blob: 151f108828204dc4d88c9d890fbcce051b2d8912 [file] [log] [blame]
Artem Bityutskiy1e517642008-07-14 19:08:37 +03001/*
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.
Artem Bityutskiy70782022009-01-19 19:57:27 +020034 *
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
Artem Bityutskiyf10770f2009-03-08 15:13:00 +020050 * which are guaranteed to be reclaimable. If LEB has less space, the GC might
Artem Bityutskiy70782022009-01-19 19:57:27 +020051 * 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.
Artem Bityutskiy1e517642008-07-14 19:08:37 +030054 */
55
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090056#include <linux/slab.h>
Artem Bityutskiy1e517642008-07-14 19:08:37 +030057#include <linux/pagemap.h>
Dave Chinner2c761272010-01-12 17:39:16 +110058#include <linux/list_sort.h>
Artem Bityutskiy1e517642008-07-14 19:08:37 +030059#include "ubifs.h"
60
61/*
Frederik Schwarzer025dfda2008-10-16 19:02:37 +020062 * GC may need to move more than one LEB to make progress. The below constants
Artem Bityutskiy1e517642008-07-14 19:08:37 +030063 * 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 */
81static 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/**
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200113 * data_nodes_cmp - compare 2 data nodes.
114 * @priv: UBIFS file-system description object
115 * @a: first data node
116 * @a: second data node
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300117 *
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200118 * This function compares data nodes @a and @b. Returns %1 if @a has greater
119 * inode or block number, and %-1 otherwise.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300120 */
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200121int 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();
Artem Bityutskiy1a9476a2010-08-08 12:45:23 +0300128 if (a == b)
129 return 0;
130
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200131 sa = list_entry(a, struct ubifs_scan_node, list);
132 sb = list_entry(b, struct ubifs_scan_node, list);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300133
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200134 ubifs_assert(key_type(c, &sa->key) == UBIFS_DATA_KEY);
135 ubifs_assert(key_type(c, &sb->key) == UBIFS_DATA_KEY);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300136 ubifs_assert(sa->type == UBIFS_DATA_NODE);
137 ubifs_assert(sb->type == UBIFS_DATA_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200138
139 inuma = key_inum(c, &sa->key);
140 inumb = key_inum(c, &sb->key);
141
142 if (inuma == inumb) {
143 unsigned int blka = key_block(c, &sa->key);
144 unsigned int blkb = key_block(c, &sb->key);
145
146 if (blka <= blkb)
147 return -1;
148 } else if (inuma <= inumb)
149 return -1;
150
151 return 1;
152}
153
154/*
155 * nondata_nodes_cmp - compare 2 non-data nodes.
156 * @priv: UBIFS file-system description object
157 * @a: first node
158 * @a: second node
159 *
160 * This function compares nodes @a and @b. It makes sure that inode nodes go
161 * first and sorted by length in descending order. Directory entry nodes go
162 * after inode nodes and are sorted in ascending hash valuer order.
163 */
164int nondata_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
165{
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200166 ino_t inuma, inumb;
167 struct ubifs_info *c = priv;
168 struct ubifs_scan_node *sa, *sb;
169
170 cond_resched();
Artem Bityutskiy1a9476a2010-08-08 12:45:23 +0300171 if (a == b)
172 return 0;
173
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200174 sa = list_entry(a, struct ubifs_scan_node, list);
175 sb = list_entry(b, struct ubifs_scan_node, list);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300176
177 ubifs_assert(key_type(c, &sa->key) != UBIFS_DATA_KEY &&
178 key_type(c, &sb->key) != UBIFS_DATA_KEY);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300179 ubifs_assert(sa->type != UBIFS_DATA_NODE &&
180 sb->type != UBIFS_DATA_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200181
182 /* Inodes go before directory entries */
Artem Bityutskiyab871182010-08-08 12:25:33 +0300183 if (sa->type == UBIFS_INO_NODE) {
184 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200185 return sb->len - sa->len;
186 return -1;
187 }
Artem Bityutskiyab871182010-08-08 12:25:33 +0300188 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200189 return 1;
190
Artem Bityutskiy66576832010-08-08 12:29:58 +0300191 ubifs_assert(key_type(c, &sa->key) == UBIFS_DENT_KEY ||
192 key_type(c, &sa->key) == UBIFS_XENT_KEY);
193 ubifs_assert(key_type(c, &sb->key) == UBIFS_DENT_KEY ||
194 key_type(c, &sb->key) == UBIFS_XENT_KEY);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300195 ubifs_assert(sa->type == UBIFS_DENT_NODE ||
196 sa->type == UBIFS_XENT_NODE);
197 ubifs_assert(sb->type == UBIFS_DENT_NODE ||
198 sb->type == UBIFS_XENT_NODE);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300199
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200200 inuma = key_inum(c, &sa->key);
201 inumb = key_inum(c, &sb->key);
202
203 if (inuma == inumb) {
204 uint32_t hasha = key_hash(c, &sa->key);
205 uint32_t hashb = key_hash(c, &sb->key);
206
207 if (hasha <= hashb)
208 return -1;
209 } else if (inuma <= inumb)
210 return -1;
211
212 return 1;
213}
214
215/**
216 * sort_nodes - sort nodes for GC.
