2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_btree.h"
33 #include "xfs_dinode.h"
34 #include "xfs_inode.h"
35 #include "xfs_inode_item.h"
36 #include "xfs_alloc.h"
37 #include "xfs_ialloc.h"
38 #include "xfs_log_priv.h"
39 #include "xfs_buf_item.h"
40 #include "xfs_log_recover.h"
41 #include "xfs_extfree_item.h"
42 #include "xfs_trans_priv.h"
43 #include "xfs_quota.h"
44 #include "xfs_utils.h"
45 #include "xfs_cksum.h"
46 #include "xfs_trace.h"
47 #include "xfs_icache.h"
49 /* Need all the magic numbers and buffer ops structures from these headers */
50 #include "xfs_symlink.h"
51 #include "xfs_da_btree.h"
52 #include "xfs_dir2_format.h"
53 #include "xfs_dir2_priv.h"
54 #include "xfs_attr_leaf.h"
55 #include "xfs_attr_remote.h"
62 xlog_clear_stale_blocks(
67 xlog_recover_check_summary(
70 #define xlog_recover_check_summary(log)
74 * This structure is used during recovery to record the buf log items which
75 * have been canceled and should not be replayed.
77 struct xfs_buf_cancel {
81 struct list_head bc_list;
85 * Sector aligned buffer routines for buffer create/read/write/access
89 * Verify the given count of basic blocks is valid number of blocks
90 * to specify for an operation involving the given XFS log buffer.
91 * Returns nonzero if the count is valid, 0 otherwise.
95 xlog_buf_bbcount_valid(
99 return bbcount > 0 && bbcount <= log->l_logBBsize;
103 * Allocate a buffer to hold log data. The buffer needs to be able
104 * to map to a range of nbblks basic blocks at any valid (basic
105 * block) offset within the log.
114 if (!xlog_buf_bbcount_valid(log, nbblks)) {
115 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
117 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
122 * We do log I/O in units of log sectors (a power-of-2
123 * multiple of the basic block size), so we round up the
124 * requested size to accommodate the basic blocks required
125 * for complete log sectors.
127 * In addition, the buffer may be used for a non-sector-
128 * aligned block offset, in which case an I/O of the
129 * requested size could extend beyond the end of the
130 * buffer. If the requested size is only 1 basic block it
131 * will never straddle a sector boundary, so this won't be
132 * an issue. Nor will this be a problem if the log I/O is
133 * done in basic blocks (sector size 1). But otherwise we
134 * extend the buffer by one extra log sector to ensure
135 * there's space to accommodate this possibility.
137 if (nbblks > 1 && log->l_sectBBsize > 1)
138 nbblks += log->l_sectBBsize;
139 nbblks = round_up(nbblks, log->l_sectBBsize);
141 bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, nbblks, 0);
155 * Return the address of the start of the given block number's data
156 * in a log buffer. The buffer covers a log sector-aligned region.
165 xfs_daddr_t offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
167 ASSERT(offset + nbblks <= bp->b_length);
168 return bp->b_addr + BBTOB(offset);
173 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
184 if (!xlog_buf_bbcount_valid(log, nbblks)) {
185 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
187 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
191 blk_no = round_down(blk_no, log->l_sectBBsize);
192 nbblks = round_up(nbblks, log->l_sectBBsize);
195 ASSERT(nbblks <= bp->b_length);
197 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
199 bp->b_io_length = nbblks;
202 xfsbdstrat(log->l_mp, bp);
203 error = xfs_buf_iowait(bp);
205 xfs_buf_ioerror_alert(bp, __func__);
219 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
223 *offset = xlog_align(log, blk_no, nbblks, bp);
228 * Read at an offset into the buffer. Returns with the buffer in it's original
229 * state regardless of the result of the read.
234 xfs_daddr_t blk_no, /* block to read from */
235 int nbblks, /* blocks to read */
239 xfs_caddr_t orig_offset = bp->b_addr;
240 int orig_len = BBTOB(bp->b_length);
243 error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
247 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
249 /* must reset buffer pointer even on error */
250 error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
257 * Write out the buffer at the given block for the given number of blocks.
258 * The buffer is kept locked across the write and is returned locked.
259 * This can only be used for synchronous log writes.
270 if (!xlog_buf_bbcount_valid(log, nbblks)) {
271 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
273 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
277 blk_no = round_down(blk_no, log->l_sectBBsize);
278 nbblks = round_up(nbblks, log->l_sectBBsize);
281 ASSERT(nbblks <= bp->b_length);
283 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
284 XFS_BUF_ZEROFLAGS(bp);
287 bp->b_io_length = nbblks;
290 error = xfs_bwrite(bp);
292 xfs_buf_ioerror_alert(bp, __func__);
299 * dump debug superblock and log record information
302 xlog_header_check_dump(
304 xlog_rec_header_t *head)
306 xfs_debug(mp, "%s: SB : uuid = %pU, fmt = %d\n",
307 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
308 xfs_debug(mp, " log : uuid = %pU, fmt = %d\n",
309 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
312 #define xlog_header_check_dump(mp, head)
316 * check log record header for recovery
319 xlog_header_check_recover(
321 xlog_rec_header_t *head)
323 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
326 * IRIX doesn't write the h_fmt field and leaves it zeroed
327 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
328 * a dirty log created in IRIX.
330 if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
332 "dirty log written in incompatible format - can't recover");
333 xlog_header_check_dump(mp, head);
334 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
335 XFS_ERRLEVEL_HIGH, mp);
336 return XFS_ERROR(EFSCORRUPTED);
337 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
339 "dirty log entry has mismatched uuid - can't recover");
340 xlog_header_check_dump(mp, head);
341 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
342 XFS_ERRLEVEL_HIGH, mp);
343 return XFS_ERROR(EFSCORRUPTED);
349 * read the head block of the log and check the header
352 xlog_header_check_mount(
354 xlog_rec_header_t *head)
356 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
358 if (uuid_is_nil(&head->h_fs_uuid)) {
360 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
361 * h_fs_uuid is nil, we assume this log was last mounted
362 * by IRIX and continue.
364 xfs_warn(mp, "nil uuid in log - IRIX style log");
365 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
366 xfs_warn(mp, "log has mismatched uuid - can't recover");
367 xlog_header_check_dump(mp, head);
368 XFS_ERROR_REPORT("xlog_header_check_mount",
369 XFS_ERRLEVEL_HIGH, mp);
370 return XFS_ERROR(EFSCORRUPTED);
381 * We're not going to bother about retrying
382 * this during recovery. One strike!
384 xfs_buf_ioerror_alert(bp, __func__);
385 xfs_force_shutdown(bp->b_target->bt_mount,
386 SHUTDOWN_META_IO_ERROR);
389 xfs_buf_ioend(bp, 0);
393 * This routine finds (to an approximation) the first block in the physical
394 * log which contains the given cycle. It uses a binary search algorithm.
395 * Note that the algorithm can not be perfect because the disk will not
396 * necessarily be perfect.
399 xlog_find_cycle_start(
402 xfs_daddr_t first_blk,
403 xfs_daddr_t *last_blk,
413 mid_blk = BLK_AVG(first_blk, end_blk);
414 while (mid_blk != first_blk && mid_blk != end_blk) {
415 error = xlog_bread(log, mid_blk, 1, bp, &offset);
418 mid_cycle = xlog_get_cycle(offset);
419 if (mid_cycle == cycle)
420 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
422 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
423 mid_blk = BLK_AVG(first_blk, end_blk);
425 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
426 (mid_blk == end_blk && mid_blk-1 == first_blk));
434 * Check that a range of blocks does not contain stop_on_cycle_no.
435 * Fill in *new_blk with the block offset where such a block is
436 * found, or with -1 (an invalid block number) if there is no such
437 * block in the range. The scan needs to occur from front to back
438 * and the pointer into the region must be updated since a later
439 * routine will need to perform another test.
442 xlog_find_verify_cycle(
444 xfs_daddr_t start_blk,
446 uint stop_on_cycle_no,
447 xfs_daddr_t *new_blk)
453 xfs_caddr_t buf = NULL;
457 * Greedily allocate a buffer big enough to handle the full
458 * range of basic blocks we'll be examining. If that fails,
459 * try a smaller size. We need to be able to read at least
460 * a log sector, or we're out of luck.
462 bufblks = 1 << ffs(nbblks);
463 while (bufblks > log->l_logBBsize)
465 while (!(bp = xlog_get_bp(log, bufblks))) {
467 if (bufblks < log->l_sectBBsize)
471 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
474 bcount = min(bufblks, (start_blk + nbblks - i));
476 error = xlog_bread(log, i, bcount, bp, &buf);
480 for (j = 0; j < bcount; j++) {
481 cycle = xlog_get_cycle(buf);
482 if (cycle == stop_on_cycle_no) {
499 * Potentially backup over partial log record write.
501 * In the typical case, last_blk is the number of the block directly after
502 * a good log record. Therefore, we subtract one to get the block number
503 * of the last block in the given buffer. extra_bblks contains the number
504 * of blocks we would have read on a previous read. This happens when the
505 * last log record is split over the end of the physical log.
507 * extra_bblks is the number of blocks potentially verified on a previous
508 * call to this routine.
511 xlog_find_verify_log_record(
513 xfs_daddr_t start_blk,
514 xfs_daddr_t *last_blk,
519 xfs_caddr_t offset = NULL;
520 xlog_rec_header_t *head = NULL;
523 int num_blks = *last_blk - start_blk;
526 ASSERT(start_blk != 0 || *last_blk != start_blk);
528 if (!(bp = xlog_get_bp(log, num_blks))) {
529 if (!(bp = xlog_get_bp(log, 1)))
533 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
536 offset += ((num_blks - 1) << BBSHIFT);
539 for (i = (*last_blk) - 1; i >= 0; i--) {
541 /* valid log record not found */
543 "Log inconsistent (didn't find previous header)");
545 error = XFS_ERROR(EIO);
550 error = xlog_bread(log, i, 1, bp, &offset);
555 head = (xlog_rec_header_t *)offset;
557 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
565 * We hit the beginning of the physical log & still no header. Return
566 * to caller. If caller can handle a return of -1, then this routine
567 * will be called again for the end of the physical log.
575 * We have the final block of the good log (the first block
576 * of the log record _before_ the head. So we check the uuid.
578 if ((error = xlog_header_check_mount(log->l_mp, head)))
582 * We may have found a log record header before we expected one.
583 * last_blk will be the 1st block # with a given cycle #. We may end
584 * up reading an entire log record. In this case, we don't want to
585 * reset last_blk. Only when last_blk points in the middle of a log
586 * record do we update last_blk.
588 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
589 uint h_size = be32_to_cpu(head->h_size);
591 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
592 if (h_size % XLOG_HEADER_CYCLE_SIZE)
598 if (*last_blk - i + extra_bblks !=
599 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
608 * Head is defined to be the point of the log where the next log write
609 * write could go. This means that incomplete LR writes at the end are
610 * eliminated when calculating the head. We aren't guaranteed that previous
611 * LR have complete transactions. We only know that a cycle number of
612 * current cycle number -1 won't be present in the log if we start writing
613 * from our current block number.
615 * last_blk contains the block number of the first block with a given
618 * Return: zero if normal, non-zero if error.
623 xfs_daddr_t *return_head_blk)
627 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
629 uint first_half_cycle, last_half_cycle;
631 int error, log_bbnum = log->l_logBBsize;
633 /* Is the end of the log device zeroed? */
634 if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
635 *return_head_blk = first_blk;
637 /* Is the whole lot zeroed? */
639 /* Linux XFS shouldn't generate totally zeroed logs -
640 * mkfs etc write a dummy unmount record to a fresh
641 * log so we can store the uuid in there
643 xfs_warn(log->l_mp, "totally zeroed log");
648 xfs_warn(log->l_mp, "empty log check failed");
652 first_blk = 0; /* get cycle # of 1st block */
653 bp = xlog_get_bp(log, 1);
657 error = xlog_bread(log, 0, 1, bp, &offset);
661 first_half_cycle = xlog_get_cycle(offset);
663 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
664 error = xlog_bread(log, last_blk, 1, bp, &offset);
668 last_half_cycle = xlog_get_cycle(offset);
669 ASSERT(last_half_cycle != 0);
672 * If the 1st half cycle number is equal to the last half cycle number,
673 * then the entire log is stamped with the same cycle number. In this
674 * case, head_blk can't be set to zero (which makes sense). The below
675 * math doesn't work out properly with head_blk equal to zero. Instead,
676 * we set it to log_bbnum which is an invalid block number, but this
677 * value makes the math correct. If head_blk doesn't changed through
678 * all the tests below, *head_blk is set to zero at the very end rather
679 * than log_bbnum. In a sense, log_bbnum and zero are the same block
680 * in a circular file.
