2 * This file is part of UBIFS.
4 * Copyright (C) 2006-2008 Nokia Corporation.
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.
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
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
19 * Authors: Adrian Hunter
20 * Artem Bityutskiy (Битюцкий Артём)
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.
36 #include <linux/pagemap.h>
40 * GC tries to optimize the way it fit nodes to available space, and it sorts
41 * nodes a little. The below constants are watermarks which define "large",
42 * "medium", and "small" nodes.
44 #define MEDIUM_NODE_WM (UBIFS_BLOCK_SIZE / 4)
45 #define SMALL_NODE_WM UBIFS_MAX_DENT_NODE_SZ
48 * GC may need to move more then one LEB to make progress. The below constants
49 * define "soft" and "hard" limits on the number of LEBs the garbage collector
52 #define SOFT_LEBS_LIMIT 4
53 #define HARD_LEBS_LIMIT 32
56 * switch_gc_head - switch the garbage collection journal head.
57 * @c: UBIFS file-system description object
58 * @buf: buffer to write
59 * @len: length of the buffer to write
60 * @lnum: LEB number written is returned here
61 * @offs: offset written is returned here
63 * This function switch the GC head to the next LEB which is reserved in
64 * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required,
65 * and other negative error code in case of failures.
67 static int switch_gc_head(struct ubifs_info
*c
)
69 int err
, gc_lnum
= c
->gc_lnum
;
70 struct ubifs_wbuf
*wbuf
= &c
->jheads
[GCHD
].wbuf
;
72 ubifs_assert(gc_lnum
!= -1);
73 dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)",
74 wbuf
->lnum
, wbuf
->offs
+ wbuf
->used
, gc_lnum
,
75 c
->leb_size
- wbuf
->offs
- wbuf
->used
);
77 err
= ubifs_wbuf_sync_nolock(wbuf
);
82 * The GC write-buffer was synchronized, we may safely unmap
85 err
= ubifs_leb_unmap(c
, gc_lnum
);
89 err
= ubifs_add_bud_to_log(c
, GCHD
, gc_lnum
, 0);
94 err
= ubifs_wbuf_seek_nolock(wbuf
, gc_lnum
, 0, UBI_LONGTERM
);
99 * move_nodes - move nodes.
100 * @c: UBIFS file-system description object
101 * @sleb: describes nodes to move
103 * This function moves valid nodes from data LEB described by @sleb to the GC
104 * journal head. The obsolete nodes are dropped.
106 * When moving nodes we have to deal with classical bin-packing problem: the
107 * space in the current GC journal head LEB and in @c->gc_lnum are the "bins",
108 * where the nodes in the @sleb->nodes list are the elements which should be
109 * fit optimally to the bins. This function uses the "first fit decreasing"
110 * strategy, although it does not really sort the nodes but just split them on
111 * 3 classes - large, medium, and small, so they are roughly sorted.
113 * This function returns zero in case of success, %-EAGAIN if commit is
114 * required, and other negative error codes in case of other failures.
116 static int move_nodes(struct ubifs_info
*c
, struct ubifs_scan_leb
*sleb
)
118 struct ubifs_scan_node
*snod
, *tmp
;
119 struct list_head large
, medium
, small
;
120 struct ubifs_wbuf
*wbuf
= &c
->jheads
[GCHD
].wbuf
;
121 int avail
, err
, min
= INT_MAX
;
123 INIT_LIST_HEAD(&large
);
124 INIT_LIST_HEAD(&medium
);
125 INIT_LIST_HEAD(&small
);
127 list_for_each_entry_safe(snod
, tmp
, &sleb
->nodes
, list
) {
128 struct list_head
*lst
;
130 ubifs_assert(snod
->type
!= UBIFS_IDX_NODE
);
131 ubifs_assert(snod
->type
!= UBIFS_REF_NODE
);
132 ubifs_assert(snod
->type
!= UBIFS_CS_NODE
);
134 err
= ubifs_tnc_has_node(c
, &snod
->key
, 0, sleb
->lnum
,
142 /* The node is obsolete, remove it from the list */
148 * Sort the list of nodes so that large nodes go first, and
149 * small nodes go last.
151 if (snod
->len
> MEDIUM_NODE_WM
)
152 list_add(lst
, &large
);
153 else if (snod
->len
> SMALL_NODE_WM
)
154 list_add(lst
, &medium
);
156 list_add(lst
, &small
);
158 /* And find the smallest node */
164 * Join the tree lists so that we'd have one roughly sorted list
165 * ('large' will be the head of the joined list).
