1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
14 #include "xfs_mount.h"
15 #include "xfs_defer.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
19 #include "xfs_log_priv.h"
20 #include "xfs_log_recover.h"
21 #include "xfs_trans_priv.h"
22 #include "xfs_alloc.h"
23 #include "xfs_ialloc.h"
24 #include "xfs_trace.h"
25 #include "xfs_icache.h"
26 #include "xfs_error.h"
27 #include "xfs_buf_item.h"
29 #define BLK_AVG(blk1, blk2) ((blk1+blk2) >> 1)
36 xlog_clear_stale_blocks(
41 xlog_recover_check_summary(
44 #define xlog_recover_check_summary(log)
47 xlog_do_recovery_pass(
48 struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
51 * Sector aligned buffer routines for buffer create/read/write/access
55 * Verify the log-relative block number and length in basic blocks are valid for
56 * an operation involving the given XFS log buffer. Returns true if the fields
57 * are valid, false otherwise.
65 if (blk_no < 0 || blk_no >= log->l_logBBsize)
67 if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
73 * Allocate a buffer to hold log data. The buffer needs to be able to map to
74 * a range of nbblks basic blocks at any valid offset within the log.
81 int align_mask = xfs_buftarg_dma_alignment(log->l_targ);
84 * Pass log block 0 since we don't have an addr yet, buffer will be
87 if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, 0, nbblks))) {
88 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
94 * We do log I/O in units of log sectors (a power-of-2 multiple of the
95 * basic block size), so we round up the requested size to accommodate
96 * the basic blocks required for complete log sectors.
98 * In addition, the buffer may be used for a non-sector-aligned block
99 * offset, in which case an I/O of the requested size could extend
100 * beyond the end of the buffer. If the requested size is only 1 basic
101 * block it will never straddle a sector boundary, so this won't be an
102 * issue. Nor will this be a problem if the log I/O is done in basic
103 * blocks (sector size 1). But otherwise we extend the buffer by one
104 * extra log sector to ensure there's space to accommodate this
107 if (nbblks > 1 && log->l_sectBBsize > 1)
108 nbblks += log->l_sectBBsize;
109 nbblks = round_up(nbblks, log->l_sectBBsize);
110 return kmem_alloc_io(BBTOB(nbblks), align_mask, KM_MAYFAIL | KM_ZERO);
114 * Return the address of the start of the given block number's data
115 * in a log buffer. The buffer covers a log sector-aligned region.
117 static inline unsigned int
122 return BBTOB(blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1));
135 if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, blk_no, nbblks))) {
137 "Invalid log block/length (0x%llx, 0x%x) for buffer",
139 return -EFSCORRUPTED;
142 blk_no = round_down(blk_no, log->l_sectBBsize);
143 nbblks = round_up(nbblks, log->l_sectBBsize);
146 error = xfs_rw_bdev(log->l_targ->bt_bdev, log->l_logBBstart + blk_no,
147 BBTOB(nbblks), data, op);
148 if (error && !XFS_FORCED_SHUTDOWN(log->l_mp)) {
150 "log recovery %s I/O error at daddr 0x%llx len %d error %d",
151 op == REQ_OP_WRITE ? "write" : "read",
152 blk_no, nbblks, error);
164 return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
177 error = xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
179 *offset = data + xlog_align(log, blk_no);
190 return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_WRITE);
195 * dump debug superblock and log record information
198 xlog_header_check_dump(
200 xlog_rec_header_t *head)
202 xfs_debug(mp, "%s: SB : uuid = %pU, fmt = %d",
203 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
204 xfs_debug(mp, " log : uuid = %pU, fmt = %d",
205 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
208 #define xlog_header_check_dump(mp, head)
212 * check log record header for recovery
215 xlog_header_check_recover(
217 xlog_rec_header_t *head)
219 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
222 * IRIX doesn't write the h_fmt field and leaves it zeroed
223 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
224 * a dirty log created in IRIX.
226 if (XFS_IS_CORRUPT(mp, head->h_fmt != cpu_to_be32(XLOG_FMT))) {
228 "dirty log written in incompatible format - can't recover");
229 xlog_header_check_dump(mp, head);
230 return -EFSCORRUPTED;
232 if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
233 &head->h_fs_uuid))) {
235 "dirty log entry has mismatched uuid - can't recover");
236 xlog_header_check_dump(mp, head);
237 return -EFSCORRUPTED;
243 * read the head block of the log and check the header
246 xlog_header_check_mount(
248 xlog_rec_header_t *head)
250 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
252 if (uuid_is_null(&head->h_fs_uuid)) {
254 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
255 * h_fs_uuid is null, we assume this log was last mounted
256 * by IRIX and continue.
258 xfs_warn(mp, "null uuid in log - IRIX style log");
259 } else if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
260 &head->h_fs_uuid))) {
261 xfs_warn(mp, "log has mismatched uuid - can't recover");
262 xlog_header_check_dump(mp, head);
263 return -EFSCORRUPTED;
274 * We're not going to bother about retrying
275 * this during recovery. One strike!
277 if (!XFS_FORCED_SHUTDOWN(bp->b_mount)) {
278 xfs_buf_ioerror_alert(bp, __this_address);
279 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
284 * On v5 supers, a bli could be attached to update the metadata LSN.
288 xfs_buf_item_relse(bp);
289 ASSERT(bp->b_log_item == NULL);
296 * This routine finds (to an approximation) the first block in the physical
297 * log which contains the given cycle. It uses a binary search algorithm.
298 * Note that the algorithm can not be perfect because the disk will not
299 * necessarily be perfect.
302 xlog_find_cycle_start(
305 xfs_daddr_t first_blk,
306 xfs_daddr_t *last_blk,
316 mid_blk = BLK_AVG(first_blk, end_blk);
317 while (mid_blk != first_blk && mid_blk != end_blk) {
318 error = xlog_bread(log, mid_blk, 1, buffer, &offset);
321 mid_cycle = xlog_get_cycle(offset);
322 if (mid_cycle == cycle)
323 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
325 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
326 mid_blk = BLK_AVG(first_blk, end_blk);
328 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
329 (mid_blk == end_blk && mid_blk-1 == first_blk));
337 * Check that a range of blocks does not contain stop_on_cycle_no.
338 * Fill in *new_blk with the block offset where such a block is
339 * found, or with -1 (an invalid block number) if there is no such
340 * block in the range. The scan needs to occur from front to back
341 * and the pointer into the region must be updated since a later
342 * routine will need to perform another test.
345 xlog_find_verify_cycle(
347 xfs_daddr_t start_blk,
349 uint stop_on_cycle_no,
350 xfs_daddr_t *new_blk)
360 * Greedily allocate a buffer big enough to handle the full
361 * range of basic blocks we'll be examining. If that fails,
362 * try a smaller size. We need to be able to read at least
363 * a log sector, or we're out of luck.
365 bufblks = 1 << ffs(nbblks);
366 while (bufblks > log->l_logBBsize)
368 while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
370 if (bufblks < log->l_sectBBsize)
374 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
377 bcount = min(bufblks, (start_blk + nbblks - i));
379 error = xlog_bread(log, i, bcount, buffer, &buf);
383 for (j = 0; j < bcount; j++) {
384 cycle = xlog_get_cycle(buf);
385 if (cycle == stop_on_cycle_no) {
402 * Potentially backup over partial log record write.
404 * In the typical case, last_blk is the number of the block directly after
405 * a good log record. Therefore, we subtract one to get the block number
406 * of the last block in the given buffer. extra_bblks contains the number
407 * of blocks we would have read on a previous read. This happens when the
408 * last log record is split over the end of the physical log.
410 * extra_bblks is the number of blocks potentially verified on a previous
411 * call to this routine.
414 xlog_find_verify_log_record(
416 xfs_daddr_t start_blk,
417 xfs_daddr_t *last_blk,
423 xlog_rec_header_t *head = NULL;
426 int num_blks = *last_blk - start_blk;
429 ASSERT(start_blk != 0 || *last_blk != start_blk);
431 buffer = xlog_alloc_buffer(log, num_blks);
433 buffer = xlog_alloc_buffer(log, 1);
438 error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
441 offset += ((num_blks - 1) << BBSHIFT);
444 for (i = (*last_blk) - 1; i >= 0; i--) {
446 /* valid log record not found */
448 "Log inconsistent (didn't find previous header)");
450 error = -EFSCORRUPTED;
455 error = xlog_bread(log, i, 1, buffer, &offset);
460 head = (xlog_rec_header_t *)offset;
462 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
470 * We hit the beginning of the physical log & still no header. Return
471 * to caller. If caller can handle a return of -1, then this routine
472 * will be called again for the end of the physical log.
480 * We have the final block of the good log (the first block
481 * of the log record _before_ the head. So we check the uuid.
483 if ((error = xlog_header_check_mount(log->l_mp, head)))
487 * We may have found a log record header before we expected one.
488 * last_blk will be the 1st block # with a given cycle #. We may end
489 * up reading an entire log record. In this case, we don't want to
490 * reset last_blk. Only when last_blk points in the middle of a log
491 * record do we update last_blk.
493 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
494 uint h_size = be32_to_cpu(head->h_size);
496 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
497 if (h_size % XLOG_HEADER_CYCLE_SIZE)
503 if (*last_blk - i + extra_bblks !=
504 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
513 * Head is defined to be the point of the log where the next log write
514 * could go. This means that incomplete LR writes at the end are
515 * eliminated when calculating the head. We aren't guaranteed that previous
516 * LR have complete transactions. We only know that a cycle number of
517 * current cycle number -1 won't be present in the log if we start writing
518 * from our current block number.
520 * last_blk contains the block number of the first block with a given
523 * Return: zero if normal, non-zero if error.
528 xfs_daddr_t *return_head_blk)
532 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
534 uint first_half_cycle, last_half_cycle;
536 int error, log_bbnum = log->l_logBBsize;
538 /* Is the end of the log device zeroed? */
539 error = xlog_find_zeroed(log, &first_blk);
541 xfs_warn(log->l_mp, "empty log check failed");
545 *return_head_blk = first_blk;
547 /* Is the whole lot zeroed? */
549 /* Linux XFS shouldn't generate totally zeroed logs -
550 * mkfs etc write a dummy unmount record to a fresh
551 * log so we can store the uuid in there
553 xfs_warn(log->l_mp, "totally zeroed log");
559 first_blk = 0; /* get cycle # of 1st block */
560 buffer = xlog_alloc_buffer(log, 1);
564 error = xlog_bread(log, 0, 1, buffer, &offset);
566 goto out_free_buffer;
568 first_half_cycle = xlog_get_cycle(offset);
570 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
571 error = xlog_bread(log, last_blk, 1, buffer, &offset);
573 goto out_free_buffer;
575 last_half_cycle = xlog_get_cycle(offset);
576 ASSERT(last_half_cycle != 0);
579 * If the 1st half cycle number is equal to the last half cycle number,
580 * then the entire log is stamped with the same cycle number. In this
581 * case, head_blk can't be set to zero (which makes sense). The below
582 * math doesn't work out properly with head_blk equal to zero. Instead,
583 * we set it to log_bbnum which is an invalid block number, but this
584 * value makes the math correct. If head_blk doesn't changed through
585 * all the tests below, *head_blk is set to zero at the very end rather
586 * than log_bbnum. In a sense, log_bbnum and zero are the same block
587 * in a circular file.
