blob: e845c64b6ce19c175b8208d448fc4f75c0c2c1b7 [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;
Richard Weinbergerb36a2612012-05-14 17:55:51 +0200108 err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300109 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
Julia Lawallec037df2016-10-01 22:53:11 +0200116 * @b: 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 Bityutskiye9ef7b5f2011-05-13 13:22:19 +0300121static int data_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200122{
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 */
Artem Bityutskiye9ef7b5f2011-05-13 13:22:19 +0300164static int nondata_nodes_cmp(void *priv, struct list_head *a,
165 struct list_head *b)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200166{
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200167 ino_t inuma, inumb;
168 struct ubifs_info *c = priv;
169 struct ubifs_scan_node *sa, *sb;
170
171 cond_resched();
Artem Bityutskiy1a9476a2010-08-08 12:45:23 +0300172 if (a == b)
173 return 0;
174
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200175 sa = list_entry(a, struct ubifs_scan_node, list);
176 sb = list_entry(b, struct ubifs_scan_node, list);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300177
178 ubifs_assert(key_type(c, &sa->key) != UBIFS_DATA_KEY &&
179 key_type(c, &sb->key) != UBIFS_DATA_KEY);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300180 ubifs_assert(sa->type != UBIFS_DATA_NODE &&
181 sb->type != UBIFS_DATA_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200182
183 /* Inodes go before directory entries */
Artem Bityutskiyab871182010-08-08 12:25:33 +0300184 if (sa->type == UBIFS_INO_NODE) {
185 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200186 return sb->len - sa->len;
187 return -1;
188 }
Artem Bityutskiyab871182010-08-08 12:25:33 +0300189 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200190 return 1;
191
Artem Bityutskiy66576832010-08-08 12:29:58 +0300192 ubifs_assert(key_type(c, &sa->key) == UBIFS_DENT_KEY ||
193 key_type(c, &sa->key) == UBIFS_XENT_KEY);
194 ubifs_assert(key_type(c, &sb->key) == UBIFS_DENT_KEY ||
195 key_type(c, &sb->key) == UBIFS_XENT_KEY);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300196 ubifs_assert(sa->type == UBIFS_DENT_NODE ||
197 sa->type == UBIFS_XENT_NODE);
198 ubifs_assert(sb->type == UBIFS_DENT_NODE ||
199 sb->type == UBIFS_XENT_NODE);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300200
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200201 inuma = key_inum(c, &sa->key);
202 inumb = key_inum(c, &sb->key);
203
204 if (inuma == inumb) {
205 uint32_t hasha = key_hash(c, &sa->key);
206 uint32_t hashb = key_hash(c, &sb->key);
207
208 if (hasha <= hashb)
209 return -1;
210 } else if (inuma <= inumb)
211 return -1;
212
213 return 1;
214}
215
216/**
217 * sort_nodes - sort nodes for GC.
218 * @c: UBIFS file-system description object
219 * @sleb: describes nodes to sort and contains the result on exit
220 * @nondata: contains non-data nodes on exit
221 * @min: minimum node size is returned here
222 *
223 * This function sorts the list of inodes to garbage collect. First of all, it
224 * kills obsolete nodes and separates data and non-data nodes to the
225 * @sleb->nodes and @nondata lists correspondingly.
226 *
227 * Data nodes are then sorted in block number order - this is important for
228 * bulk-read; data nodes with lower inode number go before data nodes with
229 * higher inode number, and data nodes with lower block number go before data
230 * nodes with higher block number;
231 *
232 * Non-data nodes are sorted as follows.
233 * o First go inode nodes - they are sorted in descending length order.
234 * o Then go directory entry nodes - they are sorted in hash order, which
235 * should supposedly optimize 'readdir()'. Direntry nodes with lower parent
236 * inode number go before direntry nodes with higher parent inode number,
237 * and direntry nodes with lower name hash values go before direntry nodes
238 * with higher name hash values.
239 *
240 * This function returns zero in case of success and a negative error code in
241 * case of failure.