682 if (first_half_cycle == last_half_cycle) {
684 * In this case we believe that the entire log should have
685 * cycle number last_half_cycle. We need to scan backwards
686 * from the end verifying that there are no holes still
687 * containing last_half_cycle - 1. If we find such a hole,
688 * then the start of that hole will be the new head. The
689 * simple case looks like
690 * x | x ... | x - 1 | x
691 * Another case that fits this picture would be
692 * x | x + 1 | x ... | x
693 * In this case the head really is somewhere at the end of the
694 * log, as one of the latest writes at the beginning was
697 * x | x + 1 | x ... | x - 1 | x
698 * This is really the combination of the above two cases, and
699 * the head has to end up at the start of the x-1 hole at the
702 * In the 256k log case, we will read from the beginning to the
703 * end of the log and search for cycle numbers equal to x-1.
704 * We don't worry about the x+1 blocks that we encounter,
705 * because we know that they cannot be the head since the log
708 head_blk = log_bbnum;
709 stop_on_cycle = last_half_cycle - 1;
712 * In this case we want to find the first block with cycle
713 * number matching last_half_cycle. We expect the log to be
715 * x + 1 ... | x ... | x
716 * The first block with cycle number x (last_half_cycle) will
717 * be where the new head belongs. First we do a binary search
718 * for the first occurrence of last_half_cycle. The binary
719 * search may not be totally accurate, so then we scan back
720 * from there looking for occurrences of last_half_cycle before
721 * us. If that backwards scan wraps around the beginning of
722 * the log, then we look for occurrences of last_half_cycle - 1
723 * at the end of the log. The cases we're looking for look
725 * v binary search stopped here
726 * x + 1 ... | x | x + 1 | x ... | x
727 * ^ but we want to locate this spot
729 * <---------> less than scan distance
730 * x + 1 ... | x ... | x - 1 | x
731 * ^ we want to locate this spot
733 stop_on_cycle = last_half_cycle;
734 if ((error = xlog_find_cycle_start(log, bp, first_blk,
735 &head_blk, last_half_cycle)))
740 * Now validate the answer. Scan back some number of maximum possible
741 * blocks and make sure each one has the expected cycle number. The
742 * maximum is determined by the total possible amount of buffering
743 * in the in-core log. The following number can be made tighter if
744 * we actually look at the block size of the filesystem.
746 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
747 if (head_blk >= num_scan_bblks) {
749 * We are guaranteed that the entire check can be performed
752 start_blk = head_blk - num_scan_bblks;
753 if ((error = xlog_find_verify_cycle(log,
754 start_blk, num_scan_bblks,
755 stop_on_cycle, &new_blk)))
759 } else { /* need to read 2 parts of log */
761 * We are going to scan backwards in the log in two parts.
762 * First we scan the physical end of the log. In this part
763 * of the log, we are looking for blocks with cycle number
764 * last_half_cycle - 1.
765 * If we find one, then we know that the log starts there, as
766 * we've found a hole that didn't get written in going around
767 * the end of the physical log. The simple case for this is
768 * x + 1 ... | x ... | x - 1 | x
769 * <---------> less than scan distance
770 * If all of the blocks at the end of the log have cycle number
771 * last_half_cycle, then we check the blocks at the start of
772 * the log looking for occurrences of last_half_cycle. If we
773 * find one, then our current estimate for the location of the
774 * first occurrence of last_half_cycle is wrong and we move
775 * back to the hole we've found. This case looks like
776 * x + 1 ... | x | x + 1 | x ...
777 * ^ binary search stopped here
778 * Another case we need to handle that only occurs in 256k
780 * x + 1 ... | x ... | x+1 | x ...
781 * ^ binary search stops here
782 * In a 256k log, the scan at the end of the log will see the
783 * x + 1 blocks. We need to skip past those since that is
784 * certainly not the head of the log. By searching for
785 * last_half_cycle-1 we accomplish that.
787 ASSERT(head_blk <= INT_MAX &&
788 (xfs_daddr_t) num_scan_bblks >= head_blk);
789 start_blk = log_bbnum - (num_scan_bblks - head_blk);
790 if ((error = xlog_find_verify_cycle(log, start_blk,
791 num_scan_bblks - (int)head_blk,
792 (stop_on_cycle - 1), &new_blk)))
800 * Scan beginning of log now. The last part of the physical
801 * log is good. This scan needs to verify that it doesn't find
802 * the last_half_cycle.
805 ASSERT(head_blk <= INT_MAX);
806 if ((error = xlog_find_verify_cycle(log,
807 start_blk, (int)head_blk,
808 stop_on_cycle, &new_blk)))
816 * Now we need to make sure head_blk is not pointing to a block in
817 * the middle of a log record.
819 num_scan_bblks = XLOG_REC_SHIFT(log);
820 if (head_blk >= num_scan_bblks) {
821 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
823 /* start ptr at last block ptr before head_blk */
824 if ((error = xlog_find_verify_log_record(log, start_blk,
825 &head_blk, 0)) == -1) {
826 error = XFS_ERROR(EIO);
832 ASSERT(head_blk <= INT_MAX);
833 if ((error = xlog_find_verify_log_record(log, start_blk,
834 &head_blk, 0)) == -1) {
835 /* We hit the beginning of the log during our search */
836 start_blk = log_bbnum - (num_scan_bblks - head_blk);
838 ASSERT(start_blk <= INT_MAX &&
839 (xfs_daddr_t) log_bbnum-start_blk >= 0);
840 ASSERT(head_blk <= INT_MAX);
841 if ((error = xlog_find_verify_log_record(log,
843 (int)head_blk)) == -1) {
844 error = XFS_ERROR(EIO);
848 if (new_blk != log_bbnum)
855 if (head_blk == log_bbnum)
856 *return_head_blk = 0;
858 *return_head_blk = head_blk;
860 * When returning here, we have a good block number. Bad block
861 * means that during a previous crash, we didn't have a clean break
862 * from cycle number N to cycle number N-1. In this case, we need
863 * to find the first block with cycle number N-1.
871 xfs_warn(log->l_mp, "failed to find log head");
876 * Find the sync block number or the tail of the log.
878 * This will be the block number of the last record to have its
879 * associated buffers synced to disk. Every log record header has
880 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
881 * to get a sync block number. The only concern is to figure out which
882 * log record header to believe.
884 * The following algorithm uses the log record header with the largest
885 * lsn. The entire log record does not need to be valid. We only care
886 * that the header is valid.
888 * We could speed up search by using current head_blk buffer, but it is not
894 xfs_daddr_t *head_blk,
895 xfs_daddr_t *tail_blk)
897 xlog_rec_header_t *rhead;
898 xlog_op_header_t *op_head;
899 xfs_caddr_t offset = NULL;
902 xfs_daddr_t umount_data_blk;
903 xfs_daddr_t after_umount_blk;
910 * Find previous log record
912 if ((error = xlog_find_head(log, head_blk)))
915 bp = xlog_get_bp(log, 1);
918 if (*head_blk == 0) { /* special case */
919 error = xlog_bread(log, 0, 1, bp, &offset);
923 if (xlog_get_cycle(offset) == 0) {
925 /* leave all other log inited values alone */
931 * Search backwards looking for log record header block
933 ASSERT(*head_blk < INT_MAX);
934 for (i = (int)(*head_blk) - 1; i >= 0; i--) {
935 error = xlog_bread(log, i, 1, bp, &offset);
939 if (*(__be32 *)offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
945 * If we haven't found the log record header block, start looking
946 * again from the end of the physical log. XXXmiken: There should be
947 * a check here to make sure we didn't search more than N blocks in
951 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
952 error = xlog_bread(log, i, 1, bp, &offset);
956 if (*(__be32 *)offset ==
957 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
964 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
966 return XFS_ERROR(EIO);
969 /* find blk_no of tail of log */
970 rhead = (xlog_rec_header_t *)offset;
971 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
974 * Reset log values according to the state of the log when we
975 * crashed. In the case where head_blk == 0, we bump curr_cycle
976 * one because the next write starts a new cycle rather than
977 * continuing the cycle of the last good log record. At this
978 * point we have guaranteed that all partial log records have been
979 * accounted for. Therefore, we know that the last good log record
980 * written was complete and ended exactly on the end boundary
981 * of the physical log.
983 log->l_prev_block = i;
984 log->l_curr_block = (int)*head_blk;
985 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
988 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
989 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
990 xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
991 BBTOB(log->l_curr_block));
992 xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
993 BBTOB(log->l_curr_block));
996 * Look for unmount record. If we find it, then we know there
997 * was a clean unmount. Since 'i' could be the last block in
998 * the physical log, we convert to a log block before comparing
1001 * Save the current tail lsn to use to pass to
1002 * xlog_clear_stale_blocks() below. We won't want to clear the
1003 * unmount record if there is one, so we pass the lsn of the
1004 * unmount record rather than the block after it.
1006 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1007 int h_size = be32_to_cpu(rhead->h_size);
1008 int h_version = be32_to_cpu(rhead->h_version);
1010 if ((h_version & XLOG_VERSION_2) &&
1011 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1012 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1013 if (h_size % XLOG_HEADER_CYCLE_SIZE)
1021 after_umount_blk = (i + hblks + (int)
1022 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
1023 tail_lsn = atomic64_read(&log->l_tail_lsn);
1024 if (*head_blk == after_umount_blk &&
1025 be32_to_cpu(rhead->h_num_logops) == 1) {
1026 umount_data_blk = (i + hblks) % log->l_logBBsize;
1027 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1031 op_head = (xlog_op_header_t *)offset;
1032 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1034 * Set tail and last sync so that newly written
1035 * log records will point recovery to after the
1036 * current unmount record.
1038 xlog_assign_atomic_lsn(&log->l_tail_lsn,
1039 log->l_curr_cycle, after_umount_blk);
1040 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1041 log->l_curr_cycle, after_umount_blk);
1042 *tail_blk = after_umount_blk;
1045 * Note that the unmount was clean. If the unmount
1046 * was not clean, we need to know this to rebuild the
1047 * superblock counters from the perag headers if we
1048 * have a filesystem using non-persistent counters.
1050 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1055 * Make sure that there are no blocks in front of the head
1056 * with the same cycle number as the head. This can happen
1057 * because we allow multiple outstanding log writes concurrently,
1058 * and the later writes might make it out before earlier ones.
1060 * We use the lsn from before modifying it so that we'll never
1061 * overwrite the unmount record after a clean unmount.
1063 * Do this only if we are going to recover the filesystem
1065 * NOTE: This used to say "if (!readonly)"
1066 * However on Linux, we can & do recover a read-only filesystem.
1067 * We only skip recovery if NORECOVERY is specified on mount,
1068 * in which case we would not be here.
1070 * But... if the -device- itself is readonly, just skip this.
1071 * We can't recover this device anyway, so it won't matter.
1073 if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1074 error = xlog_clear_stale_blocks(log, tail_lsn);
1080 xfs_warn(log->l_mp, "failed to locate log tail");
1085 * Is the log zeroed at all?
1087 * The last binary search should be changed to perform an X block read
1088 * once X becomes small enough. You can then search linearly through
1089 * the X blocks. This will cut down on the number of reads we need to do.
1091 * If the log is partially zeroed, this routine will pass back the blkno
1092 * of the first block with cycle number 0. It won't have a complete LR
1096 * 0 => the log is completely written to
1097 * -1 => use *blk_no as the first block of the log
1098 * >0 => error has occurred
1103 xfs_daddr_t *blk_no)
1107 uint first_cycle, last_cycle;
1108 xfs_daddr_t new_blk, last_blk, start_blk;
1109 xfs_daddr_t num_scan_bblks;
1110 int error, log_bbnum = log->l_logBBsize;
1114 /* check totally zeroed log */
1115 bp = xlog_get_bp(log, 1);
1118 error = xlog_bread(log, 0, 1, bp, &offset);
1122 first_cycle = xlog_get_cycle(offset);
1123 if (first_cycle == 0) { /* completely zeroed log */
1129 /* check partially zeroed log */
1130 error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1134 last_cycle = xlog_get_cycle(offset);
1135 if (last_cycle != 0) { /* log completely written to */
1138 } else if (first_cycle != 1) {
1140 * If the cycle of the last block is zero, the cycle of
1141 * the first block must be 1. If it's not, maybe we're
1142 * not looking at a log... Bail out.
