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: Artem Bityutskiy (Битюцкий Артём)
24 * This file implements UBIFS shrinker which evicts clean znodes from the TNC
25 * tree when Linux VM needs more RAM.
27 * We do not implement any LRU lists to find oldest znodes to free because it
28 * would add additional overhead to the file system fast paths. So the shrinker
29 * just walks the TNC tree when searching for znodes to free.
31 * If the root of a TNC sub-tree is clean and old enough, then the children are
32 * also clean and old enough. So the shrinker walks the TNC in level order and
33 * dumps entire sub-trees.
35 * The age of znodes is just the time-stamp when they were last looked at.
36 * The current shrinker first tries to evict old znodes, then young ones.
38 * Since the shrinker is global, it has to protect against races with FS
39 * un-mounts, which is done by the 'ubifs_infos_lock' and 'c->umount_mutex'.
44 /* List of all UBIFS file-system instances */
45 LIST_HEAD(ubifs_infos);
48 * We number each shrinker run and record the number on the ubifs_info structure
49 * so that we can easily work out which ubifs_info structures have already been
50 * done by the current run.
52 static unsigned int shrinker_run_no;
54 /* Protects 'ubifs_infos' list */
55 DEFINE_SPINLOCK(ubifs_infos_lock);
57 /* Global clean znode counter (for all mounted UBIFS instances) */
58 atomic_long_t ubifs_clean_zn_cnt;
61 * shrink_tnc - shrink TNC tree.
62 * @c: UBIFS file-system description object
63 * @nr: number of znodes to free
64 * @age: the age of znodes to free
65 * @contention: if any contention, this is set to %1
67 * This function traverses TNC tree and frees clean znodes. It does not free
68 * clean znodes which younger then @age. Returns number of freed znodes.
70 static int shrink_tnc(struct ubifs_info *c, int nr, int age, int *contention)
73 struct ubifs_znode *znode, *zprev;
74 int time = get_seconds();
76 ubifs_assert(mutex_is_locked(&c->umount_mutex));
77 ubifs_assert(mutex_is_locked(&c->tnc_mutex));
79 if (!c->zroot.znode || atomic_long_read(&c->clean_zn_cnt) == 0)
83 * Traverse the TNC tree in levelorder manner, so that it is possible
84 * to destroy large sub-trees. Indeed, if a znode is old, then all its
85 * children are older or of the same age.
87 * Note, we are holding 'c->tnc_mutex', so we do not have to lock the
88 * 'c->space_lock' when _reading_ 'c->clean_zn_cnt', because it is
89 * changed only when the 'c->tnc_mutex' is held.
92 znode = ubifs_tnc_levelorder_next(c->zroot.znode, NULL);
93 while (znode && total_freed < nr &&
94 atomic_long_read(&c->clean_zn_cnt) > 0) {
98 * If the znode is clean, but it is in the 'c->cnext' list, this
99 * means that this znode has just been written to flash as a
100 * part of commit and was marked clean. They will be removed
101 * from the list at end commit. We cannot change the list,
102 * because it is not protected by any mutex (design decision to
103 * make commit really independent and parallel to main I/O). So
104 * we just skip these znodes.
106 * Note, the 'clean_zn_cnt' counters are not updated until
107 * after the commit, so the UBIFS shrinker does not report
108 * the znodes which are in the 'c->cnext' list as freeable.
110 * Also note, if the root of a sub-tree is not in 'c->cnext',
111 * then the whole sub-tree is not in 'c->cnext' as well, so it
112 * is safe to dump whole sub-tree.
117 * Very soon these znodes will be removed from the list
118 * and become freeable.
121 } else if (!ubifs_zn_dirty(znode) &&
122 abs(time - znode->time) >= age) {
124 znode->parent->zbranch[znode->iip].znode = NULL;
126 c->zroot.znode = NULL;
128 freed = ubifs_destroy_tnc_subtree(znode);
129 atomic_long_sub(freed, &ubifs_clean_zn_cnt);
130 atomic_long_sub(freed, &c->clean_zn_cnt);
131 total_freed += freed;
135 if (unlikely(!c->zroot.znode))
139 znode = ubifs_tnc_levelorder_next(c->zroot.znode, znode);
147 * shrink_tnc_trees - shrink UBIFS TNC trees.
148 * @nr: number of znodes to free
149 * @age: the age of znodes to free
150 * @contention: if any contention, this is set to %1
152 * This function walks the list of mounted UBIFS file-systems and frees clean
153 * znodes which are older than @age, until at least @nr znodes are freed.
154 * Returns the number of freed znodes.