167 list_splice(&medium
, large
.prev
);
168 list_splice(&small
, large
.prev
);
170 if (wbuf
->lnum
== -1) {
172 * The GC journal head is not set, because it is the first GC
173 * invocation since mount.
175 err
= switch_gc_head(c
);
180 /* Write nodes to their new location. Use the first-fit strategy */
182 avail
= c
->leb_size
- wbuf
->offs
- wbuf
->used
;
183 list_for_each_entry_safe(snod
, tmp
, &large
, list
) {
184 int new_lnum
, new_offs
;
189 if (snod
->len
> avail
)
190 /* This node does not fit */
195 new_lnum
= wbuf
->lnum
;
196 new_offs
= wbuf
->offs
+ wbuf
->used
;
197 err
= ubifs_wbuf_write_nolock(wbuf
, snod
->node
,
201 err
= ubifs_tnc_replace(c
, &snod
->key
, sleb
->lnum
,
202 snod
->offs
, new_lnum
, new_offs
,
207 avail
= c
->leb_size
- wbuf
->offs
- wbuf
->used
;
208 list_del(&snod
->list
);
212 if (list_empty(&large
))
216 * Waste the rest of the space in the LEB and switch to the
219 err
= switch_gc_head(c
);
227 list_for_each_entry_safe(snod
, tmp
, &large
, list
) {
228 list_del(&snod
->list
);
235 * gc_sync_wbufs - sync write-buffers for GC.
236 * @c: UBIFS file-system description object
238 * We must guarantee that obsoleting nodes are on flash. Unfortunately they may
239 * be in a write-buffer instead. That is, a node could be written to a
240 * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is
241 * erased before the write-buffer is sync'd and then there is an unclean
242 * unmount, then an existing node is lost. To avoid this, we sync all
245 * This function returns %0 on success or a negative error code on failure.
247 static int gc_sync_wbufs(struct ubifs_info
*c
)
251 for (i
= 0; i
< c
->jhead_cnt
; i
++) {
254 err
= ubifs_wbuf_sync(&c
->jheads
[i
].wbuf
);
262 * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock.
263 * @c: UBIFS file-system description object
264 * @lp: describes the LEB to garbage collect
266 * This function garbage-collects an LEB and returns one of the @LEB_FREED,
267 * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is
268 * required, and other negative error codes in case of failures.
270 int ubifs_garbage_collect_leb(struct ubifs_info
*c
, struct ubifs_lprops
*lp
)
272 struct ubifs_scan_leb
*sleb
;
273 struct ubifs_scan_node
*snod
;
274 struct ubifs_wbuf
*wbuf
= &c
->jheads
[GCHD
].wbuf
;
275 int err
= 0, lnum
= lp
->lnum
;
277 ubifs_assert(c
->gc_lnum
!= -1 || wbuf
->offs
+ wbuf
->used
== 0 ||
279 ubifs_assert(c
->gc_lnum
!= lnum
);
280 ubifs_assert(wbuf
->lnum
!= lnum
);
283 * We scan the entire LEB even though we only really need to scan up to
284 * (c->leb_size - lp->free).
286 sleb
= ubifs_scan(c
, lnum
, 0, c
->sbuf
);
288 return PTR_ERR(sleb
);
290 ubifs_assert(!list_empty(&sleb
->nodes
));
291 snod
= list_entry(sleb
->nodes
.next
, struct ubifs_scan_node
, list
);
293 if (snod
->type
== UBIFS_IDX_NODE
) {
294 struct ubifs_gced_idx_leb
*idx_gc
;
296 dbg_gc("indexing LEB %d (free %d, dirty %d)",
297 lnum
, lp
->free
, lp
->dirty
);
298 list_for_each_entry(snod
, &sleb
->nodes
, list
) {
299 struct ubifs_idx_node
*idx
= snod
->node
;
300 int level
= le16_to_cpu(idx
->level
);
302 ubifs_assert(snod
->type
== UBIFS_IDX_NODE
);
303 key_read(c
, ubifs_idx_key(c
, idx
), &snod
->key
);
304 err
= ubifs_dirty_idx_node(c
, &snod
->key
, level
, lnum
,
310 idx_gc
= kmalloc(sizeof(struct ubifs_gced_idx_leb
), GFP_NOFS
);
318 list_add(&idx_gc
->list
, &c
->idx_gc
);
321 * Don't release the LEB until after the next commit, because
322 * it may contain date which is needed for recovery. So
323 * although we freed this LEB, it will become usable only after
326 err
= ubifs_change_one_lp(c
, lnum
, c
->leb_size
, 0, 0,
332 dbg_gc("data LEB %d (free %d, dirty %d)",
333 lnum
, lp
->free
, lp
->dirty
);
335 err
= move_nodes(c
, sleb
);
339 err
= gc_sync_wbufs(c
);
343 err
= ubifs_change_one_lp(c
, lnum
, c
->leb_size
, 0, 0, 0, 0);
347 /* Allow for races with TNC */
353 if (c
->gc_lnum
== -1) {
357 err
= ubifs_wbuf_sync_nolock(wbuf
);
361 err
= ubifs_leb_unmap(c
, lnum
);
370 ubifs_scan_destroy(sleb
);
375 * ubifs_garbage_collect - UBIFS garbage collector.
