]>
Commit | Line | Data |
---|---|---|
1e51764a AB |
1 | /* |
2 | * This file is part of UBIFS. | |
3 | * | |
4 | * Copyright (C) 2006-2008 Nokia Corporation. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify it | |
7 | * under the terms of the GNU General Public License version 2 as published by | |
8 | * the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but WITHOUT | |
11 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
12 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
13 | * more details. | |
14 | * | |
15 | * You should have received a copy of the GNU General Public License along with | |
16 | * this program; if not, write to the Free Software Foundation, Inc., 51 | |
17 | * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA | |
18 | * | |
19 | * Authors: Adrian Hunter | |
20 | * Artem Bityutskiy (Битюцкий Артём) | |
21 | */ | |
22 | ||
23 | /* | |
24 | * This file implements garbage collection. The procedure for garbage collection | |
25 | * is different depending on whether a LEB as an index LEB (contains index | |
26 | * nodes) or not. For non-index LEBs, garbage collection finds a LEB which | |
27 | * contains a lot of dirty space (obsolete nodes), and copies the non-obsolete | |
28 | * nodes to the journal, at which point the garbage-collected LEB is free to be | |
29 | * reused. For index LEBs, garbage collection marks the non-obsolete index nodes | |
30 | * dirty in the TNC, and after the next commit, the garbage-collected LEB is | |
31 | * to be reused. Garbage collection will cause the number of dirty index nodes | |
32 | * to grow, however sufficient space is reserved for the index to ensure the | |
33 | * commit will never run out of space. | |
7078202e AB |
34 | * |
35 | * Notes about dead watermark. At current UBIFS implementation we assume that | |
36 | * LEBs which have less than @c->dead_wm bytes of free + dirty space are full | |
37 | * and not worth garbage-collecting. The dead watermark is one min. I/O unit | |
38 | * size, or min. UBIFS node size, depending on what is greater. Indeed, UBIFS | |
39 | * Garbage Collector has to synchronize the GC head's write buffer before | |
40 | * returning, so this is about wasting one min. I/O unit. However, UBIFS GC can | |
41 | * actually reclaim even very small pieces of dirty space by garbage collecting | |
42 | * enough dirty LEBs, but we do not bother doing this at this implementation. | |
43 | * | |
44 | * Notes about dark watermark. The results of GC work depends on how big are | |
45 | * the UBIFS nodes GC deals with. Large nodes make GC waste more space. Indeed, | |
46 | * if GC move data from LEB A to LEB B and nodes in LEB A are large, GC would | |
47 | * have to waste large pieces of free space at the end of LEB B, because nodes | |
48 | * from LEB A would not fit. And the worst situation is when all nodes are of | |
49 | * maximum size. So dark watermark is the amount of free + dirty space in LEB | |
f10770f5 | 50 | * which are guaranteed to be reclaimable. If LEB has less space, the GC might |
7078202e AB |
51 | * be unable to reclaim it. So, LEBs with free + dirty greater than dark |
52 | * watermark are "good" LEBs from GC's point of few. The other LEBs are not so | |
53 | * good, and GC takes extra care when moving them. | |
1e51764a AB |
54 | */ |
55 | ||
56 | #include <linux/pagemap.h> | |
2c761270 | 57 | #include <linux/list_sort.h> |
1e51764a AB |
58 | #include "ubifs.h" |
59 | ||
1e51764a | 60 | /* |
025dfdaf | 61 | * GC may need to move more than one LEB to make progress. The below constants |
1e51764a AB |
62 | * define "soft" and "hard" limits on the number of LEBs the garbage collector |
63 | * may move. | |
64 | */ | |
65 | #define SOFT_LEBS_LIMIT 4 | |
66 | #define HARD_LEBS_LIMIT 32 | |
67 | ||
68 | /** | |
69 | * switch_gc_head - switch the garbage collection journal head. | |
70 | * @c: UBIFS file-system description object | |
71 | * @buf: buffer to write | |
72 | * @len: length of the buffer to write | |
73 | * @lnum: LEB number written is returned here | |
74 | * @offs: offset written is returned here | |
75 | * | |
76 | * This function switch the GC head to the next LEB which is reserved in | |
77 | * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required, | |
78 | * and other negative error code in case of failures. | |
79 | */ | |
80 | static int switch_gc_head(struct ubifs_info *c) | |
81 | { | |
