fs/ubifs/gc.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: Adrian Hunter
  20  *          Artem Bityutskiy (Битюцкий Артём)
  21  */
  22
  23 /*
  24  * This file implements garbage collection. The procedure for garbage collection
  25  * is different depending on whether a LEB as an index LEB (contains index
  26  * nodes) or not. For non-index LEBs, garbage collection finds a LEB which
  27  * contains a lot of dirty space (obsolete nodes), and copies the non-obsolete
  28  * nodes to the journal, at which point the garbage-collected LEB is free to be
  29  * reused. For index LEBs, garbage collection marks the non-obsolete index nodes
  30  * dirty in the TNC, and after the next commit, the garbage-collected LEB is
  31  * to be reused. Garbage collection will cause the number of dirty index nodes
  32  * to grow, however sufficient space is reserved for the index to ensure the
  33  * commit will never run out of space.
  34  *
  35  * Notes about dead watermark. At current UBIFS implementation we assume that
  36  * LEBs which have less than @c->dead_wm bytes of free + dirty space are full
  37  * and not worth garbage-collecting. The dead watermark is one min. I/O unit
  38  * size, or min. UBIFS node size, depending on what is greater. Indeed, UBIFS
  39  * Garbage Collector has to synchronize the GC head's write buffer before
  40  * returning, so this is about wasting one min. I/O unit. However, UBIFS GC can
  41  * actually reclaim even very small pieces of dirty space by garbage collecting
  42  * enough dirty LEBs, but we do not bother doing this at this implementation.
  43  *
  44  * Notes about dark watermark. The results of GC work depends on how big are
  45  * the UBIFS nodes GC deals with. Large nodes make GC waste more space. Indeed,
  46  * if GC move data from LEB A to LEB B and nodes in LEB A are large, GC would
  47  * have to waste large pieces of free space at the end of LEB B, because nodes
  48  * from LEB A would not fit. And the worst situation is when all nodes are of
  49  * maximum size. So dark watermark is the amount of free + dirty space in LEB
  50  * which are guaranteed to be reclaimable. If LEB has less space, the GC might
  51  * be unable to reclaim it. So, LEBs with free + dirty greater than dark
  52  * watermark are "good" LEBs from GC's point of few. The other LEBs are not so
  53  * good, and GC takes extra care when moving them.
  54  */
  55
  56 #include <linux/slab.h>
  57 #include <linux/pagemap.h>
  58 #include <linux/list_sort.h>
  59 #include "ubifs.h"
  60
  61 /*
  62  * GC may need to move more than one LEB to make progress. The below constants
  63  * define "soft" and "hard" limits on the number of LEBs the garbage collector
  64  * may move.
  65  */
  66 #define SOFT_LEBS_LIMIT 4
  67 #define HARD_LEBS_LIMIT 32
  68
  69 /**
  70  * switch_gc_head - switch the garbage collection journal head.
  71  * @c: UBIFS file-system description object
  72  * @buf: buffer to write
  73  * @len: length of the buffer to write
  74  * @lnum: LEB number written is returned here
  75  * @offs: offset written is returned here
  76  *
  77  * This function switch the GC head to the next LEB which is reserved in
  78  * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required,
  79  * and other negative error code in case of failures.
  80  */
  81 static int switch_gc_head(struct ubifs_info *c)
  82 {
  83         int err, gc_lnum = c->gc_lnum;
  84         struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
  85
  86         ubifs_assert(gc_lnum != -1);
  87         dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)",
  88                wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum,
  89                c->leb_size - wbuf->offs - wbuf->used);
  90
  91         err = ubifs_wbuf_sync_nolock(wbuf);
  92         if (err)
  93                 return err;
  94
  95         /*
  96          * The GC write-buffer was synchronized, we may safely unmap
  97          * 'c->gc_lnum'.
  98          */
  99         err = ubifs_leb_unmap(c, gc_lnum);
 100         if (err)
 101                 return err;
 102
 103         err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0);
 104         if (err)
 105                 return err;
 106
 107         c->gc_lnum = -1;
 108         err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0, UBI_LONGTERM);
 109         return err;
 110 }
 111
 112 /**
 113  * data_nodes_cmp - compare 2 data nodes.
