Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * This file is part of UBIFS.
4 *
5 * Copyright (C) 2006-2008 Nokia Corporation.
6 *
7 * Authors: Adrian Hunter
8 * Artem Bityutskiy (Битюцкий Артём)
9 */
10
11/*
12 * This file implements garbage collection. The procedure for garbage collection
13 * is different depending on whether a LEB as an index LEB (contains index
14 * nodes) or not. For non-index LEBs, garbage collection finds a LEB which
15 * contains a lot of dirty space (obsolete nodes), and copies the non-obsolete
16 * nodes to the journal, at which point the garbage-collected LEB is free to be
17 * reused. For index LEBs, garbage collection marks the non-obsolete index nodes
18 * dirty in the TNC, and after the next commit, the garbage-collected LEB is
19 * to be reused. Garbage collection will cause the number of dirty index nodes
20 * to grow, however sufficient space is reserved for the index to ensure the
21 * commit will never run out of space.
22 *
23 * Notes about dead watermark. At current UBIFS implementation we assume that
24 * LEBs which have less than @c->dead_wm bytes of free + dirty space are full
25 * and not worth garbage-collecting. The dead watermark is one min. I/O unit
26 * size, or min. UBIFS node size, depending on what is greater. Indeed, UBIFS
27 * Garbage Collector has to synchronize the GC head's write buffer before
28 * returning, so this is about wasting one min. I/O unit. However, UBIFS GC can
29 * actually reclaim even very small pieces of dirty space by garbage collecting
30 * enough dirty LEBs, but we do not bother doing this at this implementation.
31 *
32 * Notes about dark watermark. The results of GC work depends on how big are
33 * the UBIFS nodes GC deals with. Large nodes make GC waste more space. Indeed,
34 * if GC move data from LEB A to LEB B and nodes in LEB A are large, GC would
35 * have to waste large pieces of free space at the end of LEB B, because nodes
36 * from LEB A would not fit. And the worst situation is when all nodes are of
37 * maximum size. So dark watermark is the amount of free + dirty space in LEB
38 * which are guaranteed to be reclaimable. If LEB has less space, the GC might
39 * be unable to reclaim it. So, LEBs with free + dirty greater than dark
40 * watermark are "good" LEBs from GC's point of view. The other LEBs are not so
41 * good, and GC takes extra care when moving them.
42 */
43
44#include <linux/slab.h>
45#include <linux/pagemap.h>
46#include <linux/list_sort.h>
47#include "ubifs.h"
48
49/*
50 * GC may need to move more than one LEB to make progress. The below constants
51 * define "soft" and "hard" limits on the number of LEBs the garbage collector
52 * may move.
53 */
54#define SOFT_LEBS_LIMIT 4
55#define HARD_LEBS_LIMIT 32
56
57/**
58 * switch_gc_head - switch the garbage collection journal head.
59 * @c: UBIFS file-system description object
60 * @buf: buffer to write
61 * @len: length of the buffer to write
62 * @lnum: LEB number written is returned here
63 * @offs: offset written is returned here
64 *
65 * This function switch the GC head to the next LEB which is reserved in
66 * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required,
67 * and other negative error code in case of failures.
68 */
69static int switch_gc_head(struct ubifs_info *c)
70{
71 int err, gc_lnum = c->gc_lnum;
72 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
73
74 ubifs_assert(c, gc_lnum != -1);
75 dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)",
76 wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum,
77 c->leb_size - wbuf->offs - wbuf->used);
78
79 err = ubifs_wbuf_sync_nolock(wbuf);
80 if (err)
81 return err;
82
83 /*
84 * The GC write-buffer was synchronized, we may safely unmap
85 * 'c->gc_lnum'.
86 */
87 err = ubifs_leb_unmap(c, gc_lnum);
88 if (err)
89 return err;
90
91 err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0);
92 if (err)
93 return err;
94
95 c->gc_lnum = -1;
96 err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0);
97 return err;
98}
99
100/**
101 * data_nodes_cmp - compare 2 data nodes.
102 * @priv: UBIFS file-system description object
103 * @a: first data node
104 * @b: second data node
105 *
106 * This function compares data nodes @a and @b. Returns %1 if @a has greater
107 * inode or block number, and %-1 otherwise.
108 */
109static int data_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
110{
111 ino_t inuma, inumb;
112 struct ubifs_info *c = priv;
113 struct ubifs_scan_node *sa, *sb;
114
115 cond_resched();
116 if (a == b)
117 return 0;
118
119 sa = list_entry(a, struct ubifs_scan_node, list);
120 sb = list_entry(b, struct ubifs_scan_node, list);
121
122 ubifs_assert(c, key_type(c, &sa->key) == UBIFS_DATA_KEY);
123 ubifs_assert(c, key_type(c, &sb->key) == UBIFS_DATA_KEY);
124 ubifs_assert(c, sa->type == UBIFS_DATA_NODE);
125 ubifs_assert(c, sb->type == UBIFS_DATA_NODE);
126
127 inuma = key_inum(c, &sa->key);
128 inumb = key_inum(c, &sb->key);
129
130 if (inuma == inumb) {
131 unsigned int blka = key_block(c, &sa->key);
132 unsigned int blkb = key_block(c, &sb->key);
133
134 if (blka <= blkb)
135 return -1;
136 } else if (inuma <= inumb)
137 return -1;
138
139 return 1;
140}
141
142/*
143 * nondata_nodes_cmp - compare 2 non-data nodes.
144 * @priv: UBIFS file-system description object
145 * @a: first node
146 * @a: second node
147 *
148 * This function compares nodes @a and @b. It makes sure that inode nodes go
149 * first and sorted by length in descending order. Directory entry nodes go
150 * after inode nodes and are sorted in ascending hash valuer order.
151 */
152static int nondata_nodes_cmp(void *priv, struct list_head *a,
153 struct list_head *b)
154{
155 ino_t inuma, inumb;
156 struct ubifs_info *c = priv;
157 struct ubifs_scan_node *sa, *sb;
158
159 cond_resched();
160 if (a == b)
161 return 0;
162
163 sa = list_entry(a, struct ubifs_scan_node, list);
164 sb = list_entry(b, struct ubifs_scan_node, list);
165
166 ubifs_assert(c, key_type(c, &sa->key) != UBIFS_DATA_KEY &&
167 key_type(c, &sb->key) != UBIFS_DATA_KEY);
168 ubifs_assert(c, sa->type != UBIFS_DATA_NODE &&
169 sb->type != UBIFS_DATA_NODE);
170
171 /* Inodes go before directory entries */
172 if (sa->type == UBIFS_INO_NODE) {
173 if (sb->type == UBIFS_INO_NODE)
174 return sb->len - sa->len;
175 return -1;
176 }
177 if (sb->type == UBIFS_INO_NODE)
178 return 1;
179
180 ubifs_assert(c, key_type(c, &sa->key) == UBIFS_DENT_KEY ||
181 key_type(c, &sa->key) == UBIFS_XENT_KEY);
182 ubifs_assert(c, key_type(c, &sb->key) == UBIFS_DENT_KEY ||
183 key_type(c, &sb->key) == UBIFS_XENT_KEY);
184 ubifs_assert(c, sa->type == UBIFS_DENT_NODE ||
185 sa->type == UBIFS_XENT_NODE);
186 ubifs_assert(c, sb->type == UBIFS_DENT_NODE ||
187 sb->type == UBIFS_XENT_NODE);
188
189 inuma = key_inum(c, &sa->key);
190 inumb = key_inum(c, &sb->key);
191
192 if (inuma == inumb) {
193 uint32_t hasha = key_hash(c, &sa->key);
194 uint32_t hashb = key_hash(c, &sb->key);
195
196 if (hasha <= hashb)
197 return -1;
198 } else if (inuma <= inumb)
199 return -1;
200
201 return 1;
202}
203
204/**
205 * sort_nodes - sort nodes for GC.
