tree, em);
4429 free_extent_map(em);
4430 }
4431 WARN_ON(!list_empty(&extents));
4432 write_unlock(&tree->lock);
4433
4434 btrfs_release_path(path);
4435 if (!ret)
4436 ret = btrfs_log_prealloc_extents(trans, inode, path);
4437
4438 return ret;
4439}
4440
4441static int logged_inode_size(struct btrfs_root *log, struct btrfs_inode *inode,
4442 struct btrfs_path *path, u64 *size_ret)
4443{
4444 struct btrfs_key key;
4445 int ret;
4446
4447 key.objectid = btrfs_ino(inode);
4448 key.type = BTRFS_INODE_ITEM_KEY;
4449 key.offset = 0;
4450
4451 ret = btrfs_search_slot(NULL, log, &key, path, 0, 0);
4452 if (ret < 0) {
4453 return ret;
4454 } else if (ret > 0) {
4455 *size_ret = 0;
4456 } else {
4457 struct btrfs_inode_item *item;
4458
4459 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4460 struct btrfs_inode_item);
4461 *size_ret = btrfs_inode_size(path->nodes[0], item);
4462 /*
4463 * If the in-memory inode's i_size is smaller then the inode
4464 * size stored in the btree, return the inode's i_size, so
4465 * that we get a correct inode size after replaying the log
4466 * when before a power failure we had a shrinking truncate
4467 * followed by addition of a new name (rename / new hard link).
4468 * Otherwise return the inode size from the btree, to avoid
4469 * data loss when replaying a log due to previously doing a
4470 * write that expands the inode's size and logging a new name
4471 * immediately after.
4472 */
4473 if (*size_ret > inode->vfs_inode.i_size)
4474 *size_ret = inode->vfs_inode.i_size;
4475 }
4476
4477 btrfs_release_path(path);
4478 return 0;
4479}
4480
4481/*
4482 * At the moment we always log all xattrs. This is to figure out at log replay
4483 * time which xattrs must have their deletion replayed. If a xattr is missing
4484 * in the log tree and exists in the fs/subvol tree, we delete it. This is
4485 * because if a xattr is deleted, the inode is fsynced and a power failure
4486 * happens, causing the log to be replayed the next time the fs is mounted,
4487 * we want the xattr to not exist anymore (same behaviour as other filesystems
4488 * with a journal, ext3/4, xfs, f2fs, etc).
4489 */
4490static int btrfs_log_all_xattrs(struct btrfs_trans_handle *trans,
4491 struct btrfs_root *root,
4492 struct btrfs_inode *inode,
4493 struct btrfs_path *path,
4494 struct btrfs_path *dst_path)
4495{
4496 int ret;
4497 struct btrfs_key key;
4498 const u64 ino = btrfs_ino(inode);
4499 int ins_nr = 0;
4500 int start_slot = 0;
4501
4502 key.objectid = ino;
4503 key.type = BTRFS_XATTR_ITEM_KEY;
4504 key.offset = 0;
4505
4506 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4507 if (ret < 0)
4508 return ret;
4509
4510 while (true) {
4511 int slot = path->slots[0];
4512 struct extent_buffer *leaf = path->nodes[0];
4513 int nritems = btrfs_header_nritems(leaf);
4514
4515 if (slot >= nritems) {
4516 if (ins_nr > 0) {
4517 ret = copy_items(trans, inode, dst_path, path,
4518 start_slot, ins_nr, 1, 0);
4519 if (ret < 0)
4520 return ret;
4521 ins_nr = 0;
4522 }
4523 ret = btrfs_next_leaf(root, path);
4524 if (ret < 0)
4525 return ret;
4526 else if (ret > 0)
4527 break;
4528 continue;
4529 }
4530
4531 btrfs_item_key_to_cpu(leaf, &key, slot);
4532 if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY)
4533 break;
4534
4535 if (ins_nr == 0)
4536 start_slot = slot;
4537 ins_nr++;
4538 path->slots[0]++;
4539 cond_resched();
4540 }
4541 if (ins_nr > 0) {
4542 ret = copy_items(trans, inode, dst_path, path,
4543 start_slot, ins_nr, 1, 0);
4544 if (ret < 0)
4545 return ret;
4546 }
4547
4548 return 0;
4549}
4550
4551/*
4552 * When using the NO_HOLES feature if we punched a hole that causes the
4553 * deletion of entire leafs or all the extent items of the first leaf (the one
4554 * that contains the inode item and references) we may end up not processing
4555 * any extents, because there are no leafs with a generation matching the
4556 * current transaction that have extent items for our inode. So we need to find
4557 * if any holes exist and then log them. We also need to log holes after any
4558 * truncate operation that changes the inode's size.
4559 */
4560static int btrfs_log_holes(struct btrfs_trans_handle *trans,
4561 struct btrfs_root *root,
4562 struct btrfs_inode *inode,
4563 struct btrfs_path *path)
4564{
4565 struct btrfs_fs_info *fs_info = root->fs_info;
4566 struct btrfs_key key;
4567 const u64 ino = btrfs_ino(inode);
4568 const u64 i_size = i_size_read(&inode->vfs_inode);
4569 u64 prev_extent_end = 0;
4570 int ret;
4571
4572 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || i_size == 0)
4573 return 0;
4574
4575 key.objectid = ino;
4576 key.type = BTRFS_EXTENT_DATA_KEY;
4577 key.offset = 0;
4578
4579 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4580 if (ret < 0)
4581 return ret;
4582
4583 while (true) {
4584 struct extent_buffer *leaf = path->nodes[0];
4585
4586 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
4587 ret = btrfs_next_leaf(root, path);
4588 if (ret < 0)
4589 return ret;
4590 if (ret > 0) {
4591 ret = 0;
4592 break;
4593 }
4594 leaf = path->nodes[0];
4595 }
4596
4597 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4598 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
4599 break;
4600
4601 /* We have a hole, log it. */
4602 if (prev_extent_end < key.offset) {
4603 const u64 hole_len = key.offset - prev_extent_end;
4604
4605 /*
4606 * Release the path to avoid deadlocks with other code
4607 * paths that search the root while holding locks on
4608 * leafs from the log root.
4609 */
4610 btrfs_release_path(path);
4611 ret = btrfs_insert_file_extent(trans, root->log_root,
4612 ino, prev_extent_end, 0,
4613 0, hole_len, 0, hole_len,
4614 0, 0, 0);
4615 if (ret < 0)
4616 return ret;
4617
4618 /*
4619 * Search for the same key again in the root. Since it's
4620 * an extent item and we are holding the inode lock, the
4621 * key must still exist. If it doesn't just emit warning
4622 * and return an error to fall back to a transaction
4623 * commit.
4624 */
4625 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4626 if (ret < 0)
4627 return ret;
4628 if (WARN_ON(ret > 0))
4629 return -ENOENT;
4630 leaf = path->nodes[0];
4631 }
4632
4633 prev_extent_end = btrfs_file_extent_end(path);
4634 path->slots[0]++;
4635 cond_resched();
4636 }
4637
4638 if (prev_extent_end < i_size) {
4639 u64 hole_len;
4640
4641 btrfs_release_path(path);
4642 hole_len = ALIGN(i_size - prev_extent_end, fs_info->sectorsize);
4643 ret = btrfs_insert_file_extent(trans, root->log_root,
4644 ino, prev_extent_end, 0, 0,
4645 hole_len, 0, hole_len,
4646 0, 0, 0);
4647 if (ret < 0)
4648 return ret;
4649 }
4650
4651 return 0;
4652}
4653
4654/*
4655 * When we are logging a new inode X, check if it doesn't have a reference that
4656 * matches the reference from some other inode Y created in a past transaction
4657 * and that was renamed in the current transaction. If we don't do this, then at
4658 * log replay time we can lose inode Y (and all its files if it's a directory):
4659 *
4660 * mkdir /mnt/x
4661 * echo "hello world" > /mnt/x/foobar
4662 * sync
4663 * mv /mnt/x /mnt/y
4664 * mkdir /mnt/x # or touch /mnt/x
4665 * xfs_io -c fsync /mnt/x
4666 * <power fail>
4667 * mount fs, trigger log replay
4668 *
4669 * After the log replay procedure, we would lose the first directory and all its
4670 * files (file foobar).
4671 * For the case where inode Y is not a directory we simply end up losing it:
4672 *
4673 * echo "123" > /mnt/foo
4674 * sync
4675 * mv /mnt/foo /mnt/bar
4676 * echo "abc" > /mnt/foo
4677 * xfs_io -c fsync /mnt/foo
4678 * <power fail>
4679 *
4680 * We also need this for cases where a snapshot entry is replaced by some other
4681 * entry (file or directory) otherwise we end up with an unreplayable log due to
4682 * attempts to delete the snapshot entry (entry of type BTRFS_ROOT_ITEM_KEY) as
4683 * if it were a regular entry:
4684 *
4685 * mkdir /mnt/x
4686 * btrfs subvolume snapshot /mnt /mnt/x/snap
4687 * btrfs subvolume delete /mnt/x/snap
4688 * rmdir /mnt/x
4689 * mkdir /mnt/x
4690 * fsync /mnt/x or fsync some new file inside it
4691 * <power fail>
4692 *
4693 * The snapshot delete, rmdir of x, mkdir of a new x and the fsync all happen in
4694 * the same transaction.