242 */
243static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
244 struct list_head *nondata, int *min)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300245{
Artem Bityutskiy3bb66b42010-08-07 10:06:11 +0300246 int err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300247 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300248
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200249 *min = INT_MAX;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300250
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200251 /* Separate data nodes and non-data nodes */
252 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
Artem Bityutskiy44ec83b2010-08-07 08:44:13 +0300253 ubifs_assert(snod->type == UBIFS_INO_NODE ||
254 snod->type == UBIFS_DATA_NODE ||
255 snod->type == UBIFS_DENT_NODE ||
256 snod->type == UBIFS_XENT_NODE ||
257 snod->type == UBIFS_TRUN_NODE);
258
259 if (snod->type != UBIFS_INO_NODE &&
260 snod->type != UBIFS_DATA_NODE &&
261 snod->type != UBIFS_DENT_NODE &&
262 snod->type != UBIFS_XENT_NODE) {
263 /* Probably truncation node, zap it */
264 list_del(&snod->list);
265 kfree(snod);
266 continue;
267 }
268
269 ubifs_assert(key_type(c, &snod->key) == UBIFS_DATA_KEY ||
270 key_type(c, &snod->key) == UBIFS_INO_KEY ||
271 key_type(c, &snod->key) == UBIFS_DENT_KEY ||
272 key_type(c, &snod->key) == UBIFS_XENT_KEY);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300273
274 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
275 snod->offs, 0);
276 if (err < 0)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200277 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300278
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300279 if (!err) {
280 /* The node is obsolete, remove it from the list */
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200281 list_del(&snod->list);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300282 kfree(snod);
283 continue;
284 }
285
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200286 if (snod->len < *min)
287 *min = snod->len;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300288
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200289 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
290 list_move_tail(&snod->list, nondata);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300291 }
292
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200293 /* Sort data and non-data nodes */
294 list_sort(c, &sleb->nodes, &data_nodes_cmp);
295 list_sort(c, nondata, &nondata_nodes_cmp);
Artem Bityutskiy3bb66b42010-08-07 10:06:11 +0300296
297 err = dbg_check_data_nodes_order(c, &sleb->nodes);
298 if (err)
299 return err;
300 err = dbg_check_nondata_nodes_order(c, nondata);
301 if (err)
302 return err;
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200303 return 0;
304}
305
306/**
307 * move_node - move a node.
308 * @c: UBIFS file-system description object
309 * @sleb: describes the LEB to move nodes from
310 * @snod: the mode to move
311 * @wbuf: write-buffer to move node to
312 *
313 * This function moves node @snod to @wbuf, changes TNC correspondingly, and
314 * destroys @snod. Returns zero in case of success and a negative error code in
315 * case of failure.
316 */
317static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
318 struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
319{
320 int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
321
322 cond_resched();
323 err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
324 if (err)
325 return err;
326
327 err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
328 snod->offs, new_lnum, new_offs,
329 snod->len);
330 list_del(&snod->list);
331 kfree(snod);
332 return err;
333}
334
335/**
336 * move_nodes - move nodes.
337 * @c: UBIFS file-system description object
338 * @sleb: describes the LEB to move nodes from
339 *
340 * This function moves valid nodes from data LEB described by @sleb to the GC
341 * journal head. This function returns zero in case of success, %-EAGAIN if
342 * commit is required, and other negative error codes in case of other
343 * failures.
344 */
345static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
346{
347 int err, min;
348 LIST_HEAD(nondata);
349 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300350
351 if (wbuf->lnum == -1) {
352 /*
353 * The GC journal head is not set, because it is the first GC
354 * invocation since mount.
355 */
356 err = switch_gc_head(c);
357 if (err)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200358 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300359 }
360
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200361 err = sort_nodes(c, sleb, &nondata, &min);
362 if (err)
363 goto out;
364
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300365 /* Write nodes to their new location. Use the first-fit strategy */
366 while (1) {
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200367 int avail;
368 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300369
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200370 /* Move data nodes */
371 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
372 avail = c->leb_size - wbuf->offs - wbuf->used;
373 if (snod->len > avail)
374 /*
375 * Do not skip data nodes in order to optimize
376 * bulk-read.
377 */
378 break;
379
380 err = move_node(c, sleb, snod, wbuf);
381 if (err)
382 goto out;
383 }
384
385 /* Move non-data nodes */
386 list_for_each_entry_safe(snod, tmp, &nondata, list) {
387 avail = c->leb_size - wbuf->offs - wbuf->used;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300388 if (avail < min)
389 break;
390
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200391 if (snod->len > avail) {
392 /*
393 * Keep going only if this is an inode with
394 * some data. Otherwise stop and switch the GC
395 * head. IOW, we assume that data-less inode
396 * nodes and direntry nodes are roughly of the
397 * same size.
398 */
399 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
400 snod->len == UBIFS_INO_NODE_SZ)
401 break;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300402 continue;
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200403 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300404
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200405 err = move_node(c, sleb, snod, wbuf);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300406 if (err)
407 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300408 }
409
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200410 if (list_empty(&sleb->nodes) && list_empty(&nondata))
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300411 break;
412
413 /*
414 * Waste the rest of the space in the LEB and switch to the
415 * next LEB.