 114  * @priv: UBIFS file-system description object
 115  * @a: first data node
 116  * @a: second data node
 117  *
 118  * This function compares data nodes @a and @b. Returns %1 if @a has greater
 119  * inode or block number, and %-1 otherwise.
 120  */
 121 int data_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
 122 {
 123         ino_t inuma, inumb;
 124         struct ubifs_info *c = priv;
 125         struct ubifs_scan_node *sa, *sb;
 126
 127         cond_resched();
 128         sa = list_entry(a, struct ubifs_scan_node, list);
 129         sb = list_entry(b, struct ubifs_scan_node, list);
 130         ubifs_assert(key_type(c, &sa->key) == UBIFS_DATA_KEY);
 131         ubifs_assert(key_type(c, &sb->key) == UBIFS_DATA_KEY);
 132
 133         inuma = key_inum(c, &sa->key);
 134         inumb = key_inum(c, &sb->key);
 135
 136         if (inuma == inumb) {
 137                 unsigned int blka = key_block(c, &sa->key);
 138                 unsigned int blkb = key_block(c, &sb->key);
 139
 140                 if (blka <= blkb)
 141                         return -1;
 142         } else if (inuma <= inumb)
 143                 return -1;
 144
 145         return 1;
 146 }
 147
 148 /*
 149  * nondata_nodes_cmp - compare 2 non-data nodes.
 150  * @priv: UBIFS file-system description object
 151  * @a: first node
 152  * @a: second node
 153  *
 154  * This function compares nodes @a and @b. It makes sure that inode nodes go
 155  * first and sorted by length in descending order. Directory entry nodes go
 156  * after inode nodes and are sorted in ascending hash valuer order.
 157  */
 158 int nondata_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
 159 {
 160         int typea, typeb;
 161         ino_t inuma, inumb;
 162         struct ubifs_info *c = priv;
 163         struct ubifs_scan_node *sa, *sb;
 164
 165         cond_resched();
 166         sa = list_entry(a, struct ubifs_scan_node, list);
 167         sb = list_entry(b, struct ubifs_scan_node, list);
 168         typea = key_type(c, &sa->key);
 169         typeb = key_type(c, &sb->key);
 170         ubifs_assert(typea != UBIFS_DATA_KEY && typeb != UBIFS_DATA_KEY);
 171
 172         /* Inodes go before directory entries */
 173         if (typea == UBIFS_INO_KEY) {
 174                 if (typeb == UBIFS_INO_KEY)
 175                         return sb->len - sa->len;
 176                 return -1;
 177         }
 178         if (typeb == UBIFS_INO_KEY)
 179                 return 1;
 180
 181         ubifs_assert(typea == UBIFS_DENT_KEY && typeb == UBIFS_DENT_KEY);
 182         inuma = key_inum(c, &sa->key);
 183         inumb = key_inum(c, &sb->key);
 184
 185         if (inuma == inumb) {
 186                 uint32_t hasha = key_hash(c, &sa->key);
 187                 uint32_t hashb = key_hash(c, &sb->key);
 188
 189                 if (hasha <= hashb)
 190                         return -1;
 191         } else if (inuma <= inumb)
 192                 return -1;
 193
 194         return 1;
 195 }
 196
 197 /**
 198  * sort_nodes - sort nodes for GC.
 199  * @c: UBIFS file-system description object
 200  * @sleb: describes nodes to sort and contains the result on exit
 201  * @nondata: contains non-data nodes on exit
 202  * @min: minimum node size is returned here
 203  *
 204  * This function sorts the list of inodes to garbage collect. First of all, it
 205  * kills obsolete nodes and separates data and non-data nodes to the
 206  * @sleb->nodes and @nondata lists correspondingly.
 207  *
 208  * Data nodes are then sorted in block number order - this is important for
 209  * bulk-read; data nodes with lower inode number go before data nodes with
 210  * higher inode number, and data nodes with lower block number go before data
 211  * nodes with higher block number;
 212  *
 213  * Non-data nodes are sorted as follows.
 214  *   o First go inode nodes - they are sorted in descending length order.
