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 err = dbg_check_old_index(c, &zroot);
173 c->mst_node->cmt_no = cpu_to_le64(c->cmt_no);
174 c->mst_node->log_lnum = cpu_to_le32(new_ltail_lnum);
175 c->mst_node->root_lnum = cpu_to_le32(zroot.lnum);
176 c->mst_node->root_offs = cpu_to_le32(zroot.offs);
177 c->mst_node->root_len = cpu_to_le32(zroot.len);
178 c->mst_node->ihead_lnum = cpu_to_le32(c->ihead_lnum);
179 c->mst_node->ihead_offs = cpu_to_le32(c->ihead_offs);
180 c->mst_node->index_size = cpu_to_le64(c->bi.old_idx_sz);
181 c->mst_node->lpt_lnum = cpu_to_le32(c->lpt_lnum);
182 c->mst_node->lpt_offs = cpu_to_le32(c->lpt_offs);
183 c->mst_node->nhead_lnum = cpu_to_le32(c->nhead_lnum);
184 c->mst_node->nhead_offs = cpu_to_le32(c->nhead_offs);
185 c->mst_node->ltab_lnum = cpu_to_le32(c->ltab_lnum);
186 c->mst_node->ltab_offs = cpu_to_le32(c->ltab_offs);
187 c->mst_node->lsave_lnum = cpu_to_le32(c->lsave_lnum);
188 c->mst_node->lsave_offs = cpu_to_le32(c->lsave_offs);
189 c->mst_node->lscan_lnum = cpu_to_le32(c->lscan_lnum);
190 c->mst_node->empty_lebs = cpu_to_le32(lst.empty_lebs);
191 c->mst_node->idx_lebs = cpu_to_le32(lst.idx_lebs);
192 c->mst_node->total_free = cpu_to_le64(lst.total_free);
193 c->mst_node->total_dirty = cpu_to_le64(lst.total_dirty);
194 c->mst_node->total_used = cpu_to_le64(lst.total_used);
195 c->mst_node->total_dead = cpu_to_le64(lst.total_dead);
196 c->mst_node->total_dark = cpu_to_le64(lst.total_dark);
198 c->mst_node->flags |= cpu_to_le32(UBIFS_MST_NO_ORPHS);
200 c->mst_node->flags &= ~cpu_to_le32(UBIFS_MST_NO_ORPHS);
202 old_ltail_lnum = c->ltail_lnum;
203 err = ubifs_log_end_commit(c, new_ltail_lnum);
207 err = ubifs_log_post_commit(c, old_ltail_lnum);
210 err = ubifs_gc_end_commit(c);
213 err = ubifs_lpt_post_commit(c);
218 spin_lock(&c->cs_lock);
219 c->cmt_state = COMMIT_RESTING;
221 dbg_cmt("commit end");
222 spin_unlock(&c->cs_lock);
226 up_write(&c->commit_sem);
228 ubifs_err(c, "commit failed, error %d", err);
229 spin_lock(&c->cs_lock);
230 c->cmt_state = COMMIT_BROKEN;
232 spin_unlock(&c->cs_lock);
233 ubifs_ro_mode(c, err);
238 * run_bg_commit - run background commit if it is needed.
239 * @c: UBIFS file-system description object
241 * This function runs background commit if it is needed. Returns zero in case
242 * of success and a negative error code in case of failure.
244 static int run_bg_commit(struct ubifs_info *c)
246 spin_lock(&c->cs_lock);
248 * Run background commit only if background commit was requested or if
249 * commit is required.
251 if (c->cmt_state != COMMIT_BACKGROUND &&
252 c->cmt_state != COMMIT_REQUIRED)
254 spin_unlock(&c->cs_lock);
256 down_write(&c->commit_sem);
257 spin_lock(&c->cs_lock);
258 if (c->cmt_state == COMMIT_REQUIRED)
259 c->cmt_state = COMMIT_RUNNING_REQUIRED;
260 else if (c->cmt_state == COMMIT_BACKGROUND)
261 c->cmt_state = COMMIT_RUNNING_BACKGROUND;
264 spin_unlock(&c->cs_lock);
269 up_write(&c->commit_sem);
271 spin_unlock(&c->cs_lock);
276 * ubifs_bg_thread - UBIFS background thread function.
