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);
290 bp->b_flags &= ~_XBF_LOGRECOVERY;
291 xfs_buf_ioend_finish(bp);
295 * This routine finds (to an approximation) the first block in the physical
296 * log which contains the given cycle. It uses a binary search algorithm.
297 * Note that the algorithm can not be perfect because the disk will not
298 * necessarily be perfect.
301 xlog_find_cycle_start(
304 xfs_daddr_t first_blk,
305 xfs_daddr_t *last_blk,
315 mid_blk = BLK_AVG(first_blk, end_blk);
316 while (mid_blk != first_blk && mid_blk != end_blk) {
317 error = xlog_bread(log, mid_blk, 1, buffer, &offset);
320 mid_cycle = xlog_get_cycle(offset);
321 if (mid_cycle == cycle)
322 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
324 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
325 mid_blk = BLK_AVG(first_blk, end_blk);
327 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
328 (mid_blk == end_blk && mid_blk-1 == first_blk));
336 * Check that a range of blocks does not contain stop_on_cycle_no.
337 * Fill in *new_blk with the block offset where such a block is
338 * found, or with -1 (an invalid block number) if there is no such
339 * block in the range. The scan needs to occur from front to back
340 * and the pointer into the region must be updated since a later
341 * routine will need to perform another test.
344 xlog_find_verify_cycle(
346 xfs_daddr_t start_blk,
348 uint stop_on_cycle_no,
349 xfs_daddr_t *new_blk)
359 * Greedily allocate a buffer big enough to handle the full
360 * range of basic blocks we'll be examining. If that fails,
361 * try a smaller size. We need to be able to read at least
362 * a log sector, or we're out of luck.
364 bufblks = 1 << ffs(nbblks);
365 while (bufblks > log->l_logBBsize)
367 while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
369 if (bufblks < log->l_sectBBsize)
373 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
376 bcount = min(bufblks, (start_blk + nbblks - i));
378 error = xlog_bread(log, i, bcount, buffer, &buf);
382 for (j = 0; j < bcount; j++) {
383 cycle = xlog_get_cycle(buf);
384 if (cycle == stop_on_cycle_no) {
401 * Potentially backup over partial log record write.
403 * In the typical case, last_blk is the number of the block directly after
404 * a good log record. Therefore, we subtract one to get the block number
405 * of the last block in the given buffer. extra_bblks contains the number
406 * of blocks we would have read on a previous read. This happens when the
407 * last log record is split over the end of the physical log.
409 * extra_bblks is the number of blocks potentially verified on a previous
410 * call to this routine.
413 xlog_find_verify_log_record(
415 xfs_daddr_t start_blk,
416 xfs_daddr_t *last_blk,
422 xlog_rec_header_t *head = NULL;
425 int num_blks = *last_blk - start_blk;
428 ASSERT(start_blk != 0 || *last_blk != start_blk);
430 buffer = xlog_alloc_buffer(log, num_blks);
432 buffer = xlog_alloc_buffer(log, 1);
437 error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
440 offset += ((num_blks - 1) << BBSHIFT);
443 for (i = (*last_blk) - 1; i >= 0; i--) {
445 /* valid log record not found */
447 "Log inconsistent (didn't find previous header)");
449 error = -EFSCORRUPTED;
454 error = xlog_bread(log, i, 1, buffer, &offset);
459 head = (xlog_rec_header_t *)offset;
461 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
469 * We hit the beginning of the physical log & still no header. Return
470 * to caller. If caller can handle a return of -1, then this routine
471 * will be called again for the end of the physical log.
479 * We have the final block of the good log (the first block
480 * of the log record _before_ the head. So we check the uuid.
482 if ((error = xlog_header_check_mount(log->l_mp, head)))
486 * We may have found a log record header before we expected one.
487 * last_blk will be the 1st block # with a given cycle #. We may end
488 * up reading an entire log record. In this case, we don't want to
489 * reset last_blk. Only when last_blk points in the middle of a log
490 * record do we update last_blk.
492 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
493 uint h_size = be32_to_cpu(head->h_size);
495 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
496 if (h_size % XLOG_HEADER_CYCLE_SIZE)
502 if (*last_blk - i + extra_bblks !=
503 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
512 * Head is defined to be the point of the log where the next log write
513 * could go. This means that incomplete LR writes at the end are
514 * eliminated when calculating the head. We aren't guaranteed that previous
515 * LR have complete transactions. We only know that a cycle number of
516 * current cycle number -1 won't be present in the log if we start writing
517 * from our current block number.
519 * last_blk contains the block number of the first block with a given
522 * Return: zero if normal, non-zero if error.
527 xfs_daddr_t *return_head_blk)
531 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
533 uint first_half_cycle, last_half_cycle;
535 int error, log_bbnum = log->l_logBBsize;
537 /* Is the end of the log device zeroed? */
538 error = xlog_find_zeroed(log, &first_blk);
540 xfs_warn(log->l_mp, "empty log check failed");
544 *return_head_blk = first_blk;
546 /* Is the whole lot zeroed? */
548 /* Linux XFS shouldn't generate totally zeroed logs -
549 * mkfs etc write a dummy unmount record to a fresh
550 * log so we can store the uuid in there
552 xfs_warn(log->l_mp, "totally zeroed log");
558 first_blk = 0; /* get cycle # of 1st block */
559 buffer = xlog_alloc_buffer(log, 1);
563 error = xlog_bread(log, 0, 1, buffer, &offset);
565 goto out_free_buffer;
567 first_half_cycle = xlog_get_cycle(offset);
569 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
570 error = xlog_bread(log, last_blk, 1, buffer, &offset);
572 goto out_free_buffer;
574 last_half_cycle = xlog_get_cycle(offset);
575 ASSERT(last_half_cycle != 0);
578 * If the 1st half cycle number is equal to the last half cycle number,
579 * then the entire log is stamped with the same cycle number. In this
580 * case, head_blk can't be set to zero (which makes sense). The below
581 * math doesn't work out properly with head_blk equal to zero. Instead,
582 * we set it to log_bbnum which is an invalid block number, but this
583 * value makes the math correct. If head_blk doesn't changed through
584 * all the tests below, *head_blk is set to zero at the very end rather
585 * than log_bbnum. In a sense, log_bbnum and zero are the same block
586 * in a circular file.
588 if (first_half_cycle == last_half_cycle) {
590 * In this case we believe that the entire log should have
591 * cycle number last_half_cycle. We need to scan backwards
592 * from the end verifying that there are no holes still
593 * containing last_half_cycle - 1. If we find such a hole,
594 * then the start of that hole will be the new head. The
595 * simple case looks like
596 * x | x ... | x - 1 | x
597 * Another case that fits this picture would be
598 * x | x + 1 | x ... | x
599 * In this case the head really is somewhere at the end of the
600 * log, as one of the latest writes at the beginning was
603 * x | x + 1 | x ... | x - 1 | x
604 * This is really the combination of the above two cases, and
605 * the head has to end up at the start of the x-1 hole at the
608 * In the 256k log case, we will read from the beginning to the
609 * end of the log and search for cycle numbers equal to x-1.
610 * We don't worry about the x+1 blocks that we encounter,
611 * because we know that they cannot be the head since the log
614 head_blk = log_bbnum;
615 stop_on_cycle = last_half_cycle - 1;
618 * In this case we want to find the first block with cycle
619 * number matching last_half_cycle. We expect the log to be
621 * x + 1 ... | x ... | x
622 * The first block with cycle number x (last_half_cycle) will
623 * be where the new head belongs. First we do a binary search
624 * for the first occurrence of last_half_cycle. The binary
625 * search may not be totally accurate, so then we scan back
626 * from there looking for occurrences of last_half_cycle before
627 * us. If that backwards scan wraps around the beginning of
628 * the log, then we look for occurrences of last_half_cycle - 1
629 * at the end of the log. The cases we're looking for look
631 * v binary search stopped here
632 * x + 1 ... | x | x + 1 | x ... | x
633 * ^ but we want to locate this spot
635 * <---------> less than scan distance
636 * x + 1 ... | x ... | x - 1 | x
637 * ^ we want to locate this spot
639 stop_on_cycle = last_half_cycle;
640 error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
643 goto out_free_buffer;
647 * Now validate the answer. Scan back some number of maximum possible
648 * blocks and make sure each one has the expected cycle number. The
649 * maximum is determined by the total possible amount of buffering
650 * in the in-core log. The following number can be made tighter if
651 * we actually look at the block size of the filesystem.
653 num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
654 if (head_blk >= num_scan_bblks) {
656 * We are guaranteed that the entire check can be performed
659 start_blk = head_blk - num_scan_bblks;
660 if ((error = xlog_find_verify_cycle(log,
661 start_blk, num_scan_bblks,
662 stop_on_cycle, &new_blk)))
663 goto out_free_buffer;
666 } else { /* need to read 2 parts of log */
668 * We are going to scan backwards in the log in two parts.
669 * First we scan the physical end of the log. In this part
670 * of the log, we are looking for blocks with cycle number
671 * last_half_cycle - 1.
672 * If we find one, then we know that the log starts there, as
673 * we've found a hole that didn't get written in going around
674 * the end of the physical log. The simple case for this is
675 * x + 1 ... | x ... | x - 1 | x
676 * <---------> less than scan distance
677 * If all of the blocks at the end of the log have cycle number
678 * last_half_cycle, then we check the blocks at the start of
679 * the log looking for occurrences of last_half_cycle. If we
680 * find one, then our current estimate for the location of the
681 * first occurrence of last_half_cycle is wrong and we move
682 * back to the hole we've found. This case looks like
683 * x + 1 ... | x | x + 1 | x ...
684 * ^ binary search stopped here
685 * Another case we need to handle that only occurs in 256k
687 * x + 1 ... | x ... | x+1 | x ...
688 * ^ binary search stops here
689 * In a 256k log, the scan at the end of the log will see the
690 * x + 1 blocks. We need to skip past those since that is
691 * certainly not the head of the log. By searching for
692 * last_half_cycle-1 we accomplish that.