217 * @c: UBIFS file-system description object
218 * @sleb: describes nodes to sort and contains the result on exit
219 * @nondata: contains non-data nodes on exit
220 * @min: minimum node size is returned here
221 *
222 * This function sorts the list of inodes to garbage collect. First of all, it
223 * kills obsolete nodes and separates data and non-data nodes to the
224 * @sleb->nodes and @nondata lists correspondingly.
225 *
226 * Data nodes are then sorted in block number order - this is important for
227 * bulk-read; data nodes with lower inode number go before data nodes with
228 * higher inode number, and data nodes with lower block number go before data
229 * nodes with higher block number;
230 *
231 * Non-data nodes are sorted as follows.
232 * o First go inode nodes - they are sorted in descending length order.
233 * o Then go directory entry nodes - they are sorted in hash order, which
234 * should supposedly optimize 'readdir()'. Direntry nodes with lower parent
235 * inode number go before direntry nodes with higher parent inode number,
236 * and direntry nodes with lower name hash values go before direntry nodes
237 * with higher name hash values.
238 *
239 * This function returns zero in case of success and a negative error code in
240 * case of failure.
241 */
242static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
243 struct list_head *nondata, int *min)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300244{
Artem Bityutskiy3bb66b42010-08-07 10:06:11 +0300245 int err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300246 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300247
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200248 *min = INT_MAX;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300249
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200250 /* Separate data nodes and non-data nodes */
251 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
Artem Bityutskiy44ec83b2010-08-07 08:44:13 +0300252 ubifs_assert(snod->type == UBIFS_INO_NODE ||
253 snod->type == UBIFS_DATA_NODE ||
254 snod->type == UBIFS_DENT_NODE ||
255 snod->type == UBIFS_XENT_NODE ||
256 snod->type == UBIFS_TRUN_NODE);
257
258 if (snod->type != UBIFS_INO_NODE &&
259 snod->type != UBIFS_DATA_NODE &&
260 snod->type != UBIFS_DENT_NODE &&
261 snod->type != UBIFS_XENT_NODE) {
262 /* Probably truncation node, zap it */
263 list_del(&snod->list);
264 kfree(snod);
265 continue;
266 }
267
268 ubifs_assert(key_type(c, &snod->key) == UBIFS_DATA_KEY ||
269 key_type(c, &snod->key) == UBIFS_INO_KEY ||
270 key_type(c, &snod->key) == UBIFS_DENT_KEY ||
271 key_type(c, &snod->key) == UBIFS_XENT_KEY);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300272
273 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
274 snod->offs, 0);
275 if (err < 0)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200276 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300277
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300278 if (!err) {
279 /* The node is obsolete, remove it from the list */
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200280 list_del(&snod->list);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300281 kfree(snod);
282 continue;
283 }
284
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200285 if (snod->len < *min)
286 *min = snod->len;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300287
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200288 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
289 list_move_tail(&snod->list, nondata);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300290 }
291
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200292 /* Sort data and non-data nodes */
293 list_sort(c, &sleb->nodes, &data_nodes_cmp);
294 list_sort(c, nondata, &nondata_nodes_cmp);
Artem Bityutskiy3bb66b42010-08-07 10:06:11 +0300295
296 err = dbg_check_data_nodes_order(c, &sleb->nodes);
297 if (err)
298 return err;
299 err = dbg_check_nondata_nodes_order(c, nondata);
300 if (err)
301 return err;
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200302 return 0;
303}
304
305/**
306 * move_node - move a node.
