blob: 27815bb911259033218e8221f7dae13124a14761 [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();
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 */
158int nondata_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
159{
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200160 ino_t inuma, inumb;
161 struct ubifs_info *c = priv;
162 struct ubifs_scan_node *sa, *sb;
163
164 cond_resched();
165 sa = list_entry(a, struct ubifs_scan_node, list);
166 sb = list_entry(b, struct ubifs_scan_node, list);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300167 ubifs_assert(sa->type != UBIFS_DATA_NODE &&
168 sb->type != UBIFS_DATA_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200169
170 /* Inodes go before directory entries */
Artem Bityutskiyab871182010-08-08 12:25:33 +0300171 if (sa->type == UBIFS_INO_NODE) {
172 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200173 return sb->len - sa->len;
174 return -1;
175 }
Artem Bityutskiyab871182010-08-08 12:25:33 +0300176 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200177 return 1;
178
Artem Bityutskiyab871182010-08-08 12:25:33 +0300179 ubifs_assert(sa->type == UBIFS_DENT_NODE ||
180 sa->type == UBIFS_XENT_NODE);
181 ubifs_assert(sb->type == UBIFS_DENT_NODE ||
182 sb->type == UBIFS_XENT_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200183 inuma = key_inum(c, &sa->key);
184 inumb = key_inum(c, &sb->key);
185
186 if (inuma == inumb) {
187 uint32_t hasha = key_hash(c, &sa->key);
188 uint32_t hashb = key_hash(c, &sb->key);
189
190 if (hasha <= hashb)
191 return -1;
192 } else if (inuma <= inumb)
193 return -1;
194
195 return 1;
196}
197
198/**
199 * sort_nodes - sort nodes for GC.
200 * @c: UBIFS file-system description object
201 * @sleb: describes nodes to sort and contains the result on exit
202 * @nondata: contains non-data nodes on exit
203 * @min: minimum node size is returned here
204 *
205 * This function sorts the list of inodes to garbage collect. First of all, it
206 * kills obsolete nodes and separates data and non-data nodes to the
207 * @sleb->nodes and @nondata lists correspondingly.
208 *
209 * Data nodes are then sorted in block number order - this is important for
210 * bulk-read; data nodes with lower inode number go before data nodes with
211 * higher inode number, and data nodes with lower block number go before data
212 * nodes with higher block number;
213 *
214 * Non-data nodes are sorted as follows.
215 * o First go inode nodes - they are sorted in descending length order.
216 * o Then go directory entry nodes - they are sorted in hash order, which
217 * should supposedly optimize 'readdir()'. Direntry nodes with lower parent
218 * inode number go before direntry nodes with higher parent inode number,
219 * and direntry nodes with lower name hash values go before direntry nodes
220 * with higher name hash values.
221 *
222 * This function returns zero in case of success and a negative error code in
223 * case of failure.
224 */
225static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
226 struct list_head *nondata, int *min)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300227{
228 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300229
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200230 *min = INT_MAX;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300231
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200232 /* Separate data nodes and non-data nodes */
233 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
234 int err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300235
Artem Bityutskiy44ec83b2010-08-07 08:44:13 +0300236 ubifs_assert(snod->type == UBIFS_INO_NODE ||
237 snod->type == UBIFS_DATA_NODE ||
238 snod->type == UBIFS_DENT_NODE ||
239 snod->type == UBIFS_XENT_NODE ||
240 snod->type == UBIFS_TRUN_NODE);
241
242 if (snod->type != UBIFS_INO_NODE &&
243 snod->type != UBIFS_DATA_NODE &&
244 snod->type != UBIFS_DENT_NODE &&
245 snod->type != UBIFS_XENT_NODE) {
246 /* Probably truncation node, zap it */
247 list_del(&snod->list);
248 kfree(snod);
249 continue;
250 }
251
252 ubifs_assert(key_type(c, &snod->key) == UBIFS_DATA_KEY ||
253 key_type(c, &snod->key) == UBIFS_INO_KEY ||
254 key_type(c, &snod->key) == UBIFS_DENT_KEY ||
255 key_type(c, &snod->key) == UBIFS_XENT_KEY);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300256
257 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
258 snod->offs, 0);
259 if (err < 0)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200260 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300261
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300262 if (!err) {
263 /* The node is obsolete, remove it from the list */
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200264 list_del(&snod->list);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300265 kfree(snod);
266 continue;
267 }
268
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200269 if (snod->len < *min)
270 *min = snod->len;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300271
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200272 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
273 list_move_tail(&snod->list, nondata);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300274 }
275
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200276 /* Sort data and non-data nodes */
277 list_sort(c, &sleb->nodes, &data_nodes_cmp);
278 list_sort(c, nondata, &nondata_nodes_cmp);
279 return 0;
280}
281
282/**
283 * move_node - move a node.