82 | int err, gc_lnum = c->gc_lnum; | |
83 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
84 | ||
85 | ubifs_assert(gc_lnum != -1); | |
86 | dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)", | |
87 | wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum, | |
88 | c->leb_size - wbuf->offs - wbuf->used); | |
89 | ||
90 | err = ubifs_wbuf_sync_nolock(wbuf); | |
91 | if (err) | |
92 | return err; | |
93 | ||
94 | /* | |
95 | * The GC write-buffer was synchronized, we may safely unmap | |
96 | * 'c->gc_lnum'. | |
97 | */ | |
98 | err = ubifs_leb_unmap(c, gc_lnum); | |
99 | if (err) | |
100 | return err; | |
101 | ||
102 | err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0); | |
103 | if (err) | |
104 | return err; | |
105 | ||
106 | c->gc_lnum = -1; | |
107 | err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0, UBI_LONGTERM); | |
108 | return err; | |
109 | } | |
110 | ||
111 | /** | |
f10770f5 AB |
112 | * data_nodes_cmp - compare 2 data nodes. |
113 | * @priv: UBIFS file-system description object | |
114 | * @a: first data node | |
115 | * @a: second data node | |
116 | * | |
117 | * This function compares data nodes @a and @b. Returns %1 if @a has greater | |
118 | * inode or block number, and %-1 otherwise. | |
119 | */ | |
120 | int data_nodes_cmp(void *priv, struct list_head *a, struct list_head *b) | |
121 | { | |
122 | ino_t inuma, inumb; | |
123 | struct ubifs_info *c = priv; | |
124 | struct ubifs_scan_node *sa, *sb; | |
125 | ||
126 | cond_resched(); | |
127 | sa = list_entry(a, struct ubifs_scan_node, list); | |
128 | sb = list_entry(b, struct ubifs_scan_node, list); | |
129 | ubifs_assert(key_type(c, &sa->key) == UBIFS_DATA_KEY); | |
130 | ubifs_assert(key_type(c, &sb->key) == UBIFS_DATA_KEY); | |
131 | ||
132 | inuma = key_inum(c, &sa->key); | |
133 | inumb = key_inum(c, &sb->key); | |
134 | ||
135 | if (inuma == inumb) { | |
136 | unsigned int blka = key_block(c, &sa->key); | |
137 | unsigned int blkb = key_block(c, &sb->key); | |
138 | ||
139 | if (blka <= blkb) | |
140 | return -1; | |
141 | } else if (inuma <= inumb) | |
142 | return -1; | |
143 | ||
144 | return 1; | |
145 | } | |
146 | ||
147 | /* | |
148 | * nondata_nodes_cmp - compare 2 non-data nodes. | |
149 | * @priv: UBIFS file-system description object | |
150 | * @a: first node | |
151 | * @a: second node | |
152 | * | |
153 | * This function compares nodes @a and @b. It makes sure that inode nodes go | |
154 | * first and sorted by length in descending order. Directory entry nodes go | |
155 | * after inode nodes and are sorted in ascending hash valuer order. | |
156 | */ | |
157 | int nondata_nodes_cmp(void *priv, struct list_head *a, struct list_head *b) | |
158 | { | |
159 | int typea, typeb; | |
160 | 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); | |
167 | typea = key_type(c, &sa->key); | |
168 | typeb = key_type(c, &sb->key); | |
169 | ubifs_assert(typea != UBIFS_DATA_KEY && typeb != UBIFS_DATA_KEY); | |
170 | ||
171 | /* Inodes go before directory entries */ | |
172 | if (typea == UBIFS_INO_KEY) { | |
173 | if (typeb == UBIFS_INO_KEY) | |
174 | return sb->len - sa->len; | |
175 | return -1; | |
176 | } | |
177 | if (typeb == UBIFS_INO_KEY) | |
178 | return 1; | |
179 | ||
180 | ubifs_assert(typea == UBIFS_DENT_KEY && typeb == UBIFS_DENT_KEY); | |
181 | inuma = key_inum(c, &sa->key); | |
182 | inumb = key_inum(c, &sb->key); | |
183 | ||
184 | if (inuma == inumb) { | |
185 | uint32_t hasha = key_hash(c, &sa->key); | |
186 | uint32_t hashb = key_hash(c, &sb->key); | |
187 | ||
188 | if (hasha <= hashb) | |
189 | return -1; | |
190 | } else if (inuma <= inumb) | |
191 | return -1; | |
192 | ||
193 | return 1; | |
194 | } | |
195 | ||
196 | /** | |
197 | * sort_nodes - sort nodes for GC. | |
1e51764a | 198 | * @c: UBIFS file-system description object |
f10770f5 AB |
199 | * @sleb: describes nodes to sort and contains the result on exit |
200 | * @nondata: contains non-data nodes on exit | |
201 | * @min: minimum node size is returned here | |
1e51764a | 202 | * |
f10770f5 AB |
203 | * This function sorts the list of inodes to garbage collect. First of all, it |
204 | * kills obsolete nodes and separates data and non-data nodes to the | |
205 | * @sleb->nodes and @nondata lists correspondingly. | |
1e51764a | 206 | * |
f10770f5 AB |
207 | * Data nodes are then sorted in block number order - this is important for |