4695 */
4696static int btrfs_check_ref_name_override(struct extent_buffer *eb,
4697 const int slot,
4698 const struct btrfs_key *key,
4699 struct btrfs_inode *inode,
4700 u64 *other_ino, u64 *other_parent)
4701{
4702 int ret;
4703 struct btrfs_path *search_path;
4704 char *name = NULL;
4705 u32 name_len = 0;
4706 u32 item_size = btrfs_item_size_nr(eb, slot);
4707 u32 cur_offset = 0;
4708 unsigned long ptr = btrfs_item_ptr_offset(eb, slot);
4709
4710 search_path = btrfs_alloc_path();
4711 if (!search_path)
4712 return -ENOMEM;
4713 search_path->search_commit_root = 1;
4714 search_path->skip_locking = 1;
4715
4716 while (cur_offset < item_size) {
4717 u64 parent;
4718 u32 this_name_len;
4719 u32 this_len;
4720 unsigned long name_ptr;
4721 struct btrfs_dir_item *di;
4722
4723 if (key->type == BTRFS_INODE_REF_KEY) {
4724 struct btrfs_inode_ref *iref;
4725
4726 iref = (struct btrfs_inode_ref *)(ptr + cur_offset);
4727 parent = key->offset;
4728 this_name_len = btrfs_inode_ref_name_len(eb, iref);
4729 name_ptr = (unsigned long)(iref + 1);
4730 this_len = sizeof(*iref) + this_name_len;
4731 } else {
4732 struct btrfs_inode_extref *extref;
4733
4734 extref = (struct btrfs_inode_extref *)(ptr +
4735 cur_offset);
4736 parent = btrfs_inode_extref_parent(eb, extref);
4737 this_name_len = btrfs_inode_extref_name_len(eb, extref);
4738 name_ptr = (unsigned long)&extref->name;
4739 this_len = sizeof(*extref) + this_name_len;
4740 }
4741
4742 if (this_name_len > name_len) {
4743 char *new_name;
4744
4745 new_name = krealloc(name, this_name_len, GFP_NOFS);
4746 if (!new_name) {
4747 ret = -ENOMEM;
4748 goto out;
4749 }
4750 name_len = this_name_len;
4751 name = new_name;
4752 }
4753
4754 read_extent_buffer(eb, name, name_ptr, this_name_len);
4755 di = btrfs_lookup_dir_item(NULL, inode->root, search_path,
4756 parent, name, this_name_len, 0);
4757 if (di && !IS_ERR(di)) {
4758 struct btrfs_key di_key;
4759
4760 btrfs_dir_item_key_to_cpu(search_path->nodes[0],
4761 di, &di_key);
4762 if (di_key.type == BTRFS_INODE_ITEM_KEY) {
4763 if (di_key.objectid != key->objectid) {
4764 ret = 1;
4765 *other_ino = di_key.objectid;
4766 *other_parent = parent;
4767 } else {
4768 ret = 0;
4769 }
4770 } else {
4771 ret = -EAGAIN;
4772 }
4773 goto out;
4774 } else if (IS_ERR(di)) {
4775 ret = PTR_ERR(di);
4776 goto out;
4777 }
4778 btrfs_release_path(search_path);
4779
4780 cur_offset += this_len;
4781 }
4782 ret = 0;
4783out:
4784 btrfs_free_path(search_path);
4785 kfree(name);
4786 return ret;
4787}
4788
4789struct btrfs_ino_list {
4790 u64 ino;
4791 u64 parent;
4792 struct list_head list;
4793};
4794
4795static int log_conflicting_inodes(struct btrfs_trans_handle *trans,
4796 struct btrfs_root *root,
4797 struct btrfs_path *path,
4798 struct btrfs_log_ctx *ctx,
4799 u64 ino, u64 parent)
4800{
4801 struct btrfs_ino_list *ino_elem;
4802 LIST_HEAD(inode_list);
4803 int ret = 0;
4804
4805 ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
4806 if (!ino_elem)
4807 return -ENOMEM;
4808 ino_elem->ino = ino;
4809 ino_elem->parent = parent;
4810 list_add_tail(&ino_elem->list, &inode_list);
4811
4812 while (!list_empty(&inode_list)) {
4813 struct btrfs_fs_info *fs_info = root->fs_info;
4814 struct btrfs_key key;
4815 struct inode *inode;
4816
4817 ino_elem = list_first_entry(&inode_list, struct btrfs_ino_list,
4818 list);
4819 ino = ino_elem->ino;
4820 parent = ino_elem->parent;
4821 list_del(&ino_elem->list);
4822 kfree(ino_elem);
4823 if (ret)
4824 continue;
4825
4826 btrfs_release_path(path);
4827
4828 inode = btrfs_iget(fs_info->sb, ino, root);
4829 /*
4830 * If the other inode that had a conflicting dir entry was
4831 * deleted in the current transaction, we need to log its parent
4832 * directory.
4833 */
4834 if (IS_ERR(inode)) {
4835 ret = PTR_ERR(inode);
4836 if (ret == -ENOENT) {
4837 inode = btrfs_iget(fs_info->sb, parent, root);
4838 if (IS_ERR(inode)) {
4839 ret = PTR_ERR(inode);
4840 } else {
4841 ret = btrfs_log_inode(trans, root,
4842 BTRFS_I(inode),
4843 LOG_OTHER_INODE_ALL,
4844 0, LLONG_MAX, ctx);
4845 btrfs_add_delayed_iput(inode);
4846 }
4847 }
4848 continue;
4849 }
4850 /*
4851 * If the inode was already logged skip it - otherwise we can
4852 * hit an infinite loop. Example:
4853 *
4854 * From the commit root (previous transaction) we have the
4855 * following inodes:
4856 *
4857 * inode 257 a directory
4858 * inode 258 with references "zz" and "zz_link" on inode 257
4859 * inode 259 with reference "a" on inode 257
4860 *
4861 * And in the current (uncommitted) transaction we have:
4862 *
4863 * inode 257 a directory, unchanged
4864 * inode 258 with references "a" and "a2" on inode 257
4865 * inode 259 with reference "zz_link" on inode 257
4866 * inode 261 with reference "zz" on inode 257
4867 *
4868 * When logging inode 261 the following infinite loop could
4869 * happen if we don't skip already logged inodes:
4870 *
4871 * - we detect inode 258 as a conflicting inode, with inode 261
4872 * on reference "zz", and log it;
4873 *
4874 * - we detect inode 259 as a conflicting inode, with inode 258
4875 * on reference "a", and log it;
4876 *
4877 * - we detect inode 258 as a conflicting inode, with inode 259
4878 * on reference "zz_link", and log it - again! After this we
4879 * repeat the above steps forever.
4880 */
4881 spin_lock(&BTRFS_I(inode)->lock);
4882 /*
4883 * Check the inode's logged_trans only instead of
4884 * btrfs_inode_in_log(). This is because the last_log_commit of
4885 * the inode is not updated when we only log that it exists and
4886 * and it has the full sync bit set (see btrfs_log_inode()).
4887 */
4888 if (BTRFS_I(inode)->logged_trans == trans->transid) {
4889 spin_unlock(&BTRFS_I(inode)->lock);
4890 btrfs_add_delayed_iput(inode);
4891 continue;
4892 }
4893 spin_unlock(&BTRFS_I(inode)->lock);
4894 /*
4895 * We are safe logging the other inode without acquiring its
4896 * lock as long as we log with the LOG_INODE_EXISTS mode. We
4897 * are safe against concurrent renames of the other inode as
4898 * well because during a rename we pin the log and update the
4899 * log with the new name before we unpin it.