284 * @c: UBIFS file-system description object
285 * @sleb: describes the LEB to move nodes from
286 * @snod: the mode to move
287 * @wbuf: write-buffer to move node to
288 *
289 * This function moves node @snod to @wbuf, changes TNC correspondingly, and
290 * destroys @snod. Returns zero in case of success and a negative error code in
291 * case of failure.
292 */
293static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
294 struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
295{
296 int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
297
298 cond_resched();
299 err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
300 if (err)
301 return err;
302
303 err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
304 snod->offs, new_lnum, new_offs,
305 snod->len);
306 list_del(&snod->list);
307 kfree(snod);
308 return err;
309}
310
311/**
312 * move_nodes - move nodes.
313 * @c: UBIFS file-system description object
314 * @sleb: describes the LEB to move nodes from
315 *
316 * This function moves valid nodes from data LEB described by @sleb to the GC
317 * journal head. This function returns zero in case of success, %-EAGAIN if
318 * commit is required, and other negative error codes in case of other
319 * failures.
320 */
321static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
322{
323 int err, min;
324 LIST_HEAD(nondata);
325 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300326
327 if (wbuf->lnum == -1) {
328 /*
329 * The GC journal head is not set, because it is the first GC
330 * invocation since mount.
331 */
332 err = switch_gc_head(c);
333 if (err)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200334 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300335 }
336
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200337 err = sort_nodes(c, sleb, &nondata, &min);
338 if (err)
339 goto out;
340
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300341 /* Write nodes to their new location. Use the first-fit strategy */
342 while (1) {
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200343 int avail;
344 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300345
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200346 /* Move data nodes */
347 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
348 avail = c->leb_size - wbuf->offs - wbuf->used;
349 if (snod->len > avail)
350 /*
351 * Do not skip data nodes in order to optimize
352 * bulk-read.
353 */
354 break;
355
356 err = move_node(c, sleb, snod, wbuf);
357 if (err)
358 goto out;
359 }
360
361 /* Move non-data nodes */
362 list_for_each_entry_safe(snod, tmp, &nondata, list) {
363 avail = c->leb_size - wbuf->offs - wbuf->used;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300364 if (avail < min)
365 break;
366
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200367 if (snod->len > avail) {
368 /*
369 * Keep going only if this is an inode with
370 * some data. Otherwise stop and switch the GC
371 * head. IOW, we assume that data-less inode
372 * nodes and direntry nodes are roughly of the
373 * same size.
374 */
375 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
376 snod->len == UBIFS_INO_NODE_SZ)
377 break;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300378 continue;
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200379 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300380
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200381 err = move_node(c, sleb, snod, wbuf);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300382 if (err)
383 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300384 }
385
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200386 if (list_empty(&sleb->nodes) && list_empty(&nondata))
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300387 break;
388
389 /*
390 * Waste the rest of the space in the LEB and switch to the
391 * next LEB.
392 */
393 err = switch_gc_head(c);
394 if (err)
395 goto out;
396 }
397
398 return 0;
399
400out:
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200401 list_splice_tail(&nondata, &sleb->nodes);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300402 return err;
403}
404
405/**
406 * gc_sync_wbufs - sync write-buffers for GC.
407 * @c: UBIFS file-system description object
408 *
409 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
410 * be in a write-buffer instead. That is, a node could be written to a
411 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
412 * erased before the write-buffer is sync'd and then there is an unclean
413 * unmount, then an existing node is lost. To avoid this, we sync all
414 * write-buffers.
415 *
416 * This function returns %0 on success or a negative error code on failure.
