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_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
28 #include "xfs_mount.h"
29 #include "xfs_da_format.h"
30 #include "xfs_inode.h"
31 #include "xfs_trans.h"
33 #include "xfs_log_priv.h"
34 #include "xfs_log_recover.h"
35 #include "xfs_inode_item.h"
36 #include "xfs_extfree_item.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_alloc.h"
39 #include "xfs_ialloc.h"
40 #include "xfs_quota.h"
41 #include "xfs_cksum.h"
42 #include "xfs_trace.h"
43 #include "xfs_icache.h"
44 #include "xfs_bmap_btree.h"
45 #include "xfs_dinode.h"
46 #include "xfs_error.h"
49 #define BLK_AVG(blk1, blk2) ((blk1+blk2) >> 1)
56 xlog_clear_stale_blocks(
61 xlog_recover_check_summary(
64 #define xlog_recover_check_summary(log)
68 * This structure is used during recovery to record the buf log items which
69 * have been canceled and should not be replayed.
71 struct xfs_buf_cancel
{
75 struct list_head bc_list
;
79 * Sector aligned buffer routines for buffer create/read/write/access
83 * Verify the given count of basic blocks is valid number of blocks
84 * to specify for an operation involving the given XFS log buffer.
85 * Returns nonzero if the count is valid, 0 otherwise.
89 xlog_buf_bbcount_valid(
93 return bbcount
> 0 && bbcount
<= log
->l_logBBsize
;
97 * Allocate a buffer to hold log data. The buffer needs to be able
98 * to map to a range of nbblks basic blocks at any valid (basic
99 * block) offset within the log.
108 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
109 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
111 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
116 * We do log I/O in units of log sectors (a power-of-2
117 * multiple of the basic block size), so we round up the
118 * requested size to accommodate the basic blocks required
119 * for complete log sectors.
121 * In addition, the buffer may be used for a non-sector-
122 * aligned block offset, in which case an I/O of the
123 * requested size could extend beyond the end of the
124 * buffer. If the requested size is only 1 basic block it
125 * will never straddle a sector boundary, so this won't be
126 * an issue. Nor will this be a problem if the log I/O is
127 * done in basic blocks (sector size 1). But otherwise we
128 * extend the buffer by one extra log sector to ensure
129 * there's space to accommodate this possibility.
131 if (nbblks
> 1 && log
->l_sectBBsize
> 1)
132 nbblks
+= log
->l_sectBBsize
;
133 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
135 bp
= xfs_buf_get_uncached(log
->l_mp
->m_logdev_targp
, nbblks
, 0);
149 * Return the address of the start of the given block number's data
150 * in a log buffer. The buffer covers a log sector-aligned region.
159 xfs_daddr_t offset
= blk_no
& ((xfs_daddr_t
)log
->l_sectBBsize
- 1);
161 ASSERT(offset
+ nbblks
<= bp
->b_length
);
162 return bp
->b_addr
+ BBTOB(offset
);
167 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
178 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
179 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
181 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
182 return -EFSCORRUPTED
;
185 blk_no
= round_down(blk_no
, log
->l_sectBBsize
);
186 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
189 ASSERT(nbblks
<= bp
->b_length
);
191 XFS_BUF_SET_ADDR(bp
, log
->l_logBBstart
+ blk_no
);
193 bp
->b_io_length
= nbblks
;
196 if (XFS_FORCED_SHUTDOWN(log
->l_mp
))
199 xfs_buf_iorequest(bp
);
200 error
= xfs_buf_iowait(bp
);
202 xfs_buf_ioerror_alert(bp
, __func__
);
216 error
= xlog_bread_noalign(log
, blk_no
, nbblks
, bp
);
220 *offset
= xlog_align(log
, blk_no
, nbblks
, bp
);
225 * Read at an offset into the buffer. Returns with the buffer in it's original
226 * state regardless of the result of the read.
231 xfs_daddr_t blk_no
, /* block to read from */
232 int nbblks
, /* blocks to read */
236 xfs_caddr_t orig_offset
= bp
->b_addr
;
237 int orig_len
= BBTOB(bp
->b_length
);
240 error
= xfs_buf_associate_memory(bp
, offset
, BBTOB(nbblks
));
244 error
= xlog_bread_noalign(log
, blk_no
, nbblks
, bp
);
246 /* must reset buffer pointer even on error */
247 error2
= xfs_buf_associate_memory(bp
, orig_offset
, orig_len
);
254 * Write out the buffer at the given block for the given number of blocks.
255 * The buffer is kept locked across the write and is returned locked.
256 * This can only be used for synchronous log writes.
267 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
268 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
270 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
271 return -EFSCORRUPTED
;
274 blk_no
= round_down(blk_no
, log
->l_sectBBsize
);
275 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
278 ASSERT(nbblks
<= bp
->b_length
);
280 XFS_BUF_SET_ADDR(bp
, log
->l_logBBstart
+ blk_no
);
281 XFS_BUF_ZEROFLAGS(bp
);
284 bp
->b_io_length
= nbblks
;
287 error
= xfs_bwrite(bp
);
289 xfs_buf_ioerror_alert(bp
, __func__
);
296 * dump debug superblock and log record information
299 xlog_header_check_dump(
301 xlog_rec_header_t
*head
)
303 xfs_debug(mp
, "%s: SB : uuid = %pU, fmt = %d",
304 __func__
, &mp
->m_sb
.sb_uuid
, XLOG_FMT
);
305 xfs_debug(mp
, " log : uuid = %pU, fmt = %d",
306 &head
->h_fs_uuid
, be32_to_cpu(head
->h_fmt
));
309 #define xlog_header_check_dump(mp, head)
313 * check log record header for recovery
316 xlog_header_check_recover(
318 xlog_rec_header_t
*head
)
320 ASSERT(head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
));
323 * IRIX doesn't write the h_fmt field and leaves it zeroed
324 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
325 * a dirty log created in IRIX.
327 if (unlikely(head
->h_fmt
!= cpu_to_be32(XLOG_FMT
))) {
329 "dirty log written in incompatible format - can't recover");
330 xlog_header_check_dump(mp
, head
);
331 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
332 XFS_ERRLEVEL_HIGH
, mp
);
333 return -EFSCORRUPTED
;
334 } else if (unlikely(!uuid_equal(&mp
->m_sb
.sb_uuid
, &head
->h_fs_uuid
))) {
336 "dirty log entry has mismatched uuid - can't recover");
337 xlog_header_check_dump(mp
, head
);
338 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
339 XFS_ERRLEVEL_HIGH
, mp
);
340 return -EFSCORRUPTED
;
346 * read the head block of the log and check the header
349 xlog_header_check_mount(
351 xlog_rec_header_t
*head
)
353 ASSERT(head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
));
355 if (uuid_is_nil(&head
->h_fs_uuid
)) {
357 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
358 * h_fs_uuid is nil, we assume this log was last mounted
359 * by IRIX and continue.
361 xfs_warn(mp
, "nil uuid in log - IRIX style log");
362 } else if (unlikely(!uuid_equal(&mp
->m_sb
.sb_uuid
, &head
->h_fs_uuid
))) {
363 xfs_warn(mp
, "log has mismatched uuid - can't recover");
364 xlog_header_check_dump(mp
, head
);
365 XFS_ERROR_REPORT("xlog_header_check_mount",
366 XFS_ERRLEVEL_HIGH
, mp
);
367 return -EFSCORRUPTED
;
378 * We're not going to bother about retrying
379 * this during recovery. One strike!
381 xfs_buf_ioerror_alert(bp
, __func__
);
382 xfs_force_shutdown(bp
->b_target
->bt_mount
,
383 SHUTDOWN_META_IO_ERROR
);
386 xfs_buf_ioend(bp
, 0);
390 * This routine finds (to an approximation) the first block in the physical
391 * log which contains the given cycle. It uses a binary search algorithm.
392 * Note that the algorithm can not be perfect because the disk will not
393 * necessarily be perfect.