1145 "Log inconsistent or not a log (last==0, first!=1)");
1146 return XFS_ERROR(EINVAL);
1149 /* we have a partially zeroed log */
1150 last_blk = log_bbnum-1;
1151 if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1155 * Validate the answer. Because there is no way to guarantee that
1156 * the entire log is made up of log records which are the same size,
1157 * we scan over the defined maximum blocks. At this point, the maximum
1158 * is not chosen to mean anything special. XXXmiken
1160 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1161 ASSERT(num_scan_bblks <= INT_MAX);
1163 if (last_blk < num_scan_bblks)
1164 num_scan_bblks = last_blk;
1165 start_blk = last_blk - num_scan_bblks;
1168 * We search for any instances of cycle number 0 that occur before
1169 * our current estimate of the head. What we're trying to detect is
1170 * 1 ... | 0 | 1 | 0...
1171 * ^ binary search ends here
1173 if ((error = xlog_find_verify_cycle(log, start_blk,
1174 (int)num_scan_bblks, 0, &new_blk)))
1180 * Potentially backup over partial log record write. We don't need
1181 * to search the end of the log because we know it is zero.
1183 if ((error = xlog_find_verify_log_record(log, start_blk,
1184 &last_blk, 0)) == -1) {
1185 error = XFS_ERROR(EIO);
1199 * These are simple subroutines used by xlog_clear_stale_blocks() below
1200 * to initialize a buffer full of empty log record headers and write
1201 * them into the log.
1212 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1214 memset(buf, 0, BBSIZE);
1215 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1216 recp->h_cycle = cpu_to_be32(cycle);
1217 recp->h_version = cpu_to_be32(
1218 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1219 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1220 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1221 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1222 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1226 xlog_write_log_records(
1237 int sectbb = log->l_sectBBsize;
1238 int end_block = start_block + blocks;
1244 * Greedily allocate a buffer big enough to handle the full
1245 * range of basic blocks to be written. If that fails, try
1246 * a smaller size. We need to be able to write at least a
1247 * log sector, or we're out of luck.
1249 bufblks = 1 << ffs(blocks);
1250 while (bufblks > log->l_logBBsize)
1252 while (!(bp = xlog_get_bp(log, bufblks))) {
1254 if (bufblks < sectbb)
1258 /* We may need to do a read at the start to fill in part of
1259 * the buffer in the starting sector not covered by the first
1262 balign = round_down(start_block, sectbb);
1263 if (balign != start_block) {
1264 error = xlog_bread_noalign(log, start_block, 1, bp);
1268 j = start_block - balign;
1271 for (i = start_block; i < end_block; i += bufblks) {
1272 int bcount, endcount;
1274 bcount = min(bufblks, end_block - start_block);
1275 endcount = bcount - j;
1277 /* We may need to do a read at the end to fill in part of
1278 * the buffer in the final sector not covered by the write.
1279 * If this is the same sector as the above read, skip it.
1281 ealign = round_down(end_block, sectbb);
1282 if (j == 0 && (start_block + endcount > ealign)) {
1283 offset = bp->b_addr + BBTOB(ealign - start_block);
1284 error = xlog_bread_offset(log, ealign, sectbb,
1291 offset = xlog_align(log, start_block, endcount, bp);
1292 for (; j < endcount; j++) {
1293 xlog_add_record(log, offset, cycle, i+j,
1294 tail_cycle, tail_block);
1297 error = xlog_bwrite(log, start_block, endcount, bp);
1300 start_block += endcount;
1310 * This routine is called to blow away any incomplete log writes out
1311 * in front of the log head. We do this so that we won't become confused
1312 * if we come up, write only a little bit more, and then crash again.
1313 * If we leave the partial log records out there, this situation could
1314 * cause us to think those partial writes are valid blocks since they
1315 * have the current cycle number. We get rid of them by overwriting them
1316 * with empty log records with the old cycle number rather than the
1319 * The tail lsn is passed in rather than taken from
1320 * the log so that we will not write over the unmount record after a
1321 * clean unmount in a 512 block log. Doing so would leave the log without
1322 * any valid log records in it until a new one was written. If we crashed
1323 * during that time we would not be able to recover.
1326 xlog_clear_stale_blocks(
1330 int tail_cycle, head_cycle;
1331 int tail_block, head_block;
1332 int tail_distance, max_distance;
1336 tail_cycle = CYCLE_LSN(tail_lsn);
1337 tail_block = BLOCK_LSN(tail_lsn);
1338 head_cycle = log->l_curr_cycle;
1339 head_block = log->l_curr_block;
1342 * Figure out the distance between the new head of the log
1343 * and the tail. We want to write over any blocks beyond the
1344 * head that we may have written just before the crash, but
1345 * we don't want to overwrite the tail of the log.
1347 if (head_cycle == tail_cycle) {
1349 * The tail is behind the head in the physical log,
1350 * so the distance from the head to the tail is the
1351 * distance from the head to the end of the log plus
1352 * the distance from the beginning of the log to the
1355 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1356 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1357 XFS_ERRLEVEL_LOW, log->l_mp);
1358 return XFS_ERROR(EFSCORRUPTED);
1360 tail_distance = tail_block + (log->l_logBBsize - head_block);
1363 * The head is behind the tail in the physical log,
1364 * so the distance from the head to the tail is just
1365 * the tail block minus the head block.
1367 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1368 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1369 XFS_ERRLEVEL_LOW, log->l_mp);
1370 return XFS_ERROR(EFSCORRUPTED);
1372 tail_distance = tail_block - head_block;
1376 * If the head is right up against the tail, we can't clear
1379 if (tail_distance <= 0) {
1380 ASSERT(tail_distance == 0);
1384 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1386 * Take the smaller of the maximum amount of outstanding I/O
1387 * we could have and the distance to the tail to clear out.
1388 * We take the smaller so that we don't overwrite the tail and
1389 * we don't waste all day writing from the head to the tail
1392 max_distance = MIN(max_distance, tail_distance);
1394 if ((head_block + max_distance) <= log->l_logBBsize) {
1396 * We can stomp all the blocks we need to without
1397 * wrapping around the end of the log. Just do it
1398 * in a single write. Use the cycle number of the
1399 * current cycle minus one so that the log will look like:
1402 error = xlog_write_log_records(log, (head_cycle - 1),
1403 head_block, max_distance, tail_cycle,
1409 * We need to wrap around the end of the physical log in
1410 * order to clear all the blocks. Do it in two separate
1411 * I/Os. The first write should be from the head to the
1412 * end of the physical log, and it should use the current
1413 * cycle number minus one just like above.
1415 distance = log->l_logBBsize - head_block;
1416 error = xlog_write_log_records(log, (head_cycle - 1),
1417 head_block, distance, tail_cycle,
1424 * Now write the blocks at the start of the physical log.
1425 * This writes the remainder of the blocks we want to clear.
1426 * It uses the current cycle number since we're now on the
1427 * same cycle as the head so that we get:
1428 * n ... n ... | n - 1 ...
1429 * ^^^^^ blocks we're writing
1431 distance = max_distance - (log->l_logBBsize - head_block);
1432 error = xlog_write_log_records(log, head_cycle, 0, distance,
1433 tail_cycle, tail_block);
1441 /******************************************************************************
1443 * Log recover routines
1445 ******************************************************************************
1448 STATIC xlog_recover_t *
1449 xlog_recover_find_tid(
1450 struct hlist_head *head,
1453 xlog_recover_t *trans;
1455 hlist_for_each_entry(trans, head, r_list) {
1456 if (trans->r_log_tid == tid)
1463 xlog_recover_new_tid(
1464 struct hlist_head *head,
1468 xlog_recover_t *trans;
1470 trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1471 trans->r_log_tid = tid;
1473 INIT_LIST_HEAD(&trans->r_itemq);
1475 INIT_HLIST_NODE(&trans->r_list);
1476 hlist_add_head(&trans->r_list, head);
1480 xlog_recover_add_item(
1481 struct list_head *head)
1483 xlog_recover_item_t *item;
1485 item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1486 INIT_LIST_HEAD(&item->ri_list);
1487 list_add_tail(&item->ri_list, head);
1491 xlog_recover_add_to_cont_trans(
1493 struct xlog_recover *trans,
1497 xlog_recover_item_t *item;
1498 xfs_caddr_t ptr, old_ptr;
1501 if (list_empty(&trans->r_itemq)) {
1502 /* finish copying rest of trans header */
1503 xlog_recover_add_item(&trans->r_itemq);
1504 ptr = (xfs_caddr_t) &trans->r_theader +
1505 sizeof(xfs_trans_header_t) - len;
1506 memcpy(ptr, dp, len); /* d, s, l */
1509 /* take the tail entry */
1510 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1512 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1513 old_len = item->ri_buf[item->ri_cnt-1].i_len;
1515 ptr = kmem_realloc(old_ptr, len+old_len, old_len, KM_SLEEP);
1516 memcpy(&ptr[old_len], dp, len); /* d, s, l */
1517 item->ri_buf[item->ri_cnt-1].i_len += len;
1518 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1519 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
1524 * The next region to add is the start of a new region. It could be
1525 * a whole region or it could be the first part of a new region. Because
1526 * of this, the assumption here is that the type and size fields of all
1527 * format structures fit into the first 32 bits of the structure.
1529 * This works because all regions must be 32 bit aligned. Therefore, we
1530 * either have both fields or we have neither field. In the case we have
1531 * neither field, the data part of the region is zero length. We only have
1532 * a log_op_header and can throw away the header since a new one will appear
1533 * later. If we have at least 4 bytes, then we can determine how many regions
1534 * will appear in the current log item.
1537 xlog_recover_add_to_trans(
1539 struct xlog_recover *trans,
1543 xfs_inode_log_format_t *in_f; /* any will do */
1544 xlog_recover_item_t *item;
1549 if (list_empty(&trans->r_itemq)) {
1550 /* we need to catch log corruptions here */
1551 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1552 xfs_warn(log->l_mp, "%s: bad header magic number",
1555 return XFS_ERROR(EIO);
1557 if (len == sizeof(xfs_trans_header_t))
1558 xlog_recover_add_item(&trans->r_itemq);
1559 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1563 ptr = kmem_alloc(len, KM_SLEEP);
1564 memcpy(ptr, dp, len);
1565 in_f = (xfs_inode_log_format_t *)ptr;
1567 /* take the tail entry */
1568 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1569 if (item->ri_total != 0 &&
1570 item->ri_total == item->ri_cnt) {
1571 /* tail item is in use, get a new one */
1572 xlog_recover_add_item(&trans->r_itemq);
1573 item = list_entry(trans->r_itemq.prev,
1574 xlog_recover_item_t, ri_list);
1577 if (item->ri_total == 0) { /* first region to be added */
1578 if (in_f->ilf_size == 0 ||
1579 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1581 "bad number of regions (%d) in inode log format",
1584 return XFS_ERROR(EIO);
1587 item->ri_total = in_f->ilf_size;
1589 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1592 ASSERT(item->ri_total > item->ri_cnt);
1593 /* Description region is ri_buf[0] */
1594 item->ri_buf[item->ri_cnt].i_addr = ptr;
1595 item->ri_buf[item->ri_cnt].i_len = len;
1597 trace_xfs_log_recover_item_add(log, trans, item, 0);
1602 * Sort the log items in the transaction.
1604 * The ordering constraints are defined by the inode allocation and unlink
1605 * behaviour. The rules are:
1607 * 1. Every item is only logged once in a given transaction. Hence it
1608 * represents the last logged state of the item. Hence ordering is
1609 * dependent on the order in which operations need to be performed so
1610 * required initial conditions are always met.
1612 * 2. Cancelled buffers are recorded in pass 1 in a separate table and
1613 * there's nothing to replay from them so we can simply cull them
1614 * from the transaction. However, we can't do that until after we've
1615 * replayed all the other items because they may be dependent on the
1616 * cancelled buffer and replaying the cancelled buffer can remove it
1617 * form the cancelled buffer table. Hence they have tobe done last.
1619 * 3. Inode allocation buffers must be replayed before inode items that
1620 * read the buffer and replay changes into it.
1622 * 4. Inode unlink buffers must be replayed after inode items are replayed.