417 */
418static int gc_sync_wbufs(struct ubifs_info *c)
419{
420 int err, i;
421
422 for (i = 0; i < c->jhead_cnt; i++) {
423 if (i == GCHD)
424 continue;
425 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
426 if (err)
427 return err;
428 }
429 return 0;
430}
431
432/**
433 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
434 * @c: UBIFS file-system description object
435 * @lp: describes the LEB to garbage collect
436 *
437 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
438 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
439 * required, and other negative error codes in case of failures.
440 */
441int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
442{
443 struct ubifs_scan_leb *sleb;
444 struct ubifs_scan_node *snod;
445 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
446 int err = 0, lnum = lp->lnum;
447
448 ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
449 c->need_recovery);
450 ubifs_assert(c->gc_lnum != lnum);
451 ubifs_assert(wbuf->lnum != lnum);
452
453 /*
454 * We scan the entire LEB even though we only really need to scan up to
455 * (c->leb_size - lp->free).
456 */
Artem Bityutskiy348709b2009-08-25 15:00:55 +0300457 sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300458 if (IS_ERR(sleb))
459 return PTR_ERR(sleb);
460
461 ubifs_assert(!list_empty(&sleb->nodes));
462 snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
463
464 if (snod->type == UBIFS_IDX_NODE) {
465 struct ubifs_gced_idx_leb *idx_gc;
466
467 dbg_gc("indexing LEB %d (free %d, dirty %d)",
468 lnum, lp->free, lp->dirty);
469 list_for_each_entry(snod, &sleb->nodes, list) {
470 struct ubifs_idx_node *idx = snod->node;
471 int level = le16_to_cpu(idx->level);
472
473 ubifs_assert(snod->type == UBIFS_IDX_NODE);
474 key_read(c, ubifs_idx_key(c, idx), &snod->key);
475 err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
476 snod->offs);
477 if (err)
478 goto out;
479 }
480
481 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
482 if (!idx_gc) {
483 err = -ENOMEM;
484 goto out;
485 }
486
487 idx_gc->lnum = lnum;
488 idx_gc->unmap = 0;
489 list_add(&idx_gc->list, &c->idx_gc);
490
491 /*
492 * Don't release the LEB until after the next commit, because
Adrian Hunter227c75c2009-01-29 11:53:51 +0200493 * it may contain data which is needed for recovery. So
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300494 * although we freed this LEB, it will become usable only after
495 * the commit.
496 */
497 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
498 LPROPS_INDEX, 1);
499 if (err)
500 goto out;
501 err = LEB_FREED_IDX;
502 } else {
503 dbg_gc("data LEB %d (free %d, dirty %d)",
504 lnum, lp->free, lp->dirty);
505
506 err = move_nodes(c, sleb);
507 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300508 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300509
510 err = gc_sync_wbufs(c);
511 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300512 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300513
514 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
515 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300516 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300517
Adrian Hunter601c0bc2008-08-22 14:23:35 +0300518 /* Allow for races with TNC */
519 c->gced_lnum = lnum;
520 smp_wmb();
521 c->gc_seq += 1;
522 smp_wmb();
523
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300524 if (c->gc_lnum == -1) {
525 c->gc_lnum = lnum;
526 err = LEB_RETAINED;
527 } else {
528 err = ubifs_wbuf_sync_nolock(wbuf);
529 if (err)
530 goto out;
531
532 err = ubifs_leb_unmap(c, lnum);
533 if (err)
534 goto out;
535
536 err = LEB_FREED;
537 }
538 }
539
540out:
541 ubifs_scan_destroy(sleb);
542 return err;
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300543
544out_inc_seq:
545 /* We may have moved at least some nodes so allow for races with TNC */
546 c->gced_lnum = lnum;
547 smp_wmb();
548 c->gc_seq += 1;
549 smp_wmb();
550 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300551}
552
553/**
554 * ubifs_garbage_collect - UBIFS garbage collector.
555 * @c: UBIFS file-system description object
556 * @anyway: do GC even if there are free LEBs
557 *
558 * This function does out-of-place garbage collection. The return codes are:
559 * o positive LEB number if the LEB has been freed and may be used;
560 * o %-EAGAIN if the caller has to run commit;
561 * o %-ENOSPC if GC failed to make any progress;
562 * o other negative error codes in case of other errors.
563 *
564 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
565 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
566 * commit may be required. But commit cannot be run from inside GC, because the
567 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
568 * And this error code means that the caller has to run commit, and re-run GC
569 * if there is still no free space.