589 if (first_half_cycle == last_half_cycle) {
591 * In this case we believe that the entire log should have
592 * cycle number last_half_cycle. We need to scan backwards
593 * from the end verifying that there are no holes still
594 * containing last_half_cycle - 1. If we find such a hole,
595 * then the start of that hole will be the new head. The
596 * simple case looks like
597 * x | x ... | x - 1 | x
598 * Another case that fits this picture would be
599 * x | x + 1 | x ... | x
600 * In this case the head really is somewhere at the end of the
601 * log, as one of the latest writes at the beginning was
604 * x | x + 1 | x ... | x - 1 | x
605 * This is really the combination of the above two cases, and
606 * the head has to end up at the start of the x-1 hole at the
609 * In the 256k log case, we will read from the beginning to the
610 * end of the log and search for cycle numbers equal to x-1.
611 * We don't worry about the x+1 blocks that we encounter,
612 * because we know that they cannot be the head since the log
615 head_blk = log_bbnum;
616 stop_on_cycle = last_half_cycle - 1;
619 * In this case we want to find the first block with cycle
620 * number matching last_half_cycle. We expect the log to be
622 * x + 1 ... | x ... | x
623 * The first block with cycle number x (last_half_cycle) will
624 * be where the new head belongs. First we do a binary search
625 * for the first occurrence of last_half_cycle. The binary
626 * search may not be totally accurate, so then we scan back
627 * from there looking for occurrences of last_half_cycle before
628 * us. If that backwards scan wraps around the beginning of
629 * the log, then we look for occurrences of last_half_cycle - 1
630 * at the end of the log. The cases we're looking for look
632 * v binary search stopped here
633 * x + 1 ... | x | x + 1 | x ... | x
634 * ^ but we want to locate this spot
636 * <---------> less than scan distance
637 * x + 1 ... | x ... | x - 1 | x
638 * ^ we want to locate this spot
640 stop_on_cycle = last_half_cycle;
641 error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
644 goto out_free_buffer;
648 * Now validate the answer. Scan back some number of maximum possible
649 * blocks and make sure each one has the expected cycle number. The
650 * maximum is determined by the total possible amount of buffering
651 * in the in-core log. The following number can be made tighter if
652 * we actually look at the block size of the filesystem.
654 num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
655 if (head_blk >= num_scan_bblks) {
657 * We are guaranteed that the entire check can be performed
660 start_blk = head_blk - num_scan_bblks;
661 if ((error = xlog_find_verify_cycle(log,
662 start_blk, num_scan_bblks,
663 stop_on_cycle, &new_blk)))
664 goto out_free_buffer;
667 } else { /* need to read 2 parts of log */
669 * We are going to scan backwards in the log in two parts.
670 * First we scan the physical end of the log. In this part
671 * of the log, we are looking for blocks with cycle number
672 * last_half_cycle - 1.
673 * If we find one, then we know that the log starts there, as
674 * we've found a hole that didn't get written in going around
675 * the end of the physical log. The simple case for this is
676 * x + 1 ... | x ... | x - 1 | x
677 * <---------> less than scan distance
678 * If all of the blocks at the end of the log have cycle number
679 * last_half_cycle, then we check the blocks at the start of
680 * the log looking for occurrences of last_half_cycle. If we
681 * find one, then our current estimate for the location of the
682 * first occurrence of last_half_cycle is wrong and we move
683 * back to the hole we've found. This case looks like
684 * x + 1 ... | x | x + 1 | x ...
685 * ^ binary search stopped here
686 * Another case we need to handle that only occurs in 256k
688 * x + 1 ... | x ... | x+1 | x ...
689 * ^ binary search stops here
690 * In a 256k log, the scan at the end of the log will see the
691 * x + 1 blocks. We need to skip past those since that is
692 * certainly not the head of the log. By searching for
693 * last_half_cycle-1 we accomplish that.
695 ASSERT(head_blk <= INT_MAX &&
696 (xfs_daddr_t) num_scan_bblks >= head_blk);
697 start_blk = log_bbnum - (num_scan_bblks - head_blk);
698 if ((error = xlog_find_verify_cycle(log, start_blk,
699 num_scan_bblks - (int)head_blk,
700 (stop_on_cycle - 1), &new_blk)))
701 goto out_free_buffer;
708 * Scan beginning of log now. The last part of the physical
709 * log is good. This scan needs to verify that it doesn't find
710 * the last_half_cycle.
713 ASSERT(head_blk <= INT_MAX);
714 if ((error = xlog_find_verify_cycle(log,
715 start_blk, (int)head_blk,
716 stop_on_cycle, &new_blk)))
717 goto out_free_buffer;
724 * Now we need to make sure head_blk is not pointing to a block in
725 * the middle of a log record.
727 num_scan_bblks = XLOG_REC_SHIFT(log);
728 if (head_blk >= num_scan_bblks) {
729 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
731 /* start ptr at last block ptr before head_blk */
732 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
736 goto out_free_buffer;
739 ASSERT(head_blk <= INT_MAX);
740 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
742 goto out_free_buffer;
744 /* We hit the beginning of the log during our search */
745 start_blk = log_bbnum - (num_scan_bblks - head_blk);
747 ASSERT(start_blk <= INT_MAX &&
748 (xfs_daddr_t) log_bbnum-start_blk >= 0);
749 ASSERT(head_blk <= INT_MAX);
750 error = xlog_find_verify_log_record(log, start_blk,
751 &new_blk, (int)head_blk);
755 goto out_free_buffer;
756 if (new_blk != log_bbnum)
759 goto out_free_buffer;
763 if (head_blk == log_bbnum)
764 *return_head_blk = 0;
766 *return_head_blk = head_blk;
768 * When returning here, we have a good block number. Bad block
769 * means that during a previous crash, we didn't have a clean break
770 * from cycle number N to cycle number N-1. In this case, we need
771 * to find the first block with cycle number N-1.
778 xfs_warn(log->l_mp, "failed to find log head");
783 * Seek backwards in the log for log record headers.
785 * Given a starting log block, walk backwards until we find the provided number
786 * of records or hit the provided tail block. The return value is the number of
787 * records encountered or a negative error code. The log block and buffer
788 * pointer of the last record seen are returned in rblk and rhead respectively.
791 xlog_rseek_logrec_hdr(
793 xfs_daddr_t head_blk,
794 xfs_daddr_t tail_blk,
798 struct xlog_rec_header **rhead,
810 * Walk backwards from the head block until we hit the tail or the first
813 end_blk = head_blk > tail_blk ? tail_blk : 0;
814 for (i = (int) head_blk - 1; i >= end_blk; i--) {
815 error = xlog_bread(log, i, 1, buffer, &offset);
819 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
821 *rhead = (struct xlog_rec_header *) offset;
822 if (++found == count)
828 * If we haven't hit the tail block or the log record header count,
829 * start looking again from the end of the physical log. Note that
830 * callers can pass head == tail if the tail is not yet known.
832 if (tail_blk >= head_blk && found != count) {
833 for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
834 error = xlog_bread(log, i, 1, buffer, &offset);
838 if (*(__be32 *)offset ==
839 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
842 *rhead = (struct xlog_rec_header *) offset;
843 if (++found == count)
856 * Seek forward in the log for log record headers.
858 * Given head and tail blocks, walk forward from the tail block until we find
859 * the provided number of records or hit the head block. The return value is the
860 * number of records encountered or a negative error code. The log block and
861 * buffer pointer of the last record seen are returned in rblk and rhead
865 xlog_seek_logrec_hdr(
867 xfs_daddr_t head_blk,
868 xfs_daddr_t tail_blk,
872 struct xlog_rec_header **rhead,
884 * Walk forward from the tail block until we hit the head or the last
887 end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
888 for (i = (int) tail_blk; i <= end_blk; i++) {
889 error = xlog_bread(log, i, 1, buffer, &offset);
893 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
895 *rhead = (struct xlog_rec_header *) offset;
896 if (++found == count)
902 * If we haven't hit the head block or the log record header count,
903 * start looking again from the start of the physical log.
905 if (tail_blk > head_blk && found != count) {
906 for (i = 0; i < (int) head_blk; i++) {
907 error = xlog_bread(log, i, 1, buffer, &offset);
911 if (*(__be32 *)offset ==
912 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
915 *rhead = (struct xlog_rec_header *) offset;
916 if (++found == count)
929 * Calculate distance from head to tail (i.e., unused space in the log).
934 xfs_daddr_t head_blk,
935 xfs_daddr_t tail_blk)
937 if (head_blk < tail_blk)
938 return tail_blk - head_blk;
940 return tail_blk + (log->l_logBBsize - head_blk);
944 * Verify the log tail. This is particularly important when torn or incomplete
945 * writes have been detected near the front of the log and the head has been
946 * walked back accordingly.
948 * We also have to handle the case where the tail was pinned and the head
949 * blocked behind the tail right before a crash. If the tail had been pushed
950 * immediately prior to the crash and the subsequent checkpoint was only
951 * partially written, it's possible it overwrote the last referenced tail in the
952 * log with garbage. This is not a coherency problem because the tail must have
953 * been pushed before it can be overwritten, but appears as log corruption to
954 * recovery because we have no way to know the tail was updated if the
955 * subsequent checkpoint didn't write successfully.
957 * Therefore, CRC check the log from tail to head. If a failure occurs and the
958 * offending record is within max iclog bufs from the head, walk the tail
959 * forward and retry until a valid tail is found or corruption is detected out
960 * of the range of a possible overwrite.
965 xfs_daddr_t head_blk,
966 xfs_daddr_t *tail_blk,
969 struct xlog_rec_header *thead;
971 xfs_daddr_t first_bad;
974 xfs_daddr_t tmp_tail;
975 xfs_daddr_t orig_tail = *tail_blk;
977 buffer = xlog_alloc_buffer(log, 1);
982 * Make sure the tail points to a record (returns positive count on
985 error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
986 &tmp_tail, &thead, &wrapped);
989 if (*tail_blk != tmp_tail)
990 *tail_blk = tmp_tail;
993 * Run a CRC check from the tail to the head. We can't just check
994 * MAX_ICLOGS records past the tail because the tail may point to stale
995 * blocks cleared during the search for the head/tail. These blocks are
996 * overwritten with zero-length records and thus record count is not a
997 * reliable indicator of the iclog state before a crash.