396 xlog_find_cycle_start(
399 xfs_daddr_t first_blk
,
400 xfs_daddr_t
*last_blk
,
410 mid_blk
= BLK_AVG(first_blk
, end_blk
);
411 while (mid_blk
!= first_blk
&& mid_blk
!= end_blk
) {
412 error
= xlog_bread(log
, mid_blk
, 1, bp
, &offset
);
415 mid_cycle
= xlog_get_cycle(offset
);
416 if (mid_cycle
== cycle
)
417 end_blk
= mid_blk
; /* last_half_cycle == mid_cycle */
419 first_blk
= mid_blk
; /* first_half_cycle == mid_cycle */
420 mid_blk
= BLK_AVG(first_blk
, end_blk
);
422 ASSERT((mid_blk
== first_blk
&& mid_blk
+1 == end_blk
) ||
423 (mid_blk
== end_blk
&& mid_blk
-1 == first_blk
));
431 * Check that a range of blocks does not contain stop_on_cycle_no.
432 * Fill in *new_blk with the block offset where such a block is
433 * found, or with -1 (an invalid block number) if there is no such
434 * block in the range. The scan needs to occur from front to back
435 * and the pointer into the region must be updated since a later
436 * routine will need to perform another test.
439 xlog_find_verify_cycle(
441 xfs_daddr_t start_blk
,
443 uint stop_on_cycle_no
,
444 xfs_daddr_t
*new_blk
)
450 xfs_caddr_t buf
= NULL
;
454 * Greedily allocate a buffer big enough to handle the full
455 * range of basic blocks we'll be examining. If that fails,
456 * try a smaller size. We need to be able to read at least
457 * a log sector, or we're out of luck.
459 bufblks
= 1 << ffs(nbblks
);
460 while (bufblks
> log
->l_logBBsize
)
462 while (!(bp
= xlog_get_bp(log
, bufblks
))) {
464 if (bufblks
< log
->l_sectBBsize
)
468 for (i
= start_blk
; i
< start_blk
+ nbblks
; i
+= bufblks
) {
471 bcount
= min(bufblks
, (start_blk
+ nbblks
- i
));
473 error
= xlog_bread(log
, i
, bcount
, bp
, &buf
);
477 for (j
= 0; j
< bcount
; j
++) {
478 cycle
= xlog_get_cycle(buf
);
479 if (cycle
== stop_on_cycle_no
) {
496 * Potentially backup over partial log record write.
498 * In the typical case, last_blk is the number of the block directly after
499 * a good log record. Therefore, we subtract one to get the block number
500 * of the last block in the given buffer. extra_bblks contains the number
501 * of blocks we would have read on a previous read. This happens when the
502 * last log record is split over the end of the physical log.
504 * extra_bblks is the number of blocks potentially verified on a previous
505 * call to this routine.
508 xlog_find_verify_log_record(
510 xfs_daddr_t start_blk
,
511 xfs_daddr_t
*last_blk
,
516 xfs_caddr_t offset
= NULL
;
517 xlog_rec_header_t
*head
= NULL
;
520 int num_blks
= *last_blk
- start_blk
;
523 ASSERT(start_blk
!= 0 || *last_blk
!= start_blk
);
525 if (!(bp
= xlog_get_bp(log
, num_blks
))) {
526 if (!(bp
= xlog_get_bp(log
, 1)))
530 error
= xlog_bread(log
, start_blk
, num_blks
, bp
, &offset
);
533 offset
+= ((num_blks
- 1) << BBSHIFT
);
536 for (i
= (*last_blk
) - 1; i
>= 0; i
--) {
538 /* valid log record not found */
540 "Log inconsistent (didn't find previous header)");
547 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
552 head
= (xlog_rec_header_t
*)offset
;
554 if (head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
562 * We hit the beginning of the physical log & still no header. Return
563 * to caller. If caller can handle a return of -1, then this routine
564 * will be called again for the end of the physical log.
572 * We have the final block of the good log (the first block
573 * of the log record _before_ the head. So we check the uuid.
575 if ((error
= xlog_header_check_mount(log
->l_mp
, head
)))
579 * We may have found a log record header before we expected one.
580 * last_blk will be the 1st block # with a given cycle #. We may end
581 * up reading an entire log record. In this case, we don't want to
582 * reset last_blk. Only when last_blk points in the middle of a log
583 * record do we update last_blk.
585 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
586 uint h_size
= be32_to_cpu(head
->h_size
);
588 xhdrs
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
589 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
595 if (*last_blk
- i
+ extra_bblks
!=
596 BTOBB(be32_to_cpu(head
->h_len
)) + xhdrs
)
605 * Head is defined to be the point of the log where the next log write
606 * could go. This means that incomplete LR writes at the end are
607 * eliminated when calculating the head. We aren't guaranteed that previous
608 * LR have complete transactions. We only know that a cycle number of
609 * current cycle number -1 won't be present in the log if we start writing
610 * from our current block number.
612 * last_blk contains the block number of the first block with a given
615 * Return: zero if normal, non-zero if error.
620 xfs_daddr_t
*return_head_blk
)
624 xfs_daddr_t new_blk
, first_blk
, start_blk
, last_blk
, head_blk
;
626 uint first_half_cycle
, last_half_cycle
;
628 int error
, log_bbnum
= log
->l_logBBsize
;
630 /* Is the end of the log device zeroed? */
631 error
= xlog_find_zeroed(log
, &first_blk
);
633 xfs_warn(log
->l_mp
, "empty log check failed");
637 *return_head_blk
= first_blk
;
639 /* Is the whole lot zeroed? */
641 /* Linux XFS shouldn't generate totally zeroed logs -
642 * mkfs etc write a dummy unmount record to a fresh
643 * log so we can store the uuid in there
645 xfs_warn(log
->l_mp
, "totally zeroed log");
651 first_blk
= 0; /* get cycle # of 1st block */
652 bp
= xlog_get_bp(log
, 1);
656 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
660 first_half_cycle
= xlog_get_cycle(offset
);
662 last_blk
= head_blk
= log_bbnum
- 1; /* get cycle # of last block */
663 error
= xlog_bread(log
, last_blk
, 1, bp
, &offset
);
667 last_half_cycle
= xlog_get_cycle(offset
);
668 ASSERT(last_half_cycle
!= 0);
671 * If the 1st half cycle number is equal to the last half cycle number,
672 * then the entire log is stamped with the same cycle number. In this
673 * case, head_blk can't be set to zero (which makes sense). The below
674 * math doesn't work out properly with head_blk equal to zero. Instead,
675 * we set it to log_bbnum which is an invalid block number, but this
676 * value makes the math correct. If head_blk doesn't changed through
677 * all the tests below, *head_blk is set to zero at the very end rather
678 * than log_bbnum. In a sense, log_bbnum and zero are the same block
679 * in a circular file.
681 if (first_half_cycle
== last_half_cycle
) {
683 * In this case we believe that the entire log should have
684 * cycle number last_half_cycle. We need to scan backwards
685 * from the end verifying that there are no holes still
686 * containing last_half_cycle - 1. If we find such a hole,
687 * then the start of that hole will be the new head. The
688 * simple case looks like
689 * x | x ... | x - 1 | x
690 * Another case that fits this picture would be
691 * x | x + 1 | x ... | x
692 * In this case the head really is somewhere at the end of the
693 * log, as one of the latest writes at the beginning was
696 * x | x + 1 | x ... | x - 1 | x
697 * This is really the combination of the above two cases, and
698 * the head has to end up at the start of the x-1 hole at the
701 * In the 256k log case, we will read from the beginning to the
702 * end of the log and search for cycle numbers equal to x-1.