694 ASSERT(head_blk <= INT_MAX &&
695 (xfs_daddr_t) num_scan_bblks >= head_blk);
696 start_blk = log_bbnum - (num_scan_bblks - head_blk);
697 if ((error = xlog_find_verify_cycle(log, start_blk,
698 num_scan_bblks - (int)head_blk,
699 (stop_on_cycle - 1), &new_blk)))
700 goto out_free_buffer;
707 * Scan beginning of log now. The last part of the physical
708 * log is good. This scan needs to verify that it doesn't find
709 * the last_half_cycle.
712 ASSERT(head_blk <= INT_MAX);
713 if ((error = xlog_find_verify_cycle(log,
714 start_blk, (int)head_blk,
715 stop_on_cycle, &new_blk)))
716 goto out_free_buffer;
723 * Now we need to make sure head_blk is not pointing to a block in
724 * the middle of a log record.
726 num_scan_bblks = XLOG_REC_SHIFT(log);
727 if (head_blk >= num_scan_bblks) {
728 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
730 /* start ptr at last block ptr before head_blk */
731 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
735 goto out_free_buffer;
738 ASSERT(head_blk <= INT_MAX);
739 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
741 goto out_free_buffer;
743 /* We hit the beginning of the log during our search */
744 start_blk = log_bbnum - (num_scan_bblks - head_blk);
746 ASSERT(start_blk <= INT_MAX &&
747 (xfs_daddr_t) log_bbnum-start_blk >= 0);
748 ASSERT(head_blk <= INT_MAX);
749 error = xlog_find_verify_log_record(log, start_blk,
750 &new_blk, (int)head_blk);
754 goto out_free_buffer;
755 if (new_blk != log_bbnum)
758 goto out_free_buffer;
762 if (head_blk == log_bbnum)
763 *return_head_blk = 0;
765 *return_head_blk = head_blk;
767 * When returning here, we have a good block number. Bad block
768 * means that during a previous crash, we didn't have a clean break
769 * from cycle number N to cycle number N-1. In this case, we need
770 * to find the first block with cycle number N-1.
777 xfs_warn(log->l_mp, "failed to find log head");
782 * Seek backwards in the log for log record headers.
784 * Given a starting log block, walk backwards until we find the provided number
785 * of records or hit the provided tail block. The return value is the number of
786 * records encountered or a negative error code. The log block and buffer
787 * pointer of the last record seen are returned in rblk and rhead respectively.
790 xlog_rseek_logrec_hdr(
792 xfs_daddr_t head_blk,
793 xfs_daddr_t tail_blk,
797 struct xlog_rec_header **rhead,
809 * Walk backwards from the head block until we hit the tail or the first
812 end_blk = head_blk > tail_blk ? tail_blk : 0;
813 for (i = (int) head_blk - 1; i >= end_blk; i--) {
814 error = xlog_bread(log, i, 1, buffer, &offset);
818 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
820 *rhead = (struct xlog_rec_header *) offset;
821 if (++found == count)
827 * If we haven't hit the tail block or the log record header count,
828 * start looking again from the end of the physical log. Note that
829 * callers can pass head == tail if the tail is not yet known.
831 if (tail_blk >= head_blk && found != count) {
832 for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
833 error = xlog_bread(log, i, 1, buffer, &offset);
837 if (*(__be32 *)offset ==
838 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
841 *rhead = (struct xlog_rec_header *) offset;
842 if (++found == count)
855 * Seek forward in the log for log record headers.
857 * Given head and tail blocks, walk forward from the tail block until we find
858 * the provided number of records or hit the head block. The return value is the
859 * number of records encountered or a negative error code. The log block and
860 * buffer pointer of the last record seen are returned in rblk and rhead
864 xlog_seek_logrec_hdr(
866 xfs_daddr_t head_blk,
867 xfs_daddr_t tail_blk,
871 struct xlog_rec_header **rhead,
883 * Walk forward from the tail block until we hit the head or the last
886 end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
887 for (i = (int) tail_blk; i <= end_blk; i++) {
888 error = xlog_bread(log, i, 1, buffer, &offset);
892 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
894 *rhead = (struct xlog_rec_header *) offset;
895 if (++found == count)
901 * If we haven't hit the head block or the log record header count,
902 * start looking again from the start of the physical log.
904 if (tail_blk > head_blk && found != count) {
905 for (i = 0; i < (int) head_blk; i++) {
906 error = xlog_bread(log, i, 1, buffer, &offset);
910 if (*(__be32 *)offset ==
911 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
914 *rhead = (struct xlog_rec_header *) offset;
915 if (++found == count)
928 * Calculate distance from head to tail (i.e., unused space in the log).
933 xfs_daddr_t head_blk,
934 xfs_daddr_t tail_blk)
936 if (head_blk < tail_blk)
937 return tail_blk - head_blk;
939 return tail_blk + (log->l_logBBsize - head_blk);
943 * Verify the log tail. This is particularly important when torn or incomplete
944 * writes have been detected near the front of the log and the head has been
945 * walked back accordingly.
947 * We also have to handle the case where the tail was pinned and the head
948 * blocked behind the tail right before a crash. If the tail had been pushed
949 * immediately prior to the crash and the subsequent checkpoint was only
950 * partially written, it's possible it overwrote the last referenced tail in the
951 * log with garbage. This is not a coherency problem because the tail must have
952 * been pushed before it can be overwritten, but appears as log corruption to
953 * recovery because we have no way to know the tail was updated if the
954 * subsequent checkpoint didn't write successfully.
956 * Therefore, CRC check the log from tail to head. If a failure occurs and the
957 * offending record is within max iclog bufs from the head, walk the tail
958 * forward and retry until a valid tail is found or corruption is detected out
959 * of the range of a possible overwrite.
964 xfs_daddr_t head_blk,
965 xfs_daddr_t *tail_blk,
968 struct xlog_rec_header *thead;
970 xfs_daddr_t first_bad;
973 xfs_daddr_t tmp_tail;
974 xfs_daddr_t orig_tail = *tail_blk;
976 buffer = xlog_alloc_buffer(log, 1);
981 * Make sure the tail points to a record (returns positive count on
984 error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
985 &tmp_tail, &thead, &wrapped);
988 if (*tail_blk != tmp_tail)
989 *tail_blk = tmp_tail;
992 * Run a CRC check from the tail to the head. We can't just check
993 * MAX_ICLOGS records past the tail because the tail may point to stale
994 * blocks cleared during the search for the head/tail. These blocks are
995 * overwritten with zero-length records and thus record count is not a
996 * reliable indicator of the iclog state before a crash.
999 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1000 XLOG_RECOVER_CRCPASS, &first_bad);
1001 while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1005 * Is corruption within range of the head? If so, retry from
1006 * the next record. Otherwise return an error.
1008 tail_distance = xlog_tail_distance(log, head_blk, first_bad);
1009 if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
1012 /* skip to the next record; returns positive count on success */
1013 error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
1014 buffer, &tmp_tail, &thead, &wrapped);
1018 *tail_blk = tmp_tail;
1020 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1021 XLOG_RECOVER_CRCPASS, &first_bad);
1024 if (!error && *tail_blk != orig_tail)
1026 "Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1027 orig_tail, *tail_blk);
1034 * Detect and trim torn writes from the head of the log.
1036 * Storage without sector atomicity guarantees can result in torn writes in the
1037 * log in the event of a crash. Our only means to detect this scenario is via
1038 * CRC verification. While we can't always be certain that CRC verification
1039 * failure is due to a torn write vs. an unrelated corruption, we do know that
1040 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1041 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1042 * the log and treat failures in this range as torn writes as a matter of
1043 * policy. In the event of CRC failure, the head is walked back to the last good
1044 * record in the log and the tail is updated from that record and verified.
1049 xfs_daddr_t *head_blk, /* in/out: unverified head */
1050 xfs_daddr_t *tail_blk, /* out: tail block */
1052 xfs_daddr_t *rhead_blk, /* start blk of last record */
1053 struct xlog_rec_header **rhead, /* ptr to last record */
1054 bool *wrapped) /* last rec. wraps phys. log */
1056 struct xlog_rec_header *tmp_rhead;
1058 xfs_daddr_t first_bad;
1059 xfs_daddr_t tmp_rhead_blk;
1065 * Check the head of the log for torn writes. Search backwards from the
1066 * head until we hit the tail or the maximum number of log record I/Os
1067 * that could have been in flight at one time. Use a temporary buffer so
1068 * we don't trash the rhead/buffer pointers from the caller.
1070 tmp_buffer = xlog_alloc_buffer(log, 1);
1073 error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1074 XLOG_MAX_ICLOGS, tmp_buffer,
1075 &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1076 kmem_free(tmp_buffer);
1081 * Now run a CRC verification pass over the records starting at the
1082 * block found above to the current head. If a CRC failure occurs, the
1083 * log block of the first bad record is saved in first_bad.
1085 error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1086 XLOG_RECOVER_CRCPASS, &first_bad);
1087 if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1089 * We've hit a potential torn write. Reset the error and warn
1094 "Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1095 first_bad, *head_blk);
1098 * Get the header block and buffer pointer for the last good
1099 * record before the bad record.
1101 * Note that xlog_find_tail() clears the blocks at the new head
1102 * (i.e., the records with invalid CRC) if the cycle number
1103 * matches the current cycle.
1105 found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1106 buffer, rhead_blk, rhead, wrapped);
1109 if (found == 0) /* XXX: right thing to do here? */
1113 * Reset the head block to the starting block of the first bad
1114 * log record and set the tail block based on the last good
1117 * Bail out if the updated head/tail match as this indicates
1118 * possible corruption outside of the acceptable
1119 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1121 *head_blk = first_bad;
1122 *tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1123 if (*head_blk == *tail_blk) {
1131 return xlog_verify_tail(log, *head_blk, tail_blk,
1132 be32_to_cpu((*rhead)->h_size));
1136 * We need to make sure we handle log wrapping properly, so we can't use the
1137 * calculated logbno directly. Make sure it wraps to the correct bno inside the
1140 * The log is limited to 32 bit sizes, so we use the appropriate modulus
1141 * operation here and cast it back to a 64 bit daddr on return.
