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. | |
34 | */ | |
35 | ||
36 | #include <linux/pagemap.h> | |
37 | #include "ubifs.h" | |
38 | ||
39 | /* | |
40 | * GC tries to optimize the way it fit nodes to available space, and it sorts | |
41 | * nodes a little. The below constants are watermarks which define "large", | |
42 | * "medium", and "small" nodes. | |
43 | */ | |
44 | #define MEDIUM_NODE_WM (UBIFS_BLOCK_SIZE / 4) | |
45 | #define SMALL_NODE_WM UBIFS_MAX_DENT_NODE_SZ | |
46 | ||
47 | /* | |
025dfdaf | 48 | * GC may need to move more than one LEB to make progress. The below constants |
1e51764a AB |
49 | * define "soft" and "hard" limits on the number of LEBs the garbage collector |
50 | * may move. | |
51 | */ | |
52 | #define SOFT_LEBS_LIMIT 4 | |
53 | #define HARD_LEBS_LIMIT 32 | |
54 | ||
55 | /** | |
56 | * switch_gc_head - switch the garbage collection journal head. | |
57 | * @c: UBIFS file-system description object | |
58 | * @buf: buffer to write | |
59 | * @len: length of the buffer to write | |
60 | * @lnum: LEB number written is returned here | |
61 | * @offs: offset written is returned here | |
62 | * | |
63 | * This function switch the GC head to the next LEB which is reserved in | |
64 | * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required, | |
65 | * and other negative error code in case of failures. | |
66 | */ | |
67 | static int switch_gc_head(struct ubifs_info *c) | |
68 | { | |
69 | int err, gc_lnum = c->gc_lnum; | |
70 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
71 | ||
72 | ubifs_assert(gc_lnum != -1); | |
73 | dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)", | |
74 | wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum, | |
75 | c->leb_size - wbuf->offs - wbuf->used); | |
76 | ||
77 | err = ubifs_wbuf_sync_nolock(wbuf); | |
78 | if (err) | |
79 | return err; | |
80 | ||
81 | /* | |
82 | * The GC write-buffer was synchronized, we may safely unmap | |
83 | * 'c->gc_lnum'. | |
84 | */ | |
85 | err = ubifs_leb_unmap(c, gc_lnum); | |
86 | if (err) | |
87 | return err; | |
88 | ||
89 | err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0); | |
90 | if (err) | |
91 | return err; | |
92 | ||
93 | c->gc_lnum = -1; | |
94 | err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0, UBI_LONGTERM); | |
95 | return err; | |
96 | } | |
97 | ||
46773be4 AH |
98 | /** |
99 | * joinup - bring data nodes for an inode together. | |
100 | * @c: UBIFS file-system description object | |
101 | * @sleb: describes scanned LEB | |
102 | * @inum: inode number | |
103 | * @blk: block number | |
104 | * @data: list to which to add data nodes | |
105 | * | |
106 | * This function looks at the first few nodes in the scanned LEB @sleb and adds | |
107 | * them to @data if they are data nodes from @inum and have a larger block | |
108 | * number than @blk. This function returns %0 on success and a negative error | |
109 | * code on failure. | |
110 | */ | |
111 | static int joinup(struct ubifs_info *c, struct ubifs_scan_leb *sleb, ino_t inum, | |
112 | unsigned int blk, struct list_head *data) | |
113 | { | |
114 | int err, cnt = 6, lnum = sleb->lnum, offs; | |
115 | struct ubifs_scan_node *snod, *tmp; | |
116 | union ubifs_key *key; | |
117 | ||
118 | list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) { | |
119 | key = &snod->key; | |
120 | if (key_inum(c, key) == inum && | |
121 | key_type(c, key) == UBIFS_DATA_KEY && | |
122 | key_block(c, key) > blk) { | |
123 | offs = snod->offs; | |
124 | err = ubifs_tnc_has_node(c, key, 0, lnum, offs, 0); | |
125 | if (err < 0) | |
126 | return err; | |
127 | list_del(&snod->list); | |