703 * We don't worry about the x+1 blocks that we encounter,
704 * because we know that they cannot be the head since the log
707 head_blk
= log_bbnum
;
708 stop_on_cycle
= last_half_cycle
- 1;
711 * In this case we want to find the first block with cycle
712 * number matching last_half_cycle. We expect the log to be
714 * x + 1 ... | x ... | x
715 * The first block with cycle number x (last_half_cycle) will
716 * be where the new head belongs. First we do a binary search
717 * for the first occurrence of last_half_cycle. The binary
718 * search may not be totally accurate, so then we scan back
719 * from there looking for occurrences of last_half_cycle before
720 * us. If that backwards scan wraps around the beginning of
721 * the log, then we look for occurrences of last_half_cycle - 1
722 * at the end of the log. The cases we're looking for look
724 * v binary search stopped here
725 * x + 1 ... | x | x + 1 | x ... | x
726 * ^ but we want to locate this spot
728 * <---------> less than scan distance
729 * x + 1 ... | x ... | x - 1 | x
730 * ^ we want to locate this spot
732 stop_on_cycle
= last_half_cycle
;
733 if ((error
= xlog_find_cycle_start(log
, bp
, first_blk
,
734 &head_blk
, last_half_cycle
)))
739 * Now validate the answer. Scan back some number of maximum possible
740 * blocks and make sure each one has the expected cycle number. The
741 * maximum is determined by the total possible amount of buffering
742 * in the in-core log. The following number can be made tighter if
743 * we actually look at the block size of the filesystem.
745 num_scan_bblks
= XLOG_TOTAL_REC_SHIFT(log
);
746 if (head_blk
>= num_scan_bblks
) {
748 * We are guaranteed that the entire check can be performed
751 start_blk
= head_blk
- num_scan_bblks
;
752 if ((error
= xlog_find_verify_cycle(log
,
753 start_blk
, num_scan_bblks
,
754 stop_on_cycle
, &new_blk
)))
758 } else { /* need to read 2 parts of log */
760 * We are going to scan backwards in the log in two parts.
761 * First we scan the physical end of the log. In this part
762 * of the log, we are looking for blocks with cycle number
763 * last_half_cycle - 1.
764 * If we find one, then we know that the log starts there, as
765 * we've found a hole that didn't get written in going around
766 * the end of the physical log. The simple case for this is
767 * x + 1 ... | x ... | x - 1 | x
768 * <---------> less than scan distance
769 * If all of the blocks at the end of the log have cycle number
770 * last_half_cycle, then we check the blocks at the start of
771 * the log looking for occurrences of last_half_cycle. If we
772 * find one, then our current estimate for the location of the
773 * first occurrence of last_half_cycle is wrong and we move
774 * back to the hole we've found. This case looks like
775 * x + 1 ... | x | x + 1 | x ...
776 * ^ binary search stopped here
777 * Another case we need to handle that only occurs in 256k
779 * x + 1 ... | x ... | x+1 | x ...
780 * ^ binary search stops here
781 * In a 256k log, the scan at the end of the log will see the
782 * x + 1 blocks. We need to skip past those since that is
783 * certainly not the head of the log. By searching for
784 * last_half_cycle-1 we accomplish that.
786 ASSERT(head_blk
<= INT_MAX
&&
787 (xfs_daddr_t
) num_scan_bblks
>= head_blk
);
788 start_blk
= log_bbnum
- (num_scan_bblks
- head_blk
);
789 if ((error
= xlog_find_verify_cycle(log
, start_blk
,
790 num_scan_bblks
- (int)head_blk
,
791 (stop_on_cycle
- 1), &new_blk
)))
799 * Scan beginning of log now. The last part of the physical
800 * log is good. This scan needs to verify that it doesn't find
801 * the last_half_cycle.
804 ASSERT(head_blk
<= INT_MAX
);
805 if ((error
= xlog_find_verify_cycle(log
,
806 start_blk
, (int)head_blk
,
807 stop_on_cycle
, &new_blk
)))
815 * Now we need to make sure head_blk is not pointing to a block in
816 * the middle of a log record.
818 num_scan_bblks
= XLOG_REC_SHIFT(log
);
819 if (head_blk
>= num_scan_bblks
) {
820 start_blk
= head_blk
- num_scan_bblks
; /* don't read head_blk */
822 /* start ptr at last block ptr before head_blk */
823 error
= xlog_find_verify_log_record(log
, start_blk
, &head_blk
, 0);
830 ASSERT(head_blk
<= INT_MAX
);
831 error
= xlog_find_verify_log_record(log
, start_blk
, &head_blk
, 0);
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 error
= xlog_find_verify_log_record(log
, start_blk
,
842 &new_blk
, (int)head_blk
);
847 if (new_blk
!= log_bbnum
)
854 if (head_blk
== log_bbnum
)
855 *return_head_blk
= 0;
857 *return_head_blk
= head_blk
;
859 * When returning here, we have a good block number. Bad block
860 * means that during a previous crash, we didn't have a clean break
861 * from cycle number N to cycle number N-1. In this case, we need
862 * to find the first block with cycle number N-1.
870 xfs_warn(log
->l_mp
, "failed to find log head");
875 * Find the sync block number or the tail of the log.
877 * This will be the block number of the last record to have its
878 * associated buffers synced to disk. Every log record header has
879 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
880 * to get a sync block number. The only concern is to figure out which
881 * log record header to believe.
883 * The following algorithm uses the log record header with the largest
884 * lsn. The entire log record does not need to be valid. We only care
885 * that the header is valid.
887 * We could speed up search by using current head_blk buffer, but it is not
893 xfs_daddr_t
*head_blk
,
894 xfs_daddr_t
*tail_blk
)
896 xlog_rec_header_t
*rhead
;
897 xlog_op_header_t
*op_head
;
898 xfs_caddr_t offset
= NULL
;
901 xfs_daddr_t umount_data_blk
;
902 xfs_daddr_t after_umount_blk
;
909 * Find previous log record
911 if ((error
= xlog_find_head(log
, head_blk
)))
914 bp
= xlog_get_bp(log
, 1);
917 if (*head_blk
== 0) { /* special case */
918 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
922 if (xlog_get_cycle(offset
) == 0) {
924 /* leave all other log inited values alone */
930 * Search backwards looking for log record header block
932 ASSERT(*head_blk
< INT_MAX
);
933 for (i
= (int)(*head_blk
) - 1; i
>= 0; i
--) {
934 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
938 if (*(__be32
*)offset
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
)) {
944 * If we haven't found the log record header block, start looking
945 * again from the end of the physical log. XXXmiken: There should be
946 * a check here to make sure we didn't search more than N blocks in
950 for (i
= log
->l_logBBsize
- 1; i
>= (int)(*head_blk
); i
--) {
951 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
955 if (*(__be32
*)offset
==
956 cpu_to_be32(XLOG_HEADER_MAGIC_NUM
)) {
963 xfs_warn(log
->l_mp
, "%s: couldn't find sync record", __func__
);
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)");
1150 /* we have a partially zeroed log */
1151 last_blk
= log_bbnum
-1;
1152 if ((error
= xlog_find_cycle_start(log
, bp
, 0, &last_blk
, 0)))
1156 * Validate the answer. Because there is no way to guarantee that
1157 * the entire log is made up of log records which are the same size,
1158 * we scan over the defined maximum blocks. At this point, the maximum
1159 * is not chosen to mean anything special. XXXmiken
1161 num_scan_bblks
= XLOG_TOTAL_REC_SHIFT(log
);
1162 ASSERT(num_scan_bblks
<= INT_MAX
);
1164 if (last_blk
< num_scan_bblks
)
1165 num_scan_bblks
= last_blk
;
1166 start_blk
= last_blk
- num_scan_bblks
;
1169 * We search for any instances of cycle number 0 that occur before
1170 * our current estimate of the head. What we're trying to detect is
1171 * 1 ... | 0 | 1 | 0...
1172 * ^ binary search ends here
1174 if ((error
= xlog_find_verify_cycle(log
, start_blk
,
1175 (int)num_scan_bblks
, 0, &new_blk
)))
1181 * Potentially backup over partial log record write. We don't need
1182 * to search the end of the log because we know it is zero.
1184 error
= xlog_find_verify_log_record(log
, start_blk
, &last_blk
, 0);
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 -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 -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",
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",
1588 item
->ri_total
= in_f
->ilf_size
;
1590 kmem_zalloc(item
->ri_total
* sizeof(xfs_log_iovec_t
),
1593 ASSERT(item
->ri_total
> item
->ri_cnt
);
1594 /* Description region is ri_buf[0] */
1595 item
->ri_buf
[item
->ri_cnt
].i_addr
= ptr
;
1596 item
->ri_buf
[item
->ri_cnt
].i_len
= len
;
1598 trace_xfs_log_recover_item_add(log
, trans
, item
, 0);
1603 * Sort the log items in the transaction.
