xref: /openbmc/u-boot/fs/ubifs/gc.c (revision 2a046ff5)
1 // SPDX-License-Identifier: GPL-2.0
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 few. The other LEBs are not so
41  * good, and GC takes extra care when moving them.
42  */
43 #ifndef __UBOOT__
44 #include <linux/slab.h>
45 #include <linux/pagemap.h>
46 #include <linux/list_sort.h>
47 #endif
48 #include "ubifs.h"
49 
50 #ifndef __UBOOT__
51 /*
52  * GC may need to move more than one LEB to make progress. The below constants
53  * define "soft" and "hard" limits on the number of LEBs the garbage collector
54  * may move.
55  */
56 #define SOFT_LEBS_LIMIT 4
57 #define HARD_LEBS_LIMIT 32
58 
59 /**
60  * switch_gc_head - switch the garbage collection journal head.
61  * @c: UBIFS file-system description object
62  * @buf: buffer to write
63  * @len: length of the buffer to write
64  * @lnum: LEB number written is returned here
65  * @offs: offset written is returned here
66  *
67  * This function switch the GC head to the next LEB which is reserved in
68  * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required,
69  * and other negative error code in case of failures.
70  */
71 static int switch_gc_head(struct ubifs_info *c)
72 {
73 	int err, gc_lnum = c->gc_lnum;
74 	struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
75 
76 	ubifs_assert(gc_lnum != -1);
77 	dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)",
78 	       wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum,
79 	       c->leb_size - wbuf->offs - wbuf->used);
80 
81 	err = ubifs_wbuf_sync_nolock(wbuf);
82 	if (err)
83 		return err;
84 
85 	/*
86 	 * The GC write-buffer was synchronized, we may safely unmap
87 	 * 'c->gc_lnum'.
88 	 */
89 	err = ubifs_leb_unmap(c, gc_lnum);
90 	if (err)
91 		return err;
92 
93 	err = ubifs_wbuf_sync_nolock(wbuf);
94 	if (err)
95 		return err;
96 
97 	err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0);
98 	if (err)
99 		return err;
100 
101 	c->gc_lnum = -1;
102 	err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0);
103 	return err;
104 }
105 
106 /**
107  * data_nodes_cmp - compare 2 data nodes.
108  * @priv: UBIFS file-system description object
109  * @a: first data node
110  * @a: second data node
111  *
112  * This function compares data nodes @a and @b. Returns %1 if @a has greater
113  * inode or block number, and %-1 otherwise.
114  */
115 static int data_nodes_cmp(void *priv, struct list_head *a, struct list_head *b)
116 {
117 	ino_t inuma, inumb;
118 	struct ubifs_info *c = priv;
119 	struct ubifs_scan_node *sa, *sb;
120 
121 	cond_resched();
122 	if (a == b)
123 		return 0;
124 
125 	sa = list_entry(a, struct ubifs_scan_node, list);
126 	sb = list_entry(b, struct ubifs_scan_node, list);
127 
128 	ubifs_assert(key_type(c, &sa->key) == UBIFS_DATA_KEY);
129 	ubifs_assert(key_type(c, &sb->key) == UBIFS_DATA_KEY);
130 	ubifs_assert(sa->type == UBIFS_DATA_NODE);
131 	ubifs_assert(sb->type == UBIFS_DATA_NODE);
132 
133 	inuma = key_inum(c, &sa->key);
134 	inumb = key_inum(c, &sb->key);
135 
136 	if (inuma == inumb) {
137 		unsigned int blka = key_block(c, &sa->key);
138 		unsigned int blkb = key_block(c, &sb->key);
139 
140 		if (blka <= blkb)
141 			return -1;
142 	} else if (inuma <= inumb)
143 		return -1;
144 
145 	return 1;
146 }
147 
148 /*
149  * nondata_nodes_cmp - compare 2 non-data nodes.
150  * @priv: UBIFS file-system description object
151  * @a: first node
152  * @a: second node
153  *
154  * This function compares nodes @a and @b. It makes sure that inode nodes go
155  * first and sorted by length in descending order. Directory entry nodes go
156  * after inode nodes and are sorted in ascending hash valuer order.