4900 */
4901 ret = btrfs_log_inode(trans, root, BTRFS_I(inode),
4902 LOG_OTHER_INODE, 0, LLONG_MAX, ctx);
4903 if (ret) {
4904 btrfs_add_delayed_iput(inode);
4905 continue;
4906 }
4907
4908 key.objectid = ino;
4909 key.type = BTRFS_INODE_REF_KEY;
4910 key.offset = 0;
4911 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4912 if (ret < 0) {
4913 btrfs_add_delayed_iput(inode);
4914 continue;
4915 }
4916
4917 while (true) {
4918 struct extent_buffer *leaf = path->nodes[0];
4919 int slot = path->slots[0];
4920 u64 other_ino = 0;
4921 u64 other_parent = 0;
4922
4923 if (slot >= btrfs_header_nritems(leaf)) {
4924 ret = btrfs_next_leaf(root, path);
4925 if (ret < 0) {
4926 break;
4927 } else if (ret > 0) {
4928 ret = 0;
4929 break;
4930 }
4931 continue;
4932 }
4933
4934 btrfs_item_key_to_cpu(leaf, &key, slot);
4935 if (key.objectid != ino ||
4936 (key.type != BTRFS_INODE_REF_KEY &&
4937 key.type != BTRFS_INODE_EXTREF_KEY)) {
4938 ret = 0;
4939 break;
4940 }
4941
4942 ret = btrfs_check_ref_name_override(leaf, slot, &key,
4943 BTRFS_I(inode), &other_ino,
4944 &other_parent);
4945 if (ret < 0)
4946 break;
4947 if (ret > 0) {
4948 ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
4949 if (!ino_elem) {
4950 ret = -ENOMEM;
4951 break;
4952 }
4953 ino_elem->ino = other_ino;
4954 ino_elem->parent = other_parent;
4955 list_add_tail(&ino_elem->list, &inode_list);
4956 ret = 0;
4957 }
4958 path->slots[0]++;
4959 }
4960 btrfs_add_delayed_iput(inode);
4961 }
4962
4963 return ret;
4964}
4965
4966static int copy_inode_items_to_log(struct btrfs_trans_handle *trans,
4967 struct btrfs_inode *inode,
4968 struct btrfs_key *min_key,
4969 const struct btrfs_key *max_key,
4970 struct btrfs_path *path,
4971 struct btrfs_path *dst_path,
4972 const u64 logged_isize,
4973 const bool recursive_logging,
4974 const int inode_only,
4975 struct btrfs_log_ctx *ctx,
4976 bool *need_log_inode_item)
4977{
4978 struct btrfs_root *root = inode->root;
4979 int ins_start_slot = 0;
4980 int ins_nr = 0;
4981 int ret;
4982
4983 while (1) {
4984 ret = btrfs_search_forward(root, min_key, path, trans->transid);
4985 if (ret < 0)
4986 return ret;
4987 if (ret > 0) {
4988 ret = 0;
4989 break;
4990 }
4991again:
4992 /* Note, ins_nr might be > 0 here, cleanup outside the loop */
4993 if (min_key->objectid != max_key->objectid)
4994 break;
4995 if (min_key->type > max_key->type)
4996 break;
4997
4998 if (min_key->type == BTRFS_INODE_ITEM_KEY)
4999 *need_log_inode_item = false;
5000
5001 if ((min_key->type == BTRFS_INODE_REF_KEY ||
5002 min_key->type == BTRFS_INODE_EXTREF_KEY) &&
5003 inode->generation == trans->transid &&
5004 !recursive_logging) {
5005 u64 other_ino = 0;
5006 u64 other_parent = 0;
5007
5008 ret = btrfs_check_ref_name_override(path->nodes[0],
5009 path->slots[0], min_key, inode,
5010 &other_ino, &other_parent);
5011 if (ret < 0) {
5012 return ret;
5013 } else if (ret > 0 && ctx &&
5014 other_ino != btrfs_ino(BTRFS_I(ctx->inode))) {
5015 if (ins_nr > 0) {
5016 ins_nr++;
5017 } else {
5018 ins_nr = 1;
5019 ins_start_slot = path->slots[0];
5020 }
5021 ret = copy_items(trans, inode, dst_path, path,
5022 ins_start_slot, ins_nr,
5023 inode_only, logged_isize);
5024 if (ret < 0)
5025 return ret;
5026 ins_nr = 0;
5027
5028 ret = log_conflicting_inodes(trans, root, path,
5029 ctx, other_ino, other_parent);
5030 if (ret)
5031 return ret;
5032 btrfs_release_path(path);
5033 goto next_key;
5034 }
5035 }
5036
5037 /* Skip xattrs, we log them later with btrfs_log_all_xattrs() */
5038 if (min_key->type == BTRFS_XATTR_ITEM_KEY) {
5039 if (ins_nr == 0)
5040 goto next_slot;
5041 ret = copy_items(trans, inode, dst_path, path,
5042 ins_start_slot,
5043 ins_nr, inode_only, logged_isize);
5044 if (ret < 0)
5045 return ret;
5046 ins_nr = 0;
5047 goto next_slot;
5048 }
5049
5050 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
5051 ins_nr++;
5052 goto next_slot;
5053 } else if (!ins_nr) {
5054 ins_start_slot = path->slots[0];
5055 ins_nr = 1;
5056 goto next_slot;
5057 }
5058
5059 ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
5060 ins_nr, inode_only, logged_isize);
5061 if (ret < 0)
5062 return ret;
5063 ins_nr = 1;
5064 ins_start_slot = path->slots[0];
5065next_slot:
5066 path->slots[0]++;
5067 if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
5068 btrfs_item_key_to_cpu(path->nodes[0], min_key,
5069 path->slots[0]);
5070 goto again;
5071 }
5072 if (ins_nr) {
5073 ret = copy_items(trans, inode, dst_path, path,
5074 ins_start_slot, ins_nr, inode_only,
5075 logged_isize);
5076 if (ret < 0)
5077 return ret;
5078 ins_nr = 0;
5079 }
5080 btrfs_release_path(path);
5081next_key:
5082 if (min_key->offset < (u64)-1) {
5083 min_key->offset++;
5084 } else if (min_key->type < max_key->type) {
5085 min_key->type++;
5086 min_key->offset = 0;
5087 } else {
5088 break;
5089 }
5090 }
5091 if (ins_nr)
5092 ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
5093 ins_nr, inode_only, logged_isize);
5094
5095 return ret;
5096}
5097
5098/* log a single inode in the tree log.
5099 * At least one parent directory for this inode must exist in the tree
5100 * or be logged already.
5101 *
5102 * Any items from this inode changed by the current transaction are copied
5103 * to the log tree. An extra reference is taken on any extents in this
5104 * file, allowing us to avoid a whole pile of corner cases around logging
5105 * blocks that have been removed from the tree.
5106 *
5107 * See LOG_INODE_ALL and related defines for a description of what inode_only
5108 * does.
5109 *
5110 * This handles both files and directories.
5111 */
5112static int btrfs_log_inode(struct btrfs_trans_handle *trans,
5113 struct btrfs_root *root, struct btrfs_inode *inode,
5114 int inode_only,
5115 const loff_t start,
5116 const loff_t end,
5117 struct btrfs_log_ctx *ctx)
5118{
5119 struct btrfs_path *path;
5120 struct btrfs_path *dst_path;
5121 struct btrfs_key min_key;
5122 struct btrfs_key max_key;
5123 struct btrfs_root *log = root->log_root;
5124 int err = 0;
5125 int ret = 0;
5126 bool fast_search = false;
5127 u64 ino = btrfs_ino(inode);
5128 struct extent_map_tree *em_tree = &inode->extent_tree;
5129 u64 logged_isize = 0;
5130 bool need_log_inode_item = true;
5131 bool xattrs_logged = false;
5132 bool recursive_logging = false;
5133
5134 path = btrfs_alloc_path();
5135 if (!path)
5136 return -ENOMEM;
5137 dst_path = btrfs_alloc_path();
5138 if (!dst_path) {
5139 btrfs_free_path(path);
5140 return -ENOMEM;
5141 }
5142
5143 min_key.objectid = ino;
5144 min_key.type = BTRFS_INODE_ITEM_KEY;
5145 min_key.offset = 0;
5146
5147 max_key.objectid = ino;
5148
5149
5150 /* today the code can only do partial logging of directories */
5151 if (S_ISDIR(inode->vfs_inode.i_mode) ||
5152 (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5153 &inode->runtime_flags) &&
5154 inode_only >= LOG_INODE_EXISTS))
5155 max_key.type = BTRFS_XATTR_ITEM_KEY;
5156 else
5157 max_key.type = (u8)-1;
5158 max_key.offset = (u64)-1;
5159
5160 /*
5161 * Only run delayed items if we are a directory. We want to make sure
5162 * all directory indexes hit the fs/subvolume tree so we can find them
5163 * and figure out which index ranges have to be logged.
5164 *
5165 * Otherwise commit the delayed inode only if the full sync flag is set,
5166 * as we want to make sure an up to date version is in the subvolume
5167 * tree so copy_inode_items_to_log() / copy_items() can find it and copy
5168 * it to the log tree. For a non full sync, we always log the inode item
5169 * based on the in-memory struct btrfs_inode which is always up to date.
5170 */
5171 if (S_ISDIR(inode->vfs_inode.i_mode))
5172 ret = btrfs_commit_inode_delayed_items(trans, inode);
5173 else if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags))
5174 ret = btrfs_commit_inode_delayed_inode(inode);
5175
5176 if (ret) {
5177 btrfs_free_path(path);
5178 btrfs_free_path(dst_path);
5179 return ret;
5180 }
5181
5182 if (inode_only == LOG_OTHER_INODE || inode_only == LOG_OTHER_INODE_ALL) {
5183 recursive_logging = true;
5184 if (inode_only == LOG_OTHER_INODE)
5185 inode_only = LOG_INODE_EXISTS;
5186 else
5187 inode_only = LOG_INODE_ALL;
5188 mutex_lock_nested(&inode->log_mutex, SINGLE_DEPTH_NESTING);
5189 } else {
5190 mutex_lock(&inode->log_mutex);
5191 }
5192
5193 /*
5194 * a brute force approach to making sure we get the most uptodate
5195 * copies of everything.