277 * @info: points to the file-system description object
279 * This function implements various file-system background activities:
280 * o when a write-buffer timer expires it synchronizes the appropriate
282 * o when the journal is about to be full, it starts in-advance commit.
284 * Note, other stuff like background garbage collection may be added here in
287 int ubifs_bg_thread(void *info)
290 struct ubifs_info *c = info;
292 ubifs_msg(c, "background thread \"%s\" started, PID %d",
293 c->bgt_name, current->pid);
297 if (kthread_should_stop())
303 set_current_state(TASK_INTERRUPTIBLE);
304 /* Check if there is something to do */
307 * Nothing prevents us from going sleep now and
308 * be never woken up and block the task which
309 * could wait in 'kthread_stop()' forever.
311 if (kthread_should_stop())
316 __set_current_state(TASK_RUNNING);
319 err = ubifs_bg_wbufs_sync(c);
321 ubifs_ro_mode(c, err);
327 ubifs_msg(c, "background thread \"%s\" stops", c->bgt_name);
332 * ubifs_commit_required - set commit state to "required".
333 * @c: UBIFS file-system description object
335 * This function is called if a commit is required but cannot be done from the
336 * calling function, so it is just flagged instead.
338 void ubifs_commit_required(struct ubifs_info *c)
340 spin_lock(&c->cs_lock);
341 switch (c->cmt_state) {
343 case COMMIT_BACKGROUND:
344 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
345 dbg_cstate(COMMIT_REQUIRED));
346 c->cmt_state = COMMIT_REQUIRED;
348 case COMMIT_RUNNING_BACKGROUND:
349 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
350 dbg_cstate(COMMIT_RUNNING_REQUIRED));
351 c->cmt_state = COMMIT_RUNNING_REQUIRED;
353 case COMMIT_REQUIRED:
354 case COMMIT_RUNNING_REQUIRED:
358 spin_unlock(&c->cs_lock);
362 * ubifs_request_bg_commit - notify the background thread to do a commit.
363 * @c: UBIFS file-system description object
365 * This function is called if the journal is full enough to make a commit
366 * worthwhile, so background thread is kicked to start it.
368 void ubifs_request_bg_commit(struct ubifs_info *c)
370 spin_lock(&c->cs_lock);
371 if (c->cmt_state == COMMIT_RESTING) {
372 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
373 dbg_cstate(COMMIT_BACKGROUND));
374 c->cmt_state = COMMIT_BACKGROUND;
375 spin_unlock(&c->cs_lock);
376 ubifs_wake_up_bgt(c);
378 spin_unlock(&c->cs_lock);
382 * wait_for_commit - wait for commit.
383 * @c: UBIFS file-system description object
385 * This function sleeps until the commit operation is no longer running.
387 static int wait_for_commit(struct ubifs_info *c)
389 dbg_cmt("pid %d goes sleep", current->pid);
392 * The following sleeps if the condition is false, and will be woken
393 * when the commit ends. It is possible, although very unlikely, that we
394 * will wake up and see the subsequent commit running, rather than the
395 * one we were waiting for, and go back to sleep. However, we will be
396 * woken again, so there is no danger of sleeping forever.
398 wait_event(c->cmt_wq, c->cmt_state != COMMIT_RUNNING_BACKGROUND &&
399 c->cmt_state != COMMIT_RUNNING_REQUIRED);
400 dbg_cmt("commit finished, pid %d woke up", current->pid);
405 * ubifs_run_commit - run or wait for commit.
406 * @c: UBIFS file-system description object
408 * This function runs commit and returns zero in case of success and a negative
409 * error code in case of failure.
411 int ubifs_run_commit(struct ubifs_info *c)
415 spin_lock(&c->cs_lock);
416 if (c->cmt_state == COMMIT_BROKEN) {
421 if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
423 * We set the commit state to 'running required' to indicate
424 * that we want it to complete as quickly as possible.