206 * @c: UBIFS file-system description object
207 * @sleb: describes nodes to sort and contains the result on exit
208 * @nondata: contains non-data nodes on exit
209 * @min: minimum node size is returned here
210 *
211 * This function sorts the list of inodes to garbage collect. First of all, it
212 * kills obsolete nodes and separates data and non-data nodes to the
213 * @sleb->nodes and @nondata lists correspondingly.
214 *
215 * Data nodes are then sorted in block number order - this is important for
216 * bulk-read; data nodes with lower inode number go before data nodes with
217 * higher inode number, and data nodes with lower block number go before data
218 * nodes with higher block number;
219 *
220 * Non-data nodes are sorted as follows.
221 * o First go inode nodes - they are sorted in descending length order.
222 * o Then go directory entry nodes - they are sorted in hash order, which
223 * should supposedly optimize 'readdir()'. Direntry nodes with lower parent
224 * inode number go before direntry nodes with higher parent inode number,
225 * and direntry nodes with lower name hash values go before direntry nodes
226 * with higher name hash values.
227 *
228 * This function returns zero in case of success and a negative error code in
229 * case of failure.
230 */
231static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
232 struct list_head *nondata, int *min)
233{
234 int err;
235 struct ubifs_scan_node *snod, *tmp;
236
237 *min = INT_MAX;
238
239 /* Separate data nodes and non-data nodes */
240 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
241 ubifs_assert(c, snod->type == UBIFS_INO_NODE ||
242 snod->type == UBIFS_DATA_NODE ||
243 snod->type == UBIFS_DENT_NODE ||
244 snod->type == UBIFS_XENT_NODE ||
245 snod->type == UBIFS_TRUN_NODE ||
246 snod->type == UBIFS_AUTH_NODE);
247
248 if (snod->type != UBIFS_INO_NODE &&
249 snod->type != UBIFS_DATA_NODE &&
250 snod->type != UBIFS_DENT_NODE &&
251 snod->type != UBIFS_XENT_NODE) {
252 /* Probably truncation node, zap it */
253 list_del(&snod->list);
254 kfree(snod);
255 continue;
256 }
257
258 ubifs_assert(c, key_type(c, &snod->key) == UBIFS_DATA_KEY ||
259 key_type(c, &snod->key) == UBIFS_INO_KEY ||
260 key_type(c, &snod->key) == UBIFS_DENT_KEY ||
261 key_type(c, &snod->key) == UBIFS_XENT_KEY);
262
263 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
264 snod->offs, 0);
265 if (err < 0)
266 return err;
267
268 if (!err) {
269 /* The node is obsolete, remove it from the list */
270 list_del(&snod->list);
271 kfree(snod);
272 continue;
273 }
274
275 if (snod->len < *min)
276 *min = snod->len;
277
278 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
279 list_move_tail(&snod->list, nondata);
280 }
281
282 /* Sort data and non-data nodes */
283 list_sort(c, &sleb->nodes, &data_nodes_cmp);
284 list_sort(c, nondata, &nondata_nodes_cmp);
285
286 err = dbg_check_data_nodes_order(c, &sleb->nodes);
287 if (err)
288 return err;
289 err = dbg_check_nondata_nodes_order(c, nondata);
290 if (err)
291 return err;
292 return 0;
293}
294
295/**
296 * move_node - move a node.
297 * @c: UBIFS file-system description object
298 * @sleb: describes the LEB to move nodes from
299 * @snod: the mode to move
300 * @wbuf: write-buffer to move node to
301 *
302 * This function moves node @snod to @wbuf, changes TNC correspondingly, and
303 * destroys @snod. Returns zero in case of success and a negative error code in
304 * case of failure.
305 */
306static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
307 struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
308{
309 int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
310
311 cond_resched();
312 err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
313 if (err)
314 return err;
315
316 err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
317 snod->offs, new_lnum, new_offs,
318 snod->len);
319 list_del(&snod->list);
320 kfree(snod);
321 return err;
322}
323
324/**
325 * move_nodes - move nodes.
326 * @c: UBIFS file-system description object
327 * @sleb: describes the LEB to move nodes from
328 *
329 * This function moves valid nodes from data LEB described by @sleb to the GC
330 * journal head. This function returns zero in case of success, %-EAGAIN if
331 * commit is required, and other negative error codes in case of other
332 * failures.
333 */
334static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
335{
336 int err, min;
337 LIST_HEAD(nondata);
338 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
339
340 if (wbuf->lnum == -1) {
341 /*
342 * The GC journal head is not set, because it is the first GC
343 * invocation since mount.
344 */
345 err = switch_gc_head(c);
346 if (err)
347 return err;
348 }
349
350 err = sort_nodes(c, sleb, &nondata, &min);
351 if (err)
352 goto out;
353
354 /* Write nodes to their new location. Use the first-fit strategy */
355 while (1) {
356 int avail, moved = 0;
357 struct ubifs_scan_node *snod, *tmp;
358
359 /* Move data nodes */
360 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
361 avail = c->leb_size - wbuf->offs - wbuf->used -
362 ubifs_auth_node_sz(c);
363 if (snod->len > avail)
364 /*
365 * Do not skip data nodes in order to optimize
366 * bulk-read.
367 */
368 break;
369
370 err = ubifs_shash_update(c, c->jheads[GCHD].log_hash,
371 snod->node, snod->len);
372 if (err)
373 goto out;
374
375 err = move_node(c, sleb, snod, wbuf);
376 if (err)
377 goto out;
378 moved = 1;
379 }
380
381 /* Move non-data nodes */
382 list_for_each_entry_safe(snod, tmp, &nondata, list) {
383 avail = c->leb_size - wbuf->offs - wbuf->used -
384 ubifs_auth_node_sz(c);
385 if (avail < min)
386 break;
387
388 if (snod->len > avail) {
389 /*
390 * Keep going only if this is an inode with
391 * some data. Otherwise stop and switch the GC
392 * head. IOW, we assume that data-less inode
393 * nodes and direntry nodes are roughly of the
394 * same size.
395 */
396 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
397 snod->len == UBIFS_INO_NODE_SZ)
398 break;
399 continue;
400 }
401
402 err = ubifs_shash_update(c, c->jheads[GCHD].log_hash,
403 snod->node, snod->len);
404 if (err)
405 goto out;
406
407 err = move_node(c, sleb, snod, wbuf);
408 if (err)
409 goto out;
410 moved = 1;
411 }
412
413 if (ubifs_authenticated(c) && moved) {
414 struct ubifs_auth_node *auth;
415
416 auth = kmalloc(ubifs_auth_node_sz(c), GFP_NOFS);
417 if (!auth) {
418 err = -ENOMEM;
419 goto out;
420 }
421
422 err = ubifs_prepare_auth_node(c, auth,
423 c->jheads[GCHD].log_hash);
424 if (err) {
425 kfree(auth);
426 goto out;
427 }
428
429 err = ubifs_wbuf_write_nolock(wbuf, auth,
430 ubifs_auth_node_sz(c));
431 if (err) {
432 kfree(auth);
433 goto out;
434 }
435
436 ubifs_add_dirt(c, wbuf->lnum, ubifs_auth_node_sz(c));
437 }
438
439 if (list_empty(&sleb->nodes) && list_empty(&nondata))
440 break;
441
442 /*
443 * Waste the rest of the space in the LEB and switch to the
444 * next LEB.
445 */
446 err = switch_gc_head(c);
447 if (err)
448 goto out;
449 }
450
451 return 0;
452
453out:
454 list_splice_tail(&nondata, &sleb->nodes);
455 return err;
456}
457
458/**
459 * gc_sync_wbufs - sync write-buffers for GC.
