1 // SPDX-License-Identifier: GPL-2.0-only
3 * This file is part of UBIFS.
5 * Copyright (C) 2006-2008 Nokia Corporation.
7 * Authors: Adrian Hunter
8 * Artem Bityutskiy (Битюцкий Артём)
12 * This file implements functions that manage the running of the commit process.
13 * Each affected module has its own functions to accomplish their part in the
14 * commit and those functions are called here.
16 * The commit is the process whereby all updates to the index and LEB properties
17 * are written out together and the journal becomes empty. This keeps the
18 * file system consistent - at all times the state can be recreated by reading
19 * the index and LEB properties and then replaying the journal.
21 * The commit is split into two parts named "commit start" and "commit end".
22 * During commit start, the commit process has exclusive access to the journal
23 * by holding the commit semaphore down for writing. As few I/O operations as
24 * possible are performed during commit start, instead the nodes that are to be
25 * written are merely identified. During commit end, the commit semaphore is no
26 * longer held and the journal is again in operation, allowing users to continue
27 * to use the file system while the bulk of the commit I/O is performed. The
28 * purpose of this two-step approach is to prevent the commit from causing any
29 * latency blips. Note that in any case, the commit does not prevent lookups
30 * (as permitted by the TNC mutex), or access to VFS data structures e.g. page
34 #include <linux/freezer.h>
35 #include <linux/kthread.h>
36 #include <linux/slab.h>
40 * nothing_to_commit - check if there is nothing to commit.
41 * @c: UBIFS file-system description object
43 * This is a helper function which checks if there is anything to commit. It is
44 * used as an optimization to avoid starting the commit if it is not really
45 * necessary. Indeed, the commit operation always assumes flash I/O (e.g.,
46 * writing the commit start node to the log), and it is better to avoid doing
47 * this unnecessarily. E.g., 'ubifs_sync_fs()' runs the commit, but if there is
48 * nothing to commit, it is more optimal to avoid any flash I/O.
50 * This function has to be called with @c->commit_sem locked for writing -
51 * this function does not take LPT/TNC locks because the @c->commit_sem
52 * guarantees that we have exclusive access to the TNC and LPT data structures.
54 * This function returns %1 if there is nothing to commit and %0 otherwise.
56 static int nothing_to_commit(struct ubifs_info *c)
59 * During mounting or remounting from R/O mode to R/W mode we may
60 * commit for various recovery-related reasons.
62 if (c->mounting || c->remounting_rw)
66 * If the root TNC node is dirty, we definitely have something to
69 if (c->zroot.znode && ubifs_zn_dirty(c->zroot.znode))
73 * Even though the TNC is clean, the LPT tree may have dirty nodes. For
74 * example, this may happen if the budgeting subsystem invoked GC to
75 * make some free space, and the GC found an LEB with only dirty and
76 * free space. In this case GC would just change the lprops of this
77 * LEB (by turning all space into free space) and unmap it.
79 if (c->nroot && test_bit(DIRTY_CNODE, &c->nroot->flags))
82 ubifs_assert(c, atomic_long_read(&c->dirty_zn_cnt) == 0);
83 ubifs_assert(c, c->dirty_pn_cnt == 0);
84 ubifs_assert(c, c->dirty_nn_cnt == 0);
90 * do_commit - commit the journal.