1000 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1001 XLOG_RECOVER_CRCPASS, &first_bad);
1002 while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1006 * Is corruption within range of the head? If so, retry from
1007 * the next record. Otherwise return an error.
1009 tail_distance = xlog_tail_distance(log, head_blk, first_bad);
1010 if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
1013 /* skip to the next record; returns positive count on success */
1014 error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
1015 buffer, &tmp_tail, &thead, &wrapped);
1019 *tail_blk = tmp_tail;
1021 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1022 XLOG_RECOVER_CRCPASS, &first_bad);
1025 if (!error && *tail_blk != orig_tail)
1027 "Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1028 orig_tail, *tail_blk);
1035 * Detect and trim torn writes from the head of the log.
1037 * Storage without sector atomicity guarantees can result in torn writes in the
1038 * log in the event of a crash. Our only means to detect this scenario is via
1039 * CRC verification. While we can't always be certain that CRC verification
1040 * failure is due to a torn write vs. an unrelated corruption, we do know that
1041 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1042 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1043 * the log and treat failures in this range as torn writes as a matter of
1044 * policy. In the event of CRC failure, the head is walked back to the last good
1045 * record in the log and the tail is updated from that record and verified.
1050 xfs_daddr_t *head_blk, /* in/out: unverified head */
1051 xfs_daddr_t *tail_blk, /* out: tail block */
1053 xfs_daddr_t *rhead_blk, /* start blk of last record */
1054 struct xlog_rec_header **rhead, /* ptr to last record */
1055 bool *wrapped) /* last rec. wraps phys. log */
1057 struct xlog_rec_header *tmp_rhead;
1059 xfs_daddr_t first_bad;
1060 xfs_daddr_t tmp_rhead_blk;
1066 * Check the head of the log for torn writes. Search backwards from the
1067 * head until we hit the tail or the maximum number of log record I/Os
1068 * that could have been in flight at one time. Use a temporary buffer so
1069 * we don't trash the rhead/buffer pointers from the caller.
1071 tmp_buffer = xlog_alloc_buffer(log, 1);
1074 error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1075 XLOG_MAX_ICLOGS, tmp_buffer,
1076 &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1077 kmem_free(tmp_buffer);
1082 * Now run a CRC verification pass over the records starting at the
1083 * block found above to the current head. If a CRC failure occurs, the
1084 * log block of the first bad record is saved in first_bad.
1086 error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1087 XLOG_RECOVER_CRCPASS, &first_bad);
1088 if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1090 * We've hit a potential torn write. Reset the error and warn
1095 "Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1096 first_bad, *head_blk);
1099 * Get the header block and buffer pointer for the last good
1100 * record before the bad record.
1102 * Note that xlog_find_tail() clears the blocks at the new head
1103 * (i.e., the records with invalid CRC) if the cycle number
1104 * matches the the current cycle.
1106 found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1107 buffer, rhead_blk, rhead, wrapped);
1110 if (found == 0) /* XXX: right thing to do here? */
1114 * Reset the head block to the starting block of the first bad
1115 * log record and set the tail block based on the last good
1118 * Bail out if the updated head/tail match as this indicates
1119 * possible corruption outside of the acceptable
1120 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1122 *head_blk = first_bad;
1123 *tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1124 if (*head_blk == *tail_blk) {
1132 return xlog_verify_tail(log, *head_blk, tail_blk,
1133 be32_to_cpu((*rhead)->h_size));
1137 * We need to make sure we handle log wrapping properly, so we can't use the
1138 * calculated logbno directly. Make sure it wraps to the correct bno inside the
1141 * The log is limited to 32 bit sizes, so we use the appropriate modulus
1142 * operation here and cast it back to a 64 bit daddr on return.
1144 static inline xfs_daddr_t
1151 div_s64_rem(bno, log->l_logBBsize, &mod);
1156 * Check whether the head of the log points to an unmount record. In other
1157 * words, determine whether the log is clean. If so, update the in-core state
1161 xlog_check_unmount_rec(
1163 xfs_daddr_t *head_blk,
1164 xfs_daddr_t *tail_blk,
1165 struct xlog_rec_header *rhead,
1166 xfs_daddr_t rhead_blk,
1170 struct xlog_op_header *op_head;
1171 xfs_daddr_t umount_data_blk;
1172 xfs_daddr_t after_umount_blk;
1180 * Look for unmount record. If we find it, then we know there was a
1181 * clean unmount. Since 'i' could be the last block in the physical
1182 * log, we convert to a log block before comparing to the head_blk.
1184 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1185 * below. We won't want to clear the unmount record if there is one, so
1186 * we pass the lsn of the unmount record rather than the block after it.
1188 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1189 int h_size = be32_to_cpu(rhead->h_size);
1190 int h_version = be32_to_cpu(rhead->h_version);
1192 if ((h_version & XLOG_VERSION_2) &&
1193 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1194 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1195 if (h_size % XLOG_HEADER_CYCLE_SIZE)
1204 after_umount_blk = xlog_wrap_logbno(log,
1205 rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1207 if (*head_blk == after_umount_blk &&
1208 be32_to_cpu(rhead->h_num_logops) == 1) {
1209 umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1210 error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1214 op_head = (struct xlog_op_header *)offset;
1215 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1217 * Set tail and last sync so that newly written log
1218 * records will point recovery to after the current
1221 xlog_assign_atomic_lsn(&log->l_tail_lsn,
1222 log->l_curr_cycle, after_umount_blk);
1223 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1224 log->l_curr_cycle, after_umount_blk);
1225 *tail_blk = after_umount_blk;
1237 xfs_daddr_t head_blk,
1238 struct xlog_rec_header *rhead,
1239 xfs_daddr_t rhead_blk,
1243 * Reset log values according to the state of the log when we
1244 * crashed. In the case where head_blk == 0, we bump curr_cycle
1245 * one because the next write starts a new cycle rather than
1246 * continuing the cycle of the last good log record. At this
1247 * point we have guaranteed that all partial log records have been
1248 * accounted for. Therefore, we know that the last good log record
1249 * written was complete and ended exactly on the end boundary
1250 * of the physical log.
1252 log->l_prev_block = rhead_blk;
1253 log->l_curr_block = (int)head_blk;
1254 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1256 log->l_curr_cycle++;
1257 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1258 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1259 xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1260 BBTOB(log->l_curr_block));
1261 xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1262 BBTOB(log->l_curr_block));
1266 * Find the sync block number or the tail of the log.
1268 * This will be the block number of the last record to have its
1269 * associated buffers synced to disk. Every log record header has
1270 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
1271 * to get a sync block number. The only concern is to figure out which
1272 * log record header to believe.
1274 * The following algorithm uses the log record header with the largest
1275 * lsn. The entire log record does not need to be valid. We only care
1276 * that the header is valid.
1278 * We could speed up search by using current head_blk buffer, but it is not
1284 xfs_daddr_t *head_blk,
1285 xfs_daddr_t *tail_blk)
1287 xlog_rec_header_t *rhead;
1288 char *offset = NULL;
1291 xfs_daddr_t rhead_blk;
1293 bool wrapped = false;
1297 * Find previous log record
1299 if ((error = xlog_find_head(log, head_blk)))
1301 ASSERT(*head_blk < INT_MAX);
1303 buffer = xlog_alloc_buffer(log, 1);
1306 if (*head_blk == 0) { /* special case */
1307 error = xlog_bread(log, 0, 1, buffer, &offset);
1311 if (xlog_get_cycle(offset) == 0) {
1313 /* leave all other log inited values alone */
1319 * Search backwards through the log looking for the log record header
1320 * block. This wraps all the way back around to the head so something is
1321 * seriously wrong if we can't find it.
1323 error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1324 &rhead_blk, &rhead, &wrapped);
1328 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1329 error = -EFSCORRUPTED;
1332 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1335 * Set the log state based on the current head record.
1337 xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1338 tail_lsn = atomic64_read(&log->l_tail_lsn);
1341 * Look for an unmount record at the head of the log. This sets the log
1342 * state to determine whether recovery is necessary.
1344 error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1345 rhead_blk, buffer, &clean);
1350 * Verify the log head if the log is not clean (e.g., we have anything
1351 * but an unmount record at the head). This uses CRC verification to
1352 * detect and trim torn writes. If discovered, CRC failures are
1353 * considered torn writes and the log head is trimmed accordingly.
1355 * Note that we can only run CRC verification when the log is dirty
1356 * because there's no guarantee that the log data behind an unmount
1357 * record is compatible with the current architecture.
1360 xfs_daddr_t orig_head = *head_blk;
1362 error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1363 &rhead_blk, &rhead, &wrapped);
1367 /* update in-core state again if the head changed */
1368 if (*head_blk != orig_head) {
1369 xlog_set_state(log, *head_blk, rhead, rhead_blk,
1371 tail_lsn = atomic64_read(&log->l_tail_lsn);
1372 error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1373 rhead, rhead_blk, buffer,
1381 * Note that the unmount was clean. If the unmount was not clean, we
1382 * need to know this to rebuild the superblock counters from the perag
1383 * headers if we have a filesystem using non-persistent counters.
1386 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1389 * Make sure that there are no blocks in front of the head
1390 * with the same cycle number as the head. This can happen
1391 * because we allow multiple outstanding log writes concurrently,
1392 * and the later writes might make it out before earlier ones.
1394 * We use the lsn from before modifying it so that we'll never
1395 * overwrite the unmount record after a clean unmount.
1397 * Do this only if we are going to recover the filesystem
1399 * NOTE: This used to say "if (!readonly)"
1400 * However on Linux, we can & do recover a read-only filesystem.
1401 * We only skip recovery if NORECOVERY is specified on mount,
1402 * in which case we would not be here.
1404 * But... if the -device- itself is readonly, just skip this.
1405 * We can't recover this device anyway, so it won't matter.
1407 if (!xfs_readonly_buftarg(log->l_targ))
1408 error = xlog_clear_stale_blocks(log, tail_lsn);
1414 xfs_warn(log->l_mp, "failed to locate log tail");
1419 * Is the log zeroed at all?
1421 * The last binary search should be changed to perform an X block read
1422 * once X becomes small enough. You can then search linearly through
1423 * the X blocks. This will cut down on the number of reads we need to do.