 215  *   o Then go directory entry nodes - they are sorted in hash order, which
 216  *     should supposedly optimize 'readdir()'. Direntry nodes with lower parent
 217  *     inode number go before direntry nodes with higher parent inode number,
 218  *     and direntry nodes with lower name hash values go before direntry nodes
 219  *     with higher name hash values.
 220  *
 221  * This function returns zero in case of success and a negative error code in
 222  * case of failure.
 223  */
 224 static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
 225                       struct list_head *nondata, int *min)
 226 {
 227         struct ubifs_scan_node *snod, *tmp;
 228
 229         *min = INT_MAX;
 230
 231         /* Separate data nodes and non-data nodes */
 232         list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
 233                 int err;
 234
 235                 ubifs_assert(snod->type != UBIFS_IDX_NODE);
 236                 ubifs_assert(snod->type != UBIFS_REF_NODE);
 237                 ubifs_assert(snod->type != UBIFS_CS_NODE);
 238
 239                 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
 240                                          snod->offs, 0);
 241                 if (err < 0)
 242                         return err;
 243
 244                 if (!err) {
 245                         /* The node is obsolete, remove it from the list */
 246                         list_del(&snod->list);
 247                         kfree(snod);
 248                         continue;
 249                 }
 250
 251                 if (snod->len < *min)
 252                         *min = snod->len;
 253
 254                 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
 255                         list_move_tail(&snod->list, nondata);
 256         }
 257
 258         /* Sort data and non-data nodes */
 259         list_sort(c, &sleb->nodes, &data_nodes_cmp);
 260         list_sort(c, nondata, &nondata_nodes_cmp);
 261         return 0;
 262 }
 263
 264 /**
 265  * move_node - move a node.
 266  * @c: UBIFS file-system description object
 267  * @sleb: describes the LEB to move nodes from
 268  * @snod: the mode to move
 269  * @wbuf: write-buffer to move node to
 270  *
 271  * This function moves node @snod to @wbuf, changes TNC correspondingly, and
 272  * destroys @snod. Returns zero in case of success and a negative error code in
 273  * case of failure.
 274  */
 275 static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
 276                      struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
 277 {
 278         int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
 279
 280         cond_resched();
 281         err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
 282         if (err)
 283                 return err;
 284
 285         err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
 286                                 snod->offs, new_lnum, new_offs,
 287                                 snod->len);
 288         list_del(&snod->list);
 289         kfree(snod);
 290         return err;
 291 }
 292
 293 /**
 294  * move_nodes - move nodes.
 295  * @c: UBIFS file-system description object
 296  * @sleb: describes the LEB to move nodes from
 297  *
 298  * This function moves valid nodes from data LEB described by @sleb to the GC
 299  * journal head. This function returns zero in case of success, %-EAGAIN if
 300  * commit is required, and other negative error codes in case of other
 301  * failures.
 302  */
 303 static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
 304 {
 305         int err, min;
 306         LIST_HEAD(nondata);
 307         struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
 308
 309         if (wbuf->lnum == -1) {
 310                 /*
 311                  * The GC journal head is not set, because it is the first GC
 312                  * invocation since mount.
 313                  */
 314                 err = switch_gc_head(c);
 315                 if (err)
 316                         return err;
 317         }
 318
 319         err = sort_nodes(c, sleb, &nondata, &min);
 320         if (err)
 321                 goto out;
 322
 323         /* Write nodes to their new location. Use the first-fit strategy */
 324         while (1) {
 325                 int avail;
 326                 struct ubifs_scan_node *snod, *tmp;
 327
 328                 /* Move data nodes */
 329                 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
 330                         avail = c->leb_size - wbuf->offs - wbuf->used;
 331                         if  (snod->len > avail)
 332                                 /*
 333                                  * Do not skip data nodes in order to optimize
 334                                  * bulk-read.
 335                                  */
 336                                 break;
 337
 338                         err = move_node(c, sleb, snod, wbuf);
 339                         if (err)
 340                                 goto out;
 341                 }
 342
 343                 /* Move non-data nodes */
 344                 list_for_each_entry_safe(snod, tmp, &nondata, list) {
 345                         avail = c->leb_size - wbuf->offs - wbuf->used;
 346                         if (avail < min)
 347                                 break;
 348
 349                         if  (snod->len > avail) {
 350                                 /*
 351                                  * Keep going only if this is an inode with
 352                                  * some data. Otherwise stop and switch the GC
 353                                  * head. IOW, we assume that data-less inode
 354                                  * nodes and direntry nodes are roughly of the
 355                                  * same size.
