fs/ubifs/shrinker.c

   1 /*
   2  * This file is part of UBIFS.
   3  *
   4  * Copyright (C) 2006-2008 Nokia Corporation.
   5  *
   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.
   9  *
  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
  13  * more details.
  14  *
  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
  18  *
  19  * Authors: Artem Bityutskiy (Битюцкий Артём)
  20  *          Adrian Hunter
  21  */
  22
  23 /*
  24  * This file implements UBIFS shrinker which evicts clean znodes from the TNC
  25  * tree when Linux VM needs more RAM.
  26  *
  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.
  30  *
  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.
  34  *
  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.
  37  *
  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'.
  40  */
  41
  42 #include "ubifs.h"
  43
  44 /* List of all UBIFS file-system instances */
  45 LIST_HEAD(ubifs_infos);
  46
  47 /*
  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.
  51  */
  52 static unsigned int shrinker_run_no;
  53
  54 /* Protects 'ubifs_infos' list */
  55 DEFINE_SPINLOCK(ubifs_infos_lock);
  56
  57 /* Global clean znode counter (for all mounted UBIFS instances) */
  58 atomic_long_t ubifs_clean_zn_cnt;
  59
  60 /**
  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
  66  *
  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.
  69  */
  70 static int shrink_tnc(struct ubifs_info *c, int nr, int age, int *contention)
  71 {
  72         int total_freed = 0;
  73         struct ubifs_znode *znode, *zprev;
  74         int time = get_seconds();
  75
  76         ubifs_assert(mutex_is_locked(&c->umount_mutex));
  77         ubifs_assert(mutex_is_locked(&c->tnc_mutex));
  78
  79         if (!c->zroot.znode || atomic_long_read(&c->clean_zn_cnt) == 0)
  80                 return 0;
  81
  82         /*
  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.
  86          *
  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.
  90          */
  91         zprev = NULL;
  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) {
  95                 int freed;
  96
  97                 /*
  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.
 105                  *
 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.
 109                  *
 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.
 113                  */
 114
 115                 if (znode->cnext) {
 116                         /*
 117                          * Very soon these znodes will be removed from the list
 118                          * and become freeable.
 119                          */
 120                         *contention = 1;
 121                 } else if (!ubifs_zn_dirty(znode) &&
 122                            abs(time - znode->time) >= age) {
 123                         if (znode->parent)
 124                                 znode->parent->zbranch[znode->iip].znode = NULL;
 125                         else
 126                                 c->zroot.znode = NULL;
 127
 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                         ubifs_assert(atomic_long_read(&c->clean_zn_cnt) >= 0);
 132                         total_freed += freed;
 133                         znode = zprev;
 134                 }
 135
 136                 if (unlikely(!c->zroot.znode))
 137                         break;
 138
 139                 zprev = znode;
 140                 znode = ubifs_tnc_levelorder_next(c->zroot.znode, znode);
 141                 cond_resched();
 142         }
 143
 144         return total_freed;
 145 }
 146
 147 /**
 148  * shrink_tnc_trees - shrink UBIFS TNC trees.
 149  * @nr: number of znodes to free
 150  * @age: the age of znodes to free
 151  * @contention: if any contention, this is set to %1
 152  *
 153  * This function walks the list of mounted UBIFS file-systems and frees clean
 154  * znodes which are older than @age, until at least @nr znodes are freed.
 155  * Returns the number of freed znodes.
 156  */
 157 static int shrink_tnc_trees(int nr, int age, int *contention)
 158 {
 159         struct ubifs_info *c;
 160         struct list_head *p;
 161         unsigned int run_no;
 162         int freed = 0;
 163
 164         spin_lock(&ubifs_infos_lock);
 165         do {
 166                 run_no = ++shrinker_run_no;
 167         } while (run_no == 0);
 168         /* Iterate over all mounted UBIFS file-systems and try to shrink them */
 169         p = ubifs_infos.next;
 170         while (p != &ubifs_infos) {
 171                 c = list_entry(p, struct ubifs_info, infos_list);
 172                 /*
 173                  * We move the ones we do to the end of the list, so we stop
 174                  * when we see one we have already done.
 175                  */
 176                 if (c->shrinker_run_no == run_no)
 177                         break;
 178                 if (!mutex_trylock(&c->umount_mutex)) {
 179                         /* Some un-mount is in progress, try next FS */
 180                         *contention = 1;
 181                         p = p->next;
 182                         continue;
 183                 }
 184                 /*
 185                  * We're holding 'c->umount_mutex', so the file-system won't go
 186                  * away.
 187                  */
 188                 if (!mutex_trylock(&c->tnc_mutex)) {
 189                         mutex_unlock(&c->umount_mutex);
 190                         *contention = 1;
 191                         p = p->next;
 192                         continue;
 193                 }
 194                 spin_unlock(&ubifs_infos_lock);
 195                 /*
 196                  * OK, now we have TNC locked, the file-system cannot go away -
 197                  * it is safe to reap the cache.
