blob: dafef3d195d3321eb596bb38d8736ca086ed3847 [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);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300130
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200131 ubifs_assert(key_type(c, &sa->key) == UBIFS_DATA_KEY);
132 ubifs_assert(key_type(c, &sb->key) == UBIFS_DATA_KEY);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300133 ubifs_assert(sa->type == UBIFS_DATA_NODE);
134 ubifs_assert(sb->type == UBIFS_DATA_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200135
136 inuma = key_inum(c, &sa->key);
137 inumb = key_inum(c, &sb->key);
138
139 if (inuma == inumb) {
140 unsigned int blka = key_block(c, &sa->key);
141 unsigned int blkb = key_block(c, &sb->key);
142
143 if (blka <= blkb)
144 return -1;
145 } else if (inuma <= inumb)
146 return -1;
147
148 return 1;
149}
150
151/*
152 * nondata_nodes_cmp - compare 2 non-data nodes.
153 * @priv: UBIFS file-system description object
154 * @a: first node
155 * @a: second node
156 *
157 * This function compares nodes @a and @b. It makes sure that inode nodes go
158 * first and sorted by length in descending order. Directory entry nodes go
159 * after inode nodes and are sorted in ascending hash valuer order.
160 */
161int nondata_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
162{
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200163 ino_t inuma, inumb;
164 struct ubifs_info *c = priv;
165 struct ubifs_scan_node *sa, *sb;
166
167 cond_resched();
168 sa = list_entry(a, struct ubifs_scan_node, list);
169 sb = list_entry(b, struct ubifs_scan_node, list);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300170
171 ubifs_assert(key_type(c, &sa->key) != UBIFS_DATA_KEY &&
172 key_type(c, &sb->key) != UBIFS_DATA_KEY);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300173 ubifs_assert(sa->type != UBIFS_DATA_NODE &&
174 sb->type != UBIFS_DATA_NODE);
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200175
176 /* Inodes go before directory entries */
Artem Bityutskiyab871182010-08-08 12:25:33 +0300177 if (sa->type == UBIFS_INO_NODE) {
178 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200179 return sb->len - sa->len;
180 return -1;
181 }
Artem Bityutskiyab871182010-08-08 12:25:33 +0300182 if (sb->type == UBIFS_INO_NODE)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200183 return 1;
184
Artem Bityutskiy66576832010-08-08 12:29:58 +0300185 ubifs_assert(key_type(c, &sa->key) == UBIFS_DENT_KEY ||
186 key_type(c, &sa->key) == UBIFS_XENT_KEY);
187 ubifs_assert(key_type(c, &sb->key) == UBIFS_DENT_KEY ||
188 key_type(c, &sb->key) == UBIFS_XENT_KEY);
Artem Bityutskiyab871182010-08-08 12:25:33 +0300189 ubifs_assert(sa->type == UBIFS_DENT_NODE ||
190 sa->type == UBIFS_XENT_NODE);
191 ubifs_assert(sb->type == UBIFS_DENT_NODE ||
192 sb->type == UBIFS_XENT_NODE);
Artem Bityutskiy66576832010-08-08 12:29:58 +0300193
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200194 inuma = key_inum(c, &sa->key);
195 inumb = key_inum(c, &sb->key);
196
197 if (inuma == inumb) {
198 uint32_t hasha = key_hash(c, &sa->key);
199 uint32_t hashb = key_hash(c, &sb->key);
200
201 if (hasha <= hashb)
202 return -1;
203 } else if (inuma <= inumb)
204 return -1;
205
206 return 1;
207}
208
209/**
210 * sort_nodes - sort nodes for GC.
211 * @c: UBIFS file-system description object
212 * @sleb: describes nodes to sort and contains the result on exit
213 * @nondata: contains non-data nodes on exit
214 * @min: minimum node size is returned here
215 *
216 * This function sorts the list of inodes to garbage collect. First of all, it
217 * kills obsolete nodes and separates data and non-data nodes to the
218 * @sleb->nodes and @nondata lists correspondingly.