1425 * If the log is partially zeroed, this routine will pass back the blkno
1426 * of the first block with cycle number 0. It won't have a complete LR
1430 * 0 => the log is completely written to
1431 * 1 => use *blk_no as the first block of the log
1432 * <0 => error has occurred
1437 xfs_daddr_t *blk_no)
1441 uint first_cycle, last_cycle;
1442 xfs_daddr_t new_blk, last_blk, start_blk;
1443 xfs_daddr_t num_scan_bblks;
1444 int error, log_bbnum = log->l_logBBsize;
1448 /* check totally zeroed log */
1449 buffer = xlog_alloc_buffer(log, 1);
1452 error = xlog_bread(log, 0, 1, buffer, &offset);
1454 goto out_free_buffer;
1456 first_cycle = xlog_get_cycle(offset);
1457 if (first_cycle == 0) { /* completely zeroed log */
1463 /* check partially zeroed log */
1464 error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1466 goto out_free_buffer;
1468 last_cycle = xlog_get_cycle(offset);
1469 if (last_cycle != 0) { /* log completely written to */
1474 /* we have a partially zeroed log */
1475 last_blk = log_bbnum-1;
1476 error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1478 goto out_free_buffer;
1481 * Validate the answer. Because there is no way to guarantee that
1482 * the entire log is made up of log records which are the same size,
1483 * we scan over the defined maximum blocks. At this point, the maximum
1484 * is not chosen to mean anything special. XXXmiken
1486 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1487 ASSERT(num_scan_bblks <= INT_MAX);
1489 if (last_blk < num_scan_bblks)
1490 num_scan_bblks = last_blk;
1491 start_blk = last_blk - num_scan_bblks;
1494 * We search for any instances of cycle number 0 that occur before
1495 * our current estimate of the head. What we're trying to detect is
1496 * 1 ... | 0 | 1 | 0...
1497 * ^ binary search ends here
1499 if ((error = xlog_find_verify_cycle(log, start_blk,
1500 (int)num_scan_bblks, 0, &new_blk)))
1501 goto out_free_buffer;
1506 * Potentially backup over partial log record write. We don't need
1507 * to search the end of the log because we know it is zero.
1509 error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1513 goto out_free_buffer;
1524 * These are simple subroutines used by xlog_clear_stale_blocks() below
1525 * to initialize a buffer full of empty log record headers and write
1526 * them into the log.
1537 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1539 memset(buf, 0, BBSIZE);
1540 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1541 recp->h_cycle = cpu_to_be32(cycle);
1542 recp->h_version = cpu_to_be32(
1543 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1544 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1545 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1546 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1547 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1551 xlog_write_log_records(
1562 int sectbb = log->l_sectBBsize;
1563 int end_block = start_block + blocks;
1569 * Greedily allocate a buffer big enough to handle the full
1570 * range of basic blocks to be written. If that fails, try
1571 * a smaller size. We need to be able to write at least a
1572 * log sector, or we're out of luck.
1574 bufblks = 1 << ffs(blocks);
1575 while (bufblks > log->l_logBBsize)
1577 while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1579 if (bufblks < sectbb)
1583 /* We may need to do a read at the start to fill in part of
1584 * the buffer in the starting sector not covered by the first
1587 balign = round_down(start_block, sectbb);
1588 if (balign != start_block) {
1589 error = xlog_bread_noalign(log, start_block, 1, buffer);
1591 goto out_free_buffer;
1593 j = start_block - balign;
1596 for (i = start_block; i < end_block; i += bufblks) {
1597 int bcount, endcount;
1599 bcount = min(bufblks, end_block - start_block);
1600 endcount = bcount - j;
1602 /* We may need to do a read at the end to fill in part of
1603 * the buffer in the final sector not covered by the write.
1604 * If this is the same sector as the above read, skip it.
1606 ealign = round_down(end_block, sectbb);
1607 if (j == 0 && (start_block + endcount > ealign)) {
1608 error = xlog_bread_noalign(log, ealign, sectbb,
1609 buffer + BBTOB(ealign - start_block));
1615 offset = buffer + xlog_align(log, start_block);
1616 for (; j < endcount; j++) {
1617 xlog_add_record(log, offset, cycle, i+j,
1618 tail_cycle, tail_block);
1621 error = xlog_bwrite(log, start_block, endcount, buffer);
1624 start_block += endcount;
1634 * This routine is called to blow away any incomplete log writes out
1635 * in front of the log head. We do this so that we won't become confused
1636 * if we come up, write only a little bit more, and then crash again.
1637 * If we leave the partial log records out there, this situation could
1638 * cause us to think those partial writes are valid blocks since they
1639 * have the current cycle number. We get rid of them by overwriting them
1640 * with empty log records with the old cycle number rather than the
1643 * The tail lsn is passed in rather than taken from
1644 * the log so that we will not write over the unmount record after a
1645 * clean unmount in a 512 block log. Doing so would leave the log without
1646 * any valid log records in it until a new one was written. If we crashed
1647 * during that time we would not be able to recover.
1650 xlog_clear_stale_blocks(
1654 int tail_cycle, head_cycle;
1655 int tail_block, head_block;
1656 int tail_distance, max_distance;
1660 tail_cycle = CYCLE_LSN(tail_lsn);
1661 tail_block = BLOCK_LSN(tail_lsn);
1662 head_cycle = log->l_curr_cycle;
1663 head_block = log->l_curr_block;
1666 * Figure out the distance between the new head of the log
1667 * and the tail. We want to write over any blocks beyond the
1668 * head that we may have written just before the crash, but
1669 * we don't want to overwrite the tail of the log.
1671 if (head_cycle == tail_cycle) {
1673 * The tail is behind the head in the physical log,
1674 * so the distance from the head to the tail is the
1675 * distance from the head to the end of the log plus
1676 * the distance from the beginning of the log to the
1679 if (XFS_IS_CORRUPT(log->l_mp,
1680 head_block < tail_block ||
1681 head_block >= log->l_logBBsize))
1682 return -EFSCORRUPTED;
1683 tail_distance = tail_block + (log->l_logBBsize - head_block);
1686 * The head is behind the tail in the physical log,
1687 * so the distance from the head to the tail is just
1688 * the tail block minus the head block.
1690 if (XFS_IS_CORRUPT(log->l_mp,
1691 head_block >= tail_block ||
1692 head_cycle != tail_cycle + 1))
1693 return -EFSCORRUPTED;
1694 tail_distance = tail_block - head_block;
1698 * If the head is right up against the tail, we can't clear
1701 if (tail_distance <= 0) {
1702 ASSERT(tail_distance == 0);
1706 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1708 * Take the smaller of the maximum amount of outstanding I/O
1709 * we could have and the distance to the tail to clear out.
1710 * We take the smaller so that we don't overwrite the tail and
1711 * we don't waste all day writing from the head to the tail
1714 max_distance = min(max_distance, tail_distance);
1716 if ((head_block + max_distance) <= log->l_logBBsize) {
1718 * We can stomp all the blocks we need to without
1719 * wrapping around the end of the log. Just do it
1720 * in a single write. Use the cycle number of the
1721 * current cycle minus one so that the log will look like:
1724 error = xlog_write_log_records(log, (head_cycle - 1),
1725 head_block, max_distance, tail_cycle,
1731 * We need to wrap around the end of the physical log in
1732 * order to clear all the blocks. Do it in two separate
1733 * I/Os. The first write should be from the head to the
1734 * end of the physical log, and it should use the current
1735 * cycle number minus one just like above.
1737 distance = log->l_logBBsize - head_block;
1738 error = xlog_write_log_records(log, (head_cycle - 1),
1739 head_block, distance, tail_cycle,
1746 * Now write the blocks at the start of the physical log.
1747 * This writes the remainder of the blocks we want to clear.
1748 * It uses the current cycle number since we're now on the
1749 * same cycle as the head so that we get:
1750 * n ... n ... | n - 1 ...
1751 * ^^^^^ blocks we're writing
1753 distance = max_distance - (log->l_logBBsize - head_block);
1754 error = xlog_write_log_records(log, head_cycle, 0, distance,
1755 tail_cycle, tail_block);
1764 * Release the recovered intent item in the AIL that matches the given intent
1765 * type and intent id.
1768 xlog_recover_release_intent(
1770 unsigned short intent_type,
1773 struct xfs_ail_cursor cur;
1774 struct xfs_log_item *lip;
1775 struct xfs_ail *ailp = log->l_ailp;
1777 spin_lock(&ailp->ail_lock);
1778 for (lip = xfs_trans_ail_cursor_first(ailp, &cur, 0); lip != NULL;
1779 lip = xfs_trans_ail_cursor_next(ailp, &cur)) {
1780 if (lip->li_type != intent_type)
1782 if (!lip->li_ops->iop_match(lip, intent_id))
1785 spin_unlock(&ailp->ail_lock);
1786 lip->li_ops->iop_release(lip);
1787 spin_lock(&ailp->ail_lock);
1791 xfs_trans_ail_cursor_done(&cur);
1792 spin_unlock(&ailp->ail_lock);
1795 /******************************************************************************
1797 * Log recover routines
1799 ******************************************************************************
1801 static const struct xlog_recover_item_ops *xlog_recover_item_ops[] = {
1803 &xlog_inode_item_ops,
1804 &xlog_dquot_item_ops,
1805 &xlog_quotaoff_item_ops,
1806 &xlog_icreate_item_ops,
1817 static const struct xlog_recover_item_ops *
1819 struct xlog_recover_item *item)
1823 for (i = 0; i < ARRAY_SIZE(xlog_recover_item_ops); i++)
1824 if (ITEM_TYPE(item) == xlog_recover_item_ops[i]->item_type)
1825 return xlog_recover_item_ops[i];
1831 * Sort the log items in the transaction.
1833 * The ordering constraints are defined by the inode allocation and unlink
1834 * behaviour. The rules are:
1836 * 1. Every item is only logged once in a given transaction. Hence it
1837 * represents the last logged state of the item. Hence ordering is
1838 * dependent on the order in which operations need to be performed so
1839 * required initial conditions are always met.
1841 * 2. Cancelled buffers are recorded in pass 1 in a separate table and
1842 * there's nothing to replay from them so we can simply cull them
1843 * from the transaction. However, we can't do that until after we've
1844 * replayed all the other items because they may be dependent on the
1845 * cancelled buffer and replaying the cancelled buffer can remove it
1846 * form the cancelled buffer table. Hence they have tobe done last.
1848 * 3. Inode allocation buffers must be replayed before inode items that
1849 * read the buffer and replay changes into it. For filesystems using the
1850 * ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1851 * treated the same as inode allocation buffers as they create and
1852 * initialise the buffers directly.
1854 * 4. Inode unlink buffers must be replayed after inode items are replayed.
