diff options
Diffstat (limited to 'fs/ubifs/gc.c')
-rw-r--r-- | fs/ubifs/gc.c | 773 |
1 files changed, 773 insertions, 0 deletions
diff --git a/fs/ubifs/gc.c b/fs/ubifs/gc.c new file mode 100644 index 000000000000..d0f3dac29081 --- /dev/null +++ b/fs/ubifs/gc.c | |||
@@ -0,0 +1,773 @@ | |||
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 | /* | ||
48 | * GC may need to move more then one LEB to make progress. The below constants | ||
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 | |||
98 | /** | ||
99 | * move_nodes - move nodes. | ||
100 | * @c: UBIFS file-system description object | ||
101 | * @sleb: describes nodes to move | ||
102 | * | ||
103 | * This function moves valid nodes from data LEB described by @sleb to the GC | ||
104 | * journal head. The obsolete nodes are dropped. | ||
105 | * | ||
106 | * When moving nodes we have to deal with classical bin-packing problem: the | ||
107 | * space in the current GC journal head LEB and in @c->gc_lnum are the "bins", | ||
108 | * where the nodes in the @sleb->nodes list are the elements which should be | ||
109 | * fit optimally to the bins. This function uses the "first fit decreasing" | ||
110 | * strategy, although it does not really sort the nodes but just split them on | ||
111 | * 3 classes - large, medium, and small, so they are roughly sorted. | ||
112 | * | ||
113 | * This function returns zero in case of success, %-EAGAIN if commit is | ||
114 | * required, and other negative error codes in case of other failures. | ||
115 | */ | ||
116 | static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb) | ||
117 | { | ||
118 | struct ubifs_scan_node *snod, *tmp; | ||
119 | struct list_head large, medium, small; | ||
120 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | ||
121 | int avail, err, min = INT_MAX; | ||
122 | |||
123 | INIT_LIST_HEAD(&large); | ||
124 | INIT_LIST_HEAD(&medium); | ||
125 | INIT_LIST_HEAD(&small); | ||
126 | |||
127 | list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) { | ||
128 | struct list_head *lst; | ||
129 | |||
130 | ubifs_assert(snod->type != UBIFS_IDX_NODE); | ||
131 | ubifs_assert(snod->type != UBIFS_REF_NODE); | ||
132 | ubifs_assert(snod->type != UBIFS_CS_NODE); | ||
133 | |||
134 | err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum, | ||
135 | snod->offs, 0); | ||
136 | if (err < 0) | ||
137 | goto out; | ||
138 | |||
139 | lst = &snod->list; | ||
140 | list_del(lst); | ||
141 | if (!err) { | ||
142 | /* The node is obsolete, remove it from the list */ | ||
143 | kfree(snod); | ||
144 | continue; | ||
145 | } | ||
146 | |||
147 | /* | ||
148 | * Sort the list of nodes so that large nodes go first, and | ||
149 | * small nodes go last. | ||
150 | */ | ||
151 | if (snod->len > MEDIUM_NODE_WM) | ||
152 | list_add(lst, &large); | ||
153 | else if (snod->len > SMALL_NODE_WM) | ||
154 | list_add(lst, &medium); | ||
155 | else | ||
156 | list_add(lst, &small); | ||
157 | |||
158 | /* And find the smallest node */ | ||
159 | if (snod->len < min) | ||
160 | min = snod->len; | ||
161 | } | ||
162 | |||
163 | /* | ||
164 | * Join the tree lists so that we'd have one roughly sorted list | ||
165 | * ('large' will be the head of the joined list). | ||
166 | */ | ||
167 | list_splice(&medium, large.prev); | ||
168 | list_splice(&small, large.prev); | ||
169 | |||
170 | if (wbuf->lnum == -1) { | ||
171 | /* | ||
172 | * The GC journal head is not set, because it is the first GC | ||
173 | * invocation since mount. | ||
174 | */ | ||
175 | err = switch_gc_head(c); | ||