156 static int shrink_tnc_trees(int nr, int age, int *contention)
158 struct ubifs_info *c;
163 spin_lock(&ubifs_infos_lock);
165 run_no = ++shrinker_run_no;
166 } while (run_no == 0);
167 /* Iterate over all mounted UBIFS file-systems and try to shrink them */
168 p = ubifs_infos.next;
169 while (p != &ubifs_infos) {
170 c = list_entry(p, struct ubifs_info, infos_list);
172 * We move the ones we do to the end of the list, so we stop
173 * when we see one we have already done.
175 if (c->shrinker_run_no == run_no)
177 if (!mutex_trylock(&c->umount_mutex)) {
178 /* Some un-mount is in progress, try next FS */
184 * We're holding 'c->umount_mutex', so the file-system won't go
187 if (!mutex_trylock(&c->tnc_mutex)) {
188 mutex_unlock(&c->umount_mutex);
193 spin_unlock(&ubifs_infos_lock);
195 * OK, now we have TNC locked, the file-system cannot go away -
196 * it is safe to reap the cache.
198 c->shrinker_run_no = run_no;
199 freed += shrink_tnc(c, nr, age, contention);
200 mutex_unlock(&c->tnc_mutex);
201 spin_lock(&ubifs_infos_lock);
202 /* Get the next list element before we move this one */
205 * Move this one to the end of the list to provide some
208 list_move_tail(&c->infos_list, &ubifs_infos);
209 mutex_unlock(&c->umount_mutex);
213 spin_unlock(&ubifs_infos_lock);
218 * kick_a_thread - kick a background thread to start commit.
220 * This function kicks a background thread to start background commit. Returns
221 * %-1 if a thread was kicked or there is another reason to assume the memory
222 * will soon be freed or become freeable. If there are no dirty znodes, returns
225 static int kick_a_thread(void)
228 struct ubifs_info *c;
231 * Iterate over all mounted UBIFS file-systems and find out if there is
232 * already an ongoing commit operation there. If no, then iterate for
233 * the second time and initiate background commit.
235 spin_lock(&ubifs_infos_lock);
236 for (i = 0; i < 2; i++) {
237 list_for_each_entry(c, &ubifs_infos, infos_list) {
240 if (!mutex_trylock(&c->umount_mutex)) {
242 * Some un-mount is in progress, it will
243 * certainly free memory, so just return.
245 spin_unlock(&ubifs_infos_lock);
249 dirty_zn_cnt = atomic_long_read(&c->dirty_zn_cnt);
251 if (!dirty_zn_cnt || c->cmt_state == COMMIT_BROKEN ||
252 c->ro_mount || c->ro_error) {
253 mutex_unlock(&c->umount_mutex);
257 if (c->cmt_state != COMMIT_RESTING) {
258 spin_unlock(&ubifs_infos_lock);
259 mutex_unlock(&c->umount_mutex);
264 list_move_tail(&c->infos_list, &ubifs_infos);
265 spin_unlock(&ubifs_infos_lock);
267 ubifs_request_bg_commit(c);
268 mutex_unlock(&c->umount_mutex);
271 mutex_unlock(&c->umount_mutex);
274 spin_unlock(&ubifs_infos_lock);
279 unsigned long ubifs_shrink_count(struct shrinker *shrink,
280 struct shrink_control *sc)
282 long clean_zn_cnt = atomic_long_read(&ubifs_clean_zn_cnt);
285 * Due to the way UBIFS updates the clean znode counter it may
286 * temporarily be negative.
288 return clean_zn_cnt >= 0 ? clean_zn_cnt : 1;
291 unsigned long ubifs_shrink_scan(struct shrinker *shrink,
292 struct shrink_control *sc)
294 unsigned long nr = sc->nr_to_scan;
297 long clean_zn_cnt = atomic_long_read(&ubifs_clean_zn_cnt);
301 * No clean znodes, nothing to reap. All we can do in this case
302 * is to kick background threads to start commit, which will
303 * probably make clean znodes which, in turn, will be freeable.
304 * And we return -1 which means will make VM call us again
307 dbg_tnc("no clean znodes, kick a thread");
308 return kick_a_thread();
311 freed = shrink_tnc_trees(nr, OLD_ZNODE_AGE, &contention);
315 dbg_tnc("not enough old znodes, try to free young ones");
316 freed += shrink_tnc_trees(nr - freed, YOUNG_ZNODE_AGE, &contention);
320 dbg_tnc("not enough young znodes, free all");
321 freed += shrink_tnc_trees(nr - freed, 0, &contention);
323 if (!freed && contention) {
324 dbg_tnc("freed nothing, but contention");
329 dbg_tnc("%lu znodes were freed, requested %lu", freed, nr);