1 // SPDX-License-Identifier: GPL-2.0-only
3 * This file is part of UBIFS.
5 * Copyright (C) 2006-2008 Nokia Corporation
7 * Authors: Adrian Hunter
8 * Artem Bityutskiy (Битюцкий Артём)
12 * This file implements functions needed to recover from unclean un-mounts.
13 * When UBIFS is mounted, it checks a flag on the master node to determine if
14 * an un-mount was completed successfully. If not, the process of mounting
15 * incorporates additional checking and fixing of on-flash data structures.
16 * UBIFS always cleans away all remnants of an unclean un-mount, so that
17 * errors do not accumulate. However UBIFS defers recovery if it is mounted
18 * read-only, and the flash is not modified in that case.
20 * The general UBIFS approach to the recovery is that it recovers from
21 * corruptions which could be caused by power cuts, but it refuses to recover
22 * from corruption caused by other reasons. And UBIFS tries to distinguish
23 * between these 2 reasons of corruptions and silently recover in the former
24 * case and loudly complain in the latter case.
26 * UBIFS writes only to erased LEBs, so it writes only to the flash space
27 * containing only 0xFFs. UBIFS also always writes strictly from the beginning
28 * of the LEB to the end. And UBIFS assumes that the underlying flash media
29 * writes in @c->max_write_size bytes at a time.
31 * Hence, if UBIFS finds a corrupted node at offset X, it expects only the min.
32 * I/O unit corresponding to offset X to contain corrupted data, all the
33 * following min. I/O units have to contain empty space (all 0xFFs). If this is
34 * not true, the corruption cannot be the result of a power cut, and UBIFS
38 #include <linux/crc32.h>
39 #include <linux/slab.h>
43 * is_empty - determine whether a buffer is empty (contains all 0xff).
44 * @buf: buffer to clean
45 * @len: length of buffer
47 * This function returns %1 if the buffer is empty (contains all 0xff) otherwise
50 static int is_empty(void *buf, int len)
55 for (i = 0; i < len; i++)
62 * first_non_ff - find offset of the first non-0xff byte.
63 * @buf: buffer to search in
64 * @len: length of buffer
66 * This function returns offset of the first non-0xff byte in @buf or %-1 if
67 * the buffer contains only 0xff bytes.
69 static int first_non_ff(void *buf, int len)
74 for (i = 0; i < len; i++)
81 * get_master_node - get the last valid master node allowing for corruption.
82 * @c: UBIFS file-system description object
84 * @pbuf: buffer containing the LEB read, is returned here
85 * @mst: master node, if found, is returned here
86 * @cor: corruption, if found, is returned here
88 * This function allocates a buffer, reads the LEB into it, and finds and
89 * returns the last valid master node allowing for one area of corruption.
90 * The corrupt area, if there is one, must be consistent with the assumption
91 * that it is the result of an unclean unmount while the master node was being
92 * written. Under those circumstances, it is valid to use the previously written
95 * This function returns %0 on success and a negative error code on failure.
97 static int get_master_node(const struct ubifs_info *c, int lnum, void **pbuf,
98 struct ubifs_mst_node **mst, void **cor)
100 const int sz = c->mst_node_alsz;
104 sbuf = vmalloc(c->leb_size);
108 err = ubifs_leb_read(c, lnum, sbuf, 0, c->leb_size, 0);
109 if (err && err != -EBADMSG)
112 /* Find the first position that is definitely not a node */
116 while (offs + UBIFS_MST_NODE_SZ <= c->leb_size) {
117 struct ubifs_ch *ch = buf;
119 if (le32_to_cpu(ch->magic) != UBIFS_NODE_MAGIC)
125 /* See if there was a valid master node before that */
132 ret = ubifs_scan_a_node(c, buf, len, lnum, offs, 1);
133 if (ret != SCANNED_A_NODE && offs) {
134 /* Could have been corruption so check one place back */
138 ret = ubifs_scan_a_node(c, buf, len, lnum, offs, 1);
139 if (ret != SCANNED_A_NODE)
141 * We accept only one area of corruption because
142 * we are assuming that it was caused while
143 * trying to write a master node.
147 if (ret == SCANNED_A_NODE) {
148 struct ubifs_ch *ch = buf;
150 if (ch->node_type != UBIFS_MST_NODE)
152 dbg_rcvry("found a master node at %d:%d", lnum, offs);
159 /* Check for corruption */
160 if (offs < c->leb_size) {
161 if (!is_empty(buf, min_t(int, len, sz))) {
163 dbg_rcvry("found corruption at %d:%d", lnum, offs);
169 /* Check remaining empty space */
170 if (offs < c->leb_size)
171 if (!is_empty(buf, len))
186 * write_rcvrd_mst_node - write recovered master node.
187 * @c: UBIFS file-system description object
190 * This function returns %0 on success and a negative error code on failure.
192 static int write_rcvrd_mst_node(struct ubifs_info *c,
193 struct ubifs_mst_node *mst)
195 int err = 0, lnum = UBIFS_MST_LNUM, sz = c->mst_node_alsz;
198 dbg_rcvry("recovery");
200 save_flags = mst->flags;
201 mst->flags |= cpu_to_le32(UBIFS_MST_RCVRY);
203 err = ubifs_prepare_node_hmac(c, mst, UBIFS_MST_NODE_SZ,
204 offsetof(struct ubifs_mst_node, hmac), 1);
207 err = ubifs_leb_change(c, lnum, mst, sz);
210 err = ubifs_leb_change(c, lnum + 1, mst, sz);
214 mst->flags = save_flags;
219 * ubifs_recover_master_node - recover the master node.