128 | if (err) { | |
129 | list_add_tail(&snod->list, data); | |
130 | blk = key_block(c, key); | |
131 | } else | |
132 | kfree(snod); | |
133 | cnt = 6; | |
134 | } else if (--cnt == 0) | |
135 | break; | |
136 | } | |
137 | return 0; | |
138 | } | |
139 | ||
1e51764a AB |
140 | /** |
141 | * move_nodes - move nodes. | |
142 | * @c: UBIFS file-system description object | |
143 | * @sleb: describes nodes to move | |
144 | * | |
145 | * This function moves valid nodes from data LEB described by @sleb to the GC | |
146 | * journal head. The obsolete nodes are dropped. | |
147 | * | |
148 | * When moving nodes we have to deal with classical bin-packing problem: the | |
149 | * space in the current GC journal head LEB and in @c->gc_lnum are the "bins", | |
150 | * where the nodes in the @sleb->nodes list are the elements which should be | |
151 | * fit optimally to the bins. This function uses the "first fit decreasing" | |
152 | * strategy, although it does not really sort the nodes but just split them on | |
153 | * 3 classes - large, medium, and small, so they are roughly sorted. | |
154 | * | |
155 | * This function returns zero in case of success, %-EAGAIN if commit is | |
156 | * required, and other negative error codes in case of other failures. | |
157 | */ | |
158 | static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb) | |
159 | { | |
160 | struct ubifs_scan_node *snod, *tmp; | |
46773be4 | 161 | struct list_head data, large, medium, small; |
1e51764a AB |
162 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; |
163 | int avail, err, min = INT_MAX; | |
46773be4 AH |
164 | unsigned int blk = 0; |
165 | ino_t inum = 0; | |
1e51764a | 166 | |
46773be4 | 167 | INIT_LIST_HEAD(&data); |
1e51764a AB |
168 | INIT_LIST_HEAD(&large); |
169 | INIT_LIST_HEAD(&medium); | |
170 | INIT_LIST_HEAD(&small); | |
171 | ||
46773be4 AH |
172 | while (!list_empty(&sleb->nodes)) { |
173 | struct list_head *lst = sleb->nodes.next; | |
174 | ||
175 | snod = list_entry(lst, struct ubifs_scan_node, list); | |
1e51764a AB |
176 | |
177 | ubifs_assert(snod->type != UBIFS_IDX_NODE); | |
178 | ubifs_assert(snod->type != UBIFS_REF_NODE); | |
179 | ubifs_assert(snod->type != UBIFS_CS_NODE); | |
180 | ||
181 | err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum, | |
182 | snod->offs, 0); | |
183 | if (err < 0) | |
184 | goto out; | |
185 | ||
1e51764a AB |
186 | list_del(lst); |
187 | if (!err) { | |
188 | /* The node is obsolete, remove it from the list */ | |
189 | kfree(snod); | |
190 | continue; | |
191 | } | |
192 | ||
193 | /* | |
46773be4 AH |
194 | * Sort the list of nodes so that data nodes go first, large |
195 | * nodes go second, and small nodes go last. | |
1e51764a | 196 | */ |
46773be4 AH |
197 | if (key_type(c, &snod->key) == UBIFS_DATA_KEY) { |
198 | if (inum != key_inum(c, &snod->key)) { | |
199 | if (inum) { | |
200 | /* | |
201 | * Try to move data nodes from the same | |
202 | * inode together. | |
203 | */ | |
204 | err = joinup(c, sleb, inum, blk, &data); | |
205 | if (err) | |
206 | goto out; | |
207 | } | |
208 | inum = key_inum(c, &snod->key); | |
209 | blk = key_block(c, &snod->key); | |
210 | } | |
211 | list_add_tail(lst, &data); | |
212 | } else if (snod->len > MEDIUM_NODE_WM) | |
213 | list_add_tail(lst, &large); | |
1e51764a | 214 | else if (snod->len > SMALL_NODE_WM) |
46773be4 | 215 | list_add_tail(lst, &medium); |
1e51764a | 216 | else |
46773be4 | 217 | list_add_tail(lst, &small); |
1e51764a AB |
218 | |
219 | /* And find the smallest node */ | |
220 | if (snod->len < min) | |