307 * @c: UBIFS file-system description object
308 * @sleb: describes the LEB to move nodes from
309 * @snod: the mode to move
310 * @wbuf: write-buffer to move node to
311 *
312 * This function moves node @snod to @wbuf, changes TNC correspondingly, and
313 * destroys @snod. Returns zero in case of success and a negative error code in
314 * case of failure.
315 */
316static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
317 struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
318{
319 int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
320
321 cond_resched();
322 err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
323 if (err)
324 return err;
325
326 err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
327 snod->offs, new_lnum, new_offs,
328 snod->len);
329 list_del(&snod->list);
330 kfree(snod);
331 return err;
332}
333
334/**
335 * move_nodes - move nodes.
336 * @c: UBIFS file-system description object
337 * @sleb: describes the LEB to move nodes from
338 *
339 * This function moves valid nodes from data LEB described by @sleb to the GC
340 * journal head. This function returns zero in case of success, %-EAGAIN if
341 * commit is required, and other negative error codes in case of other
342 * failures.
343 */
344static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
345{
346 int err, min;
347 LIST_HEAD(nondata);
348 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300349
350 if (wbuf->lnum == -1) {
351 /*
352 * The GC journal head is not set, because it is the first GC
353 * invocation since mount.
354 */
355 err = switch_gc_head(c);
356 if (err)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200357 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300358 }
359
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200360 err = sort_nodes(c, sleb, &nondata, &min);
361 if (err)
362 goto out;
363
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300364 /* Write nodes to their new location. Use the first-fit strategy */
365 while (1) {
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200366 int avail;
367 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300368
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200369 /* Move data nodes */
370 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
371 avail = c->leb_size - wbuf->offs - wbuf->used;
372 if (snod->len > avail)
373 /*
374 * Do not skip data nodes in order to optimize
375 * bulk-read.
376 */
377 break;
378
379 err = move_node(c, sleb, snod, wbuf);
380 if (err)
381 goto out;
382 }
383
384 /* Move non-data nodes */
385 list_for_each_entry_safe(snod, tmp, &nondata, list) {
386 avail = c->leb_size - wbuf->offs - wbuf->used;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300387 if (avail < min)
388 break;
389
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200390 if (snod->len > avail) {
391 /*
392 * Keep going only if this is an inode with
393 * some data. Otherwise stop and switch the GC
394 * head. IOW, we assume that data-less inode
395 * nodes and direntry nodes are roughly of the
396 * same size.
397 */
398 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
399 snod->len == UBIFS_INO_NODE_SZ)
400 break;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300401 continue;
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200402 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300403
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200404 err = move_node(c, sleb, snod, wbuf);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300405 if (err)
406 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300407 }
408
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200409 if (list_empty(&sleb->nodes) && list_empty(&nondata))
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300410 break;
411
412 /*
413 * Waste the rest of the space in the LEB and switch to the
414 * next LEB.
415 */
416 err = switch_gc_head(c);
417 if (err)
418 goto out;
419 }
420
421 return 0;
422
423out:
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200424 list_splice_tail(&nondata, &sleb->nodes);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300425 return err;
426}
427
428/**
429 * gc_sync_wbufs - sync write-buffers for GC.
430 * @c: UBIFS file-system description object
431 *
432 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
433 * be in a write-buffer instead. That is, a node could be written to a
434 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
435 * erased before the write-buffer is sync'd and then there is an unclean
436 * unmount, then an existing node is lost. To avoid this, we sync all
437 * write-buffers.
438 *
439 * This function returns %0 on success or a negative error code on failure.
440 */
441static int gc_sync_wbufs(struct ubifs_info *c)
442{
443 int err, i;
444
445 for (i = 0; i < c->jhead_cnt; i++) {
446 if (i == GCHD)
447 continue;
448 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
449 if (err)
450 return err;
451 }
452 return 0;
453}
454
455/**
456 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
457 * @c: UBIFS file-system description object
458 * @lp: describes the LEB to garbage collect
459 *
460 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
461 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
462 * required, and other negative error codes in case of failures.