460 * @c: UBIFS file-system description object
461 *
462 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
463 * be in a write-buffer instead. That is, a node could be written to a
464 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
465 * erased before the write-buffer is sync'd and then there is an unclean
466 * unmount, then an existing node is lost. To avoid this, we sync all
467 * write-buffers.
468 *
469 * This function returns %0 on success or a negative error code on failure.
470 */
471static int gc_sync_wbufs(struct ubifs_info *c)
472{
473 int err, i;
474
475 for (i = 0; i < c->jhead_cnt; i++) {
476 if (i == GCHD)
477 continue;
478 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
479 if (err)
480 return err;
481 }
482 return 0;
483}
484
485/**
486 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
487 * @c: UBIFS file-system description object
488 * @lp: describes the LEB to garbage collect
489 *
490 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
491 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
492 * required, and other negative error codes in case of failures.
493 */
494int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
495{
496 struct ubifs_scan_leb *sleb;
497 struct ubifs_scan_node *snod;
498 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
499 int err = 0, lnum = lp->lnum;
500
501 ubifs_assert(c, c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
502 c->need_recovery);
503 ubifs_assert(c, c->gc_lnum != lnum);
504 ubifs_assert(c, wbuf->lnum != lnum);
505
506 if (lp->free + lp->dirty == c->leb_size) {
507 /* Special case - a free LEB */
508 dbg_gc("LEB %d is free, return it", lp->lnum);
509 ubifs_assert(c, !(lp->flags & LPROPS_INDEX));
510
511 if (lp->free != c->leb_size) {
512 /*
513 * Write buffers must be sync'd before unmapping
514 * freeable LEBs, because one of them may contain data
515 * which obsoletes something in 'lp->lnum'.
516 */
517 err = gc_sync_wbufs(c);
518 if (err)
519 return err;
520 err = ubifs_change_one_lp(c, lp->lnum, c->leb_size,
521 0, 0, 0, 0);
522 if (err)
523 return err;
524 }
525 err = ubifs_leb_unmap(c, lp->lnum);
526 if (err)
527 return err;
528
529 if (c->gc_lnum == -1) {
530 c->gc_lnum = lnum;
531 return LEB_RETAINED;
532 }
533
534 return LEB_FREED;
535 }
536
537 /*
538 * We scan the entire LEB even though we only really need to scan up to
539 * (c->leb_size - lp->free).
540 */
541 sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
542 if (IS_ERR(sleb))
543 return PTR_ERR(sleb);
544
545 ubifs_assert(c, !list_empty(&sleb->nodes));
546 snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
547
548 if (snod->type == UBIFS_IDX_NODE) {
549 struct ubifs_gced_idx_leb *idx_gc;
550
551 dbg_gc("indexing LEB %d (free %d, dirty %d)",
552 lnum, lp->free, lp->dirty);
553 list_for_each_entry(snod, &sleb->nodes, list) {
554 struct ubifs_idx_node *idx = snod->node;
555 int level = le16_to_cpu(idx->level);
556
557 ubifs_assert(c, snod->type == UBIFS_IDX_NODE);
558 key_read(c, ubifs_idx_key(c, idx), &snod->key);
559 err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
560 snod->offs);
561 if (err)
562 goto out;
563 }
564
565 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
566 if (!idx_gc) {
567 err = -ENOMEM;
568 goto out;
569 }
570
571 idx_gc->lnum = lnum;
572 idx_gc->unmap = 0;
573 list_add(&idx_gc->list, &c->idx_gc);
574
575 /*
576 * Don't release the LEB until after the next commit, because
577 * it may contain data which is needed for recovery. So
578 * although we freed this LEB, it will become usable only after
579 * the commit.
580 */
581 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
582 LPROPS_INDEX, 1);
583 if (err)
584 goto out;
585 err = LEB_FREED_IDX;
586 } else {
587 dbg_gc("data LEB %d (free %d, dirty %d)",
588 lnum, lp->free, lp->dirty);
589
590 err = move_nodes(c, sleb);
591 if (err)
592 goto out_inc_seq;
593
594 err = gc_sync_wbufs(c);
595 if (err)
596 goto out_inc_seq;
597
598 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
599 if (err)
600 goto out_inc_seq;
601
602 /* Allow for races with TNC */
603 c->gced_lnum = lnum;
604 smp_wmb();
605 c->gc_seq += 1;
606 smp_wmb();
607
608 if (c->gc_lnum == -1) {
609 c->gc_lnum = lnum;
610 err = LEB_RETAINED;
611 } else {
612 err = ubifs_wbuf_sync_nolock(wbuf);
613 if (err)
614 goto out;
615
616 err = ubifs_leb_unmap(c, lnum);
617 if (err)
618 goto out;
619
620 err = LEB_FREED;
621 }
622 }
623
624out:
625 ubifs_scan_destroy(sleb);
626 return err;
627
628out_inc_seq:
629 /* We may have moved at least some nodes so allow for races with TNC */
630 c->gced_lnum = lnum;
631 smp_wmb();
632 c->gc_seq += 1;
633 smp_wmb();
634 goto out;
635}
636
637/**
638 * ubifs_garbage_collect - UBIFS garbage collector.
639 * @c: UBIFS file-system description object
640 * @anyway: do GC even if there are free LEBs
641 *
642 * This function does out-of-place garbage collection. The return codes are:
643 * o positive LEB number if the LEB has been freed and may be used;
644 * o %-EAGAIN if the caller has to run commit;
645 * o %-ENOSPC if GC failed to make any progress;
646 * o other negative error codes in case of other errors.
647 *
648 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
649 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
650 * commit may be required. But commit cannot be run from inside GC, because the
651 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
652 * And this error code means that the caller has to run commit, and re-run GC
653 * if there is still no free space.
654 *
655 * There are many reasons why this function may return %-EAGAIN:
656 * o the log is full and there is no space to write an LEB reference for
657 * @c->gc_lnum;
658 * o the journal is too large and exceeds size limitations;
659 * o GC moved indexing LEBs, but they can be used only after the commit;
660 * o the shrinker fails to find clean znodes to free and requests the commit;
661 * o etc.
662 *
663 * Note, if the file-system is close to be full, this function may return
664 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
665 * the function. E.g., this happens if the limits on the journal size are too
666 * tough and GC writes too much to the journal before an LEB is freed. This
667 * might also mean that the journal is too large, and the TNC becomes to big,
668 * so that the shrinker is constantly called, finds not clean znodes to free,
669 * and requests commit. Well, this may also happen if the journal is all right,
670 * but another kernel process consumes too much memory. Anyway, infinite
671 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
672 */
673int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
674{
675 int i, err, ret, min_space = c->dead_wm;
676 struct ubifs_lprops lp;
677 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
678
679 ubifs_assert_cmt_locked(c);
680 ubifs_assert(c, !c->ro_media && !c->ro_mount);
681
682 if (ubifs_gc_should_commit(c))
683 return -EAGAIN;
684
685 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
686
687 if (c->ro_error) {
688 ret = -EROFS;
689 goto out_unlock;
690 }
691
692 /* We expect the write-buffer to be empty on entry */
693 ubifs_assert(c, !wbuf->used);
694
695 for (i = 0; ; i++) {
696 int space_before, space_after;
697
698 cond_resched();
699
700 /* Give the commit an opportunity to run */
701 if (ubifs_gc_should_commit(c)) {
702 ret = -EAGAIN;
703 break;
704 }
705
706 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
707 /*
708 * We've done enough iterations. Indexing LEBs were
709 * moved and will be available after the commit.
710 */
711 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
712 ubifs_commit_required(c);
713 ret = -EAGAIN;
714 break;
715 }
716
717 if (i > HARD_LEBS_LIMIT) {
718 /*
719 * We've moved too many LEBs and have not made
720 * progress, give up.