221 | min = snod->len; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Join the tree lists so that we'd have one roughly sorted list | |
226 | * ('large' will be the head of the joined list). | |
227 | */ | |
46773be4 | 228 | list_splice(&data, &large); |
1e51764a AB |
229 | list_splice(&medium, large.prev); |
230 | list_splice(&small, large.prev); | |
231 | ||
232 | if (wbuf->lnum == -1) { | |
233 | /* | |
234 | * The GC journal head is not set, because it is the first GC | |
235 | * invocation since mount. | |
236 | */ | |
237 | err = switch_gc_head(c); | |
238 | if (err) | |
239 | goto out; | |
240 | } | |
241 | ||
242 | /* Write nodes to their new location. Use the first-fit strategy */ | |
243 | while (1) { | |
244 | avail = c->leb_size - wbuf->offs - wbuf->used; | |
245 | list_for_each_entry_safe(snod, tmp, &large, list) { | |
246 | int new_lnum, new_offs; | |
247 | ||
248 | if (avail < min) | |
249 | break; | |
250 | ||
251 | if (snod->len > avail) | |
252 | /* This node does not fit */ | |
253 | continue; | |
254 | ||
255 | cond_resched(); | |
256 | ||
257 | new_lnum = wbuf->lnum; | |
258 | new_offs = wbuf->offs + wbuf->used; | |
259 | err = ubifs_wbuf_write_nolock(wbuf, snod->node, | |
260 | snod->len); | |
261 | if (err) | |
262 | goto out; | |
263 | err = ubifs_tnc_replace(c, &snod->key, sleb->lnum, | |
264 | snod->offs, new_lnum, new_offs, | |
265 | snod->len); | |
266 | if (err) | |
267 | goto out; | |
268 | ||
269 | avail = c->leb_size - wbuf->offs - wbuf->used; | |
270 | list_del(&snod->list); | |
271 | kfree(snod); | |
272 | } | |
273 | ||
274 | if (list_empty(&large)) | |
275 | break; | |
276 | ||
277 | /* | |
278 | * Waste the rest of the space in the LEB and switch to the | |
279 | * next LEB. | |
280 | */ | |
281 | err = switch_gc_head(c); | |
282 | if (err) | |
283 | goto out; | |
284 | } | |
285 | ||
286 | return 0; | |
287 | ||
288 | out: | |
289 | list_for_each_entry_safe(snod, tmp, &large, list) { | |
290 | list_del(&snod->list); | |
291 | kfree(snod); | |
292 | } | |
293 | return err; | |
294 | } | |
295 | ||
296 | /** | |
297 | * gc_sync_wbufs - sync write-buffers for GC. | |
298 | * @c: UBIFS file-system description object | |
299 | * | |
300 | * We must guarantee that obsoleting nodes are on flash. Unfortunately they may | |
301 | * be in a write-buffer instead. That is, a node could be written to a | |
302 | * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is | |
303 | * erased before the write-buffer is sync'd and then there is an unclean | |
304 | * unmount, then an existing node is lost. To avoid this, we sync all | |
305 | * write-buffers. | |
306 | * | |
307 | * This function returns %0 on success or a negative error code on failure. | |
308 | */ | |
309 | static int gc_sync_wbufs(struct ubifs_info *c) | |
310 | { | |
311 | int err, i; | |
312 | ||
313 | for (i = 0; i < c->jhead_cnt; i++) { | |
314 | if (i == GCHD) | |
315 | continue; | |
316 | err = ubifs_wbuf_sync(&c->jheads[i].wbuf); | |
317 | if (err) | |
318 | return err; | |
319 | } | |
320 | return 0; | |
321 | } | |
322 | ||
323 | /** | |
324 | * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock. | |
325 | * @c: UBIFS file-system description object | |
326 | * @lp: describes the LEB to garbage collect | |
327 | * | |
328 | * This function garbage-collects an LEB and returns one of the @LEB_FREED, | |
329 | * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is | |
330 | * required, and other negative error codes in case of failures. | |
331 | */ | |
332 | int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp) | |
333 | { | |
334 | struct ubifs_scan_leb *sleb; | |
335 | struct ubifs_scan_node *snod; | |