176 | if (err) | ||
177 | goto out; | ||
178 | } | ||
179 | |||
180 | /* Write nodes to their new location. Use the first-fit strategy */ | ||
181 | while (1) { | ||
182 | avail = c->leb_size - wbuf->offs - wbuf->used; | ||
183 | list_for_each_entry_safe(snod, tmp, &large, list) { | ||
184 | int new_lnum, new_offs; | ||
185 | |||
186 | if (avail < min) | ||
187 | break; | ||
188 | |||
189 | if (snod->len > avail) | ||
190 | /* This node does not fit */ | ||
191 | continue; | ||
192 | |||
193 | cond_resched(); | ||
194 | |||
195 | new_lnum = wbuf->lnum; | ||
196 | new_offs = wbuf->offs + wbuf->used; | ||
197 | err = ubifs_wbuf_write_nolock(wbuf, snod->node, | ||
198 | snod->len); | ||
199 | if (err) | ||
200 | goto out; | ||
201 | err = ubifs_tnc_replace(c, &snod->key, sleb->lnum, | ||
202 | snod->offs, new_lnum, new_offs, | ||
203 | snod->len); | ||
204 | if (err) | ||
205 | goto out; | ||
206 | |||
207 | avail = c->leb_size - wbuf->offs - wbuf->used; | ||
208 | list_del(&snod->list); | ||
209 | kfree(snod); | ||
210 | } | ||
211 | |||
212 | if (list_empty(&large)) | ||
213 | break; | ||
214 | |||
215 | /* | ||
216 | * Waste the rest of the space in the LEB and switch to the | ||
217 | * next LEB. | ||
218 | */ | ||
219 | err = switch_gc_head(c); | ||
220 | if (err) | ||
221 | goto out; | ||
222 | } | ||
223 | |||
224 | return 0; | ||
225 | |||
226 | out: | ||
227 | list_for_each_entry_safe(snod, tmp, &large, list) { | ||
228 | list_del(&snod->list); | ||
229 | kfree(snod); | ||
230 | } | ||
231 | return err; | ||
232 | } | ||
233 | |||
234 | /** | ||
235 | * gc_sync_wbufs - sync write-buffers for GC. | ||
236 | * @c: UBIFS file-system description object | ||
237 | * | ||
238 | * We must guarantee that obsoleting nodes are on flash. Unfortunately they may | ||
239 | * be in a write-buffer instead. That is, a node could be written to a | ||
240 | * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is | ||
241 | * erased before the write-buffer is sync'd and then there is an unclean | ||
242 | * unmount, then an existing node is lost. To avoid this, we sync all | ||
243 | * write-buffers. | ||
244 | * | ||
245 | * This function returns %0 on success or a negative error code on failure. | ||
246 | */ | ||
247 | static int gc_sync_wbufs(struct ubifs_info *c) | ||
248 | { | ||
249 | int err, i; | ||
250 | |||
251 | for (i = 0; i < c->jhead_cnt; i++) { | ||
252 | if (i == GCHD) | ||
253 | continue; | ||
254 | err = ubifs_wbuf_sync(&c->jheads[i].wbuf); | ||
255 | if (err) | ||
256 | return err; | ||
257 | } | ||
258 | return 0; | ||
259 | } | ||
260 | |||
261 | /** | ||
262 | * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock. | ||
263 | * @c: UBIFS file-system description object | ||
264 | * @lp: describes the LEB to garbage collect | ||
265 | * | ||
266 | * This function garbage-collects an LEB and returns one of the @LEB_FREED, | ||
267 | * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is | ||
268 | * required, and other negative error codes in case of failures. | ||
269 | */ | ||
270 | int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp) | ||
271 | { | ||
272 | struct ubifs_scan_leb *sleb; | ||
273 | struct ubifs_scan_node *snod; | ||
274 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | ||
275 | int err = 0, lnum = lp->lnum; | ||
276 | |||
277 | ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 || | ||
278 | c->need_recovery); | ||
279 | ubifs_assert(c->gc_lnum != lnum); | ||
280 | ubifs_assert(wbuf->lnum != lnum); | ||
281 | |||
282 | /* | ||