1605 * The ordering constraints are defined by the inode allocation and unlink
1606 * behaviour. The rules are:
1608 * 1. Every item is only logged once in a given transaction. Hence it
1609 * represents the last logged state of the item. Hence ordering is
1610 * dependent on the order in which operations need to be performed so
1611 * required initial conditions are always met.
1613 * 2. Cancelled buffers are recorded in pass 1 in a separate table and
1614 * there's nothing to replay from them so we can simply cull them
1615 * from the transaction. However, we can't do that until after we've
1616 * replayed all the other items because they may be dependent on the
1617 * cancelled buffer and replaying the cancelled buffer can remove it
1618 * form the cancelled buffer table. Hence they have tobe done last.
1620 * 3. Inode allocation buffers must be replayed before inode items that
1621 * read the buffer and replay changes into it. For filesystems using the
1622 * ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1623 * treated the same as inode allocation buffers as they create and
1624 * initialise the buffers directly.
1626 * 4. Inode unlink buffers must be replayed after inode items are replayed.
1627 * This ensures that inodes are completely flushed to the inode buffer
1628 * in a "free" state before we remove the unlinked inode list pointer.
1630 * Hence the ordering needs to be inode allocation buffers first, inode items
1631 * second, inode unlink buffers third and cancelled buffers last.
1633 * But there's a problem with that - we can't tell an inode allocation buffer
1634 * apart from a regular buffer, so we can't separate them. We can, however,
1635 * tell an inode unlink buffer from the others, and so we can separate them out
1636 * from all the other buffers and move them to last.
1638 * Hence, 4 lists, in order from head to tail:
1639 * - buffer_list for all buffers except cancelled/inode unlink buffers
1640 * - item_list for all non-buffer items
1641 * - inode_buffer_list for inode unlink buffers
1642 * - cancel_list for the cancelled buffers
1644 * Note that we add objects to the tail of the lists so that first-to-last
1645 * ordering is preserved within the lists. Adding objects to the head of the
1646 * list means when we traverse from the head we walk them in last-to-first
1647 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1648 * but for all other items there may be specific ordering that we need to
1652 xlog_recover_reorder_trans(
1654 struct xlog_recover
*trans
,
1657 xlog_recover_item_t
*item
, *n
;
1659 LIST_HEAD(sort_list
);
1660 LIST_HEAD(cancel_list
);
1661 LIST_HEAD(buffer_list
);
1662 LIST_HEAD(inode_buffer_list
);
1663 LIST_HEAD(inode_list
);
1665 list_splice_init(&trans
->r_itemq
, &sort_list
);
1666 list_for_each_entry_safe(item
, n
, &sort_list
, ri_list
) {
1667 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
1669 switch (ITEM_TYPE(item
)) {
1670 case XFS_LI_ICREATE
:
1671 list_move_tail(&item
->ri_list
, &buffer_list
);
1674 if (buf_f
->blf_flags
& XFS_BLF_CANCEL
) {
1675 trace_xfs_log_recover_item_reorder_head(log
,
1677 list_move(&item
->ri_list
, &cancel_list
);
1680 if (buf_f
->blf_flags
& XFS_BLF_INODE_BUF
) {
1681 list_move(&item
->ri_list
, &inode_buffer_list
);
1684 list_move_tail(&item
->ri_list
, &buffer_list
);
1688 case XFS_LI_QUOTAOFF
:
1691 trace_xfs_log_recover_item_reorder_tail(log
,
1693 list_move_tail(&item
->ri_list
, &inode_list
);
1697 "%s: unrecognized type of log operation",
1701 * return the remaining items back to the transaction
1702 * item list so they can be freed in caller.
1704 if (!list_empty(&sort_list
))
1705 list_splice_init(&sort_list
, &trans
->r_itemq
);
1711 ASSERT(list_empty(&sort_list
));
1712 if (!list_empty(&buffer_list
))
1713 list_splice(&buffer_list
, &trans
->r_itemq
);
1714 if (!list_empty(&inode_list
))
1715 list_splice_tail(&inode_list
, &trans
->r_itemq
);
1716 if (!list_empty(&inode_buffer_list
))
1717 list_splice_tail(&inode_buffer_list
, &trans
->r_itemq
);
1718 if (!list_empty(&cancel_list
))
1719 list_splice_tail(&cancel_list
, &trans
->r_itemq
);
1724 * Build up the table of buf cancel records so that we don't replay
1725 * cancelled data in the second pass. For buffer records that are
1726 * not cancel records, there is nothing to do here so we just return.
1728 * If we get a cancel record which is already in the table, this indicates
1729 * that the buffer was cancelled multiple times. In order to ensure
1730 * that during pass 2 we keep the record in the table until we reach its
1731 * last occurrence in the log, we keep a reference count in the cancel
1732 * record in the table to tell us how many times we expect to see this
1733 * record during the second pass.
1736 xlog_recover_buffer_pass1(
1738 struct xlog_recover_item
*item
)
1740 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
1741 struct list_head
*bucket
;
1742 struct xfs_buf_cancel
*bcp
;
1745 * If this isn't a cancel buffer item, then just return.
1747 if (!(buf_f
->blf_flags
& XFS_BLF_CANCEL
)) {
1748 trace_xfs_log_recover_buf_not_cancel(log
, buf_f
);
1753 * Insert an xfs_buf_cancel record into the hash table of them.
1754 * If there is already an identical record, bump its reference count.
1756 bucket
= XLOG_BUF_CANCEL_BUCKET(log
, buf_f
->blf_blkno
);
1757 list_for_each_entry(bcp
, bucket
, bc_list
) {
1758 if (bcp
->bc_blkno
== buf_f
->blf_blkno
&&
1759 bcp
->bc_len
== buf_f
->blf_len
) {
1761 trace_xfs_log_recover_buf_cancel_ref_inc(log
, buf_f
);
1766 bcp
= kmem_alloc(sizeof(struct xfs_buf_cancel
), KM_SLEEP
);
1767 bcp
->bc_blkno
= buf_f
->blf_blkno
;
1768 bcp
->bc_len
= buf_f
->blf_len
;
1769 bcp
->bc_refcount
= 1;
1770 list_add_tail(&bcp
->bc_list
, bucket
);
1772 trace_xfs_log_recover_buf_cancel_add(log
, buf_f
);
1777 * Check to see whether the buffer being recovered has a corresponding
1778 * entry in the buffer cancel record table. If it is, return the cancel
1779 * buffer structure to the caller.
1781 STATIC
struct xfs_buf_cancel
*
1782 xlog_peek_buffer_cancelled(
1788 struct list_head
*bucket
;
1789 struct xfs_buf_cancel
*bcp
;
1791 if (!log
->l_buf_cancel_table
) {
1792 /* empty table means no cancelled buffers in the log */
1793 ASSERT(!(flags
& XFS_BLF_CANCEL
));
1797 bucket
= XLOG_BUF_CANCEL_BUCKET(log
, blkno
);
1798 list_for_each_entry(bcp
, bucket
, bc_list
) {
1799 if (bcp
->bc_blkno
== blkno
&& bcp
->bc_len
== len
)
1804 * We didn't find a corresponding entry in the table, so return 0 so
1805 * that the buffer is NOT cancelled.
1807 ASSERT(!(flags
& XFS_BLF_CANCEL
));
1812 * If the buffer is being cancelled then return 1 so that it will be cancelled,
1813 * otherwise return 0. If the buffer is actually a buffer cancel item
1814 * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
1815 * table and remove it from the table if this is the last reference.
1817 * We remove the cancel record from the table when we encounter its last
1818 * occurrence in the log so that if the same buffer is re-used again after its
1819 * last cancellation we actually replay the changes made at that point.
1822 xlog_check_buffer_cancelled(
1828 struct xfs_buf_cancel
*bcp
;
1830 bcp
= xlog_peek_buffer_cancelled(log
, blkno
, len
, flags
);
1835 * We've go a match, so return 1 so that the recovery of this buffer
1836 * is cancelled. If this buffer is actually a buffer cancel log
1837 * item, then decrement the refcount on the one in the table and
1838 * remove it if this is the last reference.