157  */
158 static int nondata_nodes_cmp(void *priv, struct list_head *a,
159 			     struct list_head *b)
160 {
161 	ino_t inuma, inumb;
162 	struct ubifs_info *c = priv;
163 	struct ubifs_scan_node *sa, *sb;
164 
165 	cond_resched();
166 	if (a == b)
167 		return 0;
168 
169 	sa = list_entry(a, struct ubifs_scan_node, list);
170 	sb = list_entry(b, struct ubifs_scan_node, list);
171 
172 	ubifs_assert(key_type(c, &sa->key) != UBIFS_DATA_KEY &&
173 		     key_type(c, &sb->key) != UBIFS_DATA_KEY);
174 	ubifs_assert(sa->type != UBIFS_DATA_NODE &&
175 		     sb->type != UBIFS_DATA_NODE);
176 
177 	/* Inodes go before directory entries */
178 	if (sa->type == UBIFS_INO_NODE) {
179 		if (sb->type == UBIFS_INO_NODE)
180 			return sb->len - sa->len;
181 		return -1;
182 	}
183 	if (sb->type == UBIFS_INO_NODE)
184 		return 1;
185 
186 	ubifs_assert(key_type(c, &sa->key) == UBIFS_DENT_KEY ||
187 		     key_type(c, &sa->key) == UBIFS_XENT_KEY);
188 	ubifs_assert(key_type(c, &sb->key) == UBIFS_DENT_KEY ||
189 		     key_type(c, &sb->key) == UBIFS_XENT_KEY);
190 	ubifs_assert(sa->type == UBIFS_DENT_NODE ||
191 		     sa->type == UBIFS_XENT_NODE);
192 	ubifs_assert(sb->type == UBIFS_DENT_NODE ||
193 		     sb->type == UBIFS_XENT_NODE);
194 
195 	inuma = key_inum(c, &sa->key);
196 	inumb = key_inum(c, &sb->key);
197 
198 	if (inuma == inumb) {
199 		uint32_t hasha = key_hash(c, &sa->key);
200 		uint32_t hashb = key_hash(c, &sb->key);
201 
202 		if (hasha <= hashb)
203 			return -1;
204 	} else if (inuma <= inumb)
205 		return -1;
206 
207 	return 1;
208 }
209 
210 /**
211  * sort_nodes - sort nodes for GC.
212  * @c: UBIFS file-system description object
213  * @sleb: describes nodes to sort and contains the result on exit
214  * @nondata: contains non-data nodes on exit
215  * @min: minimum node size is returned here
216  *
217  * This function sorts the list of inodes to garbage collect. First of all, it
218  * kills obsolete nodes and separates data and non-data nodes to the
219  * @sleb->nodes and @nondata lists correspondingly.
220  *
221  * Data nodes are then sorted in block number order - this is important for
222  * bulk-read; data nodes with lower inode number go before data nodes with
223  * higher inode number, and data nodes with lower block number go before data
224  * nodes with higher block number;
225  *
226  * Non-data nodes are sorted as follows.
227  *   o First go inode nodes - they are sorted in descending length order.
228  *   o Then go directory entry nodes - they are sorted in hash order, which
229  *     should supposedly optimize 'readdir()'. Direntry nodes with lower parent
230  *     inode number go before direntry nodes with higher parent inode number,
231  *     and direntry nodes with lower name hash values go before direntry nodes
232  *     with higher name hash values.
233  *
234  * This function returns zero in case of success and a negative error code in
235  * case of failure.