283 | * We scan the entire LEB even though we only really need to scan up to | ||
284 | * (c->leb_size - lp->free). | ||
285 | */ | ||
286 | sleb = ubifs_scan(c, lnum, 0, c->sbuf); | ||
287 | if (IS_ERR(sleb)) | ||
288 | return PTR_ERR(sleb); | ||
289 | |||
290 | ubifs_assert(!list_empty(&sleb->nodes)); | ||
291 | snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list); | ||
292 | |||
293 | if (snod->type == UBIFS_IDX_NODE) { | ||
294 | struct ubifs_gced_idx_leb *idx_gc; | ||
295 | |||
296 | dbg_gc("indexing LEB %d (free %d, dirty %d)", | ||
297 | lnum, lp->free, lp->dirty); | ||
298 | list_for_each_entry(snod, &sleb->nodes, list) { | ||
299 | struct ubifs_idx_node *idx = snod->node; | ||
300 | int level = le16_to_cpu(idx->level); | ||
301 | |||
302 | ubifs_assert(snod->type == UBIFS_IDX_NODE); | ||
303 | key_read(c, ubifs_idx_key(c, idx), &snod->key); | ||
304 | err = ubifs_dirty_idx_node(c, &snod->key, level, lnum, | ||
305 | snod->offs); | ||
306 | if (err) | ||
307 | goto out; | ||
308 | } | ||
309 | |||
310 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); | ||
311 | if (!idx_gc) { | ||
312 | err = -ENOMEM; | ||
313 | goto out; | ||
314 | } | ||
315 | |||
316 | idx_gc->lnum = lnum; | ||
317 | idx_gc->unmap = 0; | ||
318 | list_add(&idx_gc->list, &c->idx_gc); | ||
319 | |||
320 | /* | ||
321 | * Don't release the LEB until after the next commit, because | ||
322 | * it may contain date which is needed for recovery. So | ||
323 | * although we freed this LEB, it will become usable only after | ||
324 | * the commit. | ||
325 | */ | ||
326 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, | ||
327 | LPROPS_INDEX, 1); | ||
328 | if (err) | ||
329 | goto out; | ||
330 | err = LEB_FREED_IDX; | ||
331 | } else { | ||
332 | dbg_gc("data LEB %d (free %d, dirty %d)", | ||
333 | lnum, lp->free, lp->dirty); | ||
334 | |||
335 | err = move_nodes(c, sleb); | ||
336 | if (err) | ||
337 | goto out; | ||
338 | |||
339 | err = gc_sync_wbufs(c); | ||
340 | if (err) | ||
341 | goto out; | ||
342 | |||
343 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0); | ||
344 | if (err) | ||
345 | goto out; | ||
346 | |||
347 | if (c->gc_lnum == -1) { | ||
348 | c->gc_lnum = lnum; | ||
349 | err = LEB_RETAINED; | ||
350 | } else { | ||
351 | err = ubifs_wbuf_sync_nolock(wbuf); | ||
352 | if (err) | ||
353 | goto out; | ||
354 | |||
355 | err = ubifs_leb_unmap(c, lnum); | ||
356 | if (err) | ||
357 | goto out; | ||
358 | |||
359 | err = LEB_FREED; | ||
360 | } | ||
361 | } | ||
362 | |||
363 | out: | ||
364 | ubifs_scan_destroy(sleb); | ||
365 | return err; | ||
366 | } | ||
367 | |||
368 | /** | ||
369 | * ubifs_garbage_collect - UBIFS garbage collector. | ||
370 | * @c: UBIFS file-system description object | ||
371 | * @anyway: do GC even if there are free LEBs | ||
372 | * | ||
373 | * This function does out-of-place garbage collection. The return codes are: | ||
374 | * o positive LEB number if the LEB has been freed and may be used; | ||
375 | * o %-EAGAIN if the caller has to run commit; | ||
376 | * o %-ENOSPC if GC failed to make any progress; | ||
377 | * o other negative error codes in case of other errors. | ||
378 | * | ||
379 | * Garbage collector writes data to the journal when GC'ing data LEBs, and just | ||
380 | * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point | ||
381 | * commit may be required. But commit cannot be run from inside GC, because the | ||
382 | * caller might be holding the commit lock, so %-EAGAIN is returned instead; | ||
383 | * And this error code means that the caller has to run commit, and re-run GC | ||