1840 if (flags
& XFS_BLF_CANCEL
) {
1841 if (--bcp
->bc_refcount
== 0) {
1842 list_del(&bcp
->bc_list
);
1850 * Perform recovery for a buffer full of inodes. In these buffers, the only
1851 * data which should be recovered is that which corresponds to the
1852 * di_next_unlinked pointers in the on disk inode structures. The rest of the
1853 * data for the inodes is always logged through the inodes themselves rather
1854 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1856 * The only time when buffers full of inodes are fully recovered is when the
1857 * buffer is full of newly allocated inodes. In this case the buffer will
1858 * not be marked as an inode buffer and so will be sent to
1859 * xlog_recover_do_reg_buffer() below during recovery.
1862 xlog_recover_do_inode_buffer(
1863 struct xfs_mount
*mp
,
1864 xlog_recover_item_t
*item
,
1866 xfs_buf_log_format_t
*buf_f
)
1872 int reg_buf_offset
= 0;
1873 int reg_buf_bytes
= 0;
1874 int next_unlinked_offset
;
1876 xfs_agino_t
*logged_nextp
;
1877 xfs_agino_t
*buffer_nextp
;
1879 trace_xfs_log_recover_buf_inode_buf(mp
->m_log
, buf_f
);
1882 * Post recovery validation only works properly on CRC enabled
1885 if (xfs_sb_version_hascrc(&mp
->m_sb
))
1886 bp
->b_ops
= &xfs_inode_buf_ops
;
1888 inodes_per_buf
= BBTOB(bp
->b_io_length
) >> mp
->m_sb
.sb_inodelog
;
1889 for (i
= 0; i
< inodes_per_buf
; i
++) {
1890 next_unlinked_offset
= (i
* mp
->m_sb
.sb_inodesize
) +
1891 offsetof(xfs_dinode_t
, di_next_unlinked
);
1893 while (next_unlinked_offset
>=
1894 (reg_buf_offset
+ reg_buf_bytes
)) {
1896 * The next di_next_unlinked field is beyond
1897 * the current logged region. Find the next
1898 * logged region that contains or is beyond
1899 * the current di_next_unlinked field.
1902 bit
= xfs_next_bit(buf_f
->blf_data_map
,
1903 buf_f
->blf_map_size
, bit
);
1906 * If there are no more logged regions in the
1907 * buffer, then we're done.
1912 nbits
= xfs_contig_bits(buf_f
->blf_data_map
,
1913 buf_f
->blf_map_size
, bit
);
1915 reg_buf_offset
= bit
<< XFS_BLF_SHIFT
;
1916 reg_buf_bytes
= nbits
<< XFS_BLF_SHIFT
;
1921 * If the current logged region starts after the current
1922 * di_next_unlinked field, then move on to the next
1923 * di_next_unlinked field.
1925 if (next_unlinked_offset
< reg_buf_offset
)
1928 ASSERT(item
->ri_buf
[item_index
].i_addr
!= NULL
);
1929 ASSERT((item
->ri_buf
[item_index
].i_len
% XFS_BLF_CHUNK
) == 0);
1930 ASSERT((reg_buf_offset
+ reg_buf_bytes
) <=
1931 BBTOB(bp
->b_io_length
));
1934 * The current logged region contains a copy of the
1935 * current di_next_unlinked field. Extract its value
1936 * and copy it to the buffer copy.
1938 logged_nextp
= item
->ri_buf
[item_index
].i_addr
+
1939 next_unlinked_offset
- reg_buf_offset
;
1940 if (unlikely(*logged_nextp
== 0)) {
1942 "Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1943 "Trying to replay bad (0) inode di_next_unlinked field.",
1945 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1946 XFS_ERRLEVEL_LOW
, mp
);
1947 return -EFSCORRUPTED
;
1950 buffer_nextp
= (xfs_agino_t
*)xfs_buf_offset(bp
,
1951 next_unlinked_offset
);
1952 *buffer_nextp
= *logged_nextp
;
1955 * If necessary, recalculate the CRC in the on-disk inode. We
1956 * have to leave the inode in a consistent state for whoever
1959 xfs_dinode_calc_crc(mp
, (struct xfs_dinode
*)
1960 xfs_buf_offset(bp
, i
* mp
->m_sb
.sb_inodesize
));
1968 * V5 filesystems know the age of the buffer on disk being recovered. We can
1969 * have newer objects on disk than we are replaying, and so for these cases we
1970 * don't want to replay the current change as that will make the buffer contents
1971 * temporarily invalid on disk.
1973 * The magic number might not match the buffer type we are going to recover
1974 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags. Hence
1975 * extract the LSN of the existing object in the buffer based on it's current
1976 * magic number. If we don't recognise the magic number in the buffer, then
1977 * return a LSN of -1 so that the caller knows it was an unrecognised block and
1978 * so can recover the buffer.
1980 * Note: we cannot rely solely on magic number matches to determine that the
1981 * buffer has a valid LSN - we also need to verify that it belongs to this
1982 * filesystem, so we need to extract the object's LSN and compare it to that
1983 * which we read from the superblock. If the UUIDs don't match, then we've got a
1984 * stale metadata block from an old filesystem instance that we need to recover
1988 xlog_recover_get_buf_lsn(
1989 struct xfs_mount
*mp
,
1995 void *blk
= bp
->b_addr
;
1999 /* v4 filesystems always recover immediately */
2000 if (!xfs_sb_version_hascrc(&mp
->m_sb
))
2001 goto recover_immediately
;
2003 magic32
= be32_to_cpu(*(__be32
*)blk
);
2005 case XFS_ABTB_CRC_MAGIC
:
2006 case XFS_ABTC_CRC_MAGIC
:
2007 case XFS_ABTB_MAGIC
:
2008 case XFS_ABTC_MAGIC
:
2009 case XFS_IBT_CRC_MAGIC
:
2010 case XFS_IBT_MAGIC
: {
2011 struct xfs_btree_block
*btb
= blk
;
2013 lsn
= be64_to_cpu(btb
->bb_u
.s
.bb_lsn
);
2014 uuid
= &btb
->bb_u
.s
.bb_uuid
;
2017 case XFS_BMAP_CRC_MAGIC
:
2018 case XFS_BMAP_MAGIC
: {
2019 struct xfs_btree_block
*btb
= blk
;
2021 lsn
= be64_to_cpu(btb
->bb_u
.l
.bb_lsn
);
2022 uuid
= &btb
->bb_u
.l
.bb_uuid
;
2026 lsn
= be64_to_cpu(((struct xfs_agf
*)blk
)->agf_lsn
);
2027 uuid
= &((struct xfs_agf
*)blk
)->agf_uuid
;
2029 case XFS_AGFL_MAGIC
:
2030 lsn
= be64_to_cpu(((struct xfs_agfl
*)blk
)->agfl_lsn
);
2031 uuid
= &((struct xfs_agfl
*)blk
)->agfl_uuid
;
2034 lsn
= be64_to_cpu(((struct xfs_agi
*)blk
)->agi_lsn
);
2035 uuid
= &((struct xfs_agi
*)blk
)->agi_uuid
;
2037 case XFS_SYMLINK_MAGIC
:
2038 lsn
= be64_to_cpu(((struct xfs_dsymlink_hdr
*)blk
)->sl_lsn
);
2039 uuid
= &((struct xfs_dsymlink_hdr
*)blk
)->sl_uuid
;
2041 case XFS_DIR3_BLOCK_MAGIC
:
2042 case XFS_DIR3_DATA_MAGIC
:
2043 case XFS_DIR3_FREE_MAGIC
:
2044 lsn
= be64_to_cpu(((struct xfs_dir3_blk_hdr
*)blk
)->lsn
);
2045 uuid
= &((struct xfs_dir3_blk_hdr
*)blk
)->uuid
;
2047 case XFS_ATTR3_RMT_MAGIC
:
2048 lsn
= be64_to_cpu(((struct xfs_attr3_rmt_hdr
*)blk
)->rm_lsn
);
2049 uuid
= &((struct xfs_attr3_rmt_hdr
*)blk
)->rm_uuid
;
2052 lsn
= be64_to_cpu(((struct xfs_dsb
*)blk
)->sb_lsn
);
2053 uuid
= &((struct xfs_dsb
*)blk
)->sb_uuid
;
2059 if (lsn
!= (xfs_lsn_t
)-1) {
2060 if (!uuid_equal(&mp
->m_sb
.sb_uuid
, uuid
))
2061 goto recover_immediately
;
2065 magicda
= be16_to_cpu(((struct xfs_da_blkinfo
*)blk
)->magic
);
2067 case XFS_DIR3_LEAF1_MAGIC
:
2068 case XFS_DIR3_LEAFN_MAGIC
:
2069 case XFS_DA3_NODE_MAGIC
:
2070 lsn
= be64_to_cpu(((struct xfs_da3_blkinfo
*)blk
)->lsn
);
2071 uuid
= &((struct xfs_da3_blkinfo
*)blk
)->uuid
;
2077 if (lsn
!= (xfs_lsn_t
)-1) {
2078 if (!uuid_equal(&mp
->m_sb
.sb_uuid
, uuid
))
2079 goto recover_immediately
;
2084 * We do individual object checks on dquot and inode buffers as they
2085 * have their own individual LSN records. Also, we could have a stale
2086 * buffer here, so we have to at least recognise these buffer types.