220 * @c: UBIFS file-system description object
222 * This function recovers the master node from corruption that may occur due to
223 * an unclean unmount.
225 * This function returns %0 on success and a negative error code on failure.
227 int ubifs_recover_master_node(struct ubifs_info *c)
229 void *buf1 = NULL, *buf2 = NULL, *cor1 = NULL, *cor2 = NULL;
230 struct ubifs_mst_node *mst1 = NULL, *mst2 = NULL, *mst;
231 const int sz = c->mst_node_alsz;
232 int err, offs1, offs2;
234 dbg_rcvry("recovery");
236 err = get_master_node(c, UBIFS_MST_LNUM, &buf1, &mst1, &cor1);
240 err = get_master_node(c, UBIFS_MST_LNUM + 1, &buf2, &mst2, &cor2);
245 offs1 = (void *)mst1 - buf1;
246 if ((le32_to_cpu(mst1->flags) & UBIFS_MST_RCVRY) &&
247 (offs1 == 0 && !cor1)) {
249 * mst1 was written by recovery at offset 0 with no
252 dbg_rcvry("recovery recovery");
255 offs2 = (void *)mst2 - buf2;
256 if (offs1 == offs2) {
257 /* Same offset, so must be the same */
258 if (ubifs_compare_master_node(c, mst1, mst2))
261 } else if (offs2 + sz == offs1) {
262 /* 1st LEB was written, 2nd was not */
266 } else if (offs1 == 0 &&
267 c->leb_size - offs2 - sz < sz) {
268 /* 1st LEB was unmapped and written, 2nd not */
276 * 2nd LEB was unmapped and about to be written, so
277 * there must be only one master node in the first LEB
280 if (offs1 != 0 || cor1)
288 * 1st LEB was unmapped and about to be written, so there must
289 * be no room left in 2nd LEB.
291 offs2 = (void *)mst2 - buf2;
292 if (offs2 + sz + sz <= c->leb_size)
297 ubifs_msg(c, "recovered master node from LEB %d",
298 (mst == mst1 ? UBIFS_MST_LNUM : UBIFS_MST_LNUM + 1));
300 memcpy(c->mst_node, mst, UBIFS_MST_NODE_SZ);
303 /* Read-only mode. Keep a copy for switching to rw mode */
304 c->rcvrd_mst_node = kmalloc(sz, GFP_KERNEL);
305 if (!c->rcvrd_mst_node) {
309 memcpy(c->rcvrd_mst_node, c->mst_node, UBIFS_MST_NODE_SZ);
312 * We had to recover the master node, which means there was an
313 * unclean reboot. However, it is possible that the master node
314 * is clean at this point, i.e., %UBIFS_MST_DIRTY is not set.
315 * E.g., consider the following chain of events:
317 * 1. UBIFS was cleanly unmounted, so the master node is clean
318 * 2. UBIFS is being mounted R/W and starts changing the master
319 * node in the first (%UBIFS_MST_LNUM). A power cut happens,
320 * so this LEB ends up with some amount of garbage at the
322 * 3. UBIFS is being mounted R/O. We reach this place and
323 * recover the master node from the second LEB
324 * (%UBIFS_MST_LNUM + 1). But we cannot update the media
325 * because we are being mounted R/O. We have to defer the
327 * 4. However, this master node (@c->mst_node) is marked as
328 * clean (since the step 1). And if we just return, the
329 * mount code will be confused and won't recover the master
330 * node when it is re-mounter R/W later.
332 * Thus, to force the recovery by marking the master node as
335 c->mst_node->flags |= cpu_to_le32(UBIFS_MST_DIRTY);
337 /* Write the recovered master node */
338 c->max_sqnum = le64_to_cpu(mst->ch.sqnum) - 1;
339 err = write_rcvrd_mst_node(c, c->mst_node);
352 ubifs_err(c, "failed to recover master node");
354 ubifs_err(c, "dumping first master node");
355 ubifs_dump_node(c, mst1, c->leb_size - ((void *)mst1 - buf1));
358 ubifs_err(c, "dumping second master node");
359 ubifs_dump_node(c, mst2, c->leb_size - ((void *)mst2 - buf2));
367 * ubifs_write_rcvrd_mst_node - write the recovered master node.
368 * @c: UBIFS file-system description object
370 * This function writes the master node that was recovered during mounting in
371 * read-only mode and must now be written because we are remounting rw.
373 * This function returns %0 on success and a negative error code on failure.
375 int ubifs_write_rcvrd_mst_node(struct ubifs_info *c)
379 if (!c->rcvrd_mst_node)
381 c->rcvrd_mst_node->flags |= cpu_to_le32(UBIFS_MST_DIRTY);
382 c->mst_node->flags |= cpu_to_le32(UBIFS_MST_DIRTY);
383 err = write_rcvrd_mst_node(c, c->rcvrd_mst_node);
386 kfree(c->rcvrd_mst_node);
387 c->rcvrd_mst_node = NULL;
392 * is_last_write - determine if an offset was in the last write to a LEB.
393 * @c: UBIFS file-system description object
394 * @buf: buffer to check
395 * @offs: offset to check
397 * This function returns %1 if @offs was in the last write to the LEB whose data
398 * is in @buf, otherwise %0 is returned. The determination is made by checking
399 * for subsequent empty space starting from the next @c->max_write_size
402 static int is_last_write(const struct ubifs_info *c, void *buf, int offs)
404 int empty_offs, check_len;
408 * Round up to the next @c->max_write_size boundary i.e. @offs is in
409 * the last wbuf written. After that should be empty space.