416 */
417 err = switch_gc_head(c);
418 if (err)
419 goto out;
420 }
421
422 return 0;
423
424out:
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200425 list_splice_tail(&nondata, &sleb->nodes);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300426 return err;
427}
428
429/**
430 * gc_sync_wbufs - sync write-buffers for GC.
431 * @c: UBIFS file-system description object
432 *
433 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
434 * be in a write-buffer instead. That is, a node could be written to a
435 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
436 * erased before the write-buffer is sync'd and then there is an unclean
437 * unmount, then an existing node is lost. To avoid this, we sync all
438 * write-buffers.
439 *
440 * This function returns %0 on success or a negative error code on failure.
441 */
442static int gc_sync_wbufs(struct ubifs_info *c)
443{
444 int err, i;
445
446 for (i = 0; i < c->jhead_cnt; i++) {
447 if (i == GCHD)
448 continue;
449 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
450 if (err)
451 return err;
452 }
453 return 0;
454}
455
456/**
457 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
458 * @c: UBIFS file-system description object
459 * @lp: describes the LEB to garbage collect
460 *
461 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
462 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
463 * required, and other negative error codes in case of failures.
464 */
465int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
466{
467 struct ubifs_scan_leb *sleb;
468 struct ubifs_scan_node *snod;
469 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
470 int err = 0, lnum = lp->lnum;
471
472 ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
473 c->need_recovery);
474 ubifs_assert(c->gc_lnum != lnum);
475 ubifs_assert(wbuf->lnum != lnum);
476
Artem Bityutskiy2405f592011-04-26 09:49:32 +0300477 if (lp->free + lp->dirty == c->leb_size) {
478 /* Special case - a free LEB */
479 dbg_gc("LEB %d is free, return it", lp->lnum);
480 ubifs_assert(!(lp->flags & LPROPS_INDEX));
481
482 if (lp->free != c->leb_size) {
483 /*
484 * Write buffers must be sync'd before unmapping
485 * freeable LEBs, because one of them may contain data
486 * which obsoletes something in 'lp->pnum'.
487 */
488 err = gc_sync_wbufs(c);
489 if (err)
490 return err;
491 err = ubifs_change_one_lp(c, lp->lnum, c->leb_size,
492 0, 0, 0, 0);
493 if (err)
494 return err;
495 }
496 err = ubifs_leb_unmap(c, lp->lnum);
497 if (err)
498 return err;
499
500 if (c->gc_lnum == -1) {
501 c->gc_lnum = lnum;
502 return LEB_RETAINED;
503 }
504
505 return LEB_FREED;
506 }
507
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300508 /*
509 * We scan the entire LEB even though we only really need to scan up to
510 * (c->leb_size - lp->free).
511 */
Artem Bityutskiy348709b2009-08-25 15:00:55 +0300512 sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300513 if (IS_ERR(sleb))
514 return PTR_ERR(sleb);
515
516 ubifs_assert(!list_empty(&sleb->nodes));
517 snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
518
519 if (snod->type == UBIFS_IDX_NODE) {
520 struct ubifs_gced_idx_leb *idx_gc;
521
522 dbg_gc("indexing LEB %d (free %d, dirty %d)",
523 lnum, lp->free, lp->dirty);
524 list_for_each_entry(snod, &sleb->nodes, list) {
525 struct ubifs_idx_node *idx = snod->node;
526 int level = le16_to_cpu(idx->level);
527
528 ubifs_assert(snod->type == UBIFS_IDX_NODE);
529 key_read(c, ubifs_idx_key(c, idx), &snod->key);
530 err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
531 snod->offs);
532 if (err)
533 goto out;
534 }
535
536 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
537 if (!idx_gc) {
538 err = -ENOMEM;
539 goto out;
540 }
541
542 idx_gc->lnum = lnum;
543 idx_gc->unmap = 0;
544 list_add(&idx_gc->list, &c->idx_gc);
545
546 /*
547 * Don't release the LEB until after the next commit, because
Adrian Hunter227c75c2009-01-29 11:53:51 +0200548 * it may contain data which is needed for recovery. So
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300549 * although we freed this LEB, it will become usable only after
550 * the commit.