91 * @c: UBIFS file-system description object
93 * This function implements UBIFS commit. It has to be called with commit lock
94 * locked. Returns zero in case of success and a negative error code in case of
97 static int do_commit(struct ubifs_info *c)
99 int err, new_ltail_lnum, old_ltail_lnum, i;
100 struct ubifs_zbranch zroot;
101 struct ubifs_lp_stats lst;
104 ubifs_assert(c, !c->ro_media && !c->ro_mount);
111 if (nothing_to_commit(c)) {
112 up_write(&c->commit_sem);
117 /* Sync all write buffers (necessary for recovery) */
118 for (i = 0; i < c->jhead_cnt; i++) {
119 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
125 err = ubifs_gc_start_commit(c);
128 err = dbg_check_lprops(c);
131 err = ubifs_log_start_commit(c, &new_ltail_lnum);
134 err = ubifs_tnc_start_commit(c, &zroot);
137 err = ubifs_lpt_start_commit(c);
140 err = ubifs_orphan_start_commit(c);
144 ubifs_get_lp_stats(c, &lst);
146 up_write(&c->commit_sem);
148 err = ubifs_tnc_end_commit(c);
151 err = ubifs_lpt_end_commit(c);
154 err = ubifs_orphan_end_commit(c);
157 err = dbg_check_old_index(c, &zroot);
161 c->mst_node->cmt_no = cpu_to_le64(c->cmt_no);
162 c->mst_node->log_lnum = cpu_to_le32(new_ltail_lnum);
163 c->mst_node->root_lnum = cpu_to_le32(zroot.lnum);
164 c->mst_node->root_offs = cpu_to_le32(zroot.offs);
165 c->mst_node->root_len = cpu_to_le32(zroot.len);
166 c->mst_node->ihead_lnum = cpu_to_le32(c->ihead_lnum);
167 c->mst_node->ihead_offs = cpu_to_le32(c->ihead_offs);
168 c->mst_node->index_size = cpu_to_le64(c->bi.old_idx_sz);
169 c->mst_node->lpt_lnum = cpu_to_le32(c->lpt_lnum);
170 c->mst_node->lpt_offs = cpu_to_le32(c->lpt_offs);
171 c->mst_node->nhead_lnum = cpu_to_le32(c->nhead_lnum);
172 c->mst_node->nhead_offs = cpu_to_le32(c->nhead_offs);
173 c->mst_node->ltab_lnum = cpu_to_le32(c->ltab_lnum);
174 c->mst_node->ltab_offs = cpu_to_le32(c->ltab_offs);
175 c->mst_node->lsave_lnum = cpu_to_le32(c->lsave_lnum);
176 c->mst_node->lsave_offs = cpu_to_le32(c->lsave_offs);
177 c->mst_node->lscan_lnum = cpu_to_le32(c->lscan_lnum);
178 c->mst_node->empty_lebs = cpu_to_le32(lst.empty_lebs);
179 c->mst_node->idx_lebs = cpu_to_le32(lst.idx_lebs);
180 c->mst_node->total_free = cpu_to_le64(lst.total_free);
181 c->mst_node->total_dirty = cpu_to_le64(lst.total_dirty);
182 c->mst_node->total_used = cpu_to_le64(lst.total_used);
183 c->mst_node->total_dead = cpu_to_le64(lst.total_dead);
184 c->mst_node->total_dark = cpu_to_le64(lst.total_dark);
186 c->mst_node->flags |= cpu_to_le32(UBIFS_MST_NO_ORPHS);
188 c->mst_node->flags &= ~cpu_to_le32(UBIFS_MST_NO_ORPHS);
190 old_ltail_lnum = c->ltail_lnum;
191 err = ubifs_log_end_commit(c, new_ltail_lnum);
195 err = ubifs_log_post_commit(c, old_ltail_lnum);
198 err = ubifs_gc_end_commit(c);
201 err = ubifs_lpt_post_commit(c);
206 spin_lock(&c->cs_lock);
207 c->cmt_state = COMMIT_RESTING;
209 dbg_cmt("commit end");
210 spin_unlock(&c->cs_lock);
214 up_write(&c->commit_sem);
216 ubifs_err(c, "commit failed, error %d", err);
217 spin_lock(&c->cs_lock);
218 c->cmt_state = COMMIT_BROKEN;
220 spin_unlock(&c->cs_lock);
221 ubifs_ro_mode(c, err);
226 * run_bg_commit - run background commit if it is needed.
227 * @c: UBIFS file-system description object
229 * This function runs background commit if it is needed. Returns zero in case
230 * of success and a negative error code in case of failure.
232 static int run_bg_commit(struct ubifs_info *c)
234 spin_lock(&c->cs_lock);
236 * Run background commit only if background commit was requested or if
237 * commit is required.