 198                  */
 199                 c->shrinker_run_no = run_no;
 200                 freed += shrink_tnc(c, nr, age, contention);
 201                 mutex_unlock(&c->tnc_mutex);
 202                 spin_lock(&ubifs_infos_lock);
 203                 /* Get the next list element before we move this one */
 204                 p = p->next;
 205                 /*
 206                  * Move this one to the end of the list to provide some
 207                  * fairness.
 208                  */
 209                 list_del(&c->infos_list);
 210                 list_add_tail(&c->infos_list, &ubifs_infos);
 211                 mutex_unlock(&c->umount_mutex);
 212                 if (freed >= nr)
 213                         break;
 214         }
 215         spin_unlock(&ubifs_infos_lock);
 216         return freed;
 217 }
 218
 219 /**
 220  * kick_a_thread - kick a background thread to start commit.
 221  *
 222  * This function kicks a background thread to start background commit. Returns
 223  * %-1 if a thread was kicked or there is another reason to assume the memory
 224  * will soon be freed or become freeable. If there are no dirty znodes, returns
 225  * %0.
 226  */
 227 static int kick_a_thread(void)
 228 {
 229         int i;
 230         struct ubifs_info *c;
 231
 232         /*
 233          * Iterate over all mounted UBIFS file-systems and find out if there is
 234          * already an ongoing commit operation there. If no, then iterate for
 235          * the second time and initiate background commit.
 236          */
 237         spin_lock(&ubifs_infos_lock);
 238         for (i = 0; i < 2; i++) {
 239                 list_for_each_entry(c, &ubifs_infos, infos_list) {
 240                         long dirty_zn_cnt;
 241
 242                         if (!mutex_trylock(&c->umount_mutex)) {
 243                                 /*
 244                                  * Some un-mount is in progress, it will
 245                                  * certainly free memory, so just return.
 246                                  */
 247                                 spin_unlock(&ubifs_infos_lock);
 248                                 return -1;
 249                         }
 250
 251                         dirty_zn_cnt = atomic_long_read(&c->dirty_zn_cnt);
 252
 253                         if (!dirty_zn_cnt || c->cmt_state == COMMIT_BROKEN ||
 254                             c->ro_media) {
 255                                 mutex_unlock(&c->umount_mutex);
 256                                 continue;
 257                         }
 258
 259                         if (c->cmt_state != COMMIT_RESTING) {
 260                                 spin_unlock(&ubifs_infos_lock);
 261                                 mutex_unlock(&c->umount_mutex);
 262                                 return -1;
 263                         }
 264
 265                         if (i == 1) {
 266                                 list_del(&c->infos_list);
 267                                 list_add_tail(&c->infos_list, &ubifs_infos);
 268                                 spin_unlock(&ubifs_infos_lock);
 269
 270                                 ubifs_request_bg_commit(c);
 271                                 mutex_unlock(&c->umount_mutex);
 272                                 return -1;
 273                         }
 274                         mutex_unlock(&c->umount_mutex);
 275                 }
 276         }
 277         spin_unlock(&ubifs_infos_lock);
 278
 279         return 0;
 280 }
 281
 282 int ubifs_shrinker(int nr, gfp_t gfp_mask)
 283 {
 284         int freed, contention = 0;
 285         long clean_zn_cnt = atomic_long_read(&ubifs_clean_zn_cnt);
 286
 287         if (nr == 0)
 288                 return clean_zn_cnt;
 289
 290         if (!clean_zn_cnt) {
 291                 /*
 292                  * No clean znodes, nothing to reap. All we can do in this case
 293                  * is to kick background threads to start commit, which will
 294                  * probably make clean znodes which, in turn, will be freeable.
 295                  * And we return -1 which means will make VM call us again
 296                  * later.
 297                  */
 298                 dbg_tnc("no clean znodes, kick a thread");
 299                 return kick_a_thread();
 300         }
 301
 302         freed = shrink_tnc_trees(nr, OLD_ZNODE_AGE, &contention);
 303         if (freed >= nr)
 304                 goto out;
 305
 306         dbg_tnc("not enough old znodes, try to free young ones");
 307         freed += shrink_tnc_trees(nr - freed, YOUNG_ZNODE_AGE, &contention);
 308         if (freed >= nr)
 309                 goto out;
 310
 311         dbg_tnc("not enough young znodes, free all");
 312         freed += shrink_tnc_trees(nr - freed, 0, &contention);
 313
 314         if (!freed && contention) {
 315                 dbg_tnc("freed nothing, but contention");
 316                 return -1;
 317         }
 318
 319 out:
 320         dbg_tnc("%d znodes were freed, requested %d", freed, nr);
 321         return freed;
 322 }