208 | * bulk-read; data nodes with lower inode number go before data nodes with | |
209 | * higher inode number, and data nodes with lower block number go before data | |
210 | * nodes with higher block number; | |
1e51764a | 211 | * |
f10770f5 AB |
212 | * Non-data nodes are sorted as follows. |
213 | * o First go inode nodes - they are sorted in descending length order. | |
214 | * o Then go directory entry nodes - they are sorted in hash order, which | |
215 | * should supposedly optimize 'readdir()'. Direntry nodes with lower parent | |
216 | * inode number go before direntry nodes with higher parent inode number, | |
217 | * and direntry nodes with lower name hash values go before direntry nodes | |
218 | * with higher name hash values. | |
219 | * | |
220 | * This function returns zero in case of success and a negative error code in | |
221 | * case of failure. | |
1e51764a | 222 | */ |
f10770f5 AB |
223 | static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb, |
224 | struct list_head *nondata, int *min) | |
1e51764a AB |
225 | { |
226 | struct ubifs_scan_node *snod, *tmp; | |
1e51764a | 227 | |
f10770f5 | 228 | *min = INT_MAX; |
1e51764a | 229 | |
f10770f5 AB |
230 | /* Separate data nodes and non-data nodes */ |
231 | list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) { | |
232 | int err; | |
1e51764a AB |
233 | |
234 | ubifs_assert(snod->type != UBIFS_IDX_NODE); | |
235 | ubifs_assert(snod->type != UBIFS_REF_NODE); | |
236 | ubifs_assert(snod->type != UBIFS_CS_NODE); | |
237 | ||
238 | err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum, | |
239 | snod->offs, 0); | |
240 | if (err < 0) | |
f10770f5 | 241 | return err; |
1e51764a | 242 | |
1e51764a AB |
243 | if (!err) { |
244 | /* The node is obsolete, remove it from the list */ | |
f10770f5 | 245 | list_del(&snod->list); |
1e51764a AB |
246 | kfree(snod); |
247 | continue; | |
248 | } | |
249 | ||
f10770f5 AB |
250 | if (snod->len < *min) |
251 | *min = snod->len; | |
252 | ||
253 | if (key_type(c, &snod->key) != UBIFS_DATA_KEY) | |
254 | list_move_tail(&snod->list, nondata); | |
1e51764a AB |
255 | } |
256 | ||
f10770f5 AB |
257 | /* Sort data and non-data nodes */ |
258 | list_sort(c, &sleb->nodes, &data_nodes_cmp); | |
259 | list_sort(c, nondata, &nondata_nodes_cmp); | |
260 | return 0; | |
261 | } | |
262 | ||
263 | /** | |
264 | * move_node - move a node. | |
265 | * @c: UBIFS file-system description object | |
266 | * @sleb: describes the LEB to move nodes from | |
267 | * @snod: the mode to move | |
268 | * @wbuf: write-buffer to move node to | |
269 | * | |
270 | * This function moves node @snod to @wbuf, changes TNC correspondingly, and | |
271 | * destroys @snod. Returns zero in case of success and a negative error code in | |
272 | * case of failure. | |
273 | */ | |
274 | static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb, | |
275 | struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf) | |
276 | { | |
277 | int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used; | |
278 | ||
279 | cond_resched(); | |
280 | err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len); | |
281 | if (err) | |
282 | return err; | |
283 | ||
284 | err = ubifs_tnc_replace(c, &snod->key, sleb->lnum, | |
285 | snod->offs, new_lnum, new_offs, | |
286 | snod->len); | |
287 | list_del(&snod->list); | |
288 | kfree(snod); | |
289 | return err; | |
290 | } | |
291 | ||
292 | /** | |
293 | * move_nodes - move nodes. | |
294 | * @c: UBIFS file-system description object | |
295 | * @sleb: describes the LEB to move nodes from | |
296 | * | |
297 | * This function moves valid nodes from data LEB described by @sleb to the GC | |
298 | * journal head. This function returns zero in case of success, %-EAGAIN if | |
299 | * commit is required, and other negative error codes in case of other | |
300 | * failures. | |
301 | */ | |
302 | static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb) | |
303 | { | |
304 | int err, min; | |
305 | LIST_HEAD(nondata); | |
306 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
1e51764a AB |
307 | |
308 | if (wbuf->lnum == -1) { | |
309 | /* | |
310 | * The GC journal head is not set, because it is the first GC | |
311 | * invocation since mount. | |
312 | */ | |
313 | err = switch_gc_head(c); | |
314 | if (err) | |
f10770f5 | 315 | return err; |
1e51764a AB |
316 | } |
317 | ||
f10770f5 AB |
