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 * Increasing @c->dirty_pn_cnt/@c->dirty_nn_cnt and marking
74 * nnodes/pnodes as dirty in run_gc() could race with following
75 * checking, which leads inconsistent states between @c->nroot
76 * and @c->dirty_pn_cnt/@c->dirty_nn_cnt, holding @c->lp_mutex
79 mutex_lock(&c->lp_mutex);
81 * Even though the TNC is clean, the LPT tree may have dirty nodes. For
82 * example, this may happen if the budgeting subsystem invoked GC to
83 * make some free space, and the GC found an LEB with only dirty and
84 * free space. In this case GC would just change the lprops of this
85 * LEB (by turning all space into free space) and unmap it.
87 if (c->nroot && test_bit(DIRTY_CNODE, &c->nroot->flags)) {
88 mutex_unlock(&c->lp_mutex);
92 ubifs_assert(c, atomic_long_read(&c->dirty_zn_cnt) == 0);
93 ubifs_assert(c, c->dirty_pn_cnt == 0);
94 ubifs_assert(c, c->dirty_nn_cnt == 0);
95 mutex_unlock(&c->lp_mutex);
101 * do_commit - commit the journal.
102 * @c: UBIFS file-system description object
104 * This function implements UBIFS commit. It has to be called with commit lock
105 * locked. Returns zero in case of success and a negative error code in case of
108 static int do_commit(struct ubifs_info *c)
110 int err, new_ltail_lnum, old_ltail_lnum, i;
111 struct ubifs_zbranch zroot;
112 struct ubifs_lp_stats lst;
115 ubifs_assert(c, !c->ro_media && !c->ro_mount);
122 if (nothing_to_commit(c)) {
123 up_write(&c->commit_sem);
128 /* Sync all write buffers (necessary for recovery) */
129 for (i = 0; i < c->jhead_cnt; i++) {
130 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
136 err = ubifs_gc_start_commit(c);
139 err = dbg_check_lprops(c);
142 err = ubifs_log_start_commit(c, &new_ltail_lnum);
145 err = ubifs_tnc_start_commit(c, &zroot);
148 err = ubifs_lpt_start_commit(c);
151 err = ubifs_orphan_start_commit(c);
155 ubifs_get_lp_stats(c, &lst);
157 up_write(&c->commit_sem);
159 err = ubifs_tnc_end_commit(c);
162 err = ubifs_lpt_end_commit(c);
165 err = ubifs_orphan_end_commit(c);
168 err = dbg_check_old_index(c, &zroot);
172 c->mst_node->cmt_no = cpu_to_le64(c->cmt_no);
173 c->mst_node->log_lnum = cpu_to_le32(new_ltail_lnum);
174 c->mst_node->root_lnum = cpu_to_le32(zroot.lnum);
175 c->mst_node->root_offs = cpu_to_le32(zroot.offs);
176 c->mst_node->root_len = cpu_to_le32(zroot.len);
177 c->mst_node->ihead_lnum = cpu_to_le32(c->ihead_lnum);
178 c->mst_node->ihead_offs = cpu_to_le32(c->ihead_offs);
179 c->mst_node->index_size = cpu_to_le64(c->bi.old_idx_sz);
180 c->mst_node->lpt_lnum = cpu_to_le32(c->lpt_lnum);
181 c->mst_node->lpt_offs = cpu_to_le32(c->lpt_offs);
182 c->mst_node->nhead_lnum = cpu_to_le32(c->nhead_lnum);
183 c->mst_node->nhead_offs = cpu_to_le32(c->nhead_offs);
184 c->mst_node->ltab_lnum = cpu_to_le32(c->ltab_lnum);
185 c->mst_node->ltab_offs = cpu_to_le32(c->ltab_offs);
186 c->mst_node->lsave_lnum = cpu_to_le32(c->lsave_lnum);
187 c->mst_node->lsave_offs = cpu_to_le32(c->lsave_offs);
188 c->mst_node->lscan_lnum = cpu_to_le32(c->lscan_lnum);
189 c->mst_node->empty_lebs = cpu_to_le32(lst.empty_lebs);
190 c->mst_node->idx_lebs = cpu_to_le32(lst.idx_lebs);
191 c->mst_node->total_free = cpu_to_le64(lst.total_free);
192 c->mst_node->total_dirty = cpu_to_le64(lst.total_dirty);
193 c->mst_node->total_used = cpu_to_le64(lst.total_used);
194 c->mst_node->total_dead = cpu_to_le64(lst.total_dead);
195 c->mst_node->total_dark = cpu_to_le64(lst.total_dark);
197 c->mst_node->flags |= cpu_to_le32(UBIFS_MST_NO_ORPHS);
199 c->mst_node->flags &= ~cpu_to_le32(UBIFS_MST_NO_ORPHS);
201 old_ltail_lnum = c->ltail_lnum;
202 err = ubifs_log_end_commit(c, new_ltail_lnum);
206 err = ubifs_log_post_commit(c, old_ltail_lnum);
209 err = ubifs_gc_end_commit(c);
212 err = ubifs_lpt_post_commit(c);
217 spin_lock(&c->cs_lock);
218 c->cmt_state = COMMIT_RESTING;
220 dbg_cmt("commit end");
221 spin_unlock(&c->cs_lock);
225 up_write(&c->commit_sem);
227 ubifs_err(c, "commit failed, error %d", err);
228 spin_lock(&c->cs_lock);
229 c->cmt_state = COMMIT_BROKEN;
231 spin_unlock(&c->cs_lock);
232 ubifs_ro_mode(c, err);
237 * run_bg_commit - run background commit if it is needed.