1623 * This ensures that inodes are completely flushed to the inode buffer
1624 * in a "free" state before we remove the unlinked inode list pointer.
1626 * Hence the ordering needs to be inode allocation buffers first, inode items
1627 * second, inode unlink buffers third and cancelled buffers last.
1629 * But there's a problem with that - we can't tell an inode allocation buffer
1630 * apart from a regular buffer, so we can't separate them. We can, however,
1631 * tell an inode unlink buffer from the others, and so we can separate them out
1632 * from all the other buffers and move them to last.
1634 * Hence, 4 lists, in order from head to tail:
1635 * - buffer_list for all buffers except cancelled/inode unlink buffers
1636 * - item_list for all non-buffer items
1637 * - inode_buffer_list for inode unlink buffers
1638 * - cancel_list for the cancelled buffers
1641 xlog_recover_reorder_trans(
1643 struct xlog_recover *trans,
1646 xlog_recover_item_t *item, *n;
1647 LIST_HEAD(sort_list);
1648 LIST_HEAD(cancel_list);
1649 LIST_HEAD(buffer_list);
1650 LIST_HEAD(inode_buffer_list);
1651 LIST_HEAD(inode_list);
1653 list_splice_init(&trans->r_itemq, &sort_list);
1654 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1655 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1657 switch (ITEM_TYPE(item)) {
1659 if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1660 trace_xfs_log_recover_item_reorder_head(log,
1662 list_move(&item->ri_list, &cancel_list);
1665 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1666 list_move(&item->ri_list, &inode_buffer_list);
1669 list_move_tail(&item->ri_list, &buffer_list);
1673 case XFS_LI_QUOTAOFF:
1676 trace_xfs_log_recover_item_reorder_tail(log,
1678 list_move_tail(&item->ri_list, &inode_list);
1682 "%s: unrecognized type of log operation",
1685 return XFS_ERROR(EIO);
1688 ASSERT(list_empty(&sort_list));
1689 if (!list_empty(&buffer_list))
1690 list_splice(&buffer_list, &trans->r_itemq);
1691 if (!list_empty(&inode_list))
1692 list_splice_tail(&inode_list, &trans->r_itemq);
1693 if (!list_empty(&inode_buffer_list))
1694 list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1695 if (!list_empty(&cancel_list))
1696 list_splice_tail(&cancel_list, &trans->r_itemq);
1701 * Build up the table of buf cancel records so that we don't replay
1702 * cancelled data in the second pass. For buffer records that are
1703 * not cancel records, there is nothing to do here so we just return.
1705 * If we get a cancel record which is already in the table, this indicates
1706 * that the buffer was cancelled multiple times. In order to ensure
1707 * that during pass 2 we keep the record in the table until we reach its
1708 * last occurrence in the log, we keep a reference count in the cancel
1709 * record in the table to tell us how many times we expect to see this
1710 * record during the second pass.
1713 xlog_recover_buffer_pass1(
1715 struct xlog_recover_item *item)
1717 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1718 struct list_head *bucket;
1719 struct xfs_buf_cancel *bcp;
1722 * If this isn't a cancel buffer item, then just return.
1724 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1725 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1730 * Insert an xfs_buf_cancel record into the hash table of them.
1731 * If there is already an identical record, bump its reference count.
1733 bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1734 list_for_each_entry(bcp, bucket, bc_list) {
1735 if (bcp->bc_blkno == buf_f->blf_blkno &&
1736 bcp->bc_len == buf_f->blf_len) {
1738 trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1743 bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
1744 bcp->bc_blkno = buf_f->blf_blkno;
1745 bcp->bc_len = buf_f->blf_len;
1746 bcp->bc_refcount = 1;
1747 list_add_tail(&bcp->bc_list, bucket);
1749 trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1754 * Check to see whether the buffer being recovered has a corresponding
1755 * entry in the buffer cancel record table. If it does then return 1
1756 * so that it will be cancelled, otherwise return 0. If the buffer is
1757 * actually a buffer cancel item (XFS_BLF_CANCEL is set), then decrement
1758 * the refcount on the entry in the table and remove it from the table
1759 * if this is the last reference.
1761 * We remove the cancel record from the table when we encounter its
1762 * last occurrence in the log so that if the same buffer is re-used
1763 * again after its last cancellation we actually replay the changes
1764 * made at that point.
1767 xlog_check_buffer_cancelled(
1773 struct list_head *bucket;
1774 struct xfs_buf_cancel *bcp;
1776 if (log->l_buf_cancel_table == NULL) {
1778 * There is nothing in the table built in pass one,
1779 * so this buffer must not be cancelled.
1781 ASSERT(!(flags & XFS_BLF_CANCEL));
1786 * Search for an entry in the cancel table that matches our buffer.
1788 bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1789 list_for_each_entry(bcp, bucket, bc_list) {
1790 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1795 * We didn't find a corresponding entry in the table, so return 0 so
1796 * that the buffer is NOT cancelled.
1798 ASSERT(!(flags & XFS_BLF_CANCEL));
1803 * We've go a match, so return 1 so that the recovery of this buffer
1804 * is cancelled. If this buffer is actually a buffer cancel log
1805 * item, then decrement the refcount on the one in the table and
1806 * remove it if this is the last reference.
1808 if (flags & XFS_BLF_CANCEL) {
1809 if (--bcp->bc_refcount == 0) {
1810 list_del(&bcp->bc_list);
1818 * Perform recovery for a buffer full of inodes. In these buffers, the only
1819 * data which should be recovered is that which corresponds to the
1820 * di_next_unlinked pointers in the on disk inode structures. The rest of the
1821 * data for the inodes is always logged through the inodes themselves rather
1822 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1824 * The only time when buffers full of inodes are fully recovered is when the
1825 * buffer is full of newly allocated inodes. In this case the buffer will
1826 * not be marked as an inode buffer and so will be sent to
1827 * xlog_recover_do_reg_buffer() below during recovery.
1830 xlog_recover_do_inode_buffer(
1831 struct xfs_mount *mp,
1832 xlog_recover_item_t *item,
1834 xfs_buf_log_format_t *buf_f)
1840 int reg_buf_offset = 0;
1841 int reg_buf_bytes = 0;
1842 int next_unlinked_offset;
1844 xfs_agino_t *logged_nextp;
1845 xfs_agino_t *buffer_nextp;
1847 trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
1848 bp->b_ops = &xfs_inode_buf_ops;
1850 inodes_per_buf = BBTOB(bp->b_io_length) >> mp->m_sb.sb_inodelog;
1851 for (i = 0; i < inodes_per_buf; i++) {
1852 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1853 offsetof(xfs_dinode_t, di_next_unlinked);
1855 while (next_unlinked_offset >=
1856 (reg_buf_offset + reg_buf_bytes)) {
1858 * The next di_next_unlinked field is beyond
1859 * the current logged region. Find the next
1860 * logged region that contains or is beyond
1861 * the current di_next_unlinked field.
1864 bit = xfs_next_bit(buf_f->blf_data_map,
1865 buf_f->blf_map_size, bit);
1868 * If there are no more logged regions in the
1869 * buffer, then we're done.
1874 nbits = xfs_contig_bits(buf_f->blf_data_map,
1875 buf_f->blf_map_size, bit);
1877 reg_buf_offset = bit << XFS_BLF_SHIFT;
1878 reg_buf_bytes = nbits << XFS_BLF_SHIFT;
1883 * If the current logged region starts after the current
1884 * di_next_unlinked field, then move on to the next
1885 * di_next_unlinked field.
1887 if (next_unlinked_offset < reg_buf_offset)
1890 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1891 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
1892 ASSERT((reg_buf_offset + reg_buf_bytes) <=
1893 BBTOB(bp->b_io_length));
1896 * The current logged region contains a copy of the
1897 * current di_next_unlinked field. Extract its value
1898 * and copy it to the buffer copy.
1900 logged_nextp = item->ri_buf[item_index].i_addr +
1901 next_unlinked_offset - reg_buf_offset;
1902 if (unlikely(*logged_nextp == 0)) {
1904 "Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1905 "Trying to replay bad (0) inode di_next_unlinked field.",
1907 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1908 XFS_ERRLEVEL_LOW, mp);
1909 return XFS_ERROR(EFSCORRUPTED);
1912 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1913 next_unlinked_offset);
1914 *buffer_nextp = *logged_nextp;
1921 * Validate the recovered buffer is of the correct type and attach the
1922 * appropriate buffer operations to them for writeback. Magic numbers are in a
1924 * the first 16 bits of the buffer (inode buffer, dquot buffer),
1925 * the first 32 bits of the buffer (most blocks),
1926 * inside a struct xfs_da_blkinfo at the start of the buffer.
1929 xlog_recovery_validate_buf_type(
1930 struct xfs_mount *mp,
1932 xfs_buf_log_format_t *buf_f)
1934 struct xfs_da_blkinfo *info = bp->b_addr;
1939 magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
1940 magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
1941 magicda = be16_to_cpu(info->magic);
1942 switch (xfs_blft_from_flags(buf_f)) {
1943 case XFS_BLFT_BTREE_BUF:
1945 case XFS_ABTB_CRC_MAGIC:
1946 case XFS_ABTC_CRC_MAGIC:
1947 case XFS_ABTB_MAGIC:
1948 case XFS_ABTC_MAGIC:
1949 bp->b_ops = &xfs_allocbt_buf_ops;
1951 case XFS_IBT_CRC_MAGIC:
1953 bp->b_ops = &xfs_inobt_buf_ops;
1955 case XFS_BMAP_CRC_MAGIC:
1956 case XFS_BMAP_MAGIC:
1957 bp->b_ops = &xfs_bmbt_buf_ops;
1960 xfs_warn(mp, "Bad btree block magic!");
1965 case XFS_BLFT_AGF_BUF:
1966 if (magic32 != XFS_AGF_MAGIC) {
1967 xfs_warn(mp, "Bad AGF block magic!");
1971 bp->b_ops = &xfs_agf_buf_ops;
1973 case XFS_BLFT_AGFL_BUF:
1974 if (!xfs_sb_version_hascrc(&mp->m_sb))
1976 if (magic32 != XFS_AGFL_MAGIC) {
1977 xfs_warn(mp, "Bad AGFL block magic!");
1981 bp->b_ops = &xfs_agfl_buf_ops;
1983 case XFS_BLFT_AGI_BUF:
1984 if (magic32 != XFS_AGI_MAGIC) {
1985 xfs_warn(mp, "Bad AGI block magic!");
1989 bp->b_ops = &xfs_agi_buf_ops;
1991 case XFS_BLFT_UDQUOT_BUF:
1992 case XFS_BLFT_PDQUOT_BUF:
1993 case XFS_BLFT_GDQUOT_BUF:
1994 #ifdef CONFIG_XFS_QUOTA
1995 if (magic16 != XFS_DQUOT_MAGIC) {
1996 xfs_warn(mp, "Bad DQUOT block magic!");
2000 bp->b_ops = &xfs_dquot_buf_ops;
2003 "Trying to recover dquots without QUOTA support built in!");
2007 case XFS_BLFT_DINO_BUF:
2009 * we get here with inode allocation buffers, not buffers that
2010 * track unlinked list changes.