1143 static inline xfs_daddr_t
1150 div_s64_rem(bno, log->l_logBBsize, &mod);
1155 * Check whether the head of the log points to an unmount record. In other
1156 * words, determine whether the log is clean. If so, update the in-core state
1160 xlog_check_unmount_rec(
1162 xfs_daddr_t *head_blk,
1163 xfs_daddr_t *tail_blk,
1164 struct xlog_rec_header *rhead,
1165 xfs_daddr_t rhead_blk,
1169 struct xlog_op_header *op_head;
1170 xfs_daddr_t umount_data_blk;
1171 xfs_daddr_t after_umount_blk;
1179 * Look for unmount record. If we find it, then we know there was a
1180 * clean unmount. Since 'i' could be the last block in the physical
1181 * log, we convert to a log block before comparing to the head_blk.
1183 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1184 * below. We won't want to clear the unmount record if there is one, so
1185 * we pass the lsn of the unmount record rather than the block after it.
1187 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1188 int h_size = be32_to_cpu(rhead->h_size);
1189 int h_version = be32_to_cpu(rhead->h_version);
1191 if ((h_version & XLOG_VERSION_2) &&
1192 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1193 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1194 if (h_size % XLOG_HEADER_CYCLE_SIZE)
1203 after_umount_blk = xlog_wrap_logbno(log,
1204 rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1206 if (*head_blk == after_umount_blk &&
1207 be32_to_cpu(rhead->h_num_logops) == 1) {
1208 umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1209 error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1213 op_head = (struct xlog_op_header *)offset;
1214 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1216 * Set tail and last sync so that newly written log
1217 * records will point recovery to after the current
1220 xlog_assign_atomic_lsn(&log->l_tail_lsn,
1221 log->l_curr_cycle, after_umount_blk);
1222 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1223 log->l_curr_cycle, after_umount_blk);
1224 *tail_blk = after_umount_blk;
1236 xfs_daddr_t head_blk,
1237 struct xlog_rec_header *rhead,
1238 xfs_daddr_t rhead_blk,
1242 * Reset log values according to the state of the log when we
1243 * crashed. In the case where head_blk == 0, we bump curr_cycle
1244 * one because the next write starts a new cycle rather than
1245 * continuing the cycle of the last good log record. At this
1246 * point we have guaranteed that all partial log records have been
1247 * accounted for. Therefore, we know that the last good log record
1248 * written was complete and ended exactly on the end boundary
1249 * of the physical log.
1251 log->l_prev_block = rhead_blk;
1252 log->l_curr_block = (int)head_blk;
1253 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1255 log->l_curr_cycle++;
1256 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1257 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1258 xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1259 BBTOB(log->l_curr_block));
1260 xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1261 BBTOB(log->l_curr_block));
1265 * Find the sync block number or the tail of the log.
1267 * This will be the block number of the last record to have its
1268 * associated buffers synced to disk. Every log record header has
1269 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
1270 * to get a sync block number. The only concern is to figure out which
1271 * log record header to believe.
1273 * The following algorithm uses the log record header with the largest
1274 * lsn. The entire log record does not need to be valid. We only care
1275 * that the header is valid.
1277 * We could speed up search by using current head_blk buffer, but it is not
1283 xfs_daddr_t *head_blk,
1284 xfs_daddr_t *tail_blk)
1286 xlog_rec_header_t *rhead;
1287 char *offset = NULL;
1290 xfs_daddr_t rhead_blk;
1292 bool wrapped = false;
1296 * Find previous log record
1298 if ((error = xlog_find_head(log, head_blk)))
1300 ASSERT(*head_blk < INT_MAX);
1302 buffer = xlog_alloc_buffer(log, 1);
1305 if (*head_blk == 0) { /* special case */
1306 error = xlog_bread(log, 0, 1, buffer, &offset);
1310 if (xlog_get_cycle(offset) == 0) {
1312 /* leave all other log inited values alone */
1318 * Search backwards through the log looking for the log record header
1319 * block. This wraps all the way back around to the head so something is
1320 * seriously wrong if we can't find it.
1322 error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1323 &rhead_blk, &rhead, &wrapped);
1327 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1328 error = -EFSCORRUPTED;
1331 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1334 * Set the log state based on the current head record.
1336 xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1337 tail_lsn = atomic64_read(&log->l_tail_lsn);
1340 * Look for an unmount record at the head of the log. This sets the log
1341 * state to determine whether recovery is necessary.
1343 error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1344 rhead_blk, buffer, &clean);
1349 * Verify the log head if the log is not clean (e.g., we have anything
1350 * but an unmount record at the head). This uses CRC verification to
1351 * detect and trim torn writes. If discovered, CRC failures are
1352 * considered torn writes and the log head is trimmed accordingly.
1354 * Note that we can only run CRC verification when the log is dirty
1355 * because there's no guarantee that the log data behind an unmount
1356 * record is compatible with the current architecture.
1359 xfs_daddr_t orig_head = *head_blk;
1361 error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1362 &rhead_blk, &rhead, &wrapped);
1366 /* update in-core state again if the head changed */
1367 if (*head_blk != orig_head) {
1368 xlog_set_state(log, *head_blk, rhead, rhead_blk,
1370 tail_lsn = atomic64_read(&log->l_tail_lsn);
1371 error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1372 rhead, rhead_blk, buffer,
1380 * Note that the unmount was clean. If the unmount was not clean, we
1381 * need to know this to rebuild the superblock counters from the perag
1382 * headers if we have a filesystem using non-persistent counters.
1385 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1388 * Make sure that there are no blocks in front of the head
1389 * with the same cycle number as the head. This can happen
1390 * because we allow multiple outstanding log writes concurrently,
1391 * and the later writes might make it out before earlier ones.
1393 * We use the lsn from before modifying it so that we'll never
1394 * overwrite the unmount record after a clean unmount.
1396 * Do this only if we are going to recover the filesystem
1398 * NOTE: This used to say "if (!readonly)"
1399 * However on Linux, we can & do recover a read-only filesystem.
1400 * We only skip recovery if NORECOVERY is specified on mount,
1401 * in which case we would not be here.
1403 * But... if the -device- itself is readonly, just skip this.
1404 * We can't recover this device anyway, so it won't matter.
1406 if (!xfs_readonly_buftarg(log->l_targ))
1407 error = xlog_clear_stale_blocks(log, tail_lsn);
1413 xfs_warn(log->l_mp, "failed to locate log tail");
1418 * Is the log zeroed at all?
1420 * The last binary search should be changed to perform an X block read
1421 * once X becomes small enough. You can then search linearly through
1422 * the X blocks. This will cut down on the number of reads we need to do.
1424 * If the log is partially zeroed, this routine will pass back the blkno
1425 * of the first block with cycle number 0. It won't have a complete LR
1429 * 0 => the log is completely written to
1430 * 1 => use *blk_no as the first block of the log
1431 * <0 => error has occurred
1436 xfs_daddr_t *blk_no)
1440 uint first_cycle, last_cycle;
1441 xfs_daddr_t new_blk, last_blk, start_blk;
1442 xfs_daddr_t num_scan_bblks;
1443 int error, log_bbnum = log->l_logBBsize;
1447 /* check totally zeroed log */
1448 buffer = xlog_alloc_buffer(log, 1);
1451 error = xlog_bread(log, 0, 1, buffer, &offset);
1453 goto out_free_buffer;
1455 first_cycle = xlog_get_cycle(offset);
1456 if (first_cycle == 0) { /* completely zeroed log */
1462 /* check partially zeroed log */
1463 error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1465 goto out_free_buffer;
1467 last_cycle = xlog_get_cycle(offset);
1468 if (last_cycle != 0) { /* log completely written to */
1473 /* we have a partially zeroed log */
1474 last_blk = log_bbnum-1;
1475 error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1477 goto out_free_buffer;
1480 * Validate the answer. Because there is no way to guarantee that
1481 * the entire log is made up of log records which are the same size,
1482 * we scan over the defined maximum blocks. At this point, the maximum
1483 * is not chosen to mean anything special. XXXmiken
1485 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1486 ASSERT(num_scan_bblks <= INT_MAX);
1488 if (last_blk < num_scan_bblks)
1489 num_scan_bblks = last_blk;
1490 start_blk = last_blk - num_scan_bblks;
1493 * We search for any instances of cycle number 0 that occur before
1494 * our current estimate of the head. What we're trying to detect is
1495 * 1 ... | 0 | 1 | 0...
1496 * ^ binary search ends here
1498 if ((error = xlog_find_verify_cycle(log, start_blk,
1499 (int)num_scan_bblks, 0, &new_blk)))
1500 goto out_free_buffer;
1505 * Potentially backup over partial log record write. We don't need
1506 * to search the end of the log because we know it is zero.
1508 error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1512 goto out_free_buffer;
1523 * These are simple subroutines used by xlog_clear_stale_blocks() below
1524 * to initialize a buffer full of empty log record headers and write
1525 * them into the log.
1536 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1538 memset(buf, 0, BBSIZE);
1539 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1540 recp->h_cycle = cpu_to_be32(cycle);
1541 recp->h_version = cpu_to_be32(
1542 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1543 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1544 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1545 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1546 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1550 xlog_write_log_records(
1561 int sectbb = log->l_sectBBsize;
1562 int end_block = start_block + blocks;
1568 * Greedily allocate a buffer big enough to handle the full
1569 * range of basic blocks to be written. If that fails, try
1570 * a smaller size. We need to be able to write at least a
1571 * log sector, or we're out of luck.
1573 bufblks = 1 << ffs(blocks);
1574 while (bufblks > log->l_logBBsize)
1576 while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1578 if (bufblks < sectbb)
1582 /* We may need to do a read at the start to fill in part of
1583 * the buffer in the starting sector not covered by the first
1586 balign = round_down(start_block, sectbb);
1587 if (balign != start_block) {
1588 error = xlog_bread_noalign(log, start_block, 1, buffer);
1590 goto out_free_buffer;
1592 j = start_block - balign;
1595 for (i = start_block; i < end_block; i += bufblks) {
1596 int bcount, endcount;
1598 bcount = min(bufblks, end_block - start_block);
1599 endcount = bcount - j;
1601 /* We may need to do a read at the end to fill in part of
1602 * the buffer in the final sector not covered by the write.