318 | err = sort_nodes(c, sleb, &nondata, &min); |
319 | if (err) | |
320 | goto out; | |
321 | ||
1e51764a AB |
322 | /* Write nodes to their new location. Use the first-fit strategy */ |
323 | while (1) { | |
f10770f5 AB |
324 | int avail; |
325 | struct ubifs_scan_node *snod, *tmp; | |
326 | ||
327 | /* Move data nodes */ | |
328 | list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) { | |
329 | avail = c->leb_size - wbuf->offs - wbuf->used; | |
330 | if (snod->len > avail) | |
331 | /* | |
332 | * Do not skip data nodes in order to optimize | |
333 | * bulk-read. | |
334 | */ | |
335 | break; | |
336 | ||
337 | err = move_node(c, sleb, snod, wbuf); | |
338 | if (err) | |
339 | goto out; | |
340 | } | |
1e51764a | 341 | |
f10770f5 AB |
342 | /* Move non-data nodes */ |
343 | list_for_each_entry_safe(snod, tmp, &nondata, list) { | |
344 | avail = c->leb_size - wbuf->offs - wbuf->used; | |
1e51764a AB |
345 | if (avail < min) |
346 | break; | |
347 | ||
f10770f5 AB |
348 | if (snod->len > avail) { |
349 | /* | |
350 | * Keep going only if this is an inode with | |
351 | * some data. Otherwise stop and switch the GC | |
352 | * head. IOW, we assume that data-less inode | |
353 | * nodes and direntry nodes are roughly of the | |
354 | * same size. | |
355 | */ | |
356 | if (key_type(c, &snod->key) == UBIFS_DENT_KEY || | |
357 | snod->len == UBIFS_INO_NODE_SZ) | |
358 | break; | |
1e51764a | 359 | continue; |
f10770f5 | 360 | } |
1e51764a | 361 | |
f10770f5 | 362 | err = move_node(c, sleb, snod, wbuf); |
1e51764a AB |
363 | if (err) |
364 | goto out; | |
1e51764a AB |
365 | } |
366 | ||
f10770f5 | 367 | if (list_empty(&sleb->nodes) && list_empty(&nondata)) |
1e51764a AB |
368 | break; |
369 | ||
370 | /* | |
371 | * Waste the rest of the space in the LEB and switch to the | |
372 | * next LEB. | |
373 | */ | |
374 | err = switch_gc_head(c); | |
375 | if (err) | |
376 | goto out; | |
377 | } | |
378 | ||
379 | return 0; | |
380 | ||
381 | out: | |
f10770f5 | 382 | list_splice_tail(&nondata, &sleb->nodes); |
1e51764a AB |
383 | return err; |
384 | } | |
385 | ||
386 | /** | |
387 | * gc_sync_wbufs - sync write-buffers for GC. | |
388 | * @c: UBIFS file-system description object | |
389 | * | |
390 | * We must guarantee that obsoleting nodes are on flash. Unfortunately they may | |
391 | * be in a write-buffer instead. That is, a node could be written to a | |
392 | * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is | |
393 | * erased before the write-buffer is sync'd and then there is an unclean | |
394 | * unmount, then an existing node is lost. To avoid this, we sync all | |
395 | * write-buffers. | |
396 | * | |
397 | * This function returns %0 on success or a negative error code on failure. | |
398 | */ | |
399 | static int gc_sync_wbufs(struct ubifs_info *c) | |
400 | { | |
401 | int err, i; | |
402 | ||
403 | for (i = 0; i < c->jhead_cnt; i++) { | |
404 | if (i == GCHD) | |
405 | continue; | |
406 | err = ubifs_wbuf_sync(&c->jheads[i].wbuf); | |
407 | if (err) | |
408 | return err; | |
409 | } | |
410 | return 0; | |
411 | } | |
412 | ||
413 | /** | |
414 | * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock. | |
415 | * @c: UBIFS file-system description object | |
416 | * @lp: describes the LEB to garbage collect | |
417 | * | |
418 | * This function garbage-collects an LEB and returns one of the @LEB_FREED, | |
419 | * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is | |
420 | * required, and other negative error codes in case of failures. | |
421 | */ | |
422 | int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp) | |
423 | { | |
424 | struct ubifs_scan_leb *sleb; | |
425 | struct ubifs_scan_node *snod; | |
426 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
427 | int err = 0, lnum = lp->lnum; | |
428 | ||
429 | ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 || | |
430 | c->need_recovery); | |
431 | ubifs_assert(c->gc_lnum != lnum); | |
432 | ubifs_assert(wbuf->lnum != lnum); | |
433 | ||
434 | /* | |
435 | * We scan the entire LEB even though we only really need to scan up to | |
436 | * (c->leb_size - lp->free). | |
437 | */ | |
348709ba | 438 | sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0); |
1e51764a AB |
439 | if (IS_ERR(sleb)) |
440 | return PTR_ERR(sleb); | |
441 | ||
442 | ubifs_assert(!list_empty(&sleb->nodes)); | |
443 | snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list); | |