411 empty_offs = ALIGN(offs + 1, c->max_write_size);
412 check_len = c->leb_size - empty_offs;
413 p = buf + empty_offs - offs;
414 return is_empty(p, check_len);
418 * clean_buf - clean the data from an LEB sitting in a buffer.
419 * @c: UBIFS file-system description object
420 * @buf: buffer to clean
421 * @lnum: LEB number to clean
422 * @offs: offset from which to clean
423 * @len: length of buffer
425 * This function pads up to the next min_io_size boundary (if there is one) and
426 * sets empty space to all 0xff. @buf, @offs and @len are updated to the next
427 * @c->min_io_size boundary.
429 static void clean_buf(const struct ubifs_info *c, void **buf, int lnum,
432 int empty_offs, pad_len;
434 dbg_rcvry("cleaning corruption at %d:%d", lnum, *offs);
436 ubifs_assert(c, !(*offs & 7));
437 empty_offs = ALIGN(*offs, c->min_io_size);
438 pad_len = empty_offs - *offs;
439 ubifs_pad(c, *buf, pad_len);
443 memset(*buf, 0xff, c->leb_size - empty_offs);
447 * no_more_nodes - determine if there are no more nodes in a buffer.
448 * @c: UBIFS file-system description object
449 * @buf: buffer to check
450 * @len: length of buffer
451 * @lnum: LEB number of the LEB from which @buf was read
452 * @offs: offset from which @buf was read
454 * This function ensures that the corrupted node at @offs is the last thing
455 * written to a LEB. This function returns %1 if more data is not found and
456 * %0 if more data is found.
458 static int no_more_nodes(const struct ubifs_info *c, void *buf, int len,
461 struct ubifs_ch *ch = buf;
462 int skip, dlen = le32_to_cpu(ch->len);
464 /* Check for empty space after the corrupt node's common header */
465 skip = ALIGN(offs + UBIFS_CH_SZ, c->max_write_size) - offs;
466 if (is_empty(buf + skip, len - skip))
469 * The area after the common header size is not empty, so the common
470 * header must be intact. Check it.
472 if (ubifs_check_node(c, buf, len, lnum, offs, 1, 0) != -EUCLEAN) {
473 dbg_rcvry("unexpected bad common header at %d:%d", lnum, offs);
476 /* Now we know the corrupt node's length we can skip over it */
477 skip = ALIGN(offs + dlen, c->max_write_size) - offs;
478 /* After which there should be empty space */
479 if (is_empty(buf + skip, len - skip))
481 dbg_rcvry("unexpected data at %d:%d", lnum, offs + skip);
486 * fix_unclean_leb - fix an unclean LEB.
487 * @c: UBIFS file-system description object
488 * @sleb: scanned LEB information
489 * @start: offset where scan started
491 static int fix_unclean_leb(struct ubifs_info *c, struct ubifs_scan_leb *sleb,
494 int lnum = sleb->lnum, endpt = start;
496 /* Get the end offset of the last node we are keeping */
497 if (!list_empty(&sleb->nodes)) {
498 struct ubifs_scan_node *snod;
500 snod = list_entry(sleb->nodes.prev,
501 struct ubifs_scan_node, list);
502 endpt = snod->offs + snod->len;
505 if (c->ro_mount && !c->remounting_rw) {
506 /* Add to recovery list */
507 struct ubifs_unclean_leb *ucleb;
509 dbg_rcvry("need to fix LEB %d start %d endpt %d",
510 lnum, start, sleb->endpt);
511 ucleb = kzalloc(sizeof(struct ubifs_unclean_leb), GFP_NOFS);
515 ucleb->endpt = endpt;
516 list_add_tail(&ucleb->list, &c->unclean_leb_list);
518 /* Write the fixed LEB back to flash */
521 dbg_rcvry("fixing LEB %d start %d endpt %d",
522 lnum, start, sleb->endpt);
524 err = ubifs_leb_unmap(c, lnum);
528 int len = ALIGN(endpt, c->min_io_size);
531 err = ubifs_leb_read(c, lnum, sleb->buf, 0,
536 /* Pad to min_io_size */
538 int pad_len = len - ALIGN(endpt, 8);
541 void *buf = sleb->buf + len - pad_len;
543 ubifs_pad(c, buf, pad_len);
546 err = ubifs_leb_change(c, lnum, sleb->buf, len);
555 * drop_last_group - drop the last group of nodes.
556 * @sleb: scanned LEB information
557 * @offs: offset of dropped nodes is returned here
559 * This is a helper function for 'ubifs_recover_leb()' which drops the last
560 * group of nodes of the scanned LEB.
562 static void drop_last_group(struct ubifs_scan_leb *sleb, int *offs)
564 while (!list_empty(&sleb->nodes)) {
565 struct ubifs_scan_node *snod;
568 snod = list_entry(sleb->nodes.prev, struct ubifs_scan_node,
571 if (ch->group_type != UBIFS_IN_NODE_GROUP)
574 dbg_rcvry("dropping grouped node at %d:%d",
575 sleb->lnum, snod->offs);
577 list_del(&snod->list);
579 sleb->nodes_cnt -= 1;
584 * drop_last_node - drop the last node.
585 * @sleb: scanned LEB information
586 * @offs: offset of dropped nodes is returned here
588 * This is a helper function for 'ubifs_recover_leb()' which drops the last
589 * node of the scanned LEB.
591 static void drop_last_node(struct ubifs_scan_leb *sleb, int *offs)
593 struct ubifs_scan_node *snod;
595 if (!list_empty(&sleb->nodes)) {
596 snod = list_entry(sleb->nodes.prev, struct ubifs_scan_node,
599 dbg_rcvry("dropping last node at %d:%d",
600 sleb->lnum, snod->offs);
602 list_del(&snod->list);
604 sleb->nodes_cnt -= 1;
609 * ubifs_recover_leb - scan and recover a LEB.