570 *
571 * There are many reasons why this function may return %-EAGAIN:
572 * o the log is full and there is no space to write an LEB reference for
573 * @c->gc_lnum;
574 * o the journal is too large and exceeds size limitations;
575 * o GC moved indexing LEBs, but they can be used only after the commit;
576 * o the shrinker fails to find clean znodes to free and requests the commit;
577 * o etc.
578 *
579 * Note, if the file-system is close to be full, this function may return
580 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
581 * the function. E.g., this happens if the limits on the journal size are too
582 * tough and GC writes too much to the journal before an LEB is freed. This
583 * might also mean that the journal is too large, and the TNC becomes to big,
584 * so that the shrinker is constantly called, finds not clean znodes to free,
585 * and requests commit. Well, this may also happen if the journal is all right,
586 * but another kernel process consumes too much memory. Anyway, infinite
587 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
588 */
589int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
590{
591 int i, err, ret, min_space = c->dead_wm;
592 struct ubifs_lprops lp;
593 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
594
595 ubifs_assert_cmt_locked(c);
596
597 if (ubifs_gc_should_commit(c))
598 return -EAGAIN;
599
600 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
601
602 if (c->ro_media) {
603 ret = -EROFS;
604 goto out_unlock;
605 }
606
607 /* We expect the write-buffer to be empty on entry */
608 ubifs_assert(!wbuf->used);
609
610 for (i = 0; ; i++) {
611 int space_before = c->leb_size - wbuf->offs - wbuf->used;
612 int space_after;
613
614 cond_resched();
615
616 /* Give the commit an opportunity to run */
617 if (ubifs_gc_should_commit(c)) {
618 ret = -EAGAIN;
619 break;
620 }
621
622 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
623 /*
624 * We've done enough iterations. Indexing LEBs were
625 * moved and will be available after the commit.
626 */
627 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
628 ubifs_commit_required(c);
629 ret = -EAGAIN;
630 break;
631 }
632
633 if (i > HARD_LEBS_LIMIT) {
634 /*
635 * We've moved too many LEBs and have not made
636 * progress, give up.
637 */
638 dbg_gc("hard limit, -ENOSPC");
639 ret = -ENOSPC;
640 break;
641 }
642
643 /*
644 * Empty and freeable LEBs can turn up while we waited for
645 * the wbuf lock, or while we have been running GC. In that
646 * case, we should just return one of those instead of
647 * continuing to GC dirty LEBs. Hence we request
648 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
649 */
650 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
651 if (ret) {
652 if (ret == -ENOSPC)
653 dbg_gc("no more dirty LEBs");
654 break;
655 }
656
657 dbg_gc("found LEB %d: free %d, dirty %d, sum %d "
658 "(min. space %d)", lp.lnum, lp.free, lp.dirty,
659 lp.free + lp.dirty, min_space);
660
661 if (lp.free + lp.dirty == c->leb_size) {
662 /* An empty LEB was returned */
663 dbg_gc("LEB %d is free, return it", lp.lnum);
664 /*
665 * ubifs_find_dirty_leb() doesn't return freeable index
666 * LEBs.
667 */
668 ubifs_assert(!(lp.flags & LPROPS_INDEX));
669 if (lp.free != c->leb_size) {
670 /*
671 * Write buffers must be sync'd before
672 * unmapping freeable LEBs, because one of them
673 * may contain data which obsoletes something
674 * in 'lp.pnum'.
