1 // SPDX-License-Identifier: GPL-2.0
4 * fs/ext4/fast_commit.c
6 * Written by Harshad Shirwadkar <harshadshirwadkar@gmail.com>
8 * Ext4 fast commits routines.
11 #include "ext4_jbd2.h"
12 #include "ext4_extents.h"
19 * Ext4 fast commits implement fine grained journalling for Ext4.
21 * Fast commits are organized as a log of tag-length-value (TLV) structs. (See
22 * struct ext4_fc_tl). Each TLV contains some delta that is replayed TLV by
23 * TLV during the recovery phase. For the scenarios for which we currently
24 * don't have replay code, fast commit falls back to full commits.
25 * Fast commits record delta in one of the following three categories.
27 * (A) Directory entry updates:
29 * - EXT4_FC_TAG_UNLINK - records directory entry unlink
30 * - EXT4_FC_TAG_LINK - records directory entry link
31 * - EXT4_FC_TAG_CREAT - records inode and directory entry creation
33 * (B) File specific data range updates:
35 * - EXT4_FC_TAG_ADD_RANGE - records addition of new blocks to an inode
36 * - EXT4_FC_TAG_DEL_RANGE - records deletion of blocks from an inode
38 * (C) Inode metadata (mtime / ctime etc):
40 * - EXT4_FC_TAG_INODE - record the inode that should be replayed
41 * during recovery. Note that iblocks field is
42 * not replayed and instead derived during
46 * With fast commits, we maintain all the directory entry operations in the
47 * order in which they are issued in an in-memory queue. This queue is flushed
48 * to disk during the commit operation. We also maintain a list of inodes
49 * that need to be committed during a fast commit in another in memory queue of
50 * inodes. During the commit operation, we commit in the following order:
52 * [1] Lock inodes for any further data updates by setting COMMITTING state
53 * [2] Submit data buffers of all the inodes
54 * [3] Wait for [2] to complete
55 * [4] Commit all the directory entry updates in the fast commit space
56 * [5] Commit all the changed inode structures
57 * [6] Write tail tag (this tag ensures the atomicity, please read the following
58 * section for more details).
59 * [7] Wait for [4], [5] and [6] to complete.
61 * All the inode updates must call ext4_fc_start_update() before starting an
62 * update. If such an ongoing update is present, fast commit waits for it to
63 * complete. The completion of such an update is marked by
64 * ext4_fc_stop_update().
66 * Fast Commit Ineligibility
67 * -------------------------
69 * Not all operations are supported by fast commits today (e.g extended
70 * attributes). Fast commit ineligibility is marked by calling
71 * ext4_fc_mark_ineligible(): This makes next fast commit operation to fall back
74 * Atomicity of commits
75 * --------------------
76 * In order to guarantee atomicity during the commit operation, fast commit
77 * uses "EXT4_FC_TAG_TAIL" tag that marks a fast commit as complete. Tail
78 * tag contains CRC of the contents and TID of the transaction after which
79 * this fast commit should be applied. Recovery code replays fast commit
80 * logs only if there's at least 1 valid tail present. For every fast commit
81 * operation, there is 1 tail. This means, we may end up with multiple tails
82 * in the fast commit space. Here's an example:
84 * - Create a new file A and remove existing file B
86 * - Append contents to file A
90 * The fast commit space at the end of above operations would look like this:
91 * [HEAD] [CREAT A] [UNLINK B] [TAIL] [ADD_RANGE A] [DEL_RANGE A] [TAIL]
92 * |<--- Fast Commit 1 --->|<--- Fast Commit 2 ---->|
94 * Replay code should thus check for all the valid tails in the FC area.
96 * Fast Commit Replay Idempotence
97 * ------------------------------
99 * Fast commits tags are idempotent in nature provided the recovery code follows
100 * certain rules. The guiding principle that the commit path follows while
101 * committing is that it stores the result of a particular operation instead of
102 * storing the procedure.
104 * Let's consider this rename operation: 'mv /a /b'. Let's assume dirent '/a'
105 * was associated with inode 10. During fast commit, instead of storing this
106 * operation as a procedure "rename a to b", we store the resulting file system
107 * state as a "series" of outcomes:
109 * - Link dirent b to inode 10
111 * - Inode <10> with valid refcount
113 * Now when recovery code runs, it needs "enforce" this state on the file
114 * system. This is what guarantees idempotence of fast commit replay.
116 * Let's take an example of a procedure that is not idempotent and see how fast
117 * commits make it idempotent. Consider following sequence of operations:
119 * rm A; mv B A; read A
122 * (x), (y) and (z) are the points at which we can crash. If we store this
123 * sequence of operations as is then the replay is not idempotent. Let's say
124 * while in replay, we crash at (z). During the second replay, file A (which was
125 * actually created as a result of "mv B A" operation) would get deleted. Thus,
126 * file named A would be absent when we try to read A. So, this sequence of
127 * operations is not idempotent. However, as mentioned above, instead of storing
128 * the procedure fast commits store the outcome of each procedure. Thus the fast
129 * commit log for above procedure would be as follows:
131 * (Let's assume dirent A was linked to inode 10 and dirent B was linked to
132 * inode 11 before the replay)
134 * [Unlink A] [Link A to inode 11] [Unlink B] [Inode 11]
137 * If we crash at (z), we will have file A linked to inode 11. During the second
138 * replay, we will remove file A (inode 11). But we will create it back and make
139 * it point to inode 11. We won't find B, so we'll just skip that step. At this
140 * point, the refcount for inode 11 is not reliable, but that gets fixed by the
141 * replay of last inode 11 tag. Crashes at points (w), (x) and (y) get handled
142 * similarly. Thus, by converting a non-idempotent procedure into a series of
143 * idempotent outcomes, fast commits ensured idempotence during the replay.
148 * 0) Fast commit replay path hardening: Fast commit replay code should use
149 * journal handles to make sure all the updates it does during the replay
150 * path are atomic. With that if we crash during fast commit replay, after
151 * trying to do recovery again, we will find a file system where fast commit
152 * area is invalid (because new full commit would be found). In order to deal
153 * with that, fast commit replay code should ensure that the "FC_REPLAY"
154 * superblock state is persisted before starting the replay, so that after
155 * the crash, fast commit recovery code can look at that flag and perform
156 * fast commit recovery even if that area is invalidated by later full
159 * 1) Fast commit's commit path locks the entire file system during fast
160 * commit. This has significant performance penalty. Instead of that, we
161 * should use ext4_fc_start/stop_update functions to start inode level
162 * updates from ext4_journal_start/stop. Once we do that we can drop file
163 * system locking during commit path.
165 * 2) Handle more ineligible cases.
168 #include <trace/events/ext4.h>
169 static struct kmem_cache *ext4_fc_dentry_cachep;
171 static void ext4_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
173 BUFFER_TRACE(bh, "");
175 ext4_debug("%s: Block %lld up-to-date",
176 __func__, bh->b_blocknr);
177 set_buffer_uptodate(bh);
179 ext4_debug("%s: Block %lld not up-to-date",
180 __func__, bh->b_blocknr);
181 clear_buffer_uptodate(bh);
187 static inline void ext4_fc_reset_inode(struct inode *inode)
189 struct ext4_inode_info *ei = EXT4_I(inode);
191 ei->i_fc_lblk_start = 0;
192 ei->i_fc_lblk_len = 0;
195 void ext4_fc_init_inode(struct inode *inode)
197 struct ext4_inode_info *ei = EXT4_I(inode);
199 ext4_fc_reset_inode(inode);
200 ext4_clear_inode_state(inode, EXT4_STATE_FC_COMMITTING);
201 INIT_LIST_HEAD(&ei->i_fc_list);
202 init_waitqueue_head(&ei->i_fc_wait);
203 atomic_set(&ei->i_fc_updates, 0);
206 /* This function must be called with sbi->s_fc_lock held. */
207 static void ext4_fc_wait_committing_inode(struct inode *inode)
208 __releases(&EXT4_SB(inode->i_sb)->s_fc_lock)
210 wait_queue_head_t *wq;
211 struct ext4_inode_info *ei = EXT4_I(inode);
213 #if (BITS_PER_LONG < 64)
214 DEFINE_WAIT_BIT(wait, &ei->i_state_flags,
215 EXT4_STATE_FC_COMMITTING);
216 wq = bit_waitqueue(&ei->i_state_flags,
217 EXT4_STATE_FC_COMMITTING);
219 DEFINE_WAIT_BIT(wait, &ei->i_flags,
220 EXT4_STATE_FC_COMMITTING);
221 wq = bit_waitqueue(&ei->i_flags,
222 EXT4_STATE_FC_COMMITTING);
224 lockdep_assert_held(&EXT4_SB(inode->i_sb)->s_fc_lock);
225 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
226 spin_unlock(&EXT4_SB(inode->i_sb)->s_fc_lock);
228 finish_wait(wq, &wait.wq_entry);
232 * Inform Ext4's fast about start of an inode update
234 * This function is called by the high level call VFS callbacks before
235 * performing any inode update. This function blocks if there's an ongoing
236 * fast commit on the inode in question.