384 | * if there is still no free space. | ||
385 | * | ||
386 | * There are many reasons why this function may return %-EAGAIN: | ||
387 | * o the log is full and there is no space to write an LEB reference for | ||
388 | * @c->gc_lnum; | ||
389 | * o the journal is too large and exceeds size limitations; | ||
390 | * o GC moved indexing LEBs, but they can be used only after the commit; | ||
391 | * o the shrinker fails to find clean znodes to free and requests the commit; | ||
392 | * o etc. | ||
393 | * | ||
394 | * Note, if the file-system is close to be full, this function may return | ||
395 | * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of | ||
396 | * the function. E.g., this happens if the limits on the journal size are too | ||
397 | * tough and GC writes too much to the journal before an LEB is freed. This | ||
398 | * might also mean that the journal is too large, and the TNC becomes to big, | ||
399 | * so that the shrinker is constantly called, finds not clean znodes to free, | ||
400 | * and requests commit. Well, this may also happen if the journal is all right, | ||
401 | * but another kernel process consumes too much memory. Anyway, infinite | ||
402 | * %-EAGAIN may happen, but in some extreme/misconfiguration cases. | ||
403 | */ | ||
404 | int ubifs_garbage_collect(struct ubifs_info *c, int anyway) | ||
405 | { | ||
406 | int i, err, ret, min_space = c->dead_wm; | ||
407 | struct ubifs_lprops lp; | ||
408 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; | ||
409 | |||
410 | ubifs_assert_cmt_locked(c); | ||
411 | |||
412 | if (ubifs_gc_should_commit(c)) | ||
413 | return -EAGAIN; | ||
414 | |||
415 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); | ||
416 | |||
417 | if (c->ro_media) { | ||
418 | ret = -EROFS; | ||
419 | goto out_unlock; | ||
420 | } | ||
421 | |||
422 | /* We expect the write-buffer to be empty on entry */ | ||
423 | ubifs_assert(!wbuf->used); | ||
424 | |||
425 | for (i = 0; ; i++) { | ||
426 | int space_before = c->leb_size - wbuf->offs - wbuf->used; | ||
427 | int space_after; | ||
428 | |||
429 | cond_resched(); | ||
430 | |||
431 | /* Give the commit an opportunity to run */ | ||
432 | if (ubifs_gc_should_commit(c)) { | ||
433 | ret = -EAGAIN; | ||
434 | break; | ||
435 | } | ||
436 | |||
437 | if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) { | ||
438 | /* | ||
439 | * We've done enough iterations. Indexing LEBs were | ||
440 | * moved and will be available after the commit. | ||
441 | */ | ||
442 | dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN"); | ||
443 | ubifs_commit_required(c); | ||
444 | ret = -EAGAIN; | ||
445 | break; | ||
446 | } | ||
447 | |||
448 | if (i > HARD_LEBS_LIMIT) { | ||
449 | /* | ||
450 | * We've moved too many LEBs and have not made | ||
451 | * progress, give up. | ||
452 | */ | ||
453 | dbg_gc("hard limit, -ENOSPC"); | ||
454 | ret = -ENOSPC; | ||
455 | break; | ||
456 | } | ||
457 | |||
458 | /* | ||
459 | * Empty and freeable LEBs can turn up while we waited for | ||
460 | * the wbuf lock, or while we have been running GC. In that | ||
461 | * case, we should just return one of those instead of | ||
462 | * continuing to GC dirty LEBs. Hence we request | ||
463 | * 'ubifs_find_dirty_leb()' to return an empty LEB if it can. | ||
464 | */ | ||
465 | ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1); | ||
466 | if (ret) { | ||
467 | if (ret == -ENOSPC) | ||
468 | dbg_gc("no more dirty LEBs"); | ||
469 | break; | ||
470 | } | ||
471 | |||