551 */
552 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
553 LPROPS_INDEX, 1);
554 if (err)
555 goto out;
556 err = LEB_FREED_IDX;
557 } else {
558 dbg_gc("data LEB %d (free %d, dirty %d)",
559 lnum, lp->free, lp->dirty);
560
561 err = move_nodes(c, sleb);
562 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300563 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300564
565 err = gc_sync_wbufs(c);
566 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300567 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300568
569 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
570 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300571 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300572
Adrian Hunter601c0bc2008-08-22 14:23:35 +0300573 /* Allow for races with TNC */
574 c->gced_lnum = lnum;
575 smp_wmb();
576 c->gc_seq += 1;
577 smp_wmb();
578
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300579 if (c->gc_lnum == -1) {
580 c->gc_lnum = lnum;
581 err = LEB_RETAINED;
582 } else {
583 err = ubifs_wbuf_sync_nolock(wbuf);
584 if (err)
585 goto out;
586
587 err = ubifs_leb_unmap(c, lnum);
588 if (err)
589 goto out;
590
591 err = LEB_FREED;
592 }
593 }
594
595out:
596 ubifs_scan_destroy(sleb);
597 return err;
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300598
599out_inc_seq:
600 /* We may have moved at least some nodes so allow for races with TNC */
601 c->gced_lnum = lnum;
602 smp_wmb();
603 c->gc_seq += 1;
604 smp_wmb();
605 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300606}
607
608/**
609 * ubifs_garbage_collect - UBIFS garbage collector.
610 * @c: UBIFS file-system description object
611 * @anyway: do GC even if there are free LEBs
612 *
613 * This function does out-of-place garbage collection. The return codes are:
614 * o positive LEB number if the LEB has been freed and may be used;
615 * o %-EAGAIN if the caller has to run commit;
616 * o %-ENOSPC if GC failed to make any progress;
617 * o other negative error codes in case of other errors.
618 *
619 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
620 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
621 * commit may be required. But commit cannot be run from inside GC, because the
622 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
623 * And this error code means that the caller has to run commit, and re-run GC
624 * if there is still no free space.
625 *
626 * There are many reasons why this function may return %-EAGAIN:
627 * o the log is full and there is no space to write an LEB reference for
628 * @c->gc_lnum;
629 * o the journal is too large and exceeds size limitations;
630 * o GC moved indexing LEBs, but they can be used only after the commit;
631 * o the shrinker fails to find clean znodes to free and requests the commit;
632 * o etc.
633 *
634 * Note, if the file-system is close to be full, this function may return
635 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
636 * the function. E.g., this happens if the limits on the journal size are too
637 * tough and GC writes too much to the journal before an LEB is freed. This
638 * might also mean that the journal is too large, and the TNC becomes to big,
639 * so that the shrinker is constantly called, finds not clean znodes to free,
640 * and requests commit. Well, this may also happen if the journal is all right,
641 * but another kernel process consumes too much memory. Anyway, infinite
642 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
643 */
644int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
645{
646 int i, err, ret, min_space = c->dead_wm;
647 struct ubifs_lprops lp;
648 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
649
650 ubifs_assert_cmt_locked(c);
Artem Bityutskiy2ef13292010-09-19 18:34:26 +0300651 ubifs_assert(!c->ro_media && !c->ro_mount);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300652
653 if (ubifs_gc_should_commit(c))
654 return -EAGAIN;
655
656 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
657
Artem Bityutskiy2680d722010-09-17 16:44:28 +0300658 if (c->ro_error) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300659 ret = -EROFS;
660 goto out_unlock;
661 }
662
663 /* We expect the write-buffer to be empty on entry */
664 ubifs_assert(!wbuf->used);
665
666 for (i = 0; ; i++) {
wang.bo116@zte.com.cne71d1a52013-08-16 15:43:36 +0800667 int space_before, space_after;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300668
669 cond_resched();
670
671 /* Give the commit an opportunity to run */
672 if (ubifs_gc_should_commit(c)) {
673 ret = -EAGAIN;
674 break;
675 }
676
677 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
678 /*
679 * We've done enough iterations. Indexing LEBs were
680 * moved and will be available after the commit.
681 */
682 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
683 ubifs_commit_required(c);
684 ret = -EAGAIN;
685 break;
686 }
687
688 if (i > HARD_LEBS_LIMIT) {
689 /*
690 * We've moved too many LEBs and have not made
691 * progress, give up.
692 */
693 dbg_gc("hard limit, -ENOSPC");
694 ret = -ENOSPC;
695 break;
696 }
697
698 /*
699 * Empty and freeable LEBs can turn up while we waited for
700 * the wbuf lock, or while we have been running GC. In that
701 * case, we should just return one of those instead of
702 * continuing to GC dirty LEBs. Hence we request
703 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
704 */
705 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
706 if (ret) {
707 if (ret == -ENOSPC)
708 dbg_gc("no more dirty LEBs");
709 break;
710 }
711
Artem Bityutskiy79fda512012-08-27 13:34:09 +0300712 dbg_gc("found LEB %d: free %d, dirty %d, sum %d (min. space %d)",
713 lp.lnum, lp.free, lp.dirty, lp.free + lp.dirty,
714 min_space);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300715
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300716 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}