238 void ext4_fc_start_update(struct inode *inode)
240 struct ext4_inode_info *ei = EXT4_I(inode);
242 if (!test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT) ||
243 (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY))
247 spin_lock(&EXT4_SB(inode->i_sb)->s_fc_lock);
248 if (list_empty(&ei->i_fc_list))
251 if (ext4_test_inode_state(inode, EXT4_STATE_FC_COMMITTING)) {
252 ext4_fc_wait_committing_inode(inode);
256 atomic_inc(&ei->i_fc_updates);
257 spin_unlock(&EXT4_SB(inode->i_sb)->s_fc_lock);
261 * Stop inode update and wake up waiting fast commits if any.
263 void ext4_fc_stop_update(struct inode *inode)
265 struct ext4_inode_info *ei = EXT4_I(inode);
267 if (!test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT) ||
268 (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY))
271 if (atomic_dec_and_test(&ei->i_fc_updates))
272 wake_up_all(&ei->i_fc_wait);
276 * Remove inode from fast commit list. If the inode is being committed
277 * we wait until inode commit is done.
279 void ext4_fc_del(struct inode *inode)
281 struct ext4_inode_info *ei = EXT4_I(inode);
283 if (!test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT) ||
284 (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY))
288 spin_lock(&EXT4_SB(inode->i_sb)->s_fc_lock);
289 if (list_empty(&ei->i_fc_list)) {
290 spin_unlock(&EXT4_SB(inode->i_sb)->s_fc_lock);
294 if (ext4_test_inode_state(inode, EXT4_STATE_FC_COMMITTING)) {
295 ext4_fc_wait_committing_inode(inode);
298 list_del_init(&ei->i_fc_list);
299 spin_unlock(&EXT4_SB(inode->i_sb)->s_fc_lock);
303 * Mark file system as fast commit ineligible. This means that next commit
304 * operation would result in a full jbd2 commit.
306 void ext4_fc_mark_ineligible(struct super_block *sb, int reason)
308 struct ext4_sb_info *sbi = EXT4_SB(sb);
310 if (!test_opt2(sb, JOURNAL_FAST_COMMIT) ||
311 (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY))
314 ext4_set_mount_flag(sb, EXT4_MF_FC_INELIGIBLE);
315 WARN_ON(reason >= EXT4_FC_REASON_MAX);
316 sbi->s_fc_stats.fc_ineligible_reason_count[reason]++;
320 * Generic fast commit tracking function. If this is the first time this we are
321 * called after a full commit, we initialize fast commit fields and then call
322 * __fc_track_fn() with update = 0. If we have already been called after a full
323 * commit, we pass update = 1. Based on that, the track function can determine
324 * if it needs to track a field for the first time or if it needs to just
325 * update the previously tracked value.
327 * If enqueue is set, this function enqueues the inode in fast commit list.
329 static int ext4_fc_track_template(
330 handle_t *handle, struct inode *inode,
331 int (*__fc_track_fn)(struct inode *, void *, bool),
332 void *args, int enqueue)
335 struct ext4_inode_info *ei = EXT4_I(inode);
336 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
340 if (!test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT) ||
341 (sbi->s_mount_state & EXT4_FC_REPLAY))
344 if (ext4_test_mount_flag(inode->i_sb, EXT4_MF_FC_INELIGIBLE))
347 tid = handle->h_transaction->t_tid;
348 mutex_lock(&ei->i_fc_lock);
349 if (tid == ei->i_sync_tid) {
352 ext4_fc_reset_inode(inode);
353 ei->i_sync_tid = tid;
355 ret = __fc_track_fn(inode, args, update);
356 mutex_unlock(&ei->i_fc_lock);
361 spin_lock(&sbi->s_fc_lock);
362 if (list_empty(&EXT4_I(inode)->i_fc_list))
363 list_add_tail(&EXT4_I(inode)->i_fc_list,
364 (ext4_test_mount_flag(inode->i_sb, EXT4_MF_FC_COMMITTING)) ?
365 &sbi->s_fc_q[FC_Q_STAGING] :
366 &sbi->s_fc_q[FC_Q_MAIN]);
367 spin_unlock(&sbi->s_fc_lock);
372 struct __track_dentry_update_args {
373 struct dentry *dentry;
377 /* __track_fn for directory entry updates. Called with ei->i_fc_lock. */
378 static int __track_dentry_update(struct inode *inode, void *arg, bool update)
380 struct ext4_fc_dentry_update *node;
381 struct ext4_inode_info *ei = EXT4_I(inode);
382 struct __track_dentry_update_args *dentry_update =
383 (struct __track_dentry_update_args *)arg;
384 struct dentry *dentry = dentry_update->dentry;
385 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
387 mutex_unlock(&ei->i_fc_lock);
388 node = kmem_cache_alloc(ext4_fc_dentry_cachep, GFP_NOFS);
390 ext4_fc_mark_ineligible(inode->i_sb, EXT4_FC_REASON_NOMEM);
391 mutex_lock(&ei->i_fc_lock);
395 node->fcd_op = dentry_update->op;
396 node->fcd_parent = dentry->d_parent->d_inode->i_ino;
397 node->fcd_ino = inode->i_ino;
398 if (dentry->d_name.len > DNAME_INLINE_LEN) {
399 node->fcd_name.name = kmalloc(dentry->d_name.len, GFP_NOFS);
400 if (!node->fcd_name.name) {
401 kmem_cache_free(ext4_fc_dentry_cachep, node);
402 ext4_fc_mark_ineligible(inode->i_sb,
403 EXT4_FC_REASON_NOMEM);
404 mutex_lock(&ei->i_fc_lock);
407 memcpy((u8 *)node->fcd_name.name, dentry->d_name.name,
410 memcpy(node->fcd_iname, dentry->d_name.name,
412 node->fcd_name.name = node->fcd_iname;
414 node->fcd_name.len = dentry->d_name.len;
416 spin_lock(&sbi->s_fc_lock);
417 if (ext4_test_mount_flag(inode->i_sb, EXT4_MF_FC_COMMITTING))
418 list_add_tail(&node->fcd_list,
419 &sbi->s_fc_dentry_q[FC_Q_STAGING]);
421 list_add_tail(&node->fcd_list, &sbi->s_fc_dentry_q[FC_Q_MAIN]);
422 spin_unlock(&sbi->s_fc_lock);
423 mutex_lock(&ei->i_fc_lock);
428 void __ext4_fc_track_unlink(handle_t *handle,
429 struct inode *inode, struct dentry *dentry)
431 struct __track_dentry_update_args args;
434 args.dentry = dentry;
435 args.op = EXT4_FC_TAG_UNLINK;
437 ret = ext4_fc_track_template(handle, inode, __track_dentry_update,
439 trace_ext4_fc_track_unlink(inode, dentry, ret);
442 void ext4_fc_track_unlink(handle_t *handle, struct dentry *dentry)
444 __ext4_fc_track_unlink(handle, d_inode(dentry), dentry);
447 void __ext4_fc_track_link(handle_t *handle,
448 struct inode *inode, struct dentry *dentry)
450 struct __track_dentry_update_args args;
453 args.dentry = dentry;
454 args.op = EXT4_FC_TAG_LINK;
456 ret = ext4_fc_track_template(handle, inode, __track_dentry_update,
458 trace_ext4_fc_track_link(inode, dentry, ret);
461 void ext4_fc_track_link(handle_t *handle, struct dentry *dentry)
463 __ext4_fc_track_link(handle, d_inode(dentry), dentry);
466 void __ext4_fc_track_create(handle_t *handle, struct inode *inode,
467 struct dentry *dentry)
469 struct __track_dentry_update_args args;
472 args.dentry = dentry;
473 args.op = EXT4_FC_TAG_CREAT;
475 ret = ext4_fc_track_template(handle, inode, __track_dentry_update,
477 trace_ext4_fc_track_create(inode, dentry, ret);
480 void ext4_fc_track_create(handle_t *handle, struct dentry *dentry)
482 __ext4_fc_track_create(handle, d_inode(dentry), dentry);
485 /* __track_fn for inode tracking */
486 static int __track_inode(struct inode *inode, void *arg, bool update)
491 EXT4_I(inode)->i_fc_lblk_len = 0;
496 void ext4_fc_track_inode(handle_t *handle, struct inode *inode)
500 if (S_ISDIR(inode->i_mode))
503 if (ext4_should_journal_data(inode)) {
504 ext4_fc_mark_ineligible(inode->i_sb,
505 EXT4_FC_REASON_INODE_JOURNAL_DATA);
509 ret = ext4_fc_track_template(handle, inode, __track_inode, NULL, 1);
510 trace_ext4_fc_track_inode(inode, ret);
513 struct __track_range_args {
514 ext4_lblk_t start, end;
517 /* __track_fn for tracking data updates */
518 static int __track_range(struct inode *inode, void *arg, bool update)
520 struct ext4_inode_info *ei = EXT4_I(inode);
521 ext4_lblk_t oldstart;
522 struct __track_range_args *__arg =
523 (struct __track_range_args *)arg;
525 if (inode->i_ino < EXT4_FIRST_INO(inode->i_sb)) {
526 ext4_debug("Special inode %ld being modified\n", inode->i_ino);
530 oldstart = ei->i_fc_lblk_start;
532 if (update && ei->i_fc_lblk_len > 0) {
533 ei->i_fc_lblk_start = min(ei->i_fc_lblk_start, __arg->start);
535 max(oldstart + ei->i_fc_lblk_len - 1, __arg->end) -
536 ei->i_fc_lblk_start + 1;
538 ei->i_fc_lblk_start = __arg->start;
539 ei->i_fc_lblk_len = __arg->end - __arg->start + 1;
545 void ext4_fc_track_range(handle_t *handle, struct inode *inode, ext4_lblk_t start,
548 struct __track_range_args args;
551 if (S_ISDIR(inode->i_mode))
557 ret = ext4_fc_track_template(handle, inode, __track_range, &args, 1);
559 trace_ext4_fc_track_range(inode, start, end, ret);
562 static void ext4_fc_submit_bh(struct super_block *sb, bool is_tail)
564 int write_flags = REQ_SYNC;
565 struct buffer_head *bh = EXT4_SB(sb)->s_fc_bh;
567 /* Add REQ_FUA | REQ_PREFLUSH only its tail */
568 if (test_opt(sb, BARRIER) && is_tail)
569 write_flags |= REQ_FUA | REQ_PREFLUSH;
571 set_buffer_dirty(bh);
572 set_buffer_uptodate(bh);
573 bh->b_end_io = ext4_end_buffer_io_sync;
574 submit_bh(REQ_OP_WRITE, write_flags, bh);
575 EXT4_SB(sb)->s_fc_bh = NULL;
578 /* Ext4 commit path routines */
580 /* memzero and update CRC */
581 static void *ext4_fc_memzero(struct super_block *sb, void *dst, int len,
586 ret = memset(dst, 0, len);
588 *crc = ext4_chksum(EXT4_SB(sb), *crc, dst, len);
593 * Allocate len bytes on a fast commit buffer.