444 | ||
445 | if (snod->type == UBIFS_IDX_NODE) { | |
446 | struct ubifs_gced_idx_leb *idx_gc; | |
447 | ||
448 | dbg_gc("indexing LEB %d (free %d, dirty %d)", | |
449 | lnum, lp->free, lp->dirty); | |
450 | list_for_each_entry(snod, &sleb->nodes, list) { | |
451 | struct ubifs_idx_node *idx = snod->node; | |
452 | int level = le16_to_cpu(idx->level); | |
453 | ||
454 | ubifs_assert(snod->type == UBIFS_IDX_NODE); | |
455 | key_read(c, ubifs_idx_key(c, idx), &snod->key); | |
456 | err = ubifs_dirty_idx_node(c, &snod->key, level, lnum, | |
457 | snod->offs); | |
458 | if (err) | |
459 | goto out; | |
460 | } | |
461 | ||
462 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); | |
463 | if (!idx_gc) { | |
464 | err = -ENOMEM; | |
465 | goto out; | |
466 | } | |
467 | ||
468 | idx_gc->lnum = lnum; | |
469 | idx_gc->unmap = 0; | |
470 | list_add(&idx_gc->list, &c->idx_gc); | |
471 | ||
472 | /* | |
473 | * Don't release the LEB until after the next commit, because | |
227c75c9 | 474 | * it may contain data which is needed for recovery. So |
1e51764a AB |
475 | * although we freed this LEB, it will become usable only after |
476 | * the commit. | |
477 | */ | |
478 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, | |
479 | LPROPS_INDEX, 1); | |
480 | if (err) | |
481 | goto out; | |
482 | err = LEB_FREED_IDX; | |
483 | } else { | |
484 | dbg_gc("data LEB %d (free %d, dirty %d)", | |
485 | lnum, lp->free, lp->dirty); | |
486 | ||
487 | err = move_nodes(c, sleb); | |
488 | if (err) | |
6dcfac4f | 489 | goto out_inc_seq; |
1e51764a AB |
490 | |
491 | err = gc_sync_wbufs(c); | |
492 | if (err) | |
6dcfac4f | 493 | goto out_inc_seq; |
1e51764a AB |
494 | |
495 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0); | |
496 | if (err) | |
6dcfac4f | 497 | goto out_inc_seq; |
1e51764a | 498 | |
601c0bc4 AH |
499 | /* Allow for races with TNC */ |
500 | c->gced_lnum = lnum; | |
501 | smp_wmb(); | |
502 | c->gc_seq += 1; | |
503 | smp_wmb(); | |
504 | ||
1e51764a AB |
505 | if (c->gc_lnum == -1) { |
506 | c->gc_lnum = lnum; | |
507 | err = LEB_RETAINED; | |
508 | } else { | |
509 | err = ubifs_wbuf_sync_nolock(wbuf); | |
510 | if (err) | |
511 | goto out; | |
512 | ||
513 | err = ubifs_leb_unmap(c, lnum); | |
514 | if (err) | |
515 | goto out; | |
516 | ||
517 | err = LEB_FREED; | |
518 | } | |
519 | } | |
520 | ||
521 | out: | |
522 | ubifs_scan_destroy(sleb); | |
523 | return err; | |
6dcfac4f AH |
524 | |
525 | out_inc_seq: | |
526 | /* We may have moved at least some nodes so allow for races with TNC */ | |
527 | c->gced_lnum = lnum; | |
528 | smp_wmb(); | |
529 | c->gc_seq += 1; | |
530 | smp_wmb(); | |
531 | goto out; | |
1e51764a AB |
532 | } |
533 | ||
534 | /** | |
535 | * ubifs_garbage_collect - UBIFS garbage collector. | |
536 | * @c: UBIFS file-system description object | |
537 | * @anyway: do GC even if there are free LEBs | |
538 | * | |
539 | * This function does out-of-place garbage collection. The return codes are: | |
540 | * o positive LEB number if the LEB has been freed and may be used; | |
541 | * o %-EAGAIN if the caller has to run commit; | |
542 | * o %-ENOSPC if GC failed to make any progress; | |
543 | * o other negative error codes in case of other errors. | |
544 | * | |
545 | * Garbage collector writes data to the journal when GC'ing data LEBs, and just | |
546 | * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point | |
547 | * commit may be required. But commit cannot be run from inside GC, because the | |
548 | * caller might be holding the commit lock, so %-EAGAIN is returned instead; | |
549 | * And this error code means that the caller has to run commit, and re-run GC | |
550 | * if there is still no free space. | |
551 | * | |
552 | * There are many reasons why this function may return %-EAGAIN: | |
553 | * o the log is full and there is no space to write an LEB reference for | |
554 | * @c->gc_lnum; | |
555 | * o the journal is too large and exceeds size limitations; | |
556 | * o GC moved indexing LEBs, but they can be used only after the commit; | |
557 | * o the shrinker fails to find clean znodes to free and requests the commit; | |
558 | * o etc. | |
559 | * | |
560 | * Note, if the file-system is close to be full, this function may return | |
561 | * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of | |
562 | * the function. E.g., this happens if the limits on the journal size are too | |