721 */
722 dbg_gc("hard limit, -ENOSPC");
723 ret = -ENOSPC;
724 break;
725 }
726
727 /*
728 * Empty and freeable LEBs can turn up while we waited for
729 * the wbuf lock, or while we have been running GC. In that
730 * case, we should just return one of those instead of
731 * continuing to GC dirty LEBs. Hence we request
732 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
733 */
734 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
735 if (ret) {
736 if (ret == -ENOSPC)
737 dbg_gc("no more dirty LEBs");
738 break;
739 }
740
741 dbg_gc("found LEB %d: free %d, dirty %d, sum %d (min. space %d)",
742 lp.lnum, lp.free, lp.dirty, lp.free + lp.dirty,
743 min_space);
744
745 space_before = c->leb_size - wbuf->offs - wbuf->used;
746 if (wbuf->lnum == -1)
747 space_before = 0;
748
749 ret = ubifs_garbage_collect_leb(c, &lp);
750 if (ret < 0) {
751 if (ret == -EAGAIN) {
752 /*
753 * This is not error, so we have to return the
754 * LEB to lprops. But if 'ubifs_return_leb()'
755 * fails, its failure code is propagated to the
756 * caller instead of the original '-EAGAIN'.
757 */
758 err = ubifs_return_leb(c, lp.lnum);
759 if (err)
760 ret = err;
761 break;
762 }
763 goto out;
764 }
765
766 if (ret == LEB_FREED) {
767 /* An LEB has been freed and is ready for use */
768 dbg_gc("LEB %d freed, return", lp.lnum);
769 ret = lp.lnum;
770 break;
771 }
772
773 if (ret == LEB_FREED_IDX) {
774 /*
775 * This was an indexing LEB and it cannot be
776 * immediately used. And instead of requesting the
777 * commit straight away, we try to garbage collect some
778 * more.
779 */
780 dbg_gc("indexing LEB %d freed, continue", lp.lnum);
781 continue;
782 }
783
784 ubifs_assert(c, ret == LEB_RETAINED);
785 space_after = c->leb_size - wbuf->offs - wbuf->used;
786 dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
787 space_after - space_before);
788
789 if (space_after > space_before) {
790 /* GC makes progress, keep working */
791 min_space >>= 1;
792 if (min_space < c->dead_wm)
793 min_space = c->dead_wm;
794 continue;
795 }
796
797 dbg_gc("did not make progress");
798
799 /*
800 * GC moved an LEB bud have not done any progress. This means
801 * that the previous GC head LEB contained too few free space
802 * and the LEB which was GC'ed contained only large nodes which
803 * did not fit that space.
804 *
805 * We can do 2 things:
806 * 1. pick another LEB in a hope it'll contain a small node
807 * which will fit the space we have at the end of current GC
808 * head LEB, but there is no guarantee, so we try this out
809 * unless we have already been working for too long;
810 * 2. request an LEB with more dirty space, which will force
811 * 'ubifs_find_dirty_leb()' to start scanning the lprops
812 * table, instead of just picking one from the heap
813 * (previously it already picked the dirtiest LEB).
814 */
815 if (i < SOFT_LEBS_LIMIT) {
816 dbg_gc("try again");
817 continue;
818 }
819
820 min_space <<= 1;
821 if (min_space > c->dark_wm)
822 min_space = c->dark_wm;
823 dbg_gc("set min. space to %d", min_space);
824 }
825
826 if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
827 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
828 ubifs_commit_required(c);
829 ret = -EAGAIN;
830 }
831
832 err = ubifs_wbuf_sync_nolock(wbuf);
833 if (!err)
834 err = ubifs_leb_unmap(c, c->gc_lnum);
835 if (err) {
836 ret = err;
837 goto out;
838 }
839out_unlock:
840 mutex_unlock(&wbuf->io_mutex);
841 return ret;
842
843out:
844 ubifs_assert(c, ret < 0);
845 ubifs_assert(c, ret != -ENOSPC && ret != -EAGAIN);
846 ubifs_wbuf_sync_nolock(wbuf);
847 ubifs_ro_mode(c, ret);
848 mutex_unlock(&wbuf->io_mutex);
849 ubifs_return_leb(c, lp.lnum);
850 return ret;
851}
852
853/**
854 * ubifs_gc_start_commit - garbage collection at start of commit.
855 * @c: UBIFS file-system description object
856 *
857 * If a LEB has only dirty and free space, then we may safely unmap it and make
858 * it free. Note, we cannot do this with indexing LEBs because dirty space may
859 * correspond index nodes that are required for recovery. In that case, the
860 * LEB cannot be unmapped until after the next commit.
861 *
862 * This function returns %0 upon success and a negative error code upon failure.
863 */
864int ubifs_gc_start_commit(struct ubifs_info *c)
865{
866 struct ubifs_gced_idx_leb *idx_gc;
867 const struct ubifs_lprops *lp;
868 int err = 0, flags;
869
870 ubifs_get_lprops(c);
871
872 /*
873 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
874 * wbufs are sync'd before this, which is done in 'do_commit()'.
875 */
876 while (1) {
877 lp = ubifs_fast_find_freeable(c);
878 if (!lp)
879 break;
880 ubifs_assert(c, !(lp->flags & LPROPS_TAKEN));
881 ubifs_assert(c, !(lp->flags & LPROPS_INDEX));
882 err = ubifs_leb_unmap(c, lp->lnum);
883 if (err)
884 goto out;
885 lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
886 if (IS_ERR(lp)) {
887 err = PTR_ERR(lp);
888 goto out;
889 }
890 ubifs_assert(c, !(lp->flags & LPROPS_TAKEN));
891 ubifs_assert(c, !(lp->flags & LPROPS_INDEX));
892 }
893
894 /* Mark GC'd index LEBs OK to unmap after this commit finishes */
895 list_for_each_entry(idx_gc, &c->idx_gc, list)
896 idx_gc->unmap = 1;
897
898 /* Record index freeable LEBs for unmapping after commit */
899 while (1) {
900 lp = ubifs_fast_find_frdi_idx(c);
901 if (IS_ERR(lp)) {
902 err = PTR_ERR(lp);
903 goto out;
904 }
905 if (!lp)
906 break;
907 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
908 if (!idx_gc) {
909 err = -ENOMEM;
910 goto out;
911 }
912 ubifs_assert(c, !(lp->flags & LPROPS_TAKEN));
913 ubifs_assert(c, lp->flags & LPROPS_INDEX);
914 /* Don't release the LEB until after the next commit */
915 flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
916 lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
917 if (IS_ERR(lp)) {
918 err = PTR_ERR(lp);
919 kfree(idx_gc);
920 goto out;
921 }
922 ubifs_assert(c, lp->flags & LPROPS_TAKEN);
923 ubifs_assert(c, !(lp->flags & LPROPS_INDEX));
924 idx_gc->lnum = lp->lnum;
925 idx_gc->unmap = 1;
926 list_add(&idx_gc->list, &c->idx_gc);
927 }
928out:
929 ubifs_release_lprops(c);
930 return err;
931}
932
933/**
934 * ubifs_gc_end_commit - garbage collection at end of commit.
935 * @c: UBIFS file-system description object
936 *
937 * This function completes out-of-place garbage collection of index LEBs.