472 | dbg_gc("found LEB %d: free %d, dirty %d, sum %d " | ||
473 | "(min. space %d)", lp.lnum, lp.free, lp.dirty, | ||
474 | lp.free + lp.dirty, min_space); | ||
475 | |||
476 | if (lp.free + lp.dirty == c->leb_size) { | ||
477 | /* An empty LEB was returned */ | ||
478 | dbg_gc("LEB %d is free, return it", lp.lnum); | ||
479 | /* | ||
480 | * ubifs_find_dirty_leb() doesn't return freeable index | ||
481 | * LEBs. | ||
482 | */ | ||
483 | ubifs_assert(!(lp.flags & LPROPS_INDEX)); | ||
484 | if (lp.free != c->leb_size) { | ||
485 | /* | ||
486 | * Write buffers must be sync'd before | ||
487 | * unmapping freeable LEBs, because one of them | ||
488 | * may contain data which obsoletes something | ||
489 | * in 'lp.pnum'. | ||
490 | */ | ||
491 | ret = gc_sync_wbufs(c); | ||
492 | if (ret) | ||
493 | goto out; | ||
494 | ret = ubifs_change_one_lp(c, lp.lnum, | ||
495 | c->leb_size, 0, 0, 0, | ||
496 | 0); | ||
497 | if (ret) | ||
498 | goto out; | ||
499 | } | ||
500 | ret = ubifs_leb_unmap(c, lp.lnum); | ||
501 | if (ret) | ||
502 | goto out; | ||
503 | ret = lp.lnum; | ||
504 | break; | ||
505 | } | ||
506 | |||
507 | space_before = c->leb_size - wbuf->offs - wbuf->used; | ||
508 | if (wbuf->lnum == -1) | ||
509 | space_before = 0; | ||
510 | |||
511 | ret = ubifs_garbage_collect_leb(c, &lp); | ||
512 | if (ret < 0) { | ||
513 | if (ret == -EAGAIN || ret == -ENOSPC) { | ||
514 | /* | ||
515 | * These codes are not errors, so we have to | ||
516 | * return the LEB to lprops. But if the | ||
517 | * 'ubifs_return_leb()' function fails, its | ||
518 | * failure code is propagated to the caller | ||
519 | * instead of the original '-EAGAIN' or | ||
520 | * '-ENOSPC'. | ||
521 | */ | ||
522 | err = ubifs_return_leb(c, lp.lnum); | ||
523 | if (err) | ||
524 | ret = err; | ||
525 | break; | ||
526 | } | ||
527 | goto out; | ||
528 | } | ||
529 | |||
530 | if (ret == LEB_FREED) { | ||
531 | /* An LEB has been freed and is ready for use */ | ||
532 | dbg_gc("LEB %d freed, return", lp.lnum); | ||
533 | ret = lp.lnum; | ||
534 | break; | ||
535 | } | ||
536 | |||
537 | if (ret == LEB_FREED_IDX) { | ||
538 | /* | ||
539 | * This was an indexing LEB and it cannot be | ||
540 | * immediately used. And instead of requesting the | ||
541 | * commit straight away, we try to garbage collect some | ||
542 | * more. | ||
543 | */ | ||
544 | dbg_gc("indexing LEB %d freed, continue", lp.lnum); | ||
545 | continue; | ||
546 | } | ||
547 | |||
548 | ubifs_assert(ret == LEB_RETAINED); | ||
549 | space_after = c->leb_size - wbuf->offs - wbuf->used; | ||
550 | dbg_gc("LEB %d retained, freed %d bytes", lp.lnum, | ||
551 | space_after - space_before); | ||
552 | |||
553 | if (space_after > space_before) { | ||
554 | /* GC makes progress, keep working */ | ||
555 | min_space >>= 1; | ||
556 | if (min_space < c->dead_wm) | ||
557 | min_space = c->dead_wm; | ||
558 | continue; | ||
559 | } | ||
560 | |||
561 | dbg_gc("did not make progress"); | ||
562 | |||
563 | /* | ||
564 | * GC moved an LEB bud have not done any progress. This means | ||
565 | * that the previous GC head LEB contained too few free space | ||
566 | * and the LEB which was GC'ed contained only large nodes which | ||
567 | * did not fit that space. | ||
568 | * | ||
569 | * We can do 2 things: | ||
570 | * 1. pick another LEB in a hope it'll contain a small node | ||
571 | * which will fit the space we have at the end of current GC | ||
572 | * head LEB, but there is no guarantee, so we try this out | ||
573 | * unless we have already been working for too long; | ||