236  */
237 static int sort_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
238 		      struct list_head *nondata, int *min)
239 {
240 	int err;
241 	struct ubifs_scan_node *snod, *tmp;
242 
243 	*min = INT_MAX;
244 
245 	/* Separate data nodes and non-data nodes */
246 	list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
247 		ubifs_assert(snod->type == UBIFS_INO_NODE  ||
248 			     snod->type == UBIFS_DATA_NODE ||
249 			     snod->type == UBIFS_DENT_NODE ||
250 			     snod->type == UBIFS_XENT_NODE ||
251 			     snod->type == UBIFS_TRUN_NODE);
252 
253 		if (snod->type != UBIFS_INO_NODE  &&
254 		    snod->type != UBIFS_DATA_NODE &&
255 		    snod->type != UBIFS_DENT_NODE &&
256 		    snod->type != UBIFS_XENT_NODE) {
257 			/* Probably truncation node, zap it */
258 			list_del(&snod->list);
259 			kfree(snod);
260 			continue;
261 		}
262 
263 		ubifs_assert(key_type(c, &snod->key) == UBIFS_DATA_KEY ||
264 			     key_type(c, &snod->key) == UBIFS_INO_KEY  ||
265 			     key_type(c, &snod->key) == UBIFS_DENT_KEY ||
266 			     key_type(c, &snod->key) == UBIFS_XENT_KEY);
267 
268 		err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum,
269 					 snod->offs, 0);
270 		if (err < 0)
271 			return err;
272 
273 		if (!err) {
274 			/* The node is obsolete, remove it from the list */
275 			list_del(&snod->list);
276 			kfree(snod);
277 			continue;
278 		}
279 
280 		if (snod->len < *min)
281 			*min = snod->len;
282 
283 		if (key_type(c, &snod->key) != UBIFS_DATA_KEY)
284 			list_move_tail(&snod->list, nondata);
285 	}
286 
287 	/* Sort data and non-data nodes */
288 	list_sort(c, &sleb->nodes, &data_nodes_cmp);
289 	list_sort(c, nondata, &nondata_nodes_cmp);
290 
291 	err = dbg_check_data_nodes_order(c, &sleb->nodes);
292 	if (err)
293 		return err;
294 	err = dbg_check_nondata_nodes_order(c, nondata);
295 	if (err)
296 		return err;
297 	return 0;
298 }
299 
300 /**
301  * move_node - move a node.
302  * @c: UBIFS file-system description object
303  * @sleb: describes the LEB to move nodes from
304  * @snod: the mode to move
305  * @wbuf: write-buffer to move node to
306  *
307  * This function moves node @snod to @wbuf, changes TNC correspondingly, and
308  * destroys @snod. Returns zero in case of success and a negative error code in
309  * case of failure.
310  */
311 static int move_node(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
312 		     struct ubifs_scan_node *snod, struct ubifs_wbuf *wbuf)
313 {
314 	int err, new_lnum = wbuf->lnum, new_offs = wbuf->offs + wbuf->used;
315 
316 	cond_resched();
317 	err = ubifs_wbuf_write_nolock(wbuf, snod->node, snod->len);
318 	if (err)
319 		return err;
320 
321 	err = ubifs_tnc_replace(c, &snod->key, sleb->lnum,
322 				snod->offs, new_lnum, new_offs,
323 				snod->len);
324 	list_del(&snod->list);
325 	kfree(snod);
326 	return err;
327 }
328 
329 /**
330  * move_nodes - move nodes.
331  * @c: UBIFS file-system description object
332  * @sleb: describes the LEB to move nodes from
333  *
334  * This function moves valid nodes from data LEB described by @sleb to the GC
335  * journal head. This function returns zero in case of success, %-EAGAIN if
336  * commit is required, and other negative error codes in case of other
337  * failures.
338  */
339 static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb)
340 {
341 	int err, min;
342 	LIST_HEAD(nondata);
343 	struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
344 
345 	if (wbuf->lnum == -1) {
346 		/*
347 		 * The GC journal head is not set, because it is the first GC
348 		 * invocation since mount.
349 		 */
350 		err = switch_gc_head(c);
351 		if (err)
352 			return err;
353 	}
354 
355 	err = sort_nodes(c, sleb, &nondata, &min);
356 	if (err)
357 		goto out;
358 
359 	/* Write nodes to their new location. Use the first-fit strategy */
360 	while (1) {
361 		int avail;
362 		struct ubifs_scan_node *snod, *tmp;
363 
364 		/* Move data nodes */
365 		list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) {
366 			avail = c->leb_size - wbuf->offs - wbuf->used;
367 			if  (snod->len > avail)
368 				/*
369 				 * Do not skip data nodes in order to optimize
370 				 * bulk-read.