2012 if (magic16 != XFS_DINODE_MAGIC) {
2013 xfs_warn(mp, "Bad INODE block magic!");
2017 bp->b_ops = &xfs_inode_buf_ops;
2019 case XFS_BLFT_SYMLINK_BUF:
2020 if (magic32 != XFS_SYMLINK_MAGIC) {
2021 xfs_warn(mp, "Bad symlink block magic!");
2025 bp->b_ops = &xfs_symlink_buf_ops;
2027 case XFS_BLFT_DIR_BLOCK_BUF:
2028 if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2029 magic32 != XFS_DIR3_BLOCK_MAGIC) {
2030 xfs_warn(mp, "Bad dir block magic!");
2034 bp->b_ops = &xfs_dir3_block_buf_ops;
2036 case XFS_BLFT_DIR_DATA_BUF:
2037 if (magic32 != XFS_DIR2_DATA_MAGIC &&
2038 magic32 != XFS_DIR3_DATA_MAGIC) {
2039 xfs_warn(mp, "Bad dir data magic!");
2043 bp->b_ops = &xfs_dir3_data_buf_ops;
2045 case XFS_BLFT_DIR_FREE_BUF:
2046 if (magic32 != XFS_DIR2_FREE_MAGIC &&
2047 magic32 != XFS_DIR3_FREE_MAGIC) {
2048 xfs_warn(mp, "Bad dir3 free magic!");
2052 bp->b_ops = &xfs_dir3_free_buf_ops;
2054 case XFS_BLFT_DIR_LEAF1_BUF:
2055 if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2056 magicda != XFS_DIR3_LEAF1_MAGIC) {
2057 xfs_warn(mp, "Bad dir leaf1 magic!");
2061 bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2063 case XFS_BLFT_DIR_LEAFN_BUF:
2064 if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2065 magicda != XFS_DIR3_LEAFN_MAGIC) {
2066 xfs_warn(mp, "Bad dir leafn magic!");
2070 bp->b_ops = &xfs_dir3_leafn_buf_ops;
2072 case XFS_BLFT_DA_NODE_BUF:
2073 if (magicda != XFS_DA_NODE_MAGIC &&
2074 magicda != XFS_DA3_NODE_MAGIC) {
2075 xfs_warn(mp, "Bad da node magic!");
2079 bp->b_ops = &xfs_da3_node_buf_ops;
2081 case XFS_BLFT_ATTR_LEAF_BUF:
2082 if (magicda != XFS_ATTR_LEAF_MAGIC &&
2083 magicda != XFS_ATTR3_LEAF_MAGIC) {
2084 xfs_warn(mp, "Bad attr leaf magic!");
2088 bp->b_ops = &xfs_attr3_leaf_buf_ops;
2090 case XFS_BLFT_ATTR_RMT_BUF:
2091 if (!xfs_sb_version_hascrc(&mp->m_sb))
2093 if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2094 xfs_warn(mp, "Bad attr remote magic!");
2098 bp->b_ops = &xfs_attr3_rmt_buf_ops;
2100 case XFS_BLFT_SB_BUF:
2101 if (magic32 != XFS_SB_MAGIC) {
2102 xfs_warn(mp, "Bad SB block magic!");
2106 bp->b_ops = &xfs_sb_buf_ops;
2109 xfs_warn(mp, "Unknown buffer type %d!",
2110 xfs_blft_from_flags(buf_f));
2116 * Perform a 'normal' buffer recovery. Each logged region of the
2117 * buffer should be copied over the corresponding region in the
2118 * given buffer. The bitmap in the buf log format structure indicates
2119 * where to place the logged data.
2122 xlog_recover_do_reg_buffer(
2123 struct xfs_mount *mp,
2124 xlog_recover_item_t *item,
2126 xfs_buf_log_format_t *buf_f)
2133 trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2136 i = 1; /* 0 is the buf format structure */
2138 bit = xfs_next_bit(buf_f->blf_data_map,
2139 buf_f->blf_map_size, bit);
2142 nbits = xfs_contig_bits(buf_f->blf_data_map,
2143 buf_f->blf_map_size, bit);
2145 ASSERT(item->ri_buf[i].i_addr != NULL);
2146 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2147 ASSERT(BBTOB(bp->b_io_length) >=
2148 ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2151 * The dirty regions logged in the buffer, even though
2152 * contiguous, may span multiple chunks. This is because the
2153 * dirty region may span a physical page boundary in a buffer
2154 * and hence be split into two separate vectors for writing into
2155 * the log. Hence we need to trim nbits back to the length of
2156 * the current region being copied out of the log.
2158 if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2159 nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2162 * Do a sanity check if this is a dquot buffer. Just checking
2163 * the first dquot in the buffer should do. XXXThis is
2164 * probably a good thing to do for other buf types also.
2167 if (buf_f->blf_flags &
2168 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2169 if (item->ri_buf[i].i_addr == NULL) {
2171 "XFS: NULL dquot in %s.", __func__);
2174 if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2176 "XFS: dquot too small (%d) in %s.",
2177 item->ri_buf[i].i_len, __func__);
2180 error = xfs_qm_dqcheck(mp, item->ri_buf[i].i_addr,
2181 -1, 0, XFS_QMOPT_DOWARN,
2182 "dquot_buf_recover");
2187 memcpy(xfs_buf_offset(bp,
2188 (uint)bit << XFS_BLF_SHIFT), /* dest */
2189 item->ri_buf[i].i_addr, /* source */
2190 nbits<<XFS_BLF_SHIFT); /* length */
2196 /* Shouldn't be any more regions */
2197 ASSERT(i == item->ri_total);
2199 xlog_recovery_validate_buf_type(mp, bp, buf_f);
2203 * Do some primitive error checking on ondisk dquot data structures.
2207 struct xfs_mount *mp,
2208 xfs_disk_dquot_t *ddq,
2210 uint type, /* used only when IO_dorepair is true */
2214 xfs_dqblk_t *d = (xfs_dqblk_t *)ddq;
2218 * We can encounter an uninitialized dquot buffer for 2 reasons:
2219 * 1. If we crash while deleting the quotainode(s), and those blks got
2220 * used for user data. This is because we take the path of regular
2221 * file deletion; however, the size field of quotainodes is never
2222 * updated, so all the tricks that we play in itruncate_finish
2223 * don't quite matter.
2225 * 2. We don't play the quota buffers when there's a quotaoff logitem.
2226 * But the allocation will be replayed so we'll end up with an
2227 * uninitialized quota block.
2229 * This is all fine; things are still consistent, and we haven't lost
2230 * any quota information. Just don't complain about bad dquot blks.
2232 if (ddq->d_magic != cpu_to_be16(XFS_DQUOT_MAGIC)) {
2233 if (flags & XFS_QMOPT_DOWARN)
2235 "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
2236 str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
2239 if (ddq->d_version != XFS_DQUOT_VERSION) {
2240 if (flags & XFS_QMOPT_DOWARN)
2242 "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
2243 str, id, ddq->d_version, XFS_DQUOT_VERSION);
2247 if (ddq->d_flags != XFS_DQ_USER &&
2248 ddq->d_flags != XFS_DQ_PROJ &&
2249 ddq->d_flags != XFS_DQ_GROUP) {
2250 if (flags & XFS_QMOPT_DOWARN)
2252 "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
2253 str, id, ddq->d_flags);
2257 if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
2258 if (flags & XFS_QMOPT_DOWARN)
2260 "%s : ondisk-dquot 0x%p, ID mismatch: "
2261 "0x%x expected, found id 0x%x",
2262 str, ddq, id, be32_to_cpu(ddq->d_id));
2266 if (!errs && ddq->d_id) {
2267 if (ddq->d_blk_softlimit &&
2268 be64_to_cpu(ddq->d_bcount) >
2269 be64_to_cpu(ddq->d_blk_softlimit)) {
2270 if (!ddq->d_btimer) {
2271 if (flags & XFS_QMOPT_DOWARN)
2273 "%s : Dquot ID 0x%x (0x%p) BLK TIMER NOT STARTED",
2274 str, (int)be32_to_cpu(ddq->d_id), ddq);
2278 if (ddq->d_ino_softlimit &&
2279 be64_to_cpu(ddq->d_icount) >
2280 be64_to_cpu(ddq->d_ino_softlimit)) {
2281 if (!ddq->d_itimer) {
2282 if (flags & XFS_QMOPT_DOWARN)
2284 "%s : Dquot ID 0x%x (0x%p) INODE TIMER NOT STARTED",
2285 str, (int)be32_to_cpu(ddq->d_id), ddq);
2289 if (ddq->d_rtb_softlimit &&
2290 be64_to_cpu(ddq->d_rtbcount) >
2291 be64_to_cpu(ddq->d_rtb_softlimit)) {
2292 if (!ddq->d_rtbtimer) {
2293 if (flags & XFS_QMOPT_DOWARN)
2295 "%s : Dquot ID 0x%x (0x%p) RTBLK TIMER NOT STARTED",
2296 str, (int)be32_to_cpu(ddq->d_id), ddq);
2302 if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2305 if (flags & XFS_QMOPT_DOWARN)
2306 xfs_notice(mp, "Re-initializing dquot ID 0x%x", id);
2309 * Typically, a repair is only requested by quotacheck.
2312 ASSERT(flags & XFS_QMOPT_DQREPAIR);
2313 memset(d, 0, sizeof(xfs_dqblk_t));
2315 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2316 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2317 d->dd_diskdq.d_flags = type;
2318 d->dd_diskdq.d_id = cpu_to_be32(id);
2320 if (xfs_sb_version_hascrc(&mp->m_sb)) {
2321 uuid_copy(&d->dd_uuid, &mp->m_sb.sb_uuid);
2322 xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
2330 * Perform a dquot buffer recovery.
2331 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2332 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2333 * Else, treat it as a regular buffer and do recovery.
2336 xlog_recover_do_dquot_buffer(
2337 struct xfs_mount *mp,
2339 struct xlog_recover_item *item,
2341 struct xfs_buf_log_format *buf_f)
2345 trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2348 * Filesystems are required to send in quota flags at mount time.
2350 if (mp->m_qflags == 0) {
2355 if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2356 type |= XFS_DQ_USER;
2357 if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2358 type |= XFS_DQ_PROJ;
2359 if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2360 type |= XFS_DQ_GROUP;
2362 * This type of quotas was turned off, so ignore this buffer
2364 if (log->l_quotaoffs_flag & type)
2367 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2371 * This routine replays a modification made to a buffer at runtime.
2372 * There are actually two types of buffer, regular and inode, which
2373 * are handled differently. Inode buffers are handled differently
2374 * in that we only recover a specific set of data from them, namely
2375 * the inode di_next_unlinked fields. This is because all other inode
2376 * data is actually logged via inode records and any data we replay
2377 * here which overlaps that may be stale.
2379 * When meta-data buffers are freed at run time we log a buffer item
2380 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2381 * of the buffer in the log should not be replayed at recovery time.
2382 * This is so that if the blocks covered by the buffer are reused for
2383 * file data before we crash we don't end up replaying old, freed
2384 * meta-data into a user's file.
2386 * To handle the cancellation of buffer log items, we make two passes
2387 * over the log during recovery. During the first we build a table of
2388 * those buffers which have been cancelled, and during the second we
2389 * only replay those buffers which do not have corresponding cancel
2390 * records in the table. See xlog_recover_do_buffer_pass[1,2] above
2391 * for more details on the implementation of the table of cancel records.
2394 xlog_recover_buffer_pass2(
2396 struct list_head *buffer_list,
2397 struct xlog_recover_item *item)
2399 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
2400 xfs_mount_t *mp = log->l_mp;
2406 * In this pass we only want to recover all the buffers which have
2407 * not been cancelled and are not cancellation buffers themselves.
2409 if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2410 buf_f->blf_len, buf_f->blf_flags)) {
2411 trace_xfs_log_recover_buf_cancel(log, buf_f);
2415 trace_xfs_log_recover_buf_recover(log, buf_f);
2418 if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2419 buf_flags |= XBF_UNMAPPED;
2421 bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2424 return XFS_ERROR(ENOMEM);
2425 error = bp->b_error;
2427 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2432 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2433 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2434 } else if (buf_f->blf_flags &
2435 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2436 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2438 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2441 return XFS_ERROR(error);
2444 * Perform delayed write on the buffer. Asynchronous writes will be
2445 * slower when taking into account all the buffers to be flushed.
2447 * Also make sure that only inode buffers with good sizes stay in
2448 * the buffer cache. The kernel moves inodes in buffers of 1 block
2449 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger. The inode
2450 * buffers in the log can be a different size if the log was generated
2451 * by an older kernel using unclustered inode buffers or a newer kernel
2452 * running with a different inode cluster size. Regardless, if the
2453 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2454 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2455 * the buffer out of the buffer cache so that the buffer won't
2456 * overlap with future reads of those inodes.
2458 if (XFS_DINODE_MAGIC ==
2459 be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2460 (BBTOB(bp->b_io_length) != MAX(log->l_mp->m_sb.sb_blocksize,
2461 (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2463 error = xfs_bwrite(bp);
2465 ASSERT(bp->b_target->bt_mount == mp);
2466 bp->b_iodone = xlog_recover_iodone;
2467 xfs_buf_delwri_queue(bp, buffer_list);
2475 xlog_recover_inode_pass2(
2477 struct list_head *buffer_list,
2478 struct xlog_recover_item *item)
2480 xfs_inode_log_format_t *in_f;
2481 xfs_mount_t *mp = log->l_mp;
2490 xfs_icdinode_t *dicp;
2494 if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2495 in_f = item->ri_buf[0].i_addr;
2497 in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2499 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2505 * Inode buffers can be freed, look out for it,
2506 * and do not replay the inode.