574 | * 2. request an LEB with more dirty space, which will force | ||
575 | * 'ubifs_find_dirty_leb()' to start scanning the lprops | ||
576 | * table, instead of just picking one from the heap | ||
577 | * (previously it already picked the dirtiest LEB). | ||
578 | */ | ||
579 | if (i < SOFT_LEBS_LIMIT) { | ||
580 | dbg_gc("try again"); | ||
581 | continue; | ||
582 | } | ||
583 | |||
584 | min_space <<= 1; | ||
585 | if (min_space > c->dark_wm) | ||
586 | min_space = c->dark_wm; | ||
587 | dbg_gc("set min. space to %d", min_space); | ||
588 | } | ||
589 | |||
590 | if (ret == -ENOSPC && !list_empty(&c->idx_gc)) { | ||
591 | dbg_gc("no space, some index LEBs GC'ed, -EAGAIN"); | ||
592 | ubifs_commit_required(c); | ||
593 | ret = -EAGAIN; | ||
594 | } | ||
595 | |||
596 | err = ubifs_wbuf_sync_nolock(wbuf); | ||
597 | if (!err) | ||
598 | err = ubifs_leb_unmap(c, c->gc_lnum); | ||
599 | if (err) { | ||
600 | ret = err; | ||
601 | goto out; | ||
602 | } | ||
603 | out_unlock: | ||
604 | mutex_unlock(&wbuf->io_mutex); | ||
605 | return ret; | ||
606 | |||
607 | out: | ||
608 | ubifs_assert(ret < 0); | ||
609 | ubifs_assert(ret != -ENOSPC && ret != -EAGAIN); | ||
610 | ubifs_ro_mode(c, ret); | ||
611 | ubifs_wbuf_sync_nolock(wbuf); | ||
612 | mutex_unlock(&wbuf->io_mutex); | ||
613 | ubifs_return_leb(c, lp.lnum); | ||
614 | return ret; | ||
615 | } | ||
616 | |||
617 | /** | ||
618 | * ubifs_gc_start_commit - garbage collection at start of commit. | ||
619 | * @c: UBIFS file-system description object | ||
620 | * | ||
621 | * If a LEB has only dirty and free space, then we may safely unmap it and make | ||
622 | * it free. Note, we cannot do this with indexing LEBs because dirty space may | ||
623 | * correspond index nodes that are required for recovery. In that case, the | ||
624 | * LEB cannot be unmapped until after the next commit. | ||
625 | * | ||
626 | * This function returns %0 upon success and a negative error code upon failure. | ||
627 | */ | ||
628 | int ubifs_gc_start_commit(struct ubifs_info *c) | ||
629 | { | ||
630 | struct ubifs_gced_idx_leb *idx_gc; | ||
631 | const struct ubifs_lprops *lp; | ||
632 | int err = 0, flags; | ||
633 | |||
634 | ubifs_get_lprops(c); | ||
635 | |||
636 | /* | ||
637 | * Unmap (non-index) freeable LEBs. Note that recovery requires that all | ||
638 | * wbufs are sync'd before this, which is done in 'do_commit()'. | ||
639 | */ | ||
640 | while (1) { | ||
641 | lp = ubifs_fast_find_freeable(c); | ||
642 | if (unlikely(IS_ERR(lp))) { | ||
643 | err = PTR_ERR(lp); | ||
644 | goto out; | ||
645 | } | ||
646 | if (!lp) | ||
647 | break; | ||
648 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | ||
649 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | ||
650 | err = ubifs_leb_unmap(c, lp->lnum); | ||
651 | if (err) | ||
652 | goto out; | ||
653 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0); | ||
654 | if (unlikely(IS_ERR(lp))) { | ||
655 | err = PTR_ERR(lp); | ||
656 | goto out; | ||
657 | } | ||
658 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | ||
659 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | ||
660 | } | ||
661 | |||
662 | /* Mark GC'd index LEBs OK to unmap after this commit finishes */ | ||
663 | list_for_each_entry(idx_gc, &c->idx_gc, list) | ||
664 | idx_gc->unmap = 1; | ||
665 | |||
666 | /* Record index freeable LEBs for unmapping after commit */ | ||
667 | while (1) { | ||
668 | lp = ubifs_fast_find_frdi_idx(c); | ||
669 | if (unlikely(IS_ERR(lp))) { | ||
670 | err = PTR_ERR(lp); | ||