371 				 */
372 				break;
373 
374 			err = move_node(c, sleb, snod, wbuf);
375 			if (err)
376 				goto out;
377 		}
378 
379 		/* Move non-data nodes */
380 		list_for_each_entry_safe(snod, tmp, &nondata, list) {
381 			avail = c->leb_size - wbuf->offs - wbuf->used;
382 			if (avail < min)
383 				break;
384 
385 			if  (snod->len > avail) {
386 				/*
387 				 * Keep going only if this is an inode with
388 				 * some data. Otherwise stop and switch the GC
389 				 * head. IOW, we assume that data-less inode
390 				 * nodes and direntry nodes are roughly of the
391 				 * same size.
392 				 */
393 				if (key_type(c, &snod->key) == UBIFS_DENT_KEY ||
394 				    snod->len == UBIFS_INO_NODE_SZ)
395 					break;
396 				continue;
397 			}
398 
399 			err = move_node(c, sleb, snod, wbuf);
400 			if (err)
401 				goto out;
402 		}
403 
404 		if (list_empty(&sleb->nodes) && list_empty(&nondata))
405 			break;
406 
407 		/*
408 		 * Waste the rest of the space in the LEB and switch to the
409 		 * next LEB.
410 		 */
411 		err = switch_gc_head(c);
412 		if (err)
413 			goto out;
414 	}
415 
416 	return 0;
417 
418 out:
419 	list_splice_tail(&nondata, &sleb->nodes);
420 	return err;
421 }
422 
423 /**
424  * gc_sync_wbufs - sync write-buffers for GC.
425  * @c: UBIFS file-system description object
426  *
427  * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
428  * be in a write-buffer instead. That is, a node could be written to a
429  * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
430  * erased before the write-buffer is sync'd and then there is an unclean
431  * unmount, then an existing node is lost. To avoid this, we sync all
432  * write-buffers.
433  *
434  * This function returns %0 on success or a negative error code on failure.
435  */
436 static int gc_sync_wbufs(struct ubifs_info *c)
437 {
438 	int err, i;
439 
440 	for (i = 0; i < c->jhead_cnt; i++) {
441 		if (i == GCHD)
442 			continue;
443 		err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
444 		if (err)
445 			return err;
446 	}
447 	return 0;
448 }
449 
450 /**
451  * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
452  * @c: UBIFS file-system description object
453  * @lp: describes the LEB to garbage collect
454  *
455  * This function garbage-collects an LEB and returns one of the @LEB_FREED,
456  * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
457  * required, and other negative error codes in case of failures.
458  */
459 int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp)
460 {
461 	struct ubifs_scan_leb *sleb;
462 	struct ubifs_scan_node *snod;
463 	struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
464 	int err = 0, lnum = lp->lnum;
465 
466 	ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 ||
467 		     c->need_recovery);
468 	ubifs_assert(c->gc_lnum != lnum);
469 	ubifs_assert(wbuf->lnum != lnum);
470 
471 	if (lp->free + lp->dirty == c->leb_size) {
472 		/* Special case - a free LEB  */
473 		dbg_gc("LEB %d is free, return it", lp->lnum);
474 		ubifs_assert(!(lp->flags & LPROPS_INDEX));
475 
476 		if (lp->free != c->leb_size) {
477 			/*
478 			 * Write buffers must be sync'd before unmapping
479 			 * freeable LEBs, because one of them may contain data
480 			 * which obsoletes something in 'lp->pnum'.
481 			 */
482 			err = gc_sync_wbufs(c);
483 			if (err)
484 				return err;
485 			err = ubifs_change_one_lp(c, lp->lnum, c->leb_size,
486 						  0, 0, 0, 0);
487 			if (err)
488 				return err;
489 		}
490 		err = ubifs_leb_unmap(c, lp->lnum);
491 		if (err)
492 			return err;
493 
494 		if (c->gc_lnum == -1) {
495 			c->gc_lnum = lnum;
496 			return LEB_RETAINED;
497 		}
498 
499 		return LEB_FREED;
500 	}
501 
502 	/*
503 	 * We scan the entire LEB even though we only really need to scan up to
504 	 * (c->leb_size - lp->free).