376 * @c: UBIFS file-system description object
377 * @anyway: do GC even if there are free LEBs
379 * This function does out-of-place garbage collection. The return codes are:
380 * o positive LEB number if the LEB has been freed and may be used;
381 * o %-EAGAIN if the caller has to run commit;
382 * o %-ENOSPC if GC failed to make any progress;
383 * o other negative error codes in case of other errors.
385 * Garbage collector writes data to the journal when GC'ing data LEBs, and just
386 * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point
387 * commit may be required. But commit cannot be run from inside GC, because the
388 * caller might be holding the commit lock, so %-EAGAIN is returned instead;
389 * And this error code means that the caller has to run commit, and re-run GC
390 * if there is still no free space.
392 * There are many reasons why this function may return %-EAGAIN:
393 * o the log is full and there is no space to write an LEB reference for
395 * o the journal is too large and exceeds size limitations;
396 * o GC moved indexing LEBs, but they can be used only after the commit;
397 * o the shrinker fails to find clean znodes to free and requests the commit;
400 * Note, if the file-system is close to be full, this function may return
401 * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of
402 * the function. E.g., this happens if the limits on the journal size are too
403 * tough and GC writes too much to the journal before an LEB is freed. This
404 * might also mean that the journal is too large, and the TNC becomes to big,
405 * so that the shrinker is constantly called, finds not clean znodes to free,
406 * and requests commit. Well, this may also happen if the journal is all right,
407 * but another kernel process consumes too much memory. Anyway, infinite
408 * %-EAGAIN may happen, but in some extreme/misconfiguration cases.
410 int ubifs_garbage_collect(struct ubifs_info
*c
, int anyway
)
412 int i
, err
, ret
, min_space
= c
->dead_wm
;
413 struct ubifs_lprops lp
;
414 struct ubifs_wbuf
*wbuf
= &c
->jheads
[GCHD
].wbuf
;
416 ubifs_assert_cmt_locked(c
);
418 if (ubifs_gc_should_commit(c
))
421 mutex_lock_nested(&wbuf
->io_mutex
, wbuf
->jhead
);
428 /* We expect the write-buffer to be empty on entry */
429 ubifs_assert(!wbuf
->used
);
432 int space_before
= c
->leb_size
- wbuf
->offs
- wbuf
->used
;
437 /* Give the commit an opportunity to run */
438 if (ubifs_gc_should_commit(c
)) {
443 if (i
> SOFT_LEBS_LIMIT
&& !list_empty(&c
->idx_gc
)) {
445 * We've done enough iterations. Indexing LEBs were
446 * moved and will be available after the commit.
448 dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN");
449 ubifs_commit_required(c
);
454 if (i
> HARD_LEBS_LIMIT
) {
456 * We've moved too many LEBs and have not made
459 dbg_gc("hard limit, -ENOSPC");
465 * Empty and freeable LEBs can turn up while we waited for
466 * the wbuf lock, or while we have been running GC. In that
467 * case, we should just return one of those instead of
468 * continuing to GC dirty LEBs. Hence we request
469 * 'ubifs_find_dirty_leb()' to return an empty LEB if it can.