 356                                  */
 357                                 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
 358                                     snod->len == UBIFS_INO_NODE_SZ)
 359                                         break;
 360                                 continue;
 361                         }
 362
 363                         err = move_node(c, sleb, snod, wbuf);
 364                         if (err)
 365                                 goto out;
 366                 }
 367
 368                 if (list_empty(&sleb->nodes) && list_empty(&nondata))
 369                         break;
 370
 371                 /*
 372                  * Waste the rest of the space in the LEB and switch to the
 373                  * next LEB.
 374                  */
 375                 err = switch_gc_head(c);
 376                 if (err)
 377                         goto out;
 378         }
 379
 380         return 0;
 381
 382 out:
 383         list_splice_tail(&nondata, &sleb->nodes);
 384         return err;
 385 }
 386
 387 /**
 388  * gc_sync_wbufs - sync write-buffers for GC.
 389  * @c: UBIFS file-system description object
 390  *
 391  * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
 392  * be in a write-buffer instead. That is, a node could be written to a
 393  * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
 394  * erased before the write-buffer is sync'd and then there is an unclean
 395  * unmount, then an existing node is lost. To avoid this, we sync all
 396  * write-buffers.
 397  *
 398  * This function returns %0 on success or a negative error code on failure.
 399  */
 400 static int gc_sync_wbufs(struct ubifs_info *c)
 401 {
 402         int err, i;
 403
 404         for (i = 0; i < c->jhead_cnt; i++) {
 405                 if (i == GCHD)
 406                         continue;
 407                 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
 408                 if (err)
 409                         return err;
 410         }
 411         return 0;
 412 }
 413
 414 /**
 415  * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
 416  * @c: UBIFS file-system description object
 417  * @lp: describes the LEB to garbage collect
 418  *
 419  * This function garbage-collects an LEB and returns one of the @LEB_FREED,
 420  * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
 421  * required, and other negative error codes in case of failures.
 422  */
 423 int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
 424 {
 425         struct ubifs_scan_leb *sleb;
 426         struct ubifs_scan_node *snod;
 427         struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
 428         int err = 0, lnum = lp->lnum;
 429
 430         ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
 431                      c->need_recovery);
 432         ubifs_assert(c->gc_lnum != lnum);
 433         ubifs_assert(wbuf->lnum != lnum);
 434
 435         /*
 436          * We scan the entire LEB even though we only really need to scan up to
 437          * (c->leb_size - lp->free).
 438          */
 439         sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
 440         if (IS_ERR(sleb))
 441                 return PTR_ERR(sleb);
 442
 443         ubifs_assert(!list_empty(&sleb->nodes));
 444         snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
 445
 446         if (snod->type == UBIFS_IDX_NODE) {
 447                 struct ubifs_gced_idx_leb *idx_gc;
 448
 449                 dbg_gc("indexing LEB %d (free %d, dirty %d)",
 450                        lnum, lp->free, lp->dirty);
 451                 list_for_each_entry(snod, &sleb->nodes, list) {
 452                         struct ubifs_idx_node *idx = snod->node;
 453                         int level = le16_to_cpu(idx->level);
 454
 455                         ubifs_assert(snod->type == UBIFS_IDX_NODE);
 456                         key_read(c, ubifs_idx_key(c, idx), &snod->key);
 457                         err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
 458                                                    snod->offs);
 459                         if (err)
 460                                 goto out;
 461                 }
 462
 463                 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
 464                 if (!idx_gc) {
 465                         err = -ENOMEM;
 466                         goto out;
 467                 }
 468
 469                 idx_gc->lnum = lnum;
 470                 idx_gc->unmap = 0;
 471                 list_add(&idx_gc->list, &c->idx_gc);
 472
 473                 /*
 474                  * Don't release the LEB until after the next commit, because
 475                  * it may contain data which is needed for recovery. So
 476                  * although we freed this LEB, it will become usable only after
 477                  * the commit.