595 * During the commit time this function is used to manage fast commit
596 * block space. We don't split a fast commit log onto different
597 * blocks. So this function makes sure that if there's not enough space
598 * on the current block, the remaining space in the current block is
599 * marked as unused by adding EXT4_FC_TAG_PAD tag. In that case,
600 * new block is from jbd2 and CRC is updated to reflect the padding
603 static u8 *ext4_fc_reserve_space(struct super_block *sb, int len, u32 *crc)
605 struct ext4_fc_tl *tl;
606 struct ext4_sb_info *sbi = EXT4_SB(sb);
607 struct buffer_head *bh;
608 int bsize = sbi->s_journal->j_blocksize;
609 int ret, off = sbi->s_fc_bytes % bsize;
613 * After allocating len, we should have space at least for a 0 byte
616 if (len + sizeof(struct ext4_fc_tl) > bsize)
619 if (bsize - off - 1 > len + sizeof(struct ext4_fc_tl)) {
621 * Only allocate from current buffer if we have enough space for
622 * this request AND we have space to add a zero byte padding.
625 ret = jbd2_fc_get_buf(EXT4_SB(sb)->s_journal, &bh);
630 sbi->s_fc_bytes += len;
631 return sbi->s_fc_bh->b_data + off;
633 /* Need to add PAD tag */
634 tl = (struct ext4_fc_tl *)(sbi->s_fc_bh->b_data + off);
635 tl->fc_tag = cpu_to_le16(EXT4_FC_TAG_PAD);
636 pad_len = bsize - off - 1 - sizeof(struct ext4_fc_tl);
637 tl->fc_len = cpu_to_le16(pad_len);
639 *crc = ext4_chksum(sbi, *crc, tl, sizeof(*tl));
641 ext4_fc_memzero(sb, tl + 1, pad_len, crc);
642 ext4_fc_submit_bh(sb, false);
644 ret = jbd2_fc_get_buf(EXT4_SB(sb)->s_journal, &bh);
648 sbi->s_fc_bytes = (sbi->s_fc_bytes / bsize + 1) * bsize + len;
649 return sbi->s_fc_bh->b_data;
652 /* memcpy to fc reserved space and update CRC */
653 static void *ext4_fc_memcpy(struct super_block *sb, void *dst, const void *src,
657 *crc = ext4_chksum(EXT4_SB(sb), *crc, src, len);
658 return memcpy(dst, src, len);
662 * Complete a fast commit by writing tail tag.
664 * Writing tail tag marks the end of a fast commit. In order to guarantee
665 * atomicity, after writing tail tag, even if there's space remaining
666 * in the block, next commit shouldn't use it. That's why tail tag
667 * has the length as that of the remaining space on the block.
669 static int ext4_fc_write_tail(struct super_block *sb, u32 crc)
671 struct ext4_sb_info *sbi = EXT4_SB(sb);
672 struct ext4_fc_tl tl;
673 struct ext4_fc_tail tail;
674 int off, bsize = sbi->s_journal->j_blocksize;
678 * ext4_fc_reserve_space takes care of allocating an extra block if
679 * there's no enough space on this block for accommodating this tail.
681 dst = ext4_fc_reserve_space(sb, sizeof(tl) + sizeof(tail), &crc);
685 off = sbi->s_fc_bytes % bsize;
687 tl.fc_tag = cpu_to_le16(EXT4_FC_TAG_TAIL);
688 tl.fc_len = cpu_to_le16(bsize - off - 1 + sizeof(struct ext4_fc_tail));
689 sbi->s_fc_bytes = round_up(sbi->s_fc_bytes, bsize);
691 ext4_fc_memcpy(sb, dst, &tl, sizeof(tl), &crc);
693 tail.fc_tid = cpu_to_le32(sbi->s_journal->j_running_transaction->t_tid);
694 ext4_fc_memcpy(sb, dst, &tail.fc_tid, sizeof(tail.fc_tid), &crc);
695 dst += sizeof(tail.fc_tid);
696 tail.fc_crc = cpu_to_le32(crc);
697 ext4_fc_memcpy(sb, dst, &tail.fc_crc, sizeof(tail.fc_crc), NULL);
699 ext4_fc_submit_bh(sb, true);
705 * Adds tag, length, value and updates CRC. Returns true if tlv was added.
706 * Returns false if there's not enough space.
708 static bool ext4_fc_add_tlv(struct super_block *sb, u16 tag, u16 len, u8 *val,
711 struct ext4_fc_tl tl;
714 dst = ext4_fc_reserve_space(sb, sizeof(tl) + len, crc);
718 tl.fc_tag = cpu_to_le16(tag);
719 tl.fc_len = cpu_to_le16(len);
721 ext4_fc_memcpy(sb, dst, &tl, sizeof(tl), crc);
722 ext4_fc_memcpy(sb, dst + sizeof(tl), val, len, crc);
727 /* Same as above, but adds dentry tlv. */
728 static bool ext4_fc_add_dentry_tlv(struct super_block *sb, u32 *crc,
729 struct ext4_fc_dentry_update *fc_dentry)
731 struct ext4_fc_dentry_info fcd;
732 struct ext4_fc_tl tl;
733 int dlen = fc_dentry->fcd_name.len;
734 u8 *dst = ext4_fc_reserve_space(sb, sizeof(tl) + sizeof(fcd) + dlen,
740 fcd.fc_parent_ino = cpu_to_le32(fc_dentry->fcd_parent);
741 fcd.fc_ino = cpu_to_le32(fc_dentry->fcd_ino);
742 tl.fc_tag = cpu_to_le16(fc_dentry->fcd_op);
743 tl.fc_len = cpu_to_le16(sizeof(fcd) + dlen);
744 ext4_fc_memcpy(sb, dst, &tl, sizeof(tl), crc);
746 ext4_fc_memcpy(sb, dst, &fcd, sizeof(fcd), crc);
748 ext4_fc_memcpy(sb, dst, fc_dentry->fcd_name.name, dlen, crc);
754 * Writes inode in the fast commit space under TLV with tag @tag.
755 * Returns 0 on success, error on failure.
757 static int ext4_fc_write_inode(struct inode *inode, u32 *crc)
759 struct ext4_inode_info *ei = EXT4_I(inode);
760 int inode_len = EXT4_GOOD_OLD_INODE_SIZE;
762 struct ext4_iloc iloc;
763 struct ext4_fc_inode fc_inode;
764 struct ext4_fc_tl tl;
767 ret = ext4_get_inode_loc(inode, &iloc);
771 if (ext4_test_inode_flag(inode, EXT4_INODE_INLINE_DATA))
772 inode_len = EXT4_INODE_SIZE(inode->i_sb);
773 else if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE)
774 inode_len += ei->i_extra_isize;
776 fc_inode.fc_ino = cpu_to_le32(inode->i_ino);
777 tl.fc_tag = cpu_to_le16(EXT4_FC_TAG_INODE);
778 tl.fc_len = cpu_to_le16(inode_len + sizeof(fc_inode.fc_ino));
780 dst = ext4_fc_reserve_space(inode->i_sb,
781 sizeof(tl) + inode_len + sizeof(fc_inode.fc_ino), crc);
785 if (!ext4_fc_memcpy(inode->i_sb, dst, &tl, sizeof(tl), crc))
788 if (!ext4_fc_memcpy(inode->i_sb, dst, &fc_inode, sizeof(fc_inode), crc))
790 dst += sizeof(fc_inode);
791 if (!ext4_fc_memcpy(inode->i_sb, dst, (u8 *)ext4_raw_inode(&iloc),
799 * Writes updated data ranges for the inode in question. Updates CRC.
800 * Returns 0 on success, error otherwise.