563 | * tough and GC writes too much to the journal before an LEB is freed. This | |
564 | * might also mean that the journal is too large, and the TNC becomes to big, | |
565 | * so that the shrinker is constantly called, finds not clean znodes to free, | |
566 | * and requests commit. Well, this may also happen if the journal is all right, | |
567 | * but another kernel process consumes too much memory. Anyway, infinite | |
568 | * %-EAGAIN may happen, but in some extreme/misconfiguration cases. | |
569 | */ | |
570 | int ubifs_garbage_collect(struct ubifs_info *c, int anyway) | |
571 | { | |
572 | int i, err, ret, min_space = c->dead_wm; | |
573 | struct ubifs_lprops lp; | |
574 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
575 | ||
576 | ubifs_assert_cmt_locked(c); | |
577 | ||
578 | if (ubifs_gc_should_commit(c)) | |
579 | return -EAGAIN; | |
580 | ||
581 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); | |
582 | ||
583 | if (c->ro_media) { | |
584 | ret = -EROFS; | |
585 | goto out_unlock; | |
586 | } | |
587 | ||
588 | /* We expect the write-buffer to be empty on entry */ | |
589 | ubifs_assert(!wbuf->used); | |
590 | ||
591 | for (i = 0; ; i++) { | |
592 | int space_before = c->leb_size - wbuf->offs - wbuf->used; | |
593 | int space_after; | |
594 | ||
595 | cond_resched(); | |
596 | ||
597 | /* Give the commit an opportunity to run */ | |
598 | if (ubifs_gc_should_commit(c)) { | |
599 | ret = -EAGAIN; | |
600 | break; | |
601 | } | |
602 | ||
603 | if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) { | |
604 | /* | |
605 | * We've done enough iterations. Indexing LEBs were | |
606 | * moved and will be available after the commit. | |
607 | */ | |
608 | dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN"); | |
609 | ubifs_commit_required(c); | |
610 | ret = -EAGAIN; | |
611 | break; | |
612 | } | |
613 | ||
614 | if (i > HARD_LEBS_LIMIT) { | |
615 | /* | |
616 | * We've moved too many LEBs and have not made | |
617 | * progress, give up. | |
618 | */ | |
619 | dbg_gc("hard limit, -ENOSPC"); | |
620 | ret = -ENOSPC; | |
621 | break; | |
622 | } | |
623 | ||
624 | /* | |
625 | * Empty and freeable LEBs can turn up while we waited for | |
626 | * the wbuf lock, or while we have been running GC. In that | |
627 | * case, we should just return one of those instead of | |
628 | * continuing to GC dirty LEBs. Hence we request | |
629 | * 'ubifs_find_dirty_leb()' to return an empty LEB if it can. | |
630 | */ | |
631 | ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1); | |
632 | if (ret) { | |
633 | if (ret == -ENOSPC) | |
634 | dbg_gc("no more dirty LEBs"); | |
635 | break; | |
636 | } | |
637 | ||
638 | dbg_gc("found LEB %d: free %d, dirty %d, sum %d " | |
639 | "(min. space %d)", lp.lnum, lp.free, lp.dirty, | |
640 | lp.free + lp.dirty, min_space); | |
641 | ||
642 | if (lp.free + lp.dirty == c->leb_size) { | |
643 | /* An empty LEB was returned */ | |
644 | dbg_gc("LEB %d is free, return it", lp.lnum); | |
645 | /* | |
646 | * ubifs_find_dirty_leb() doesn't return freeable index | |
647 | * LEBs. | |
648 | */ | |
649 | ubifs_assert(!(lp.flags & LPROPS_INDEX)); | |
650 | if (lp.free != c->leb_size) { | |
651 | /* | |
652 | * Write buffers must be sync'd before | |
653 | * unmapping freeable LEBs, because one of them | |
654 | * may contain data which obsoletes something | |
655 | * in 'lp.pnum'. | |
656 | */ | |
657 | ret = gc_sync_wbufs(c); | |
658 | if (ret) | |
659 | goto out; | |
660 | ret = ubifs_change_one_lp(c, lp.lnum, | |
661 | c->leb_size, 0, 0, 0, | |
662 | 0); | |
663 | if (ret) | |
664 | goto out; | |
665 | } | |
666 | ret = ubifs_leb_unmap(c, lp.lnum); | |
667 | if (ret) | |
668 | goto out; | |
669 | ret = lp.lnum; | |
670 | break; | |
671 | } | |
672 | ||
673 | space_before = c->leb_size - wbuf->offs - wbuf->used; | |
674 | if (wbuf->lnum == -1) | |
675 | space_before = 0; | |
676 | ||
677 | ret = ubifs_garbage_collect_leb(c, &lp); | |
678 | if (ret < 0) { | |
679 | if (ret == -EAGAIN || ret == -ENOSPC) { | |
680 | /* | |
681 | * These codes are not errors, so we have to | |
682 | * return the LEB to lprops. But if the | |
683 | * 'ubifs_return_leb()' function fails, its | |
684 | * failure code is propagated to the caller | |
685 | * instead of the original '-EAGAIN' or | |
686 | * '-ENOSPC'. | |
687 | */ | |
688 | err = ubifs_return_leb(c, lp.lnum); | |