239 if (c->cmt_state != COMMIT_BACKGROUND &&
240 c->cmt_state != COMMIT_REQUIRED)
242 spin_unlock(&c->cs_lock);
244 down_write(&c->commit_sem);
245 spin_lock(&c->cs_lock);
246 if (c->cmt_state == COMMIT_REQUIRED)
247 c->cmt_state = COMMIT_RUNNING_REQUIRED;
248 else if (c->cmt_state == COMMIT_BACKGROUND)
249 c->cmt_state = COMMIT_RUNNING_BACKGROUND;
252 spin_unlock(&c->cs_lock);
257 up_write(&c->commit_sem);
259 spin_unlock(&c->cs_lock);
264 * ubifs_bg_thread - UBIFS background thread function.
265 * @info: points to the file-system description object
267 * This function implements various file-system background activities:
268 * o when a write-buffer timer expires it synchronizes the appropriate
270 * o when the journal is about to be full, it starts in-advance commit.
272 * Note, other stuff like background garbage collection may be added here in
275 int ubifs_bg_thread(void *info)
278 struct ubifs_info *c = info;
280 ubifs_msg(c, "background thread \"%s\" started, PID %d",
281 c->bgt_name, current->pid);
285 if (kthread_should_stop())
291 set_current_state(TASK_INTERRUPTIBLE);
292 /* Check if there is something to do */
295 * Nothing prevents us from going sleep now and
296 * be never woken up and block the task which
297 * could wait in 'kthread_stop()' forever.
299 if (kthread_should_stop())
304 __set_current_state(TASK_RUNNING);
307 err = ubifs_bg_wbufs_sync(c);
309 ubifs_ro_mode(c, err);
315 ubifs_msg(c, "background thread \"%s\" stops", c->bgt_name);
320 * ubifs_commit_required - set commit state to "required".
321 * @c: UBIFS file-system description object
323 * This function is called if a commit is required but cannot be done from the
324 * calling function, so it is just flagged instead.
326 void ubifs_commit_required(struct ubifs_info *c)
328 spin_lock(&c->cs_lock);
329 switch (c->cmt_state) {
331 case COMMIT_BACKGROUND:
332 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
333 dbg_cstate(COMMIT_REQUIRED));
334 c->cmt_state = COMMIT_REQUIRED;
336 case COMMIT_RUNNING_BACKGROUND:
337 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
338 dbg_cstate(COMMIT_RUNNING_REQUIRED));
339 c->cmt_state = COMMIT_RUNNING_REQUIRED;
341 case COMMIT_REQUIRED:
342 case COMMIT_RUNNING_REQUIRED:
346 spin_unlock(&c->cs_lock);
350 * ubifs_request_bg_commit - notify the background thread to do a commit.
351 * @c: UBIFS file-system description object
353 * This function is called if the journal is full enough to make a commit
354 * worthwhile, so background thread is kicked to start it.
356 void ubifs_request_bg_commit(struct ubifs_info *c)
358 spin_lock(&c->cs_lock);
359 if (c->cmt_state == COMMIT_RESTING) {
360 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
361 dbg_cstate(COMMIT_BACKGROUND));
362 c->cmt_state = COMMIT_BACKGROUND;
363 spin_unlock(&c->cs_lock);
364 ubifs_wake_up_bgt(c);
366 spin_unlock(&c->cs_lock);
370 * wait_for_commit - wait for commit.
371 * @c: UBIFS file-system description object
373 * This function sleeps until the commit operation is no longer running.
375 static int wait_for_commit(struct ubifs_info *c)
377 dbg_cmt("pid %d goes sleep", current->pid);
380 * The following sleeps if the condition is false, and will be woken
381 * when the commit ends. It is possible, although very unlikely, that we
382 * will wake up and see the subsequent commit running, rather than the
383 * one we were waiting for, and go back to sleep. However, we will be
384 * woken again, so there is no danger of sleeping forever.
386 wait_event(c->cmt_wq, c->cmt_state != COMMIT_RUNNING_BACKGROUND &&
387 c->cmt_state != COMMIT_RUNNING_REQUIRED);
388 dbg_cmt("commit finished, pid %d woke up", current->pid);
393 * ubifs_run_commit - run or wait for commit.
394 * @c: UBIFS file-system description object
396 * This function runs commit and returns zero in case of success and a negative
397 * error code in case of failure.