675 */
676 ret = gc_sync_wbufs(c);
677 if (ret)
678 goto out;
679 ret = ubifs_change_one_lp(c, lp.lnum,
680 c->leb_size, 0, 0, 0,
681 0);
682 if (ret)
683 goto out;
684 }
685 ret = ubifs_leb_unmap(c, lp.lnum);
686 if (ret)
687 goto out;
688 ret = lp.lnum;
689 break;
690 }
691
692 space_before = c->leb_size - wbuf->offs - wbuf->used;
693 if (wbuf->lnum == -1)
694 space_before = 0;
695
696 ret = ubifs_garbage_collect_leb(c, &lp);
697 if (ret < 0) {
Artem Bityutskiyefe18812010-08-04 14:06:06 +0300698 if (ret == -EAGAIN) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300699 /*
Artem Bityutskiyefe18812010-08-04 14:06:06 +0300700 * This is not error, so we have to return the
701 * LEB to lprops. But if 'ubifs_return_leb()'
702 * fails, its failure code is propagated to the
703 * caller instead of the original '-EAGAIN'.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300704 */
705 err = ubifs_return_leb(c, lp.lnum);
706 if (err)
707 ret = err;
708 break;
709 }
710 goto out;
711 }
712
713 if (ret == LEB_FREED) {
714 /* An LEB has been freed and is ready for use */
715 dbg_gc("LEB %d freed, return", lp.lnum);
716 ret = lp.lnum;
717 break;
718 }
719
720 if (ret == LEB_FREED_IDX) {
721 /*
722 * This was an indexing LEB and it cannot be
723 * immediately used. And instead of requesting the
724 * commit straight away, we try to garbage collect some
725 * more.
726 */
727 dbg_gc("indexing LEB %d freed, continue", lp.lnum);
728 continue;
729 }
730
731 ubifs_assert(ret == LEB_RETAINED);
732 space_after = c->leb_size - wbuf->offs - wbuf->used;
733 dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
734 space_after - space_before);
735
736 if (space_after > space_before) {
737 /* GC makes progress, keep working */
738 min_space >>= 1;
739 if (min_space < c->dead_wm)
740 min_space = c->dead_wm;
741 continue;
742 }
743
744 dbg_gc("did not make progress");
745
746 /*
747 * GC moved an LEB bud have not done any progress. This means
748 * that the previous GC head LEB contained too few free space
749 * and the LEB which was GC'ed contained only large nodes which
750 * did not fit that space.
751 *
752 * We can do 2 things:
753 * 1. pick another LEB in a hope it'll contain a small node
754 * which will fit the space we have at the end of current GC
755 * head LEB, but there is no guarantee, so we try this out
756 * unless we have already been working for too long;
757 * 2. request an LEB with more dirty space, which will force
758 * 'ubifs_find_dirty_leb()' to start scanning the lprops
759 * table, instead of just picking one from the heap
760 * (previously it already picked the dirtiest LEB).
761 */
762 if (i < SOFT_LEBS_LIMIT) {
763 dbg_gc("try again");
764 continue;
765 }
766
767 min_space <<= 1;
768 if (min_space > c->dark_wm)
769 min_space = c->dark_wm;
770 dbg_gc("set min. space to %d", min_space);
771 }
772
773 if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
774 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
775 ubifs_commit_required(c);
776 ret = -EAGAIN;
777 }
778
779 err = ubifs_wbuf_sync_nolock(wbuf);
780 if (!err)
781 err = ubifs_leb_unmap(c, c->gc_lnum);
782 if (err) {
783 ret = err;
784 goto out;
785 }
786out_unlock:
787 mutex_unlock(&wbuf->io_mutex);
788 return ret;
789
790out:
791 ubifs_assert(ret < 0);
792 ubifs_assert(ret != -ENOSPC && ret != -EAGAIN);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300793 ubifs_wbuf_sync_nolock(wbuf);
Artem Bityutskiy5ffef882010-08-04 10:14:47 +0300794 ubifs_ro_mode(c, ret);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300795 mutex_unlock(&wbuf->io_mutex);
796 ubifs_return_leb(c, lp.lnum);
797 return ret;
798}
799
800/**
801 * ubifs_gc_start_commit - garbage collection at start of commit.
802 * @c: UBIFS file-system description object
803 *
804 * If a LEB has only dirty and free space, then we may safely unmap it and make
805 * it free. Note, we cannot do this with indexing LEBs because dirty space may
806 * correspond index nodes that are required for recovery. In that case, the
807 * LEB cannot be unmapped until after the next commit.
808 *
809 * This function returns %0 upon success and a negative error code upon failure.
810 */
811int ubifs_gc_start_commit(struct ubifs_info *c)
812{
813 struct ubifs_gced_idx_leb *idx_gc;
814 const struct ubifs_lprops *lp;
815 int err = 0, flags;
816
817 ubifs_get_lprops(c);
818
819 /*
820 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
821 * wbufs are sync'd before this, which is done in 'do_commit()'.