689 | if (err) | |
690 | ret = err; | |
691 | break; | |
692 | } | |
693 | goto out; | |
694 | } | |
695 | ||
696 | if (ret == LEB_FREED) { | |
697 | /* An LEB has been freed and is ready for use */ | |
698 | dbg_gc("LEB %d freed, return", lp.lnum); | |
699 | ret = lp.lnum; | |
700 | break; | |
701 | } | |
702 | ||
703 | if (ret == LEB_FREED_IDX) { | |
704 | /* | |
705 | * This was an indexing LEB and it cannot be | |
706 | * immediately used. And instead of requesting the | |
707 | * commit straight away, we try to garbage collect some | |
708 | * more. | |
709 | */ | |
710 | dbg_gc("indexing LEB %d freed, continue", lp.lnum); | |
711 | continue; | |
712 | } | |
713 | ||
714 | ubifs_assert(ret == LEB_RETAINED); | |
715 | space_after = c->leb_size - wbuf->offs - wbuf->used; | |
716 | dbg_gc("LEB %d retained, freed %d bytes", lp.lnum, | |
717 | space_after - space_before); | |
718 | ||
719 | if (space_after > space_before) { | |
720 | /* GC makes progress, keep working */ | |
721 | min_space >>= 1; | |
722 | if (min_space < c->dead_wm) | |
723 | min_space = c->dead_wm; | |
724 | continue; | |
725 | } | |
726 | ||
727 | dbg_gc("did not make progress"); | |
728 | ||
729 | /* | |
730 | * GC moved an LEB bud have not done any progress. This means | |
731 | * that the previous GC head LEB contained too few free space | |
732 | * and the LEB which was GC'ed contained only large nodes which | |
733 | * did not fit that space. | |
734 | * | |
735 | * We can do 2 things: | |
736 | * 1. pick another LEB in a hope it'll contain a small node | |
737 | * which will fit the space we have at the end of current GC | |
738 | * head LEB, but there is no guarantee, so we try this out | |
739 | * unless we have already been working for too long; | |
740 | * 2. request an LEB with more dirty space, which will force | |
741 | * 'ubifs_find_dirty_leb()' to start scanning the lprops | |
742 | * table, instead of just picking one from the heap | |
743 | * (previously it already picked the dirtiest LEB). | |
744 | */ | |
745 | if (i < SOFT_LEBS_LIMIT) { | |
746 | dbg_gc("try again"); | |
747 | continue; | |
748 | } | |
749 | ||
750 | min_space <<= 1; | |
751 | if (min_space > c->dark_wm) | |
752 | min_space = c->dark_wm; | |
753 | dbg_gc("set min. space to %d", min_space); | |
754 | } | |
755 | ||
756 | if (ret == -ENOSPC && !list_empty(&c->idx_gc)) { | |
757 | dbg_gc("no space, some index LEBs GC'ed, -EAGAIN"); | |
758 | ubifs_commit_required(c); | |
759 | ret = -EAGAIN; | |
760 | } | |
761 | ||
762 | err = ubifs_wbuf_sync_nolock(wbuf); | |
763 | if (!err) | |
764 | err = ubifs_leb_unmap(c, c->gc_lnum); | |
765 | if (err) { | |
766 | ret = err; | |
767 | goto out; | |
768 | } | |
769 | out_unlock: | |
770 | mutex_unlock(&wbuf->io_mutex); | |
771 | return ret; | |
772 | ||
773 | out: | |
774 | ubifs_assert(ret < 0); | |
775 | ubifs_assert(ret != -ENOSPC && ret != -EAGAIN); | |
776 | ubifs_ro_mode(c, ret); | |
777 | ubifs_wbuf_sync_nolock(wbuf); | |
778 | mutex_unlock(&wbuf->io_mutex); | |
779 | ubifs_return_leb(c, lp.lnum); | |
780 | return ret; | |
781 | } | |
782 | ||
783 | /** | |
784 | * ubifs_gc_start_commit - garbage collection at start of commit. | |
785 | * @c: UBIFS file-system description object | |
786 | * | |
787 | * If a LEB has only dirty and free space, then we may safely unmap it and make | |
788 | * it free. Note, we cannot do this with indexing LEBs because dirty space may | |
789 | * correspond index nodes that are required for recovery. In that case, the | |
790 | * LEB cannot be unmapped until after the next commit. | |
791 | * | |
792 | * This function returns %0 upon success and a negative error code upon failure. | |
793 | */ | |
794 | int ubifs_gc_start_commit(struct ubifs_info *c) | |
795 | { | |
796 | struct ubifs_gced_idx_leb *idx_gc; | |
797 | const struct ubifs_lprops *lp; | |
798 | int err = 0, flags; | |
799 | ||
800 | ubifs_get_lprops(c); | |
801 | ||
802 | /* | |
803 | * Unmap (non-index) freeable LEBs. Note that recovery requires that all | |
804 | * wbufs are sync'd before this, which is done in 'do_commit()'. | |
805 | */ | |
806 | while (1) { | |
807 | lp = ubifs_fast_find_freeable(c); | |
8d47aef4 | 808 | if (IS_ERR(lp)) { |
1e51764a AB |
809 | err = PTR_ERR(lp); |
810 | goto out; | |
811 | } | |
812 | if (!lp) | |
813 | break; | |
814 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | |
815 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | |
816 | err = ubifs_leb_unmap(c, lp->lnum); | |
817 | if (err) | |
818 | goto out; | |
819 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0); | |
8d47aef4 | 820 | if (IS_ERR(lp)) { |
1e51764a AB |
821 | err = PTR_ERR(lp); |
822 | goto out; | |
823 | } | |
824 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | |
825 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | |
826 | } | |
827 | ||
828 | /* Mark GC'd index LEBs OK to unmap after this commit finishes */ | |
829 | list_for_each_entry(idx_gc, &c->idx_gc, list) | |
830 | idx_gc->unmap = 1; | |
831 | ||
832 | /* Record index freeable LEBs for unmapping after commit */ | |
833 | while (1) { | |
834 | lp = ubifs_fast_find_frdi_idx(c); | |
8d47aef4 | 835 | if (IS_ERR(lp)) { |
1e51764a AB |
836 | err = PTR_ERR(lp); |
837 | goto out; | |
838 | } | |
839 | if (!lp) | |
840 | break; | |
841 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); | |
842 | if (!idx_gc) { | |
843 | err = -ENOMEM; | |
844 | goto out; | |
845 | } | |
846 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | |
847 | ubifs_assert(lp->flags & LPROPS_INDEX); | |
848 | /* Don't release the LEB until after the next commit */ | |
849 | flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX; | |
850 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1); | |
8d47aef4 | 851 | if (IS_ERR(lp)) { |
1e51764a AB |
852 | err = PTR_ERR(lp); |
853 | kfree(idx_gc); | |
854 | goto out; | |
855 | } | |
856 | ubifs_assert(lp->flags & LPROPS_TAKEN); | |
857 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | |
858 | idx_gc->lnum = lp->lnum; | |
859 | idx_gc->unmap = 1; | |
860 | list_add(&idx_gc->list, &c->idx_gc); | |
861 | } | |
862 | out: | |
863 | ubifs_release_lprops(c); | |
864 | return err; | |
865 | } | |
866 | ||
867 | /** | |
868 | * ubifs_gc_end_commit - garbage collection at end of commit. | |
869 | * @c: UBIFS file-system description object | |
870 | * | |
871 | * This function completes out-of-place garbage collection of index LEBs. | |
872 | */ | |
873 | int ubifs_gc_end_commit(struct ubifs_info *c) | |
874 | { | |
875 | struct ubifs_gced_idx_leb *idx_gc, *tmp; | |
876 | struct ubifs_wbuf *wbuf; | |
877 | int err = 0; | |
878 | ||
879 | wbuf = &c->jheads[GCHD].wbuf; | |
880 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); | |
881 | list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list) | |
882 | if (idx_gc->unmap) { | |
883 | dbg_gc("LEB %d", idx_gc->lnum); | |
884 | err = ubifs_leb_unmap(c, idx_gc->lnum); | |
885 | if (err) | |
886 | goto out; | |
887 | err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC, | |
888 | LPROPS_NC, 0, LPROPS_TAKEN, -1); | |
889 | if (err) | |
890 | goto out; | |
891 | list_del(&idx_gc->list); | |
892 | kfree(idx_gc); | |
893 | } | |
894 | out: | |
895 | mutex_unlock(&wbuf->io_mutex); | |
896 | return err; | |
897 | } | |
898 | ||
899 | /** | |
900 | * ubifs_destroy_idx_gc - destroy idx_gc list. | |
901 | * @c: UBIFS file-system description object | |
902 | * | |
b466f17d AH |
903 | * This function destroys the @c->idx_gc list. It is called when unmounting |
904 | * so locks are not needed. Returns zero in case of success and a negative | |
905 | * error code in case of failure. | |
1e51764a | 906 | */ |
b466f17d | 907 | void ubifs_destroy_idx_gc(struct ubifs_info *c) |
1e51764a AB |
908 | { |
909 | while (!list_empty(&c->idx_gc)) { | |
910 | struct ubifs_gced_idx_leb *idx_gc; | |
911 | ||
912 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, | |
913 | list); | |
b466f17d | 914 | c->idx_gc_cnt -= 1; |
1e51764a AB |
915 | list_del(&idx_gc->list); |
916 | kfree(idx_gc); | |
917 | } | |
1e51764a AB |
918 | } |
919 | ||
920 | /** | |
921 | * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list. | |
922 | * @c: UBIFS file-system description object | |
923 | * | |
924 | * Called during start commit so locks are not needed. | |
925 | */ | |
926 | int ubifs_get_idx_gc_leb(struct ubifs_info *c) | |
927 | { | |
928 | struct ubifs_gced_idx_leb *idx_gc; | |
929 | int lnum; | |
930 | ||
931 | if (list_empty(&c->idx_gc)) | |
932 | return -ENOSPC; | |
933 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list); | |
934 | lnum = idx_gc->lnum; | |
935 | /* c->idx_gc_cnt is updated by the caller when lprops are updated */ | |
936 | list_del(&idx_gc->list); | |
937 | kfree(idx_gc); | |
938 | return lnum; | |
939 | } |