802 static int ext4_fc_write_inode_data(struct inode *inode, u32 *crc)
804 ext4_lblk_t old_blk_size, cur_lblk_off, new_blk_size;
805 struct ext4_inode_info *ei = EXT4_I(inode);
806 struct ext4_map_blocks map;
807 struct ext4_fc_add_range fc_ext;
808 struct ext4_fc_del_range lrange;
809 struct ext4_extent *ex;
812 mutex_lock(&ei->i_fc_lock);
813 if (ei->i_fc_lblk_len == 0) {
814 mutex_unlock(&ei->i_fc_lock);
817 old_blk_size = ei->i_fc_lblk_start;
818 new_blk_size = ei->i_fc_lblk_start + ei->i_fc_lblk_len - 1;
819 ei->i_fc_lblk_len = 0;
820 mutex_unlock(&ei->i_fc_lock);
822 cur_lblk_off = old_blk_size;
823 jbd_debug(1, "%s: will try writing %d to %d for inode %ld\n",
824 __func__, cur_lblk_off, new_blk_size, inode->i_ino);
826 while (cur_lblk_off <= new_blk_size) {
827 map.m_lblk = cur_lblk_off;
828 map.m_len = new_blk_size - cur_lblk_off + 1;
829 ret = ext4_map_blocks(NULL, inode, &map, 0);
833 if (map.m_len == 0) {
839 lrange.fc_ino = cpu_to_le32(inode->i_ino);
840 lrange.fc_lblk = cpu_to_le32(map.m_lblk);
841 lrange.fc_len = cpu_to_le32(map.m_len);
842 if (!ext4_fc_add_tlv(inode->i_sb, EXT4_FC_TAG_DEL_RANGE,
843 sizeof(lrange), (u8 *)&lrange, crc))
846 unsigned int max = (map.m_flags & EXT4_MAP_UNWRITTEN) ?
847 EXT_UNWRITTEN_MAX_LEN : EXT_INIT_MAX_LEN;
849 /* Limit the number of blocks in one extent */
850 map.m_len = min(max, map.m_len);
852 fc_ext.fc_ino = cpu_to_le32(inode->i_ino);
853 ex = (struct ext4_extent *)&fc_ext.fc_ex;
854 ex->ee_block = cpu_to_le32(map.m_lblk);
855 ex->ee_len = cpu_to_le16(map.m_len);
856 ext4_ext_store_pblock(ex, map.m_pblk);
857 if (map.m_flags & EXT4_MAP_UNWRITTEN)
858 ext4_ext_mark_unwritten(ex);
860 ext4_ext_mark_initialized(ex);
861 if (!ext4_fc_add_tlv(inode->i_sb, EXT4_FC_TAG_ADD_RANGE,
862 sizeof(fc_ext), (u8 *)&fc_ext, crc))
866 cur_lblk_off += map.m_len;
873 /* Submit data for all the fast commit inodes */
874 static int ext4_fc_submit_inode_data_all(journal_t *journal)
876 struct super_block *sb = (struct super_block *)(journal->j_private);
877 struct ext4_sb_info *sbi = EXT4_SB(sb);
878 struct ext4_inode_info *ei;
881 spin_lock(&sbi->s_fc_lock);
882 ext4_set_mount_flag(sb, EXT4_MF_FC_COMMITTING);
883 list_for_each_entry(ei, &sbi->s_fc_q[FC_Q_MAIN], i_fc_list) {
884 ext4_set_inode_state(&ei->vfs_inode, EXT4_STATE_FC_COMMITTING);
885 while (atomic_read(&ei->i_fc_updates)) {
888 prepare_to_wait(&ei->i_fc_wait, &wait,
889 TASK_UNINTERRUPTIBLE);
890 if (atomic_read(&ei->i_fc_updates)) {
891 spin_unlock(&sbi->s_fc_lock);
893 spin_lock(&sbi->s_fc_lock);
895 finish_wait(&ei->i_fc_wait, &wait);
897 spin_unlock(&sbi->s_fc_lock);
898 ret = jbd2_submit_inode_data(ei->jinode);
901 spin_lock(&sbi->s_fc_lock);
903 spin_unlock(&sbi->s_fc_lock);
908 /* Wait for completion of data for all the fast commit inodes */
909 static int ext4_fc_wait_inode_data_all(journal_t *journal)
911 struct super_block *sb = (struct super_block *)(journal->j_private);
912 struct ext4_sb_info *sbi = EXT4_SB(sb);
913 struct ext4_inode_info *pos, *n;
916 spin_lock(&sbi->s_fc_lock);
917 list_for_each_entry_safe(pos, n, &sbi->s_fc_q[FC_Q_MAIN], i_fc_list) {
918 if (!ext4_test_inode_state(&pos->vfs_inode,
919 EXT4_STATE_FC_COMMITTING))
921 spin_unlock(&sbi->s_fc_lock);
923 ret = jbd2_wait_inode_data(journal, pos->jinode);
926 spin_lock(&sbi->s_fc_lock);
928 spin_unlock(&sbi->s_fc_lock);
933 /* Commit all the directory entry updates */
934 static int ext4_fc_commit_dentry_updates(journal_t *journal, u32 *crc)
935 __acquires(&sbi->s_fc_lock)
936 __releases(&sbi->s_fc_lock)
938 struct super_block *sb = (struct super_block *)(journal->j_private);
939 struct ext4_sb_info *sbi = EXT4_SB(sb);
940 struct ext4_fc_dentry_update *fc_dentry, *fc_dentry_n;
942 struct ext4_inode_info *ei, *ei_n;
945 if (list_empty(&sbi->s_fc_dentry_q[FC_Q_MAIN]))
947 list_for_each_entry_safe(fc_dentry, fc_dentry_n,
948 &sbi->s_fc_dentry_q[FC_Q_MAIN], fcd_list) {
949 if (fc_dentry->fcd_op != EXT4_FC_TAG_CREAT) {
950 spin_unlock(&sbi->s_fc_lock);
951 if (!ext4_fc_add_dentry_tlv(sb, crc, fc_dentry)) {
955 spin_lock(&sbi->s_fc_lock);
960 list_for_each_entry_safe(ei, ei_n, &sbi->s_fc_q[FC_Q_MAIN],
962 if (ei->vfs_inode.i_ino == fc_dentry->fcd_ino) {
963 inode = &ei->vfs_inode;
968 * If we don't find inode in our list, then it was deleted,
969 * in which case, we don't need to record it's create tag.
973 spin_unlock(&sbi->s_fc_lock);
976 * We first write the inode and then the create dirent. This
977 * allows the recovery code to create an unnamed inode first
978 * and then link it to a directory entry. This allows us
979 * to use namei.c routines almost as is and simplifies
982 ret = ext4_fc_write_inode(inode, crc);
986 ret = ext4_fc_write_inode_data(inode, crc);
990 if (!ext4_fc_add_dentry_tlv(sb, crc, fc_dentry)) {
995 spin_lock(&sbi->s_fc_lock);
999 spin_lock(&sbi->s_fc_lock);
1003 static int ext4_fc_perform_commit(journal_t *journal)
1005 struct super_block *sb = (struct super_block *)(journal->j_private);
1006 struct ext4_sb_info *sbi = EXT4_SB(sb);
1007 struct ext4_inode_info *iter;
1008 struct ext4_fc_head head;
1009 struct inode *inode;
1010 struct blk_plug plug;
1014 ret = ext4_fc_submit_inode_data_all(journal);
1018 ret = ext4_fc_wait_inode_data_all(journal);
1023 * If file system device is different from journal device, issue a cache
1024 * flush before we start writing fast commit blocks.
1026 if (journal->j_fs_dev != journal->j_dev)
1027 blkdev_issue_flush(journal->j_fs_dev);
1029 blk_start_plug(&plug);
1030 if (sbi->s_fc_bytes == 0) {
1032 * Add a head tag only if this is the first fast commit
1035 head.fc_features = cpu_to_le32(EXT4_FC_SUPPORTED_FEATURES);
1036 head.fc_tid = cpu_to_le32(
1037 sbi->s_journal->j_running_transaction->t_tid);
1038 if (!ext4_fc_add_tlv(sb, EXT4_FC_TAG_HEAD, sizeof(head),
1039 (u8 *)&head, &crc)) {
1045 spin_lock(&sbi->s_fc_lock);
1046 ret = ext4_fc_commit_dentry_updates(journal, &crc);
1048 spin_unlock(&sbi->s_fc_lock);
1052 list_for_each_entry(iter, &sbi->s_fc_q[FC_Q_MAIN], i_fc_list) {
1053 inode = &iter->vfs_inode;
1054 if (!ext4_test_inode_state(inode, EXT4_STATE_FC_COMMITTING))
1057 spin_unlock(&sbi->s_fc_lock);
1058 ret = ext4_fc_write_inode_data(inode, &crc);
1061 ret = ext4_fc_write_inode(inode, &crc);
1064 spin_lock(&sbi->s_fc_lock);
1066 spin_unlock(&sbi->s_fc_lock);
1068 ret = ext4_fc_write_tail(sb, crc);
1071 blk_finish_plug(&plug);
1075 static void ext4_fc_update_stats(struct super_block *sb, int status,
1076 u64 commit_time, int nblks)
1078 struct ext4_fc_stats *stats = &EXT4_SB(sb)->s_fc_stats;
1080 jbd_debug(1, "Fast commit ended with status = %d", status);
1081 if (status == EXT4_FC_STATUS_OK) {
1082 stats->fc_num_commits++;
1083 stats->fc_numblks += nblks;
1084 if (likely(stats->s_fc_avg_commit_time))
1085 stats->s_fc_avg_commit_time =
1087 stats->s_fc_avg_commit_time * 3) / 4;
1089 stats->s_fc_avg_commit_time = commit_time;
1090 } else if (status == EXT4_FC_STATUS_FAILED ||
1091 status == EXT4_FC_STATUS_INELIGIBLE) {
1092 if (status == EXT4_FC_STATUS_FAILED)
1093 stats->fc_failed_commits++;
1094 stats->fc_ineligible_commits++;
1096 stats->fc_skipped_commits++;
1098 trace_ext4_fc_commit_stop(sb, nblks, status);
1102 * The main commit entry point. Performs a fast commit for transaction
1103 * commit_tid if needed. If it's not possible to perform a fast commit
1104 * due to various reasons, we fall back to full commit. Returns 0
1105 * on success, error otherwise.
