2 * This file is part of UBIFS.
4 * Copyright (C) 2006-2008 Nokia Corporation.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published by
8 * the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along with
16 * this program; if not, write to the Free Software Foundation, Inc., 51
17 * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 * Authors: Adrian Hunter
20 * Artem Bityutskiy (Битюцкий Артём)
24 * This file implements functions that manage the running of the commit process.
25 * Each affected module has its own functions to accomplish their part in the
26 * commit and those functions are called here.
28 * The commit is the process whereby all updates to the index and LEB properties
29 * are written out together and the journal becomes empty. This keeps the
30 * file system consistent - at all times the state can be recreated by reading
31 * the index and LEB properties and then replaying the journal.
33 * The commit is split into two parts named "commit start" and "commit end".
34 * During commit start, the commit process has exclusive access to the journal
35 * by holding the commit semaphore down for writing. As few I/O operations as
36 * possible are performed during commit start, instead the nodes that are to be
37 * written are merely identified. During commit end, the commit semaphore is no
38 * longer held and the journal is again in operation, allowing users to continue
39 * to use the file system while the bulk of the commit I/O is performed. The
40 * purpose of this two-step approach is to prevent the commit from causing any
41 * latency blips. Note that in any case, the commit does not prevent lookups
42 * (as permitted by the TNC mutex), or access to VFS data structures e.g. page
46 #include <linux/freezer.h>
47 #include <linux/kthread.h>
48 #include <linux/slab.h>
52 * nothing_to_commit - check if there is nothing to commit.
53 * @c: UBIFS file-system description object
55 * This is a helper function which checks if there is anything to commit. It is
56 * used as an optimization to avoid starting the commit if it is not really
57 * necessary. Indeed, the commit operation always assumes flash I/O (e.g.,
58 * writing the commit start node to the log), and it is better to avoid doing
59 * this unnecessarily. E.g., 'ubifs_sync_fs()' runs the commit, but if there is
60 * nothing to commit, it is more optimal to avoid any flash I/O.
62 * This function has to be called with @c->commit_sem locked for writing -
63 * this function does not take LPT/TNC locks because the @c->commit_sem
64 * guarantees that we have exclusive access to the TNC and LPT data structures.
66 * This function returns %1 if there is nothing to commit and %0 otherwise.
68 static int nothing_to_commit(struct ubifs_info
*c
)
71 * During mounting or remounting from R/O mode to R/W mode we may
72 * commit for various recovery-related reasons.
74 if (c
->mounting
|| c
->remounting_rw
)
78 * If the root TNC node is dirty, we definitely have something to
81 if (c
->zroot
.znode
&& ubifs_zn_dirty(c
->zroot
.znode
))
85 * Even though the TNC is clean, the LPT tree may have dirty nodes. For
86 * example, this may happen if the budgeting subsystem invoked GC to
87 * make some free space, and the GC found an LEB with only dirty and
88 * free space. In this case GC would just change the lprops of this
89 * LEB (by turning all space into free space) and unmap it.
91 if (c
->nroot
&& test_bit(DIRTY_CNODE
, &c
->nroot
->flags
))
94 ubifs_assert(atomic_long_read(&c
->dirty_zn_cnt
) == 0);
95 ubifs_assert(c
->dirty_pn_cnt
== 0);
96 ubifs_assert(c
->dirty_nn_cnt
== 0);
102 * do_commit - commit the journal.