1855 * This ensures that inodes are completely flushed to the inode buffer
1856 * in a "free" state before we remove the unlinked inode list pointer.
1858 * Hence the ordering needs to be inode allocation buffers first, inode items
1859 * second, inode unlink buffers third and cancelled buffers last.
1861 * But there's a problem with that - we can't tell an inode allocation buffer
1862 * apart from a regular buffer, so we can't separate them. We can, however,
1863 * tell an inode unlink buffer from the others, and so we can separate them out
1864 * from all the other buffers and move them to last.
1866 * Hence, 4 lists, in order from head to tail:
1867 * - buffer_list for all buffers except cancelled/inode unlink buffers
1868 * - item_list for all non-buffer items
1869 * - inode_buffer_list for inode unlink buffers
1870 * - cancel_list for the cancelled buffers
1872 * Note that we add objects to the tail of the lists so that first-to-last
1873 * ordering is preserved within the lists. Adding objects to the head of the
1874 * list means when we traverse from the head we walk them in last-to-first
1875 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1876 * but for all other items there may be specific ordering that we need to
1880 xlog_recover_reorder_trans(
1882 struct xlog_recover *trans,
1885 struct xlog_recover_item *item, *n;
1887 LIST_HEAD(sort_list);
1888 LIST_HEAD(cancel_list);
1889 LIST_HEAD(buffer_list);
1890 LIST_HEAD(inode_buffer_list);
1891 LIST_HEAD(item_list);
1893 list_splice_init(&trans->r_itemq, &sort_list);
1894 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1895 enum xlog_recover_reorder fate = XLOG_REORDER_ITEM_LIST;
1897 item->ri_ops = xlog_find_item_ops(item);
1898 if (!item->ri_ops) {
1900 "%s: unrecognized type of log operation (%d)",
1901 __func__, ITEM_TYPE(item));
1904 * return the remaining items back to the transaction
1905 * item list so they can be freed in caller.
1907 if (!list_empty(&sort_list))
1908 list_splice_init(&sort_list, &trans->r_itemq);
1909 error = -EFSCORRUPTED;
1913 if (item->ri_ops->reorder)
1914 fate = item->ri_ops->reorder(item);
1917 case XLOG_REORDER_BUFFER_LIST:
1918 list_move_tail(&item->ri_list, &buffer_list);
1920 case XLOG_REORDER_CANCEL_LIST:
1921 trace_xfs_log_recover_item_reorder_head(log,
1923 list_move(&item->ri_list, &cancel_list);
1925 case XLOG_REORDER_INODE_BUFFER_LIST:
1926 list_move(&item->ri_list, &inode_buffer_list);
1928 case XLOG_REORDER_ITEM_LIST:
1929 trace_xfs_log_recover_item_reorder_tail(log,
1931 list_move_tail(&item->ri_list, &item_list);
1936 ASSERT(list_empty(&sort_list));
1937 if (!list_empty(&buffer_list))
1938 list_splice(&buffer_list, &trans->r_itemq);
1939 if (!list_empty(&item_list))
1940 list_splice_tail(&item_list, &trans->r_itemq);
1941 if (!list_empty(&inode_buffer_list))
1942 list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1943 if (!list_empty(&cancel_list))
1944 list_splice_tail(&cancel_list, &trans->r_itemq);
1953 const struct xfs_buf_ops *ops)
1955 if (!xlog_is_buffer_cancelled(log, blkno, len))
1956 xfs_buf_readahead(log->l_mp->m_ddev_targp, blkno, len, ops);
1960 xlog_recover_items_pass2(
1962 struct xlog_recover *trans,
1963 struct list_head *buffer_list,
1964 struct list_head *item_list)
1966 struct xlog_recover_item *item;
1969 list_for_each_entry(item, item_list, ri_list) {
1970 trace_xfs_log_recover_item_recover(log, trans, item,
1971 XLOG_RECOVER_PASS2);
1973 if (item->ri_ops->commit_pass2)
1974 error = item->ri_ops->commit_pass2(log, buffer_list,
1975 item, trans->r_lsn);
1984 * Perform the transaction.
1986 * If the transaction modifies a buffer or inode, do it now. Otherwise,
1987 * EFIs and EFDs get queued up by adding entries into the AIL for them.
1990 xlog_recover_commit_trans(
1992 struct xlog_recover *trans,
1994 struct list_head *buffer_list)
1997 int items_queued = 0;
1998 struct xlog_recover_item *item;
1999 struct xlog_recover_item *next;
2000 LIST_HEAD (ra_list);
2001 LIST_HEAD (done_list);
2003 #define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
2005 hlist_del_init(&trans->r_list);
2007 error = xlog_recover_reorder_trans(log, trans, pass);
2011 list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
2012 trace_xfs_log_recover_item_recover(log, trans, item, pass);
2015 case XLOG_RECOVER_PASS1:
2016 if (item->ri_ops->commit_pass1)
2017 error = item->ri_ops->commit_pass1(log, item);
2019 case XLOG_RECOVER_PASS2:
2020 if (item->ri_ops->ra_pass2)
2021 item->ri_ops->ra_pass2(log, item);
2022 list_move_tail(&item->ri_list, &ra_list);
2024 if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
2025 error = xlog_recover_items_pass2(log, trans,
2026 buffer_list, &ra_list);
2027 list_splice_tail_init(&ra_list, &done_list);
2041 if (!list_empty(&ra_list)) {
2043 error = xlog_recover_items_pass2(log, trans,
2044 buffer_list, &ra_list);
2045 list_splice_tail_init(&ra_list, &done_list);
2048 if (!list_empty(&done_list))
2049 list_splice_init(&done_list, &trans->r_itemq);
2055 xlog_recover_add_item(
2056 struct list_head *head)
2058 struct xlog_recover_item *item;
2060 item = kmem_zalloc(sizeof(struct xlog_recover_item), 0);
2061 INIT_LIST_HEAD(&item->ri_list);
2062 list_add_tail(&item->ri_list, head);
2066 xlog_recover_add_to_cont_trans(
2068 struct xlog_recover *trans,
2072 struct xlog_recover_item *item;
2073 char *ptr, *old_ptr;
2077 * If the transaction is empty, the header was split across this and the
2078 * previous record. Copy the rest of the header.
2080 if (list_empty(&trans->r_itemq)) {
2081 ASSERT(len <= sizeof(struct xfs_trans_header));
2082 if (len > sizeof(struct xfs_trans_header)) {
2083 xfs_warn(log->l_mp, "%s: bad header length", __func__);
2084 return -EFSCORRUPTED;
2087 xlog_recover_add_item(&trans->r_itemq);
2088 ptr = (char *)&trans->r_theader +
2089 sizeof(struct xfs_trans_header) - len;
2090 memcpy(ptr, dp, len);
2094 /* take the tail entry */
2095 item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2098 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
2099 old_len = item->ri_buf[item->ri_cnt-1].i_len;
2101 ptr = kmem_realloc(old_ptr, len + old_len, 0);
2102 memcpy(&ptr[old_len], dp, len);
2103 item->ri_buf[item->ri_cnt-1].i_len += len;
2104 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
2105 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
2110 * The next region to add is the start of a new region. It could be
2111 * a whole region or it could be the first part of a new region. Because
2112 * of this, the assumption here is that the type and size fields of all
2113 * format structures fit into the first 32 bits of the structure.
2115 * This works because all regions must be 32 bit aligned. Therefore, we
2116 * either have both fields or we have neither field. In the case we have
2117 * neither field, the data part of the region is zero length. We only have
2118 * a log_op_header and can throw away the header since a new one will appear
2119 * later. If we have at least 4 bytes, then we can determine how many regions
2120 * will appear in the current log item.
2123 xlog_recover_add_to_trans(
2125 struct xlog_recover *trans,
2129 struct xfs_inode_log_format *in_f; /* any will do */
2130 struct xlog_recover_item *item;
2135 if (list_empty(&trans->r_itemq)) {
2136 /* we need to catch log corruptions here */
2137 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
2138 xfs_warn(log->l_mp, "%s: bad header magic number",
2141 return -EFSCORRUPTED;
2144 if (len > sizeof(struct xfs_trans_header)) {
2145 xfs_warn(log->l_mp, "%s: bad header length", __func__);
2147 return -EFSCORRUPTED;
2151 * The transaction header can be arbitrarily split across op
2152 * records. If we don't have the whole thing here, copy what we
2153 * do have and handle the rest in the next record.
2155 if (len == sizeof(struct xfs_trans_header))
2156 xlog_recover_add_item(&trans->r_itemq);
2157 memcpy(&trans->r_theader, dp, len);
2161 ptr = kmem_alloc(len, 0);
2162 memcpy(ptr, dp, len);
2163 in_f = (struct xfs_inode_log_format *)ptr;
2165 /* take the tail entry */
2166 item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2168 if (item->ri_total != 0 &&
2169 item->ri_total == item->ri_cnt) {
2170 /* tail item is in use, get a new one */
2171 xlog_recover_add_item(&trans->r_itemq);
2172 item = list_entry(trans->r_itemq.prev,
2173 struct xlog_recover_item, ri_list);
2176 if (item->ri_total == 0) { /* first region to be added */
2177 if (in_f->ilf_size == 0 ||
2178 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
2180 "bad number of regions (%d) in inode log format",
2184 return -EFSCORRUPTED;
2187 item->ri_total = in_f->ilf_size;
2189 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
2193 if (item->ri_total <= item->ri_cnt) {
2195 "log item region count (%d) overflowed size (%d)",
2196 item->ri_cnt, item->ri_total);
2199 return -EFSCORRUPTED;
2202 /* Description region is ri_buf[0] */
2203 item->ri_buf[item->ri_cnt].i_addr = ptr;
2204 item->ri_buf[item->ri_cnt].i_len = len;
2206 trace_xfs_log_recover_item_add(log, trans, item, 0);
2211 * Free up any resources allocated by the transaction
2213 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2216 xlog_recover_free_trans(
2217 struct xlog_recover *trans)
2219 struct xlog_recover_item *item, *n;
2222 hlist_del_init(&trans->r_list);
2224 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2225 /* Free the regions in the item. */
2226 list_del(&item->ri_list);
2227 for (i = 0; i < item->ri_cnt; i++)
2228 kmem_free(item->ri_buf[i].i_addr);
2229 /* Free the item itself */
2230 kmem_free(item->ri_buf);
2233 /* Free the transaction recover structure */
2238 * On error or completion, trans is freed.
2241 xlog_recovery_process_trans(
2243 struct xlog_recover *trans,
2248 struct list_head *buffer_list)
2251 bool freeit = false;
2253 /* mask off ophdr transaction container flags */
2254 flags &= ~XLOG_END_TRANS;
2255 if (flags & XLOG_WAS_CONT_TRANS)
2256 flags &= ~XLOG_CONTINUE_TRANS;
2259 * Callees must not free the trans structure. We'll decide if we need to
2260 * free it or not based on the operation being done and it's result.