2088 * A notd complexity here is inode unlinked list processing - it logs
2089 * the inode directly in the buffer, but we don't know which inodes have
2090 * been modified, and there is no global buffer LSN. Hence we need to
2091 * recover all inode buffer types immediately. This problem will be
2092 * fixed by logical logging of the unlinked list modifications.
2094 magic16
= be16_to_cpu(*(__be16
*)blk
);
2096 case XFS_DQUOT_MAGIC
:
2097 case XFS_DINODE_MAGIC
:
2098 goto recover_immediately
;
2103 /* unknown buffer contents, recover immediately */
2105 recover_immediately
:
2106 return (xfs_lsn_t
)-1;
2111 * Validate the recovered buffer is of the correct type and attach the
2112 * appropriate buffer operations to them for writeback. Magic numbers are in a
2114 * the first 16 bits of the buffer (inode buffer, dquot buffer),
2115 * the first 32 bits of the buffer (most blocks),
2116 * inside a struct xfs_da_blkinfo at the start of the buffer.
2119 xlog_recover_validate_buf_type(
2120 struct xfs_mount
*mp
,
2122 xfs_buf_log_format_t
*buf_f
)
2124 struct xfs_da_blkinfo
*info
= bp
->b_addr
;
2129 magic32
= be32_to_cpu(*(__be32
*)bp
->b_addr
);
2130 magic16
= be16_to_cpu(*(__be16
*)bp
->b_addr
);
2131 magicda
= be16_to_cpu(info
->magic
);
2132 switch (xfs_blft_from_flags(buf_f
)) {
2133 case XFS_BLFT_BTREE_BUF
:
2135 case XFS_ABTB_CRC_MAGIC
:
2136 case XFS_ABTC_CRC_MAGIC
:
2137 case XFS_ABTB_MAGIC
:
2138 case XFS_ABTC_MAGIC
:
2139 bp
->b_ops
= &xfs_allocbt_buf_ops
;
2141 case XFS_IBT_CRC_MAGIC
:
2142 case XFS_FIBT_CRC_MAGIC
:
2144 case XFS_FIBT_MAGIC
:
2145 bp
->b_ops
= &xfs_inobt_buf_ops
;
2147 case XFS_BMAP_CRC_MAGIC
:
2148 case XFS_BMAP_MAGIC
:
2149 bp
->b_ops
= &xfs_bmbt_buf_ops
;
2152 xfs_warn(mp
, "Bad btree block magic!");
2157 case XFS_BLFT_AGF_BUF
:
2158 if (magic32
!= XFS_AGF_MAGIC
) {
2159 xfs_warn(mp
, "Bad AGF block magic!");
2163 bp
->b_ops
= &xfs_agf_buf_ops
;
2165 case XFS_BLFT_AGFL_BUF
:
2166 if (!xfs_sb_version_hascrc(&mp
->m_sb
))
2168 if (magic32
!= XFS_AGFL_MAGIC
) {
2169 xfs_warn(mp
, "Bad AGFL block magic!");
2173 bp
->b_ops
= &xfs_agfl_buf_ops
;
2175 case XFS_BLFT_AGI_BUF
:
2176 if (magic32
!= XFS_AGI_MAGIC
) {
2177 xfs_warn(mp
, "Bad AGI block magic!");
2181 bp
->b_ops
= &xfs_agi_buf_ops
;
2183 case XFS_BLFT_UDQUOT_BUF
:
2184 case XFS_BLFT_PDQUOT_BUF
:
2185 case XFS_BLFT_GDQUOT_BUF
:
2186 #ifdef CONFIG_XFS_QUOTA
2187 if (magic16
!= XFS_DQUOT_MAGIC
) {
2188 xfs_warn(mp
, "Bad DQUOT block magic!");
2192 bp
->b_ops
= &xfs_dquot_buf_ops
;
2195 "Trying to recover dquots without QUOTA support built in!");
2199 case XFS_BLFT_DINO_BUF
:
2201 * we get here with inode allocation buffers, not buffers that
2202 * track unlinked list changes.
2204 if (magic16
!= XFS_DINODE_MAGIC
) {
2205 xfs_warn(mp
, "Bad INODE block magic!");
2209 bp
->b_ops
= &xfs_inode_buf_ops
;
2211 case XFS_BLFT_SYMLINK_BUF
:
2212 if (magic32
!= XFS_SYMLINK_MAGIC
) {
2213 xfs_warn(mp
, "Bad symlink block magic!");
2217 bp
->b_ops
= &xfs_symlink_buf_ops
;
2219 case XFS_BLFT_DIR_BLOCK_BUF
:
2220 if (magic32
!= XFS_DIR2_BLOCK_MAGIC
&&
2221 magic32
!= XFS_DIR3_BLOCK_MAGIC
) {
2222 xfs_warn(mp
, "Bad dir block magic!");
2226 bp
->b_ops
= &xfs_dir3_block_buf_ops
;
2228 case XFS_BLFT_DIR_DATA_BUF
:
2229 if (magic32
!= XFS_DIR2_DATA_MAGIC
&&
2230 magic32
!= XFS_DIR3_DATA_MAGIC
) {
2231 xfs_warn(mp
, "Bad dir data magic!");
2235 bp
->b_ops
= &xfs_dir3_data_buf_ops
;
2237 case XFS_BLFT_DIR_FREE_BUF
:
2238 if (magic32
!= XFS_DIR2_FREE_MAGIC
&&
2239 magic32
!= XFS_DIR3_FREE_MAGIC
) {
2240 xfs_warn(mp
, "Bad dir3 free magic!");
2244 bp
->b_ops
= &xfs_dir3_free_buf_ops
;
2246 case XFS_BLFT_DIR_LEAF1_BUF
:
2247 if (magicda
!= XFS_DIR2_LEAF1_MAGIC
&&
2248 magicda
!= XFS_DIR3_LEAF1_MAGIC
) {
2249 xfs_warn(mp
, "Bad dir leaf1 magic!");
2253 bp
->b_ops
= &xfs_dir3_leaf1_buf_ops
;
2255 case XFS_BLFT_DIR_LEAFN_BUF
:
2256 if (magicda
!= XFS_DIR2_LEAFN_MAGIC
&&
2257 magicda
!= XFS_DIR3_LEAFN_MAGIC
) {
2258 xfs_warn(mp
, "Bad dir leafn magic!");
2262 bp
->b_ops
= &xfs_dir3_leafn_buf_ops
;
2264 case XFS_BLFT_DA_NODE_BUF
:
2265 if (magicda
!= XFS_DA_NODE_MAGIC
&&
2266 magicda
!= XFS_DA3_NODE_MAGIC
) {
2267 xfs_warn(mp
, "Bad da node magic!");
2271 bp
->b_ops
= &xfs_da3_node_buf_ops
;
2273 case XFS_BLFT_ATTR_LEAF_BUF
:
2274 if (magicda
!= XFS_ATTR_LEAF_MAGIC
&&
2275 magicda
!= XFS_ATTR3_LEAF_MAGIC
) {
2276 xfs_warn(mp
, "Bad attr leaf magic!");
2280 bp
->b_ops
= &xfs_attr3_leaf_buf_ops
;
2282 case XFS_BLFT_ATTR_RMT_BUF
:
2283 if (!xfs_sb_version_hascrc(&mp
->m_sb
))
2285 if (magic32
!= XFS_ATTR3_RMT_MAGIC
) {
2286 xfs_warn(mp
, "Bad attr remote magic!");
2290 bp
->b_ops
= &xfs_attr3_rmt_buf_ops
;
2292 case XFS_BLFT_SB_BUF
:
2293 if (magic32
!= XFS_SB_MAGIC
) {
2294 xfs_warn(mp
, "Bad SB block magic!");
2298 bp
->b_ops
= &xfs_sb_buf_ops
;
2301 xfs_warn(mp
, "Unknown buffer type %d!",
2302 xfs_blft_from_flags(buf_f
));
2308 * Perform a 'normal' buffer recovery. Each logged region of the
2309 * buffer should be copied over the corresponding region in the
2310 * given buffer. The bitmap in the buf log format structure indicates
2311 * where to place the logged data.