5196 */
5197 if (S_ISDIR(inode->vfs_inode.i_mode)) {
5198 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
5199
5200 if (inode_only == LOG_INODE_EXISTS)
5201 max_key_type = BTRFS_XATTR_ITEM_KEY;
5202 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
5203 } else {
5204 if (inode_only == LOG_INODE_EXISTS) {
5205 /*
5206 * Make sure the new inode item we write to the log has
5207 * the same isize as the current one (if it exists).
5208 * This is necessary to prevent data loss after log
5209 * replay, and also to prevent doing a wrong expanding
5210 * truncate - for e.g. create file, write 4K into offset
5211 * 0, fsync, write 4K into offset 4096, add hard link,
5212 * fsync some other file (to sync log), power fail - if
5213 * we use the inode's current i_size, after log replay
5214 * we get a 8Kb file, with the last 4Kb extent as a hole
5215 * (zeroes), as if an expanding truncate happened,
5216 * instead of getting a file of 4Kb only.
5217 */
5218 err = logged_inode_size(log, inode, path, &logged_isize);
5219 if (err)
5220 goto out_unlock;
5221 }
5222 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5223 &inode->runtime_flags)) {
5224 if (inode_only == LOG_INODE_EXISTS) {
5225 max_key.type = BTRFS_XATTR_ITEM_KEY;
5226 ret = drop_objectid_items(trans, log, path, ino,
5227 max_key.type);
5228 } else {
5229 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5230 &inode->runtime_flags);
5231 clear_bit(BTRFS_INODE_COPY_EVERYTHING,
5232 &inode->runtime_flags);
5233 while(1) {
5234 ret = btrfs_truncate_inode_items(trans,
5235 log, &inode->vfs_inode, 0, 0);
5236 if (ret != -EAGAIN)
5237 break;
5238 }
5239 }
5240 } else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
5241 &inode->runtime_flags) ||
5242 inode_only == LOG_INODE_EXISTS) {
5243 if (inode_only == LOG_INODE_ALL)
5244 fast_search = true;
5245 max_key.type = BTRFS_XATTR_ITEM_KEY;
5246 ret = drop_objectid_items(trans, log, path, ino,
5247 max_key.type);
5248 } else {
5249 if (inode_only == LOG_INODE_ALL)
5250 fast_search = true;
5251 goto log_extents;
5252 }
5253
5254 }
5255 if (ret) {
5256 err = ret;
5257 goto out_unlock;
5258 }
5259
5260 err = copy_inode_items_to_log(trans, inode, &min_key, &max_key,
5261 path, dst_path, logged_isize,
5262 recursive_logging, inode_only, ctx,
5263 &need_log_inode_item);
5264 if (err)
5265 goto out_unlock;
5266
5267 btrfs_release_path(path);
5268 btrfs_release_path(dst_path);
5269 err = btrfs_log_all_xattrs(trans, root, inode, path, dst_path);
5270 if (err)
5271 goto out_unlock;
5272 xattrs_logged = true;
5273 if (max_key.type >= BTRFS_EXTENT_DATA_KEY && !fast_search) {
5274 btrfs_release_path(path);
5275 btrfs_release_path(dst_path);
5276 err = btrfs_log_holes(trans, root, inode, path);
5277 if (err)
5278 goto out_unlock;
5279 }
5280log_extents:
5281 btrfs_release_path(path);
5282 btrfs_release_path(dst_path);
5283 if (need_log_inode_item) {
5284 err = log_inode_item(trans, log, dst_path, inode);
5285 if (!err && !xattrs_logged) {
5286 err = btrfs_log_all_xattrs(trans, root, inode, path,
5287 dst_path);
5288 btrfs_release_path(path);
5289 }
5290 if (err)
5291 goto out_unlock;
5292 }
5293 if (fast_search) {
5294 ret = btrfs_log_changed_extents(trans, root, inode, dst_path,
5295 ctx, start, end);
5296 if (ret) {
5297 err = ret;
5298 goto out_unlock;
5299 }
5300 } else if (inode_only == LOG_INODE_ALL) {
5301 struct extent_map *em, *n;
5302
5303 write_lock(&em_tree->lock);
5304 /*
5305 * We can't just remove every em if we're called for a ranged
5306 * fsync - that is, one that doesn't cover the whole possible
5307 * file range (0 to LLONG_MAX). This is because we can have
5308 * em's that fall outside the range we're logging and therefore
5309 * their ordered operations haven't completed yet
5310 * (btrfs_finish_ordered_io() not invoked yet). This means we
5311 * didn't get their respective file extent item in the fs/subvol
5312 * tree yet, and need to let the next fast fsync (one which
5313 * consults the list of modified extent maps) find the em so
5314 * that it logs a matching file extent item and waits for the
5315 * respective ordered operation to complete (if it's still
5316 * running).
5317 *
5318 * Removing every em outside the range we're logging would make
5319 * the next fast fsync not log their matching file extent items,
5320 * therefore making us lose data after a log replay.
5321 */
5322 list_for_each_entry_safe(em, n, &em_tree->modified_extents,
5323 list) {
5324 const u64 mod_end = em->mod_start + em->mod_len - 1;
5325
5326 if (em->mod_start >= start && mod_end <= end)
5327 list_del_init(&em->list);
5328 }
5329 write_unlock(&em_tree->lock);
5330 }
5331
5332 if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->vfs_inode.i_mode)) {
5333 ret = log_directory_changes(trans, root, inode, path, dst_path,
5334 ctx);
5335 if (ret) {
5336 err = ret;
5337 goto out_unlock;
5338 }
5339 }
5340
5341 /*
5342 * Don't update last_log_commit if we logged that an inode exists after
5343 * it was loaded to memory (full_sync bit set).
5344 * This is to prevent data loss when we do a write to the inode, then
5345 * the inode gets evicted after all delalloc was flushed, then we log
5346 * it exists (due to a rename for example) and then fsync it. This last
5347 * fsync would do nothing (not logging the extents previously written).
5348 */
5349 spin_lock(&inode->lock);
5350 inode->logged_trans = trans->transid;
5351 if (inode_only != LOG_INODE_EXISTS ||
5352 !test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags))
5353 inode->last_log_commit = inode->last_sub_trans;
5354 spin_unlock(&inode->lock);
5355out_unlock:
5356 mutex_unlock(&inode->log_mutex);
5357
5358 btrfs_free_path(path);
5359 btrfs_free_path(dst_path);
5360 return err;
5361}
5362
5363/*
5364 * Check if we must fallback to a transaction commit when logging an inode.
5365 * This must be called after logging the inode and is used only in the context
5366 * when fsyncing an inode requires the need to log some other inode - in which
5367 * case we can't lock the i_mutex of each other inode we need to log as that
5368 * can lead to deadlocks with concurrent fsync against other inodes (as we can
5369 * log inodes up or down in the hierarchy) or rename operations for example. So
5370 * we take the log_mutex of the inode after we have logged it and then check for
5371 * its last_unlink_trans value - this is safe because any task setting
5372 * last_unlink_trans must take the log_mutex and it must do this before it does
5373 * the actual unlink operation, so if we do this check before a concurrent task
5374 * sets last_unlink_trans it means we've logged a consistent version/state of
5375 * all the inode items, otherwise we are not sure and must do a transaction
5376 * commit (the concurrent task might have only updated last_unlink_trans before
5377 * we logged the inode or it might have also done the unlink).
5378 */
5379static bool btrfs_must_commit_transaction(struct btrfs_trans_handle *trans,
5380 struct btrfs_inode *inode)
5381{
5382 struct btrfs_fs_info *fs_info = inode->root->fs_info;
5383 bool ret = false;
5384
5385 mutex_lock(&inode->log_mutex);
5386 if (inode->last_unlink_trans > fs_info->last_trans_committed) {
5387 /*
5388 * Make sure any commits to the log are forced to be full
5389 * commits.
5390 */
5391 btrfs_set_log_full_commit(trans);
5392 ret = true;
5393 }
5394 mutex_unlock(&inode->log_mutex);
5395
5396 return ret;
5397}
5398
5399/*
5400 * follow the dentry parent pointers up the chain and see if any
5401 * of the directories in it require a full commit before they can
5402 * be logged. Returns zero if nothing special needs to be done or 1 if
5403 * a full commit is required.
5404 */
5405static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
5406 struct btrfs_inode *inode,
5407 struct dentry *parent,
5408 struct super_block *sb,
5409 u64 last_committed)
5410{
5411 int ret = 0;
5412 struct dentry *old_parent = NULL;
5413
5414 /*
5415 * for regular files, if its inode is already on disk, we don't
5416 * have to worry about the parents at all. This is because
5417 * we can use the last_unlink_trans field to record renames
5418 * and other fun in this file.