103 * @c: UBIFS file-system description object
105 * This function implements UBIFS commit. It has to be called with commit lock
106 * locked. Returns zero in case of success and a negative error code in case of
109 static int do_commit(struct ubifs_info
*c
)
111 int err
, new_ltail_lnum
, old_ltail_lnum
, i
;
112 struct ubifs_zbranch zroot
;
113 struct ubifs_lp_stats lst
;
116 ubifs_assert(!c
->ro_media
&& !c
->ro_mount
);
123 if (nothing_to_commit(c
)) {
124 up_write(&c
->commit_sem
);
129 /* Sync all write buffers (necessary for recovery) */
130 for (i
= 0; i
< c
->jhead_cnt
; i
++) {
131 err
= ubifs_wbuf_sync(&c
->jheads
[i
].wbuf
);
137 err
= ubifs_gc_start_commit(c
);
140 err
= dbg_check_lprops(c
);
143 err
= ubifs_log_start_commit(c
, &new_ltail_lnum
);
146 err
= ubifs_tnc_start_commit(c
, &zroot
);
149 err
= ubifs_lpt_start_commit(c
);
152 err
= ubifs_orphan_start_commit(c
);
156 ubifs_get_lp_stats(c
, &lst
);
158 up_write(&c
->commit_sem
);
160 err
= ubifs_tnc_end_commit(c
);
163 err
= ubifs_lpt_end_commit(c
);
166 err
= ubifs_orphan_end_commit(c
);
169 old_ltail_lnum
= c
->ltail_lnum
;
170 err
= ubifs_log_end_commit(c
, new_ltail_lnum
);
173 err
= dbg_check_old_index(c
, &zroot
);
177 mutex_lock(&c
->mst_mutex
);
178 c
->mst_node
->cmt_no
= cpu_to_le64(c
->cmt_no
);
179 c
->mst_node
->log_lnum
= cpu_to_le32(new_ltail_lnum
);
180 c
->mst_node
->root_lnum
= cpu_to_le32(zroot
.lnum
);
181 c
->mst_node
->root_offs
= cpu_to_le32(zroot
.offs
);
182 c
->mst_node
->root_len
= cpu_to_le32(zroot
.len
);
183 c
->mst_node
->ihead_lnum
= cpu_to_le32(c
->ihead_lnum
);
184 c
->mst_node
->ihead_offs
= cpu_to_le32(c
->ihead_offs
);
185 c
->mst_node
->index_size
= cpu_to_le64(c
->bi
.old_idx_sz
);
186 c
->mst_node
->lpt_lnum
= cpu_to_le32(c
->lpt_lnum
);
187 c
->mst_node
->lpt_offs
= cpu_to_le32(c
->lpt_offs
);
188 c
->mst_node
->nhead_lnum
= cpu_to_le32(c
->nhead_lnum
);
189 c
->mst_node
->nhead_offs
= cpu_to_le32(c
->nhead_offs
);
190 c
->mst_node
->ltab_lnum
= cpu_to_le32(c
->ltab_lnum
);
191 c
->mst_node
->ltab_offs
= cpu_to_le32(c
->ltab_offs
);
192 c
->mst_node
->lsave_lnum
= cpu_to_le32(c
->lsave_lnum
);
193 c
->mst_node
->lsave_offs
= cpu_to_le32(c
->lsave_offs
);
194 c
->mst_node
->lscan_lnum
= cpu_to_le32(c
->lscan_lnum
);
195 c
->mst_node
->empty_lebs
= cpu_to_le32(lst
.empty_lebs
);
196 c
->mst_node
->idx_lebs
= cpu_to_le32(lst
.idx_lebs
);
197 c
->mst_node
->total_free
= cpu_to_le64(lst
.total_free
);
198 c
->mst_node
->total_dirty
= cpu_to_le64(lst
.total_dirty
);
199 c
->mst_node
->total_used
= cpu_to_le64(lst
.total_used
);
200 c
->mst_node
->total_dead
= cpu_to_le64(lst
.total_dead
);
201 c
->mst_node
->total_dark
= cpu_to_le64(lst
.total_dark
);
203 c
->mst_node
->flags
|= cpu_to_le32(UBIFS_MST_NO_ORPHS
);
205 c
->mst_node
->flags
&= ~cpu_to_le32(UBIFS_MST_NO_ORPHS
);
206 err
= ubifs_write_master(c
);
207 mutex_unlock(&c
->mst_mutex
);
211 err
= ubifs_log_post_commit(c
, old_ltail_lnum
);
214 err
= ubifs_gc_end_commit(c
);
217 err
= ubifs_lpt_post_commit(c
);
222 spin_lock(&c
->cs_lock
);
223 c
->cmt_state
= COMMIT_RESTING
;
225 dbg_cmt("commit end");
226 spin_unlock(&c
->cs_lock
);
230 up_write(&c
->commit_sem
);
232 ubifs_err("commit failed, error %d", err
);
233 spin_lock(&c
->cs_lock
);
234 c
->cmt_state
= COMMIT_BROKEN
;
236 spin_unlock(&c
->cs_lock
);
237 ubifs_ro_mode(c
, err
);
242 * run_bg_commit - run background commit if it is needed.