1107 int ext4_fc_commit(journal_t *journal, tid_t commit_tid)
1109 struct super_block *sb = (struct super_block *)(journal->j_private);
1110 struct ext4_sb_info *sbi = EXT4_SB(sb);
1111 int nblks = 0, ret, bsize = journal->j_blocksize;
1112 int subtid = atomic_read(&sbi->s_fc_subtid);
1113 int status = EXT4_FC_STATUS_OK, fc_bufs_before = 0;
1114 ktime_t start_time, commit_time;
1116 trace_ext4_fc_commit_start(sb);
1118 start_time = ktime_get();
1120 if (!test_opt2(sb, JOURNAL_FAST_COMMIT))
1121 return jbd2_complete_transaction(journal, commit_tid);
1124 ret = jbd2_fc_begin_commit(journal, commit_tid);
1125 if (ret == -EALREADY) {
1126 /* There was an ongoing commit, check if we need to restart */
1127 if (atomic_read(&sbi->s_fc_subtid) <= subtid &&
1128 commit_tid > journal->j_commit_sequence)
1130 ext4_fc_update_stats(sb, EXT4_FC_STATUS_SKIPPED, 0, 0);
1134 * Commit couldn't start. Just update stats and perform a
1137 ext4_fc_update_stats(sb, EXT4_FC_STATUS_FAILED, 0, 0);
1138 return jbd2_complete_transaction(journal, commit_tid);
1142 * After establishing journal barrier via jbd2_fc_begin_commit(), check
1143 * if we are fast commit ineligible.
1145 if (ext4_test_mount_flag(sb, EXT4_MF_FC_INELIGIBLE)) {
1146 status = EXT4_FC_STATUS_INELIGIBLE;
1150 fc_bufs_before = (sbi->s_fc_bytes + bsize - 1) / bsize;
1151 ret = ext4_fc_perform_commit(journal);
1153 status = EXT4_FC_STATUS_FAILED;
1156 nblks = (sbi->s_fc_bytes + bsize - 1) / bsize - fc_bufs_before;
1157 ret = jbd2_fc_wait_bufs(journal, nblks);
1159 status = EXT4_FC_STATUS_FAILED;
1162 atomic_inc(&sbi->s_fc_subtid);
1163 ret = jbd2_fc_end_commit(journal);
1165 * weight the commit time higher than the average time so we
1166 * don't react too strongly to vast changes in the commit time
1168 commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1169 ext4_fc_update_stats(sb, status, commit_time, nblks);
1173 ret = jbd2_fc_end_commit_fallback(journal);
1174 ext4_fc_update_stats(sb, status, 0, 0);
1179 * Fast commit cleanup routine. This is called after every fast commit and
1180 * full commit. full is true if we are called after a full commit.
1182 static void ext4_fc_cleanup(journal_t *journal, int full)
1184 struct super_block *sb = journal->j_private;
1185 struct ext4_sb_info *sbi = EXT4_SB(sb);
1186 struct ext4_inode_info *iter, *iter_n;
1187 struct ext4_fc_dentry_update *fc_dentry;
1189 if (full && sbi->s_fc_bh)
1190 sbi->s_fc_bh = NULL;
1192 jbd2_fc_release_bufs(journal);
1194 spin_lock(&sbi->s_fc_lock);
1195 list_for_each_entry_safe(iter, iter_n, &sbi->s_fc_q[FC_Q_MAIN],
1197 list_del_init(&iter->i_fc_list);
1198 ext4_clear_inode_state(&iter->vfs_inode,
1199 EXT4_STATE_FC_COMMITTING);
1200 ext4_fc_reset_inode(&iter->vfs_inode);
1201 /* Make sure EXT4_STATE_FC_COMMITTING bit is clear */
1203 #if (BITS_PER_LONG < 64)
1204 wake_up_bit(&iter->i_state_flags, EXT4_STATE_FC_COMMITTING);
1206 wake_up_bit(&iter->i_flags, EXT4_STATE_FC_COMMITTING);
1210 while (!list_empty(&sbi->s_fc_dentry_q[FC_Q_MAIN])) {
1211 fc_dentry = list_first_entry(&sbi->s_fc_dentry_q[FC_Q_MAIN],
1212 struct ext4_fc_dentry_update,
1214 list_del_init(&fc_dentry->fcd_list);
1215 spin_unlock(&sbi->s_fc_lock);
1217 if (fc_dentry->fcd_name.name &&
1218 fc_dentry->fcd_name.len > DNAME_INLINE_LEN)
1219 kfree(fc_dentry->fcd_name.name);
1220 kmem_cache_free(ext4_fc_dentry_cachep, fc_dentry);
1221 spin_lock(&sbi->s_fc_lock);
1224 list_splice_init(&sbi->s_fc_dentry_q[FC_Q_STAGING],
1225 &sbi->s_fc_dentry_q[FC_Q_MAIN]);
1226 list_splice_init(&sbi->s_fc_q[FC_Q_STAGING],
1227 &sbi->s_fc_q[FC_Q_MAIN]);
1229 ext4_clear_mount_flag(sb, EXT4_MF_FC_COMMITTING);
1230 ext4_clear_mount_flag(sb, EXT4_MF_FC_INELIGIBLE);
1233 sbi->s_fc_bytes = 0;
1234 spin_unlock(&sbi->s_fc_lock);
1235 trace_ext4_fc_stats(sb);
1238 /* Ext4 Replay Path Routines */
1240 /* Helper struct for dentry replay routines */
1241 struct dentry_info_args {
1242 int parent_ino, dname_len, ino, inode_len;
1246 static inline void tl_to_darg(struct dentry_info_args *darg,
1247 struct ext4_fc_tl *tl, u8 *val)
1249 struct ext4_fc_dentry_info fcd;
1251 memcpy(&fcd, val, sizeof(fcd));
1253 darg->parent_ino = le32_to_cpu(fcd.fc_parent_ino);
1254 darg->ino = le32_to_cpu(fcd.fc_ino);
1255 darg->dname = val + offsetof(struct ext4_fc_dentry_info, fc_dname);
1256 darg->dname_len = le16_to_cpu(tl->fc_len) -
1257 sizeof(struct ext4_fc_dentry_info);
1260 /* Unlink replay function */
1261 static int ext4_fc_replay_unlink(struct super_block *sb, struct ext4_fc_tl *tl,
1264 struct inode *inode, *old_parent;
1266 struct dentry_info_args darg;
1269 tl_to_darg(&darg, tl, val);
1271 trace_ext4_fc_replay(sb, EXT4_FC_TAG_UNLINK, darg.ino,
1272 darg.parent_ino, darg.dname_len);
1274 entry.name = darg.dname;
1275 entry.len = darg.dname_len;
1276 inode = ext4_iget(sb, darg.ino, EXT4_IGET_NORMAL);
1278 if (IS_ERR(inode)) {
1279 jbd_debug(1, "Inode %d not found", darg.ino);
1283 old_parent = ext4_iget(sb, darg.parent_ino,
1285 if (IS_ERR(old_parent)) {
1286 jbd_debug(1, "Dir with inode %d not found", darg.parent_ino);
1291 ret = __ext4_unlink(NULL, old_parent, &entry, inode);
1292 /* -ENOENT ok coz it might not exist anymore. */
1300 static int ext4_fc_replay_link_internal(struct super_block *sb,
1301 struct dentry_info_args *darg,
1302 struct inode *inode)
1304 struct inode *dir = NULL;
1305 struct dentry *dentry_dir = NULL, *dentry_inode = NULL;
1306 struct qstr qstr_dname = QSTR_INIT(darg->dname, darg->dname_len);
1309 dir = ext4_iget(sb, darg->parent_ino, EXT4_IGET_NORMAL);
1311 jbd_debug(1, "Dir with inode %d not found.", darg->parent_ino);
1316 dentry_dir = d_obtain_alias(dir);
1317 if (IS_ERR(dentry_dir)) {
1318 jbd_debug(1, "Failed to obtain dentry");
1323 dentry_inode = d_alloc(dentry_dir, &qstr_dname);
1324 if (!dentry_inode) {
1325 jbd_debug(1, "Inode dentry not created.");
1330 ret = __ext4_link(dir, inode, dentry_inode);
1332 * It's possible that link already existed since data blocks
1333 * for the dir in question got persisted before we crashed OR
1334 * we replayed this tag and crashed before the entire replay
1337 if (ret && ret != -EEXIST) {
1338 jbd_debug(1, "Failed to link\n");
1351 d_drop(dentry_inode);
1358 /* Link replay function */
1359 static int ext4_fc_replay_link(struct super_block *sb, struct ext4_fc_tl *tl,
1362 struct inode *inode;
1363 struct dentry_info_args darg;
1366 tl_to_darg(&darg, tl, val);
1367 trace_ext4_fc_replay(sb, EXT4_FC_TAG_LINK, darg.ino,
1368 darg.parent_ino, darg.dname_len);
1370 inode = ext4_iget(sb, darg.ino, EXT4_IGET_NORMAL);
1371 if (IS_ERR(inode)) {
1372 jbd_debug(1, "Inode not found.");
1376 ret = ext4_fc_replay_link_internal(sb, &darg, inode);
1382 * Record all the modified inodes during replay. We use this later to setup
1383 * block bitmaps correctly.