2263 /* expected flag values */
2265 case XLOG_CONTINUE_TRANS:
2266 error = xlog_recover_add_to_trans(log, trans, dp, len);
2268 case XLOG_WAS_CONT_TRANS:
2269 error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
2271 case XLOG_COMMIT_TRANS:
2272 error = xlog_recover_commit_trans(log, trans, pass,
2274 /* success or fail, we are now done with this transaction. */
2278 /* unexpected flag values */
2279 case XLOG_UNMOUNT_TRANS:
2280 /* just skip trans */
2281 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2284 case XLOG_START_TRANS:
2286 xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
2288 error = -EFSCORRUPTED;
2291 if (error || freeit)
2292 xlog_recover_free_trans(trans);
2297 * Lookup the transaction recovery structure associated with the ID in the
2298 * current ophdr. If the transaction doesn't exist and the start flag is set in
2299 * the ophdr, then allocate a new transaction for future ID matches to find.
2300 * Either way, return what we found during the lookup - an existing transaction
2303 STATIC struct xlog_recover *
2304 xlog_recover_ophdr_to_trans(
2305 struct hlist_head rhash[],
2306 struct xlog_rec_header *rhead,
2307 struct xlog_op_header *ohead)
2309 struct xlog_recover *trans;
2311 struct hlist_head *rhp;
2313 tid = be32_to_cpu(ohead->oh_tid);
2314 rhp = &rhash[XLOG_RHASH(tid)];
2315 hlist_for_each_entry(trans, rhp, r_list) {
2316 if (trans->r_log_tid == tid)
2321 * skip over non-start transaction headers - we could be
2322 * processing slack space before the next transaction starts
2324 if (!(ohead->oh_flags & XLOG_START_TRANS))
2327 ASSERT(be32_to_cpu(ohead->oh_len) == 0);
2330 * This is a new transaction so allocate a new recovery container to
2331 * hold the recovery ops that will follow.
2333 trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
2334 trans->r_log_tid = tid;
2335 trans->r_lsn = be64_to_cpu(rhead->h_lsn);
2336 INIT_LIST_HEAD(&trans->r_itemq);
2337 INIT_HLIST_NODE(&trans->r_list);
2338 hlist_add_head(&trans->r_list, rhp);
2341 * Nothing more to do for this ophdr. Items to be added to this new
2342 * transaction will be in subsequent ophdr containers.
2348 xlog_recover_process_ophdr(
2350 struct hlist_head rhash[],
2351 struct xlog_rec_header *rhead,
2352 struct xlog_op_header *ohead,
2356 struct list_head *buffer_list)
2358 struct xlog_recover *trans;
2362 /* Do we understand who wrote this op? */
2363 if (ohead->oh_clientid != XFS_TRANSACTION &&
2364 ohead->oh_clientid != XFS_LOG) {
2365 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2366 __func__, ohead->oh_clientid);
2368 return -EFSCORRUPTED;
2372 * Check the ophdr contains all the data it is supposed to contain.
2374 len = be32_to_cpu(ohead->oh_len);
2375 if (dp + len > end) {
2376 xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
2378 return -EFSCORRUPTED;
2381 trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
2383 /* nothing to do, so skip over this ophdr */
2388 * The recovered buffer queue is drained only once we know that all
2389 * recovery items for the current LSN have been processed. This is
2392 * - Buffer write submission updates the metadata LSN of the buffer.
2393 * - Log recovery skips items with a metadata LSN >= the current LSN of
2394 * the recovery item.
2395 * - Separate recovery items against the same metadata buffer can share
2396 * a current LSN. I.e., consider that the LSN of a recovery item is
2397 * defined as the starting LSN of the first record in which its
2398 * transaction appears, that a record can hold multiple transactions,
2399 * and/or that a transaction can span multiple records.
2401 * In other words, we are allowed to submit a buffer from log recovery
2402 * once per current LSN. Otherwise, we may incorrectly skip recovery
2403 * items and cause corruption.
2405 * We don't know up front whether buffers are updated multiple times per
2406 * LSN. Therefore, track the current LSN of each commit log record as it
2407 * is processed and drain the queue when it changes. Use commit records
2408 * because they are ordered correctly by the logging code.
2410 if (log->l_recovery_lsn != trans->r_lsn &&
2411 ohead->oh_flags & XLOG_COMMIT_TRANS) {
2412 error = xfs_buf_delwri_submit(buffer_list);
2415 log->l_recovery_lsn = trans->r_lsn;
2418 return xlog_recovery_process_trans(log, trans, dp, len,
2419 ohead->oh_flags, pass, buffer_list);
2423 * There are two valid states of the r_state field. 0 indicates that the
2424 * transaction structure is in a normal state. We have either seen the
2425 * start of the transaction or the last operation we added was not a partial
2426 * operation. If the last operation we added to the transaction was a
2427 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2429 * NOTE: skip LRs with 0 data length.
2432 xlog_recover_process_data(
2434 struct hlist_head rhash[],
2435 struct xlog_rec_header *rhead,
2438 struct list_head *buffer_list)
2440 struct xlog_op_header *ohead;
2445 end = dp + be32_to_cpu(rhead->h_len);
2446 num_logops = be32_to_cpu(rhead->h_num_logops);
2448 /* check the log format matches our own - else we can't recover */
2449 if (xlog_header_check_recover(log->l_mp, rhead))
2452 trace_xfs_log_recover_record(log, rhead, pass);
2453 while ((dp < end) && num_logops) {
2455 ohead = (struct xlog_op_header *)dp;
2456 dp += sizeof(*ohead);
2459 /* errors will abort recovery */
2460 error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
2461 dp, end, pass, buffer_list);
2465 dp += be32_to_cpu(ohead->oh_len);
2471 /* Take all the collected deferred ops and finish them in order. */
2473 xlog_finish_defer_ops(
2474 struct xfs_trans *parent_tp)
2476 struct xfs_mount *mp = parent_tp->t_mountp;
2477 struct xfs_trans *tp;
2483 * We're finishing the defer_ops that accumulated as a result of
2484 * recovering unfinished intent items during log recovery. We
2485 * reserve an itruncate transaction because it is the largest
2486 * permanent transaction type. Since we're the only user of the fs
2487 * right now, take 93% (15/16) of the available free blocks. Use
2488 * weird math to avoid a 64-bit division.
2490 freeblks = percpu_counter_sum(&mp->m_fdblocks);
2493 resblks = min_t(int64_t, UINT_MAX, freeblks);
2494 resblks = (resblks * 15) >> 4;
2495 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, resblks,
2496 0, XFS_TRANS_RESERVE, &tp);
2499 /* transfer all collected dfops to this transaction */
2500 xfs_defer_move(tp, parent_tp);
2502 return xfs_trans_commit(tp);
2505 /* Is this log item a deferred action intent? */
2506 static inline bool xlog_item_is_intent(struct xfs_log_item *lip)
2508 return lip->li_ops->iop_recover != NULL &&
2509 lip->li_ops->iop_match != NULL;
2513 * When this is called, all of the log intent items which did not have
2514 * corresponding log done items should be in the AIL. What we do now
2515 * is update the data structures associated with each one.
2517 * Since we process the log intent items in normal transactions, they
2518 * will be removed at some point after the commit. This prevents us
2519 * from just walking down the list processing each one. We'll use a
2520 * flag in the intent item to skip those that we've already processed
2521 * and use the AIL iteration mechanism's generation count to try to
2522 * speed this up at least a bit.
2524 * When we start, we know that the intents are the only things in the
2525 * AIL. As we process them, however, other items are added to the
2529 xlog_recover_process_intents(
2532 struct xfs_trans *parent_tp;
2533 struct xfs_ail_cursor cur;
2534 struct xfs_log_item *lip;
2535 struct xfs_ail *ailp;
2537 #if defined(DEBUG) || defined(XFS_WARN)
2542 * The intent recovery handlers commit transactions to complete recovery
2543 * for individual intents, but any new deferred operations that are
2544 * queued during that process are held off until the very end. The
2545 * purpose of this transaction is to serve as a container for deferred
2546 * operations. Each intent recovery handler must transfer dfops here
2547 * before its local transaction commits, and we'll finish the entire
2550 error = xfs_trans_alloc_empty(log->l_mp, &parent_tp);
2555 spin_lock(&ailp->ail_lock);
2556 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2557 #if defined(DEBUG) || defined(XFS_WARN)
2558 last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
2560 while (lip != NULL) {
2562 * We're done when we see something other than an intent.
2563 * There should be no intents left in the AIL now.
2565 if (!xlog_item_is_intent(lip)) {
2567 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2568 ASSERT(!xlog_item_is_intent(lip));
2574 * We should never see a redo item with a LSN higher than
2575 * the last transaction we found in the log at the start
2578 ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
2581 * NOTE: If your intent processing routine can create more
2582 * deferred ops, you /must/ attach them to the transaction in
2583 * this routine or else those subsequent intents will get
2584 * replayed in the wrong order!
2586 if (!test_and_set_bit(XFS_LI_RECOVERED, &lip->li_flags)) {
2587 spin_unlock(&ailp->ail_lock);
2588 error = lip->li_ops->iop_recover(lip, parent_tp);
2589 spin_lock(&ailp->ail_lock);
2593 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2596 xfs_trans_ail_cursor_done(&cur);
2597 spin_unlock(&ailp->ail_lock);
2599 error = xlog_finish_defer_ops(parent_tp);
2600 xfs_trans_cancel(parent_tp);
2606 * A cancel occurs when the mount has failed and we're bailing out.
2607 * Release all pending log intent items so they don't pin the AIL.
2610 xlog_recover_cancel_intents(
2613 struct xfs_log_item *lip;
2614 struct xfs_ail_cursor cur;
2615 struct xfs_ail *ailp;
2618 spin_lock(&ailp->ail_lock);
2619 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2620 while (lip != NULL) {
2622 * We're done when we see something other than an intent.
2623 * There should be no intents left in the AIL now.
2625 if (!xlog_item_is_intent(lip)) {
2627 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2628 ASSERT(!xlog_item_is_intent(lip));
2633 spin_unlock(&ailp->ail_lock);
2634 lip->li_ops->iop_release(lip);
2635 spin_lock(&ailp->ail_lock);
2636 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2639 xfs_trans_ail_cursor_done(&cur);
2640 spin_unlock(&ailp->ail_lock);
2644 * This routine performs a transaction to null out a bad inode pointer
2645 * in an agi unlinked inode hash bucket.