463 */
464int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
465{
466 struct ubifs_scan_leb *sleb;
467 struct ubifs_scan_node *snod;
468 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
469 int err = 0, lnum = lp->lnum;
470
471 ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
472 c->need_recovery);
473 ubifs_assert(c->gc_lnum != lnum);
474 ubifs_assert(wbuf->lnum != lnum);
475
476 /*
477 * We scan the entire LEB even though we only really need to scan up to
478 * (c->leb_size - lp->free).
479 */
Artem Bityutskiy348709b2009-08-25 15:00:55 +0300480 sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300481 if (IS_ERR(sleb))
482 return PTR_ERR(sleb);
483
484 ubifs_assert(!list_empty(&sleb->nodes));
485 snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
486
487 if (snod->type == UBIFS_IDX_NODE) {
488 struct ubifs_gced_idx_leb *idx_gc;
489
490 dbg_gc("indexing LEB %d (free %d, dirty %d)",
491 lnum, lp->free, lp->dirty);
492 list_for_each_entry(snod, &sleb->nodes, list) {
493 struct ubifs_idx_node *idx = snod->node;
494 int level = le16_to_cpu(idx->level);
495
496 ubifs_assert(snod->type == UBIFS_IDX_NODE);
497 key_read(c, ubifs_idx_key(c, idx), &snod->key);
498 err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
499 snod->offs);
500 if (err)
501 goto out;
502 }
503
504 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
505 if (!idx_gc) {
506 err = -ENOMEM;
507 goto out;
508 }
509
510 idx_gc->lnum = lnum;
511 idx_gc->unmap = 0;
512 list_add(&idx_gc->list, &c->idx_gc);
513
514 /*
515 * Don't release the LEB until after the next commit, because
Adrian Hunter227c75c2009-01-29 11:53:51 +0200516 * it may contain data which is needed for recovery. So
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300517 * although we freed this LEB, it will become usable only after
518 * the commit.
519 */
520 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
521 LPROPS_INDEX, 1);
522 if (err)
523 goto out;
524 err = LEB_FREED_IDX;
525 } else {
526 dbg_gc("data LEB %d (free %d, dirty %d)",
527 lnum, lp->free, lp->dirty);
528
529 err = move_nodes(c, sleb);
530 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300531 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300532
533 err = gc_sync_wbufs(c);
534 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300535 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300536
537 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
538 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300539 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300540
Adrian Hunter601c0bc2008-08-22 14:23:35 +0300541 /* Allow for races with TNC */
542 c->gced_lnum = lnum;
543 smp_wmb();
544 c->gc_seq += 1;
545 smp_wmb();
546
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300547 if (c->gc_lnum == -1) {
548 c->gc_lnum = lnum;
549 err = LEB_RETAINED;
550 } else {
551 err = ubifs_wbuf_sync_nolock(wbuf);
552 if (err)
553 goto out;
554
555 err = ubifs_leb_unmap(c, lnum);
556 if (err)
557 goto out;
558
559 err = LEB_FREED;
560 }
561 }
562
563out:
564 ubifs_scan_destroy(sleb);
565 return err;
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300566
567out_inc_seq:
568 /* We may have moved at least some nodes so allow for races with TNC */
569 c->gced_lnum = lnum;
570 smp_wmb();
571 c->gc_seq += 1;
572 smp_wmb();
573 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300574}
575
576/**
577 * ubifs_garbage_collect - UBIFS garbage collector.
578 * @c: UBIFS file-system description object
579 * @anyway: do GC even if there are free LEBs
580 *
581 * This function does out-of-place garbage collection. The return codes are:
582 * o positive LEB number if the LEB has been freed and may be used;
583 * o %-EAGAIN if the caller has to run commit;
584 * o %-ENOSPC if GC failed to make any progress;
585 * o other negative error codes in case of other errors.
586 *
587 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
588 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
589 * commit may be required. But commit cannot be run from inside GC, because the
590 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
591 * And this error code means that the caller has to run commit, and re-run GC
592 * if there is still no free space.
593 *
594 * There are many reasons why this function may return %-EAGAIN:
595 * o the log is full and there is no space to write an LEB reference for
596 * @c->gc_lnum;
597 * o the journal is too large and exceeds size limitations;
598 * o GC moved indexing LEBs, but they can be used only after the commit;
599 * o the shrinker fails to find clean znodes to free and requests the commit;
600 * o etc.