1385 static int ext4_fc_record_modified_inode(struct super_block *sb, int ino)
1387 struct ext4_fc_replay_state *state;
1390 state = &EXT4_SB(sb)->s_fc_replay_state;
1391 for (i = 0; i < state->fc_modified_inodes_used; i++)
1392 if (state->fc_modified_inodes[i] == ino)
1394 if (state->fc_modified_inodes_used == state->fc_modified_inodes_size) {
1395 state->fc_modified_inodes_size +=
1396 EXT4_FC_REPLAY_REALLOC_INCREMENT;
1397 state->fc_modified_inodes = krealloc(
1398 state->fc_modified_inodes, sizeof(int) *
1399 state->fc_modified_inodes_size,
1401 if (!state->fc_modified_inodes)
1404 state->fc_modified_inodes[state->fc_modified_inodes_used++] = ino;
1409 * Inode replay function
1411 static int ext4_fc_replay_inode(struct super_block *sb, struct ext4_fc_tl *tl,
1414 struct ext4_fc_inode fc_inode;
1415 struct ext4_inode *raw_inode;
1416 struct ext4_inode *raw_fc_inode;
1417 struct inode *inode = NULL;
1418 struct ext4_iloc iloc;
1419 int inode_len, ino, ret, tag = le16_to_cpu(tl->fc_tag);
1420 struct ext4_extent_header *eh;
1422 memcpy(&fc_inode, val, sizeof(fc_inode));
1424 ino = le32_to_cpu(fc_inode.fc_ino);
1425 trace_ext4_fc_replay(sb, tag, ino, 0, 0);
1427 inode = ext4_iget(sb, ino, EXT4_IGET_NORMAL);
1428 if (!IS_ERR(inode)) {
1429 ext4_ext_clear_bb(inode);
1434 ext4_fc_record_modified_inode(sb, ino);
1436 raw_fc_inode = (struct ext4_inode *)
1437 (val + offsetof(struct ext4_fc_inode, fc_raw_inode));
1438 ret = ext4_get_fc_inode_loc(sb, ino, &iloc);
1442 inode_len = le16_to_cpu(tl->fc_len) - sizeof(struct ext4_fc_inode);
1443 raw_inode = ext4_raw_inode(&iloc);
1445 memcpy(raw_inode, raw_fc_inode, offsetof(struct ext4_inode, i_block));
1446 memcpy(&raw_inode->i_generation, &raw_fc_inode->i_generation,
1447 inode_len - offsetof(struct ext4_inode, i_generation));
1448 if (le32_to_cpu(raw_inode->i_flags) & EXT4_EXTENTS_FL) {
1449 eh = (struct ext4_extent_header *)(&raw_inode->i_block[0]);
1450 if (eh->eh_magic != EXT4_EXT_MAGIC) {
1451 memset(eh, 0, sizeof(*eh));
1452 eh->eh_magic = EXT4_EXT_MAGIC;
1453 eh->eh_max = cpu_to_le16(
1454 (sizeof(raw_inode->i_block) -
1455 sizeof(struct ext4_extent_header))
1456 / sizeof(struct ext4_extent));
1458 } else if (le32_to_cpu(raw_inode->i_flags) & EXT4_INLINE_DATA_FL) {
1459 memcpy(raw_inode->i_block, raw_fc_inode->i_block,
1460 sizeof(raw_inode->i_block));
1463 /* Immediately update the inode on disk. */
1464 ret = ext4_handle_dirty_metadata(NULL, NULL, iloc.bh);
1467 ret = sync_dirty_buffer(iloc.bh);
1470 ret = ext4_mark_inode_used(sb, ino);
1474 /* Given that we just wrote the inode on disk, this SHOULD succeed. */
1475 inode = ext4_iget(sb, ino, EXT4_IGET_NORMAL);
1476 if (IS_ERR(inode)) {
1477 jbd_debug(1, "Inode not found.");
1478 return -EFSCORRUPTED;
1482 * Our allocator could have made different decisions than before
1483 * crashing. This should be fixed but until then, we calculate
1484 * the number of blocks the inode.
1486 if (!ext4_test_inode_flag(inode, EXT4_INODE_INLINE_DATA))
1487 ext4_ext_replay_set_iblocks(inode);
1489 inode->i_generation = le32_to_cpu(ext4_raw_inode(&iloc)->i_generation);
1490 ext4_reset_inode_seed(inode);
1492 ext4_inode_csum_set(inode, ext4_raw_inode(&iloc), EXT4_I(inode));
1493 ret = ext4_handle_dirty_metadata(NULL, NULL, iloc.bh);
1494 sync_dirty_buffer(iloc.bh);
1499 blkdev_issue_flush(sb->s_bdev);
1505 * Dentry create replay function.
1507 * EXT4_FC_TAG_CREAT is preceded by EXT4_FC_TAG_INODE_FULL. Which means, the
1508 * inode for which we are trying to create a dentry here, should already have
1509 * been replayed before we start here.
1511 static int ext4_fc_replay_create(struct super_block *sb, struct ext4_fc_tl *tl,
1515 struct inode *inode = NULL;
1516 struct inode *dir = NULL;
1517 struct dentry_info_args darg;
1519 tl_to_darg(&darg, tl, val);
1521 trace_ext4_fc_replay(sb, EXT4_FC_TAG_CREAT, darg.ino,
1522 darg.parent_ino, darg.dname_len);
1524 /* This takes care of update group descriptor and other metadata */
1525 ret = ext4_mark_inode_used(sb, darg.ino);
1529 inode = ext4_iget(sb, darg.ino, EXT4_IGET_NORMAL);
1530 if (IS_ERR(inode)) {
1531 jbd_debug(1, "inode %d not found.", darg.ino);
1537 if (S_ISDIR(inode->i_mode)) {
1539 * If we are creating a directory, we need to make sure that the
1540 * dot and dot dot dirents are setup properly.
1542 dir = ext4_iget(sb, darg.parent_ino, EXT4_IGET_NORMAL);
1544 jbd_debug(1, "Dir %d not found.", darg.ino);
1547 ret = ext4_init_new_dir(NULL, dir, inode);
1554 ret = ext4_fc_replay_link_internal(sb, &darg, inode);
1557 set_nlink(inode, 1);
1558 ext4_mark_inode_dirty(NULL, inode);
1566 * Record physical disk regions which are in use as per fast commit area,
1567 * and used by inodes during replay phase. Our simple replay phase
1568 * allocator excludes these regions from allocation.
1570 int ext4_fc_record_regions(struct super_block *sb, int ino,
1571 ext4_lblk_t lblk, ext4_fsblk_t pblk, int len, int replay)
1573 struct ext4_fc_replay_state *state;
1574 struct ext4_fc_alloc_region *region;
1576 state = &EXT4_SB(sb)->s_fc_replay_state;
1578 * during replay phase, the fc_regions_valid may not same as
1579 * fc_regions_used, update it when do new additions.
1581 if (replay && state->fc_regions_used != state->fc_regions_valid)
1582 state->fc_regions_used = state->fc_regions_valid;
1583 if (state->fc_regions_used == state->fc_regions_size) {
1584 state->fc_regions_size +=
1585 EXT4_FC_REPLAY_REALLOC_INCREMENT;
1586 state->fc_regions = krealloc(
1588 state->fc_regions_size *
1589 sizeof(struct ext4_fc_alloc_region),
1591 if (!state->fc_regions)
1594 region = &state->fc_regions[state->fc_regions_used++];
1596 region->lblk = lblk;
1597 region->pblk = pblk;
1601 state->fc_regions_valid++;
1606 /* Replay add range tag */
1607 static int ext4_fc_replay_add_range(struct super_block *sb,
1608 struct ext4_fc_tl *tl, u8 *val)
1610 struct ext4_fc_add_range fc_add_ex;
1611 struct ext4_extent newex, *ex;
1612 struct inode *inode;
1613 ext4_lblk_t start, cur;
1615 ext4_fsblk_t start_pblk;
1616 struct ext4_map_blocks map;
1617 struct ext4_ext_path *path = NULL;
1620 memcpy(&fc_add_ex, val, sizeof(fc_add_ex));
1621 ex = (struct ext4_extent *)&fc_add_ex.fc_ex;
1623 trace_ext4_fc_replay(sb, EXT4_FC_TAG_ADD_RANGE,
1624 le32_to_cpu(fc_add_ex.fc_ino), le32_to_cpu(ex->ee_block),
1625 ext4_ext_get_actual_len(ex));
1627 inode = ext4_iget(sb, le32_to_cpu(fc_add_ex.fc_ino), EXT4_IGET_NORMAL);
1628 if (IS_ERR(inode)) {
1629 jbd_debug(1, "Inode not found.");
1633 ret = ext4_fc_record_modified_inode(sb, inode->i_ino);
1635 start = le32_to_cpu(ex->ee_block);
1636 start_pblk = ext4_ext_pblock(ex);
1637 len = ext4_ext_get_actual_len(ex);
1641 jbd_debug(1, "ADD_RANGE, lblk %d, pblk %lld, len %d, unwritten %d, inode %ld\n",
1642 start, start_pblk, len, ext4_ext_is_unwritten(ex),
1645 while (remaining > 0) {
1647 map.m_len = remaining;
1649 ret = ext4_map_blocks(NULL, inode, &map, 0);
1657 /* Range is not mapped */
1658 path = ext4_find_extent(inode, cur, NULL, 0);
1663 memset(&newex, 0, sizeof(newex));
1664 newex.ee_block = cpu_to_le32(cur);
1665 ext4_ext_store_pblock(
1666 &newex, start_pblk + cur - start);
1667 newex.ee_len = cpu_to_le16(map.m_len);
1668 if (ext4_ext_is_unwritten(ex))
1669 ext4_ext_mark_unwritten(&newex);
1670 down_write(&EXT4_I(inode)->i_data_sem);
1671 ret = ext4_ext_insert_extent(
1672 NULL, inode, &path, &newex, 0);
1673 up_write((&EXT4_I(inode)->i_data_sem));
1674 ext4_ext_drop_refs(path);
1683 if (start_pblk + cur - start != map.m_pblk) {
1685 * Logical to physical mapping changed. This can happen
1686 * if this range was removed and then reallocated to
1687 * map to new physical blocks during a fast commit.