2508 if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2509 in_f->ilf_len, 0)) {
2511 trace_xfs_log_recover_inode_cancel(log, in_f);
2514 trace_xfs_log_recover_inode_recover(log, in_f);
2516 bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
2517 &xfs_inode_buf_ops);
2522 error = bp->b_error;
2524 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
2528 ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2529 dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2532 * Make sure the place we're flushing out to really looks
2535 if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
2538 "%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2539 __func__, dip, bp, in_f->ilf_ino);
2540 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2541 XFS_ERRLEVEL_LOW, mp);
2542 error = EFSCORRUPTED;
2545 dicp = item->ri_buf[1].i_addr;
2546 if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2549 "%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2550 __func__, item, in_f->ilf_ino);
2551 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2552 XFS_ERRLEVEL_LOW, mp);
2553 error = EFSCORRUPTED;
2557 /* Skip replay when the on disk inode is newer than the log one */
2558 if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2560 * Deal with the wrap case, DI_MAX_FLUSH is less
2561 * than smaller numbers
2563 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2564 dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2568 trace_xfs_log_recover_inode_skip(log, in_f);
2573 /* Take the opportunity to reset the flush iteration count */
2574 dicp->di_flushiter = 0;
2576 if (unlikely(S_ISREG(dicp->di_mode))) {
2577 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2578 (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2579 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2580 XFS_ERRLEVEL_LOW, mp, dicp);
2583 "%s: Bad regular inode log record, rec ptr 0x%p, "
2584 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2585 __func__, item, dip, bp, in_f->ilf_ino);
2586 error = EFSCORRUPTED;
2589 } else if (unlikely(S_ISDIR(dicp->di_mode))) {
2590 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2591 (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2592 (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2593 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2594 XFS_ERRLEVEL_LOW, mp, dicp);
2597 "%s: Bad dir inode log record, rec ptr 0x%p, "
2598 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2599 __func__, item, dip, bp, in_f->ilf_ino);
2600 error = EFSCORRUPTED;
2604 if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2605 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2606 XFS_ERRLEVEL_LOW, mp, dicp);
2609 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2610 "dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2611 __func__, item, dip, bp, in_f->ilf_ino,
2612 dicp->di_nextents + dicp->di_anextents,
2614 error = EFSCORRUPTED;
2617 if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2618 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2619 XFS_ERRLEVEL_LOW, mp, dicp);
2622 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2623 "dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
2624 item, dip, bp, in_f->ilf_ino, dicp->di_forkoff);
2625 error = EFSCORRUPTED;
2628 isize = xfs_icdinode_size(dicp->di_version);
2629 if (unlikely(item->ri_buf[1].i_len > isize)) {
2630 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2631 XFS_ERRLEVEL_LOW, mp, dicp);
2634 "%s: Bad inode log record length %d, rec ptr 0x%p",
2635 __func__, item->ri_buf[1].i_len, item);
2636 error = EFSCORRUPTED;
2640 /* The core is in in-core format */
2641 xfs_dinode_to_disk(dip, dicp);
2643 /* the rest is in on-disk format */
2644 if (item->ri_buf[1].i_len > isize) {
2645 memcpy((char *)dip + isize,
2646 item->ri_buf[1].i_addr + isize,
2647 item->ri_buf[1].i_len - isize);
2650 fields = in_f->ilf_fields;
2651 switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2653 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2656 memcpy(XFS_DFORK_DPTR(dip),
2657 &in_f->ilf_u.ilfu_uuid,
2662 if (in_f->ilf_size == 2)
2663 goto write_inode_buffer;
2664 len = item->ri_buf[2].i_len;
2665 src = item->ri_buf[2].i_addr;
2666 ASSERT(in_f->ilf_size <= 4);
2667 ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2668 ASSERT(!(fields & XFS_ILOG_DFORK) ||
2669 (len == in_f->ilf_dsize));
2671 switch (fields & XFS_ILOG_DFORK) {
2672 case XFS_ILOG_DDATA:
2674 memcpy(XFS_DFORK_DPTR(dip), src, len);
2677 case XFS_ILOG_DBROOT:
2678 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2679 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2680 XFS_DFORK_DSIZE(dip, mp));
2685 * There are no data fork flags set.
2687 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2692 * If we logged any attribute data, recover it. There may or
2693 * may not have been any other non-core data logged in this
2696 if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2697 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2702 len = item->ri_buf[attr_index].i_len;
2703 src = item->ri_buf[attr_index].i_addr;
2704 ASSERT(len == in_f->ilf_asize);
2706 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2707 case XFS_ILOG_ADATA:
2709 dest = XFS_DFORK_APTR(dip);
2710 ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2711 memcpy(dest, src, len);
2714 case XFS_ILOG_ABROOT:
2715 dest = XFS_DFORK_APTR(dip);
2716 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2717 len, (xfs_bmdr_block_t*)dest,
2718 XFS_DFORK_ASIZE(dip, mp));
2722 xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
2731 /* re-generate the checksum. */
2732 xfs_dinode_calc_crc(log->l_mp, dip);
2734 ASSERT(bp->b_target->bt_mount == mp);
2735 bp->b_iodone = xlog_recover_iodone;
2736 xfs_buf_delwri_queue(bp, buffer_list);
2741 return XFS_ERROR(error);
2745 * Recover QUOTAOFF records. We simply make a note of it in the xlog
2746 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2750 xlog_recover_quotaoff_pass1(
2752 struct xlog_recover_item *item)
2754 xfs_qoff_logformat_t *qoff_f = item->ri_buf[0].i_addr;
2758 * The logitem format's flag tells us if this was user quotaoff,
2759 * group/project quotaoff or both.
2761 if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2762 log->l_quotaoffs_flag |= XFS_DQ_USER;
2763 if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2764 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2765 if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2766 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2772 * Recover a dquot record
2775 xlog_recover_dquot_pass2(
2777 struct list_head *buffer_list,
2778 struct xlog_recover_item *item)
2780 xfs_mount_t *mp = log->l_mp;
2782 struct xfs_disk_dquot *ddq, *recddq;
2784 xfs_dq_logformat_t *dq_f;
2789 * Filesystems are required to send in quota flags at mount time.
2791 if (mp->m_qflags == 0)
2794 recddq = item->ri_buf[1].i_addr;
2795 if (recddq == NULL) {
2796 xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
2797 return XFS_ERROR(EIO);
2799 if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2800 xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
2801 item->ri_buf[1].i_len, __func__);
2802 return XFS_ERROR(EIO);
2806 * This type of quotas was turned off, so ignore this record.
2808 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2810 if (log->l_quotaoffs_flag & type)
2814 * At this point we know that quota was _not_ turned off.
2815 * Since the mount flags are not indicating to us otherwise, this
2816 * must mean that quota is on, and the dquot needs to be replayed.
2817 * Remember that we may not have fully recovered the superblock yet,
2818 * so we can't do the usual trick of looking at the SB quota bits.
2820 * The other possibility, of course, is that the quota subsystem was
2821 * removed since the last mount - ENOSYS.
2823 dq_f = item->ri_buf[0].i_addr;
2825 error = xfs_qm_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2826 "xlog_recover_dquot_pass2 (log copy)");
2828 return XFS_ERROR(EIO);
2829 ASSERT(dq_f->qlf_len == 1);
2831 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
2832 XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
2838 ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2841 * At least the magic num portion should be on disk because this
2842 * was among a chunk of dquots created earlier, and we did some
2843 * minimal initialization then.
2845 error = xfs_qm_dqcheck(mp, ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2846 "xlog_recover_dquot_pass2");
2849 return XFS_ERROR(EIO);
2852 memcpy(ddq, recddq, item->ri_buf[1].i_len);
2853 if (xfs_sb_version_hascrc(&mp->m_sb)) {
2854 xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
2858 ASSERT(dq_f->qlf_size == 2);
2859 ASSERT(bp->b_target->bt_mount == mp);
2860 bp->b_iodone = xlog_recover_iodone;
2861 xfs_buf_delwri_queue(bp, buffer_list);
2868 * This routine is called to create an in-core extent free intent
2869 * item from the efi format structure which was logged on disk.
2870 * It allocates an in-core efi, copies the extents from the format
2871 * structure into it, and adds the efi to the AIL with the given
2875 xlog_recover_efi_pass2(
2877 struct xlog_recover_item *item,
2881 xfs_mount_t *mp = log->l_mp;
2882 xfs_efi_log_item_t *efip;
2883 xfs_efi_log_format_t *efi_formatp;
2885 efi_formatp = item->ri_buf[0].i_addr;
2887 efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2888 if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2889 &(efip->efi_format)))) {
2890 xfs_efi_item_free(efip);
2893 atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
2895 spin_lock(&log->l_ailp->xa_lock);
2897 * xfs_trans_ail_update() drops the AIL lock.
2899 xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
2905 * This routine is called when an efd format structure is found in
2906 * a committed transaction in the log. It's purpose is to cancel
2907 * the corresponding efi if it was still in the log. To do this
2908 * it searches the AIL for the efi with an id equal to that in the
2909 * efd format structure. If we find it, we remove the efi from the
2913 xlog_recover_efd_pass2(
2915 struct xlog_recover_item *item)
2917 xfs_efd_log_format_t *efd_formatp;
2918 xfs_efi_log_item_t *efip = NULL;
2919 xfs_log_item_t *lip;
2921 struct xfs_ail_cursor cur;
2922 struct xfs_ail *ailp = log->l_ailp;
2924 efd_formatp = item->ri_buf[0].i_addr;
2925 ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2926 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2927 (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2928 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2929 efi_id = efd_formatp->efd_efi_id;
2932 * Search for the efi with the id in the efd format structure
2935 spin_lock(&ailp->xa_lock);
2936 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2937 while (lip != NULL) {
2938 if (lip->li_type == XFS_LI_EFI) {
2939 efip = (xfs_efi_log_item_t *)lip;
2940 if (efip->efi_format.efi_id == efi_id) {
2942 * xfs_trans_ail_delete() drops the
2945 xfs_trans_ail_delete(ailp, lip,
2946 SHUTDOWN_CORRUPT_INCORE);
2947 xfs_efi_item_free(efip);
2948 spin_lock(&ailp->xa_lock);
2952 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2954 xfs_trans_ail_cursor_done(ailp, &cur);
2955 spin_unlock(&ailp->xa_lock);
2961 * Free up any resources allocated by the transaction
2963 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2966 xlog_recover_free_trans(
2967 struct xlog_recover *trans)
2969 xlog_recover_item_t *item, *n;
2972 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2973 /* Free the regions in the item. */
2974 list_del(&item->ri_list);
2975 for (i = 0; i < item->ri_cnt; i++)
2976 kmem_free(item->ri_buf[i].i_addr);
2977 /* Free the item itself */
2978 kmem_free(item->ri_buf);
2981 /* Free the transaction recover structure */
2986 xlog_recover_commit_pass1(
2988 struct xlog_recover *trans,
2989 struct xlog_recover_item *item)
2991 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
2993 switch (ITEM_TYPE(item)) {
2995 return xlog_recover_buffer_pass1(log, item);
2996 case XFS_LI_QUOTAOFF:
2997 return xlog_recover_quotaoff_pass1(log, item);
3002 /* nothing to do in pass 1 */
3005 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
3006 __func__, ITEM_TYPE(item));
3008 return XFS_ERROR(EIO);
3013 xlog_recover_commit_pass2(
3015 struct xlog_recover *trans,
3016 struct list_head *buffer_list,
3017 struct xlog_recover_item *item)
3019 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
3021 switch (ITEM_TYPE(item)) {
3023 return xlog_recover_buffer_pass2(log, buffer_list, item);
3025 return xlog_recover_inode_pass2(log, buffer_list, item);
3027 return xlog_recover_efi_pass2(log, item, trans->r_lsn);
3029 return xlog_recover_efd_pass2(log, item);
3031 return xlog_recover_dquot_pass2(log, buffer_list, item);
3032 case XFS_LI_QUOTAOFF:
3033 /* nothing to do in pass2 */
3036 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
3037 __func__, ITEM_TYPE(item));
3039 return XFS_ERROR(EIO);
3044 * Perform the transaction.
3046 * If the transaction modifies a buffer or inode, do it now. Otherwise,
3047 * EFIs and EFDs get queued up by adding entries into the AIL for them.