2648 xlog_recover_clear_agi_bucket(
2650 xfs_agnumber_t agno,
2659 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
2663 error = xfs_read_agi(mp, tp, agno, &agibp);
2667 agi = agibp->b_addr;
2668 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
2669 offset = offsetof(xfs_agi_t, agi_unlinked) +
2670 (sizeof(xfs_agino_t) * bucket);
2671 xfs_trans_log_buf(tp, agibp, offset,
2672 (offset + sizeof(xfs_agino_t) - 1));
2674 error = xfs_trans_commit(tp);
2680 xfs_trans_cancel(tp);
2682 xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
2687 xlog_recover_process_one_iunlink(
2688 struct xfs_mount *mp,
2689 xfs_agnumber_t agno,
2693 struct xfs_buf *ibp;
2694 struct xfs_dinode *dip;
2695 struct xfs_inode *ip;
2699 ino = XFS_AGINO_TO_INO(mp, agno, agino);
2700 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
2705 * Get the on disk inode to find the next inode in the bucket.
2707 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0);
2711 xfs_iflags_clear(ip, XFS_IRECOVERY);
2712 ASSERT(VFS_I(ip)->i_nlink == 0);
2713 ASSERT(VFS_I(ip)->i_mode != 0);
2715 /* setup for the next pass */
2716 agino = be32_to_cpu(dip->di_next_unlinked);
2720 * Prevent any DMAPI event from being sent when the reference on
2721 * the inode is dropped.
2723 ip->i_d.di_dmevmask = 0;
2732 * We can't read in the inode this bucket points to, or this inode
2733 * is messed up. Just ditch this bucket of inodes. We will lose
2734 * some inodes and space, but at least we won't hang.
2736 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
2737 * clear the inode pointer in the bucket.
2739 xlog_recover_clear_agi_bucket(mp, agno, bucket);
2744 * Recover AGI unlinked lists
2746 * This is called during recovery to process any inodes which we unlinked but
2747 * not freed when the system crashed. These inodes will be on the lists in the
2748 * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
2749 * any inodes found on the lists. Each inode is removed from the lists when it
2750 * has been fully truncated and is freed. The freeing of the inode and its
2751 * removal from the list must be atomic.
2753 * If everything we touch in the agi processing loop is already in memory, this
2754 * loop can hold the cpu for a long time. It runs without lock contention,
2755 * memory allocation contention, the need wait for IO, etc, and so will run
2756 * until we either run out of inodes to process, run low on memory or we run out
2759 * This behaviour is bad for latency on single CPU and non-preemptible kernels,
2760 * and can prevent other filesytem work (such as CIL pushes) from running. This
2761 * can lead to deadlocks if the recovery process runs out of log reservation
2762 * space. Hence we need to yield the CPU when there is other kernel work
2763 * scheduled on this CPU to ensure other scheduled work can run without undue
2767 xlog_recover_process_iunlinks(
2771 xfs_agnumber_t agno;
2780 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
2782 * Find the agi for this ag.
2784 error = xfs_read_agi(mp, NULL, agno, &agibp);
2787 * AGI is b0rked. Don't process it.
2789 * We should probably mark the filesystem as corrupt
2790 * after we've recovered all the ag's we can....
2795 * Unlock the buffer so that it can be acquired in the normal
2796 * course of the transaction to truncate and free each inode.
2797 * Because we are not racing with anyone else here for the AGI
2798 * buffer, we don't even need to hold it locked to read the
2799 * initial unlinked bucket entries out of the buffer. We keep
2800 * buffer reference though, so that it stays pinned in memory
2801 * while we need the buffer.
2803 agi = agibp->b_addr;
2804 xfs_buf_unlock(agibp);
2806 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
2807 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2808 while (agino != NULLAGINO) {
2809 agino = xlog_recover_process_one_iunlink(mp,
2810 agno, agino, bucket);
2814 xfs_buf_rele(agibp);
2820 struct xlog_rec_header *rhead,
2826 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
2827 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
2828 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
2832 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
2833 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
2834 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
2835 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2836 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2837 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
2844 * CRC check, unpack and process a log record.
2847 xlog_recover_process(
2849 struct hlist_head rhash[],
2850 struct xlog_rec_header *rhead,
2853 struct list_head *buffer_list)
2855 __le32 old_crc = rhead->h_crc;
2858 crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
2861 * Nothing else to do if this is a CRC verification pass. Just return
2862 * if this a record with a non-zero crc. Unfortunately, mkfs always
2863 * sets old_crc to 0 so we must consider this valid even on v5 supers.
2864 * Otherwise, return EFSBADCRC on failure so the callers up the stack
2865 * know precisely what failed.
2867 if (pass == XLOG_RECOVER_CRCPASS) {
2868 if (old_crc && crc != old_crc)
2874 * We're in the normal recovery path. Issue a warning if and only if the
2875 * CRC in the header is non-zero. This is an advisory warning and the
2876 * zero CRC check prevents warnings from being emitted when upgrading
2877 * the kernel from one that does not add CRCs by default.
2879 if (crc != old_crc) {
2880 if (old_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
2881 xfs_alert(log->l_mp,
2882 "log record CRC mismatch: found 0x%x, expected 0x%x.",
2883 le32_to_cpu(old_crc),
2885 xfs_hex_dump(dp, 32);
2889 * If the filesystem is CRC enabled, this mismatch becomes a
2890 * fatal log corruption failure.
2892 if (xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
2893 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
2894 return -EFSCORRUPTED;
2898 xlog_unpack_data(rhead, dp, log);
2900 return xlog_recover_process_data(log, rhash, rhead, dp, pass,
2905 xlog_valid_rec_header(
2907 struct xlog_rec_header *rhead,
2912 if (XFS_IS_CORRUPT(log->l_mp,
2913 rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)))
2914 return -EFSCORRUPTED;
2915 if (XFS_IS_CORRUPT(log->l_mp,
2916 (!rhead->h_version ||
2917 (be32_to_cpu(rhead->h_version) &
2918 (~XLOG_VERSION_OKBITS))))) {
2919 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
2920 __func__, be32_to_cpu(rhead->h_version));
2921 return -EFSCORRUPTED;
2924 /* LR body must have data or it wouldn't have been written */
2925 hlen = be32_to_cpu(rhead->h_len);
2926 if (XFS_IS_CORRUPT(log->l_mp, hlen <= 0 || hlen > INT_MAX))
2927 return -EFSCORRUPTED;
2928 if (XFS_IS_CORRUPT(log->l_mp,
2929 blkno > log->l_logBBsize || blkno > INT_MAX))
2930 return -EFSCORRUPTED;
2935 * Read the log from tail to head and process the log records found.
2936 * Handle the two cases where the tail and head are in the same cycle
2937 * and where the active portion of the log wraps around the end of
2938 * the physical log separately. The pass parameter is passed through
2939 * to the routines called to process the data and is not looked at
2943 xlog_do_recovery_pass(
2945 xfs_daddr_t head_blk,
2946 xfs_daddr_t tail_blk,
2948 xfs_daddr_t *first_bad) /* out: first bad log rec */
2950 xlog_rec_header_t *rhead;
2951 xfs_daddr_t blk_no, rblk_no;
2952 xfs_daddr_t rhead_blk;
2955 int error = 0, h_size, h_len;
2957 int bblks, split_bblks;
2958 int hblks, split_hblks, wrapped_hblks;
2960 struct hlist_head rhash[XLOG_RHASH_SIZE];
2961 LIST_HEAD (buffer_list);
2963 ASSERT(head_blk != tail_blk);
2964 blk_no = rhead_blk = tail_blk;
2966 for (i = 0; i < XLOG_RHASH_SIZE; i++)
2967 INIT_HLIST_HEAD(&rhash[i]);
2970 * Read the header of the tail block and get the iclog buffer size from
2971 * h_size. Use this to tell how many sectors make up the log header.
2973 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
2975 * When using variable length iclogs, read first sector of
2976 * iclog header and extract the header size from it. Get a
2977 * new hbp that is the correct size.
2979 hbp = xlog_alloc_buffer(log, 1);
2983 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
2987 rhead = (xlog_rec_header_t *)offset;
2988 error = xlog_valid_rec_header(log, rhead, tail_blk);
2993 * xfsprogs has a bug where record length is based on lsunit but
2994 * h_size (iclog size) is hardcoded to 32k. Now that we
2995 * unconditionally CRC verify the unmount record, this means the
2996 * log buffer can be too small for the record and cause an
2999 * Detect this condition here. Use lsunit for the buffer size as
3000 * long as this looks like the mkfs case. Otherwise, return an
3001 * error to avoid a buffer overrun.
3003 h_size = be32_to_cpu(rhead->h_size);
3004 h_len = be32_to_cpu(rhead->h_len);
3005 if (h_len > h_size) {
3006 if (h_len <= log->l_mp->m_logbsize &&
3007 be32_to_cpu(rhead->h_num_logops) == 1) {
3009 "invalid iclog size (%d bytes), using lsunit (%d bytes)",
3010 h_size, log->l_mp->m_logbsize);
3011 h_size = log->l_mp->m_logbsize;
3013 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW,
3015 error = -EFSCORRUPTED;
3020 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3021 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3022 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3023 if (h_size % XLOG_HEADER_CYCLE_SIZE)
3026 hbp = xlog_alloc_buffer(log, hblks);
3031 ASSERT(log->l_sectBBsize == 1);
3033 hbp = xlog_alloc_buffer(log, 1);
3034 h_size = XLOG_BIG_RECORD_BSIZE;
3039 dbp = xlog_alloc_buffer(log, BTOBB(h_size));
3045 memset(rhash, 0, sizeof(rhash));
3046 if (tail_blk > head_blk) {
3048 * Perform recovery around the end of the physical log.
3049 * When the head is not on the same cycle number as the tail,
3050 * we can't do a sequential recovery.
3052 while (blk_no < log->l_logBBsize) {
3054 * Check for header wrapping around physical end-of-log
3059 if (blk_no + hblks <= log->l_logBBsize) {
3060 /* Read header in one read */
3061 error = xlog_bread(log, blk_no, hblks, hbp,
3066 /* This LR is split across physical log end */
3067 if (blk_no != log->l_logBBsize) {
3068 /* some data before physical log end */
3069 ASSERT(blk_no <= INT_MAX);
3070 split_hblks = log->l_logBBsize - (int)blk_no;
3071 ASSERT(split_hblks > 0);
3072 error = xlog_bread(log, blk_no,
3080 * Note: this black magic still works with
3081 * large sector sizes (non-512) only because:
3082 * - we increased the buffer size originally
3083 * by 1 sector giving us enough extra space
3084 * for the second read;
3085 * - the log start is guaranteed to be sector
3087 * - we read the log end (LR header start)
3088 * _first_, then the log start (LR header end)
3089 * - order is important.