336 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
337 | int err = 0, lnum = lp->lnum; | |
338 | ||
339 | ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 || | |
340 | c->need_recovery); | |
341 | ubifs_assert(c->gc_lnum != lnum); | |
342 | ubifs_assert(wbuf->lnum != lnum); | |
343 | ||
344 | /* | |
345 | * We scan the entire LEB even though we only really need to scan up to | |
346 | * (c->leb_size - lp->free). | |
347 | */ | |
348 | sleb = ubifs_scan(c, lnum, 0, c->sbuf); | |
349 | if (IS_ERR(sleb)) | |
350 | return PTR_ERR(sleb); | |
351 | ||
352 | ubifs_assert(!list_empty(&sleb->nodes)); | |
353 | snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list); | |
354 | ||
355 | if (snod->type == UBIFS_IDX_NODE) { | |
356 | struct ubifs_gced_idx_leb *idx_gc; | |
357 | ||
358 | dbg_gc("indexing LEB %d (free %d, dirty %d)", | |
359 | lnum, lp->free, lp->dirty); | |
360 | list_for_each_entry(snod, &sleb->nodes, list) { | |
361 | struct ubifs_idx_node *idx = snod->node; | |
362 | int level = le16_to_cpu(idx->level); | |
363 | ||
364 | ubifs_assert(snod->type == UBIFS_IDX_NODE); | |
365 | key_read(c, ubifs_idx_key(c, idx), &snod->key); | |
366 | err = ubifs_dirty_idx_node(c, &snod->key, level, lnum, | |
367 | snod->offs); | |
368 | if (err) | |
369 | goto out; | |
370 | } | |
371 | ||
372 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); | |
373 | if (!idx_gc) { | |
374 | err = -ENOMEM; | |
375 | goto out; | |
376 | } | |
377 | ||
378 | idx_gc->lnum = lnum; | |
379 | idx_gc->unmap = 0; | |
380 | list_add(&idx_gc->list, &c->idx_gc); | |
381 | ||
382 | /* | |
383 | * Don't release the LEB until after the next commit, because | |
384 | * it may contain date which is needed for recovery. So | |
385 | * although we freed this LEB, it will become usable only after | |
386 | * the commit. | |
387 | */ | |
388 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, | |
389 | LPROPS_INDEX, 1); | |
390 | if (err) | |
391 | goto out; | |
392 | err = LEB_FREED_IDX; | |
393 | } else { | |
394 | dbg_gc("data LEB %d (free %d, dirty %d)", | |
395 | lnum, lp->free, lp->dirty); | |
396 | ||
397 | err = move_nodes(c, sleb); | |
398 | if (err) | |
6dcfac4f | 399 | goto out_inc_seq; |
1e51764a AB |
400 | |
401 | err = gc_sync_wbufs(c); | |
402 | if (err) | |
6dcfac4f | 403 | goto out_inc_seq; |
1e51764a AB |
404 | |
405 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0); | |
406 | if (err) | |
6dcfac4f | 407 | goto out_inc_seq; |
1e51764a | 408 | |
601c0bc4 AH |
409 | /* Allow for races with TNC */ |
410 | c->gced_lnum = lnum; | |
411 | smp_wmb(); | |
412 | c->gc_seq += 1; | |
413 | smp_wmb(); | |
414 | ||
1e51764a AB |
415 | if (c->gc_lnum == -1) { |
416 | c->gc_lnum = lnum; | |
417 | err = LEB_RETAINED; | |
418 | } else { | |
419 | err = ubifs_wbuf_sync_nolock(wbuf); | |
420 | if (err) | |
421 | goto out; | |
422 | ||
423 | err = ubifs_leb_unmap(c, lnum); | |
424 | if (err) | |
425 | goto out; | |
426 | ||
427 | err = LEB_FREED; | |
428 | } | |
429 | } | |
430 | ||
431 | out: | |
432 | ubifs_scan_destroy(sleb); | |
433 | return err; | |
6dcfac4f AH |
434 | |
435 | out_inc_seq: | |
436 | /* We may have moved at least some nodes so allow for races with TNC */ | |
437 | c->gced_lnum = lnum; | |
438 | smp_wmb(); | |
439 | c->gc_seq += 1; | |
440 | smp_wmb(); | |
441 | goto out; | |
1e51764a AB |
442 | } |
443 | ||
444 | /** | |
445 | * ubifs_garbage_collect - UBIFS garbage collector. | |
446 | * @c: UBIFS file-system description object | |
447 | * @anyway: do GC even if there are free LEBs | |
448 | * | |
449 | * This function does out-of-place garbage collection. The return codes are: | |