1689 ret = ext4_ext_replay_update_ex(inode, cur, map.m_len,
1690 ext4_ext_is_unwritten(ex),
1691 start_pblk + cur - start);
1697 * Mark the old blocks as free since they aren't used
1698 * anymore. We maintain an array of all the modified
1699 * inodes. In case these blocks are still used at either
1700 * a different logical range in the same inode or in
1701 * some different inode, we will mark them as allocated
1702 * at the end of the FC replay using our array of
1705 ext4_mb_mark_bb(inode->i_sb, map.m_pblk, map.m_len, 0);
1709 /* Range is mapped and needs a state change */
1710 jbd_debug(1, "Converting from %ld to %d %lld",
1711 map.m_flags & EXT4_MAP_UNWRITTEN,
1712 ext4_ext_is_unwritten(ex), map.m_pblk);
1713 ret = ext4_ext_replay_update_ex(inode, cur, map.m_len,
1714 ext4_ext_is_unwritten(ex), map.m_pblk);
1720 * We may have split the extent tree while toggling the state.
1721 * Try to shrink the extent tree now.
1723 ext4_ext_replay_shrink_inode(inode, start + len);
1726 remaining -= map.m_len;
1728 ext4_ext_replay_shrink_inode(inode, i_size_read(inode) >>
1729 sb->s_blocksize_bits);
1734 /* Replay DEL_RANGE tag */
1736 ext4_fc_replay_del_range(struct super_block *sb, struct ext4_fc_tl *tl,
1739 struct inode *inode;
1740 struct ext4_fc_del_range lrange;
1741 struct ext4_map_blocks map;
1742 ext4_lblk_t cur, remaining;
1745 memcpy(&lrange, val, sizeof(lrange));
1746 cur = le32_to_cpu(lrange.fc_lblk);
1747 remaining = le32_to_cpu(lrange.fc_len);
1749 trace_ext4_fc_replay(sb, EXT4_FC_TAG_DEL_RANGE,
1750 le32_to_cpu(lrange.fc_ino), cur, remaining);
1752 inode = ext4_iget(sb, le32_to_cpu(lrange.fc_ino), EXT4_IGET_NORMAL);
1753 if (IS_ERR(inode)) {
1754 jbd_debug(1, "Inode %d not found", le32_to_cpu(lrange.fc_ino));
1758 ret = ext4_fc_record_modified_inode(sb, inode->i_ino);
1760 jbd_debug(1, "DEL_RANGE, inode %ld, lblk %d, len %d\n",
1761 inode->i_ino, le32_to_cpu(lrange.fc_lblk),
1762 le32_to_cpu(lrange.fc_len));
1763 while (remaining > 0) {
1765 map.m_len = remaining;
1767 ret = ext4_map_blocks(NULL, inode, &map, 0);
1775 ext4_mb_mark_bb(inode->i_sb, map.m_pblk, map.m_len, 0);
1777 remaining -= map.m_len;
1782 down_write(&EXT4_I(inode)->i_data_sem);
1783 ret = ext4_ext_remove_space(inode, lrange.fc_lblk,
1784 lrange.fc_lblk + lrange.fc_len - 1);
1785 up_write(&EXT4_I(inode)->i_data_sem);
1790 ext4_ext_replay_shrink_inode(inode,
1791 i_size_read(inode) >> sb->s_blocksize_bits);
1792 ext4_mark_inode_dirty(NULL, inode);
1798 static void ext4_fc_set_bitmaps_and_counters(struct super_block *sb)
1800 struct ext4_fc_replay_state *state;
1801 struct inode *inode;
1802 struct ext4_ext_path *path = NULL;
1803 struct ext4_map_blocks map;
1805 ext4_lblk_t cur, end;
1807 state = &EXT4_SB(sb)->s_fc_replay_state;
1808 for (i = 0; i < state->fc_modified_inodes_used; i++) {
1809 inode = ext4_iget(sb, state->fc_modified_inodes[i],
1811 if (IS_ERR(inode)) {
1812 jbd_debug(1, "Inode %d not found.",
1813 state->fc_modified_inodes[i]);
1817 end = EXT_MAX_BLOCKS;
1818 if (ext4_test_inode_flag(inode, EXT4_INODE_INLINE_DATA)) {
1824 map.m_len = end - cur;
1826 ret = ext4_map_blocks(NULL, inode, &map, 0);
1831 path = ext4_find_extent(inode, map.m_lblk, NULL, 0);
1832 if (!IS_ERR(path)) {
1833 for (j = 0; j < path->p_depth; j++)
1834 ext4_mb_mark_bb(inode->i_sb,
1835 path[j].p_block, 1, 1);
1836 ext4_ext_drop_refs(path);
1840 ext4_mb_mark_bb(inode->i_sb, map.m_pblk,
1843 cur = cur + (map.m_len ? map.m_len : 1);
1851 * Check if block is in excluded regions for block allocation. The simple
1852 * allocator that runs during replay phase is calls this function to see
1853 * if it is okay to use a block.
1855 bool ext4_fc_replay_check_excluded(struct super_block *sb, ext4_fsblk_t blk)
1858 struct ext4_fc_replay_state *state;
1860 state = &EXT4_SB(sb)->s_fc_replay_state;
1861 for (i = 0; i < state->fc_regions_valid; i++) {
1862 if (state->fc_regions[i].ino == 0 ||
1863 state->fc_regions[i].len == 0)
1865 if (blk >= state->fc_regions[i].pblk &&
1866 blk < state->fc_regions[i].pblk + state->fc_regions[i].len)
1872 /* Cleanup function called after replay */
1873 void ext4_fc_replay_cleanup(struct super_block *sb)
1875 struct ext4_sb_info *sbi = EXT4_SB(sb);
1877 sbi->s_mount_state &= ~EXT4_FC_REPLAY;
1878 kfree(sbi->s_fc_replay_state.fc_regions);
1879 kfree(sbi->s_fc_replay_state.fc_modified_inodes);
1883 * Recovery Scan phase handler
1885 * This function is called during the scan phase and is responsible
1886 * for doing following things:
1887 * - Make sure the fast commit area has valid tags for replay
1888 * - Count number of tags that need to be replayed by the replay handler
1890 * - Create a list of excluded blocks for allocation during replay phase
1892 * This function returns JBD2_FC_REPLAY_CONTINUE to indicate that SCAN is
1893 * incomplete and JBD2 should send more blocks. It returns JBD2_FC_REPLAY_STOP
1894 * to indicate that scan has finished and JBD2 can now start replay phase.
1895 * It returns a negative error to indicate that there was an error. At the end
1896 * of a successful scan phase, sbi->s_fc_replay_state.fc_replay_num_tags is set
1897 * to indicate the number of tags that need to replayed during the replay phase.