243 * @c: UBIFS file-system description object
245 * This function runs background commit if it is needed. Returns zero in case
246 * of success and a negative error code in case of failure.
248 static int run_bg_commit(struct ubifs_info
*c
)
250 spin_lock(&c
->cs_lock
);
252 * Run background commit only if background commit was requested or if
253 * commit is required.
255 if (c
->cmt_state
!= COMMIT_BACKGROUND
&&
256 c
->cmt_state
!= COMMIT_REQUIRED
)
258 spin_unlock(&c
->cs_lock
);
260 down_write(&c
->commit_sem
);
261 spin_lock(&c
->cs_lock
);
262 if (c
->cmt_state
== COMMIT_REQUIRED
)
263 c
->cmt_state
= COMMIT_RUNNING_REQUIRED
;
264 else if (c
->cmt_state
== COMMIT_BACKGROUND
)
265 c
->cmt_state
= COMMIT_RUNNING_BACKGROUND
;
268 spin_unlock(&c
->cs_lock
);
273 up_write(&c
->commit_sem
);
275 spin_unlock(&c
->cs_lock
);
280 * ubifs_bg_thread - UBIFS background thread function.
281 * @info: points to the file-system description object
283 * This function implements various file-system background activities:
284 * o when a write-buffer timer expires it synchronizes the appropriate
286 * o when the journal is about to be full, it starts in-advance commit.
288 * Note, other stuff like background garbage collection may be added here in
291 int ubifs_bg_thread(void *info
)
294 struct ubifs_info
*c
= info
;
296 ubifs_msg("background thread \"%s\" started, PID %d",
297 c
->bgt_name
, current
->pid
);
301 if (kthread_should_stop())
307 set_current_state(TASK_INTERRUPTIBLE
);
308 /* Check if there is something to do */
311 * Nothing prevents us from going sleep now and
312 * be never woken up and block the task which
313 * could wait in 'kthread_stop()' forever.
315 if (kthread_should_stop())
320 __set_current_state(TASK_RUNNING
);
323 err
= ubifs_bg_wbufs_sync(c
);
325 ubifs_ro_mode(c
, err
);
331 ubifs_msg("background thread \"%s\" stops", c
->bgt_name
);
336 * ubifs_commit_required - set commit state to "required".
337 * @c: UBIFS file-system description object
339 * This function is called if a commit is required but cannot be done from the
340 * calling function, so it is just flagged instead.
342 void ubifs_commit_required(struct ubifs_info
*c
)
344 spin_lock(&c
->cs_lock
);
345 switch (c
->cmt_state
) {
347 case COMMIT_BACKGROUND
:
348 dbg_cmt("old: %s, new: %s", dbg_cstate(c
->cmt_state
),
349 dbg_cstate(COMMIT_REQUIRED
));
350 c
->cmt_state
= COMMIT_REQUIRED
;
352 case COMMIT_RUNNING_BACKGROUND
:
353 dbg_cmt("old: %s, new: %s", dbg_cstate(c
->cmt_state
),
354 dbg_cstate(COMMIT_RUNNING_REQUIRED
));
355 c
->cmt_state
= COMMIT_RUNNING_REQUIRED
;
357 case COMMIT_REQUIRED
:
358 case COMMIT_RUNNING_REQUIRED
:
362 spin_unlock(&c
->cs_lock
);
366 * ubifs_request_bg_commit - notify the background thread to do a commit.