219 *
220 * Data nodes are then sorted in block number order - this is important for
221 * bulk-read; data nodes with lower inode number go before data nodes with
222 * higher inode number, and data nodes with lower block number go before data
223 * nodes with higher block number;
224 *
225 * Non-data nodes are sorted as follows.
226 * o First go inode nodes - they are sorted in descending length order.
227 * o Then go directory entry nodes - they are sorted in hash order, which
228 * should supposedly optimize 'readdir()'. Direntry nodes with lower parent
229 * inode number go before direntry nodes with higher parent inode number,
230 * and direntry nodes with lower name hash values go before direntry nodes
231 * with higher name hash values.
232 *
233 * This function returns zero in case of success and a negative error code in
234 * case of failure.
235 */
236static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
237 struct list_head *nondata, int *min)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300238{
239 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300240
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200241 *min = INT_MAX;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300242
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200243 /* Separate data nodes and non-data nodes */
244 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
245 int err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300246
Artem Bityutskiy44ec83b2010-08-07 08:44:13 +0300247 ubifs_assert(snod->type == UBIFS_INO_NODE ||
248 snod->type == UBIFS_DATA_NODE ||
249 snod->type == UBIFS_DENT_NODE ||
250 snod->type == UBIFS_XENT_NODE ||
251 snod->type == UBIFS_TRUN_NODE);
252
253 if (snod->type != UBIFS_INO_NODE &&
254 snod->type != UBIFS_DATA_NODE &&
255 snod->type != UBIFS_DENT_NODE &&
256 snod->type != UBIFS_XENT_NODE) {
257 /* Probably truncation node, zap it */
258 list_del(&snod->list);
259 kfree(snod);
260 continue;
261 }
262
263 ubifs_assert(key_type(c, &snod->key) == UBIFS_DATA_KEY ||
264 key_type(c, &snod->key) == UBIFS_INO_KEY ||
265 key_type(c, &snod->key) == UBIFS_DENT_KEY ||
266 key_type(c, &snod->key) == UBIFS_XENT_KEY);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300267
268 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
269 snod->offs, 0);
270 if (err < 0)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200271 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300272
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300273 if (!err) {
274 /* The node is obsolete, remove it from the list */
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200275 list_del(&snod->list);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300276 kfree(snod);
277 continue;
278 }
279
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200280 if (snod->len < *min)
281 *min = snod->len;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300282
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200283 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
284 list_move_tail(&snod->list, nondata);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300285 }
286
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200287 /* Sort data and non-data nodes */
288 list_sort(c, &sleb->nodes, &data_nodes_cmp);
289 list_sort(c, nondata, &nondata_nodes_cmp);
290 return 0;
291}
292
293/**
294 * move_node - move a node.
295 * @c: UBIFS file-system description object
296 * @sleb: describes the LEB to move nodes from
297 * @snod: the mode to move
298 * @wbuf: write-buffer to move node to
299 *
300 * This function moves node @snod to @wbuf, changes TNC correspondingly, and
301 * destroys @snod. Returns zero in case of success and a negative error code in
302 * case of failure.
303 */
304static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
305 struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
306{
307 int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
308
309 cond_resched();
310 err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
311 if (err)
312 return err;
313
314 err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
315 snod->offs, new_lnum, new_offs,
316 snod->len);
317 list_del(&snod->list);
318 kfree(snod);
319 return err;
320}
321
322/**
323 * move_nodes - move nodes.
324 * @c: UBIFS file-system description object
325 * @sleb: describes the LEB to move nodes from
326 *
327 * This function moves valid nodes from data LEB described by @sleb to the GC
328 * journal head. This function returns zero in case of success, %-EAGAIN if
329 * commit is required, and other negative error codes in case of other
330 * failures.
331 */
332static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
333{
334 int err, min;
335 LIST_HEAD(nondata);
336 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300337
338 if (wbuf->lnum == -1) {
339 /*
340 * The GC journal head is not set, because it is the first GC
341 * invocation since mount.