450 | * o positive LEB number if the LEB has been freed and may be used; | |
451 | * o %-EAGAIN if the caller has to run commit; | |
452 | * o %-ENOSPC if GC failed to make any progress; | |
453 | * o other negative error codes in case of other errors. | |
454 | * | |
455 | * Garbage collector writes data to the journal when GC'ing data LEBs, and just | |
456 | * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point | |
457 | * commit may be required. But commit cannot be run from inside GC, because the | |
458 | * caller might be holding the commit lock, so %-EAGAIN is returned instead; | |
459 | * And this error code means that the caller has to run commit, and re-run GC | |
460 | * if there is still no free space. | |
461 | * | |
462 | * There are many reasons why this function may return %-EAGAIN: | |
463 | * o the log is full and there is no space to write an LEB reference for | |
464 | * @c->gc_lnum; | |
465 | * o the journal is too large and exceeds size limitations; | |
466 | * o GC moved indexing LEBs, but they can be used only after the commit; | |
467 | * o the shrinker fails to find clean znodes to free and requests the commit; | |
468 | * o etc. | |
469 | * | |
470 | * Note, if the file-system is close to be full, this function may return | |
471 | * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of | |
472 | * the function. E.g., this happens if the limits on the journal size are too | |
473 | * tough and GC writes too much to the journal before an LEB is freed. This | |
474 | * might also mean that the journal is too large, and the TNC becomes to big, | |
475 | * so that the shrinker is constantly called, finds not clean znodes to free, | |
476 | * and requests commit. Well, this may also happen if the journal is all right, | |
477 | * but another kernel process consumes too much memory. Anyway, infinite | |
478 | * %-EAGAIN may happen, but in some extreme/misconfiguration cases. | |
479 | */ | |
480 | int ubifs_garbage_collect(struct ubifs_info *c, int anyway) | |
481 | { | |
482 | int i, err, ret, min_space = c->dead_wm; | |
483 | struct ubifs_lprops lp; | |
484 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | |
485 | ||
486 | ubifs_assert_cmt_locked(c); | |
487 | ||
488 | if (ubifs_gc_should_commit(c)) | |
489 | return -EAGAIN; | |
490 | ||
491 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); | |
492 | ||
493 | if (c->ro_media) { | |
494 | ret = -EROFS; | |
495 | goto out_unlock; | |
496 | } | |
497 | ||
498 | /* We expect the write-buffer to be empty on entry */ | |
499 | ubifs_assert(!wbuf->used); | |
500 | ||
501 | for (i = 0; ; i++) { | |
502 | int space_before = c->leb_size - wbuf->offs - wbuf->used; | |
503 | int space_after; | |
504 | ||
505 | cond_resched(); | |
506 | ||
507 | /* Give the commit an opportunity to run */ | |
508 | if (ubifs_gc_should_commit(c)) { | |
509 | ret = -EAGAIN; | |
510 | break; | |
511 | } | |
512 | ||
513 | if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) { | |
514 | /* | |
515 | * We've done enough iterations. Indexing LEBs were | |
516 | * moved and will be available after the commit. | |
517 | */ | |
518 | dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN"); | |
519 | ubifs_commit_required(c); | |
520 | ret = -EAGAIN; | |
521 | break; | |
522 | } | |
523 | ||
524 | if (i > HARD_LEBS_LIMIT) { | |
525 | /* | |
526 | * We've moved too many LEBs and have not made | |
527 | * progress, give up. | |
528 | */ | |
529 | dbg_gc("hard limit, -ENOSPC"); | |
530 | ret = -ENOSPC; | |
531 | break; | |
532 | } | |
533 | ||
534 | /* | |
535 | * Empty and freeable LEBs can turn up while we waited for | |