505 	 */
506 	sleb = ubifs_scan(c, lnum, 0, c->sbuf, 0);
507 	if (IS_ERR(sleb))
508 		return PTR_ERR(sleb);
509 
510 	ubifs_assert(!list_empty(&sleb->nodes));
511 	snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list);
512 
513 	if (snod->type == UBIFS_IDX_NODE) {
514 		struct ubifs_gced_idx_leb *idx_gc;
515 
516 		dbg_gc("indexing LEB %d (free %d, dirty %d)",
517 		       lnum, lp->free, lp->dirty);
518 		list_for_each_entry(snod, &sleb->nodes, list) {
519 			struct ubifs_idx_node *idx = snod->node;
520 			int level = le16_to_cpu(idx->level);
521 
522 			ubifs_assert(snod->type == UBIFS_IDX_NODE);
523 			key_read(c, ubifs_idx_key(c, idx), &snod->key);
524 			err = ubifs_dirty_idx_node(c, &snod->key, level, lnum,
525 						   snod->offs);
526 			if (err)
527 				goto out;
528 		}
529 
530 		idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
531 		if (!idx_gc) {
532 			err = -ENOMEM;
533 			goto out;
534 		}
535 
536 		idx_gc->lnum = lnum;
537 		idx_gc->unmap = 0;
538 		list_add(&idx_gc->list, &c->idx_gc);
539 
540 		/*
541 		 * Don't release the LEB until after the next commit, because
542 		 * it may contain data which is needed for recovery. So
543 		 * although we freed this LEB, it will become usable only after
544 		 * the commit.
545 		 */
546 		err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0,
547 					  LPROPS_INDEX, 1);
548 		if (err)
549 			goto out;
550 		err = LEB_FREED_IDX;
551 	} else {
552 		dbg_gc("data LEB %d (free %d, dirty %d)",
553 		       lnum, lp->free, lp->dirty);
554 
555 		err = move_nodes(c, sleb);
556 		if (err)
557 			goto out_inc_seq;
558 
559 		err = gc_sync_wbufs(c);
560 		if (err)
561 			goto out_inc_seq;
562 
563 		err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0);
564 		if (err)
565 			goto out_inc_seq;
566 
567 		/* Allow for races with TNC */
568 		c->gced_lnum = lnum;
569 		smp_wmb();
570 		c->gc_seq += 1;
571 		smp_wmb();
572 
573 		if (c->gc_lnum == -1) {
574 			c->gc_lnum = lnum;
575 			err = LEB_RETAINED;
576 		} else {
577 			err = ubifs_wbuf_sync_nolock(wbuf);
578 			if (err)
579 				goto out;
580 
581 			err = ubifs_leb_unmap(c, lnum);
582 			if (err)
583 				goto out;
584 
585 			err = LEB_FREED;
586 		}
587 	}
588 
589 out:
590 	ubifs_scan_destroy(sleb);
591 	return err;
592 
593 out_inc_seq:
594 	/* We may have moved at least some nodes so allow for races with TNC */
595 	c->gced_lnum = lnum;
596 	smp_wmb();
597 	c->gc_seq += 1;
598 	smp_wmb();
599 	goto out;
600 }
601 
602 /**
603  * ubifs_garbage_collect - UBIFS garbage collector.
604  * @c: UBIFS file-system description object
605  * @anyway: do GC even if there are free LEBs
606  *
607  * This function does out-of-place garbage collection. The return codes are:
608  *   o positive LEB number if the LEB has been freed and may be used;
609  *   o %-EAGAIN if the caller has to run commit;
610  *   o %-ENOSPC if GC failed to make any progress;
611  *   o other negative error codes in case of other errors.
612  *
613  * Garbage collector writes data to the journal when GC'ing data LEBs, and just
614  * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
615  * commit may be required. But commit cannot be run from inside GC, because the
616  * caller might be holding the commit lock, so %-EAGAIN is returned instead;
617  * And this error code means that the caller has to run commit, and re-run GC
618  * if there is still no free space.
619  *
620  * There are many reasons why this function may return %-EAGAIN:
621  * o the log is full and there is no space to write an LEB reference for
622  *   @c->gc_lnum;
623  * o the journal is too large and exceeds size limitations;
624  * o GC moved indexing LEBs, but they can be used only after the commit;
625  * o the shrinker fails to find clean znodes to free and requests the commit;
626  * o etc.