3050 xlog_recover_commit_trans(
3052 struct xlog_recover *trans,
3055 int error = 0, error2;
3056 xlog_recover_item_t *item;
3057 LIST_HEAD (buffer_list);
3059 hlist_del(&trans->r_list);
3061 error = xlog_recover_reorder_trans(log, trans, pass);
3065 list_for_each_entry(item, &trans->r_itemq, ri_list) {
3067 case XLOG_RECOVER_PASS1:
3068 error = xlog_recover_commit_pass1(log, trans, item);
3070 case XLOG_RECOVER_PASS2:
3071 error = xlog_recover_commit_pass2(log, trans,
3072 &buffer_list, item);
3082 xlog_recover_free_trans(trans);
3085 error2 = xfs_buf_delwri_submit(&buffer_list);
3086 return error ? error : error2;
3090 xlog_recover_unmount_trans(
3092 struct xlog_recover *trans)
3094 /* Do nothing now */
3095 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
3100 * There are two valid states of the r_state field. 0 indicates that the
3101 * transaction structure is in a normal state. We have either seen the
3102 * start of the transaction or the last operation we added was not a partial
3103 * operation. If the last operation we added to the transaction was a
3104 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
3106 * NOTE: skip LRs with 0 data length.
3109 xlog_recover_process_data(
3111 struct hlist_head rhash[],
3112 struct xlog_rec_header *rhead,
3118 xlog_op_header_t *ohead;
3119 xlog_recover_t *trans;
3125 lp = dp + be32_to_cpu(rhead->h_len);
3126 num_logops = be32_to_cpu(rhead->h_num_logops);
3128 /* check the log format matches our own - else we can't recover */
3129 if (xlog_header_check_recover(log->l_mp, rhead))
3130 return (XFS_ERROR(EIO));
3132 while ((dp < lp) && num_logops) {
3133 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
3134 ohead = (xlog_op_header_t *)dp;
3135 dp += sizeof(xlog_op_header_t);
3136 if (ohead->oh_clientid != XFS_TRANSACTION &&
3137 ohead->oh_clientid != XFS_LOG) {
3138 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
3139 __func__, ohead->oh_clientid);
3141 return (XFS_ERROR(EIO));
3143 tid = be32_to_cpu(ohead->oh_tid);
3144 hash = XLOG_RHASH(tid);
3145 trans = xlog_recover_find_tid(&rhash[hash], tid);
3146 if (trans == NULL) { /* not found; add new tid */
3147 if (ohead->oh_flags & XLOG_START_TRANS)
3148 xlog_recover_new_tid(&rhash[hash], tid,
3149 be64_to_cpu(rhead->h_lsn));
3151 if (dp + be32_to_cpu(ohead->oh_len) > lp) {
3152 xfs_warn(log->l_mp, "%s: bad length 0x%x",
3153 __func__, be32_to_cpu(ohead->oh_len));
3155 return (XFS_ERROR(EIO));
3157 flags = ohead->oh_flags & ~XLOG_END_TRANS;
3158 if (flags & XLOG_WAS_CONT_TRANS)
3159 flags &= ~XLOG_CONTINUE_TRANS;
3161 case XLOG_COMMIT_TRANS:
3162 error = xlog_recover_commit_trans(log,
3165 case XLOG_UNMOUNT_TRANS:
3166 error = xlog_recover_unmount_trans(log, trans);
3168 case XLOG_WAS_CONT_TRANS:
3169 error = xlog_recover_add_to_cont_trans(log,
3171 be32_to_cpu(ohead->oh_len));
3173 case XLOG_START_TRANS:
3174 xfs_warn(log->l_mp, "%s: bad transaction",
3177 error = XFS_ERROR(EIO);
3180 case XLOG_CONTINUE_TRANS:
3181 error = xlog_recover_add_to_trans(log, trans,
3182 dp, be32_to_cpu(ohead->oh_len));
3185 xfs_warn(log->l_mp, "%s: bad flag 0x%x",
3188 error = XFS_ERROR(EIO);
3194 dp += be32_to_cpu(ohead->oh_len);
3201 * Process an extent free intent item that was recovered from
3202 * the log. We need to free the extents that it describes.
3205 xlog_recover_process_efi(
3207 xfs_efi_log_item_t *efip)
3209 xfs_efd_log_item_t *efdp;
3214 xfs_fsblock_t startblock_fsb;
3216 ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
3219 * First check the validity of the extents described by the
3220 * EFI. If any are bad, then assume that all are bad and
3221 * just toss the EFI.
3223 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3224 extp = &(efip->efi_format.efi_extents[i]);
3225 startblock_fsb = XFS_BB_TO_FSB(mp,
3226 XFS_FSB_TO_DADDR(mp, extp->ext_start));
3227 if ((startblock_fsb == 0) ||
3228 (extp->ext_len == 0) ||
3229 (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3230 (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3232 * This will pull the EFI from the AIL and
3233 * free the memory associated with it.
3235 set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
3236 xfs_efi_release(efip, efip->efi_format.efi_nextents);
3237 return XFS_ERROR(EIO);
3241 tp = xfs_trans_alloc(mp, 0);
3242 error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3245 efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3247 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3248 extp = &(efip->efi_format.efi_extents[i]);
3249 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3252 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3256 set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
3257 error = xfs_trans_commit(tp, 0);
3261 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3266 * When this is called, all of the EFIs which did not have
3267 * corresponding EFDs should be in the AIL. What we do now
3268 * is free the extents associated with each one.
3270 * Since we process the EFIs in normal transactions, they
3271 * will be removed at some point after the commit. This prevents
3272 * us from just walking down the list processing each one.
3273 * We'll use a flag in the EFI to skip those that we've already
3274 * processed and use the AIL iteration mechanism's generation
3275 * count to try to speed this up at least a bit.
3277 * When we start, we know that the EFIs are the only things in
3278 * the AIL. As we process them, however, other items are added
3279 * to the AIL. Since everything added to the AIL must come after
3280 * everything already in the AIL, we stop processing as soon as
3281 * we see something other than an EFI in the AIL.
3284 xlog_recover_process_efis(
3287 xfs_log_item_t *lip;
3288 xfs_efi_log_item_t *efip;
3290 struct xfs_ail_cursor cur;
3291 struct xfs_ail *ailp;
3294 spin_lock(&ailp->xa_lock);
3295 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3296 while (lip != NULL) {
3298 * We're done when we see something other than an EFI.
3299 * There should be no EFIs left in the AIL now.
3301 if (lip->li_type != XFS_LI_EFI) {
3303 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
3304 ASSERT(lip->li_type != XFS_LI_EFI);
3310 * Skip EFIs that we've already processed.
3312 efip = (xfs_efi_log_item_t *)lip;
3313 if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
3314 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3318 spin_unlock(&ailp->xa_lock);
3319 error = xlog_recover_process_efi(log->l_mp, efip);
3320 spin_lock(&ailp->xa_lock);
3323 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3326 xfs_trans_ail_cursor_done(ailp, &cur);
3327 spin_unlock(&ailp->xa_lock);
3332 * This routine performs a transaction to null out a bad inode pointer
3333 * in an agi unlinked inode hash bucket.
3336 xlog_recover_clear_agi_bucket(
3338 xfs_agnumber_t agno,
3347 tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3348 error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp),
3353 error = xfs_read_agi(mp, tp, agno, &agibp);
3357 agi = XFS_BUF_TO_AGI(agibp);
3358 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3359 offset = offsetof(xfs_agi_t, agi_unlinked) +
3360 (sizeof(xfs_agino_t) * bucket);
3361 xfs_trans_log_buf(tp, agibp, offset,
3362 (offset + sizeof(xfs_agino_t) - 1));
3364 error = xfs_trans_commit(tp, 0);
3370 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3372 xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
3377 xlog_recover_process_one_iunlink(
3378 struct xfs_mount *mp,
3379 xfs_agnumber_t agno,
3383 struct xfs_buf *ibp;
3384 struct xfs_dinode *dip;
3385 struct xfs_inode *ip;
3389 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3390 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
3395 * Get the on disk inode to find the next inode in the bucket.
3397 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
3401 ASSERT(ip->i_d.di_nlink == 0);
3402 ASSERT(ip->i_d.di_mode != 0);
3404 /* setup for the next pass */
3405 agino = be32_to_cpu(dip->di_next_unlinked);
3409 * Prevent any DMAPI event from being sent when the reference on
3410 * the inode is dropped.
3412 ip->i_d.di_dmevmask = 0;
3421 * We can't read in the inode this bucket points to, or this inode
3422 * is messed up. Just ditch this bucket of inodes. We will lose
3423 * some inodes and space, but at least we won't hang.
3425 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3426 * clear the inode pointer in the bucket.
3428 xlog_recover_clear_agi_bucket(mp, agno, bucket);
3433 * xlog_iunlink_recover
3435 * This is called during recovery to process any inodes which
3436 * we unlinked but not freed when the system crashed. These
3437 * inodes will be on the lists in the AGI blocks. What we do
3438 * here is scan all the AGIs and fully truncate and free any
3439 * inodes found on the lists. Each inode is removed from the
3440 * lists when it has been fully truncated and is freed. The
3441 * freeing of the inode and its removal from the list must be
3445 xlog_recover_process_iunlinks(
3449 xfs_agnumber_t agno;
3460 * Prevent any DMAPI event from being sent while in this function.
3462 mp_dmevmask = mp->m_dmevmask;
3465 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3467 * Find the agi for this ag.
3469 error = xfs_read_agi(mp, NULL, agno, &agibp);
3472 * AGI is b0rked. Don't process it.
3474 * We should probably mark the filesystem as corrupt
3475 * after we've recovered all the ag's we can....
3480 * Unlock the buffer so that it can be acquired in the normal
3481 * course of the transaction to truncate and free each inode.
3482 * Because we are not racing with anyone else here for the AGI
3483 * buffer, we don't even need to hold it locked to read the
3484 * initial unlinked bucket entries out of the buffer. We keep
3485 * buffer reference though, so that it stays pinned in memory
3486 * while we need the buffer.
3488 agi = XFS_BUF_TO_AGI(agibp);
3489 xfs_buf_unlock(agibp);
3491 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3492 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3493 while (agino != NULLAGINO) {
3494 agino = xlog_recover_process_one_iunlink(mp,
3495 agno, agino, bucket);
3498 xfs_buf_rele(agibp);
3501 mp->m_dmevmask = mp_dmevmask;
3505 * Upack the log buffer data and crc check it. If the check fails, issue a
3506 * warning if and only if the CRC in the header is non-zero. This makes the
3507 * check an advisory warning, and the zero CRC check will prevent failure
3508 * warnings from being emitted when upgrading the kernel from one that does not
3509 * add CRCs by default.
3511 * When filesystems are CRC enabled, this CRC mismatch becomes a fatal log
3512 * corruption failure
3515 xlog_unpack_data_crc(
3516 struct xlog_rec_header *rhead,
3522 crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
3523 if (crc != rhead->h_crc) {
3524 if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
3525 xfs_alert(log->l_mp,
3526 "log record CRC mismatch: found 0x%x, expected 0x%x.\n",
3527 le32_to_cpu(rhead->h_crc),
3529 xfs_hex_dump(dp, 32);
3533 * If we've detected a log record corruption, then we can't
3534 * recover past this point. Abort recovery if we are enforcing
3535 * CRC protection by punting an error back up the stack.
3537 if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
3538 return EFSCORRUPTED;
3546 struct xlog_rec_header *rhead,
3553 error = xlog_unpack_data_crc(rhead, dp, log);
3557 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3558 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3559 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3563 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3564 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3565 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3566 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3567 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3568 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3577 xlog_valid_rec_header(
3579 struct xlog_rec_header *rhead,
3584 if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
3585 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3586 XFS_ERRLEVEL_LOW, log->l_mp);
3587 return XFS_ERROR(EFSCORRUPTED);
3590 (!rhead->h_version ||
3591 (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3592 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
3593 __func__, be32_to_cpu(rhead->h_version));
3594 return XFS_ERROR(EIO);
3597 /* LR body must have data or it wouldn't have been written */
3598 hlen = be32_to_cpu(rhead->h_len);
3599 if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3600 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3601 XFS_ERRLEVEL_LOW, log->l_mp);
3602 return XFS_ERROR(EFSCORRUPTED);
3604 if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3605 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3606 XFS_ERRLEVEL_LOW, log->l_mp);
3607 return XFS_ERROR(EFSCORRUPTED);
3613 * Read the log from tail to head and process the log records found.