1603 * If this is the same sector as the above read, skip it.
1605 ealign = round_down(end_block, sectbb);
1606 if (j == 0 && (start_block + endcount > ealign)) {
1607 error = xlog_bread_noalign(log, ealign, sectbb,
1608 buffer + BBTOB(ealign - start_block));
1614 offset = buffer + xlog_align(log, start_block);
1615 for (; j < endcount; j++) {
1616 xlog_add_record(log, offset, cycle, i+j,
1617 tail_cycle, tail_block);
1620 error = xlog_bwrite(log, start_block, endcount, buffer);
1623 start_block += endcount;
1633 * This routine is called to blow away any incomplete log writes out
1634 * in front of the log head. We do this so that we won't become confused
1635 * if we come up, write only a little bit more, and then crash again.
1636 * If we leave the partial log records out there, this situation could
1637 * cause us to think those partial writes are valid blocks since they
1638 * have the current cycle number. We get rid of them by overwriting them
1639 * with empty log records with the old cycle number rather than the
1642 * The tail lsn is passed in rather than taken from
1643 * the log so that we will not write over the unmount record after a
1644 * clean unmount in a 512 block log. Doing so would leave the log without
1645 * any valid log records in it until a new one was written. If we crashed
1646 * during that time we would not be able to recover.
1649 xlog_clear_stale_blocks(
1653 int tail_cycle, head_cycle;
1654 int tail_block, head_block;
1655 int tail_distance, max_distance;
1659 tail_cycle = CYCLE_LSN(tail_lsn);
1660 tail_block = BLOCK_LSN(tail_lsn);
1661 head_cycle = log->l_curr_cycle;
1662 head_block = log->l_curr_block;
1665 * Figure out the distance between the new head of the log
1666 * and the tail. We want to write over any blocks beyond the
1667 * head that we may have written just before the crash, but
1668 * we don't want to overwrite the tail of the log.
1670 if (head_cycle == tail_cycle) {
1672 * The tail is behind the head in the physical log,
1673 * so the distance from the head to the tail is the
1674 * distance from the head to the end of the log plus
1675 * the distance from the beginning of the log to the
1678 if (XFS_IS_CORRUPT(log->l_mp,
1679 head_block < tail_block ||
1680 head_block >= log->l_logBBsize))
1681 return -EFSCORRUPTED;
1682 tail_distance = tail_block + (log->l_logBBsize - head_block);
1685 * The head is behind the tail in the physical log,
1686 * so the distance from the head to the tail is just
1687 * the tail block minus the head block.
1689 if (XFS_IS_CORRUPT(log->l_mp,
1690 head_block >= tail_block ||
1691 head_cycle != tail_cycle + 1))
1692 return -EFSCORRUPTED;
1693 tail_distance = tail_block - head_block;
1697 * If the head is right up against the tail, we can't clear
1700 if (tail_distance <= 0) {
1701 ASSERT(tail_distance == 0);
1705 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1707 * Take the smaller of the maximum amount of outstanding I/O
1708 * we could have and the distance to the tail to clear out.
1709 * We take the smaller so that we don't overwrite the tail and
1710 * we don't waste all day writing from the head to the tail
1713 max_distance = min(max_distance, tail_distance);
1715 if ((head_block + max_distance) <= log->l_logBBsize) {
1717 * We can stomp all the blocks we need to without
1718 * wrapping around the end of the log. Just do it
1719 * in a single write. Use the cycle number of the
1720 * current cycle minus one so that the log will look like:
1723 error = xlog_write_log_records(log, (head_cycle - 1),
1724 head_block, max_distance, tail_cycle,
1730 * We need to wrap around the end of the physical log in
1731 * order to clear all the blocks. Do it in two separate
1732 * I/Os. The first write should be from the head to the
1733 * end of the physical log, and it should use the current
1734 * cycle number minus one just like above.
1736 distance = log->l_logBBsize - head_block;
1737 error = xlog_write_log_records(log, (head_cycle - 1),
1738 head_block, distance, tail_cycle,
1745 * Now write the blocks at the start of the physical log.
1746 * This writes the remainder of the blocks we want to clear.
1747 * It uses the current cycle number since we're now on the
1748 * same cycle as the head so that we get:
1749 * n ... n ... | n - 1 ...
1750 * ^^^^^ blocks we're writing
1752 distance = max_distance - (log->l_logBBsize - head_block);
1753 error = xlog_write_log_records(log, head_cycle, 0, distance,
1754 tail_cycle, tail_block);
1763 * Release the recovered intent item in the AIL that matches the given intent
1764 * type and intent id.
1767 xlog_recover_release_intent(
1769 unsigned short intent_type,
1772 struct xfs_ail_cursor cur;
1773 struct xfs_log_item *lip;
1774 struct xfs_ail *ailp = log->l_ailp;
1776 spin_lock(&ailp->ail_lock);
1777 for (lip = xfs_trans_ail_cursor_first(ailp, &cur, 0); lip != NULL;
1778 lip = xfs_trans_ail_cursor_next(ailp, &cur)) {
1779 if (lip->li_type != intent_type)
1781 if (!lip->li_ops->iop_match(lip, intent_id))
1784 spin_unlock(&ailp->ail_lock);
1785 lip->li_ops->iop_release(lip);
1786 spin_lock(&ailp->ail_lock);
1790 xfs_trans_ail_cursor_done(&cur);
1791 spin_unlock(&ailp->ail_lock);
1794 /******************************************************************************
1796 * Log recover routines
1798 ******************************************************************************
1800 static const struct xlog_recover_item_ops *xlog_recover_item_ops[] = {
1802 &xlog_inode_item_ops,
1803 &xlog_dquot_item_ops,
1804 &xlog_quotaoff_item_ops,
1805 &xlog_icreate_item_ops,
1816 static const struct xlog_recover_item_ops *
1818 struct xlog_recover_item *item)
1822 for (i = 0; i < ARRAY_SIZE(xlog_recover_item_ops); i++)
1823 if (ITEM_TYPE(item) == xlog_recover_item_ops[i]->item_type)
1824 return xlog_recover_item_ops[i];
1830 * Sort the log items in the transaction.
1832 * The ordering constraints are defined by the inode allocation and unlink
1833 * behaviour. The rules are:
1835 * 1. Every item is only logged once in a given transaction. Hence it
1836 * represents the last logged state of the item. Hence ordering is
1837 * dependent on the order in which operations need to be performed so
1838 * required initial conditions are always met.
1840 * 2. Cancelled buffers are recorded in pass 1 in a separate table and
1841 * there's nothing to replay from them so we can simply cull them
1842 * from the transaction. However, we can't do that until after we've
1843 * replayed all the other items because they may be dependent on the
1844 * cancelled buffer and replaying the cancelled buffer can remove it
1845 * form the cancelled buffer table. Hence they have tobe done last.
1847 * 3. Inode allocation buffers must be replayed before inode items that
1848 * read the buffer and replay changes into it. For filesystems using the
1849 * ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1850 * treated the same as inode allocation buffers as they create and
1851 * initialise the buffers directly.
1853 * 4. Inode unlink buffers must be replayed after inode items are replayed.
1854 * This ensures that inodes are completely flushed to the inode buffer
1855 * in a "free" state before we remove the unlinked inode list pointer.
1857 * Hence the ordering needs to be inode allocation buffers first, inode items
1858 * second, inode unlink buffers third and cancelled buffers last.
1860 * But there's a problem with that - we can't tell an inode allocation buffer
1861 * apart from a regular buffer, so we can't separate them. We can, however,
1862 * tell an inode unlink buffer from the others, and so we can separate them out
1863 * from all the other buffers and move them to last.
1865 * Hence, 4 lists, in order from head to tail:
1866 * - buffer_list for all buffers except cancelled/inode unlink buffers
1867 * - item_list for all non-buffer items
1868 * - inode_buffer_list for inode unlink buffers
1869 * - cancel_list for the cancelled buffers
1871 * Note that we add objects to the tail of the lists so that first-to-last
1872 * ordering is preserved within the lists. Adding objects to the head of the
1873 * list means when we traverse from the head we walk them in last-to-first
1874 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1875 * but for all other items there may be specific ordering that we need to
1879 xlog_recover_reorder_trans(
1881 struct xlog_recover *trans,
1884 struct xlog_recover_item *item, *n;
1886 LIST_HEAD(sort_list);
1887 LIST_HEAD(cancel_list);
1888 LIST_HEAD(buffer_list);
1889 LIST_HEAD(inode_buffer_list);
1890 LIST_HEAD(item_list);
1892 list_splice_init(&trans->r_itemq, &sort_list);
1893 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1894 enum xlog_recover_reorder fate = XLOG_REORDER_ITEM_LIST;
1896 item->ri_ops = xlog_find_item_ops(item);
1897 if (!item->ri_ops) {
1899 "%s: unrecognized type of log operation (%d)",
1900 __func__, ITEM_TYPE(item));
1903 * return the remaining items back to the transaction
1904 * item list so they can be freed in caller.