399 int ubifs_run_commit(struct ubifs_info *c)
403 spin_lock(&c->cs_lock);
404 if (c->cmt_state == COMMIT_BROKEN) {
409 if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
411 * We set the commit state to 'running required' to indicate
412 * that we want it to complete as quickly as possible.
414 c->cmt_state = COMMIT_RUNNING_REQUIRED;
416 if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
417 spin_unlock(&c->cs_lock);
418 return wait_for_commit(c);
420 spin_unlock(&c->cs_lock);
422 /* Ok, the commit is indeed needed */
424 down_write(&c->commit_sem);
425 spin_lock(&c->cs_lock);
427 * Since we unlocked 'c->cs_lock', the state may have changed, so
430 if (c->cmt_state == COMMIT_BROKEN) {
435 if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
436 c->cmt_state = COMMIT_RUNNING_REQUIRED;
438 if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
439 up_write(&c->commit_sem);
440 spin_unlock(&c->cs_lock);
441 return wait_for_commit(c);
443 c->cmt_state = COMMIT_RUNNING_REQUIRED;
444 spin_unlock(&c->cs_lock);
450 up_write(&c->commit_sem);
452 spin_unlock(&c->cs_lock);
457 * ubifs_gc_should_commit - determine if it is time for GC to run commit.
458 * @c: UBIFS file-system description object
460 * This function is called by garbage collection to determine if commit should
461 * be run. If commit state is @COMMIT_BACKGROUND, which means that the journal
462 * is full enough to start commit, this function returns true. It is not
463 * absolutely necessary to commit yet, but it feels like this should be better
464 * then to keep doing GC. This function returns %1 if GC has to initiate commit
467 int ubifs_gc_should_commit(struct ubifs_info *c)
471 spin_lock(&c->cs_lock);
472 if (c->cmt_state == COMMIT_BACKGROUND) {
473 dbg_cmt("commit required now");
474 c->cmt_state = COMMIT_REQUIRED;
476 dbg_cmt("commit not requested");
477 if (c->cmt_state == COMMIT_REQUIRED)
479 spin_unlock(&c->cs_lock);
484 * Everything below is related to debugging.
488 * struct idx_node - hold index nodes during index tree traversal.
490 * @iip: index in parent (slot number of this indexing node in the parent
492 * @upper_key: all keys in this indexing node have to be less or equivalent to
494 * @idx: index node (8-byte aligned because all node structures must be 8-byte
498 struct list_head list;
500 union ubifs_key upper_key;
501 struct ubifs_idx_node idx __aligned(8);
505 * dbg_old_index_check_init - get information for the next old index check.
506 * @c: UBIFS file-system description object
507 * @zroot: root of the index
509 * This function records information about the index that will be needed for the
510 * next old index check i.e. 'dbg_check_old_index()'.
512 * This function returns %0 on success and a negative error code on failure.
514 int dbg_old_index_check_init(struct ubifs_info *c, struct ubifs_zbranch *zroot)
516 struct ubifs_idx_node *idx;
517 int lnum, offs, len, err = 0;
518 struct ubifs_debug_info *d = c->dbg;
520 d->old_zroot = *zroot;
521 lnum = d->old_zroot.lnum;
522 offs = d->old_zroot.offs;
523 len = d->old_zroot.len;
525 idx = kmalloc(c->max_idx_node_sz, GFP_NOFS);
529 err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
533 d->old_zroot_level = le16_to_cpu(idx->level);
534 d->old_zroot_sqnum = le64_to_cpu(idx->ch.sqnum);
541 * dbg_check_old_index - check the old copy of the index.
542 * @c: UBIFS file-system description object
543 * @zroot: root of the new index
545 * In order to be able to recover from an unclean unmount, a complete copy of
546 * the index must exist on flash. This is the "old" index. The commit process
547 * must write the "new" index to flash without overwriting or destroying any
548 * part of the old index. This function is run at commit end in order to check
549 * that the old index does indeed exist completely intact.
551 * This function returns %0 on success and a negative error code on failure.