627  *
628  * Note, if the file-system is close to be full, this function may return
629  * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
630  * the function. E.g., this happens if the limits on the journal size are too
631  * tough and GC writes too much to the journal before an LEB is freed. This
632  * might also mean that the journal is too large, and the TNC becomes to big,
633  * so that the shrinker is constantly called, finds not clean znodes to free,
634  * and requests commit. Well, this may also happen if the journal is all right,
635  * but another kernel process consumes too much memory. Anyway, infinite
636  * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
637  */
638 int ubifs_garbage_collect(struct ubifs_info *c, int anyway)
639 {
640 	int i, err, ret, min_space = c->dead_wm;
641 	struct ubifs_lprops lp;
642 	struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
643 
644 	ubifs_assert_cmt_locked(c);
645 	ubifs_assert(!c->ro_media && !c->ro_mount);
646 
647 	if (ubifs_gc_should_commit(c))
648 		return -EAGAIN;
649 
650 	mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
651 
652 	if (c->ro_error) {
653 		ret = -EROFS;
654 		goto out_unlock;
655 	}
656 
657 	/* We expect the write-buffer to be empty on entry */
658 	ubifs_assert(!wbuf->used);
659 
660 	for (i = 0; ; i++) {
661 		int space_before, space_after;
662 
663 		cond_resched();
664 
665 		/* Give the commit an opportunity to run */
666 		if (ubifs_gc_should_commit(c)) {
667 			ret = -EAGAIN;
668 			break;
669 		}
670 
671 		if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) {
672 			/*
673 			 * We've done enough iterations. Indexing LEBs were
674 			 * moved and will be available after the commit.
675 			 */
676 			dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
677 			ubifs_commit_required(c);
678 			ret = -EAGAIN;
679 			break;
680 		}
681 
682 		if (i > HARD_LEBS_LIMIT) {
683 			/*
684 			 * We've moved too many LEBs and have not made
685 			 * progress, give up.
686 			 */
687 			dbg_gc("hard limit, -ENOSPC");
688 			ret = -ENOSPC;
689 			break;
690 		}
691 
692 		/*
693 		 * Empty and freeable LEBs can turn up while we waited for
694 		 * the wbuf lock, or while we have been running GC. In that
695 		 * case, we should just return one of those instead of
696 		 * continuing to GC dirty LEBs. Hence we request
697 		 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
698 		 */
699 		ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1);
700 		if (ret) {
701 			if (ret == -ENOSPC)
702 				dbg_gc("no more dirty LEBs");
703 			break;
704 		}
705 
706 		dbg_gc("found LEB %d: free %d, dirty %d, sum %d (min. space %d)",
707 		       lp.lnum, lp.free, lp.dirty, lp.free + lp.dirty,
708 		       min_space);
709 
710 		space_before = c->leb_size - wbuf->offs - wbuf->used;
711 		if (wbuf->lnum == -1)
712 			space_before = 0;
713 
714 		ret = ubifs_garbage_collect_leb(c, &lp);
715 		if (ret < 0) {
716 			if (ret == -EAGAIN) {
717 				/*
718 				 * This is not error, so we have to return the
719 				 * LEB to lprops. But if 'ubifs_return_leb()'
720 				 * fails, its failure code is propagated to the
721 				 * caller instead of the original '-EAGAIN'.
722 				 */
723 				err = ubifs_return_leb(c, lp.lnum);
724 				if (err)
725 					ret = err;
726 				break;
727 			}
728 			goto out;
729 		}
730 
731 		if (ret == LEB_FREED) {
732 			/* An LEB has been freed and is ready for use */
733 			dbg_gc("LEB %d freed, return", lp.lnum);
734 			ret = lp.lnum;
735 			break;
736 		}
737 
738 		if (ret == LEB_FREED_IDX) {
739 			/*
740 			 * This was an indexing LEB and it cannot be
741 			 * immediately used. And instead of requesting the
742 			 * commit straight away, we try to garbage collect some
743 			 * more.