822 */
823 while (1) {
824 lp = ubifs_fast_find_freeable(c);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300825 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300826 err = PTR_ERR(lp);
827 goto out;
828 }
829 if (!lp)
830 break;
831 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
832 ubifs_assert(!(lp->flags & LPROPS_INDEX));
833 err = ubifs_leb_unmap(c, lp->lnum);
834 if (err)
835 goto out;
836 lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300837 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300838 err = PTR_ERR(lp);
839 goto out;
840 }
841 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
842 ubifs_assert(!(lp->flags & LPROPS_INDEX));
843 }
844
845 /* Mark GC'd index LEBs OK to unmap after this commit finishes */
846 list_for_each_entry(idx_gc, &c->idx_gc, list)
847 idx_gc->unmap = 1;
848
849 /* Record index freeable LEBs for unmapping after commit */
850 while (1) {
851 lp = ubifs_fast_find_frdi_idx(c);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300852 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300853 err = PTR_ERR(lp);
854 goto out;
855 }
856 if (!lp)
857 break;
858 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
859 if (!idx_gc) {
860 err = -ENOMEM;
861 goto out;
862 }
863 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
864 ubifs_assert(lp->flags & LPROPS_INDEX);
865 /* Don't release the LEB until after the next commit */
866 flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
867 lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300868 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300869 err = PTR_ERR(lp);
870 kfree(idx_gc);
871 goto out;
872 }
873 ubifs_assert(lp->flags & LPROPS_TAKEN);
874 ubifs_assert(!(lp->flags & LPROPS_INDEX));
875 idx_gc->lnum = lp->lnum;
876 idx_gc->unmap = 1;
877 list_add(&idx_gc->list, &c->idx_gc);
878 }
879out:
880 ubifs_release_lprops(c);
881 return err;
882}
883
884/**
885 * ubifs_gc_end_commit - garbage collection at end of commit.
886 * @c: UBIFS file-system description object
887 *
888 * This function completes out-of-place garbage collection of index LEBs.
889 */
890int ubifs_gc_end_commit(struct ubifs_info *c)
891{
892 struct ubifs_gced_idx_leb *idx_gc, *tmp;
893 struct ubifs_wbuf *wbuf;
894 int err = 0;
895
896 wbuf = &c->jheads[GCHD].wbuf;
897 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
898 list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
899 if (idx_gc->unmap) {
900 dbg_gc("LEB %d", idx_gc->lnum);
901 err = ubifs_leb_unmap(c, idx_gc->lnum);
902 if (err)
903 goto out;
904 err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
905 LPROPS_NC, 0, LPROPS_TAKEN, -1);
906 if (err)
907 goto out;
908 list_del(&idx_gc->list);
909 kfree(idx_gc);
910 }
911out:
912 mutex_unlock(&wbuf->io_mutex);
913 return err;
914}
915
916/**
917 * ubifs_destroy_idx_gc - destroy idx_gc list.
918 * @c: UBIFS file-system description object
919 *
Adrian Hunterb466f172009-01-29 12:59:33 +0200920 * This function destroys the @c->idx_gc list. It is called when unmounting
921 * so locks are not needed. Returns zero in case of success and a negative
922 * error code in case of failure.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300923 */
Adrian Hunterb466f172009-01-29 12:59:33 +0200924void ubifs_destroy_idx_gc(struct ubifs_info *c)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300925{
926 while (!list_empty(&c->idx_gc)) {
927 struct ubifs_gced_idx_leb *idx_gc;
928
929 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
930 list);
Adrian Hunterb466f172009-01-29 12:59:33 +0200931 c->idx_gc_cnt -= 1;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300932 list_del(&idx_gc->list);
933 kfree(idx_gc);
934 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300935}
936
937/**
938 * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
939 * @c: UBIFS file-system description object
940 *
941 * Called during start commit so locks are not needed.
942 */
943int ubifs_get_idx_gc_leb(struct ubifs_info *c)
944{
945 struct ubifs_gced_idx_leb *idx_gc;
946 int lnum;
947
948 if (list_empty(&c->idx_gc))
949 return -ENOSPC;
950 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
951 lnum = idx_gc->lnum;
952 /* c->idx_gc_cnt is updated by the caller when lprops are updated */
953 list_del(&idx_gc->list);
954 kfree(idx_gc);
955 return lnum;
956}