1906 if (!list_empty(&sort_list))
1907 list_splice_init(&sort_list, &trans->r_itemq);
1908 error = -EFSCORRUPTED;
1912 if (item->ri_ops->reorder)
1913 fate = item->ri_ops->reorder(item);
1916 case XLOG_REORDER_BUFFER_LIST:
1917 list_move_tail(&item->ri_list, &buffer_list);
1919 case XLOG_REORDER_CANCEL_LIST:
1920 trace_xfs_log_recover_item_reorder_head(log,
1922 list_move(&item->ri_list, &cancel_list);
1924 case XLOG_REORDER_INODE_BUFFER_LIST:
1925 list_move(&item->ri_list, &inode_buffer_list);
1927 case XLOG_REORDER_ITEM_LIST:
1928 trace_xfs_log_recover_item_reorder_tail(log,
1930 list_move_tail(&item->ri_list, &item_list);
1935 ASSERT(list_empty(&sort_list));
1936 if (!list_empty(&buffer_list))
1937 list_splice(&buffer_list, &trans->r_itemq);
1938 if (!list_empty(&item_list))
1939 list_splice_tail(&item_list, &trans->r_itemq);
1940 if (!list_empty(&inode_buffer_list))
1941 list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1942 if (!list_empty(&cancel_list))
1943 list_splice_tail(&cancel_list, &trans->r_itemq);
1952 const struct xfs_buf_ops *ops)
1954 if (!xlog_is_buffer_cancelled(log, blkno, len))
1955 xfs_buf_readahead(log->l_mp->m_ddev_targp, blkno, len, ops);
1959 xlog_recover_items_pass2(
1961 struct xlog_recover *trans,
1962 struct list_head *buffer_list,
1963 struct list_head *item_list)
1965 struct xlog_recover_item *item;
1968 list_for_each_entry(item, item_list, ri_list) {
1969 trace_xfs_log_recover_item_recover(log, trans, item,
1970 XLOG_RECOVER_PASS2);
1972 if (item->ri_ops->commit_pass2)
1973 error = item->ri_ops->commit_pass2(log, buffer_list,
1974 item, trans->r_lsn);
1983 * Perform the transaction.
1985 * If the transaction modifies a buffer or inode, do it now. Otherwise,
1986 * EFIs and EFDs get queued up by adding entries into the AIL for them.
1989 xlog_recover_commit_trans(
1991 struct xlog_recover *trans,
1993 struct list_head *buffer_list)
1996 int items_queued = 0;
1997 struct xlog_recover_item *item;
1998 struct xlog_recover_item *next;
1999 LIST_HEAD (ra_list);
2000 LIST_HEAD (done_list);
2002 #define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
2004 hlist_del_init(&trans->r_list);
2006 error = xlog_recover_reorder_trans(log, trans, pass);
2010 list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
2011 trace_xfs_log_recover_item_recover(log, trans, item, pass);
2014 case XLOG_RECOVER_PASS1:
2015 if (item->ri_ops->commit_pass1)
2016 error = item->ri_ops->commit_pass1(log, item);
2018 case XLOG_RECOVER_PASS2:
2019 if (item->ri_ops->ra_pass2)
2020 item->ri_ops->ra_pass2(log, item);
2021 list_move_tail(&item->ri_list, &ra_list);
2023 if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
2024 error = xlog_recover_items_pass2(log, trans,
2025 buffer_list, &ra_list);
2026 list_splice_tail_init(&ra_list, &done_list);
2040 if (!list_empty(&ra_list)) {
2042 error = xlog_recover_items_pass2(log, trans,
2043 buffer_list, &ra_list);
2044 list_splice_tail_init(&ra_list, &done_list);
2047 if (!list_empty(&done_list))
2048 list_splice_init(&done_list, &trans->r_itemq);
2054 xlog_recover_add_item(
2055 struct list_head *head)
2057 struct xlog_recover_item *item;
2059 item = kmem_zalloc(sizeof(struct xlog_recover_item), 0);
2060 INIT_LIST_HEAD(&item->ri_list);
2061 list_add_tail(&item->ri_list, head);
2065 xlog_recover_add_to_cont_trans(
2067 struct xlog_recover *trans,
2071 struct xlog_recover_item *item;
2072 char *ptr, *old_ptr;
2076 * If the transaction is empty, the header was split across this and the
2077 * previous record. Copy the rest of the header.
2079 if (list_empty(&trans->r_itemq)) {
2080 ASSERT(len <= sizeof(struct xfs_trans_header));
2081 if (len > sizeof(struct xfs_trans_header)) {
2082 xfs_warn(log->l_mp, "%s: bad header length", __func__);
2083 return -EFSCORRUPTED;
2086 xlog_recover_add_item(&trans->r_itemq);
2087 ptr = (char *)&trans->r_theader +
2088 sizeof(struct xfs_trans_header) - len;
2089 memcpy(ptr, dp, len);
2093 /* take the tail entry */
2094 item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2097 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
2098 old_len = item->ri_buf[item->ri_cnt-1].i_len;
2100 ptr = kmem_realloc(old_ptr, len + old_len, 0);
2101 memcpy(&ptr[old_len], dp, len);
2102 item->ri_buf[item->ri_cnt-1].i_len += len;
2103 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
2104 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
2109 * The next region to add is the start of a new region. It could be
2110 * a whole region or it could be the first part of a new region. Because
2111 * of this, the assumption here is that the type and size fields of all
2112 * format structures fit into the first 32 bits of the structure.
2114 * This works because all regions must be 32 bit aligned. Therefore, we
2115 * either have both fields or we have neither field. In the case we have
2116 * neither field, the data part of the region is zero length. We only have
2117 * a log_op_header and can throw away the header since a new one will appear
2118 * later. If we have at least 4 bytes, then we can determine how many regions
2119 * will appear in the current log item.
2122 xlog_recover_add_to_trans(
2124 struct xlog_recover *trans,
2128 struct xfs_inode_log_format *in_f; /* any will do */
2129 struct xlog_recover_item *item;
2134 if (list_empty(&trans->r_itemq)) {
2135 /* we need to catch log corruptions here */
2136 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
2137 xfs_warn(log->l_mp, "%s: bad header magic number",
2140 return -EFSCORRUPTED;
2143 if (len > sizeof(struct xfs_trans_header)) {
2144 xfs_warn(log->l_mp, "%s: bad header length", __func__);
2146 return -EFSCORRUPTED;
2150 * The transaction header can be arbitrarily split across op
2151 * records. If we don't have the whole thing here, copy what we
2152 * do have and handle the rest in the next record.
2154 if (len == sizeof(struct xfs_trans_header))
2155 xlog_recover_add_item(&trans->r_itemq);
2156 memcpy(&trans->r_theader, dp, len);
2160 ptr = kmem_alloc(len, 0);
2161 memcpy(ptr, dp, len);
2162 in_f = (struct xfs_inode_log_format *)ptr;
2164 /* take the tail entry */
2165 item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2167 if (item->ri_total != 0 &&
2168 item->ri_total == item->ri_cnt) {
2169 /* tail item is in use, get a new one */
2170 xlog_recover_add_item(&trans->r_itemq);
2171 item = list_entry(trans->r_itemq.prev,
2172 struct xlog_recover_item, ri_list);
2175 if (item->ri_total == 0) { /* first region to be added */
2176 if (in_f->ilf_size == 0 ||
2177 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
2179 "bad number of regions (%d) in inode log format",
2183 return -EFSCORRUPTED;
2186 item->ri_total = in_f->ilf_size;
2188 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
2192 if (item->ri_total <= item->ri_cnt) {
2194 "log item region count (%d) overflowed size (%d)",
2195 item->ri_cnt, item->ri_total);
2198 return -EFSCORRUPTED;
2201 /* Description region is ri_buf[0] */
2202 item->ri_buf[item->ri_cnt].i_addr = ptr;
2203 item->ri_buf[item->ri_cnt].i_len = len;
2205 trace_xfs_log_recover_item_add(log, trans, item, 0);
2210 * Free up any resources allocated by the transaction
2212 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2215 xlog_recover_free_trans(
2216 struct xlog_recover *trans)
2218 struct xlog_recover_item *item, *n;
2221 hlist_del_init(&trans->r_list);
2223 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2224 /* Free the regions in the item. */
2225 list_del(&item->ri_list);
2226 for (i = 0; i < item->ri_cnt; i++)
2227 kmem_free(item->ri_buf[i].i_addr);
2228 /* Free the item itself */
2229 kmem_free(item->ri_buf);
2232 /* Free the transaction recover structure */
2237 * On error or completion, trans is freed.
2240 xlog_recovery_process_trans(
2242 struct xlog_recover *trans,
2247 struct list_head *buffer_list)
2250 bool freeit = false;
2252 /* mask off ophdr transaction container flags */
2253 flags &= ~XLOG_END_TRANS;
2254 if (flags & XLOG_WAS_CONT_TRANS)
2255 flags &= ~XLOG_CONTINUE_TRANS;
2258 * Callees must not free the trans structure. We'll decide if we need to
2259 * free it or not based on the operation being done and it's result.
2262 /* expected flag values */
2264 case XLOG_CONTINUE_TRANS:
2265 error = xlog_recover_add_to_trans(log, trans, dp, len);
2267 case XLOG_WAS_CONT_TRANS:
2268 error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
2270 case XLOG_COMMIT_TRANS:
2271 error = xlog_recover_commit_trans(log, trans, pass,
2273 /* success or fail, we are now done with this transaction. */
2277 /* unexpected flag values */
2278 case XLOG_UNMOUNT_TRANS:
2279 /* just skip trans */
2280 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2283 case XLOG_START_TRANS:
2285 xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
2287 error = -EFSCORRUPTED;
2290 if (error || freeit)
2291 xlog_recover_free_trans(trans);
2296 * Lookup the transaction recovery structure associated with the ID in the
2297 * current ophdr. If the transaction doesn't exist and the start flag is set in
2298 * the ophdr, then allocate a new transaction for future ID matches to find.
2299 * Either way, return what we found during the lookup - an existing transaction
2302 STATIC struct xlog_recover *
2303 xlog_recover_ophdr_to_trans(
2304 struct hlist_head rhash[],
2305 struct xlog_rec_header *rhead,
2306 struct xlog_op_header *ohead)
2308 struct xlog_recover *trans;
2310 struct hlist_head *rhp;
2312 tid = be32_to_cpu(ohead->oh_tid);
2313 rhp = &rhash[XLOG_RHASH(tid)];
2314 hlist_for_each_entry(trans, rhp, r_list) {
2315 if (trans->r_log_tid == tid)
2320 * skip over non-start transaction headers - we could be
2321 * processing slack space before the next transaction starts
2323 if (!(ohead->oh_flags & XLOG_START_TRANS))
2326 ASSERT(be32_to_cpu(ohead->oh_len) == 0);
2329 * This is a new transaction so allocate a new recovery container to
2330 * hold the recovery ops that will follow.