 478                  */
 479                 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
 480                                           LPROPS_INDEX, 1);
 481                 if (err)
 482                         goto out;
 483                 err = LEB_FREED_IDX;
 484         } else {
 485                 dbg_gc("data LEB %d (free %d, dirty %d)",
 486                        lnum, lp->free, lp->dirty);
 487
 488                 err = move_nodes(c, sleb);
 489                 if (err)
 490                         goto out_inc_seq;
 491
 492                 err = gc_sync_wbufs(c);
 493                 if (err)
 494                         goto out_inc_seq;
 495
 496                 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
 497                 if (err)
 498                         goto out_inc_seq;
 499
 500                 /* Allow for races with TNC */
 501                 c->gced_lnum = lnum;
 502                 smp_wmb();
 503                 c->gc_seq += 1;
 504                 smp_wmb();
 505
 506                 if (c->gc_lnum == -1) {
 507                         c->gc_lnum = lnum;
 508                         err = LEB_RETAINED;
 509                 } else {
 510                         err = ubifs_wbuf_sync_nolock(wbuf);
 511                         if (err)
 512                                 goto out;
 513
 514                         err = ubifs_leb_unmap(c, lnum);
 515                         if (err)
 516                                 goto out;
 517
 518                         err = LEB_FREED;
 519                 }
 520         }
 521
 522 out:
 523         ubifs_scan_destroy(sleb);
 524         return err;
 525
 526 out_inc_seq:
 527         /* We may have moved at least some nodes so allow for races with TNC */
 528         c->gced_lnum = lnum;
 529         smp_wmb();
 530         c->gc_seq += 1;
 531         smp_wmb();
 532         goto out;
 533 }
 534
 535 /**
 536  * ubifs_garbage_collect - UBIFS garbage collector.
 537  * @c: UBIFS file-system description object
 538  * @anyway: do GC even if there are free LEBs
 539  *
 540  * This function does out-of-place garbage collection. The return codes are:
 541  *   o positive LEB number if the LEB has been freed and may be used;
 542  *   o %-EAGAIN if the caller has to run commit;
 543  *   o %-ENOSPC if GC failed to make any progress;
 544  *   o other negative error codes in case of other errors.
 545  *
 546  * Garbage collector writes data to the journal when GC'ing data LEBs, and just
 547  * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
 548  * commit may be required. But commit cannot be run from inside GC, because the
 549  * caller might be holding the commit lock, so %-EAGAIN is returned instead;
 550  * And this error code means that the caller has to run commit, and re-run GC
 551  * if there is still no free space.
 552  *
 553  * There are many reasons why this function may return %-EAGAIN:
 554  * o the log is full and there is no space to write an LEB reference for
 555  *   @c->gc_lnum;
 556  * o the journal is too large and exceeds size limitations;
 557  * o GC moved indexing LEBs, but they can be used only after the commit;
 558  * o the shrinker fails to find clean znodes to free and requests the commit;
 559  * o etc.
 560  *
 561  * Note, if the file-system is close to be full, this function may return
 562  * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
 563  * the function. E.g., this happens if the limits on the journal size are too
 564  * tough and GC writes too much to the journal before an LEB is freed. This
 565  * might also mean that the journal is too large, and the TNC becomes to big,
 566  * so that the shrinker is constantly called, finds not clean znodes to free,
 567  * and requests commit. Well, this may also happen if the journal is all right,
 568  * but another kernel process consumes too much memory. Anyway, infinite
 569  * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
 570  */
 571 int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
 572 {
 573         int i, err, ret, min_space = c->dead_wm;
 574         struct ubifs_lprops lp;
 575         struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
 576
 577         ubifs_assert_cmt_locked(c);
 578
 579         if (ubifs_gc_should_commit(c))
 580                 return -EAGAIN;
 581
 582         mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
 583
 584         if (c->ro_media) {
 585                 ret = -EROFS;
 586                 goto out_unlock;
 587         }
 588
 589         /* We expect the write-buffer to be empty on entry */
 590         ubifs_assert(!wbuf->used);
 591
 592         for (i = 0; ; i++) {
 593                 int space_before = c->leb_size - wbuf->offs - wbuf->used;
 594                 int space_after;
 595
 596                 cond_resched();
 597
 598                 /* Give the commit an opportunity to run */
 599                 if (ubifs_gc_should_commit(c)) {
 600                         ret = -EAGAIN;
 601                         break;
 602                 }
 603
 604                 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
 605                         /*
 606                          * We've done enough iterations. Indexing LEBs were
 607                          * moved and will be available after the commit.