610 * @c: UBIFS file-system description object
613 * @sbuf: LEB-sized buffer to use
614 * @jhead: journal head number this LEB belongs to (%-1 if the LEB does not
615 * belong to any journal head)
617 * This function does a scan of a LEB, but caters for errors that might have
618 * been caused by the unclean unmount from which we are attempting to recover.
619 * Returns the scanned information on success and a negative error code on
622 struct ubifs_scan_leb *ubifs_recover_leb(struct ubifs_info *c, int lnum,
623 int offs, void *sbuf, int jhead)
625 int ret = 0, err, len = c->leb_size - offs, start = offs, min_io_unit;
626 int grouped = jhead == -1 ? 0 : c->jheads[jhead].grouped;
627 struct ubifs_scan_leb *sleb;
628 void *buf = sbuf + offs;
630 dbg_rcvry("%d:%d, jhead %d, grouped %d", lnum, offs, jhead, grouped);
632 sleb = ubifs_start_scan(c, lnum, offs, sbuf);
636 ubifs_assert(c, len >= 8);
638 dbg_scan("look at LEB %d:%d (%d bytes left)",
644 * Scan quietly until there is an error from which we cannot
647 ret = ubifs_scan_a_node(c, buf, len, lnum, offs, 1);
648 if (ret == SCANNED_A_NODE) {
649 /* A valid node, and not a padding node */
650 struct ubifs_ch *ch = buf;
653 err = ubifs_add_snod(c, sleb, buf, offs);
656 node_len = ALIGN(le32_to_cpu(ch->len), 8);
660 } else if (ret > 0) {
661 /* Padding bytes or a valid padding node */
665 } else if (ret == SCANNED_EMPTY_SPACE ||
666 ret == SCANNED_GARBAGE ||
667 ret == SCANNED_A_BAD_PAD_NODE ||
668 ret == SCANNED_A_CORRUPT_NODE) {
669 dbg_rcvry("found corruption (%d) at %d:%d",
673 ubifs_err(c, "unexpected return value %d", ret);
679 if (ret == SCANNED_GARBAGE || ret == SCANNED_A_BAD_PAD_NODE) {
680 if (!is_last_write(c, buf, offs))
681 goto corrupted_rescan;
682 } else if (ret == SCANNED_A_CORRUPT_NODE) {
683 if (!no_more_nodes(c, buf, len, lnum, offs))
684 goto corrupted_rescan;
685 } else if (!is_empty(buf, len)) {
686 if (!is_last_write(c, buf, offs)) {
687 int corruption = first_non_ff(buf, len);
690 * See header comment for this file for more
691 * explanations about the reasons we have this check.
693 ubifs_err(c, "corrupt empty space LEB %d:%d, corruption starts at %d",
694 lnum, offs, corruption);
695 /* Make sure we dump interesting non-0xFF data */
702 min_io_unit = round_down(offs, c->min_io_size);
705 * If nodes are grouped, always drop the incomplete group at
708 drop_last_group(sleb, &offs);
712 * If this LEB belongs to the GC head then while we are in the
713 * middle of the same min. I/O unit keep dropping nodes. So
714 * basically, what we want is to make sure that the last min.
715 * I/O unit where we saw the corruption is dropped completely
716 * with all the uncorrupted nodes which may possibly sit there.
718 * In other words, let's name the min. I/O unit where the
719 * corruption starts B, and the previous min. I/O unit A. The
720 * below code tries to deal with a situation when half of B
721 * contains valid nodes or the end of a valid node, and the
722 * second half of B contains corrupted data or garbage. This
723 * means that UBIFS had been writing to B just before the power
724 * cut happened. I do not know how realistic is this scenario
725 * that half of the min. I/O unit had been written successfully
726 * and the other half not, but this is possible in our 'failure
727 * mode emulation' infrastructure at least.
729 * So what is the problem, why we need to drop those nodes? Why
730 * can't we just clean-up the second half of B by putting a
731 * padding node there? We can, and this works fine with one
732 * exception which was reproduced with power cut emulation
733 * testing and happens extremely rarely.
735 * Imagine the file-system is full, we run GC which starts
736 * moving valid nodes from LEB X to LEB Y (obviously, LEB Y is
737 * the current GC head LEB). The @c->gc_lnum is -1, which means
738 * that GC will retain LEB X and will try to continue. Imagine
739 * that LEB X is currently the dirtiest LEB, and the amount of
740 * used space in LEB Y is exactly the same as amount of free
743 * And a power cut happens when nodes are moved from LEB X to
744 * LEB Y. We are here trying to recover LEB Y which is the GC
745 * head LEB. We find the min. I/O unit B as described above.
746 * Then we clean-up LEB Y by padding min. I/O unit. And later
747 * 'ubifs_rcvry_gc_commit()' function fails, because it cannot
748 * find a dirty LEB which could be GC'd into LEB Y! Even LEB X
749 * does not match because the amount of valid nodes there does
750 * not fit the free space in LEB Y any more! And this is
751 * because of the padding node which we added to LEB Y. The
752 * user-visible effect of this which I once observed and
753 * analysed is that we cannot mount the file-system with
756 * So obviously, to make sure that situation does not happen we
757 * should free min. I/O unit B in LEB Y completely and the last
758 * used min. I/O unit in LEB Y should be A. This is basically
759 * what the below code tries to do.