2332 trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
2333 trans->r_log_tid = tid;
2334 trans->r_lsn = be64_to_cpu(rhead->h_lsn);
2335 INIT_LIST_HEAD(&trans->r_itemq);
2336 INIT_HLIST_NODE(&trans->r_list);
2337 hlist_add_head(&trans->r_list, rhp);
2340 * Nothing more to do for this ophdr. Items to be added to this new
2341 * transaction will be in subsequent ophdr containers.
2347 xlog_recover_process_ophdr(
2349 struct hlist_head rhash[],
2350 struct xlog_rec_header *rhead,
2351 struct xlog_op_header *ohead,
2355 struct list_head *buffer_list)
2357 struct xlog_recover *trans;
2361 /* Do we understand who wrote this op? */
2362 if (ohead->oh_clientid != XFS_TRANSACTION &&
2363 ohead->oh_clientid != XFS_LOG) {
2364 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2365 __func__, ohead->oh_clientid);
2367 return -EFSCORRUPTED;
2371 * Check the ophdr contains all the data it is supposed to contain.
2373 len = be32_to_cpu(ohead->oh_len);
2374 if (dp + len > end) {
2375 xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
2377 return -EFSCORRUPTED;
2380 trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
2382 /* nothing to do, so skip over this ophdr */
2387 * The recovered buffer queue is drained only once we know that all
2388 * recovery items for the current LSN have been processed. This is
2391 * - Buffer write submission updates the metadata LSN of the buffer.
2392 * - Log recovery skips items with a metadata LSN >= the current LSN of
2393 * the recovery item.
2394 * - Separate recovery items against the same metadata buffer can share
2395 * a current LSN. I.e., consider that the LSN of a recovery item is
2396 * defined as the starting LSN of the first record in which its
2397 * transaction appears, that a record can hold multiple transactions,
2398 * and/or that a transaction can span multiple records.
2400 * In other words, we are allowed to submit a buffer from log recovery
2401 * once per current LSN. Otherwise, we may incorrectly skip recovery
2402 * items and cause corruption.
2404 * We don't know up front whether buffers are updated multiple times per
2405 * LSN. Therefore, track the current LSN of each commit log record as it
2406 * is processed and drain the queue when it changes. Use commit records
2407 * because they are ordered correctly by the logging code.
2409 if (log->l_recovery_lsn != trans->r_lsn &&
2410 ohead->oh_flags & XLOG_COMMIT_TRANS) {
2411 error = xfs_buf_delwri_submit(buffer_list);
2414 log->l_recovery_lsn = trans->r_lsn;
2417 return xlog_recovery_process_trans(log, trans, dp, len,
2418 ohead->oh_flags, pass, buffer_list);
2422 * There are two valid states of the r_state field. 0 indicates that the
2423 * transaction structure is in a normal state. We have either seen the
2424 * start of the transaction or the last operation we added was not a partial
2425 * operation. If the last operation we added to the transaction was a
2426 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2428 * NOTE: skip LRs with 0 data length.
2431 xlog_recover_process_data(
2433 struct hlist_head rhash[],
2434 struct xlog_rec_header *rhead,
2437 struct list_head *buffer_list)
2439 struct xlog_op_header *ohead;
2444 end = dp + be32_to_cpu(rhead->h_len);
2445 num_logops = be32_to_cpu(rhead->h_num_logops);
2447 /* check the log format matches our own - else we can't recover */
2448 if (xlog_header_check_recover(log->l_mp, rhead))
2451 trace_xfs_log_recover_record(log, rhead, pass);
2452 while ((dp < end) && num_logops) {
2454 ohead = (struct xlog_op_header *)dp;
2455 dp += sizeof(*ohead);
2458 /* errors will abort recovery */
2459 error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
2460 dp, end, pass, buffer_list);
2464 dp += be32_to_cpu(ohead->oh_len);
2470 /* Take all the collected deferred ops and finish them in order. */
2472 xlog_finish_defer_ops(
2473 struct xfs_trans *parent_tp)
2475 struct xfs_mount *mp = parent_tp->t_mountp;
2476 struct xfs_trans *tp;
2482 * We're finishing the defer_ops that accumulated as a result of
2483 * recovering unfinished intent items during log recovery. We
2484 * reserve an itruncate transaction because it is the largest
2485 * permanent transaction type. Since we're the only user of the fs
2486 * right now, take 93% (15/16) of the available free blocks. Use
2487 * weird math to avoid a 64-bit division.
2489 freeblks = percpu_counter_sum(&mp->m_fdblocks);
2492 resblks = min_t(int64_t, UINT_MAX, freeblks);
2493 resblks = (resblks * 15) >> 4;
2494 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, resblks,
2495 0, XFS_TRANS_RESERVE, &tp);
2498 /* transfer all collected dfops to this transaction */
2499 xfs_defer_move(tp, parent_tp);
2501 return xfs_trans_commit(tp);
2504 /* Is this log item a deferred action intent? */
2505 static inline bool xlog_item_is_intent(struct xfs_log_item *lip)
2507 return lip->li_ops->iop_recover != NULL &&
2508 lip->li_ops->iop_match != NULL;
2512 * When this is called, all of the log intent items which did not have
2513 * corresponding log done items should be in the AIL. What we do now
2514 * is update the data structures associated with each one.
2516 * Since we process the log intent items in normal transactions, they
2517 * will be removed at some point after the commit. This prevents us
2518 * from just walking down the list processing each one. We'll use a
2519 * flag in the intent item to skip those that we've already processed
2520 * and use the AIL iteration mechanism's generation count to try to
2521 * speed this up at least a bit.
2523 * When we start, we know that the intents are the only things in the
2524 * AIL. As we process them, however, other items are added to the
2528 xlog_recover_process_intents(
2531 struct xfs_trans *parent_tp;
2532 struct xfs_ail_cursor cur;
2533 struct xfs_log_item *lip;
2534 struct xfs_ail *ailp;
2536 #if defined(DEBUG) || defined(XFS_WARN)
2541 * The intent recovery handlers commit transactions to complete recovery
2542 * for individual intents, but any new deferred operations that are
2543 * queued during that process are held off until the very end. The
2544 * purpose of this transaction is to serve as a container for deferred
2545 * operations. Each intent recovery handler must transfer dfops here
2546 * before its local transaction commits, and we'll finish the entire
2549 error = xfs_trans_alloc_empty(log->l_mp, &parent_tp);
2554 spin_lock(&ailp->ail_lock);
2555 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2556 #if defined(DEBUG) || defined(XFS_WARN)
2557 last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
2559 while (lip != NULL) {
2561 * We're done when we see something other than an intent.
2562 * There should be no intents left in the AIL now.
2564 if (!xlog_item_is_intent(lip)) {
2566 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2567 ASSERT(!xlog_item_is_intent(lip));
2573 * We should never see a redo item with a LSN higher than
2574 * the last transaction we found in the log at the start
2577 ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
2580 * NOTE: If your intent processing routine can create more
2581 * deferred ops, you /must/ attach them to the transaction in
2582 * this routine or else those subsequent intents will get
2583 * replayed in the wrong order!
2585 if (!test_and_set_bit(XFS_LI_RECOVERED, &lip->li_flags)) {
2586 spin_unlock(&ailp->ail_lock);
2587 error = lip->li_ops->iop_recover(lip, parent_tp);
2588 spin_lock(&ailp->ail_lock);
2592 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2595 xfs_trans_ail_cursor_done(&cur);
2596 spin_unlock(&ailp->ail_lock);
2598 error = xlog_finish_defer_ops(parent_tp);
2599 xfs_trans_cancel(parent_tp);
2605 * A cancel occurs when the mount has failed and we're bailing out.
2606 * Release all pending log intent items so they don't pin the AIL.
2609 xlog_recover_cancel_intents(
2612 struct xfs_log_item *lip;
2613 struct xfs_ail_cursor cur;
2614 struct xfs_ail *ailp;
2617 spin_lock(&ailp->ail_lock);
2618 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2619 while (lip != NULL) {
2621 * We're done when we see something other than an intent.
2622 * There should be no intents left in the AIL now.
2624 if (!xlog_item_is_intent(lip)) {
2626 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2627 ASSERT(!xlog_item_is_intent(lip));
2632 spin_unlock(&ailp->ail_lock);
2633 lip->li_ops->iop_release(lip);
2634 spin_lock(&ailp->ail_lock);
2635 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2638 xfs_trans_ail_cursor_done(&cur);
2639 spin_unlock(&ailp->ail_lock);
2643 * This routine performs a transaction to null out a bad inode pointer
2644 * in an agi unlinked inode hash bucket.
2647 xlog_recover_clear_agi_bucket(
2649 xfs_agnumber_t agno,
2658 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
2662 error = xfs_read_agi(mp, tp, agno, &agibp);
2666 agi = agibp->b_addr;
2667 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
2668 offset = offsetof(xfs_agi_t, agi_unlinked) +
2669 (sizeof(xfs_agino_t) * bucket);
2670 xfs_trans_log_buf(tp, agibp, offset,
2671 (offset + sizeof(xfs_agino_t) - 1));
2673 error = xfs_trans_commit(tp);
2679 xfs_trans_cancel(tp);
2681 xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
2686 xlog_recover_process_one_iunlink(
2687 struct xfs_mount *mp,
2688 xfs_agnumber_t agno,
2692 struct xfs_buf *ibp;
2693 struct xfs_dinode *dip;
2694 struct xfs_inode *ip;
2698 ino = XFS_AGINO_TO_INO(mp, agno, agino);
2699 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
2704 * Get the on disk inode to find the next inode in the bucket.
2706 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0);
2710 xfs_iflags_clear(ip, XFS_IRECOVERY);
2711 ASSERT(VFS_I(ip)->i_nlink == 0);
2712 ASSERT(VFS_I(ip)->i_mode != 0);
2714 /* setup for the next pass */
2715 agino = be32_to_cpu(dip->di_next_unlinked);
2719 * Prevent any DMAPI event from being sent when the reference on
2720 * the inode is dropped.