426 c->cmt_state = COMMIT_RUNNING_REQUIRED;
428 if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
429 spin_unlock(&c->cs_lock);
430 return wait_for_commit(c);
432 spin_unlock(&c->cs_lock);
434 /* Ok, the commit is indeed needed */
436 down_write(&c->commit_sem);
437 spin_lock(&c->cs_lock);
439 * Since we unlocked 'c->cs_lock', the state may have changed, so
442 if (c->cmt_state == COMMIT_BROKEN) {
447 if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
448 c->cmt_state = COMMIT_RUNNING_REQUIRED;
450 if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
451 up_write(&c->commit_sem);
452 spin_unlock(&c->cs_lock);
453 return wait_for_commit(c);
455 c->cmt_state = COMMIT_RUNNING_REQUIRED;
456 spin_unlock(&c->cs_lock);
462 up_write(&c->commit_sem);
464 spin_unlock(&c->cs_lock);
469 * ubifs_gc_should_commit - determine if it is time for GC to run commit.
470 * @c: UBIFS file-system description object
472 * This function is called by garbage collection to determine if commit should
473 * be run. If commit state is @COMMIT_BACKGROUND, which means that the journal
474 * is full enough to start commit, this function returns true. It is not
475 * absolutely necessary to commit yet, but it feels like this should be better
476 * then to keep doing GC. This function returns %1 if GC has to initiate commit
479 int ubifs_gc_should_commit(struct ubifs_info *c)
483 spin_lock(&c->cs_lock);
484 if (c->cmt_state == COMMIT_BACKGROUND) {
485 dbg_cmt("commit required now");
486 c->cmt_state = COMMIT_REQUIRED;
488 dbg_cmt("commit not requested");
489 if (c->cmt_state == COMMIT_REQUIRED)
491 spin_unlock(&c->cs_lock);
496 * Everything below is related to debugging.
500 * struct idx_node - hold index nodes during index tree traversal.
502 * @iip: index in parent (slot number of this indexing node in the parent
504 * @upper_key: all keys in this indexing node have to be less or equivalent to
506 * @idx: index node (8-byte aligned because all node structures must be 8-byte
510 struct list_head list;
512 union ubifs_key upper_key;
513 struct ubifs_idx_node idx __aligned(8);
517 * dbg_old_index_check_init - get information for the next old index check.
518 * @c: UBIFS file-system description object
519 * @zroot: root of the index
521 * This function records information about the index that will be needed for the
522 * next old index check i.e. 'dbg_check_old_index()'.
524 * This function returns %0 on success and a negative error code on failure.
526 int dbg_old_index_check_init(struct ubifs_info *c, struct ubifs_zbranch *zroot)
528 struct ubifs_idx_node *idx;
529 int lnum, offs, len, err = 0;
530 struct ubifs_debug_info *d = c->dbg;
532 d->old_zroot = *zroot;
533 lnum = d->old_zroot.lnum;
534 offs = d->old_zroot.offs;
535 len = d->old_zroot.len;
537 idx = kmalloc(c->max_idx_node_sz, GFP_NOFS);
541 err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
545 d->old_zroot_level = le16_to_cpu(idx->level);
546 d->old_zroot_sqnum = le64_to_cpu(idx->ch.sqnum);
553 * dbg_check_old_index - check the old copy of the index.
554 * @c: UBIFS file-system description object
555 * @zroot: root of the new index
557 * In order to be able to recover from an unclean unmount, a complete copy of
558 * the index must exist on flash. This is the "old" index. The commit process
559 * must write the "new" index to flash without overwriting or destroying any
560 * part of the old index. This function is run at commit end in order to check
561 * that the old index does indeed exist completely intact.
563 * This function returns %0 on success and a negative error code on failure.