761 while (offs > min_io_unit)
762 drop_last_node(sleb, &offs);
766 len = c->leb_size - offs;
768 clean_buf(c, &buf, lnum, &offs, &len);
769 ubifs_end_scan(c, sleb, lnum, offs);
771 err = fix_unclean_leb(c, sleb, start);
778 /* Re-scan the corrupted data with verbose messages */
779 ubifs_err(c, "corruption %d", ret);
780 ubifs_scan_a_node(c, buf, len, lnum, offs, 0);
782 ubifs_scanned_corruption(c, lnum, offs, buf);
785 ubifs_err(c, "LEB %d scanning failed", lnum);
786 ubifs_scan_destroy(sleb);
791 * get_cs_sqnum - get commit start sequence number.
792 * @c: UBIFS file-system description object
793 * @lnum: LEB number of commit start node
794 * @offs: offset of commit start node
795 * @cs_sqnum: commit start sequence number is returned here
797 * This function returns %0 on success and a negative error code on failure.
799 static int get_cs_sqnum(struct ubifs_info *c, int lnum, int offs,
800 unsigned long long *cs_sqnum)
802 struct ubifs_cs_node *cs_node = NULL;
805 dbg_rcvry("at %d:%d", lnum, offs);
806 cs_node = kmalloc(UBIFS_CS_NODE_SZ, GFP_KERNEL);
809 if (c->leb_size - offs < UBIFS_CS_NODE_SZ)
811 err = ubifs_leb_read(c, lnum, (void *)cs_node, offs,
812 UBIFS_CS_NODE_SZ, 0);
813 if (err && err != -EBADMSG)
815 ret = ubifs_scan_a_node(c, cs_node, UBIFS_CS_NODE_SZ, lnum, offs, 0);
816 if (ret != SCANNED_A_NODE) {
817 ubifs_err(c, "Not a valid node");
820 if (cs_node->ch.node_type != UBIFS_CS_NODE) {
821 ubifs_err(c, "Not a CS node, type is %d", cs_node->ch.node_type);
824 if (le64_to_cpu(cs_node->cmt_no) != c->cmt_no) {
825 ubifs_err(c, "CS node cmt_no %llu != current cmt_no %llu",
826 (unsigned long long)le64_to_cpu(cs_node->cmt_no),
830 *cs_sqnum = le64_to_cpu(cs_node->ch.sqnum);
831 dbg_rcvry("commit start sqnum %llu", *cs_sqnum);
838 ubifs_err(c, "failed to get CS sqnum");
844 * ubifs_recover_log_leb - scan and recover a log LEB.
845 * @c: UBIFS file-system description object
848 * @sbuf: LEB-sized buffer to use
850 * This function does a scan of a LEB, but caters for errors that might have
851 * been caused by unclean reboots from which we are attempting to recover
852 * (assume that only the last log LEB can be corrupted by an unclean reboot).
854 * This function returns %0 on success and a negative error code on failure.
856 struct ubifs_scan_leb *ubifs_recover_log_leb(struct ubifs_info *c, int lnum,
857 int offs, void *sbuf)
859 struct ubifs_scan_leb *sleb;
862 dbg_rcvry("LEB %d", lnum);
863 next_lnum = lnum + 1;
864 if (next_lnum >= UBIFS_LOG_LNUM + c->log_lebs)
865 next_lnum = UBIFS_LOG_LNUM;
866 if (next_lnum != c->ltail_lnum) {
868 * We can only recover at the end of the log, so check that the
869 * next log LEB is empty or out of date.
871 sleb = ubifs_scan(c, next_lnum, 0, sbuf, 0);
874 if (sleb->nodes_cnt) {
875 struct ubifs_scan_node *snod;
876 unsigned long long cs_sqnum = c->cs_sqnum;
878 snod = list_entry(sleb->nodes.next,
879 struct ubifs_scan_node, list);
883 err = get_cs_sqnum(c, lnum, offs, &cs_sqnum);
885 ubifs_scan_destroy(sleb);
889 if (snod->sqnum > cs_sqnum) {
890 ubifs_err(c, "unrecoverable log corruption in LEB %d",
892 ubifs_scan_destroy(sleb);
893 return ERR_PTR(-EUCLEAN);
896 ubifs_scan_destroy(sleb);
898 return ubifs_recover_leb(c, lnum, offs, sbuf, -1);
902 * recover_head - recover a head.
903 * @c: UBIFS file-system description object
904 * @lnum: LEB number of head to recover
905 * @offs: offset of head to recover
906 * @sbuf: LEB-sized buffer to use
908 * This function ensures that there is no data on the flash at a head location.
910 * This function returns %0 on success and a negative error code on failure.
912 static int recover_head(struct ubifs_info *c, int lnum, int offs, void *sbuf)
914 int len = c->max_write_size, err;
916 if (offs + len > c->leb_size)
917 len = c->leb_size - offs;
922 /* Read at the head location and check it is empty flash */
923 err = ubifs_leb_read(c, lnum, sbuf, offs, len, 1);
924 if (err || !is_empty(sbuf, len)) {
925 dbg_rcvry("cleaning head at %d:%d", lnum, offs);
927 return ubifs_leb_unmap(c, lnum);
928 err = ubifs_leb_read(c, lnum, sbuf, 0, offs, 1);
931 return ubifs_leb_change(c, lnum, sbuf, offs);
938 * ubifs_recover_inl_heads - recover index and LPT heads.
939 * @c: UBIFS file-system description object
940 * @sbuf: LEB-sized buffer to use
942 * This function ensures that there is no data on the flash at the index and
943 * LPT head locations.