601 *
602 * Note, if the file-system is close to be full, this function may return
603 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
604 * the function. E.g., this happens if the limits on the journal size are too
605 * tough and GC writes too much to the journal before an LEB is freed. This
606 * might also mean that the journal is too large, and the TNC becomes to big,
607 * so that the shrinker is constantly called, finds not clean znodes to free,
608 * and requests commit. Well, this may also happen if the journal is all right,
609 * but another kernel process consumes too much memory. Anyway, infinite
610 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
611 */
612int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
613{
614 int i, err, ret, min_space = c->dead_wm;
615 struct ubifs_lprops lp;
616 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
617
618 ubifs_assert_cmt_locked(c);
Artem Bityutskiy2ef13292010-09-19 18:34:26 +0300619 ubifs_assert(!c->ro_media && !c->ro_mount);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300620
621 if (ubifs_gc_should_commit(c))
622 return -EAGAIN;
623
624 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
625
Artem Bityutskiy2680d722010-09-17 16:44:28 +0300626 if (c->ro_error) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300627 ret = -EROFS;
628 goto out_unlock;
629 }
630
631 /* We expect the write-buffer to be empty on entry */
632 ubifs_assert(!wbuf->used);
633
634 for (i = 0; ; i++) {
635 int space_before = c->leb_size - wbuf->offs - wbuf->used;
636 int space_after;
637
638 cond_resched();
639
640 /* Give the commit an opportunity to run */
641 if (ubifs_gc_should_commit(c)) {
642 ret = -EAGAIN;
643 break;
644 }
645
646 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
647 /*
648 * We've done enough iterations. Indexing LEBs were
649 * moved and will be available after the commit.
650 */
651 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
652 ubifs_commit_required(c);
653 ret = -EAGAIN;
654 break;
655 }
656
657 if (i > HARD_LEBS_LIMIT) {
658 /*
659 * We've moved too many LEBs and have not made
660 * progress, give up.
661 */
662 dbg_gc("hard limit, -ENOSPC");
663 ret = -ENOSPC;
664 break;
665 }
666
667 /*
668 * Empty and freeable LEBs can turn up while we waited for
669 * the wbuf lock, or while we have been running GC. In that
670 * case, we should just return one of those instead of
671 * continuing to GC dirty LEBs. Hence we request
672 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
673 */
674 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
675 if (ret) {
676 if (ret == -ENOSPC)
677 dbg_gc("no more dirty LEBs");
678 break;
679 }
680
681 dbg_gc("found LEB %d: free %d, dirty %d, sum %d "
682 "(min. space %d)", lp.lnum, lp.free, lp.dirty,
683 lp.free + lp.dirty, min_space);
684
685 if (lp.free + lp.dirty == c->leb_size) {
686 /* An empty LEB was returned */
687 dbg_gc("LEB %d is free, return it", lp.lnum);
688 /*
689 * ubifs_find_dirty_leb() doesn't return freeable index
690 * LEBs.
691 */
692 ubifs_assert(!(lp.flags & LPROPS_INDEX));
693 if (lp.free != c->leb_size) {
694 /*
695 * Write buffers must be sync'd before
696 * unmapping freeable LEBs, because one of them
697 * may contain data which obsoletes something
698 * in 'lp.pnum'.