5419 */
5420 if (S_ISREG(inode->vfs_inode.i_mode) &&
5421 inode->generation <= last_committed &&
5422 inode->last_unlink_trans <= last_committed)
5423 goto out;
5424
5425 if (!S_ISDIR(inode->vfs_inode.i_mode)) {
5426 if (!parent || d_really_is_negative(parent) || sb != parent->d_sb)
5427 goto out;
5428 inode = BTRFS_I(d_inode(parent));
5429 }
5430
5431 while (1) {
5432 if (btrfs_must_commit_transaction(trans, inode)) {
5433 ret = 1;
5434 break;
5435 }
5436
5437 if (!parent || d_really_is_negative(parent) || sb != parent->d_sb)
5438 break;
5439
5440 if (IS_ROOT(parent)) {
5441 inode = BTRFS_I(d_inode(parent));
5442 if (btrfs_must_commit_transaction(trans, inode))
5443 ret = 1;
5444 break;
5445 }
5446
5447 parent = dget_parent(parent);
5448 dput(old_parent);
5449 old_parent = parent;
5450 inode = BTRFS_I(d_inode(parent));
5451
5452 }
5453 dput(old_parent);
5454out:
5455 return ret;
5456}
5457
5458struct btrfs_dir_list {
5459 u64 ino;
5460 struct list_head list;
5461};
5462
5463/*
5464 * Log the inodes of the new dentries of a directory. See log_dir_items() for
5465 * details about the why it is needed.
5466 * This is a recursive operation - if an existing dentry corresponds to a
5467 * directory, that directory's new entries are logged too (same behaviour as
5468 * ext3/4, xfs, f2fs, reiserfs, nilfs2). Note that when logging the inodes
5469 * the dentries point to we do not lock their i_mutex, otherwise lockdep
5470 * complains about the following circular lock dependency / possible deadlock:
5471 *
5472 * CPU0 CPU1
5473 * ---- ----
5474 * lock(&type->i_mutex_dir_key#3/2);
5475 * lock(sb_internal#2);
5476 * lock(&type->i_mutex_dir_key#3/2);
5477 * lock(&sb->s_type->i_mutex_key#14);
5478 *
5479 * Where sb_internal is the lock (a counter that works as a lock) acquired by
5480 * sb_start_intwrite() in btrfs_start_transaction().
5481 * Not locking i_mutex of the inodes is still safe because:
5482 *
5483 * 1) For regular files we log with a mode of LOG_INODE_EXISTS. It's possible
5484 * that while logging the inode new references (names) are added or removed
5485 * from the inode, leaving the logged inode item with a link count that does
5486 * not match the number of logged inode reference items. This is fine because
5487 * at log replay time we compute the real number of links and correct the
5488 * link count in the inode item (see replay_one_buffer() and
5489 * link_to_fixup_dir());
5490 *
5491 * 2) For directories we log with a mode of LOG_INODE_ALL. It's possible that
5492 * while logging the inode's items new items with keys BTRFS_DIR_ITEM_KEY and
5493 * BTRFS_DIR_INDEX_KEY are added to fs/subvol tree and the logged inode item
5494 * has a size that doesn't match the sum of the lengths of all the logged
5495 * names. This does not result in a problem because if a dir_item key is
5496 * logged but its matching dir_index key is not logged, at log replay time we
5497 * don't use it to replay the respective name (see replay_one_name()). On the
5498 * other hand if only the dir_index key ends up being logged, the respective
5499 * name is added to the fs/subvol tree with both the dir_item and dir_index
5500 * keys created (see replay_one_name()).
5501 * The directory's inode item with a wrong i_size is not a problem as well,
5502 * since we don't use it at log replay time to set the i_size in the inode
5503 * item of the fs/subvol tree (see overwrite_item()).
5504 */
5505static int log_new_dir_dentries(struct btrfs_trans_handle *trans,
5506 struct btrfs_root *root,
5507 struct btrfs_inode *start_inode,
5508 struct btrfs_log_ctx *ctx)
5509{
5510 struct btrfs_fs_info *fs_info = root->fs_info;
5511 struct btrfs_root *log = root->log_root;
5512 struct btrfs_path *path;
5513 LIST_HEAD(dir_list);
5514 struct btrfs_dir_list *dir_elem;
5515 int ret = 0;
5516
5517 path = btrfs_alloc_path();
5518 if (!path)
5519 return -ENOMEM;
5520
5521 dir_elem = kmalloc(sizeof(*dir_elem), GFP_NOFS);
5522 if (!dir_elem) {
5523 btrfs_free_path(path);
5524 return -ENOMEM;
5525 }
5526 dir_elem->ino = btrfs_ino(start_inode);
5527 list_add_tail(&dir_elem->list, &dir_list);
5528
5529 while (!list_empty(&dir_list)) {
5530 struct extent_buffer *leaf;
5531 struct btrfs_key min_key;
5532 int nritems;
5533 int i;
5534
5535 dir_elem = list_first_entry(&dir_list, struct btrfs_dir_list,
5536 list);
5537 if (ret)
5538 goto next_dir_inode;
5539
5540 min_key.objectid = dir_elem->ino;
5541 min_key.type = BTRFS_DIR_ITEM_KEY;
5542 min_key.offset = 0;
5543again:
5544 btrfs_release_path(path);
5545 ret = btrfs_search_forward(log, &min_key, path, trans->transid);
5546 if (ret < 0) {
5547 goto next_dir_inode;
5548 } else if (ret > 0) {
5549 ret = 0;
5550 goto next_dir_inode;
5551 }
5552
5553process_leaf:
5554 leaf = path->nodes[0];
5555 nritems = btrfs_header_nritems(leaf);
5556 for (i = path->slots[0]; i < nritems; i++) {
5557 struct btrfs_dir_item *di;
5558 struct btrfs_key di_key;
5559 struct inode *di_inode;
5560 struct btrfs_dir_list *new_dir_elem;
5561 int log_mode = LOG_INODE_EXISTS;
5562 int type;
5563
5564 btrfs_item_key_to_cpu(leaf, &min_key, i);
5565 if (min_key.objectid != dir_elem->ino ||
5566 min_key.type != BTRFS_DIR_ITEM_KEY)
5567 goto next_dir_inode;
5568
5569 di = btrfs_item_ptr(leaf, i, struct btrfs_dir_item);
5570 type = btrfs_dir_type(leaf, di);
5571 if (btrfs_dir_transid(leaf, di) < trans->transid &&
5572 type != BTRFS_FT_DIR)
5573 continue;
5574 btrfs_dir_item_key_to_cpu(leaf, di, &di_key);
5575 if (di_key.type == BTRFS_ROOT_ITEM_KEY)
5576 continue;
5577
5578 btrfs_release_path(path);
5579 di_inode = btrfs_iget(fs_info->sb, di_key.objectid, root);
5580 if (IS_ERR(di_inode)) {
5581 ret = PTR_ERR(di_inode);
5582 goto next_dir_inode;
5583 }
5584
5585 if (btrfs_inode_in_log(BTRFS_I(di_inode), trans->transid)) {
5586 btrfs_add_delayed_iput(di_inode);
5587 break;
5588 }
5589
5590 ctx->log_new_dentries = false;
5591 if (type == BTRFS_FT_DIR || type == BTRFS_FT_SYMLINK)
5592 log_mode = LOG_INODE_ALL;
5593 ret = btrfs_log_inode(trans, root, BTRFS_I(di_inode),
5594 log_mode, 0, LLONG_MAX, ctx);
5595 if (!ret &&
5596 btrfs_must_commit_transaction(trans, BTRFS_I(di_inode)))
5597 ret = 1;
5598 btrfs_add_delayed_iput(di_inode);
5599 if (ret)
5600 goto next_dir_inode;
5601 if (ctx->log_new_dentries) {
5602 new_dir_elem = kmalloc(sizeof(*new_dir_elem),
5603 GFP_NOFS);
5604 if (!new_dir_elem) {
5605 ret = -ENOMEM;
5606 goto next_dir_inode;
5607 }
5608 new_dir_elem->ino = di_key.objectid;
5609 list_add_tail(&new_dir_elem->list, &dir_list);
5610 }
5611 break;
5612 }
5613 if (i == nritems) {
5614 ret = btrfs_next_leaf(log, path);
5615 if (ret < 0) {
5616 goto next_dir_inode;
5617 } else if (ret > 0) {
5618 ret = 0;
5619 goto next_dir_inode;
5620 }
5621 goto process_leaf;
5622 }
5623 if (min_key.offset < (u64)-1) {
5624 min_key.offset++;
5625 goto again;
5626 }
5627next_dir_inode:
5628 list_del(&dir_elem->list);
5629 kfree(dir_elem);
5630 }
5631
5632 btrfs_free_path(path);
5633 return ret;
5634}
5635
5636static int btrfs_log_all_parents(struct btrfs_trans_handle *trans,
5637 struct btrfs_inode *inode,
5638 struct btrfs_log_ctx *ctx)
5639{
5640 struct btrfs_fs_info *fs_info = trans->fs_info;
5641 int ret;
5642 struct btrfs_path *path;
5643 struct btrfs_key key;
5644 struct btrfs_root *root = inode->root;
5645 const u64 ino = btrfs_ino(inode);
5646
5647 path = btrfs_alloc_path();
5648 if (!path)
5649 return -ENOMEM;
5650 path->skip_locking = 1;
5651 path->search_commit_root = 1;
5652
5653 key.objectid = ino;
5654 key.type = BTRFS_INODE_REF_KEY;
5655 key.offset = 0;
5656 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5657 if (ret < 0)
5658 goto out;
5659
5660 while (true) {
5661 struct extent_buffer *leaf = path->nodes[0];
5662 int slot = path->slots[0];
5663 u32 cur_offset = 0;
5664 u32 item_size;
5665 unsigned long ptr;
5666
5667 if (slot >= btrfs_header_nritems(leaf)) {
5668 ret = btrfs_next_leaf(root, path);
5669 if (ret < 0)
5670 goto out;
5671 else if (ret > 0)
5672 break;
5673 continue;
5674 }
5675
5676 btrfs_item_key_to_cpu(leaf, &key, slot);
5677 /* BTRFS_INODE_EXTREF_KEY is BTRFS_INODE_REF_KEY + 1 */
5678 if (key.objectid != ino || key.type > BTRFS_INODE_EXTREF_KEY)
5679 break;
5680
5681 item_size = btrfs_item_size_nr(leaf, slot);
5682 ptr = btrfs_item_ptr_offset(leaf, slot);
5683 while (cur_offset < item_size) {
5684 struct btrfs_key inode_key;
5685 struct inode *dir_inode;
5686
5687 inode_key.type = BTRFS_INODE_ITEM_KEY;
5688 inode_key.offset = 0;
5689
5690 if (key.type == BTRFS_INODE_EXTREF_KEY) {
5691 struct btrfs_inode_extref *extref;
5692
5693 extref = (struct btrfs_inode_extref *)
5694 (ptr + cur_offset);
5695 inode_key.objectid = btrfs_inode_extref_parent(
5696 leaf, extref);
5697 cur_offset += sizeof(*extref);
5698 cur_offset += btrfs_inode_extref_name_len(leaf,
5699 extref);
5700 } else {
5701 inode_key.objectid = key.offset;
5702 cur_offset = item_size;
5703 }
5704
5705 dir_inode = btrfs_iget(fs_info->sb, inode_key.objectid,
5706 root);
5707 /*
5708 * If the parent inode was deleted, return an error to
5709 * fallback to a transaction commit. This is to prevent
5710 * getting an inode that was moved from one parent A to
5711 * a parent B, got its former parent A deleted and then
5712 * it got fsync'ed, from existing at both parents after
5713 * a log replay (and the old parent still existing).