553 int dbg_check_old_index(struct ubifs_info *c, struct ubifs_zbranch *zroot)
555 int lnum, offs, len, err = 0, last_level, child_cnt;
557 struct ubifs_debug_info *d = c->dbg;
558 union ubifs_key lower_key, upper_key, l_key, u_key;
559 unsigned long long last_sqnum;
560 struct ubifs_idx_node *idx;
561 struct list_head list;
565 if (!dbg_is_chk_index(c))
568 INIT_LIST_HEAD(&list);
570 sz = sizeof(struct idx_node) + ubifs_idx_node_sz(c, c->fanout) -
573 /* Start at the old zroot */
574 lnum = d->old_zroot.lnum;
575 offs = d->old_zroot.offs;
576 len = d->old_zroot.len;
580 * Traverse the index tree preorder depth-first i.e. do a node and then
581 * its subtrees from left to right.
584 struct ubifs_branch *br;
586 /* Get the next index node */
587 i = kmalloc(sz, GFP_NOFS);
593 /* Keep the index nodes on our path in a linked list */
594 list_add_tail(&i->list, &list);
595 /* Read the index node */
597 err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
600 /* Validate index node */
601 child_cnt = le16_to_cpu(idx->child_cnt);
602 if (child_cnt < 1 || child_cnt > c->fanout) {
608 /* Check root level and sqnum */
609 if (le16_to_cpu(idx->level) != d->old_zroot_level) {
613 if (le64_to_cpu(idx->ch.sqnum) != d->old_zroot_sqnum) {
617 /* Set last values as though root had a parent */
618 last_level = le16_to_cpu(idx->level) + 1;
619 last_sqnum = le64_to_cpu(idx->ch.sqnum) + 1;
620 key_read(c, ubifs_idx_key(c, idx), &lower_key);
621 highest_ino_key(c, &upper_key, INUM_WATERMARK);
623 key_copy(c, &upper_key, &i->upper_key);
624 if (le16_to_cpu(idx->level) != last_level - 1) {
629 * The index is always written bottom up hence a child's sqnum
630 * is always less than the parents.
632 if (le64_to_cpu(idx->ch.sqnum) >= last_sqnum) {
636 /* Check key range */
637 key_read(c, ubifs_idx_key(c, idx), &l_key);
638 br = ubifs_idx_branch(c, idx, child_cnt - 1);
639 key_read(c, &br->key, &u_key);
640 if (keys_cmp(c, &lower_key, &l_key) > 0) {
644 if (keys_cmp(c, &upper_key, &u_key) < 0) {
648 if (keys_cmp(c, &upper_key, &u_key) == 0)
649 if (!is_hash_key(c, &u_key)) {
653 /* Go to next index node */
654 if (le16_to_cpu(idx->level) == 0) {
655 /* At the bottom, so go up until can go right */
657 /* Drop the bottom of the list */
660 /* No more list means we are done */
661 if (list_empty(&list))
663 /* Look at the new bottom */
664 i = list_entry(list.prev, struct idx_node,
667 /* Can we go right */
668 if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
672 /* Nope, so go up again */
679 * We have the parent in 'idx' and now we set up for reading the
680 * child pointed to by slot 'iip'.
682 last_level = le16_to_cpu(idx->level);
683 last_sqnum = le64_to_cpu(idx->ch.sqnum);
684 br = ubifs_idx_branch(c, idx, iip);
685 lnum = le32_to_cpu(br->lnum);
686 offs = le32_to_cpu(br->offs);
687 len = le32_to_cpu(br->len);
688 key_read(c, &br->key, &lower_key);
689 if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
690 br = ubifs_idx_branch(c, idx, iip + 1);
691 key_read(c, &br->key, &upper_key);
693 key_copy(c, &i->upper_key, &upper_key);
696 err = dbg_old_index_check_init(c, zroot);
703 ubifs_err(c, "dumping index node (iip=%d)", i->iip);
704 ubifs_dump_node(c, idx, ubifs_idx_node_sz(c, c->fanout));
707 if (!list_empty(&list)) {
708 i = list_entry(list.prev, struct idx_node, list);
709 ubifs_err(c, "dumping parent index node");
710 ubifs_dump_node(c, &i->idx, ubifs_idx_node_sz(c, c->fanout));
713 while (!list_empty(&list)) {
714 i = list_entry(list.next, struct idx_node, list);
718 ubifs_err(c, "failed, error %d", err);