938 */
939int ubifs_gc_end_commit(struct ubifs_info *c)
940{
941 struct ubifs_gced_idx_leb *idx_gc, *tmp;
942 struct ubifs_wbuf *wbuf;
943 int err = 0;
944
945 wbuf = &c->jheads[GCHD].wbuf;
946 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
947 list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
948 if (idx_gc->unmap) {
949 dbg_gc("LEB %d", idx_gc->lnum);
950 err = ubifs_leb_unmap(c, idx_gc->lnum);
951 if (err)
952 goto out;
953 err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
954 LPROPS_NC, 0, LPROPS_TAKEN, -1);
955 if (err)
956 goto out;
957 list_del(&idx_gc->list);
958 kfree(idx_gc);
959 }
960out:
961 mutex_unlock(&wbuf->io_mutex);
962 return err;
963}
964
965/**
966 * ubifs_destroy_idx_gc - destroy idx_gc list.
967 * @c: UBIFS file-system description object
968 *
969 * This function destroys the @c->idx_gc list. It is called when unmounting
970 * so locks are not needed. Returns zero in case of success and a negative
971 * error code in case of failure.
972 */
973void ubifs_destroy_idx_gc(struct ubifs_info *c)
974{
975 while (!list_empty(&c->idx_gc)) {
976 struct ubifs_gced_idx_leb *idx_gc;
977
978 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
979 list);
980 c->idx_gc_cnt -= 1;
981 list_del(&idx_gc->list);
982 kfree(idx_gc);
983 }
984}
985
986/**
987 * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
988 * @c: UBIFS file-system description object
989 *
990 * Called during start commit so locks are not needed.
991 */
992int ubifs_get_idx_gc_leb(struct ubifs_info *c)
993{
994 struct ubifs_gced_idx_leb *idx_gc;
995 int lnum;
996
997 if (list_empty(&c->idx_gc))
998 return -ENOSPC;
999 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
1000 lnum = idx_gc->lnum;
1001 /* c->idx_gc_cnt is updated by the caller when lprops are updated */
1002 list_del(&idx_gc->list);
1003 kfree(idx_gc);
1004 return lnum;
1005}
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 * 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
50 * which are guaranteed to be reclaimable. If LEB has less space, the GC might
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.
54 */
55
56#include <linux/slab.h>
57#include <linux/pagemap.h>
58#include <linux/list_sort.h>
59#include "ubifs.h"
60
61/*
62 * GC may need to move more than one LEB to make progress. The below constants
63 * define "soft" and "hard" limits on the number of LEBs the garbage collector
64 * may move.
65 */
66#define SOFT_LEBS_LIMIT 4
67#define HARD_LEBS_LIMIT 32
68
69/**
70 * switch_gc_head - switch the garbage collection journal head.
71 * @c: UBIFS file-system description object
72 * @buf: buffer to write
73 * @len: length of the buffer to write
74 * @lnum: LEB number written is returned here
75 * @offs: offset written is returned here
76 *
77 * This function switch the GC head to the next LEB which is reserved in
78 * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required,
79 * and other negative error code in case of failures.
80 */
81static int switch_gc_head(struct ubifs_info *c)
82{
83 int err, gc_lnum = c->gc_lnum;
84 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
85
86 ubifs_assert(gc_lnum != -1);
87 dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)",
88 wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum,
89 c->leb_size - wbuf->offs - wbuf->used);
90
91 err = ubifs_wbuf_sync_nolock(wbuf);
92 if (err)
93 return err;
94
95 /*
96 * The GC write-buffer was synchronized, we may safely unmap
97 * 'c->gc_lnum'.
98 */
99 err = ubifs_leb_unmap(c, gc_lnum);
100 if (err)
101 return err;
102
103 err = ubifs_wbuf_sync_nolock(wbuf);
104 if (err)
105 return err;
106
107 err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0);
108 if (err)
109 return err;
110
111 c->gc_lnum = -1;
112 err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0, UBI_LONGTERM);
113 return err;
114}
115
116/**
117 * data_nodes_cmp - compare 2 data nodes.
118 * @priv: UBIFS file-system description object
119 * @a: first data node
120 * @a: second data node
121 *
122 * This function compares data nodes @a and @b. Returns %1 if @a has greater
123 * inode or block number, and %-1 otherwise.
124 */
125static int data_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
126{
127 ino_t inuma, inumb;
128 struct ubifs_info *c = priv;
129 struct ubifs_scan_node *sa, *sb;
130
131 cond_resched();
132 if (a == b)
133 return 0;
134
135 sa = list_entry(a, struct ubifs_scan_node, list);
136 sb = list_entry(b, struct ubifs_scan_node, list);
137
138 ubifs_assert(key_type(c, &sa->key) == UBIFS_DATA_KEY);
139 ubifs_assert(key_type(c, &sb->key) == UBIFS_DATA_KEY);
140 ubifs_assert(sa->type == UBIFS_DATA_NODE);
141 ubifs_assert(sb->type == UBIFS_DATA_NODE);
142
143 inuma = key_inum(c, &sa->key);
144 inumb = key_inum(c, &sb->key);
145
146 if (inuma == inumb) {
147 unsigned int blka = key_block(c, &sa->key);
148 unsigned int blkb = key_block(c, &sb->key);
149
150 if (blka <= blkb)
151 return -1;
152 } else if (inuma <= inumb)
153 return -1;
154
155 return 1;
156}
157
158/*
159 * nondata_nodes_cmp - compare 2 non-data nodes.
160 * @priv: UBIFS file-system description object
161 * @a: first node
162 * @a: second node
163 *
164 * This function compares nodes @a and @b. It makes sure that inode nodes go
165 * first and sorted by length in descending order. Directory entry nodes go
166 * after inode nodes and are sorted in ascending hash valuer order.
167 */
168static int nondata_nodes_cmp(void *priv, struct list_head *a,
169 struct list_head *b)
170{
171 ino_t inuma, inumb;
172 struct ubifs_info *c = priv;
173 struct ubifs_scan_node *sa, *sb;
174
175 cond_resched();
176 if (a == b)
177 return 0;
178
179 sa = list_entry(a, struct ubifs_scan_node, list);
180 sb = list_entry(b, struct ubifs_scan_node, list);
181
182 ubifs_assert(key_type(c, &sa->key) != UBIFS_DATA_KEY &&
183 key_type(c, &sb->key) != UBIFS_DATA_KEY);
184 ubifs_assert(sa->type != UBIFS_DATA_NODE &&
185 sb->type != UBIFS_DATA_NODE);
186
187 /* Inodes go before directory entries */
188 if (sa->type == UBIFS_INO_NODE) {
189 if (sb->type == UBIFS_INO_NODE)
190 return sb->len - sa->len;
191 return -1;
192 }
193 if (sb->type == UBIFS_INO_NODE)
194 return 1;
195
196 ubifs_assert(key_type(c, &sa->key) == UBIFS_DENT_KEY ||
197 key_type(c, &sa->key) == UBIFS_XENT_KEY);
198 ubifs_assert(key_type(c, &sb->key) == UBIFS_DENT_KEY ||
199 key_type(c, &sb->key) == UBIFS_XENT_KEY);
200 ubifs_assert(sa->type == UBIFS_DENT_NODE ||
201 sa->type == UBIFS_XENT_NODE);
202 ubifs_assert(sb->type == UBIFS_DENT_NODE ||
203 sb->type == UBIFS_XENT_NODE);
204
205 inuma = key_inum(c, &sa->key);
206 inumb = key_inum(c, &sb->key);
207
208 if (inuma == inumb) {
209 uint32_t hasha = key_hash(c, &sa->key);
210 uint32_t hashb = key_hash(c, &sb->key);
211
212 if (hasha <= hashb)
213 return -1;
214 } else if (inuma <= inumb)
215 return -1;
216
217 return 1;
218}
219
220/**
221 * sort_nodes - sort nodes for GC.
222 * @c: UBIFS file-system description object
223 * @sleb: describes nodes to sort and contains the result on exit
224 * @nondata: contains non-data nodes on exit
225 * @min: minimum node size is returned here
226 *
227 * This function sorts the list of inodes to garbage collect. First of all, it
228 * kills obsolete nodes and separates data and non-data nodes to the
229 * @sleb->nodes and @nondata lists correspondingly.