367 * @c: UBIFS file-system description object
369 * This function is called if the journal is full enough to make a commit
370 * worthwhile, so background thread is kicked to start it.
372 void ubifs_request_bg_commit(struct ubifs_info
*c
)
374 spin_lock(&c
->cs_lock
);
375 if (c
->cmt_state
== COMMIT_RESTING
) {
376 dbg_cmt("old: %s, new: %s", dbg_cstate(c
->cmt_state
),
377 dbg_cstate(COMMIT_BACKGROUND
));
378 c
->cmt_state
= COMMIT_BACKGROUND
;
379 spin_unlock(&c
->cs_lock
);
380 ubifs_wake_up_bgt(c
);
382 spin_unlock(&c
->cs_lock
);
386 * wait_for_commit - wait for commit.
387 * @c: UBIFS file-system description object
389 * This function sleeps until the commit operation is no longer running.
391 static int wait_for_commit(struct ubifs_info
*c
)
393 dbg_cmt("pid %d goes sleep", current
->pid
);
396 * The following sleeps if the condition is false, and will be woken
397 * when the commit ends. It is possible, although very unlikely, that we
398 * will wake up and see the subsequent commit running, rather than the
399 * one we were waiting for, and go back to sleep. However, we will be
400 * woken again, so there is no danger of sleeping forever.
402 wait_event(c
->cmt_wq
, c
->cmt_state
!= COMMIT_RUNNING_BACKGROUND
&&
403 c
->cmt_state
!= COMMIT_RUNNING_REQUIRED
);
404 dbg_cmt("commit finished, pid %d woke up", current
->pid
);
409 * ubifs_run_commit - run or wait for commit.
410 * @c: UBIFS file-system description object
412 * This function runs commit and returns zero in case of success and a negative
413 * error code in case of failure.
415 int ubifs_run_commit(struct ubifs_info
*c
)
419 spin_lock(&c
->cs_lock
);
420 if (c
->cmt_state
== COMMIT_BROKEN
) {
425 if (c
->cmt_state
== COMMIT_RUNNING_BACKGROUND
)
427 * We set the commit state to 'running required' to indicate
428 * that we want it to complete as quickly as possible.
430 c
->cmt_state
= COMMIT_RUNNING_REQUIRED
;
432 if (c
->cmt_state
== COMMIT_RUNNING_REQUIRED
) {
433 spin_unlock(&c
->cs_lock
);
434 return wait_for_commit(c
);
436 spin_unlock(&c
->cs_lock
);
438 /* Ok, the commit is indeed needed */
440 down_write(&c
->commit_sem
);
441 spin_lock(&c
->cs_lock
);
443 * Since we unlocked 'c->cs_lock', the state may have changed, so
446 if (c
->cmt_state
== COMMIT_BROKEN
) {
451 if (c
->cmt_state
== COMMIT_RUNNING_BACKGROUND
)
452 c
->cmt_state
= COMMIT_RUNNING_REQUIRED
;
454 if (c
->cmt_state
== COMMIT_RUNNING_REQUIRED
) {
455 up_write(&c
->commit_sem
);
456 spin_unlock(&c
->cs_lock
);
457 return wait_for_commit(c
);
459 c
->cmt_state
= COMMIT_RUNNING_REQUIRED
;
460 spin_unlock(&c
->cs_lock
);
466 up_write(&c
->commit_sem
);
468 spin_unlock(&c
->cs_lock
);
473 * ubifs_gc_should_commit - determine if it is time for GC to run commit.