 608                          */
 609                         dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
 610                         ubifs_commit_required(c);
 611                         ret = -EAGAIN;
 612                         break;
 613                 }
 614
 615                 if (i > HARD_LEBS_LIMIT) {
 616                         /*
 617                          * We've moved too many LEBs and have not made
 618                          * progress, give up.
 619                          */
 620                         dbg_gc("hard limit, -ENOSPC");
 621                         ret = -ENOSPC;
 622                         break;
 623                 }
 624
 625                 /*
 626                  * Empty and freeable LEBs can turn up while we waited for
 627                  * the wbuf lock, or while we have been running GC. In that
 628                  * case, we should just return one of those instead of
 629                  * continuing to GC dirty LEBs. Hence we request
 630                  * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
 631                  */
 632                 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
 633                 if (ret) {
 634                         if (ret == -ENOSPC)
 635                                 dbg_gc("no more dirty LEBs");
 636                         break;
 637                 }
 638
 639                 dbg_gc("found LEB %d: free %d, dirty %d, sum %d "
 640                        "(min. space %d)", lp.lnum, lp.free, lp.dirty,
 641                        lp.free + lp.dirty, min_space);
 642
 643                 if (lp.free + lp.dirty == c->leb_size) {
 644                         /* An empty LEB was returned */
 645                         dbg_gc("LEB %d is free, return it", lp.lnum);
 646                         /*
 647                          * ubifs_find_dirty_leb() doesn't return freeable index
 648                          * LEBs.
 649                          */
 650                         ubifs_assert(!(lp.flags & LPROPS_INDEX));
 651                         if (lp.free != c->leb_size) {
 652                                 /*
 653                                  * Write buffers must be sync'd before
 654                                  * unmapping freeable LEBs, because one of them
 655                                  * may contain data which obsoletes something
 656                                  * in 'lp.pnum'.
 657                                  */
 658                                 ret = gc_sync_wbufs(c);
 659                                 if (ret)
 660                                         goto out;
 661                                 ret = ubifs_change_one_lp(c, lp.lnum,
 662                                                           c->leb_size, 0, 0, 0,
 663                                                           0);
 664                                 if (ret)
 665                                         goto out;
 666                         }
 667                         ret = ubifs_leb_unmap(c, lp.lnum);
 668                         if (ret)
 669                                 goto out;
 670                         ret = lp.lnum;
 671                         break;
 672                 }
 673
 674                 space_before = c->leb_size - wbuf->offs - wbuf->used;
 675                 if (wbuf->lnum == -1)
 676                         space_before = 0;
 677
 678                 ret = ubifs_garbage_collect_leb(c, &lp);
 679                 if (ret < 0) {
 680                         if (ret == -EAGAIN || ret == -ENOSPC) {
 681                                 /*
 682                                  * These codes are not errors, so we have to
 683                                  * return the LEB to lprops. But if the
 684                                  * 'ubifs_return_leb()' function fails, its
 685                                  * failure code is propagated to the caller
 686                                  * instead of the original '-EAGAIN' or
 687                                  * '-ENOSPC'.
 688                                  */
 689                                 err = ubifs_return_leb(c, lp.lnum);
 690                                 if (err)
 691                                         ret = err;
 692                                 break;
 693                         }
 694                         goto out;
 695                 }
 696
 697                 if (ret == LEB_FREED) {
 698                         /* An LEB has been freed and is ready for use */
 699                         dbg_gc("LEB %d freed, return", lp.lnum);
 700                         ret = lp.lnum;
 701                         break;
 702                 }
 703
 704                 if (ret == LEB_FREED_IDX) {
 705                         /*
 706                          * This was an indexing LEB and it cannot be
 707                          * immediately used. And instead of requesting the
 708                          * commit straight away, we try to garbage collect some
 709                          * more.