2722 ip->i_d.di_dmevmask = 0;
2731 * We can't read in the inode this bucket points to, or this inode
2732 * is messed up. Just ditch this bucket of inodes. We will lose
2733 * some inodes and space, but at least we won't hang.
2735 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
2736 * clear the inode pointer in the bucket.
2738 xlog_recover_clear_agi_bucket(mp, agno, bucket);
2743 * Recover AGI unlinked lists
2745 * This is called during recovery to process any inodes which we unlinked but
2746 * not freed when the system crashed. These inodes will be on the lists in the
2747 * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
2748 * any inodes found on the lists. Each inode is removed from the lists when it
2749 * has been fully truncated and is freed. The freeing of the inode and its
2750 * removal from the list must be atomic.
2752 * If everything we touch in the agi processing loop is already in memory, this
2753 * loop can hold the cpu for a long time. It runs without lock contention,
2754 * memory allocation contention, the need wait for IO, etc, and so will run
2755 * until we either run out of inodes to process, run low on memory or we run out
2758 * This behaviour is bad for latency on single CPU and non-preemptible kernels,
2759 * and can prevent other filesytem work (such as CIL pushes) from running. This
2760 * can lead to deadlocks if the recovery process runs out of log reservation
2761 * space. Hence we need to yield the CPU when there is other kernel work
2762 * scheduled on this CPU to ensure other scheduled work can run without undue
2766 xlog_recover_process_iunlinks(
2770 xfs_agnumber_t agno;
2779 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
2781 * Find the agi for this ag.
2783 error = xfs_read_agi(mp, NULL, agno, &agibp);
2786 * AGI is b0rked. Don't process it.
2788 * We should probably mark the filesystem as corrupt
2789 * after we've recovered all the ag's we can....
2794 * Unlock the buffer so that it can be acquired in the normal
2795 * course of the transaction to truncate and free each inode.
2796 * Because we are not racing with anyone else here for the AGI
2797 * buffer, we don't even need to hold it locked to read the
2798 * initial unlinked bucket entries out of the buffer. We keep
2799 * buffer reference though, so that it stays pinned in memory
2800 * while we need the buffer.
2802 agi = agibp->b_addr;
2803 xfs_buf_unlock(agibp);
2805 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
2806 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2807 while (agino != NULLAGINO) {
2808 agino = xlog_recover_process_one_iunlink(mp,
2809 agno, agino, bucket);
2813 xfs_buf_rele(agibp);
2819 struct xlog_rec_header *rhead,
2825 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
2826 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
2827 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
2831 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
2832 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
2833 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
2834 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2835 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2836 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
2843 * CRC check, unpack and process a log record.
2846 xlog_recover_process(
2848 struct hlist_head rhash[],
2849 struct xlog_rec_header *rhead,
2852 struct list_head *buffer_list)
2854 __le32 old_crc = rhead->h_crc;
2857 crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
2860 * Nothing else to do if this is a CRC verification pass. Just return
2861 * if this a record with a non-zero crc. Unfortunately, mkfs always
2862 * sets old_crc to 0 so we must consider this valid even on v5 supers.
2863 * Otherwise, return EFSBADCRC on failure so the callers up the stack
2864 * know precisely what failed.
2866 if (pass == XLOG_RECOVER_CRCPASS) {
2867 if (old_crc && crc != old_crc)
2873 * We're in the normal recovery path. Issue a warning if and only if the
2874 * CRC in the header is non-zero. This is an advisory warning and the
2875 * zero CRC check prevents warnings from being emitted when upgrading
2876 * the kernel from one that does not add CRCs by default.
2878 if (crc != old_crc) {
2879 if (old_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
2880 xfs_alert(log->l_mp,
2881 "log record CRC mismatch: found 0x%x, expected 0x%x.",
2882 le32_to_cpu(old_crc),
2884 xfs_hex_dump(dp, 32);
2888 * If the filesystem is CRC enabled, this mismatch becomes a
2889 * fatal log corruption failure.
2891 if (xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
2892 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
2893 return -EFSCORRUPTED;
2897 xlog_unpack_data(rhead, dp, log);
2899 return xlog_recover_process_data(log, rhash, rhead, dp, pass,
2904 xlog_valid_rec_header(
2906 struct xlog_rec_header *rhead,
2911 if (XFS_IS_CORRUPT(log->l_mp,
2912 rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)))
2913 return -EFSCORRUPTED;
2914 if (XFS_IS_CORRUPT(log->l_mp,
2915 (!rhead->h_version ||
2916 (be32_to_cpu(rhead->h_version) &
2917 (~XLOG_VERSION_OKBITS))))) {
2918 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
2919 __func__, be32_to_cpu(rhead->h_version));
2920 return -EFSCORRUPTED;
2923 /* LR body must have data or it wouldn't have been written */
2924 hlen = be32_to_cpu(rhead->h_len);
2925 if (XFS_IS_CORRUPT(log->l_mp, hlen <= 0 || hlen > INT_MAX))
2926 return -EFSCORRUPTED;
2927 if (XFS_IS_CORRUPT(log->l_mp,
2928 blkno > log->l_logBBsize || blkno > INT_MAX))
2929 return -EFSCORRUPTED;
2934 * Read the log from tail to head and process the log records found.
2935 * Handle the two cases where the tail and head are in the same cycle
2936 * and where the active portion of the log wraps around the end of
2937 * the physical log separately. The pass parameter is passed through
2938 * to the routines called to process the data and is not looked at
2942 xlog_do_recovery_pass(
2944 xfs_daddr_t head_blk,
2945 xfs_daddr_t tail_blk,
2947 xfs_daddr_t *first_bad) /* out: first bad log rec */
2949 xlog_rec_header_t *rhead;
2950 xfs_daddr_t blk_no, rblk_no;
2951 xfs_daddr_t rhead_blk;
2954 int error = 0, h_size, h_len;
2956 int bblks, split_bblks;
2957 int hblks, split_hblks, wrapped_hblks;
2959 struct hlist_head rhash[XLOG_RHASH_SIZE];
2960 LIST_HEAD (buffer_list);
2962 ASSERT(head_blk != tail_blk);
2963 blk_no = rhead_blk = tail_blk;
2965 for (i = 0; i < XLOG_RHASH_SIZE; i++)
2966 INIT_HLIST_HEAD(&rhash[i]);
2969 * Read the header of the tail block and get the iclog buffer size from
2970 * h_size. Use this to tell how many sectors make up the log header.
2972 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
2974 * When using variable length iclogs, read first sector of
2975 * iclog header and extract the header size from it. Get a
2976 * new hbp that is the correct size.
2978 hbp = xlog_alloc_buffer(log, 1);
2982 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
2986 rhead = (xlog_rec_header_t *)offset;
2987 error = xlog_valid_rec_header(log, rhead, tail_blk);
2992 * xfsprogs has a bug where record length is based on lsunit but
2993 * h_size (iclog size) is hardcoded to 32k. Now that we
2994 * unconditionally CRC verify the unmount record, this means the
2995 * log buffer can be too small for the record and cause an
2998 * Detect this condition here. Use lsunit for the buffer size as
2999 * long as this looks like the mkfs case. Otherwise, return an
3000 * error to avoid a buffer overrun.
3002 h_size = be32_to_cpu(rhead->h_size);
3003 h_len = be32_to_cpu(rhead->h_len);
3004 if (h_len > h_size) {
3005 if (h_len <= log->l_mp->m_logbsize &&
3006 be32_to_cpu(rhead->h_num_logops) == 1) {
3008 "invalid iclog size (%d bytes), using lsunit (%d bytes)",
3009 h_size, log->l_mp->m_logbsize);
3010 h_size = log->l_mp->m_logbsize;
3012 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW,
3014 error = -EFSCORRUPTED;
3019 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3020 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3021 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3022 if (h_size % XLOG_HEADER_CYCLE_SIZE)
3025 hbp = xlog_alloc_buffer(log, hblks);
3030 ASSERT(log->l_sectBBsize == 1);
3032 hbp = xlog_alloc_buffer(log, 1);
3033 h_size = XLOG_BIG_RECORD_BSIZE;
3038 dbp = xlog_alloc_buffer(log, BTOBB(h_size));
3044 memset(rhash, 0, sizeof(rhash));
3045 if (tail_blk > head_blk) {
3047 * Perform recovery around the end of the physical log.
3048 * When the head is not on the same cycle number as the tail,
3049 * we can't do a sequential recovery.
3051 while (blk_no < log->l_logBBsize) {
3053 * Check for header wrapping around physical end-of-log
3058 if (blk_no + hblks <= log->l_logBBsize) {
3059 /* Read header in one read */
3060 error = xlog_bread(log, blk_no, hblks, hbp,
3065 /* This LR is split across physical log end */
3066 if (blk_no != log->l_logBBsize) {
3067 /* some data before physical log end */
3068 ASSERT(blk_no <= INT_MAX);
3069 split_hblks = log->l_logBBsize - (int)blk_no;
3070 ASSERT(split_hblks > 0);
3071 error = xlog_bread(log, blk_no,
3079 * Note: this black magic still works with
3080 * large sector sizes (non-512) only because:
3081 * - we increased the buffer size originally
3082 * by 1 sector giving us enough extra space
3083 * for the second read;
3084 * - the log start is guaranteed to be sector
3086 * - we read the log end (LR header start)
3087 * _first_, then the log start (LR header end)
3088 * - order is important.