238 * @c: UBIFS file-system description object
240 * This function runs background commit if it is needed. Returns zero in case
241 * of success and a negative error code in case of failure.
243 static int run_bg_commit(struct ubifs_info *c)
245 spin_lock(&c->cs_lock);
247 * Run background commit only if background commit was requested or if
248 * commit is required.
250 if (c->cmt_state != COMMIT_BACKGROUND &&
251 c->cmt_state != COMMIT_REQUIRED)
253 spin_unlock(&c->cs_lock);
255 down_write(&c->commit_sem);
256 spin_lock(&c->cs_lock);
257 if (c->cmt_state == COMMIT_REQUIRED)
258 c->cmt_state = COMMIT_RUNNING_REQUIRED;
259 else if (c->cmt_state == COMMIT_BACKGROUND)
260 c->cmt_state = COMMIT_RUNNING_BACKGROUND;
263 spin_unlock(&c->cs_lock);
268 up_write(&c->commit_sem);
270 spin_unlock(&c->cs_lock);
275 * ubifs_bg_thread - UBIFS background thread function.
276 * @info: points to the file-system description object
278 * This function implements various file-system background activities:
279 * o when a write-buffer timer expires it synchronizes the appropriate
281 * o when the journal is about to be full, it starts in-advance commit.
283 * Note, other stuff like background garbage collection may be added here in
286 int ubifs_bg_thread(void *info)
289 struct ubifs_info *c = info;
291 ubifs_msg(c, "background thread \"%s\" started, PID %d",
292 c->bgt_name, current->pid);
296 if (kthread_should_stop())
302 set_current_state(TASK_INTERRUPTIBLE);
303 /* Check if there is something to do */
306 * Nothing prevents us from going sleep now and
307 * be never woken up and block the task which
308 * could wait in 'kthread_stop()' forever.
310 if (kthread_should_stop())
315 __set_current_state(TASK_RUNNING);
318 err = ubifs_bg_wbufs_sync(c);
320 ubifs_ro_mode(c, err);
326 ubifs_msg(c, "background thread \"%s\" stops", c->bgt_name);
331 * ubifs_commit_required - set commit state to "required".
332 * @c: UBIFS file-system description object
334 * This function is called if a commit is required but cannot be done from the
335 * calling function, so it is just flagged instead.
337 void ubifs_commit_required(struct ubifs_info *c)
339 spin_lock(&c->cs_lock);
340 switch (c->cmt_state) {
342 case COMMIT_BACKGROUND:
343 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
344 dbg_cstate(COMMIT_REQUIRED));
345 c->cmt_state = COMMIT_REQUIRED;
347 case COMMIT_RUNNING_BACKGROUND:
348 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
349 dbg_cstate(COMMIT_RUNNING_REQUIRED));
350 c->cmt_state = COMMIT_RUNNING_REQUIRED;
352 case COMMIT_REQUIRED:
353 case COMMIT_RUNNING_REQUIRED:
357 spin_unlock(&c->cs_lock);
361 * ubifs_request_bg_commit - notify the background thread to do a commit.
362 * @c: UBIFS file-system description object
364 * This function is called if the journal is full enough to make a commit
365 * worthwhile, so background thread is kicked to start it.
367 void ubifs_request_bg_commit(struct ubifs_info *c)
369 spin_lock(&c->cs_lock);
370 if (c->cmt_state == COMMIT_RESTING) {
371 dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
372 dbg_cstate(COMMIT_BACKGROUND));
373 c->cmt_state = COMMIT_BACKGROUND;
374 spin_unlock(&c->cs_lock);
375 ubifs_wake_up_bgt(c);
377 spin_unlock(&c->cs_lock);
381 * wait_for_commit - wait for commit.