5714 * Example:
5715 *
5716 * mkdir /mnt/A
5717 * mkdir /mnt/B
5718 * touch /mnt/B/bar
5719 * sync
5720 * mv /mnt/B/bar /mnt/A/bar
5721 * mv -T /mnt/A /mnt/B
5722 * fsync /mnt/B/bar
5723 * <power fail>
5724 *
5725 * If we ignore the old parent B which got deleted,
5726 * after a log replay we would have file bar linked
5727 * at both parents and the old parent B would still
5728 * exist.
5729 */
5730 if (IS_ERR(dir_inode)) {
5731 ret = PTR_ERR(dir_inode);
5732 goto out;
5733 }
5734
5735 if (ctx)
5736 ctx->log_new_dentries = false;
5737 ret = btrfs_log_inode(trans, root, BTRFS_I(dir_inode),
5738 LOG_INODE_ALL, 0, LLONG_MAX, ctx);
5739 if (!ret &&
5740 btrfs_must_commit_transaction(trans, BTRFS_I(dir_inode)))
5741 ret = 1;
5742 if (!ret && ctx && ctx->log_new_dentries)
5743 ret = log_new_dir_dentries(trans, root,
5744 BTRFS_I(dir_inode), ctx);
5745 btrfs_add_delayed_iput(dir_inode);
5746 if (ret)
5747 goto out;
5748 }
5749 path->slots[0]++;
5750 }
5751 ret = 0;
5752out:
5753 btrfs_free_path(path);
5754 return ret;
5755}
5756
5757static int log_new_ancestors(struct btrfs_trans_handle *trans,
5758 struct btrfs_root *root,
5759 struct btrfs_path *path,
5760 struct btrfs_log_ctx *ctx)
5761{
5762 struct btrfs_key found_key;
5763
5764 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
5765
5766 while (true) {
5767 struct btrfs_fs_info *fs_info = root->fs_info;
5768 const u64 last_committed = fs_info->last_trans_committed;
5769 struct extent_buffer *leaf = path->nodes[0];
5770 int slot = path->slots[0];
5771 struct btrfs_key search_key;
5772 struct inode *inode;
5773 u64 ino;
5774 int ret = 0;
5775
5776 btrfs_release_path(path);
5777
5778 ino = found_key.offset;
5779
5780 search_key.objectid = found_key.offset;
5781 search_key.type = BTRFS_INODE_ITEM_KEY;
5782 search_key.offset = 0;
5783 inode = btrfs_iget(fs_info->sb, ino, root);
5784 if (IS_ERR(inode))
5785 return PTR_ERR(inode);
5786
5787 if (BTRFS_I(inode)->generation > last_committed)
5788 ret = btrfs_log_inode(trans, root, BTRFS_I(inode),
5789 LOG_INODE_EXISTS,
5790 0, LLONG_MAX, ctx);
5791 btrfs_add_delayed_iput(inode);
5792 if (ret)
5793 return ret;
5794
5795 if (search_key.objectid == BTRFS_FIRST_FREE_OBJECTID)
5796 break;
5797
5798 search_key.type = BTRFS_INODE_REF_KEY;
5799 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
5800 if (ret < 0)
5801 return ret;
5802
5803 leaf = path->nodes[0];
5804 slot = path->slots[0];
5805 if (slot >= btrfs_header_nritems(leaf)) {
5806 ret = btrfs_next_leaf(root, path);
5807 if (ret < 0)
5808 return ret;
5809 else if (ret > 0)
5810 return -ENOENT;
5811 leaf = path->nodes[0];
5812 slot = path->slots[0];
5813 }
5814
5815 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5816 if (found_key.objectid != search_key.objectid ||
5817 found_key.type != BTRFS_INODE_REF_KEY)
5818 return -ENOENT;
5819 }
5820 return 0;
5821}
5822
5823static int log_new_ancestors_fast(struct btrfs_trans_handle *trans,
5824 struct btrfs_inode *inode,
5825 struct dentry *parent,
5826 struct btrfs_log_ctx *ctx)
5827{
5828 struct btrfs_root *root = inode->root;
5829 struct btrfs_fs_info *fs_info = root->fs_info;
5830 struct dentry *old_parent = NULL;
5831 struct super_block *sb = inode->vfs_inode.i_sb;
5832 int ret = 0;
5833
5834 while (true) {
5835 if (!parent || d_really_is_negative(parent) ||
5836 sb != parent->d_sb)
5837 break;
5838
5839 inode = BTRFS_I(d_inode(parent));
5840 if (root != inode->root)
5841 break;
5842
5843 if (inode->generation > fs_info->last_trans_committed) {
5844 ret = btrfs_log_inode(trans, root, inode,
5845 LOG_INODE_EXISTS, 0, LLONG_MAX, ctx);
5846 if (ret)
5847 break;
5848 }
5849 if (IS_ROOT(parent))
5850 break;
5851
5852 parent = dget_parent(parent);
5853 dput(old_parent);
5854 old_parent = parent;
5855 }
5856 dput(old_parent);
5857
5858 return ret;
5859}
5860
5861static int log_all_new_ancestors(struct btrfs_trans_handle *trans,
5862 struct btrfs_inode *inode,
5863 struct dentry *parent,
5864 struct btrfs_log_ctx *ctx)
5865{
5866 struct btrfs_root *root = inode->root;
5867 const u64 ino = btrfs_ino(inode);
5868 struct btrfs_path *path;
5869 struct btrfs_key search_key;
5870 int ret;
5871
5872 /*
5873 * For a single hard link case, go through a fast path that does not
5874 * need to iterate the fs/subvolume tree.
5875 */
5876 if (inode->vfs_inode.i_nlink < 2)
5877 return log_new_ancestors_fast(trans, inode, parent, ctx);
5878
5879 path = btrfs_alloc_path();
5880 if (!path)
5881 return -ENOMEM;
5882
5883 search_key.objectid = ino;
5884 search_key.type = BTRFS_INODE_REF_KEY;
5885 search_key.offset = 0;
5886again:
5887 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
5888 if (ret < 0)
5889 goto out;
5890 if (ret == 0)
5891 path->slots[0]++;
5892
5893 while (true) {
5894 struct extent_buffer *leaf = path->nodes[0];
5895 int slot = path->slots[0];
5896 struct btrfs_key found_key;
5897
5898 if (slot >= btrfs_header_nritems(leaf)) {
5899 ret = btrfs_next_leaf(root, path);
5900 if (ret < 0)
5901 goto out;
5902 else if (ret > 0)
5903 break;
5904 continue;
5905 }
5906
5907 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5908 if (found_key.objectid != ino ||
5909 found_key.type > BTRFS_INODE_EXTREF_KEY)
5910 break;
5911
5912 /*
5913 * Don't deal with extended references because they are rare
5914 * cases and too complex to deal with (we would need to keep
5915 * track of which subitem we are processing for each item in
5916 * this loop, etc). So just return some error to fallback to
5917 * a transaction commit.