230 *
231 * Data nodes are then sorted in block number order - this is important for
232 * bulk-read; data nodes with lower inode number go before data nodes with
233 * higher inode number, and data nodes with lower block number go before data
234 * nodes with higher block number;
235 *
236 * Non-data nodes are sorted as follows.
237 * o First go inode nodes - they are sorted in descending length order.
238 * o Then go directory entry nodes - they are sorted in hash order, which
239 * should supposedly optimize 'readdir()'. Direntry nodes with lower parent
240 * inode number go before direntry nodes with higher parent inode number,
241 * and direntry nodes with lower name hash values go before direntry nodes
242 * with higher name hash values.
243 *
244 * This function returns zero in case of success and a negative error code in
245 * case of failure.
246 */
247static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
248 struct list_head *nondata, int *min)
249{
250 int err;
251 struct ubifs_scan_node *snod, *tmp;
252
253 *min = INT_MAX;
254
255 /* Separate data nodes and non-data nodes */
256 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
257 ubifs_assert(snod->type == UBIFS_INO_NODE ||
258 snod->type == UBIFS_DATA_NODE ||
259 snod->type == UBIFS_DENT_NODE ||
260 snod->type == UBIFS_XENT_NODE ||
261 snod->type == UBIFS_TRUN_NODE);
262
263 if (snod->type != UBIFS_INO_NODE &&
264 snod->type != UBIFS_DATA_NODE &&
265 snod->type != UBIFS_DENT_NODE &&
266 snod->type != UBIFS_XENT_NODE) {
267 /* Probably truncation node, zap it */
268 list_del(&snod->list);
269 kfree(snod);
270 continue;
271 }
272
273 ubifs_assert(key_type(c, &snod->key) == UBIFS_DATA_KEY ||
274 key_type(c, &snod->key) == UBIFS_INO_KEY ||
275 key_type(c, &snod->key) == UBIFS_DENT_KEY ||
276 key_type(c, &snod->key) == UBIFS_XENT_KEY);
277
278 err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
279 snod->offs, 0);
280 if (err < 0)
281 return err;
282
283 if (!err) {
284 /* The node is obsolete, remove it from the list */
285 list_del(&snod->list);
286 kfree(snod);
287 continue;
288 }
289
290 if (snod->len < *min)
291 *min = snod->len;
292
293 if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
294 list_move_tail(&snod->list, nondata);
295 }
296
297 /* Sort data and non-data nodes */
298 list_sort(c, &sleb->nodes, &data_nodes_cmp);
299 list_sort(c, nondata, &nondata_nodes_cmp);
300
301 err = dbg_check_data_nodes_order(c, &sleb->nodes);
302 if (err)
303 return err;
304 err = dbg_check_nondata_nodes_order(c, nondata);
305 if (err)
306 return err;
307 return 0;
308}
309
310/**
311 * move_node - move a node.
312 * @c: UBIFS file-system description object
313 * @sleb: describes the LEB to move nodes from
314 * @snod: the mode to move
315 * @wbuf: write-buffer to move node to
316 *
317 * This function moves node @snod to @wbuf, changes TNC correspondingly, and
318 * destroys @snod. Returns zero in case of success and a negative error code in
319 * case of failure.
320 */
321static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
322 struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
323{
324 int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
325
326 cond_resched();
327 err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
328 if (err)
329 return err;
330
331 err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
332 snod->offs, new_lnum, new_offs,
333 snod->len);
334 list_del(&snod->list);
335 kfree(snod);
336 return err;
337}
338
339/**
340 * move_nodes - move nodes.
341 * @c: UBIFS file-system description object
342 * @sleb: describes the LEB to move nodes from
343 *
344 * This function moves valid nodes from data LEB described by @sleb to the GC
345 * journal head. This function returns zero in case of success, %-EAGAIN if
346 * commit is required, and other negative error codes in case of other
347 * failures.
348 */
349static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
350{
351 int err, min;
352 LIST_HEAD(nondata);
353 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
354
355 if (wbuf->lnum == -1) {
356 /*
357 * The GC journal head is not set, because it is the first GC
358 * invocation since mount.
359 */
360 err = switch_gc_head(c);
361 if (err)
362 return err;
363 }
364
365 err = sort_nodes(c, sleb, &nondata, &min);
366 if (err)
367 goto out;
368
369 /* Write nodes to their new location. Use the first-fit strategy */
370 while (1) {
371 int avail;
372 struct ubifs_scan_node *snod, *tmp;
373
374 /* Move data nodes */
375 list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
376 avail = c->leb_size - wbuf->offs - wbuf->used;
377 if (snod->len > avail)
378 /*
379 * Do not skip data nodes in order to optimize
380 * bulk-read.
381 */
382 break;
383
384 err = move_node(c, sleb, snod, wbuf);
385 if (err)
386 goto out;
387 }
388
389 /* Move non-data nodes */
390 list_for_each_entry_safe(snod, tmp, &nondata, list) {
391 avail = c->leb_size - wbuf->offs - wbuf->used;
392 if (avail < min)
393 break;
394
395 if (snod->len > avail) {
396 /*
397 * Keep going only if this is an inode with
398 * some data. Otherwise stop and switch the GC
399 * head. IOW, we assume that data-less inode
400 * nodes and direntry nodes are roughly of the
401 * same size.
402 */
403 if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
404 snod->len == UBIFS_INO_NODE_SZ)
405 break;
406 continue;
407 }
408
409 err = move_node(c, sleb, snod, wbuf);
410 if (err)
411 goto out;
412 }
413
414 if (list_empty(&sleb->nodes) && list_empty(&nondata))
415 break;
416
417 /*
418 * Waste the rest of the space in the LEB and switch to the
419 * next LEB.
420 */
421 err = switch_gc_head(c);
422 if (err)
423 goto out;
424 }
425
426 return 0;
427
428out:
429 list_splice_tail(&nondata, &sleb->nodes);
430 return err;
431}
432
433/**
434 * gc_sync_wbufs - sync write-buffers for GC.
435 * @c: UBIFS file-system description object
436 *
437 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
438 * be in a write-buffer instead. That is, a node could be written to a
439 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
440 * erased before the write-buffer is sync'd and then there is an unclean
441 * unmount, then an existing node is lost. To avoid this, we sync all
442 * write-buffers.
443 *
444 * This function returns %0 on success or a negative error code on failure.
445 */
446static int gc_sync_wbufs(struct ubifs_info *c)
447{
448 int err, i;
449
450 for (i = 0; i < c->jhead_cnt; i++) {
451 if (i == GCHD)
452 continue;
453 err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
454 if (err)
455 return err;
456 }
457 return 0;
458}
459
460/**
461 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
462 * @c: UBIFS file-system description object
463 * @lp: describes the LEB to garbage collect
464 *
465 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
466 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
467 * required, and other negative error codes in case of failures.
468 */
469int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
470{
471 struct ubifs_scan_leb *sleb;
472 struct ubifs_scan_node *snod;
473 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
474 int err = 0, lnum = lp->lnum;
475
476 ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
477 c->need_recovery);
478 ubifs_assert(c->gc_lnum != lnum);
479 ubifs_assert(wbuf->lnum != lnum);
480
481 if (lp->free + lp->dirty == c->leb_size) {
482 /* Special case - a free LEB */
483 dbg_gc("LEB %d is free, return it", lp->lnum);
484 ubifs_assert(!(lp->flags & LPROPS_INDEX));
485
486 if (lp->free != c->leb_size) {
487 /*
488 * Write buffers must be sync'd before unmapping
489 * freeable LEBs, because one of them may contain data
490 * which obsoletes something in 'lp->pnum'.