565 int dbg_check_old_index(struct ubifs_info *c, struct ubifs_zbranch *zroot)
567 int lnum, offs, len, err = 0, uninitialized_var(last_level), child_cnt;
569 struct ubifs_debug_info *d = c->dbg;
570 union ubifs_key uninitialized_var(lower_key), upper_key, l_key, u_key;
571 unsigned long long uninitialized_var(last_sqnum);
572 struct ubifs_idx_node *idx;
573 struct list_head list;
577 if (!dbg_is_chk_index(c))
580 INIT_LIST_HEAD(&list);
582 sz = sizeof(struct idx_node) + ubifs_idx_node_sz(c, c->fanout) -
585 /* Start at the old zroot */
586 lnum = d->old_zroot.lnum;
587 offs = d->old_zroot.offs;
588 len = d->old_zroot.len;
592 * Traverse the index tree preorder depth-first i.e. do a node and then
593 * its subtrees from left to right.
596 struct ubifs_branch *br;
598 /* Get the next index node */
599 i = kmalloc(sz, GFP_NOFS);
605 /* Keep the index nodes on our path in a linked list */
606 list_add_tail(&i->list, &list);
607 /* Read the index node */
609 err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
612 /* Validate index node */
613 child_cnt = le16_to_cpu(idx->child_cnt);
614 if (child_cnt < 1 || child_cnt > c->fanout) {
620 /* Check root level and sqnum */
621 if (le16_to_cpu(idx->level) != d->old_zroot_level) {
625 if (le64_to_cpu(idx->ch.sqnum) != d->old_zroot_sqnum) {
629 /* Set last values as though root had a parent */
630 last_level = le16_to_cpu(idx->level) + 1;
631 last_sqnum = le64_to_cpu(idx->ch.sqnum) + 1;
632 key_read(c, ubifs_idx_key(c, idx), &lower_key);
633 highest_ino_key(c, &upper_key, INUM_WATERMARK);
635 key_copy(c, &upper_key, &i->upper_key);
636 if (le16_to_cpu(idx->level) != last_level - 1) {
641 * The index is always written bottom up hence a child's sqnum
642 * is always less than the parents.
644 if (le64_to_cpu(idx->ch.sqnum) >= last_sqnum) {
648 /* Check key range */
649 key_read(c, ubifs_idx_key(c, idx), &l_key);
650 br = ubifs_idx_branch(c, idx, child_cnt - 1);
651 key_read(c, &br->key, &u_key);
652 if (keys_cmp(c, &lower_key, &l_key) > 0) {
656 if (keys_cmp(c, &upper_key, &u_key) < 0) {
660 if (keys_cmp(c, &upper_key, &u_key) == 0)
661 if (!is_hash_key(c, &u_key)) {
665 /* Go to next index node */
666 if (le16_to_cpu(idx->level) == 0) {
667 /* At the bottom, so go up until can go right */
669 /* Drop the bottom of the list */
672 /* No more list means we are done */
673 if (list_empty(&list))
675 /* Look at the new bottom */
676 i = list_entry(list.prev, struct idx_node,
679 /* Can we go right */
680 if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
684 /* Nope, so go up again */
691 * We have the parent in 'idx' and now we set up for reading the
692 * child pointed to by slot 'iip'.
694 last_level = le16_to_cpu(idx->level);
695 last_sqnum = le64_to_cpu(idx->ch.sqnum);
696 br = ubifs_idx_branch(c, idx, iip);
697 lnum = le32_to_cpu(br->lnum);
698 offs = le32_to_cpu(br->offs);
699 len = le32_to_cpu(br->len);
700 key_read(c, &br->key, &lower_key);
701 if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
702 br = ubifs_idx_branch(c, idx, iip + 1);
703 key_read(c, &br->key, &upper_key);
705 key_copy(c, &i->upper_key, &upper_key);
708 err = dbg_old_index_check_init(c, zroot);
715 ubifs_err(c, "dumping index node (iip=%d)", i->iip);
716 ubifs_dump_node(c, idx);
719 if (!list_empty(&list)) {
720 i = list_entry(list.prev, struct idx_node, list);
721 ubifs_err(c, "dumping parent index node");
722 ubifs_dump_node(c, &i->idx);
725 while (!list_empty(&list)) {
726 i = list_entry(list.next, struct idx_node, list);
730 ubifs_err(c, "failed, error %d", err);