945 * This deals with the recovery of a half-completed journal commit. UBIFS is
946 * careful never to overwrite the last version of the index or the LPT. Because
947 * the index and LPT are wandering trees, data from a half-completed commit will
948 * not be referenced anywhere in UBIFS. The data will be either in LEBs that are
949 * assumed to be empty and will be unmapped anyway before use, or in the index
952 * This function returns %0 on success and a negative error code on failure.
954 int ubifs_recover_inl_heads(struct ubifs_info *c, void *sbuf)
958 ubifs_assert(c, !c->ro_mount || c->remounting_rw);
960 dbg_rcvry("checking index head at %d:%d", c->ihead_lnum, c->ihead_offs);
961 err = recover_head(c, c->ihead_lnum, c->ihead_offs, sbuf);
965 dbg_rcvry("checking LPT head at %d:%d", c->nhead_lnum, c->nhead_offs);
967 return recover_head(c, c->nhead_lnum, c->nhead_offs, sbuf);
971 * clean_an_unclean_leb - read and write a LEB to remove corruption.
972 * @c: UBIFS file-system description object
973 * @ucleb: unclean LEB information
974 * @sbuf: LEB-sized buffer to use
976 * This function reads a LEB up to a point pre-determined by the mount recovery,
977 * checks the nodes, and writes the result back to the flash, thereby cleaning
978 * off any following corruption, or non-fatal ECC errors.
980 * This function returns %0 on success and a negative error code on failure.
982 static int clean_an_unclean_leb(struct ubifs_info *c,
983 struct ubifs_unclean_leb *ucleb, void *sbuf)
985 int err, lnum = ucleb->lnum, offs = 0, len = ucleb->endpt, quiet = 1;
988 dbg_rcvry("LEB %d len %d", lnum, len);
991 /* Nothing to read, just unmap it */
992 return ubifs_leb_unmap(c, lnum);
995 err = ubifs_leb_read(c, lnum, buf, offs, len, 0);
996 if (err && err != -EBADMSG)
1004 /* Scan quietly until there is an error */
1005 ret = ubifs_scan_a_node(c, buf, len, lnum, offs, quiet);
1007 if (ret == SCANNED_A_NODE) {
1008 /* A valid node, and not a padding node */
1009 struct ubifs_ch *ch = buf;
1012 node_len = ALIGN(le32_to_cpu(ch->len), 8);
1020 /* Padding bytes or a valid padding node */
1027 if (ret == SCANNED_EMPTY_SPACE) {
1028 ubifs_err(c, "unexpected empty space at %d:%d",
1034 /* Redo the last scan but noisily */
1039 ubifs_scanned_corruption(c, lnum, offs, buf);
1043 /* Pad to min_io_size */
1044 len = ALIGN(ucleb->endpt, c->min_io_size);
1045 if (len > ucleb->endpt) {
1046 int pad_len = len - ALIGN(ucleb->endpt, 8);
1049 buf = c->sbuf + len - pad_len;
1050 ubifs_pad(c, buf, pad_len);
1054 /* Write back the LEB atomically */
1055 err = ubifs_leb_change(c, lnum, sbuf, len);
1059 dbg_rcvry("cleaned LEB %d", lnum);
1065 * ubifs_clean_lebs - clean LEBs recovered during read-only mount.
1066 * @c: UBIFS file-system description object
1067 * @sbuf: LEB-sized buffer to use
1069 * This function cleans a LEB identified during recovery that needs to be
1070 * written but was not because UBIFS was mounted read-only. This happens when
1071 * remounting to read-write mode.
1073 * This function returns %0 on success and a negative error code on failure.
1075 int ubifs_clean_lebs(struct ubifs_info *c, void *sbuf)
1077 dbg_rcvry("recovery");
1078 while (!list_empty(&c->unclean_leb_list)) {
1079 struct ubifs_unclean_leb *ucleb;
1082 ucleb = list_entry(c->unclean_leb_list.next,
1083 struct ubifs_unclean_leb, list);
1084 err = clean_an_unclean_leb(c, ucleb, sbuf);
1087 list_del(&ucleb->list);
1094 * grab_empty_leb - grab an empty LEB to use as GC LEB and run commit.
1095 * @c: UBIFS file-system description object
1097 * This is a helper function for 'ubifs_rcvry_gc_commit()' which grabs an empty
1098 * LEB to be used as GC LEB (@c->gc_lnum), and then runs the commit. Returns
1099 * zero in case of success and a negative error code in case of failure.
1101 static int grab_empty_leb(struct ubifs_info *c)
1106 * Note, it is very important to first search for an empty LEB and then
1107 * run the commit, not vice-versa. The reason is that there might be
1108 * only one empty LEB at the moment, the one which has been the
1109 * @c->gc_lnum just before the power cut happened. During the regular
1110 * UBIFS operation (not now) @c->gc_lnum is marked as "taken", so no
1111 * one but GC can grab it. But at this moment this single empty LEB is
1112 * not marked as taken, so if we run commit - what happens? Right, the
1113 * commit will grab it and write the index there. Remember that the
1114 * index always expands as long as there is free space, and it only
1115 * starts consolidating when we run out of space.
1117 * IOW, if we run commit now, we might not be able to find a free LEB
1120 lnum = ubifs_find_free_leb_for_idx(c);
1122 ubifs_err(c, "could not find an empty LEB");
1123 ubifs_dump_lprops(c);
1124 ubifs_dump_budg(c, &c->bi);
1128 /* Reset the index flag */
1129 err = ubifs_change_one_lp(c, lnum, LPROPS_NC, LPROPS_NC, 0,
1135 dbg_rcvry("found empty LEB %d, run commit", lnum);
1137 return ubifs_run_commit(c);
1141 * ubifs_rcvry_gc_commit - recover the GC LEB number and run the commit.