744 			 */
745 			dbg_gc("indexing LEB %d freed, continue", lp.lnum);
746 			continue;
747 		}
748 
749 		ubifs_assert(ret == LEB_RETAINED);
750 		space_after = c->leb_size - wbuf->offs - wbuf->used;
751 		dbg_gc("LEB %d retained, freed %d bytes", lp.lnum,
752 		       space_after - space_before);
753 
754 		if (space_after > space_before) {
755 			/* GC makes progress, keep working */
756 			min_space >>= 1;
757 			if (min_space < c->dead_wm)
758 				min_space = c->dead_wm;
759 			continue;
760 		}
761 
762 		dbg_gc("did not make progress");
763 
764 		/*
765 		 * GC moved an LEB bud have not done any progress. This means
766 		 * that the previous GC head LEB contained too few free space
767 		 * and the LEB which was GC'ed contained only large nodes which
768 		 * did not fit that space.
769 		 *
770 		 * We can do 2 things:
771 		 * 1. pick another LEB in a hope it'll contain a small node
772 		 *    which will fit the space we have at the end of current GC
773 		 *    head LEB, but there is no guarantee, so we try this out
774 		 *    unless we have already been working for too long;
775 		 * 2. request an LEB with more dirty space, which will force
776 		 *    'ubifs_find_dirty_leb()' to start scanning the lprops
777 		 *    table, instead of just picking one from the heap
778 		 *    (previously it already picked the dirtiest LEB).
779 		 */
780 		if (i < SOFT_LEBS_LIMIT) {
781 			dbg_gc("try again");
782 			continue;
783 		}
784 
785 		min_space <<= 1;
786 		if (min_space > c->dark_wm)
787 			min_space = c->dark_wm;
788 		dbg_gc("set min. space to %d", min_space);
789 	}
790 
791 	if (ret == -ENOSPC && !list_empty(&c->idx_gc)) {
792 		dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
793 		ubifs_commit_required(c);
794 		ret = -EAGAIN;
795 	}
796 
797 	err = ubifs_wbuf_sync_nolock(wbuf);
798 	if (!err)
799 		err = ubifs_leb_unmap(c, c->gc_lnum);
800 	if (err) {
801 		ret = err;
802 		goto out;
803 	}
804 out_unlock:
805 	mutex_unlock(&wbuf->io_mutex);
806 	return ret;
807 
808 out:
809 	ubifs_assert(ret < 0);
810 	ubifs_assert(ret != -ENOSPC && ret != -EAGAIN);
811 	ubifs_wbuf_sync_nolock(wbuf);
812 	ubifs_ro_mode(c, ret);
813 	mutex_unlock(&wbuf->io_mutex);
814 	ubifs_return_leb(c, lp.lnum);
815 	return ret;
816 }
817 
818 /**
819  * ubifs_gc_start_commit - garbage collection at start of commit.
820  * @c: UBIFS file-system description object
821  *
822  * If a LEB has only dirty and free space, then we may safely unmap it and make
823  * it free.  Note, we cannot do this with indexing LEBs because dirty space may
824  * correspond index nodes that are required for recovery.  In that case, the
825  * LEB cannot be unmapped until after the next commit.
826  *
827  * This function returns %0 upon success and a negative error code upon failure.
828  */
829 int ubifs_gc_start_commit(struct ubifs_info *c)
830 {
831 	struct ubifs_gced_idx_leb *idx_gc;
832 	const struct ubifs_lprops *lp;
833 	int err = 0, flags;
834 
835 	ubifs_get_lprops(c);
836 
837 	/*
838 	 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
839 	 * wbufs are sync'd before this, which is done in 'do_commit()'.