471 ret
= ubifs_find_dirty_leb(c
, &lp
, min_space
, anyway
? 0 : 1);
474 dbg_gc("no more dirty LEBs");
478 dbg_gc("found LEB %d: free %d, dirty %d, sum %d "
479 "(min. space %d)", lp
.lnum
, lp
.free
, lp
.dirty
,
480 lp
.free
+ lp
.dirty
, min_space
);
482 if (lp
.free
+ lp
.dirty
== c
->leb_size
) {
483 /* An empty LEB was returned */
484 dbg_gc("LEB %d is free, return it", lp
.lnum
);
486 * ubifs_find_dirty_leb() doesn't return freeable index
489 ubifs_assert(!(lp
.flags
& LPROPS_INDEX
));
490 if (lp
.free
!= c
->leb_size
) {
492 * Write buffers must be sync'd before
493 * unmapping freeable LEBs, because one of them
494 * may contain data which obsoletes something
497 ret
= gc_sync_wbufs(c
);
500 ret
= ubifs_change_one_lp(c
, lp
.lnum
,
501 c
->leb_size
, 0, 0, 0,
506 ret
= ubifs_leb_unmap(c
, lp
.lnum
);
513 space_before
= c
->leb_size
- wbuf
->offs
- wbuf
->used
;
514 if (wbuf
->lnum
== -1)
517 ret
= ubifs_garbage_collect_leb(c
, &lp
);
519 if (ret
== -EAGAIN
|| ret
== -ENOSPC
) {
521 * These codes are not errors, so we have to
522 * return the LEB to lprops. But if the
523 * 'ubifs_return_leb()' function fails, its
524 * failure code is propagated to the caller
525 * instead of the original '-EAGAIN' or
528 err
= ubifs_return_leb(c
, lp
.lnum
);
536 if (ret
== LEB_FREED
) {
537 /* An LEB has been freed and is ready for use */
538 dbg_gc("LEB %d freed, return", lp
.lnum
);
543 if (ret
== LEB_FREED_IDX
) {
545 * This was an indexing LEB and it cannot be
546 * immediately used. And instead of requesting the
547 * commit straight away, we try to garbage collect some
550 dbg_gc("indexing LEB %d freed, continue", lp
.lnum
);
554 ubifs_assert(ret
== LEB_RETAINED
);
555 space_after
= c
->leb_size
- wbuf
->offs
- wbuf
->used
;
556 dbg_gc("LEB %d retained, freed %d bytes", lp
.lnum
,
557 space_after
- space_before
);
559 if (space_after
> space_before
) {
560 /* GC makes progress, keep working */
562 if (min_space
< c
->dead_wm
)
563 min_space
= c
->dead_wm
;
567 dbg_gc("did not make progress");
570 * GC moved an LEB bud have not done any progress. This means
571 * that the previous GC head LEB contained too few free space
572 * and the LEB which was GC'ed contained only large nodes which
573 * did not fit that space.
575 * We can do 2 things:
576 * 1. pick another LEB in a hope it'll contain a small node
577 * which will fit the space we have at the end of current GC
578 * head LEB, but there is no guarantee, so we try this out
579 * unless we have already been working for too long;
580 * 2. request an LEB with more dirty space, which will force
581 * 'ubifs_find_dirty_leb()' to start scanning the lprops
582 * table, instead of just picking one from the heap
583 * (previously it already picked the dirtiest LEB).
585 if (i
< SOFT_LEBS_LIMIT
) {
591 if (min_space
> c
->dark_wm
)
592 min_space
= c
->dark_wm
;
593 dbg_gc("set min. space to %d", min_space
);
596 if (ret
== -ENOSPC
&& !list_empty(&c
->idx_gc
)) {
597 dbg_gc("no space, some index LEBs GC'ed, -EAGAIN");
598 ubifs_commit_required(c
);
602 err
= ubifs_wbuf_sync_nolock(wbuf
);
604 err
= ubifs_leb_unmap(c
, c
->gc_lnum
);
610 mutex_unlock(&wbuf
->io_mutex
);
614 ubifs_assert(ret
< 0);
615 ubifs_assert(ret
!= -ENOSPC
&& ret
!= -EAGAIN
);
616 ubifs_ro_mode(c
, ret
);
617 ubifs_wbuf_sync_nolock(wbuf
);
618 mutex_unlock(&wbuf
->io_mutex
);
619 ubifs_return_leb(c
, lp
.lnum
);
624 * ubifs_gc_start_commit - garbage collection at start of commit.
625 * @c: UBIFS file-system description object
627 * If a LEB has only dirty and free space, then we may safely unmap it and make
628 * it free. Note, we cannot do this with indexing LEBs because dirty space may
629 * correspond index nodes that are required for recovery. In that case, the
630 * LEB cannot be unmapped until after the next commit.
632 * This function returns %0 upon success and a negative error code upon failure.