342 */
343 err = switch_gc_head(c);
344 if (err)
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200345 return err;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300346 }
347
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200348 err = sort_nodes(c, sleb, &nondata, &min);
349 if (err)
350 goto out;
351
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300352 /* Write nodes to their new location. Use the first-fit strategy */
353 while (1) {
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200354 int avail;
355 struct ubifs_scan_node *snod, *tmp;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300356
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200357 /* Move data nodes */
358 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
359 avail = c->leb_size - wbuf->offs - wbuf->used;
360 if (snod->len > avail)
361 /*
362 * Do not skip data nodes in order to optimize
363 * bulk-read.
364 */
365 break;
366
367 err = move_node(c, sleb, snod, wbuf);
368 if (err)
369 goto out;
370 }
371
372 /* Move non-data nodes */
373 list_for_each_entry_safe(snod, tmp, &nondata, list) {
374 avail = c->leb_size - wbuf->offs - wbuf->used;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300375 if (avail < min)
376 break;
377
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200378 if (snod->len > avail) {
379 /*
380 * Keep going only if this is an inode with
381 * some data. Otherwise stop and switch the GC
382 * head. IOW, we assume that data-less inode
383 * nodes and direntry nodes are roughly of the
384 * same size.
385 */
386 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
387 snod->len == UBIFS_INO_NODE_SZ)
388 break;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300389 continue;
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200390 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300391
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200392 err = move_node(c, sleb, snod, wbuf);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300393 if (err)
394 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300395 }
396
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200397 if (list_empty(&sleb->nodes) && list_empty(&nondata))
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300398 break;
399
400 /*
401 * Waste the rest of the space in the LEB and switch to the
402 * next LEB.
403 */
404 err = switch_gc_head(c);
405 if (err)
406 goto out;
407 }
408
409 return 0;
410
411out:
Artem Bityutskiyf10770f2009-03-08 15:13:00 +0200412 list_splice_tail(&nondata, &sleb->nodes);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300413 return err;
414}
415
416/**
417 * gc_sync_wbufs - sync write-buffers for GC.
418 * @c: UBIFS file-system description object
419 *
420 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
421 * be in a write-buffer instead. That is, a node could be written to a
422 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
423 * erased before the write-buffer is sync'd and then there is an unclean
424 * unmount, then an existing node is lost. To avoid this, we sync all
425 * write-buffers.
426 *
427 * This function returns %0 on success or a negative error code on failure.
428 */
429static int gc_sync_wbufs(struct ubifs_info *c)
430{
431 int err, i;
432
433 for (i = 0; i < c->jhead_cnt; i++) {
434 if (i == GCHD)
435 continue;
436 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
437 if (err)
438 return err;
439 }
440 return 0;
441}
442
443/**
444 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
445 * @c: UBIFS file-system description object
446 * @lp: describes the LEB to garbage collect
447 *
448 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
449 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
450 * required, and other negative error codes in case of failures.
451 */
452int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
453{
454 struct ubifs_scan_leb *sleb;
455 struct ubifs_scan_node *snod;
456 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
457 int err = 0, lnum = lp->lnum;
458
459 ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
460 c->need_recovery);
461 ubifs_assert(c->gc_lnum != lnum);
462 ubifs_assert(wbuf->lnum != lnum);
463
464 /*
465 * We scan the entire LEB even though we only really need to scan up to
466 * (c->leb_size - lp->free).