 710                          */
 711                         dbg_gc("indexing LEB %d freed, continue", lp.lnum);
 712                         continue;
 713                 }
 714
 715                 ubifs_assert(ret == LEB_RETAINED);
 716                 space_after = c->leb_size - wbuf->offs - wbuf->used;
 717                 dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
 718                        space_after - space_before);
 719
 720                 if (space_after > space_before) {
 721                         /* GC makes progress, keep working */
 722                         min_space >>= 1;
 723                         if (min_space < c->dead_wm)
 724                                 min_space = c->dead_wm;
 725                         continue;
 726                 }
 727
 728                 dbg_gc("did not make progress");
 729
 730                 /*
 731                  * GC moved an LEB bud have not done any progress. This means
 732                  * that the previous GC head LEB contained too few free space
 733                  * and the LEB which was GC'ed contained only large nodes which
 734                  * did not fit that space.
 735                  *
 736                  * We can do 2 things:
 737                  * 1. pick another LEB in a hope it'll contain a small node
 738                  *    which will fit the space we have at the end of current GC
 739                  *    head LEB, but there is no guarantee, so we try this out
 740                  *    unless we have already been working for too long;
 741                  * 2. request an LEB with more dirty space, which will force
 742                  *    'ubifs_find_dirty_leb()' to start scanning the lprops
 743                  *    table, instead of just picking one from the heap
 744                  *    (previously it already picked the dirtiest LEB).
 745                  */
 746                 if (i < SOFT_LEBS_LIMIT) {
 747                         dbg_gc("try again");
 748                         continue;
 749                 }
 750
 751                 min_space <<= 1;
 752                 if (min_space > c->dark_wm)
 753                         min_space = c->dark_wm;
 754                 dbg_gc("set min. space to %d", min_space);
 755         }
 756
 757         if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
 758                 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
 759                 ubifs_commit_required(c);
 760                 ret = -EAGAIN;
 761         }
 762
 763         err = ubifs_wbuf_sync_nolock(wbuf);
 764         if (!err)
 765                 err = ubifs_leb_unmap(c, c->gc_lnum);
 766         if (err) {
 767                 ret = err;
 768                 goto out;
 769         }
 770 out_unlock:
 771         mutex_unlock(&wbuf->io_mutex);
 772         return ret;
 773
 774 out:
 775         ubifs_assert(ret < 0);
 776         ubifs_assert(ret != -ENOSPC && ret != -EAGAIN);
 777         ubifs_ro_mode(c, ret);
 778         ubifs_wbuf_sync_nolock(wbuf);
 779         mutex_unlock(&wbuf->io_mutex);
 780         ubifs_return_leb(c, lp.lnum);
 781         return ret;
 782 }
 783
 784 /**
 785  * ubifs_gc_start_commit - garbage collection at start of commit.
 786  * @c: UBIFS file-system description object
 787  *
 788  * If a LEB has only dirty and free space, then we may safely unmap it and make
 789  * it free.  Note, we cannot do this with indexing LEBs because dirty space may
 790  * correspond index nodes that are required for recovery.  In that case, the
 791  * LEB cannot be unmapped until after the next commit.
 792  *
 793  * This function returns %0 upon success and a negative error code upon failure.
 794  */
 795 int ubifs_gc_start_commit(struct ubifs_info *c)
 796 {
 797         struct ubifs_gced_idx_leb *idx_gc;
 798         const struct ubifs_lprops *lp;
 799         int err = 0, flags;
 800
 801         ubifs_get_lprops(c);
 802
 803         /*
 804          * Unmap (non-index) freeable LEBs. Note that recovery requires that all
 805          * wbufs are sync'd before this, which is done in 'do_commit()'.