536 | * the wbuf lock, or while we have been running GC. In that | |
537 | * case, we should just return one of those instead of | |
538 | * continuing to GC dirty LEBs. Hence we request | |
539 | * 'ubifs_find_dirty_leb()' to return an empty LEB if it can. | |
540 | */ | |
541 | ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1); | |
542 | if (ret) { | |
543 | if (ret == -ENOSPC) | |
544 | dbg_gc("no more dirty LEBs"); | |
545 | break; | |
546 | } | |
547 | ||
548 | dbg_gc("found LEB %d: free %d, dirty %d, sum %d " | |
549 | "(min. space %d)", lp.lnum, lp.free, lp.dirty, | |
550 | lp.free + lp.dirty, min_space); | |
551 | ||
552 | if (lp.free + lp.dirty == c->leb_size) { | |
553 | /* An empty LEB was returned */ | |
554 | dbg_gc("LEB %d is free, return it", lp.lnum); | |
555 | /* | |
556 | * ubifs_find_dirty_leb() doesn't return freeable index | |
557 | * LEBs. | |
558 | */ | |
559 | ubifs_assert(!(lp.flags & LPROPS_INDEX)); | |
560 | if (lp.free != c->leb_size) { | |
561 | /* | |
562 | * Write buffers must be sync'd before | |
563 | * unmapping freeable LEBs, because one of them | |
564 | * may contain data which obsoletes something | |
565 | * in 'lp.pnum'. | |
566 | */ | |
567 | ret = gc_sync_wbufs(c); | |
568 | if (ret) | |
569 | goto out; | |
570 | ret = ubifs_change_one_lp(c, lp.lnum, | |
571 | c->leb_size, 0, 0, 0, | |
572 | 0); | |
573 | if (ret) | |
574 | goto out; | |
575 | } | |
576 | ret = ubifs_leb_unmap(c, lp.lnum); | |
577 | if (ret) | |
578 | goto out; | |
579 | ret = lp.lnum; | |
580 | break; | |
581 | } | |
582 | ||
583 | space_before = c->leb_size - wbuf->offs - wbuf->used; | |
584 | if (wbuf->lnum == -1) | |
585 | space_before = 0; | |
586 | ||
587 | ret = ubifs_garbage_collect_leb(c, &lp); | |
588 | if (ret < 0) { | |
589 | if (ret == -EAGAIN || ret == -ENOSPC) { | |
590 | /* | |
591 | * These codes are not errors, so we have to | |
592 | * return the LEB to lprops. But if the | |
593 | * 'ubifs_return_leb()' function fails, its | |
594 | * failure code is propagated to the caller | |
595 | * instead of the original '-EAGAIN' or | |
596 | * '-ENOSPC'. | |
597 | */ | |
598 | err = ubifs_return_leb(c, lp.lnum); | |
599 | if (err) | |
600 | ret = err; | |
601 | break; | |
602 | } | |
603 | goto out; | |
604 | } | |
605 | ||
606 | if (ret == LEB_FREED) { | |
607 | /* An LEB has been freed and is ready for use */ | |
608 | dbg_gc("LEB %d freed, return", lp.lnum); | |
609 | ret = lp.lnum; | |
610 | break; | |
611 | } | |
612 | ||
613 | if (ret == LEB_FREED_IDX) { | |
614 | /* | |
615 | * This was an indexing LEB and it cannot be | |
616 | * immediately used. And instead of requesting the | |
617 | * commit straight away, we try to garbage collect some | |
618 | * more. | |
619 | */ | |
620 | dbg_gc("indexing LEB %d freed, continue", lp.lnum); | |
621 | continue; | |
622 | } | |
623 | ||
624 | ubifs_assert(ret == LEB_RETAINED); | |
625 | space_after = c->leb_size - wbuf->offs - wbuf->used; | |
626 | dbg_gc("LEB %d retained, freed %d bytes", lp.lnum, | |
627 | space_after - space_before); | |
628 | ||
629 | if (space_after > space_before) { | |
630 | /* GC makes progress, keep working */ | |
631 | min_space >>= 1; | |
632 | if (min_space < c->dead_wm) | |
633 | min_space = c->dead_wm; | |
634 | continue; | |
635 | } | |
636 | ||
637 | dbg_gc("did not make progress"); | |
638 | ||
639 | /* | |
640 | * GC moved an LEB bud have not done any progress. This means | |
641 | * that the previous GC head LEB contained too few free space | |
642 | * and the LEB which was GC'ed contained only large nodes which | |