467 */
Artem Bityutskiy348709b2009-08-25 15:00:55 +0300468 sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300469 if (IS_ERR(sleb))
470 return PTR_ERR(sleb);
471
472 ubifs_assert(!list_empty(&sleb->nodes));
473 snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
474
475 if (snod->type == UBIFS_IDX_NODE) {
476 struct ubifs_gced_idx_leb *idx_gc;
477
478 dbg_gc("indexing LEB %d (free %d, dirty %d)",
479 lnum, lp->free, lp->dirty);
480 list_for_each_entry(snod, &sleb->nodes, list) {
481 struct ubifs_idx_node *idx = snod->node;
482 int level = le16_to_cpu(idx->level);
483
484 ubifs_assert(snod->type == UBIFS_IDX_NODE);
485 key_read(c, ubifs_idx_key(c, idx), &snod->key);
486 err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
487 snod->offs);
488 if (err)
489 goto out;
490 }
491
492 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
493 if (!idx_gc) {
494 err = -ENOMEM;
495 goto out;
496 }
497
498 idx_gc->lnum = lnum;
499 idx_gc->unmap = 0;
500 list_add(&idx_gc->list, &c->idx_gc);
501
502 /*
503 * Don't release the LEB until after the next commit, because
Adrian Hunter227c75c2009-01-29 11:53:51 +0200504 * it may contain data which is needed for recovery. So
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300505 * although we freed this LEB, it will become usable only after
506 * the commit.
507 */
508 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
509 LPROPS_INDEX, 1);
510 if (err)
511 goto out;
512 err = LEB_FREED_IDX;
513 } else {
514 dbg_gc("data LEB %d (free %d, dirty %d)",
515 lnum, lp->free, lp->dirty);
516
517 err = move_nodes(c, sleb);
518 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300519 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300520
521 err = gc_sync_wbufs(c);
522 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300523 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300524
525 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
526 if (err)
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300527 goto out_inc_seq;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300528
Adrian Hunter601c0bc2008-08-22 14:23:35 +0300529 /* Allow for races with TNC */
530 c->gced_lnum = lnum;
531 smp_wmb();
532 c->gc_seq += 1;
533 smp_wmb();
534
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300535 if (c->gc_lnum == -1) {
536 c->gc_lnum = lnum;
537 err = LEB_RETAINED;
538 } else {
539 err = ubifs_wbuf_sync_nolock(wbuf);
540 if (err)
541 goto out;
542
543 err = ubifs_leb_unmap(c, lnum);
544 if (err)
545 goto out;
546
547 err = LEB_FREED;
548 }
549 }
550
551out:
552 ubifs_scan_destroy(sleb);
553 return err;
Adrian Hunter6dcfac42008-09-12 12:27:47 +0300554
555out_inc_seq:
556 /* We may have moved at least some nodes so allow for races with TNC */
557 c->gced_lnum = lnum;
558 smp_wmb();
559 c->gc_seq += 1;
560 smp_wmb();
561 goto out;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300562}
563
564/**
565 * ubifs_garbage_collect - UBIFS garbage collector.
566 * @c: UBIFS file-system description object
567 * @anyway: do GC even if there are free LEBs
568 *
569 * This function does out-of-place garbage collection. The return codes are:
570 * o positive LEB number if the LEB has been freed and may be used;
571 * o %-EAGAIN if the caller has to run commit;
572 * o %-ENOSPC if GC failed to make any progress;
573 * o other negative error codes in case of other errors.
574 *
575 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
576 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
577 * commit may be required. But commit cannot be run from inside GC, because the
578 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
579 * And this error code means that the caller has to run commit, and re-run GC
580 * if there is still no free space.
581 *
582 * There are many reasons why this function may return %-EAGAIN:
583 * o the log is full and there is no space to write an LEB reference for
584 * @c->gc_lnum;
585 * o the journal is too large and exceeds size limitations;
586 * o GC moved indexing LEBs, but they can be used only after the commit;
587 * o the shrinker fails to find clean znodes to free and requests the commit;
588 * o etc.