634 int ubifs_gc_start_commit(struct ubifs_info
*c
)
636 struct ubifs_gced_idx_leb
*idx_gc
;
637 const struct ubifs_lprops
*lp
;
643 * Unmap (non-index) freeable LEBs. Note that recovery requires that all
644 * wbufs are sync'd before this, which is done in 'do_commit()'.
647 lp
= ubifs_fast_find_freeable(c
);
648 if (unlikely(IS_ERR(lp
))) {
654 ubifs_assert(!(lp
->flags
& LPROPS_TAKEN
));
655 ubifs_assert(!(lp
->flags
& LPROPS_INDEX
));
656 err
= ubifs_leb_unmap(c
, lp
->lnum
);
659 lp
= ubifs_change_lp(c
, lp
, c
->leb_size
, 0, lp
->flags
, 0);
660 if (unlikely(IS_ERR(lp
))) {
664 ubifs_assert(!(lp
->flags
& LPROPS_TAKEN
));
665 ubifs_assert(!(lp
->flags
& LPROPS_INDEX
));
668 /* Mark GC'd index LEBs OK to unmap after this commit finishes */
669 list_for_each_entry(idx_gc
, &c
->idx_gc
, list
)
672 /* Record index freeable LEBs for unmapping after commit */
674 lp
= ubifs_fast_find_frdi_idx(c
);
675 if (unlikely(IS_ERR(lp
))) {
681 idx_gc
= kmalloc(sizeof(struct ubifs_gced_idx_leb
), GFP_NOFS
);
686 ubifs_assert(!(lp
->flags
& LPROPS_TAKEN
));
687 ubifs_assert(lp
->flags
& LPROPS_INDEX
);
688 /* Don't release the LEB until after the next commit */
689 flags
= (lp
->flags
| LPROPS_TAKEN
) ^ LPROPS_INDEX
;
690 lp
= ubifs_change_lp(c
, lp
, c
->leb_size
, 0, flags
, 1);
691 if (unlikely(IS_ERR(lp
))) {
696 ubifs_assert(lp
->flags
& LPROPS_TAKEN
);
697 ubifs_assert(!(lp
->flags
& LPROPS_INDEX
));
698 idx_gc
->lnum
= lp
->lnum
;
700 list_add(&idx_gc
->list
, &c
->idx_gc
);
703 ubifs_release_lprops(c
);
708 * ubifs_gc_end_commit - garbage collection at end of commit.
709 * @c: UBIFS file-system description object
711 * This function completes out-of-place garbage collection of index LEBs.
713 int ubifs_gc_end_commit(struct ubifs_info
*c
)
715 struct ubifs_gced_idx_leb
*idx_gc
, *tmp
;
716 struct ubifs_wbuf
*wbuf
;
719 wbuf
= &c
->jheads
[GCHD
].wbuf
;
720 mutex_lock_nested(&wbuf
->io_mutex
, wbuf
->jhead
);
721 list_for_each_entry_safe(idx_gc
, tmp
, &c
->idx_gc
, list
)
723 dbg_gc("LEB %d", idx_gc
->lnum
);
724 err
= ubifs_leb_unmap(c
, idx_gc
->lnum
);
727 err
= ubifs_change_one_lp(c
, idx_gc
->lnum
, LPROPS_NC
,
728 LPROPS_NC
, 0, LPROPS_TAKEN
, -1);
731 list_del(&idx_gc
->list
);
735 mutex_unlock(&wbuf
->io_mutex
);
740 * ubifs_destroy_idx_gc - destroy idx_gc list.
741 * @c: UBIFS file-system description object
743 * This function destroys the idx_gc list. It is called when unmounting or
744 * remounting read-only so locks are not needed.
746 void ubifs_destroy_idx_gc(struct ubifs_info
*c
)
748 while (!list_empty(&c
->idx_gc
)) {
749 struct ubifs_gced_idx_leb
*idx_gc
;
751 idx_gc
= list_entry(c
->idx_gc
.next
, struct ubifs_gced_idx_leb
,
754 list_del(&idx_gc
->list
);
761 * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list.
762 * @c: UBIFS file-system description object
764 * Called during start commit so locks are not needed.
766 int ubifs_get_idx_gc_leb(struct ubifs_info
*c
)
768 struct ubifs_gced_idx_leb
*idx_gc
;
771 if (list_empty(&c
->idx_gc
))
773 idx_gc
= list_entry(c
->idx_gc
.next
, struct ubifs_gced_idx_leb
, list
);
775 /* c->idx_gc_cnt is updated by the caller when lprops are updated */
776 list_del(&idx_gc
->list
);