382 * @c: UBIFS file-system description object
384 * This function sleeps until the commit operation is no longer running.
386 static int wait_for_commit(struct ubifs_info *c)
388 dbg_cmt("pid %d goes sleep", current->pid);
391 * The following sleeps if the condition is false, and will be woken
392 * when the commit ends. It is possible, although very unlikely, that we
393 * will wake up and see the subsequent commit running, rather than the
394 * one we were waiting for, and go back to sleep. However, we will be
395 * woken again, so there is no danger of sleeping forever.
397 wait_event(c->cmt_wq, c->cmt_state != COMMIT_RUNNING_BACKGROUND &&
398 c->cmt_state != COMMIT_RUNNING_REQUIRED);
399 dbg_cmt("commit finished, pid %d woke up", current->pid);
404 * ubifs_run_commit - run or wait for commit.
405 * @c: UBIFS file-system description object
407 * This function runs commit and returns zero in case of success and a negative
408 * error code in case of failure.
410 int ubifs_run_commit(struct ubifs_info *c)
414 spin_lock(&c->cs_lock);
415 if (c->cmt_state == COMMIT_BROKEN) {
420 if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
422 * We set the commit state to 'running required' to indicate
423 * that we want it to complete as quickly as possible.
425 c->cmt_state = COMMIT_RUNNING_REQUIRED;
427 if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
428 spin_unlock(&c->cs_lock);
429 return wait_for_commit(c);
431 spin_unlock(&c->cs_lock);
433 /* Ok, the commit is indeed needed */
435 down_write(&c->commit_sem);
436 spin_lock(&c->cs_lock);
438 * Since we unlocked 'c->cs_lock', the state may have changed, so
441 if (c->cmt_state == COMMIT_BROKEN) {
446 if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
447 c->cmt_state = COMMIT_RUNNING_REQUIRED;
449 if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
450 up_write(&c->commit_sem);
451 spin_unlock(&c->cs_lock);
452 return wait_for_commit(c);
454 c->cmt_state = COMMIT_RUNNING_REQUIRED;
455 spin_unlock(&c->cs_lock);
461 up_write(&c->commit_sem);
463 spin_unlock(&c->cs_lock);
468 * ubifs_gc_should_commit - determine if it is time for GC to run commit.
469 * @c: UBIFS file-system description object
471 * This function is called by garbage collection to determine if commit should
472 * be run. If commit state is @COMMIT_BACKGROUND, which means that the journal
473 * is full enough to start commit, this function returns true. It is not
474 * absolutely necessary to commit yet, but it feels like this should be better
475 * then to keep doing GC. This function returns %1 if GC has to initiate commit
478 int ubifs_gc_should_commit(struct ubifs_info *c)
482 spin_lock(&c->cs_lock);
483 if (c->cmt_state == COMMIT_BACKGROUND) {
484 dbg_cmt("commit required now");
485 c->cmt_state = COMMIT_REQUIRED;
487 dbg_cmt("commit not requested");
488 if (c->cmt_state == COMMIT_REQUIRED)
490 spin_unlock(&c->cs_lock);
495 * Everything below is related to debugging.
499 * struct idx_node - hold index nodes during index tree traversal.
501 * @iip: index in parent (slot number of this indexing node in the parent
503 * @upper_key: all keys in this indexing node have to be less or equivalent to
505 * @idx: index node (8-byte aligned because all node structures must be 8-byte
509 struct list_head list;
511 union ubifs_key upper_key;
512 struct ubifs_idx_node idx __aligned(8);
516 * dbg_old_index_check_init - get information for the next old index check.
517 * @c: UBIFS file-system description object
518 * @zroot: root of the index
520 * This function records information about the index that will be needed for the
521 * next old index check i.e. 'dbg_check_old_index()'.
523 * This function returns %0 on success and a negative error code on failure.
525 int dbg_old_index_check_init(struct ubifs_info *c, struct ubifs_zbranch *zroot)
527 struct ubifs_idx_node *idx;
528 int lnum, offs, len, err = 0;
529 struct ubifs_debug_info *d = c->dbg;
531 d->old_zroot = *zroot;
532 lnum = d->old_zroot.lnum;
533 offs = d->old_zroot.offs;
534 len = d->old_zroot.len;
536 idx = kmalloc(c->max_idx_node_sz, GFP_NOFS);
540 err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
544 d->old_zroot_level = le16_to_cpu(idx->level);
545 d->old_zroot_sqnum = le64_to_cpu(idx->ch.sqnum);
552 * dbg_check_old_index - check the old copy of the index.