671 | goto out; | ||
672 | } | ||
673 | if (!lp) | ||
674 | break; | ||
675 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); | ||
676 | if (!idx_gc) { | ||
677 | err = -ENOMEM; | ||
678 | goto out; | ||
679 | } | ||
680 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); | ||
681 | ubifs_assert(lp->flags & LPROPS_INDEX); | ||
682 | /* Don't release the LEB until after the next commit */ | ||
683 | flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX; | ||
684 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1); | ||
685 | if (unlikely(IS_ERR(lp))) { | ||
686 | err = PTR_ERR(lp); | ||
687 | kfree(idx_gc); | ||
688 | goto out; | ||
689 | } | ||
690 | ubifs_assert(lp->flags & LPROPS_TAKEN); | ||
691 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); | ||
692 | idx_gc->lnum = lp->lnum; | ||
693 | idx_gc->unmap = 1; | ||
694 | list_add(&idx_gc->list, &c->idx_gc); | ||
695 | } | ||
696 | out: | ||
697 | ubifs_release_lprops(c); | ||
698 | return err; | ||
699 | } | ||
700 | |||
701 | /** | ||
702 | * ubifs_gc_end_commit - garbage collection at end of commit. | ||
703 | * @c: UBIFS file-system description object | ||
704 | * | ||
705 | * This function completes out-of-place garbage collection of index LEBs. | ||
706 | */ | ||
707 | int ubifs_gc_end_commit(struct ubifs_info *c) | ||
708 | { | ||
709 | struct ubifs_gced_idx_leb *idx_gc, *tmp; | ||
710 | struct ubifs_wbuf *wbuf; | ||
711 | int err = 0; | ||
712 | |||
713 | wbuf = &c->jheads[GCHD].wbuf; | ||
714 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); | ||
715 | list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list) | ||
716 | if (idx_gc->unmap) { | ||
717 | dbg_gc("LEB %d", idx_gc->lnum); | ||
718 | err = ubifs_leb_unmap(c, idx_gc->lnum); | ||
719 | if (err) | ||
720 | goto out; | ||
721 | err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC, | ||
722 | LPROPS_NC, 0, LPROPS_TAKEN, -1); | ||
723 | if (err) | ||
724 | goto out; | ||
725 | list_del(&idx_gc->list); | ||
726 | kfree(idx_gc); | ||
727 | } | ||
728 | out: | ||
729 | mutex_unlock(&wbuf->io_mutex); | ||
730 | return err; | ||
731 | } | ||
732 | |||
733 | /** | ||
734 | * ubifs_destroy_idx_gc - destroy idx_gc list. | ||
735 | * @c: UBIFS file-system description object | ||
736 | * | ||
737 | * This function destroys the idx_gc list. It is called when unmounting or | ||
738 | * remounting read-only so locks are not needed. | ||
739 | */ | ||
740 | void ubifs_destroy_idx_gc(struct ubifs_info *c) | ||
741 | { | ||
742 | while (!list_empty(&c->idx_gc)) { | ||
743 | struct ubifs_gced_idx_leb *idx_gc; | ||
744 | |||
745 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, | ||
746 | list); | ||
747 | c->idx_gc_cnt -= 1; | ||
748 | list_del(&idx_gc->list); | ||
749 | kfree(idx_gc); | ||
750 | } | ||
751 | |||
752 | } | ||
753 | |||
754 | /** | ||
755 | * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list. | ||
756 | * @c: UBIFS file-system description object | ||
757 | * | ||
758 | * Called during start commit so locks are not needed. | ||
759 | */ | ||
760 | int ubifs_get_idx_gc_leb(struct ubifs_info *c) | ||
761 | { | ||
762 | struct ubifs_gced_idx_leb *idx_gc; | ||
763 | int lnum; | ||
764 | |||
765 | if (list_empty(&c->idx_gc)) | ||
766 | return -ENOSPC; | ||
767 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list); | ||
768 | lnum = idx_gc->lnum; | ||
769 | /* c->idx_gc_cnt is updated by the caller when lprops are updated */ | ||
770 | list_del(&idx_gc->list); | ||
771 | kfree(idx_gc); | ||
772 | return lnum; | ||
773 | } | ||