1142 * @c: UBIFS file-system description object
1144 * Out-of-place garbage collection requires always one empty LEB with which to
1145 * start garbage collection. The LEB number is recorded in c->gc_lnum and is
1146 * written to the master node on unmounting. In the case of an unclean unmount
1147 * the value of gc_lnum recorded in the master node is out of date and cannot
1148 * be used. Instead, recovery must allocate an empty LEB for this purpose.
1149 * However, there may not be enough empty space, in which case it must be
1150 * possible to GC the dirtiest LEB into the GC head LEB.
1152 * This function also runs the commit which causes the TNC updates from
1153 * size-recovery and orphans to be written to the flash. That is important to
1154 * ensure correct replay order for subsequent mounts.
1156 * This function returns %0 on success and a negative error code on failure.
1158 int ubifs_rcvry_gc_commit(struct ubifs_info *c)
1160 struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf;
1161 struct ubifs_lprops lp;
1164 dbg_rcvry("GC head LEB %d, offs %d", wbuf->lnum, wbuf->offs);
1167 if (wbuf->lnum == -1 || wbuf->offs == c->leb_size)
1168 return grab_empty_leb(c);
1170 err = ubifs_find_dirty_leb(c, &lp, wbuf->offs, 2);
1175 dbg_rcvry("could not find a dirty LEB");
1176 return grab_empty_leb(c);
1179 ubifs_assert(c, !(lp.flags & LPROPS_INDEX));
1180 ubifs_assert(c, lp.free + lp.dirty >= wbuf->offs);
1183 * We run the commit before garbage collection otherwise subsequent
1184 * mounts will see the GC and orphan deletion in a different order.
1186 dbg_rcvry("committing");
1187 err = ubifs_run_commit(c);
1191 dbg_rcvry("GC'ing LEB %d", lp.lnum);
1192 mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead);
1193 err = ubifs_garbage_collect_leb(c, &lp);
1195 int err2 = ubifs_wbuf_sync_nolock(wbuf);
1200 mutex_unlock(&wbuf->io_mutex);
1202 ubifs_err(c, "GC failed, error %d", err);
1208 ubifs_assert(c, err == LEB_RETAINED);
1209 if (err != LEB_RETAINED)
1212 err = ubifs_leb_unmap(c, c->gc_lnum);
1216 dbg_rcvry("allocated LEB %d for GC", lp.lnum);
1221 * struct size_entry - inode size information for recovery.
1222 * @rb: link in the RB-tree of sizes
1223 * @inum: inode number
1224 * @i_size: size on inode
1225 * @d_size: maximum size based on data nodes
1226 * @exists: indicates whether the inode exists
1227 * @inode: inode if pinned in memory awaiting rw mode to fix it
1235 struct inode *inode;
1239 * add_ino - add an entry to the size tree.
1240 * @c: UBIFS file-system description object
1241 * @inum: inode number
1242 * @i_size: size on inode
1243 * @d_size: maximum size based on data nodes
1244 * @exists: indicates whether the inode exists
1246 static int add_ino(struct ubifs_info *c, ino_t inum, loff_t i_size,
1247 loff_t d_size, int exists)
1249 struct rb_node **p = &c->size_tree.rb_node, *parent = NULL;
1250 struct size_entry *e;
1254 e = rb_entry(parent, struct size_entry, rb);
1258 p = &(*p)->rb_right;
1261 e = kzalloc(sizeof(struct size_entry), GFP_KERNEL);
1270 rb_link_node(&e->rb, parent, p);
1271 rb_insert_color(&e->rb, &c->size_tree);
1277 * find_ino - find an entry on the size tree.
1278 * @c: UBIFS file-system description object
1279 * @inum: inode number
1281 static struct size_entry *find_ino(struct ubifs_info *c, ino_t inum)
1283 struct rb_node *p = c->size_tree.rb_node;
1284 struct size_entry *e;
1287 e = rb_entry(p, struct size_entry, rb);
1290 else if (inum > e->inum)
1299 * remove_ino - remove an entry from the size tree.
1300 * @c: UBIFS file-system description object
1301 * @inum: inode number
1303 static void remove_ino(struct ubifs_info *c, ino_t inum)
1305 struct size_entry *e = find_ino(c, inum);
1309 rb_erase(&e->rb, &c->size_tree);
1314 * ubifs_destroy_size_tree - free resources related to the size tree.
1315 * @c: UBIFS file-system description object
1317 void ubifs_destroy_size_tree(struct ubifs_info *c)
1319 struct size_entry *e, *n;
1321 rbtree_postorder_for_each_entry_safe(e, n, &c->size_tree, rb) {
1326 c->size_tree = RB_ROOT;
1330 * ubifs_recover_size_accum - accumulate inode sizes for recovery.
1331 * @c: UBIFS file-system description object
1333 * @deletion: node is for a deletion
1334 * @new_size: inode size
1336 * This function has two purposes:
1337 * 1) to ensure there are no data nodes that fall outside the inode size
1338 * 2) to ensure there are no data nodes for inodes that do not exist
1339 * To accomplish those purposes, a rb-tree is constructed containing an entry
1340 * for each inode number in the journal that has not been deleted, and recording
1341 * the size from the inode node, the maximum size of any data node (also altered
1342 * by truncations) and a flag indicating a inode number for which no inode node
1343 * was present in the journal.
1345 * Note that there is still the possibility that there are data nodes that have
1346 * been committed that are beyond the inode size, however the only way to find
1347 * them would be to scan the entire index. Alternatively, some provision could
1348 * be made to record the size of inodes at the start of commit, which would seem
1349 * very cumbersome for a scenario that is quite unlikely and the only negative
1350 * consequence of which is wasted space.