3091 wrapped_hblks = hblks - split_hblks;
3092 error = xlog_bread_noalign(log, 0,
3094 offset + BBTOB(split_hblks));
3098 rhead = (xlog_rec_header_t *)offset;
3099 error = xlog_valid_rec_header(log, rhead,
3100 split_hblks ? blk_no : 0);
3104 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3108 * Read the log record data in multiple reads if it
3109 * wraps around the end of the log. Note that if the
3110 * header already wrapped, blk_no could point past the
3111 * end of the log. The record data is contiguous in
3114 if (blk_no + bblks <= log->l_logBBsize ||
3115 blk_no >= log->l_logBBsize) {
3116 rblk_no = xlog_wrap_logbno(log, blk_no);
3117 error = xlog_bread(log, rblk_no, bblks, dbp,
3122 /* This log record is split across the
3123 * physical end of log */
3126 if (blk_no != log->l_logBBsize) {
3127 /* some data is before the physical
3129 ASSERT(!wrapped_hblks);
3130 ASSERT(blk_no <= INT_MAX);
3132 log->l_logBBsize - (int)blk_no;
3133 ASSERT(split_bblks > 0);
3134 error = xlog_bread(log, blk_no,
3142 * Note: this black magic still works with
3143 * large sector sizes (non-512) only because:
3144 * - we increased the buffer size originally
3145 * by 1 sector giving us enough extra space
3146 * for the second read;
3147 * - the log start is guaranteed to be sector
3149 * - we read the log end (LR header start)
3150 * _first_, then the log start (LR header end)
3151 * - order is important.
3153 error = xlog_bread_noalign(log, 0,
3154 bblks - split_bblks,
3155 offset + BBTOB(split_bblks));
3160 error = xlog_recover_process(log, rhash, rhead, offset,
3161 pass, &buffer_list);
3169 ASSERT(blk_no >= log->l_logBBsize);
3170 blk_no -= log->l_logBBsize;
3174 /* read first part of physical log */
3175 while (blk_no < head_blk) {
3176 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3180 rhead = (xlog_rec_header_t *)offset;
3181 error = xlog_valid_rec_header(log, rhead, blk_no);
3185 /* blocks in data section */
3186 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3187 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3192 error = xlog_recover_process(log, rhash, rhead, offset, pass,
3197 blk_no += bblks + hblks;
3207 * Submit buffers that have been added from the last record processed,
3208 * regardless of error status.
3210 if (!list_empty(&buffer_list))
3211 error2 = xfs_buf_delwri_submit(&buffer_list);
3213 if (error && first_bad)
3214 *first_bad = rhead_blk;
3217 * Transactions are freed at commit time but transactions without commit
3218 * records on disk are never committed. Free any that may be left in the
3221 for (i = 0; i < XLOG_RHASH_SIZE; i++) {
3222 struct hlist_node *tmp;
3223 struct xlog_recover *trans;
3225 hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
3226 xlog_recover_free_trans(trans);
3229 return error ? error : error2;
3233 * Do the recovery of the log. We actually do this in two phases.
3234 * The two passes are necessary in order to implement the function
3235 * of cancelling a record written into the log. The first pass
3236 * determines those things which have been cancelled, and the
3237 * second pass replays log items normally except for those which
3238 * have been cancelled. The handling of the replay and cancellations
3239 * takes place in the log item type specific routines.
3241 * The table of items which have cancel records in the log is allocated
3242 * and freed at this level, since only here do we know when all of
3243 * the log recovery has been completed.
3246 xlog_do_log_recovery(
3248 xfs_daddr_t head_blk,
3249 xfs_daddr_t tail_blk)
3253 ASSERT(head_blk != tail_blk);
3256 * First do a pass to find all of the cancelled buf log items.
3257 * Store them in the buf_cancel_table for use in the second pass.
3259 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3260 sizeof(struct list_head),
3262 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3263 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3265 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3266 XLOG_RECOVER_PASS1, NULL);
3268 kmem_free(log->l_buf_cancel_table);
3269 log->l_buf_cancel_table = NULL;
3273 * Then do a second pass to actually recover the items in the log.
3274 * When it is complete free the table of buf cancel items.
3276 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3277 XLOG_RECOVER_PASS2, NULL);
3282 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3283 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3287 kmem_free(log->l_buf_cancel_table);
3288 log->l_buf_cancel_table = NULL;
3294 * Do the actual recovery
3299 xfs_daddr_t head_blk,
3300 xfs_daddr_t tail_blk)
3302 struct xfs_mount *mp = log->l_mp;
3307 trace_xfs_log_recover(log, head_blk, tail_blk);
3310 * First replay the images in the log.
3312 error = xlog_do_log_recovery(log, head_blk, tail_blk);
3317 * If IO errors happened during recovery, bail out.
3319 if (XFS_FORCED_SHUTDOWN(mp)) {
3324 * We now update the tail_lsn since much of the recovery has completed
3325 * and there may be space available to use. If there were no extent
3326 * or iunlinks, we can free up the entire log and set the tail_lsn to
3327 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3328 * lsn of the last known good LR on disk. If there are extent frees
3329 * or iunlinks they will have some entries in the AIL; so we look at
3330 * the AIL to determine how to set the tail_lsn.
3332 xlog_assign_tail_lsn(mp);
3335 * Now that we've finished replaying all buffer and inode
3336 * updates, re-read in the superblock and reverify it.
3339 bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
3340 ASSERT(!(bp->b_flags & XBF_WRITE));
3341 bp->b_flags |= XBF_READ;
3342 bp->b_ops = &xfs_sb_buf_ops;
3344 error = xfs_buf_submit(bp);
3346 if (!XFS_FORCED_SHUTDOWN(mp)) {
3347 xfs_buf_ioerror_alert(bp, __this_address);
3354 /* Convert superblock from on-disk format */
3356 xfs_sb_from_disk(sbp, bp->b_addr);
3359 /* re-initialise in-core superblock and geometry structures */
3360 xfs_reinit_percpu_counters(mp);
3361 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
3363 xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
3366 mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
3368 xlog_recover_check_summary(log);
3370 /* Normal transactions can now occur */
3371 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3376 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3378 * Return error or zero.
3384 xfs_daddr_t head_blk, tail_blk;
3387 /* find the tail of the log */
3388 error = xlog_find_tail(log, &head_blk, &tail_blk);
3393 * The superblock was read before the log was available and thus the LSN
3394 * could not be verified. Check the superblock LSN against the current
3395 * LSN now that it's known.
3397 if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
3398 !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
3401 if (tail_blk != head_blk) {
3402 /* There used to be a comment here:
3404 * disallow recovery on read-only mounts. note -- mount
3405 * checks for ENOSPC and turns it into an intelligent
3407 * ...but this is no longer true. Now, unless you specify
3408 * NORECOVERY (in which case this function would never be
3409 * called), we just go ahead and recover. We do this all
3410 * under the vfs layer, so we can get away with it unless
3411 * the device itself is read-only, in which case we fail.
3413 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3418 * Version 5 superblock log feature mask validation. We know the
3419 * log is dirty so check if there are any unknown log features
3420 * in what we need to recover. If there are unknown features
3421 * (e.g. unsupported transactions, then simply reject the
3422 * attempt at recovery before touching anything.
3424 if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
3425 xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
3426 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
3428 "Superblock has unknown incompatible log features (0x%x) enabled.",
3429 (log->l_mp->m_sb.sb_features_log_incompat &
3430 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
3432 "The log can not be fully and/or safely recovered by this kernel.");
3434 "Please recover the log on a kernel that supports the unknown features.");
3439 * Delay log recovery if the debug hook is set. This is debug
3440 * instrumention to coordinate simulation of I/O failures with
3443 if (xfs_globals.log_recovery_delay) {
3444 xfs_notice(log->l_mp,
3445 "Delaying log recovery for %d seconds.",
3446 xfs_globals.log_recovery_delay);
3447 msleep(xfs_globals.log_recovery_delay * 1000);
3450 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3451 log->l_mp->m_logname ? log->l_mp->m_logname
3454 error = xlog_do_recover(log, head_blk, tail_blk);
3455 log->l_flags |= XLOG_RECOVERY_NEEDED;
3461 * In the first part of recovery we replay inodes and buffers and build
3462 * up the list of extent free items which need to be processed. Here
3463 * we process the extent free items and clean up the on disk unlinked
3464 * inode lists. This is separated from the first part of recovery so
3465 * that the root and real-time bitmap inodes can be read in from disk in
3466 * between the two stages. This is necessary so that we can free space
3467 * in the real-time portion of the file system.
3470 xlog_recover_finish(
3474 * Now we're ready to do the transactions needed for the
3475 * rest of recovery. Start with completing all the extent
3476 * free intent records and then process the unlinked inode
3477 * lists. At this point, we essentially run in normal mode
3478 * except that we're still performing recovery actions
3479 * rather than accepting new requests.
3481 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3483 error = xlog_recover_process_intents(log);
3485 xfs_alert(log->l_mp, "Failed to recover intents");
3490 * Sync the log to get all the intents out of the AIL.
3491 * This isn't absolutely necessary, but it helps in
3492 * case the unlink transactions would have problems
3493 * pushing the intents out of the way.
3495 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3497 xlog_recover_process_iunlinks(log);
3499 xlog_recover_check_summary(log);
3501 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
3502 log->l_mp->m_logname ? log->l_mp->m_logname
3504 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3506 xfs_info(log->l_mp, "Ending clean mount");
3512 xlog_recover_cancel(
3515 if (log->l_flags & XLOG_RECOVERY_NEEDED)
3516 xlog_recover_cancel_intents(log);
3521 * Read all of the agf and agi counters and check that they
3522 * are consistent with the superblock counters.
3525 xlog_recover_check_summary(
3531 xfs_agnumber_t agno;
3542 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3543 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
3545 xfs_alert(mp, "%s agf read failed agno %d error %d",
3546 __func__, agno, error);
3548 struct xfs_agf *agfp = agfbp->b_addr;
3550 freeblks += be32_to_cpu(agfp->agf_freeblks) +
3551 be32_to_cpu(agfp->agf_flcount);
3552 xfs_buf_relse(agfbp);
3555 error = xfs_read_agi(mp, NULL, agno, &agibp);
3557 xfs_alert(mp, "%s agi read failed agno %d error %d",
3558 __func__, agno, error);
3560 struct xfs_agi *agi = agibp->b_addr;
3562 itotal += be32_to_cpu(agi->agi_count);
3563 ifree += be32_to_cpu(agi->agi_freecount);
3564 xfs_buf_relse(agibp);