699 */
700 ret = gc_sync_wbufs(c);
701 if (ret)
702 goto out;
703 ret = ubifs_change_one_lp(c, lp.lnum,
704 c->leb_size, 0, 0, 0,
705 0);
706 if (ret)
707 goto out;
708 }
709 ret = ubifs_leb_unmap(c, lp.lnum);
710 if (ret)
711 goto out;
712 ret = lp.lnum;
713 break;
714 }
715
716 space_before = c->leb_size - wbuf->offs - wbuf->used;
717 if (wbuf->lnum == -1)
718 space_before = 0;
719
720 ret = ubifs_garbage_collect_leb(c, &lp);
721 if (ret < 0) {
Artem Bityutskiyefe18812010-08-04 14:06:06 +0300722 if (ret == -EAGAIN) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300723 /*
Artem Bityutskiyefe18812010-08-04 14:06:06 +0300724 * This is not error, so we have to return the
725 * LEB to lprops. But if 'ubifs_return_leb()'
726 * fails, its failure code is propagated to the
727 * caller instead of the original '-EAGAIN'.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300728 */
729 err = ubifs_return_leb(c, lp.lnum);
730 if (err)
731 ret = err;
732 break;
733 }
734 goto out;
735 }
736
737 if (ret == LEB_FREED) {
738 /* An LEB has been freed and is ready for use */
739 dbg_gc("LEB %d freed, return", lp.lnum);
740 ret = lp.lnum;
741 break;
742 }
743
744 if (ret == LEB_FREED_IDX) {
745 /*
746 * This was an indexing LEB and it cannot be
747 * immediately used. And instead of requesting the
748 * commit straight away, we try to garbage collect some
749 * more.
750 */
751 dbg_gc("indexing LEB %d freed, continue", lp.lnum);
752 continue;
753 }
754
755 ubifs_assert(ret == LEB_RETAINED);
756 space_after = c->leb_size - wbuf->offs - wbuf->used;
757 dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
758 space_after - space_before);
759
760 if (space_after > space_before) {
761 /* GC makes progress, keep working */
762 min_space >>= 1;
763 if (min_space < c->dead_wm)
764 min_space = c->dead_wm;
765 continue;
766 }
767
768 dbg_gc("did not make progress");
769
770 /*
771 * GC moved an LEB bud have not done any progress. This means
772 * that the previous GC head LEB contained too few free space
773 * and the LEB which was GC'ed contained only large nodes which
774 * did not fit that space.
775 *
776 * We can do 2 things:
777 * 1. pick another LEB in a hope it'll contain a small node
778 * which will fit the space we have at the end of current GC
779 * head LEB, but there is no guarantee, so we try this out
780 * unless we have already been working for too long;
781 * 2. request an LEB with more dirty space, which will force
782 * 'ubifs_find_dirty_leb()' to start scanning the lprops
783 * table, instead of just picking one from the heap
784 * (previously it already picked the dirtiest LEB).
785 */
786 if (i < SOFT_LEBS_LIMIT) {
787 dbg_gc("try again");
788 continue;
789 }
790
791 min_space <<= 1;
792 if (min_space > c->dark_wm)
793 min_space = c->dark_wm;
794 dbg_gc("set min. space to %d", min_space);
795 }
796
797 if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
798 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
799 ubifs_commit_required(c);
800 ret = -EAGAIN;
801 }
802
803 err = ubifs_wbuf_sync_nolock(wbuf);
804 if (!err)
805 err = ubifs_leb_unmap(c, c->gc_lnum);
806 if (err) {
807 ret = err;
808 goto out;
809 }
810out_unlock:
811 mutex_unlock(&wbuf->io_mutex);
812 return ret;
813
814out:
815 ubifs_assert(ret < 0);
816 ubifs_assert(ret != -ENOSPC && ret != -EAGAIN);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300817 ubifs_wbuf_sync_nolock(wbuf);
Artem Bityutskiy5ffef882010-08-04 10:14:47 +0300818 ubifs_ro_mode(c, ret);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300819 mutex_unlock(&wbuf->io_mutex);
820 ubifs_return_leb(c, lp.lnum);
821 return ret;
822}
823
824/**
825 * ubifs_gc_start_commit - garbage collection at start of commit.
826 * @c: UBIFS file-system description object
827 *
828 * If a LEB has only dirty and free space, then we may safely unmap it and make
829 * it free. Note, we cannot do this with indexing LEBs because dirty space may
830 * correspond index nodes that are required for recovery. In that case, the
831 * LEB cannot be unmapped until after the next commit.
832 *
833 * This function returns %0 upon success and a negative error code upon failure.
834 */
835int ubifs_gc_start_commit(struct ubifs_info *c)
836{
837 struct ubifs_gced_idx_leb *idx_gc;
838 const struct ubifs_lprops *lp;
839 int err = 0, flags;
840
841 ubifs_get_lprops(c);
842
843 /*
844 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
845 * wbufs are sync'd before this, which is done in 'do_commit()'.