1352 * This functions returns %0 on success and a negative error code on failure.
1354 int ubifs_recover_size_accum(struct ubifs_info *c, union ubifs_key *key,
1355 int deletion, loff_t new_size)
1357 ino_t inum = key_inum(c, key);
1358 struct size_entry *e;
1361 switch (key_type(c, key)) {
1364 remove_ino(c, inum);
1366 e = find_ino(c, inum);
1368 e->i_size = new_size;
1371 err = add_ino(c, inum, new_size, 0, 1);
1377 case UBIFS_DATA_KEY:
1378 e = find_ino(c, inum);
1380 if (new_size > e->d_size)
1381 e->d_size = new_size;
1383 err = add_ino(c, inum, 0, new_size, 0);
1388 case UBIFS_TRUN_KEY:
1389 e = find_ino(c, inum);
1391 e->d_size = new_size;
1398 * fix_size_in_place - fix inode size in place on flash.
1399 * @c: UBIFS file-system description object
1400 * @e: inode size information for recovery
1402 static int fix_size_in_place(struct ubifs_info *c, struct size_entry *e)
1404 struct ubifs_ino_node *ino = c->sbuf;
1406 union ubifs_key key;
1407 int err, lnum, offs, len;
1411 /* Locate the inode node LEB number and offset */
1412 ino_key_init(c, &key, e->inum);
1413 err = ubifs_tnc_locate(c, &key, ino, &lnum, &offs);
1417 * If the size recorded on the inode node is greater than the size that
1418 * was calculated from nodes in the journal then don't change the inode.
1420 i_size = le64_to_cpu(ino->size);
1421 if (i_size >= e->d_size)
1424 err = ubifs_leb_read(c, lnum, c->sbuf, 0, c->leb_size, 1);
1427 /* Change the size field and recalculate the CRC */
1428 ino = c->sbuf + offs;
1429 ino->size = cpu_to_le64(e->d_size);
1430 len = le32_to_cpu(ino->ch.len);
1431 crc = crc32(UBIFS_CRC32_INIT, (void *)ino + 8, len - 8);
1432 ino->ch.crc = cpu_to_le32(crc);
1433 /* Work out where data in the LEB ends and free space begins */
1435 len = c->leb_size - 1;
1436 while (p[len] == 0xff)
1438 len = ALIGN(len + 1, c->min_io_size);
1439 /* Atomically write the fixed LEB back again */
1440 err = ubifs_leb_change(c, lnum, c->sbuf, len);
1443 dbg_rcvry("inode %lu at %d:%d size %lld -> %lld",
1444 (unsigned long)e->inum, lnum, offs, i_size, e->d_size);
1448 ubifs_warn(c, "inode %lu failed to fix size %lld -> %lld error %d",
1449 (unsigned long)e->inum, e->i_size, e->d_size, err);
1454 * inode_fix_size - fix inode size
1455 * @c: UBIFS file-system description object
1456 * @e: inode size information for recovery
1458 static int inode_fix_size(struct ubifs_info *c, struct size_entry *e)
1460 struct inode *inode;
1461 struct ubifs_inode *ui;
1465 ubifs_assert(c, !e->inode);
1468 /* Remounting rw, pick up inode we stored earlier */
1471 inode = ubifs_iget(c->vfs_sb, e->inum);
1473 return PTR_ERR(inode);
1475 if (inode->i_size >= e->d_size) {
1477 * The original inode in the index already has a size
1478 * big enough, nothing to do
1484 dbg_rcvry("ino %lu size %lld -> %lld",
1485 (unsigned long)e->inum,
1486 inode->i_size, e->d_size);
1488 ui = ubifs_inode(inode);
1490 inode->i_size = e->d_size;
1491 ui->ui_size = e->d_size;
1492 ui->synced_i_size = e->d_size;
1498 * In readonly mode just keep the inode pinned in memory until we go
1499 * readwrite. In readwrite mode write the inode to the journal with the
1505 err = ubifs_jnl_write_inode(c, inode);
1512 rb_erase(&e->rb, &c->size_tree);
1519 * ubifs_recover_size - recover inode size.
1520 * @c: UBIFS file-system description object
1521 * @in_place: If true, do a in-place size fixup
1523 * This function attempts to fix inode size discrepancies identified by the
1524 * 'ubifs_recover_size_accum()' function.
1526 * This functions returns %0 on success and a negative error code on failure.
1528 int ubifs_recover_size(struct ubifs_info *c, bool in_place)
1530 struct rb_node *this = rb_first(&c->size_tree);
1533 struct size_entry *e;
1536 e = rb_entry(this, struct size_entry, rb);
1538 this = rb_next(this);
1541 union ubifs_key key;
1543 ino_key_init(c, &key, e->inum);
1544 err = ubifs_tnc_lookup(c, &key, c->sbuf);
1545 if (err && err != -ENOENT)
1547 if (err == -ENOENT) {
1548 /* Remove data nodes that have no inode */
1549 dbg_rcvry("removing ino %lu",
1550 (unsigned long)e->inum);
1551 err = ubifs_tnc_remove_ino(c, e->inum);
1555 struct ubifs_ino_node *ino = c->sbuf;
1558 e->i_size = le64_to_cpu(ino->size);
1562 if (e->exists && e->i_size < e->d_size) {
1563 ubifs_assert(c, !(c->ro_mount && in_place));
1566 * We found data that is outside the found inode size,
1567 * fixup the inode size
1571 err = fix_size_in_place(c, e);
1576 err = inode_fix_size(c, e);
1583 rb_erase(&e->rb, &c->size_tree);