589 *
590 * Note, if the file-system is close to be full, this function may return
591 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
592 * the function. E.g., this happens if the limits on the journal size are too
593 * tough and GC writes too much to the journal before an LEB is freed. This
594 * might also mean that the journal is too large, and the TNC becomes to big,
595 * so that the shrinker is constantly called, finds not clean znodes to free,
596 * and requests commit. Well, this may also happen if the journal is all right,
597 * but another kernel process consumes too much memory. Anyway, infinite
598 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
599 */
600int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
601{
602 int i, err, ret, min_space = c->dead_wm;
603 struct ubifs_lprops lp;
604 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
605
606 ubifs_assert_cmt_locked(c);
607
608 if (ubifs_gc_should_commit(c))
609 return -EAGAIN;
610
611 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
612
613 if (c->ro_media) {
614 ret = -EROFS;
615 goto out_unlock;
616 }
617
618 /* We expect the write-buffer to be empty on entry */
619 ubifs_assert(!wbuf->used);
620
621 for (i = 0; ; i++) {
622 int space_before = c->leb_size - wbuf->offs - wbuf->used;
623 int space_after;
624
625 cond_resched();
626
627 /* Give the commit an opportunity to run */
628 if (ubifs_gc_should_commit(c)) {
629 ret = -EAGAIN;
630 break;
631 }
632
633 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
634 /*
635 * We've done enough iterations. Indexing LEBs were
636 * moved and will be available after the commit.
637 */
638 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
639 ubifs_commit_required(c);
640 ret = -EAGAIN;
641 break;
642 }
643
644 if (i > HARD_LEBS_LIMIT) {
645 /*
646 * We've moved too many LEBs and have not made
647 * progress, give up.
648 */
649 dbg_gc("hard limit, -ENOSPC");
650 ret = -ENOSPC;
651 break;
652 }
653
654 /*
655 * Empty and freeable LEBs can turn up while we waited for
656 * the wbuf lock, or while we have been running GC. In that
657 * case, we should just return one of those instead of
658 * continuing to GC dirty LEBs. Hence we request
659 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
660 */
661 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
662 if (ret) {
663 if (ret == -ENOSPC)
664 dbg_gc("no more dirty LEBs");
665 break;
666 }
667
668 dbg_gc("found LEB %d: free %d, dirty %d, sum %d "
669 "(min. space %d)", lp.lnum, lp.free, lp.dirty,
670 lp.free + lp.dirty, min_space);
671
672 if (lp.free + lp.dirty == c->leb_size) {
673 /* An empty LEB was returned */
674 dbg_gc("LEB %d is free, return it", lp.lnum);
675 /*
676 * ubifs_find_dirty_leb() doesn't return freeable index
677 * LEBs.
678 */
679 ubifs_assert(!(lp.flags & LPROPS_INDEX));
680 if (lp.free != c->leb_size) {
681 /*
682 * Write buffers must be sync'd before
683 * unmapping freeable LEBs, because one of them
684 * may contain data which obsoletes something
685 * in 'lp.pnum'.
686 */
687 ret = gc_sync_wbufs(c);
688 if (ret)
689 goto out;
690 ret = ubifs_change_one_lp(c, lp.lnum,
691 c->leb_size, 0, 0, 0,
692 0);
693 if (ret)
694 goto out;
695 }
696 ret = ubifs_leb_unmap(c, lp.lnum);
697 if (ret)
698 goto out;
699 ret = lp.lnum;
700 break;
701 }
702
703 space_before = c->leb_size - wbuf->offs - wbuf->used;
704 if (wbuf->lnum == -1)
705 space_before = 0;
706
707 ret = ubifs_garbage_collect_leb(c, &lp);
708 if (ret < 0) {
Artem Bityutskiyefe18812010-08-04 14:06:06 +0300709 if (ret == -EAGAIN) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300710 /*
Artem Bityutskiyefe18812010-08-04 14:06:06 +0300711 * This is not error, so we have to return the
712 * LEB to lprops. But if 'ubifs_return_leb()'
713 * fails, its failure code is propagated to the
714 * caller instead of the original '-EAGAIN'.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300715 */
716 err = ubifs_return_leb(c, lp.lnum);
717 if (err)
718 ret = err;
719 break;
720 }
721 goto out;
722 }
723
724 if (ret == LEB_FREED) {
725 /* An LEB has been freed and is ready for use */
726 dbg_gc("LEB %d freed, return", lp.lnum);
727 ret = lp.lnum;
728 break;
729 }
730
731 if (ret == LEB_FREED_IDX) {
732 /*
733 * This was an indexing LEB and it cannot be
734 * immediately used. And instead of requesting the
735 * commit straight away, we try to garbage collect some
736 * more.