2314 xlog_recover_do_reg_buffer(
2315 struct xfs_mount
*mp
,
2316 xlog_recover_item_t
*item
,
2318 xfs_buf_log_format_t
*buf_f
)
2325 trace_xfs_log_recover_buf_reg_buf(mp
->m_log
, buf_f
);
2328 i
= 1; /* 0 is the buf format structure */
2330 bit
= xfs_next_bit(buf_f
->blf_data_map
,
2331 buf_f
->blf_map_size
, bit
);
2334 nbits
= xfs_contig_bits(buf_f
->blf_data_map
,
2335 buf_f
->blf_map_size
, bit
);
2337 ASSERT(item
->ri_buf
[i
].i_addr
!= NULL
);
2338 ASSERT(item
->ri_buf
[i
].i_len
% XFS_BLF_CHUNK
== 0);
2339 ASSERT(BBTOB(bp
->b_io_length
) >=
2340 ((uint
)bit
<< XFS_BLF_SHIFT
) + (nbits
<< XFS_BLF_SHIFT
));
2343 * The dirty regions logged in the buffer, even though
2344 * contiguous, may span multiple chunks. This is because the
2345 * dirty region may span a physical page boundary in a buffer
2346 * and hence be split into two separate vectors for writing into
2347 * the log. Hence we need to trim nbits back to the length of
2348 * the current region being copied out of the log.
2350 if (item
->ri_buf
[i
].i_len
< (nbits
<< XFS_BLF_SHIFT
))
2351 nbits
= item
->ri_buf
[i
].i_len
>> XFS_BLF_SHIFT
;
2354 * Do a sanity check if this is a dquot buffer. Just checking
2355 * the first dquot in the buffer should do. XXXThis is
2356 * probably a good thing to do for other buf types also.
2359 if (buf_f
->blf_flags
&
2360 (XFS_BLF_UDQUOT_BUF
|XFS_BLF_PDQUOT_BUF
|XFS_BLF_GDQUOT_BUF
)) {
2361 if (item
->ri_buf
[i
].i_addr
== NULL
) {
2363 "XFS: NULL dquot in %s.", __func__
);
2366 if (item
->ri_buf
[i
].i_len
< sizeof(xfs_disk_dquot_t
)) {
2368 "XFS: dquot too small (%d) in %s.",
2369 item
->ri_buf
[i
].i_len
, __func__
);
2372 error
= xfs_dqcheck(mp
, item
->ri_buf
[i
].i_addr
,
2373 -1, 0, XFS_QMOPT_DOWARN
,
2374 "dquot_buf_recover");
2379 memcpy(xfs_buf_offset(bp
,
2380 (uint
)bit
<< XFS_BLF_SHIFT
), /* dest */
2381 item
->ri_buf
[i
].i_addr
, /* source */
2382 nbits
<<XFS_BLF_SHIFT
); /* length */
2388 /* Shouldn't be any more regions */
2389 ASSERT(i
== item
->ri_total
);
2392 * We can only do post recovery validation on items on CRC enabled
2393 * fielsystems as we need to know when the buffer was written to be able
2394 * to determine if we should have replayed the item. If we replay old
2395 * metadata over a newer buffer, then it will enter a temporarily
2396 * inconsistent state resulting in verification failures. Hence for now
2397 * just avoid the verification stage for non-crc filesystems
2399 if (xfs_sb_version_hascrc(&mp
->m_sb
))
2400 xlog_recover_validate_buf_type(mp
, bp
, buf_f
);
2404 * Perform a dquot buffer recovery.
2405 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2406 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2407 * Else, treat it as a regular buffer and do recovery.
2410 xlog_recover_do_dquot_buffer(
2411 struct xfs_mount
*mp
,
2413 struct xlog_recover_item
*item
,
2415 struct xfs_buf_log_format
*buf_f
)
2419 trace_xfs_log_recover_buf_dquot_buf(log
, buf_f
);
2422 * Filesystems are required to send in quota flags at mount time.
2424 if (mp
->m_qflags
== 0) {
2429 if (buf_f
->blf_flags
& XFS_BLF_UDQUOT_BUF
)
2430 type
|= XFS_DQ_USER
;
2431 if (buf_f
->blf_flags
& XFS_BLF_PDQUOT_BUF
)
2432 type
|= XFS_DQ_PROJ
;
2433 if (buf_f
->blf_flags
& XFS_BLF_GDQUOT_BUF
)
2434 type
|= XFS_DQ_GROUP
;
2436 * This type of quotas was turned off, so ignore this buffer
2438 if (log
->l_quotaoffs_flag
& type
)
2441 xlog_recover_do_reg_buffer(mp
, item
, bp
, buf_f
);
2445 * This routine replays a modification made to a buffer at runtime.
2446 * There are actually two types of buffer, regular and inode, which
2447 * are handled differently. Inode buffers are handled differently
2448 * in that we only recover a specific set of data from them, namely
2449 * the inode di_next_unlinked fields. This is because all other inode
2450 * data is actually logged via inode records and any data we replay
2451 * here which overlaps that may be stale.
2453 * When meta-data buffers are freed at run time we log a buffer item
2454 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2455 * of the buffer in the log should not be replayed at recovery time.
2456 * This is so that if the blocks covered by the buffer are reused for
2457 * file data before we crash we don't end up replaying old, freed
2458 * meta-data into a user's file.
2460 * To handle the cancellation of buffer log items, we make two passes
2461 * over the log during recovery. During the first we build a table of
2462 * those buffers which have been cancelled, and during the second we
2463 * only replay those buffers which do not have corresponding cancel
2464 * records in the table. See xlog_recover_buffer_pass[1,2] above
2465 * for more details on the implementation of the table of cancel records.
2468 xlog_recover_buffer_pass2(
2470 struct list_head
*buffer_list
,
2471 struct xlog_recover_item
*item
,
2472 xfs_lsn_t current_lsn
)
2474 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
2475 xfs_mount_t
*mp
= log
->l_mp
;
2482 * In this pass we only want to recover all the buffers which have
2483 * not been cancelled and are not cancellation buffers themselves.
2485 if (xlog_check_buffer_cancelled(log
, buf_f
->blf_blkno
,
2486 buf_f
->blf_len
, buf_f
->blf_flags
)) {
2487 trace_xfs_log_recover_buf_cancel(log
, buf_f
);
2491 trace_xfs_log_recover_buf_recover(log
, buf_f
);
2494 if (buf_f
->blf_flags
& XFS_BLF_INODE_BUF
)
2495 buf_flags
|= XBF_UNMAPPED
;
2497 bp
= xfs_buf_read(mp
->m_ddev_targp
, buf_f
->blf_blkno
, buf_f
->blf_len
,
2501 error
= bp
->b_error
;
2503 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#1)");
2508 * recover the buffer only if we get an LSN from it and it's less than
2509 * the lsn of the transaction we are replaying.