3614 * Handle the two cases where the tail and head are in the same cycle
3615 * and where the active portion of the log wraps around the end of
3616 * the physical log separately. The pass parameter is passed through
3617 * to the routines called to process the data and is not looked at
3621 xlog_do_recovery_pass(
3623 xfs_daddr_t head_blk,
3624 xfs_daddr_t tail_blk,
3627 xlog_rec_header_t *rhead;
3630 xfs_buf_t *hbp, *dbp;
3631 int error = 0, h_size;
3632 int bblks, split_bblks;
3633 int hblks, split_hblks, wrapped_hblks;
3634 struct hlist_head rhash[XLOG_RHASH_SIZE];
3636 ASSERT(head_blk != tail_blk);
3639 * Read the header of the tail block and get the iclog buffer size from
3640 * h_size. Use this to tell how many sectors make up the log header.
3642 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3644 * When using variable length iclogs, read first sector of
3645 * iclog header and extract the header size from it. Get a
3646 * new hbp that is the correct size.
3648 hbp = xlog_get_bp(log, 1);
3652 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3656 rhead = (xlog_rec_header_t *)offset;
3657 error = xlog_valid_rec_header(log, rhead, tail_blk);
3660 h_size = be32_to_cpu(rhead->h_size);
3661 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3662 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3663 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3664 if (h_size % XLOG_HEADER_CYCLE_SIZE)
3667 hbp = xlog_get_bp(log, hblks);
3672 ASSERT(log->l_sectBBsize == 1);
3674 hbp = xlog_get_bp(log, 1);
3675 h_size = XLOG_BIG_RECORD_BSIZE;
3680 dbp = xlog_get_bp(log, BTOBB(h_size));
3686 memset(rhash, 0, sizeof(rhash));
3687 if (tail_blk <= head_blk) {
3688 for (blk_no = tail_blk; blk_no < head_blk; ) {
3689 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3693 rhead = (xlog_rec_header_t *)offset;
3694 error = xlog_valid_rec_header(log, rhead, blk_no);
3698 /* blocks in data section */
3699 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3700 error = xlog_bread(log, blk_no + hblks, bblks, dbp,
3705 error = xlog_unpack_data(rhead, offset, log);
3709 error = xlog_recover_process_data(log,
3710 rhash, rhead, offset, pass);
3713 blk_no += bblks + hblks;
3717 * Perform recovery around the end of the physical log.
3718 * When the head is not on the same cycle number as the tail,
3719 * we can't do a sequential recovery as above.
3722 while (blk_no < log->l_logBBsize) {
3724 * Check for header wrapping around physical end-of-log
3726 offset = hbp->b_addr;
3729 if (blk_no + hblks <= log->l_logBBsize) {
3730 /* Read header in one read */
3731 error = xlog_bread(log, blk_no, hblks, hbp,
3736 /* This LR is split across physical log end */
3737 if (blk_no != log->l_logBBsize) {
3738 /* some data before physical log end */
3739 ASSERT(blk_no <= INT_MAX);
3740 split_hblks = log->l_logBBsize - (int)blk_no;
3741 ASSERT(split_hblks > 0);
3742 error = xlog_bread(log, blk_no,
3750 * Note: this black magic still works with
3751 * large sector sizes (non-512) only because:
3752 * - we increased the buffer size originally
3753 * by 1 sector giving us enough extra space
3754 * for the second read;
3755 * - the log start is guaranteed to be sector
3757 * - we read the log end (LR header start)
3758 * _first_, then the log start (LR header end)
3759 * - order is important.
3761 wrapped_hblks = hblks - split_hblks;
3762 error = xlog_bread_offset(log, 0,
3764 offset + BBTOB(split_hblks));
3768 rhead = (xlog_rec_header_t *)offset;
3769 error = xlog_valid_rec_header(log, rhead,
3770 split_hblks ? blk_no : 0);
3774 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3777 /* Read in data for log record */
3778 if (blk_no + bblks <= log->l_logBBsize) {
3779 error = xlog_bread(log, blk_no, bblks, dbp,
3784 /* This log record is split across the
3785 * physical end of log */
3786 offset = dbp->b_addr;
3788 if (blk_no != log->l_logBBsize) {
3789 /* some data is before the physical
3791 ASSERT(!wrapped_hblks);
3792 ASSERT(blk_no <= INT_MAX);
3794 log->l_logBBsize - (int)blk_no;
3795 ASSERT(split_bblks > 0);
3796 error = xlog_bread(log, blk_no,
3804 * Note: this black magic still works with
3805 * large sector sizes (non-512) only because:
3806 * - we increased the buffer size originally
3807 * by 1 sector giving us enough extra space
3808 * for the second read;
3809 * - the log start is guaranteed to be sector
3811 * - we read the log end (LR header start)
3812 * _first_, then the log start (LR header end)
3813 * - order is important.
3815 error = xlog_bread_offset(log, 0,
3816 bblks - split_bblks, dbp,
3817 offset + BBTOB(split_bblks));
3822 error = xlog_unpack_data(rhead, offset, log);
3826 error = xlog_recover_process_data(log, rhash,
3827 rhead, offset, pass);
3833 ASSERT(blk_no >= log->l_logBBsize);
3834 blk_no -= log->l_logBBsize;
3836 /* read first part of physical log */
3837 while (blk_no < head_blk) {
3838 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3842 rhead = (xlog_rec_header_t *)offset;
3843 error = xlog_valid_rec_header(log, rhead, blk_no);
3847 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3848 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3853 error = xlog_unpack_data(rhead, offset, log);
3857 error = xlog_recover_process_data(log, rhash,
3858 rhead, offset, pass);
3861 blk_no += bblks + hblks;
3873 * Do the recovery of the log. We actually do this in two phases.
3874 * The two passes are necessary in order to implement the function
3875 * of cancelling a record written into the log. The first pass
3876 * determines those things which have been cancelled, and the
3877 * second pass replays log items normally except for those which
3878 * have been cancelled. The handling of the replay and cancellations
3879 * takes place in the log item type specific routines.
3881 * The table of items which have cancel records in the log is allocated
3882 * and freed at this level, since only here do we know when all of
3883 * the log recovery has been completed.
3886 xlog_do_log_recovery(
3888 xfs_daddr_t head_blk,
3889 xfs_daddr_t tail_blk)
3893 ASSERT(head_blk != tail_blk);
3896 * First do a pass to find all of the cancelled buf log items.
3897 * Store them in the buf_cancel_table for use in the second pass.
3899 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3900 sizeof(struct list_head),
3902 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3903 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3905 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3906 XLOG_RECOVER_PASS1);
3908 kmem_free(log->l_buf_cancel_table);
3909 log->l_buf_cancel_table = NULL;
3913 * Then do a second pass to actually recover the items in the log.
3914 * When it is complete free the table of buf cancel items.
3916 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3917 XLOG_RECOVER_PASS2);
3922 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3923 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3927 kmem_free(log->l_buf_cancel_table);
3928 log->l_buf_cancel_table = NULL;
3934 * Do the actual recovery
3939 xfs_daddr_t head_blk,
3940 xfs_daddr_t tail_blk)
3947 * First replay the images in the log.
3949 error = xlog_do_log_recovery(log, head_blk, tail_blk);
3954 * If IO errors happened during recovery, bail out.
3956 if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3961 * We now update the tail_lsn since much of the recovery has completed
3962 * and there may be space available to use. If there were no extent
3963 * or iunlinks, we can free up the entire log and set the tail_lsn to
3964 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3965 * lsn of the last known good LR on disk. If there are extent frees
3966 * or iunlinks they will have some entries in the AIL; so we look at
3967 * the AIL to determine how to set the tail_lsn.
3969 xlog_assign_tail_lsn(log->l_mp);
3972 * Now that we've finished replaying all buffer and inode
3973 * updates, re-read in the superblock and reverify it.
3975 bp = xfs_getsb(log->l_mp, 0);
3977 ASSERT(!(XFS_BUF_ISWRITE(bp)));
3979 XFS_BUF_UNASYNC(bp);
3980 bp->b_ops = &xfs_sb_buf_ops;
3981 xfsbdstrat(log->l_mp, bp);
3982 error = xfs_buf_iowait(bp);
3984 xfs_buf_ioerror_alert(bp, __func__);
3990 /* Convert superblock from on-disk format */
3991 sbp = &log->l_mp->m_sb;
3992 xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3993 ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3994 ASSERT(xfs_sb_good_version(sbp));
3997 /* We've re-read the superblock so re-initialize per-cpu counters */
3998 xfs_icsb_reinit_counters(log->l_mp);
4000 xlog_recover_check_summary(log);
4002 /* Normal transactions can now occur */
4003 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
4008 * Perform recovery and re-initialize some log variables in xlog_find_tail.
4010 * Return error or zero.
4016 xfs_daddr_t head_blk, tail_blk;
4019 /* find the tail of the log */
4020 if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
4023 if (tail_blk != head_blk) {
4024 /* There used to be a comment here:
4026 * disallow recovery on read-only mounts. note -- mount
4027 * checks for ENOSPC and turns it into an intelligent
4029 * ...but this is no longer true. Now, unless you specify
4030 * NORECOVERY (in which case this function would never be
4031 * called), we just go ahead and recover. We do this all
4032 * under the vfs layer, so we can get away with it unless
4033 * the device itself is read-only, in which case we fail.
4035 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
4040 * Version 5 superblock log feature mask validation. We know the
4041 * log is dirty so check if there are any unknown log features
4042 * in what we need to recover. If there are unknown features
4043 * (e.g. unsupported transactions, then simply reject the
4044 * attempt at recovery before touching anything.
4046 if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
4047 xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
4048 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
4050 "Superblock has unknown incompatible log features (0x%x) enabled.\n"
4051 "The log can not be fully and/or safely recovered by this kernel.\n"
4052 "Please recover the log on a kernel that supports the unknown features.",
4053 (log->l_mp->m_sb.sb_features_log_incompat &
4054 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
4058 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
4059 log->l_mp->m_logname ? log->l_mp->m_logname
4062 error = xlog_do_recover(log, head_blk, tail_blk);
4063 log->l_flags |= XLOG_RECOVERY_NEEDED;
4069 * In the first part of recovery we replay inodes and buffers and build
4070 * up the list of extent free items which need to be processed. Here
4071 * we process the extent free items and clean up the on disk unlinked
4072 * inode lists. This is separated from the first part of recovery so
4073 * that the root and real-time bitmap inodes can be read in from disk in
4074 * between the two stages. This is necessary so that we can free space
4075 * in the real-time portion of the file system.
4078 xlog_recover_finish(
4082 * Now we're ready to do the transactions needed for the
4083 * rest of recovery. Start with completing all the extent
4084 * free intent records and then process the unlinked inode
4085 * lists. At this point, we essentially run in normal mode
4086 * except that we're still performing recovery actions
4087 * rather than accepting new requests.
4089 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
4091 error = xlog_recover_process_efis(log);
4093 xfs_alert(log->l_mp, "Failed to recover EFIs");
4097 * Sync the log to get all the EFIs out of the AIL.
4098 * This isn't absolutely necessary, but it helps in
4099 * case the unlink transactions would have problems
4100 * pushing the EFIs out of the way.
4102 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
4104 xlog_recover_process_iunlinks(log);
4106 xlog_recover_check_summary(log);
4108 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
4109 log->l_mp->m_logname ? log->l_mp->m_logname
4111 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
4113 xfs_info(log->l_mp, "Ending clean mount");
4121 * Read all of the agf and agi counters and check that they
4122 * are consistent with the superblock counters.
4125 xlog_recover_check_summary(
4132 xfs_agnumber_t agno;
4133 __uint64_t freeblks;
4143 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4144 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
4146 xfs_alert(mp, "%s agf read failed agno %d error %d",
4147 __func__, agno, error);
4149 agfp = XFS_BUF_TO_AGF(agfbp);
4150 freeblks += be32_to_cpu(agfp->agf_freeblks) +
4151 be32_to_cpu(agfp->agf_flcount);
4152 xfs_buf_relse(agfbp);
4155 error = xfs_read_agi(mp, NULL, agno, &agibp);
4157 xfs_alert(mp, "%s agi read failed agno %d error %d",
4158 __func__, agno, error);
4160 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
4162 itotal += be32_to_cpu(agi->agi_count);
4163 ifree += be32_to_cpu(agi->agi_freecount);
4164 xfs_buf_relse(agibp);