474 * @c: UBIFS file-system description object
476 * This function is called by garbage collection to determine if commit should
477 * be run. If commit state is @COMMIT_BACKGROUND, which means that the journal
478 * is full enough to start commit, this function returns true. It is not
479 * absolutely necessary to commit yet, but it feels like this should be better
480 * then to keep doing GC. This function returns %1 if GC has to initiate commit
483 int ubifs_gc_should_commit(struct ubifs_info
*c
)
487 spin_lock(&c
->cs_lock
);
488 if (c
->cmt_state
== COMMIT_BACKGROUND
) {
489 dbg_cmt("commit required now");
490 c
->cmt_state
= COMMIT_REQUIRED
;
492 dbg_cmt("commit not requested");
493 if (c
->cmt_state
== COMMIT_REQUIRED
)
495 spin_unlock(&c
->cs_lock
);
500 * Everything below is related to debugging.
504 * struct idx_node - hold index nodes during index tree traversal.
506 * @iip: index in parent (slot number of this indexing node in the parent
508 * @upper_key: all keys in this indexing node have to be less or equivalent to
510 * @idx: index node (8-byte aligned because all node structures must be 8-byte
514 struct list_head list
;
516 union ubifs_key upper_key
;
517 struct ubifs_idx_node idx
__aligned(8);
521 * dbg_old_index_check_init - get information for the next old index check.
522 * @c: UBIFS file-system description object
523 * @zroot: root of the index
525 * This function records information about the index that will be needed for the
526 * next old index check i.e. 'dbg_check_old_index()'.
528 * This function returns %0 on success and a negative error code on failure.
530 int dbg_old_index_check_init(struct ubifs_info
*c
, struct ubifs_zbranch
*zroot
)
532 struct ubifs_idx_node
*idx
;
533 int lnum
, offs
, len
, err
= 0;
534 struct ubifs_debug_info
*d
= c
->dbg
;
536 d
->old_zroot
= *zroot
;
537 lnum
= d
->old_zroot
.lnum
;
538 offs
= d
->old_zroot
.offs
;
539 len
= d
->old_zroot
.len
;
541 idx
= kmalloc(c
->max_idx_node_sz
, GFP_NOFS
);
545 err
= ubifs_read_node(c
, idx
, UBIFS_IDX_NODE
, len
, lnum
, offs
);
549 d
->old_zroot_level
= le16_to_cpu(idx
->level
);
550 d
->old_zroot_sqnum
= le64_to_cpu(idx
->ch
.sqnum
);
557 * dbg_check_old_index - check the old copy of the index.
558 * @c: UBIFS file-system description object
559 * @zroot: root of the new index
561 * In order to be able to recover from an unclean unmount, a complete copy of
562 * the index must exist on flash. This is the "old" index. The commit process
563 * must write the "new" index to flash without overwriting or destroying any
564 * part of the old index. This function is run at commit end in order to check
565 * that the old index does indeed exist completely intact.
567 * This function returns %0 on success and a negative error code on failure.
569 int dbg_check_old_index(struct ubifs_info
*c
, struct ubifs_zbranch
*zroot
)
571 int lnum
, offs
, len
, err
= 0, uninitialized_var(last_level
), child_cnt
;
573 struct ubifs_debug_info
*d
= c
->dbg
;
574 union ubifs_key
uninitialized_var(lower_key
), upper_key
, l_key
, u_key
;
575 unsigned long long uninitialized_var(last_sqnum
);
576 struct ubifs_idx_node
*idx
;
577 struct list_head list
;
581 if (!dbg_is_chk_index(c
))
584 INIT_LIST_HEAD(&list
);
586 sz
= sizeof(struct idx_node
) + ubifs_idx_node_sz(c
, c
->fanout
) -
589 /* Start at the old zroot */
590 lnum
= d
->old_zroot
.lnum
;
591 offs
= d
->old_zroot
.offs
;
592 len
= d
->old_zroot
.len
;
596 * Traverse the index tree preorder depth-first i.e. do a node and then
597 * its subtrees from left to right.
