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