737 */
738 dbg_gc("indexing LEB %d freed, continue", lp.lnum);
739 continue;
740 }
741
742 ubifs_assert(ret == LEB_RETAINED);
743 space_after = c->leb_size - wbuf->offs - wbuf->used;
744 dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
745 space_after - space_before);
746
747 if (space_after > space_before) {
748 /* GC makes progress, keep working */
749 min_space >>= 1;
750 if (min_space < c->dead_wm)
751 min_space = c->dead_wm;
752 continue;
753 }
754
755 dbg_gc("did not make progress");
756
757 /*
758 * GC moved an LEB bud have not done any progress. This means
759 * that the previous GC head LEB contained too few free space
760 * and the LEB which was GC'ed contained only large nodes which
761 * did not fit that space.
762 *
763 * We can do 2 things:
764 * 1. pick another LEB in a hope it'll contain a small node
765 * which will fit the space we have at the end of current GC
766 * head LEB, but there is no guarantee, so we try this out
767 * unless we have already been working for too long;
768 * 2. request an LEB with more dirty space, which will force
769 * 'ubifs_find_dirty_leb()' to start scanning the lprops
770 * table, instead of just picking one from the heap
771 * (previously it already picked the dirtiest LEB).
772 */
773 if (i < SOFT_LEBS_LIMIT) {
774 dbg_gc("try again");
775 continue;
776 }
777
778 min_space <<= 1;
779 if (min_space > c->dark_wm)
780 min_space = c->dark_wm;
781 dbg_gc("set min. space to %d", min_space);
782 }
783
784 if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
785 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
786 ubifs_commit_required(c);
787 ret = -EAGAIN;
788 }
789
790 err = ubifs_wbuf_sync_nolock(wbuf);
791 if (!err)
792 err = ubifs_leb_unmap(c, c->gc_lnum);
793 if (err) {
794 ret = err;
795 goto out;
796 }
797out_unlock:
798 mutex_unlock(&wbuf->io_mutex);
799 return ret;
800
801out:
802 ubifs_assert(ret < 0);
803 ubifs_assert(ret != -ENOSPC && ret != -EAGAIN);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300804 ubifs_wbuf_sync_nolock(wbuf);
Artem Bityutskiy5ffef882010-08-04 10:14:47 +0300805 ubifs_ro_mode(c, ret);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300806 mutex_unlock(&wbuf->io_mutex);
807 ubifs_return_leb(c, lp.lnum);
808 return ret;
809}
810
811/**
812 * ubifs_gc_start_commit - garbage collection at start of commit.
813 * @c: UBIFS file-system description object
814 *
815 * If a LEB has only dirty and free space, then we may safely unmap it and make
816 * it free. Note, we cannot do this with indexing LEBs because dirty space may
817 * correspond index nodes that are required for recovery. In that case, the
818 * LEB cannot be unmapped until after the next commit.
819 *
820 * This function returns %0 upon success and a negative error code upon failure.
821 */
822int ubifs_gc_start_commit(struct ubifs_info *c)
823{
824 struct ubifs_gced_idx_leb *idx_gc;
825 const struct ubifs_lprops *lp;
826 int err = 0, flags;
827
828 ubifs_get_lprops(c);
829
830 /*
831 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
832 * wbufs are sync'd before this, which is done in 'do_commit()'.