600 struct ubifs_branch
*br
;
602 /* Get the next index node */
603 i
= kmalloc(sz
, GFP_NOFS
);
609 /* Keep the index nodes on our path in a linked list */
610 list_add_tail(&i
->list
, &list
);
611 /* Read the index node */
613 err
= ubifs_read_node(c
, idx
, UBIFS_IDX_NODE
, len
, lnum
, offs
);
616 /* Validate index node */
617 child_cnt
= le16_to_cpu(idx
->child_cnt
);
618 if (child_cnt
< 1 || child_cnt
> c
->fanout
) {
624 /* Check root level and sqnum */
625 if (le16_to_cpu(idx
->level
) != d
->old_zroot_level
) {
629 if (le64_to_cpu(idx
->ch
.sqnum
) != d
->old_zroot_sqnum
) {
633 /* Set last values as though root had a parent */
634 last_level
= le16_to_cpu(idx
->level
) + 1;
635 last_sqnum
= le64_to_cpu(idx
->ch
.sqnum
) + 1;
636 key_read(c
, ubifs_idx_key(c
, idx
), &lower_key
);
637 highest_ino_key(c
, &upper_key
, INUM_WATERMARK
);
639 key_copy(c
, &upper_key
, &i
->upper_key
);
640 if (le16_to_cpu(idx
->level
) != last_level
- 1) {
645 * The index is always written bottom up hence a child's sqnum
646 * is always less than the parents.
648 if (le64_to_cpu(idx
->ch
.sqnum
) >= last_sqnum
) {
652 /* Check key range */
653 key_read(c
, ubifs_idx_key(c
, idx
), &l_key
);
654 br
= ubifs_idx_branch(c
, idx
, child_cnt
- 1);
655 key_read(c
, &br
->key
, &u_key
);
656 if (keys_cmp(c
, &lower_key
, &l_key
) > 0) {
660 if (keys_cmp(c
, &upper_key
, &u_key
) < 0) {
664 if (keys_cmp(c
, &upper_key
, &u_key
) == 0)
665 if (!is_hash_key(c
, &u_key
)) {
669 /* Go to next index node */
670 if (le16_to_cpu(idx
->level
) == 0) {
671 /* At the bottom, so go up until can go right */
673 /* Drop the bottom of the list */
676 /* No more list means we are done */
677 if (list_empty(&list
))
679 /* Look at the new bottom */
680 i
= list_entry(list
.prev
, struct idx_node
,
683 /* Can we go right */
684 if (iip
+ 1 < le16_to_cpu(idx
->child_cnt
)) {
688 /* Nope, so go up again */
695 * We have the parent in 'idx' and now we set up for reading the
696 * child pointed to by slot 'iip'.
698 last_level
= le16_to_cpu(idx
->level
);
699 last_sqnum
= le64_to_cpu(idx
->ch
.sqnum
);
700 br
= ubifs_idx_branch(c
, idx
, iip
);
701 lnum
= le32_to_cpu(br
->lnum
);
702 offs
= le32_to_cpu(br
->offs
);
703 len
= le32_to_cpu(br
->len
);
704 key_read(c
, &br
->key
, &lower_key
);
705 if (iip
+ 1 < le16_to_cpu(idx
->child_cnt
)) {
706 br
= ubifs_idx_branch(c
, idx
, iip
+ 1);
707 key_read(c
, &br
->key
, &upper_key
);
709 key_copy(c
, &i
->upper_key
, &upper_key
);
712 err
= dbg_old_index_check_init(c
, zroot
);
719 ubifs_err("dumping index node (iip=%d)", i
->iip
);
720 ubifs_dump_node(c
, idx
);
723 if (!list_empty(&list
)) {
724 i
= list_entry(list
.prev
, struct idx_node
, list
);
725 ubifs_err("dumping parent index node");
726 ubifs_dump_node(c
, &i
->idx
);
729 while (!list_empty(&list
)) {
730 i
= list_entry(list
.next
, struct idx_node
, list
);
734 ubifs_err("failed, error %d", err
);