643 | * did not fit that space. | |
644 | * | |
645 | * We can do 2 things: | |
646 | * 1. pick another LEB in a hope it'll contain a small node | |
647 | * which will fit the space we have at the end of current GC | |
648 | * head LEB, but there is no guarantee, so we try this out | |
649 | * unless we have already been working for too long; | |
650 | * 2. request an LEB with more dirty space, which will force | |
651 | * 'ubifs_find_dirty_leb()' to start scanning the lprops | |
652 | * table, instead of just picking one from the heap | |
653 | * (previously it already picked the dirtiest LEB). | |
654 | */ | |
655 | if (i < SOFT_LEBS_LIMIT) { | |
656 | dbg_gc("try again"); | |
657 | continue; | |
658 | } | |
659 | ||
660 | min_space <<= 1; | |
661 | if (min_space > c->dark_wm) | |
662 | min_space = c->dark_wm; | |
663 | dbg_gc("set min. space to %d", min_space); | |
664 | } | |
665 | ||
666 | if (ret == -ENOSPC && !list_empty(&c->idx_gc)) { | |
667 | dbg_gc("no space, some index LEBs GC'ed, -EAGAIN"); | |
668 | ubifs_commit_required(c); | |
669 | ret = -EAGAIN; | |
670 | } | |
671 | ||
672 | err = ubifs_wbuf_sync_nolock(wbuf); | |
673 | if (!err) | |
674 | err = ubifs_leb_unmap(c, c->gc_lnum); | |
675 | if (err) { | |
676 | ret = err; | |
677 | goto out; | |
678 | } | |
679 | out_unlock: | |
680 | mutex_unlock(&wbuf->io_mutex); | |
681 | return ret; | |
682 | ||
683 | out: | |
684 | ubifs_assert(ret < 0); | |
685 | ubifs_assert(ret != -ENOSPC && ret != -EAGAIN); | |
686 | ubifs_ro_mode(c, ret); | |
687 | ubifs_wbuf_sync_nolock(wbuf); | |
688 | mutex_unlock(&wbuf->io_mutex); | |
689 | ubifs_return_leb(c, lp.lnum); | |
690 | return ret; | |
691 | } | |
692 | ||
693 | /** | |
694 | * ubifs_gc_start_commit - garbage collection at start of commit. | |
695 | * @c: UBIFS file-system description object | |
696 | * | |
697 | * If a LEB has only dirty and free space, then we may safely unmap it and make | |
698 | * it free. Note, we cannot do this with indexing LEBs because dirty space may | |
699 | * correspond index nodes that are required for recovery. In that case, the | |
700 | * LEB cannot be unmapped until after the next commit. | |
701 | * | |
702 | * This function returns %0 upon success and a negative error code upon failure. | |
703 | */ | |
704 | int ubifs_gc_start_commit(struct ubifs_info *c) | |
705 | { | |
706 | struct ubifs_gced_idx_leb *idx_gc; | |
707 | const struct ubifs_lprops *lp; | |
708 | int err = 0, flags; | |
709 | ||
710 | ubifs_get_lprops(c); | |
711 | ||
712 | /* | |
713 | * Unmap (non-index) freeable LEBs. Note that recovery requires that all | |
714 | * wbufs are sync'd before this, which is done in 'do_commit()'. | |
715 | */ | |
716 | while (1) { | |
717 | lp = ubifs_fast_find_freeable(c); | |
8d47aef4 | 718 | if (IS_ERR(lp)) { |
1e51764a AB |
719 | err = PTR_ERR(lp); |
720 | goto out; | |
721 | } | |
722 | if (!lp) | |
723 | break; | |
724 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | |
725 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | |
726 | err = ubifs_leb_unmap(c, lp->lnum); | |
727 | if (err) | |
728 | goto out; | |
729 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0); | |
8d47aef4 | 730 | if (IS_ERR(lp)) { |
1e51764a AB |
731 | err = PTR_ERR(lp); |
732 | goto out; | |
733 | } | |
734 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | |
735 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | |
736 | } | |
737 | ||
738 | /* Mark GC'd index LEBs OK to unmap after this commit finishes */ | |
739 | list_for_each_entry(idx_gc, &c->idx_gc, list) | |
740 | idx_gc->unmap = 1; | |
741 | ||