5918 */
5919 if (found_key.type == BTRFS_INODE_EXTREF_KEY) {
5920 ret = -EMLINK;
5921 goto out;
5922 }
5923
5924 /*
5925 * Logging ancestors needs to do more searches on the fs/subvol
5926 * tree, so it releases the path as needed to avoid deadlocks.
5927 * Keep track of the last inode ref key and resume from that key
5928 * after logging all new ancestors for the current hard link.
5929 */
5930 memcpy(&search_key, &found_key, sizeof(search_key));
5931
5932 ret = log_new_ancestors(trans, root, path, ctx);
5933 if (ret)
5934 goto out;
5935 btrfs_release_path(path);
5936 goto again;
5937 }
5938 ret = 0;
5939out:
5940 btrfs_free_path(path);
5941 return ret;
5942}
5943
5944/*
5945 * helper function around btrfs_log_inode to make sure newly created
5946 * parent directories also end up in the log. A minimal inode and backref
5947 * only logging is done of any parent directories that are older than
5948 * the last committed transaction
5949 */
5950static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
5951 struct btrfs_inode *inode,
5952 struct dentry *parent,
5953 const loff_t start,
5954 const loff_t end,
5955 int inode_only,
5956 struct btrfs_log_ctx *ctx)
5957{
5958 struct btrfs_root *root = inode->root;
5959 struct btrfs_fs_info *fs_info = root->fs_info;
5960 struct super_block *sb;
5961 int ret = 0;
5962 u64 last_committed = fs_info->last_trans_committed;
5963 bool log_dentries = false;
5964
5965 sb = inode->vfs_inode.i_sb;
5966
5967 if (btrfs_test_opt(fs_info, NOTREELOG)) {
5968 ret = 1;
5969 goto end_no_trans;
5970 }
5971
5972 /*
5973 * The prev transaction commit doesn't complete, we need do
5974 * full commit by ourselves.
5975 */
5976 if (fs_info->last_trans_log_full_commit >
5977 fs_info->last_trans_committed) {
5978 ret = 1;
5979 goto end_no_trans;
5980 }
5981
5982 if (btrfs_root_refs(&root->root_item) == 0) {
5983 ret = 1;
5984 goto end_no_trans;
5985 }
5986
5987 ret = check_parent_dirs_for_sync(trans, inode, parent, sb,
5988 last_committed);
5989 if (ret)
5990 goto end_no_trans;
5991
5992 /*
5993 * Skip already logged inodes or inodes corresponding to tmpfiles
5994 * (since logging them is pointless, a link count of 0 means they
5995 * will never be accessible).
5996 */
5997 if (btrfs_inode_in_log(inode, trans->transid) ||
5998 inode->vfs_inode.i_nlink == 0) {
5999 ret = BTRFS_NO_LOG_SYNC;
6000 goto end_no_trans;
6001 }
6002
6003 ret = start_log_trans(trans, root, ctx);
6004 if (ret)
6005 goto end_no_trans;
6006
6007 ret = btrfs_log_inode(trans, root, inode, inode_only, start, end, ctx);
6008 if (ret)
6009 goto end_trans;
6010
6011 /*
6012 * for regular files, if its inode is already on disk, we don't
6013 * have to worry about the parents at all. This is because
6014 * we can use the last_unlink_trans field to record renames
6015 * and other fun in this file.
6016 */
6017 if (S_ISREG(inode->vfs_inode.i_mode) &&
6018 inode->generation <= last_committed &&
6019 inode->last_unlink_trans <= last_committed) {
6020 ret = 0;
6021 goto end_trans;
6022 }
6023
6024 if (S_ISDIR(inode->vfs_inode.i_mode) && ctx && ctx->log_new_dentries)
6025 log_dentries = true;
6026
6027 /*
6028 * On unlink we must make sure all our current and old parent directory
6029 * inodes are fully logged. This is to prevent leaving dangling
6030 * directory index entries in directories that were our parents but are
6031 * not anymore. Not doing this results in old parent directory being
6032 * impossible to delete after log replay (rmdir will always fail with
6033 * error -ENOTEMPTY).
6034 *
6035 * Example 1:
6036 *
6037 * mkdir testdir
6038 * touch testdir/foo
6039 * ln testdir/foo testdir/bar
6040 * sync
6041 * unlink testdir/bar
6042 * xfs_io -c fsync testdir/foo
6043 * <power failure>
6044 * mount fs, triggers log replay
6045 *
6046 * If we don't log the parent directory (testdir), after log replay the
6047 * directory still has an entry pointing to the file inode using the bar
6048 * name, but a matching BTRFS_INODE_[REF|EXTREF]_KEY does not exist and
6049 * the file inode has a link count of 1.
6050 *
6051 * Example 2:
6052 *
6053 * mkdir testdir
6054 * touch foo
6055 * ln foo testdir/foo2
6056 * ln foo testdir/foo3
6057 * sync
6058 * unlink testdir/foo3
6059 * xfs_io -c fsync foo
6060 * <power failure>
6061 * mount fs, triggers log replay
6062 *
6063 * Similar as the first example, after log replay the parent directory
6064 * testdir still has an entry pointing to the inode file with name foo3
6065 * but the file inode does not have a matching BTRFS_INODE_REF_KEY item
6066 * and has a link count of 2.
6067 */
6068 if (inode->last_unlink_trans > last_committed) {
6069 ret = btrfs_log_all_parents(trans, inode, ctx);
6070 if (ret)
6071 goto end_trans;
6072 }
6073
6074 ret = log_all_new_ancestors(trans, inode, parent, ctx);
6075 if (ret)
6076 goto end_trans;
6077
6078 if (log_dentries)
6079 ret = log_new_dir_dentries(trans, root, inode, ctx);
6080 else
6081 ret = 0;
6082end_trans:
6083 if (ret < 0) {
6084 btrfs_set_log_full_commit(trans);
6085 ret = 1;
6086 }
6087
6088 if (ret)
6089 btrfs_remove_log_ctx(root, ctx);
6090 btrfs_end_log_trans(root);
6091end_no_trans:
6092 return ret;
6093}
6094
6095/*
6096 * it is not safe to log dentry if the chunk root has added new
6097 * chunks. This returns 0 if the dentry was logged, and 1 otherwise.
6098 * If this returns 1, you must commit the transaction to safely get your
6099 * data on disk.
6100 */
6101int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
6102 struct dentry *dentry,
6103 const loff_t start,
6104 const loff_t end,
6105 struct btrfs_log_ctx *ctx)
6106{
6107 struct dentry *parent = dget_parent(dentry);
6108 int ret;
6109
6110 ret = btrfs_log_inode_parent(trans, BTRFS_I(d_inode(dentry)), parent,
6111 start, end, LOG_INODE_ALL, ctx);
6112 dput(parent);
6113
6114 return ret;
6115}
6116
6117/*
6118 * should be called during mount to recover any replay any log trees
6119 * from the FS
6120 */
6121int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
6122{
6123 int ret;
6124 struct btrfs_path *path;
6125 struct btrfs_trans_handle *trans;
6126 struct btrfs_key key;
6127 struct btrfs_key found_key;
6128 struct btrfs_root *log;
6129 struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
6130 struct walk_control wc = {
6131 .process_func = process_one_buffer,
6132 .stage = LOG_WALK_PIN_ONLY,
6133 };
6134
6135 path = btrfs_alloc_path();
6136 if (!path)
6137 return -ENOMEM;
6138
6139 set_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
6140
6141 trans = btrfs_start_transaction(fs_info->tree_root, 0);
6142 if (IS_ERR(trans)) {
6143 ret = PTR_ERR(trans);
6144 goto error;
6145 }
6146
6147 wc.trans = trans;
6148 wc.pin = 1;
6149
6150 ret = walk_log_tree(trans, log_root_tree, &wc);
6151 if (ret) {
6152 btrfs_handle_fs_error(fs_info, ret,
6153 "Failed to pin buffers while recovering log root tree.");
6154 goto error;
6155 }
6156
6157again:
6158 key.objectid = BTRFS_TREE_LOG_OBJECTID;
6159 key.offset = (u64)-1;
6160 key.type = BTRFS_ROOT_ITEM_KEY;
6161
6162 while (1) {
6163 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
6164
6165 if (ret < 0) {
6166 btrfs_handle_fs_error(fs_info, ret,
6167 "Couldn't find tree log root.");
6168 goto error;
6169 }
6170 if (ret > 0) {
6171 if (path->slots[0] == 0)
6172 break;
6173 path->slots[0]--;
6174 }
6175 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
6176 path->slots[0]);
6177 btrfs_release_path(path);
6178 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
6179 break;
6180
6181 log = btrfs_read_tree_root(log_root_tree, &found_key);
6182 if (IS_ERR(log)) {
6183 ret = PTR_ERR(log);
6184 btrfs_handle_fs_error(fs_info, ret,
6185 "Couldn't read tree log root.");
6186 goto error;
6187 }
6188
6189 wc.replay_dest = btrfs_get_fs_root(fs_info, found_key.offset,
6190 true);
6191 if (IS_ERR(wc.replay_dest)) {
6192 ret = PTR_ERR(wc.replay_dest);
6193
6194 /*
6195 * We didn't find the subvol, likely because it was
6196 * deleted. This is ok, simply skip this log and go to
6197 * the next one.