1899 static int ext4_fc_replay_scan(journal_t *journal,
1900 struct buffer_head *bh, int off,
1903 struct super_block *sb = journal->j_private;
1904 struct ext4_sb_info *sbi = EXT4_SB(sb);
1905 struct ext4_fc_replay_state *state;
1906 int ret = JBD2_FC_REPLAY_CONTINUE;
1907 struct ext4_fc_add_range ext;
1908 struct ext4_fc_tl tl;
1909 struct ext4_fc_tail tail;
1910 __u8 *start, *end, *cur, *val;
1911 struct ext4_fc_head head;
1912 struct ext4_extent *ex;
1914 state = &sbi->s_fc_replay_state;
1916 start = (u8 *)bh->b_data;
1917 end = (__u8 *)bh->b_data + journal->j_blocksize - 1;
1919 if (state->fc_replay_expected_off == 0) {
1920 state->fc_cur_tag = 0;
1921 state->fc_replay_num_tags = 0;
1923 state->fc_regions = NULL;
1924 state->fc_regions_valid = state->fc_regions_used =
1925 state->fc_regions_size = 0;
1926 /* Check if we can stop early */
1927 if (le16_to_cpu(((struct ext4_fc_tl *)start)->fc_tag)
1928 != EXT4_FC_TAG_HEAD)
1932 if (off != state->fc_replay_expected_off) {
1933 ret = -EFSCORRUPTED;
1937 state->fc_replay_expected_off++;
1938 for (cur = start; cur < end; cur = cur + sizeof(tl) + le16_to_cpu(tl.fc_len)) {
1939 memcpy(&tl, cur, sizeof(tl));
1940 val = cur + sizeof(tl);
1941 jbd_debug(3, "Scan phase, tag:%s, blk %lld\n",
1942 tag2str(le16_to_cpu(tl.fc_tag)), bh->b_blocknr);
1943 switch (le16_to_cpu(tl.fc_tag)) {
1944 case EXT4_FC_TAG_ADD_RANGE:
1945 memcpy(&ext, val, sizeof(ext));
1946 ex = (struct ext4_extent *)&ext.fc_ex;
1947 ret = ext4_fc_record_regions(sb,
1948 le32_to_cpu(ext.fc_ino),
1949 le32_to_cpu(ex->ee_block), ext4_ext_pblock(ex),
1950 ext4_ext_get_actual_len(ex), 0);
1953 ret = JBD2_FC_REPLAY_CONTINUE;
1955 case EXT4_FC_TAG_DEL_RANGE:
1956 case EXT4_FC_TAG_LINK:
1957 case EXT4_FC_TAG_UNLINK:
1958 case EXT4_FC_TAG_CREAT:
1959 case EXT4_FC_TAG_INODE:
1960 case EXT4_FC_TAG_PAD:
1961 state->fc_cur_tag++;
1962 state->fc_crc = ext4_chksum(sbi, state->fc_crc, cur,
1963 sizeof(tl) + le16_to_cpu(tl.fc_len));
1965 case EXT4_FC_TAG_TAIL:
1966 state->fc_cur_tag++;
1967 memcpy(&tail, val, sizeof(tail));
1968 state->fc_crc = ext4_chksum(sbi, state->fc_crc, cur,
1970 offsetof(struct ext4_fc_tail,
1972 if (le32_to_cpu(tail.fc_tid) == expected_tid &&
1973 le32_to_cpu(tail.fc_crc) == state->fc_crc) {
1974 state->fc_replay_num_tags = state->fc_cur_tag;
1975 state->fc_regions_valid =
1976 state->fc_regions_used;
1978 ret = state->fc_replay_num_tags ?
1979 JBD2_FC_REPLAY_STOP : -EFSBADCRC;
1983 case EXT4_FC_TAG_HEAD:
1984 memcpy(&head, val, sizeof(head));
1985 if (le32_to_cpu(head.fc_features) &
1986 ~EXT4_FC_SUPPORTED_FEATURES) {
1990 if (le32_to_cpu(head.fc_tid) != expected_tid) {
1991 ret = JBD2_FC_REPLAY_STOP;
1994 state->fc_cur_tag++;
1995 state->fc_crc = ext4_chksum(sbi, state->fc_crc, cur,
1996 sizeof(tl) + le16_to_cpu(tl.fc_len));
1999 ret = state->fc_replay_num_tags ?
2000 JBD2_FC_REPLAY_STOP : -ECANCELED;
2002 if (ret < 0 || ret == JBD2_FC_REPLAY_STOP)
2007 trace_ext4_fc_replay_scan(sb, ret, off);
2012 * Main recovery path entry point.
2013 * The meaning of return codes is similar as above.
2015 static int ext4_fc_replay(journal_t *journal, struct buffer_head *bh,
2016 enum passtype pass, int off, tid_t expected_tid)
2018 struct super_block *sb = journal->j_private;
2019 struct ext4_sb_info *sbi = EXT4_SB(sb);
2020 struct ext4_fc_tl tl;
2021 __u8 *start, *end, *cur, *val;
2022 int ret = JBD2_FC_REPLAY_CONTINUE;
2023 struct ext4_fc_replay_state *state = &sbi->s_fc_replay_state;
2024 struct ext4_fc_tail tail;
2026 if (pass == PASS_SCAN) {
2027 state->fc_current_pass = PASS_SCAN;
2028 return ext4_fc_replay_scan(journal, bh, off, expected_tid);
2031 if (state->fc_current_pass != pass) {
2032 state->fc_current_pass = pass;
2033 sbi->s_mount_state |= EXT4_FC_REPLAY;
2035 if (!sbi->s_fc_replay_state.fc_replay_num_tags) {
2036 jbd_debug(1, "Replay stops\n");
2037 ext4_fc_set_bitmaps_and_counters(sb);
2041 #ifdef CONFIG_EXT4_DEBUG
2042 if (sbi->s_fc_debug_max_replay && off >= sbi->s_fc_debug_max_replay) {
2043 pr_warn("Dropping fc block %d because max_replay set\n", off);
2044 return JBD2_FC_REPLAY_STOP;
2048 start = (u8 *)bh->b_data;
2049 end = (__u8 *)bh->b_data + journal->j_blocksize - 1;
2051 for (cur = start; cur < end; cur = cur + sizeof(tl) + le16_to_cpu(tl.fc_len)) {
2052 memcpy(&tl, cur, sizeof(tl));
2053 val = cur + sizeof(tl);
2055 if (state->fc_replay_num_tags == 0) {
2056 ret = JBD2_FC_REPLAY_STOP;
2057 ext4_fc_set_bitmaps_and_counters(sb);
2060 jbd_debug(3, "Replay phase, tag:%s\n",
2061 tag2str(le16_to_cpu(tl.fc_tag)));
2062 state->fc_replay_num_tags--;
2063 switch (le16_to_cpu(tl.fc_tag)) {
2064 case EXT4_FC_TAG_LINK:
2065 ret = ext4_fc_replay_link(sb, &tl, val);
2067 case EXT4_FC_TAG_UNLINK:
2068 ret = ext4_fc_replay_unlink(sb, &tl, val);
2070 case EXT4_FC_TAG_ADD_RANGE:
2071 ret = ext4_fc_replay_add_range(sb, &tl, val);
2073 case EXT4_FC_TAG_CREAT:
2074 ret = ext4_fc_replay_create(sb, &tl, val);
2076 case EXT4_FC_TAG_DEL_RANGE:
2077 ret = ext4_fc_replay_del_range(sb, &tl, val);
2079 case EXT4_FC_TAG_INODE:
2080 ret = ext4_fc_replay_inode(sb, &tl, val);
2082 case EXT4_FC_TAG_PAD:
2083 trace_ext4_fc_replay(sb, EXT4_FC_TAG_PAD, 0,
2084 le16_to_cpu(tl.fc_len), 0);
2086 case EXT4_FC_TAG_TAIL:
2087 trace_ext4_fc_replay(sb, EXT4_FC_TAG_TAIL, 0,
2088 le16_to_cpu(tl.fc_len), 0);
2089 memcpy(&tail, val, sizeof(tail));
2090 WARN_ON(le32_to_cpu(tail.fc_tid) != expected_tid);
2092 case EXT4_FC_TAG_HEAD:
2095 trace_ext4_fc_replay(sb, le16_to_cpu(tl.fc_tag), 0,
2096 le16_to_cpu(tl.fc_len), 0);
2102 ret = JBD2_FC_REPLAY_CONTINUE;
2107 void ext4_fc_init(struct super_block *sb, journal_t *journal)
2110 * We set replay callback even if fast commit disabled because we may
2111 * could still have fast commit blocks that need to be replayed even if
2112 * fast commit has now been turned off.
2114 journal->j_fc_replay_callback = ext4_fc_replay;
2115 if (!test_opt2(sb, JOURNAL_FAST_COMMIT))
2117 journal->j_fc_cleanup_callback = ext4_fc_cleanup;
2120 static const char *fc_ineligible_reasons[] = {
2121 "Extended attributes changed",
2123 "Journal flag changed",
2124 "Insufficient memory",
2133 int ext4_fc_info_show(struct seq_file *seq, void *v)
2135 struct ext4_sb_info *sbi = EXT4_SB((struct super_block *)seq->private);
2136 struct ext4_fc_stats *stats = &sbi->s_fc_stats;
2139 if (v != SEQ_START_TOKEN)
2143 "fc stats:\n%ld commits\n%ld ineligible\n%ld numblks\n%lluus avg_commit_time\n",
2144 stats->fc_num_commits, stats->fc_ineligible_commits,
2146 div_u64(stats->s_fc_avg_commit_time, 1000));
2147 seq_puts(seq, "Ineligible reasons:\n");
2148 for (i = 0; i < EXT4_FC_REASON_MAX; i++)
2149 seq_printf(seq, "\"%s\":\t%d\n", fc_ineligible_reasons[i],
2150 stats->fc_ineligible_reason_count[i]);
2155 int __init ext4_fc_init_dentry_cache(void)
2157 ext4_fc_dentry_cachep = KMEM_CACHE(ext4_fc_dentry_update,
2158 SLAB_RECLAIM_ACCOUNT);
2160 if (ext4_fc_dentry_cachep == NULL)
2166 void ext4_fc_destroy_dentry_cache(void)
2168 kmem_cache_destroy(ext4_fc_dentry_cachep);