846 */
847 while (1) {
848 lp = ubifs_fast_find_freeable(c);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300849 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300850 err = PTR_ERR(lp);
851 goto out;
852 }
853 if (!lp)
854 break;
855 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
856 ubifs_assert(!(lp->flags & LPROPS_INDEX));
857 err = ubifs_leb_unmap(c, lp->lnum);
858 if (err)
859 goto out;
860 lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300861 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300862 err = PTR_ERR(lp);
863 goto out;
864 }
865 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
866 ubifs_assert(!(lp->flags & LPROPS_INDEX));
867 }
868
869 /* Mark GC'd index LEBs OK to unmap after this commit finishes */
870 list_for_each_entry(idx_gc, &c->idx_gc, list)
871 idx_gc->unmap = 1;
872
873 /* Record index freeable LEBs for unmapping after commit */
874 while (1) {
875 lp = ubifs_fast_find_frdi_idx(c);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300876 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300877 err = PTR_ERR(lp);
878 goto out;
879 }
880 if (!lp)
881 break;
882 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
883 if (!idx_gc) {
884 err = -ENOMEM;
885 goto out;
886 }
887 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
888 ubifs_assert(lp->flags & LPROPS_INDEX);
889 /* Don't release the LEB until after the next commit */
890 flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
891 lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300892 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300893 err = PTR_ERR(lp);
894 kfree(idx_gc);
895 goto out;
896 }
897 ubifs_assert(lp->flags & LPROPS_TAKEN);
898 ubifs_assert(!(lp->flags & LPROPS_INDEX));
899 idx_gc->lnum = lp->lnum;
900 idx_gc->unmap = 1;
901 list_add(&idx_gc->list, &c->idx_gc);
902 }
903out:
904 ubifs_release_lprops(c);
905 return err;
906}
907
908/**
909 * ubifs_gc_end_commit - garbage collection at end of commit.
910 * @c: UBIFS file-system description object
911 *
912 * This function completes out-of-place garbage collection of index LEBs.
913 */
914int ubifs_gc_end_commit(struct ubifs_info *c)
915{
916 struct ubifs_gced_idx_leb *idx_gc, *tmp;
917 struct ubifs_wbuf *wbuf;
918 int err = 0;
919
920 wbuf = &c->jheads[GCHD].wbuf;
921 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
922 list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
923 if (idx_gc->unmap) {
924 dbg_gc("LEB %d", idx_gc->lnum);
925 err = ubifs_leb_unmap(c, idx_gc->lnum);
926 if (err)
927 goto out;
928 err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
929 LPROPS_NC, 0, LPROPS_TAKEN, -1);
930 if (err)
931 goto out;
932 list_del(&idx_gc->list);
933 kfree(idx_gc);
934 }
935out:
936 mutex_unlock(&wbuf->io_mutex);
937 return err;
938}
939
940/**
941 * ubifs_destroy_idx_gc - destroy idx_gc list.
942 * @c: UBIFS file-system description object
943 *
Adrian Hunterb466f172009-01-29 12:59:33 +0200944 * This function destroys the @c->idx_gc list. It is called when unmounting
945 * so locks are not needed. Returns zero in case of success and a negative
946 * error code in case of failure.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300947 */
Adrian Hunterb466f172009-01-29 12:59:33 +0200948void ubifs_destroy_idx_gc(struct ubifs_info *c)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300949{
950 while (!list_empty(&c->idx_gc)) {
951 struct ubifs_gced_idx_leb *idx_gc;
952
953 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
954 list);
Adrian Hunterb466f172009-01-29 12:59:33 +0200955 c->idx_gc_cnt -= 1;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300956 list_del(&idx_gc->list);
957 kfree(idx_gc);
958 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300959}
960
961/**
962 * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
963 * @c: UBIFS file-system description object
964 *
965 * Called during start commit so locks are not needed.
966 */
967int ubifs_get_idx_gc_leb(struct ubifs_info *c)
968{
969 struct ubifs_gced_idx_leb *idx_gc;
970 int lnum;
971
972 if (list_empty(&c->idx_gc))
973 return -ENOSPC;
974 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
975 lnum = idx_gc->lnum;
976 /* c->idx_gc_cnt is updated by the caller when lprops are updated */
977 list_del(&idx_gc->list);
978 kfree(idx_gc);
979 return lnum;
980}