 806          */
 807         while (1) {
 808                 lp = ubifs_fast_find_freeable(c);
 809                 if (IS_ERR(lp)) {
 810                         err = PTR_ERR(lp);
 811                         goto out;
 812                 }
 813                 if (!lp)
 814                         break;
 815                 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
 816                 ubifs_assert(!(lp->flags & LPROPS_INDEX));
 817                 err = ubifs_leb_unmap(c, lp->lnum);
 818                 if (err)
 819                         goto out;
 820                 lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
 821                 if (IS_ERR(lp)) {
 822                         err = PTR_ERR(lp);
 823                         goto out;
 824                 }
 825                 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
 826                 ubifs_assert(!(lp->flags & LPROPS_INDEX));
 827         }
 828
 829         /* Mark GC'd index LEBs OK to unmap after this commit finishes */
 830         list_for_each_entry(idx_gc, &c->idx_gc, list)
 831                 idx_gc->unmap = 1;
 832
 833         /* Record index freeable LEBs for unmapping after commit */
 834         while (1) {
 835                 lp = ubifs_fast_find_frdi_idx(c);
 836                 if (IS_ERR(lp)) {
 837                         err = PTR_ERR(lp);
 838                         goto out;
 839                 }
 840                 if (!lp)
 841                         break;
 842                 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
 843                 if (!idx_gc) {
 844                         err = -ENOMEM;
 845                         goto out;
 846                 }
 847                 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
 848                 ubifs_assert(lp->flags & LPROPS_INDEX);
 849                 /* Don't release the LEB until after the next commit */
 850                 flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
 851                 lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
 852                 if (IS_ERR(lp)) {
 853                         err = PTR_ERR(lp);
 854                         kfree(idx_gc);
 855                         goto out;
 856                 }
 857                 ubifs_assert(lp->flags & LPROPS_TAKEN);
 858                 ubifs_assert(!(lp->flags & LPROPS_INDEX));
 859                 idx_gc->lnum = lp->lnum;
 860                 idx_gc->unmap = 1;
 861                 list_add(&idx_gc->list, &c->idx_gc);
 862         }
 863 out:
 864         ubifs_release_lprops(c);
 865         return err;
 866 }
 867
 868 /**
 869  * ubifs_gc_end_commit - garbage collection at end of commit.
 870  * @c: UBIFS file-system description object
 871  *
 872  * This function completes out-of-place garbage collection of index LEBs.
 873  */
 874 int ubifs_gc_end_commit(struct ubifs_info *c)
 875 {
 876         struct ubifs_gced_idx_leb *idx_gc, *tmp;
 877         struct ubifs_wbuf *wbuf;
 878         int err = 0;
 879
 880         wbuf = &c->jheads[GCHD].wbuf;
 881         mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
 882         list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
 883                 if (idx_gc->unmap) {
 884                         dbg_gc("LEB %d", idx_gc->lnum);
 885                         err = ubifs_leb_unmap(c, idx_gc->lnum);
 886                         if (err)
 887                                 goto out;
 888                         err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
 889                                           LPROPS_NC, 0, LPROPS_TAKEN, -1);
 890                         if (err)
 891                                 goto out;
 892                         list_del(&idx_gc->list);
 893                         kfree(idx_gc);
 894                 }
 895 out:
 896         mutex_unlock(&wbuf->io_mutex);
 897         return err;
 898 }
 899
 900 /**
 901  * ubifs_destroy_idx_gc - destroy idx_gc list.
 902  * @c: UBIFS file-system description object
 903  *
 904  * This function destroys the @c->idx_gc list. It is called when unmounting
 905  * so locks are not needed. Returns zero in case of success and a negative
 906  * error code in case of failure.
 907  */
 908 void ubifs_destroy_idx_gc(struct ubifs_info *c)
 909 {
 910         while (!list_empty(&c->idx_gc)) {
 911                 struct ubifs_gced_idx_leb *idx_gc;
 912
 913                 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
 914                                     list);
 915                 c->idx_gc_cnt -= 1;
 916                 list_del(&idx_gc->list);
 917                 kfree(idx_gc);
 918         }
 919 }
 920
 921 /**
 922  * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
 923  * @c: UBIFS file-system description object
 924  *
 925  * Called during start commit so locks are not needed.
 926  */
 927 int ubifs_get_idx_gc_leb(struct ubifs_info *c)
 928 {
 929         struct ubifs_gced_idx_leb *idx_gc;
 930         int lnum;
 931
 932         if (list_empty(&c->idx_gc))
 933                 return -ENOSPC;
 934         idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
 935         lnum = idx_gc->lnum;
 936         /* c->idx_gc_cnt is updated by the caller when lprops are updated */
 937         list_del(&idx_gc->list);
 938         kfree(idx_gc);
 939         return lnum;
 940 }