553 * @c: UBIFS file-system description object
554 * @zroot: root of the new index
556 * In order to be able to recover from an unclean unmount, a complete copy of
557 * the index must exist on flash. This is the "old" index. The commit process
558 * must write the "new" index to flash without overwriting or destroying any
559 * part of the old index. This function is run at commit end in order to check
560 * that the old index does indeed exist completely intact.
562 * This function returns %0 on success and a negative error code on failure.
564 int dbg_check_old_index(struct ubifs_info *c, struct ubifs_zbranch *zroot)
566 int lnum, offs, len, err = 0, last_level, child_cnt;
568 struct ubifs_debug_info *d = c->dbg;
569 union ubifs_key lower_key, upper_key, l_key, u_key;
570 unsigned long long last_sqnum;
571 struct ubifs_idx_node *idx;
572 struct list_head list;
576 if (!dbg_is_chk_index(c))
579 INIT_LIST_HEAD(&list);
581 sz = sizeof(struct idx_node) + ubifs_idx_node_sz(c, c->fanout) -
584 /* Start at the old zroot */
585 lnum = d->old_zroot.lnum;
586 offs = d->old_zroot.offs;
587 len = d->old_zroot.len;
591 * Traverse the index tree preorder depth-first i.e. do a node and then
592 * its subtrees from left to right.
595 struct ubifs_branch *br;
597 /* Get the next index node */
598 i = kmalloc(sz, GFP_NOFS);
604 /* Keep the index nodes on our path in a linked list */
605 list_add_tail(&i->list, &list);
606 /* Read the index node */
608 err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
611 /* Validate index node */
612 child_cnt = le16_to_cpu(idx->child_cnt);
613 if (child_cnt < 1 || child_cnt > c->fanout) {
619 /* Check root level and sqnum */
620 if (le16_to_cpu(idx->level) != d->old_zroot_level) {
624 if (le64_to_cpu(idx->ch.sqnum) != d->old_zroot_sqnum) {
628 /* Set last values as though root had a parent */
629 last_level = le16_to_cpu(idx->level) + 1;
630 last_sqnum = le64_to_cpu(idx->ch.sqnum) + 1;
631 key_read(c, ubifs_idx_key(c, idx), &lower_key);
632 highest_ino_key(c, &upper_key, INUM_WATERMARK);
634 key_copy(c, &upper_key, &i->upper_key);
635 if (le16_to_cpu(idx->level) != last_level - 1) {
640 * The index is always written bottom up hence a child's sqnum
641 * is always less than the parents.
643 if (le64_to_cpu(idx->ch.sqnum) >= last_sqnum) {
647 /* Check key range */
648 key_read(c, ubifs_idx_key(c, idx), &l_key);
649 br = ubifs_idx_branch(c, idx, child_cnt - 1);
650 key_read(c, &br->key, &u_key);
651 if (keys_cmp(c, &lower_key, &l_key) > 0) {
655 if (keys_cmp(c, &upper_key, &u_key) < 0) {
659 if (keys_cmp(c, &upper_key, &u_key) == 0)
660 if (!is_hash_key(c, &u_key)) {
664 /* Go to next index node */
665 if (le16_to_cpu(idx->level) == 0) {
666 /* At the bottom, so go up until can go right */
668 /* Drop the bottom of the list */
671 /* No more list means we are done */
672 if (list_empty(&list))
674 /* Look at the new bottom */
675 i = list_entry(list.prev, struct idx_node,
678 /* Can we go right */
679 if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
683 /* Nope, so go up again */
690 * We have the parent in 'idx' and now we set up for reading the
691 * child pointed to by slot 'iip'.
693 last_level = le16_to_cpu(idx->level);
694 last_sqnum = le64_to_cpu(idx->ch.sqnum);
695 br = ubifs_idx_branch(c, idx, iip);
696 lnum = le32_to_cpu(br->lnum);
697 offs = le32_to_cpu(br->offs);
698 len = le32_to_cpu(br->len);
699 key_read(c, &br->key, &lower_key);
700 if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
701 br = ubifs_idx_branch(c, idx, iip + 1);
702 key_read(c, &br->key, &upper_key);
704 key_copy(c, &i->upper_key, &upper_key);
707 err = dbg_old_index_check_init(c, zroot);
714 ubifs_err(c, "dumping index node (iip=%d)", i->iip);
715 ubifs_dump_node(c, idx, ubifs_idx_node_sz(c, c->fanout));
718 if (!list_empty(&list)) {
719 i = list_entry(list.prev, struct idx_node, list);
720 ubifs_err(c, "dumping parent index node");
721 ubifs_dump_node(c, &i->idx, ubifs_idx_node_sz(c, c->fanout));
724 while (!list_empty(&list)) {
725 i = list_entry(list.next, struct idx_node, list);
729 ubifs_err(c, "failed, error %d", err);