3090 wrapped_hblks = hblks - split_hblks;
3091 error = xlog_bread_noalign(log, 0,
3093 offset + BBTOB(split_hblks));
3097 rhead = (xlog_rec_header_t *)offset;
3098 error = xlog_valid_rec_header(log, rhead,
3099 split_hblks ? blk_no : 0);
3103 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3107 * Read the log record data in multiple reads if it
3108 * wraps around the end of the log. Note that if the
3109 * header already wrapped, blk_no could point past the
3110 * end of the log. The record data is contiguous in
3113 if (blk_no + bblks <= log->l_logBBsize ||
3114 blk_no >= log->l_logBBsize) {
3115 rblk_no = xlog_wrap_logbno(log, blk_no);
3116 error = xlog_bread(log, rblk_no, bblks, dbp,
3121 /* This log record is split across the
3122 * physical end of log */
3125 if (blk_no != log->l_logBBsize) {
3126 /* some data is before the physical
3128 ASSERT(!wrapped_hblks);
3129 ASSERT(blk_no <= INT_MAX);
3131 log->l_logBBsize - (int)blk_no;
3132 ASSERT(split_bblks > 0);
3133 error = xlog_bread(log, blk_no,
3141 * Note: this black magic still works with
3142 * large sector sizes (non-512) only because:
3143 * - we increased the buffer size originally
3144 * by 1 sector giving us enough extra space
3145 * for the second read;
3146 * - the log start is guaranteed to be sector
3148 * - we read the log end (LR header start)
3149 * _first_, then the log start (LR header end)
3150 * - order is important.
3152 error = xlog_bread_noalign(log, 0,
3153 bblks - split_bblks,
3154 offset + BBTOB(split_bblks));
3159 error = xlog_recover_process(log, rhash, rhead, offset,
3160 pass, &buffer_list);
3168 ASSERT(blk_no >= log->l_logBBsize);
3169 blk_no -= log->l_logBBsize;
3173 /* read first part of physical log */
3174 while (blk_no < head_blk) {
3175 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3179 rhead = (xlog_rec_header_t *)offset;
3180 error = xlog_valid_rec_header(log, rhead, blk_no);
3184 /* blocks in data section */
3185 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3186 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3191 error = xlog_recover_process(log, rhash, rhead, offset, pass,
3196 blk_no += bblks + hblks;
3206 * Submit buffers that have been added from the last record processed,
3207 * regardless of error status.
3209 if (!list_empty(&buffer_list))
3210 error2 = xfs_buf_delwri_submit(&buffer_list);
3212 if (error && first_bad)
3213 *first_bad = rhead_blk;
3216 * Transactions are freed at commit time but transactions without commit
3217 * records on disk are never committed. Free any that may be left in the
3220 for (i = 0; i < XLOG_RHASH_SIZE; i++) {
3221 struct hlist_node *tmp;
3222 struct xlog_recover *trans;
3224 hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
3225 xlog_recover_free_trans(trans);
3228 return error ? error : error2;
3232 * Do the recovery of the log. We actually do this in two phases.
3233 * The two passes are necessary in order to implement the function
3234 * of cancelling a record written into the log. The first pass
3235 * determines those things which have been cancelled, and the
3236 * second pass replays log items normally except for those which
3237 * have been cancelled. The handling of the replay and cancellations
3238 * takes place in the log item type specific routines.
3240 * The table of items which have cancel records in the log is allocated
3241 * and freed at this level, since only here do we know when all of
3242 * the log recovery has been completed.
3245 xlog_do_log_recovery(
3247 xfs_daddr_t head_blk,
3248 xfs_daddr_t tail_blk)
3252 ASSERT(head_blk != tail_blk);
3255 * First do a pass to find all of the cancelled buf log items.
3256 * Store them in the buf_cancel_table for use in the second pass.
3258 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3259 sizeof(struct list_head),
3261 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3262 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3264 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3265 XLOG_RECOVER_PASS1, NULL);
3267 kmem_free(log->l_buf_cancel_table);
3268 log->l_buf_cancel_table = NULL;
3272 * Then do a second pass to actually recover the items in the log.
3273 * When it is complete free the table of buf cancel items.
3275 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3276 XLOG_RECOVER_PASS2, NULL);
3281 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3282 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3286 kmem_free(log->l_buf_cancel_table);
3287 log->l_buf_cancel_table = NULL;
3293 * Do the actual recovery
3298 xfs_daddr_t head_blk,
3299 xfs_daddr_t tail_blk)
3301 struct xfs_mount *mp = log->l_mp;
3306 trace_xfs_log_recover(log, head_blk, tail_blk);
3309 * First replay the images in the log.
3311 error = xlog_do_log_recovery(log, head_blk, tail_blk);
3316 * If IO errors happened during recovery, bail out.
3318 if (XFS_FORCED_SHUTDOWN(mp)) {
3323 * We now update the tail_lsn since much of the recovery has completed
3324 * and there may be space available to use. If there were no extent
3325 * or iunlinks, we can free up the entire log and set the tail_lsn to
3326 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3327 * lsn of the last known good LR on disk. If there are extent frees
3328 * or iunlinks they will have some entries in the AIL; so we look at
3329 * the AIL to determine how to set the tail_lsn.
3331 xlog_assign_tail_lsn(mp);
3334 * Now that we've finished replaying all buffer and inode
3335 * updates, re-read in the superblock and reverify it.
3338 bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
3339 ASSERT(!(bp->b_flags & XBF_WRITE));
3340 bp->b_flags |= XBF_READ;
3341 bp->b_ops = &xfs_sb_buf_ops;
3343 error = xfs_buf_submit(bp);
3345 if (!XFS_FORCED_SHUTDOWN(mp)) {
3346 xfs_buf_ioerror_alert(bp, __this_address);
3353 /* Convert superblock from on-disk format */
3355 xfs_sb_from_disk(sbp, bp->b_addr);
3358 /* re-initialise in-core superblock and geometry structures */
3359 xfs_reinit_percpu_counters(mp);
3360 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
3362 xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
3365 mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
3367 xlog_recover_check_summary(log);
3369 /* Normal transactions can now occur */
3370 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3375 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3377 * Return error or zero.
3383 xfs_daddr_t head_blk, tail_blk;
3386 /* find the tail of the log */
3387 error = xlog_find_tail(log, &head_blk, &tail_blk);
3392 * The superblock was read before the log was available and thus the LSN
3393 * could not be verified. Check the superblock LSN against the current
3394 * LSN now that it's known.
3396 if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
3397 !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
3400 if (tail_blk != head_blk) {
3401 /* There used to be a comment here:
3403 * disallow recovery on read-only mounts. note -- mount
3404 * checks for ENOSPC and turns it into an intelligent
3406 * ...but this is no longer true. Now, unless you specify
3407 * NORECOVERY (in which case this function would never be
3408 * called), we just go ahead and recover. We do this all
3409 * under the vfs layer, so we can get away with it unless
3410 * the device itself is read-only, in which case we fail.
3412 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3417 * Version 5 superblock log feature mask validation. We know the
3418 * log is dirty so check if there are any unknown log features
3419 * in what we need to recover. If there are unknown features
3420 * (e.g. unsupported transactions, then simply reject the
3421 * attempt at recovery before touching anything.
3423 if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
3424 xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
3425 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
3427 "Superblock has unknown incompatible log features (0x%x) enabled.",
3428 (log->l_mp->m_sb.sb_features_log_incompat &
3429 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
3431 "The log can not be fully and/or safely recovered by this kernel.");
3433 "Please recover the log on a kernel that supports the unknown features.");
3438 * Delay log recovery if the debug hook is set. This is debug
3439 * instrumention to coordinate simulation of I/O failures with
3442 if (xfs_globals.log_recovery_delay) {
3443 xfs_notice(log->l_mp,
3444 "Delaying log recovery for %d seconds.",
3445 xfs_globals.log_recovery_delay);
3446 msleep(xfs_globals.log_recovery_delay * 1000);
3449 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3450 log->l_mp->m_logname ? log->l_mp->m_logname
3453 error = xlog_do_recover(log, head_blk, tail_blk);
3454 log->l_flags |= XLOG_RECOVERY_NEEDED;
3460 * In the first part of recovery we replay inodes and buffers and build
3461 * up the list of extent free items which need to be processed. Here
3462 * we process the extent free items and clean up the on disk unlinked
3463 * inode lists. This is separated from the first part of recovery so
3464 * that the root and real-time bitmap inodes can be read in from disk in
3465 * between the two stages. This is necessary so that we can free space
3466 * in the real-time portion of the file system.
3469 xlog_recover_finish(
3473 * Now we're ready to do the transactions needed for the
3474 * rest of recovery. Start with completing all the extent
3475 * free intent records and then process the unlinked inode
3476 * lists. At this point, we essentially run in normal mode
3477 * except that we're still performing recovery actions
3478 * rather than accepting new requests.
3480 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3482 error = xlog_recover_process_intents(log);
3484 xfs_alert(log->l_mp, "Failed to recover intents");
3489 * Sync the log to get all the intents out of the AIL.
3490 * This isn't absolutely necessary, but it helps in
3491 * case the unlink transactions would have problems
3492 * pushing the intents out of the way.
3494 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3496 xlog_recover_process_iunlinks(log);
3498 xlog_recover_check_summary(log);
3500 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
3501 log->l_mp->m_logname ? log->l_mp->m_logname
3503 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3505 xfs_info(log->l_mp, "Ending clean mount");
3511 xlog_recover_cancel(
3514 if (log->l_flags & XLOG_RECOVERY_NEEDED)
3515 xlog_recover_cancel_intents(log);
3520 * Read all of the agf and agi counters and check that they
3521 * are consistent with the superblock counters.
3524 xlog_recover_check_summary(
3530 xfs_agnumber_t agno;
3541 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3542 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
3544 xfs_alert(mp, "%s agf read failed agno %d error %d",
3545 __func__, agno, error);
3547 struct xfs_agf *agfp = agfbp->b_addr;
3549 freeblks += be32_to_cpu(agfp->agf_freeblks) +
3550 be32_to_cpu(agfp->agf_flcount);
3551 xfs_buf_relse(agfbp);
3554 error = xfs_read_agi(mp, NULL, agno, &agibp);
3556 xfs_alert(mp, "%s agi read failed agno %d error %d",
3557 __func__, agno, error);
3559 struct xfs_agi *agi = agibp->b_addr;
3561 itotal += be32_to_cpu(agi->agi_count);
3562 ifree += be32_to_cpu(agi->agi_freecount);
3563 xfs_buf_relse(agibp);
projects / linux.git / blob