840 	 */
841 	while (1) {
842 		lp = ubifs_fast_find_freeable(c);
843 		if (IS_ERR(lp)) {
844 			err = PTR_ERR(lp);
845 			goto out;
846 		}
847 		if (!lp)
848 			break;
849 		ubifs_assert(!(lp->flags & LPROPS_TAKEN));
850 		ubifs_assert(!(lp->flags & LPROPS_INDEX));
851 		err = ubifs_leb_unmap(c, lp->lnum);
852 		if (err)
853 			goto out;
854 		lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0);
855 		if (IS_ERR(lp)) {
856 			err = PTR_ERR(lp);
857 			goto out;
858 		}
859 		ubifs_assert(!(lp->flags & LPROPS_TAKEN));
860 		ubifs_assert(!(lp->flags & LPROPS_INDEX));
861 	}
862 
863 	/* Mark GC'd index LEBs OK to unmap after this commit finishes */
864 	list_for_each_entry(idx_gc, &c->idx_gc, list)
865 		idx_gc->unmap = 1;
866 
867 	/* Record index freeable LEBs for unmapping after commit */
868 	while (1) {
869 		lp = ubifs_fast_find_frdi_idx(c);
870 		if (IS_ERR(lp)) {
871 			err = PTR_ERR(lp);
872 			goto out;
873 		}
874 		if (!lp)
875 			break;
876 		idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS);
877 		if (!idx_gc) {
878 			err = -ENOMEM;
879 			goto out;
880 		}
881 		ubifs_assert(!(lp->flags & LPROPS_TAKEN));
882 		ubifs_assert(lp->flags & LPROPS_INDEX);
883 		/* Don't release the LEB until after the next commit */
884 		flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX;
885 		lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1);
886 		if (IS_ERR(lp)) {
887 			err = PTR_ERR(lp);
888 			kfree(idx_gc);
889 			goto out;
890 		}
891 		ubifs_assert(lp->flags & LPROPS_TAKEN);
892 		ubifs_assert(!(lp->flags & LPROPS_INDEX));
893 		idx_gc->lnum = lp->lnum;
894 		idx_gc->unmap = 1;
895 		list_add(&idx_gc->list, &c->idx_gc);
896 	}
897 out:
898 	ubifs_release_lprops(c);
899 	return err;
900 }
901 
902 /**
903  * ubifs_gc_end_commit - garbage collection at end of commit.
904  * @c: UBIFS file-system description object
905  *
906  * This function completes out-of-place garbage collection of index LEBs.
907  */
908 int ubifs_gc_end_commit(struct ubifs_info *c)
909 {
910 	struct ubifs_gced_idx_leb *idx_gc, *tmp;
911 	struct ubifs_wbuf *wbuf;
912 	int err = 0;
913 
914 	wbuf = &c->jheads[GCHD].wbuf;
915 	mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
916 	list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list)
917 		if (idx_gc->unmap) {
918 			dbg_gc("LEB %d", idx_gc->lnum);
919 			err = ubifs_leb_unmap(c, idx_gc->lnum);
920 			if (err)
921 				goto out;
922 			err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC,
923 					  LPROPS_NC, 0, LPROPS_TAKEN, -1);
924 			if (err)
925 				goto out;
926 			list_del(&idx_gc->list);
927 			kfree(idx_gc);
928 		}
929 out:
930 	mutex_unlock(&wbuf->io_mutex);
931 	return err;
932 }
933 #endif
934 /**
935  * ubifs_destroy_idx_gc - destroy idx_gc list.
936  * @c: UBIFS file-system description object
937  *
938  * This function destroys the @c->idx_gc list. It is called when unmounting
939  * so locks are not needed. Returns zero in case of success and a negative
940  * error code in case of failure.
941  */
942 void ubifs_destroy_idx_gc(struct ubifs_info *c)
943 {
944 	while (!list_empty(&c->idx_gc)) {
945 		struct ubifs_gced_idx_leb *idx_gc;
946 
947 		idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb,
948 				    list);
949 		c->idx_gc_cnt -= 1;
950 		list_del(&idx_gc->list);
951 		kfree(idx_gc);
952 	}
953 }
954 #ifndef __UBOOT__
955 /**
956  * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
957  * @c: UBIFS file-system description object
958  *
959  * Called during start commit so locks are not needed.
960  */
961 int ubifs_get_idx_gc_leb(struct ubifs_info *c)
962 {
963 	struct ubifs_gced_idx_leb *idx_gc;
964 	int lnum;
965 
966 	if (list_empty(&c->idx_gc))
967 		return -ENOSPC;
968 	idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list);
969 	lnum = idx_gc->lnum;
970 	/* c->idx_gc_cnt is updated by the caller when lprops are updated */
971 	list_del(&idx_gc->list);
972 	kfree(idx_gc);
973 	return lnum;
974 }
975 #endif
976