742 | /* Record index freeable LEBs for unmapping after commit */ | |
743 | while (1) { | |
744 | lp = ubifs_fast_find_frdi_idx(c); | |
8d47aef4 | 745 | if (IS_ERR(lp)) { |
1e51764a AB |
746 | err = PTR_ERR(lp); |
747 | goto out; | |
748 | } | |
749 | if (!lp) | |
750 | break; | |
751 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); | |
752 | if (!idx_gc) { | |
753 | err = -ENOMEM; | |
754 | goto out; | |
755 | } | |
756 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | |
757 | ubifs_assert(lp->flags & LPROPS_INDEX); | |
758 | /* Don't release the LEB until after the next commit */ | |
759 | flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX; | |
760 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1); | |
8d47aef4 | 761 | if (IS_ERR(lp)) { |
1e51764a AB |
762 | err = PTR_ERR(lp); |
763 | kfree(idx_gc); | |
764 | goto out; | |
765 | } | |
766 | ubifs_assert(lp->flags & LPROPS_TAKEN); | |
767 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | |
768 | idx_gc->lnum = lp->lnum; | |
769 | idx_gc->unmap = 1; | |
770 | list_add(&idx_gc->list, &c->idx_gc); | |
771 | } | |
772 | out: | |
773 | ubifs_release_lprops(c); | |
774 | return err; | |
775 | } | |
776 | ||
777 | /** | |
778 | * ubifs_gc_end_commit - garbage collection at end of commit. | |
779 | * @c: UBIFS file-system description object | |
780 | * | |
781 | * This function completes out-of-place garbage collection of index LEBs. | |
782 | */ | |
783 | int ubifs_gc_end_commit(struct ubifs_info *c) | |
784 | { | |
785 | struct ubifs_gced_idx_leb *idx_gc, *tmp; | |
786 | struct ubifs_wbuf *wbuf; | |
787 | int err = 0; | |
788 | ||
789 | wbuf = &c->jheads[GCHD].wbuf; | |
790 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); | |
791 | list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list) | |
792 | if (idx_gc->unmap) { | |
793 | dbg_gc("LEB %d", idx_gc->lnum); | |
794 | err = ubifs_leb_unmap(c, idx_gc->lnum); | |
795 | if (err) | |
796 | goto out; | |
797 | err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC, | |
798 | LPROPS_NC, 0, LPROPS_TAKEN, -1); | |
799 | if (err) | |
800 | goto out; | |
801 | list_del(&idx_gc->list); | |
802 | kfree(idx_gc); | |
803 | } | |
804 | out: | |
805 | mutex_unlock(&wbuf->io_mutex); | |
806 | return err; | |
807 | } | |
808 | ||
809 | /** | |
810 | * ubifs_destroy_idx_gc - destroy idx_gc list. | |
811 | * @c: UBIFS file-system description object | |
812 | * | |
813 | * This function destroys the idx_gc list. It is called when unmounting or | |
814 | * remounting read-only so locks are not needed. | |
815 | */ | |
816 | void ubifs_destroy_idx_gc(struct ubifs_info *c) | |
817 | { | |
818 | while (!list_empty(&c->idx_gc)) { | |
819 | struct ubifs_gced_idx_leb *idx_gc; | |
820 | ||
821 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, | |
822 | list); | |
823 | c->idx_gc_cnt -= 1; | |
824 | list_del(&idx_gc->list); | |
825 | kfree(idx_gc); | |
826 | } | |
827 | ||
828 | } | |
829 | ||
830 | /** | |
831 | * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list. | |
832 | * @c: UBIFS file-system description object | |
833 | * | |
834 | * Called during start commit so locks are not needed. | |
835 | */ | |
836 | int ubifs_get_idx_gc_leb(struct ubifs_info *c) | |
837 | { | |
838 | struct ubifs_gced_idx_leb *idx_gc; | |
839 | int lnum; | |
840 | ||
841 | if (list_empty(&c->idx_gc)) | |
842 | return -ENOSPC; | |
843 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list); | |
844 | lnum = idx_gc->lnum; | |
845 | /* c->idx_gc_cnt is updated by the caller when lprops are updated */ | |
846 | list_del(&idx_gc->list); | |
847 | kfree(idx_gc); | |
848 | return lnum; | |
849 | } |