491 */
492 err = gc_sync_wbufs(c);
493 if (err)
494 return err;
495 err = ubifs_change_one_lp(c, lp->lnum, c->leb_size,
496 0, 0, 0, 0);
497 if (err)
498 return err;
499 }
500 err = ubifs_leb_unmap(c, lp->lnum);
501 if (err)
502 return err;
503
504 if (c->gc_lnum == -1) {
505 c->gc_lnum = lnum;
506 return LEB_RETAINED;
507 }
508
509 return LEB_FREED;
510 }
511
512 /*
513 * We scan the entire LEB even though we only really need to scan up to
514 * (c->leb_size - lp->free).
515 */
516 sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
517 if (IS_ERR(sleb))
518 return PTR_ERR(sleb);
519
520 ubifs_assert(!list_empty(&sleb->nodes));
521 snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
522
523 if (snod->type == UBIFS_IDX_NODE) {
524 struct ubifs_gced_idx_leb *idx_gc;
525
526 dbg_gc("indexing LEB %d (free %d, dirty %d)",
527 lnum, lp->free, lp->dirty);
528 list_for_each_entry(snod, &sleb->nodes, list) {
529 struct ubifs_idx_node *idx = snod->node;
530 int level = le16_to_cpu(idx->level);
531
532 ubifs_assert(snod->type == UBIFS_IDX_NODE);
533 key_read(c, ubifs_idx_key(c, idx), &snod->key);
534 err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
535 snod->offs);
536 if (err)
537 goto out;
538 }
539
540 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
541 if (!idx_gc) {
542 err = -ENOMEM;
543 goto out;
544 }
545
546 idx_gc->lnum = lnum;
547 idx_gc->unmap = 0;
548 list_add(&idx_gc->list, &c->idx_gc);
549
550 /*
551 * Don't release the LEB until after the next commit, because
552 * it may contain data which is needed for recovery. So
553 * although we freed this LEB, it will become usable only after
554 * the commit.
555 */
556 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
557 LPROPS_INDEX, 1);
558 if (err)
559 goto out;
560 err = LEB_FREED_IDX;
561 } else {
562 dbg_gc("data LEB %d (free %d, dirty %d)",
563 lnum, lp->free, lp->dirty);
564
565 err = move_nodes(c, sleb);
566 if (err)
567 goto out_inc_seq;
568
569 err = gc_sync_wbufs(c);
570 if (err)
571 goto out_inc_seq;
572
573 err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
574 if (err)
575 goto out_inc_seq;
576
577 /* Allow for races with TNC */
578 c->gced_lnum = lnum;
579 smp_wmb();
580 c->gc_seq += 1;
581 smp_wmb();
582
583 if (c->gc_lnum == -1) {
584 c->gc_lnum = lnum;
585 err = LEB_RETAINED;
586 } else {
587 err = ubifs_wbuf_sync_nolock(wbuf);
588 if (err)
589 goto out;
590
591 err = ubifs_leb_unmap(c, lnum);
592 if (err)
593 goto out;
594
595 err = LEB_FREED;
596 }
597 }
598
599out:
600 ubifs_scan_destroy(sleb);
601 return err;
602
603out_inc_seq:
604 /* We may have moved at least some nodes so allow for races with TNC */
605 c->gced_lnum = lnum;
606 smp_wmb();
607 c->gc_seq += 1;
608 smp_wmb();
609 goto out;
610}
611
612/**
613 * ubifs_garbage_collect - UBIFS garbage collector.
614 * @c: UBIFS file-system description object
615 * @anyway: do GC even if there are free LEBs
616 *
617 * This function does out-of-place garbage collection. The return codes are:
618 * o positive LEB number if the LEB has been freed and may be used;
619 * o %-EAGAIN if the caller has to run commit;
620 * o %-ENOSPC if GC failed to make any progress;
621 * o other negative error codes in case of other errors.
622 *
623 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
624 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
625 * commit may be required. But commit cannot be run from inside GC, because the
626 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
627 * And this error code means that the caller has to run commit, and re-run GC
628 * if there is still no free space.
629 *
630 * There are many reasons why this function may return %-EAGAIN:
631 * o the log is full and there is no space to write an LEB reference for
632 * @c->gc_lnum;
633 * o the journal is too large and exceeds size limitations;
634 * o GC moved indexing LEBs, but they can be used only after the commit;
635 * o the shrinker fails to find clean znodes to free and requests the commit;
636 * o etc.
637 *
638 * Note, if the file-system is close to be full, this function may return
639 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
640 * the function. E.g., this happens if the limits on the journal size are too
641 * tough and GC writes too much to the journal before an LEB is freed. This
642 * might also mean that the journal is too large, and the TNC becomes to big,
643 * so that the shrinker is constantly called, finds not clean znodes to free,
644 * and requests commit. Well, this may also happen if the journal is all right,
645 * but another kernel process consumes too much memory. Anyway, infinite
646 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
647 */
648int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
649{
650 int i, err, ret, min_space = c->dead_wm;
651 struct ubifs_lprops lp;
652 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
653
654 ubifs_assert_cmt_locked(c);
655 ubifs_assert(!c->ro_media && !c->ro_mount);
656
657 if (ubifs_gc_should_commit(c))
658 return -EAGAIN;
659
660 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
661
662 if (c->ro_error) {
663 ret = -EROFS;
664 goto out_unlock;
665 }
666
667 /* We expect the write-buffer to be empty on entry */
668 ubifs_assert(!wbuf->used);
669
670 for (i = 0; ; i++) {
671 int space_before = c->leb_size - wbuf->offs - wbuf->used;
672 int space_after;
673
674 cond_resched();
675
676 /* Give the commit an opportunity to run */
677 if (ubifs_gc_should_commit(c)) {
678 ret = -EAGAIN;
679 break;
680 }
681
682 if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
683 /*
684 * We've done enough iterations. Indexing LEBs were
685 * moved and will be available after the commit.
686 */
687 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
688 ubifs_commit_required(c);
689 ret = -EAGAIN;
690 break;
691 }
692
693 if (i > HARD_LEBS_LIMIT) {
694 /*
695 * We've moved too many LEBs and have not made
696 * progress, give up.
697 */
698 dbg_gc("hard limit, -ENOSPC");
699 ret = -ENOSPC;
700 break;
701 }
702
703 /*
704 * Empty and freeable LEBs can turn up while we waited for
705 * the wbuf lock, or while we have been running GC. In that
706 * case, we should just return one of those instead of
707 * continuing to GC dirty LEBs. Hence we request
708 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
709 */
710 ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
711 if (ret) {
712 if (ret == -ENOSPC)
713 dbg_gc("no more dirty LEBs");
714 break;
715 }
716
717 dbg_gc("found LEB %d: free %d, dirty %d, sum %d "
718 "(min. space %d)", lp.lnum, lp.free, lp.dirty,
719 lp.free + lp.dirty, min_space);
720
721 space_before = c->leb_size - wbuf->offs - wbuf->used;
722 if (wbuf->lnum == -1)
723 space_before = 0;
724
725 ret = ubifs_garbage_collect_leb(c, &lp);
726 if (ret < 0) {
727 if (ret == -EAGAIN) {
728 /*
729 * This is not error, so we have to return the
730 * LEB to lprops. But if 'ubifs_return_leb()'
731 * fails, its failure code is propagated to the
732 * caller instead of the original '-EAGAIN'.
733 */
734 err = ubifs_return_leb(c, lp.lnum);
735 if (err)
736 ret = err;
737 break;
738 }
739 goto out;
740 }
741
742 if (ret == LEB_FREED) {
743 /* An LEB has been freed and is ready for use */
744 dbg_gc("LEB %d freed, return", lp.lnum);
745 ret = lp.lnum;
746 break;
747 }
748
749 if (ret == LEB_FREED_IDX) {
750 /*
751 * This was an indexing LEB and it cannot be
752 * immediately used. And instead of requesting the
753 * commit straight away, we try to garbage collect some
754 * more.