833 */
834 while (1) {
835 lp = ubifs_fast_find_freeable(c);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300836 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300837 err = PTR_ERR(lp);
838 goto out;
839 }
840 if (!lp)
841 break;
842 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
843 ubifs_assert(!(lp->flags & LPROPS_INDEX));
844 err = ubifs_leb_unmap(c, lp->lnum);
845 if (err)
846 goto out;
847 lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300848 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300849 err = PTR_ERR(lp);
850 goto out;
851 }
852 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
853 ubifs_assert(!(lp->flags & LPROPS_INDEX));
854 }
855
856 /* Mark GC'd index LEBs OK to unmap after this commit finishes */
857 list_for_each_entry(idx_gc, &c->idx_gc, list)
858 idx_gc->unmap = 1;
859
860 /* Record index freeable LEBs for unmapping after commit */
861 while (1) {
862 lp = ubifs_fast_find_frdi_idx(c);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300863 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300864 err = PTR_ERR(lp);
865 goto out;
866 }
867 if (!lp)
868 break;
869 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
870 if (!idx_gc) {
871 err = -ENOMEM;
872 goto out;
873 }
874 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
875 ubifs_assert(lp->flags & LPROPS_INDEX);
876 /* Don't release the LEB until after the next commit */
877 flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
878 lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
Hirofumi Nakagawa8d47aef2008-08-21 17:16:40 +0300879 if (IS_ERR(lp)) {
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300880 err = PTR_ERR(lp);
881 kfree(idx_gc);
882 goto out;
883 }
884 ubifs_assert(lp->flags & LPROPS_TAKEN);
885 ubifs_assert(!(lp->flags & LPROPS_INDEX));
886 idx_gc->lnum = lp->lnum;
887 idx_gc->unmap = 1;
888 list_add(&idx_gc->list, &c->idx_gc);
889 }
890out:
891 ubifs_release_lprops(c);
892 return err;
893}
894
895/**
896 * ubifs_gc_end_commit - garbage collection at end of commit.
897 * @c: UBIFS file-system description object
898 *
899 * This function completes out-of-place garbage collection of index LEBs.
900 */
901int ubifs_gc_end_commit(struct ubifs_info *c)
902{
903 struct ubifs_gced_idx_leb *idx_gc, *tmp;
904 struct ubifs_wbuf *wbuf;
905 int err = 0;
906
907 wbuf = &c->jheads[GCHD].wbuf;
908 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
909 list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
910 if (idx_gc->unmap) {
911 dbg_gc("LEB %d", idx_gc->lnum);
912 err = ubifs_leb_unmap(c, idx_gc->lnum);
913 if (err)
914 goto out;
915 err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
916 LPROPS_NC, 0, LPROPS_TAKEN, -1);
917 if (err)
918 goto out;
919 list_del(&idx_gc->list);
920 kfree(idx_gc);
921 }
922out:
923 mutex_unlock(&wbuf->io_mutex);
924 return err;
925}
926
927/**
928 * ubifs_destroy_idx_gc - destroy idx_gc list.
929 * @c: UBIFS file-system description object
930 *
Adrian Hunterb466f172009-01-29 12:59:33 +0200931 * This function destroys the @c->idx_gc list. It is called when unmounting
932 * so locks are not needed. Returns zero in case of success and a negative
933 * error code in case of failure.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300934 */
Adrian Hunterb466f172009-01-29 12:59:33 +0200935void ubifs_destroy_idx_gc(struct ubifs_info *c)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300936{
937 while (!list_empty(&c->idx_gc)) {
938 struct ubifs_gced_idx_leb *idx_gc;
939
940 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
941 list);
Adrian Hunterb466f172009-01-29 12:59:33 +0200942 c->idx_gc_cnt -= 1;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300943 list_del(&idx_gc->list);
944 kfree(idx_gc);
945 }
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300946}
947
948/**
949 * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
950 * @c: UBIFS file-system description object
951 *
952 * Called during start commit so locks are not needed.
953 */
954int ubifs_get_idx_gc_leb(struct ubifs_info *c)
955{
956 struct ubifs_gced_idx_leb *idx_gc;
957 int lnum;
958
959 if (list_empty(&c->idx_gc))
960 return -ENOSPC;
961 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
962 lnum = idx_gc->lnum;
963 /* c->idx_gc_cnt is updated by the caller when lprops are updated */
964 list_del(&idx_gc->list);
965 kfree(idx_gc);
966 return lnum;
967}