6198 *
6199 * We need to exclude the root because we can't have
6200 * other log replays overwriting this log as we'll read
6201 * it back in a few more times. This will keep our
6202 * block from being modified, and we'll just bail for
6203 * each subsequent pass.
6204 */
6205 if (ret == -ENOENT)
6206 ret = btrfs_pin_extent_for_log_replay(trans,
6207 log->node->start,
6208 log->node->len);
6209 btrfs_put_root(log);
6210
6211 if (!ret)
6212 goto next;
6213 btrfs_handle_fs_error(fs_info, ret,
6214 "Couldn't read target root for tree log recovery.");
6215 goto error;
6216 }
6217
6218 wc.replay_dest->log_root = log;
6219 btrfs_record_root_in_trans(trans, wc.replay_dest);
6220 ret = walk_log_tree(trans, log, &wc);
6221
6222 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
6223 ret = fixup_inode_link_counts(trans, wc.replay_dest,
6224 path);
6225 }
6226
6227 if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
6228 struct btrfs_root *root = wc.replay_dest;
6229
6230 btrfs_release_path(path);
6231
6232 /*
6233 * We have just replayed everything, and the highest
6234 * objectid of fs roots probably has changed in case
6235 * some inode_item's got replayed.
6236 *
6237 * root->objectid_mutex is not acquired as log replay
6238 * could only happen during mount.
6239 */
6240 ret = btrfs_find_highest_objectid(root,
6241 &root->highest_objectid);
6242 }
6243
6244 wc.replay_dest->log_root = NULL;
6245 btrfs_put_root(wc.replay_dest);
6246 btrfs_put_root(log);
6247
6248 if (ret)
6249 goto error;
6250next:
6251 if (found_key.offset == 0)
6252 break;
6253 key.offset = found_key.offset - 1;
6254 }
6255 btrfs_release_path(path);
6256
6257 /* step one is to pin it all, step two is to replay just inodes */
6258 if (wc.pin) {
6259 wc.pin = 0;
6260 wc.process_func = replay_one_buffer;
6261 wc.stage = LOG_WALK_REPLAY_INODES;
6262 goto again;
6263 }
6264 /* step three is to replay everything */
6265 if (wc.stage < LOG_WALK_REPLAY_ALL) {
6266 wc.stage++;
6267 goto again;
6268 }
6269
6270 btrfs_free_path(path);
6271
6272 /* step 4: commit the transaction, which also unpins the blocks */
6273 ret = btrfs_commit_transaction(trans);
6274 if (ret)
6275 return ret;
6276
6277 log_root_tree->log_root = NULL;
6278 clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
6279 btrfs_put_root(log_root_tree);
6280
6281 return 0;
6282error:
6283 if (wc.trans)
6284 btrfs_end_transaction(wc.trans);
6285 btrfs_free_path(path);
6286 return ret;
6287}
6288
6289/*
6290 * there are some corner cases where we want to force a full
6291 * commit instead of allowing a directory to be logged.
6292 *
6293 * They revolve around files there were unlinked from the directory, and
6294 * this function updates the parent directory so that a full commit is
6295 * properly done if it is fsync'd later after the unlinks are done.
6296 *
6297 * Must be called before the unlink operations (updates to the subvolume tree,
6298 * inodes, etc) are done.
6299 */
6300void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
6301 struct btrfs_inode *dir, struct btrfs_inode *inode,
6302 int for_rename)
6303{
6304 /*
6305 * when we're logging a file, if it hasn't been renamed
6306 * or unlinked, and its inode is fully committed on disk,
6307 * we don't have to worry about walking up the directory chain
6308 * to log its parents.
6309 *
6310 * So, we use the last_unlink_trans field to put this transid
6311 * into the file. When the file is logged we check it and
6312 * don't log the parents if the file is fully on disk.
6313 */
6314 mutex_lock(&inode->log_mutex);
6315 inode->last_unlink_trans = trans->transid;
6316 mutex_unlock(&inode->log_mutex);
6317
6318 /*
6319 * if this directory was already logged any new
6320 * names for this file/dir will get recorded
6321 */
6322 if (dir->logged_trans == trans->transid)
6323 return;
6324
6325 /*
6326 * if the inode we're about to unlink was logged,
6327 * the log will be properly updated for any new names
6328 */
6329 if (inode->logged_trans == trans->transid)
6330 return;
6331
6332 /*
6333 * when renaming files across directories, if the directory
6334 * there we're unlinking from gets fsync'd later on, there's
6335 * no way to find the destination directory later and fsync it
6336 * properly. So, we have to be conservative and force commits
6337 * so the new name gets discovered.
6338 */
6339 if (for_rename)
6340 goto record;
6341
6342 /* we can safely do the unlink without any special recording */
6343 return;
6344
6345record:
6346 mutex_lock(&dir->log_mutex);
6347 dir->last_unlink_trans = trans->transid;
6348 mutex_unlock(&dir->log_mutex);
6349}
6350
6351/*
6352 * Make sure that if someone attempts to fsync the parent directory of a deleted
6353 * snapshot, it ends up triggering a transaction commit. This is to guarantee
6354 * that after replaying the log tree of the parent directory's root we will not
6355 * see the snapshot anymore and at log replay time we will not see any log tree
6356 * corresponding to the deleted snapshot's root, which could lead to replaying
6357 * it after replaying the log tree of the parent directory (which would replay
6358 * the snapshot delete operation).
6359 *
6360 * Must be called before the actual snapshot destroy operation (updates to the
6361 * parent root and tree of tree roots trees, etc) are done.
6362 */
6363void btrfs_record_snapshot_destroy(struct btrfs_trans_handle *trans,
6364 struct btrfs_inode *dir)
6365{
6366 mutex_lock(&dir->log_mutex);
6367 dir->last_unlink_trans = trans->transid;
6368 mutex_unlock(&dir->log_mutex);
6369}
6370
6371/*
6372 * Call this after adding a new name for a file and it will properly
6373 * update the log to reflect the new name.
6374 *
6375 * @ctx can not be NULL when @sync_log is false, and should be NULL when it's
6376 * true (because it's not used).
6377 *
6378 * Return value depends on whether @sync_log is true or false.
6379 * When true: returns BTRFS_NEED_TRANS_COMMIT if the transaction needs to be
6380 * committed by the caller, and BTRFS_DONT_NEED_TRANS_COMMIT
6381 * otherwise.
6382 * When false: returns BTRFS_DONT_NEED_LOG_SYNC if the caller does not need to
6383 * to sync the log, BTRFS_NEED_LOG_SYNC if it needs to sync the log,
6384 * or BTRFS_NEED_TRANS_COMMIT if the transaction needs to be
6385 * committed (without attempting to sync the log).
6386 */
6387int btrfs_log_new_name(struct btrfs_trans_handle *trans,
6388 struct btrfs_inode *inode, struct btrfs_inode *old_dir,
6389 struct dentry *parent,
6390 bool sync_log, struct btrfs_log_ctx *ctx)
6391{
6392 struct btrfs_fs_info *fs_info = trans->fs_info;
6393 int ret;
6394
6395 /*
6396 * this will force the logging code to walk the dentry chain
6397 * up for the file
6398 */
6399 if (!S_ISDIR(inode->vfs_inode.i_mode))
6400 inode->last_unlink_trans = trans->transid;
6401
6402 /*
6403 * if this inode hasn't been logged and directory we're renaming it
6404 * from hasn't been logged, we don't need to log it
6405 */
6406 if (inode->logged_trans <= fs_info->last_trans_committed &&
6407 (!old_dir || old_dir->logged_trans <= fs_info->last_trans_committed))
6408 return sync_log ? BTRFS_DONT_NEED_TRANS_COMMIT :
6409 BTRFS_DONT_NEED_LOG_SYNC;
6410
6411 if (sync_log) {
6412 struct btrfs_log_ctx ctx2;
6413
6414 btrfs_init_log_ctx(&ctx2, &inode->vfs_inode);
6415 ret = btrfs_log_inode_parent(trans, inode, parent, 0, LLONG_MAX,
6416 LOG_INODE_EXISTS, &ctx2);
6417 if (ret == BTRFS_NO_LOG_SYNC)
6418 return BTRFS_DONT_NEED_TRANS_COMMIT;
6419 else if (ret)
6420 return BTRFS_NEED_TRANS_COMMIT;
6421
6422 ret = btrfs_sync_log(trans, inode->root, &ctx2);
6423 if (ret)
6424 return BTRFS_NEED_TRANS_COMMIT;
6425 return BTRFS_DONT_NEED_TRANS_COMMIT;
6426 }
6427
6428 ASSERT(ctx);
6429 ret = btrfs_log_inode_parent(trans, inode, parent, 0, LLONG_MAX,
6430 LOG_INODE_EXISTS, ctx);
6431 if (ret == BTRFS_NO_LOG_SYNC)
6432 return BTRFS_DONT_NEED_LOG_SYNC;
6433 else if (ret)
6434 return BTRFS_NEED_TRANS_COMMIT;
6435
6436 return BTRFS_NEED_LOG_SYNC;
6437}
6438