755 */
756 dbg_gc("indexing LEB %d freed, continue", lp.lnum);
757 continue;
758 }
759
760 ubifs_assert(ret == LEB_RETAINED);
761 space_after = c->leb_size - wbuf->offs - wbuf->used;
762 dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
763 space_after - space_before);
764
765 if (space_after > space_before) {
766 /* GC makes progress, keep working */
767 min_space >>= 1;
768 if (min_space < c->dead_wm)
769 min_space = c->dead_wm;
770 continue;
771 }
772
773 dbg_gc("did not make progress");
774
775 /*
776 * GC moved an LEB bud have not done any progress. This means
777 * that the previous GC head LEB contained too few free space
778 * and the LEB which was GC'ed contained only large nodes which
779 * did not fit that space.
780 *
781 * We can do 2 things:
782 * 1. pick another LEB in a hope it'll contain a small node
783 * which will fit the space we have at the end of current GC
784 * head LEB, but there is no guarantee, so we try this out
785 * unless we have already been working for too long;
786 * 2. request an LEB with more dirty space, which will force
787 * 'ubifs_find_dirty_leb()' to start scanning the lprops
788 * table, instead of just picking one from the heap
789 * (previously it already picked the dirtiest LEB).
790 */
791 if (i < SOFT_LEBS_LIMIT) {
792 dbg_gc("try again");
793 continue;
794 }
795
796 min_space <<= 1;
797 if (min_space > c->dark_wm)
798 min_space = c->dark_wm;
799 dbg_gc("set min. space to %d", min_space);
800 }
801
802 if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
803 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
804 ubifs_commit_required(c);
805 ret = -EAGAIN;
806 }
807
808 err = ubifs_wbuf_sync_nolock(wbuf);
809 if (!err)
810 err = ubifs_leb_unmap(c, c->gc_lnum);
811 if (err) {
812 ret = err;
813 goto out;
814 }
815out_unlock:
816 mutex_unlock(&wbuf->io_mutex);
817 return ret;
818
819out:
820 ubifs_assert(ret < 0);
821 ubifs_assert(ret != -ENOSPC && ret != -EAGAIN);
822 ubifs_wbuf_sync_nolock(wbuf);
823 ubifs_ro_mode(c, ret);
824 mutex_unlock(&wbuf->io_mutex);
825 ubifs_return_leb(c, lp.lnum);
826 return ret;
827}
828
829/**
830 * ubifs_gc_start_commit - garbage collection at start of commit.
831 * @c: UBIFS file-system description object
832 *
833 * If a LEB has only dirty and free space, then we may safely unmap it and make
834 * it free. Note, we cannot do this with indexing LEBs because dirty space may
835 * correspond index nodes that are required for recovery. In that case, the
836 * LEB cannot be unmapped until after the next commit.
837 *
838 * This function returns %0 upon success and a negative error code upon failure.
839 */
840int ubifs_gc_start_commit(struct ubifs_info *c)
841{
842 struct ubifs_gced_idx_leb *idx_gc;
843 const struct ubifs_lprops *lp;
844 int err = 0, flags;
845
846 ubifs_get_lprops(c);
847
848 /*
849 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
850 * wbufs are sync'd before this, which is done in 'do_commit()'.
851 */
852 while (1) {
853 lp = ubifs_fast_find_freeable(c);
854 if (IS_ERR(lp)) {
855 err = PTR_ERR(lp);
856 goto out;
857 }
858 if (!lp)
859 break;
860 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
861 ubifs_assert(!(lp->flags & LPROPS_INDEX));
862 err = ubifs_leb_unmap(c, lp->lnum);
863 if (err)
864 goto out;
865 lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
866 if (IS_ERR(lp)) {
867 err = PTR_ERR(lp);
868 goto out;
869 }
870 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
871 ubifs_assert(!(lp->flags & LPROPS_INDEX));
872 }
873
874 /* Mark GC'd index LEBs OK to unmap after this commit finishes */
875 list_for_each_entry(idx_gc, &c->idx_gc, list)
876 idx_gc->unmap = 1;
877
878 /* Record index freeable LEBs for unmapping after commit */
879 while (1) {
880 lp = ubifs_fast_find_frdi_idx(c);
881 if (IS_ERR(lp)) {
882 err = PTR_ERR(lp);
883 goto out;
884 }
885 if (!lp)
886 break;
887 idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
888 if (!idx_gc) {
889 err = -ENOMEM;
890 goto out;
891 }
892 ubifs_assert(!(lp->flags & LPROPS_TAKEN));
893 ubifs_assert(lp->flags & LPROPS_INDEX);
894 /* Don't release the LEB until after the next commit */
895 flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
896 lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
897 if (IS_ERR(lp)) {
898 err = PTR_ERR(lp);
899 kfree(idx_gc);
900 goto out;
901 }
902 ubifs_assert(lp->flags & LPROPS_TAKEN);
903 ubifs_assert(!(lp->flags & LPROPS_INDEX));
904 idx_gc->lnum = lp->lnum;
905 idx_gc->unmap = 1;
906 list_add(&idx_gc->list, &c->idx_gc);
907 }
908out:
909 ubifs_release_lprops(c);
910 return err;
911}
912
913/**
914 * ubifs_gc_end_commit - garbage collection at end of commit.
915 * @c: UBIFS file-system description object
916 *
917 * This function completes out-of-place garbage collection of index LEBs.
918 */
919int ubifs_gc_end_commit(struct ubifs_info *c)
920{
921 struct ubifs_gced_idx_leb *idx_gc, *tmp;
922 struct ubifs_wbuf *wbuf;
923 int err = 0;
924
925 wbuf = &c->jheads[GCHD].wbuf;
926 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
927 list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
928 if (idx_gc->unmap) {
929 dbg_gc("LEB %d", idx_gc->lnum);
930 err = ubifs_leb_unmap(c, idx_gc->lnum);
931 if (err)
932 goto out;
933 err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
934 LPROPS_NC, 0, LPROPS_TAKEN, -1);
935 if (err)
936 goto out;
937 list_del(&idx_gc->list);
938 kfree(idx_gc);
939 }
940out:
941 mutex_unlock(&wbuf->io_mutex);
942 return err;
943}
944
945/**
946 * ubifs_destroy_idx_gc - destroy idx_gc list.
947 * @c: UBIFS file-system description object
948 *
949 * This function destroys the @c->idx_gc list. It is called when unmounting
950 * so locks are not needed. Returns zero in case of success and a negative
951 * error code in case of failure.
952 */
953void ubifs_destroy_idx_gc(struct ubifs_info *c)
954{
955 while (!list_empty(&c->idx_gc)) {
956 struct ubifs_gced_idx_leb *idx_gc;
957
958 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
959 list);
960 c->idx_gc_cnt -= 1;
961 list_del(&idx_gc->list);
962 kfree(idx_gc);
963 }
964}
965
966/**
967 * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
968 * @c: UBIFS file-system description object
969 *
970 * Called during start commit so locks are not needed.
971 */
972int ubifs_get_idx_gc_leb(struct ubifs_info *c)
973{
974 struct ubifs_gced_idx_leb *idx_gc;
975 int lnum;
976
977 if (list_empty(&c->idx_gc))
978 return -ENOSPC;
979 idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
980 lnum = idx_gc->lnum;
981 /* c->idx_gc_cnt is updated by the caller when lprops are updated */
982 list_del(&idx_gc->list);
983 kfree(idx_gc);
984 return lnum;
985}