2511 lsn
= xlog_recover_get_buf_lsn(mp
, bp
);
2512 if (lsn
&& lsn
!= -1 && XFS_LSN_CMP(lsn
, current_lsn
) >= 0)
2515 if (buf_f
->blf_flags
& XFS_BLF_INODE_BUF
) {
2516 error
= xlog_recover_do_inode_buffer(mp
, item
, bp
, buf_f
);
2517 } else if (buf_f
->blf_flags
&
2518 (XFS_BLF_UDQUOT_BUF
|XFS_BLF_PDQUOT_BUF
|XFS_BLF_GDQUOT_BUF
)) {
2519 xlog_recover_do_dquot_buffer(mp
, log
, item
, bp
, buf_f
);
2521 xlog_recover_do_reg_buffer(mp
, item
, bp
, buf_f
);
2527 * Perform delayed write on the buffer. Asynchronous writes will be
2528 * slower when taking into account all the buffers to be flushed.
2530 * Also make sure that only inode buffers with good sizes stay in
2531 * the buffer cache. The kernel moves inodes in buffers of 1 block
2532 * or mp->m_inode_cluster_size bytes, whichever is bigger. The inode
2533 * buffers in the log can be a different size if the log was generated
2534 * by an older kernel using unclustered inode buffers or a newer kernel
2535 * running with a different inode cluster size. Regardless, if the
2536 * the inode buffer size isn't MAX(blocksize, mp->m_inode_cluster_size)
2537 * for *our* value of mp->m_inode_cluster_size, then we need to keep
2538 * the buffer out of the buffer cache so that the buffer won't
2539 * overlap with future reads of those inodes.
2541 if (XFS_DINODE_MAGIC
==
2542 be16_to_cpu(*((__be16
*)xfs_buf_offset(bp
, 0))) &&
2543 (BBTOB(bp
->b_io_length
) != MAX(log
->l_mp
->m_sb
.sb_blocksize
,
2544 (__uint32_t
)log
->l_mp
->m_inode_cluster_size
))) {
2546 error
= xfs_bwrite(bp
);
2548 ASSERT(bp
->b_target
->bt_mount
== mp
);
2549 bp
->b_iodone
= xlog_recover_iodone
;
2550 xfs_buf_delwri_queue(bp
, buffer_list
);
2559 * Inode fork owner changes
2561 * If we have been told that we have to reparent the inode fork, it's because an
2562 * extent swap operation on a CRC enabled filesystem has been done and we are
2563 * replaying it. We need to walk the BMBT of the appropriate fork and change the
2566 * The complexity here is that we don't have an inode context to work with, so
2567 * after we've replayed the inode we need to instantiate one. This is where the
2570 * We are in the middle of log recovery, so we can't run transactions. That
2571 * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2572 * that will result in the corresponding iput() running the inode through
2573 * xfs_inactive(). If we've just replayed an inode core that changes the link
2574 * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2575 * transactions (bad!).
2577 * So, to avoid this, we instantiate an inode directly from the inode core we've
2578 * just recovered. We have the buffer still locked, and all we really need to
2579 * instantiate is the inode core and the forks being modified. We can do this
2580 * manually, then run the inode btree owner change, and then tear down the
2581 * xfs_inode without having to run any transactions at all.
2583 * Also, because we don't have a transaction context available here but need to
2584 * gather all the buffers we modify for writeback so we pass the buffer_list
2585 * instead for the operation to use.
2589 xfs_recover_inode_owner_change(
2590 struct xfs_mount
*mp
,
2591 struct xfs_dinode
*dip
,
2592 struct xfs_inode_log_format
*in_f
,
2593 struct list_head
*buffer_list
)
2595 struct xfs_inode
*ip
;
2598 ASSERT(in_f
->ilf_fields
& (XFS_ILOG_DOWNER
|XFS_ILOG_AOWNER
));
2600 ip
= xfs_inode_alloc(mp
, in_f
->ilf_ino
);
2604 /* instantiate the inode */
2605 xfs_dinode_from_disk(&ip
->i_d
, dip
);
2606 ASSERT(ip
->i_d
.di_version
>= 3);
2608 error
= xfs_iformat_fork(ip
, dip
);
2613 if (in_f
->ilf_fields
& XFS_ILOG_DOWNER
) {
2614 ASSERT(in_f
->ilf_fields
& XFS_ILOG_DBROOT
);
2615 error
= xfs_bmbt_change_owner(NULL
, ip
, XFS_DATA_FORK
,
2616 ip
->i_ino
, buffer_list
);
2621 if (in_f
->ilf_fields
& XFS_ILOG_AOWNER
) {
2622 ASSERT(in_f
->ilf_fields
& XFS_ILOG_ABROOT
);
2623 error
= xfs_bmbt_change_owner(NULL
, ip
, XFS_ATTR_FORK
,
2624 ip
->i_ino
, buffer_list
);
2635 xlog_recover_inode_pass2(
2637 struct list_head
*buffer_list
,
2638 struct xlog_recover_item
*item
,
2639 xfs_lsn_t current_lsn
)
2641 xfs_inode_log_format_t
*in_f
;
2642 xfs_mount_t
*mp
= log
->l_mp
;
2651 xfs_icdinode_t
*dicp
;
2655 if (item
->ri_buf
[0].i_len
== sizeof(xfs_inode_log_format_t
)) {
2656 in_f
= item
->ri_buf
[0].i_addr
;
2658 in_f
= kmem_alloc(sizeof(xfs_inode_log_format_t
), KM_SLEEP
);
2660 error
= xfs_inode_item_format_convert(&item
->ri_buf
[0], in_f
);
2666 * Inode buffers can be freed, look out for it,
2667 * and do not replay the inode.
2669 if (xlog_check_buffer_cancelled(log
, in_f
->ilf_blkno
,
2670 in_f
->ilf_len
, 0)) {
2672 trace_xfs_log_recover_inode_cancel(log
, in_f
);
2675 trace_xfs_log_recover_inode_recover(log
, in_f
);
2677 bp
= xfs_buf_read(mp
->m_ddev_targp
, in_f
->ilf_blkno
, in_f
->ilf_len
, 0,
2678 &xfs_inode_buf_ops
);
2683 error
= bp
->b_error
;
2685 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#2)");
2688 ASSERT(in_f
->ilf_fields
& XFS_ILOG_CORE
);
2689 dip
= (xfs_dinode_t
*)xfs_buf_offset(bp
, in_f
->ilf_boffset
);
2692 * Make sure the place we're flushing out to really looks
2695 if (unlikely(dip
->di_magic
!= cpu_to_be16(XFS_DINODE_MAGIC
))) {
2697 "%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2698 __func__
, dip
, bp
, in_f
->ilf_ino
);
2699 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2700 XFS_ERRLEVEL_LOW
, mp
);
2701 error
= -EFSCORRUPTED
;
2704 dicp
= item
->ri_buf
[1].i_addr
;
2705 if (unlikely(dicp
->di_magic
!= XFS_DINODE_MAGIC
)) {
2707 "%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2708 __func__
, item
, in_f
->ilf_ino
);
2709 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2710 XFS_ERRLEVEL_LOW
, mp
);
2711 error
= -EFSCORRUPTED
;
2716 * If the inode has an LSN in it, recover the inode only if it's less
2717 * than the lsn of the transaction we are replaying. Note: we still
2718 * need to replay an owner change even though the inode is more recent
2719 * than the transaction as there is no guarantee that all the btree
2720 * blocks are more recent than this transaction, too.
2722 if (dip
->di_version
>= 3) {
2723 xfs_lsn_t lsn
= be64_to_cpu(dip
->di_lsn
);
2725 if (lsn
&& lsn
!= -1 && XFS_LSN_CMP(lsn
, current_lsn
) >= 0) {
2726 trace_xfs_log_recover_inode_skip(log
, in_f
);
2728 goto out_owner_change
;
2733 * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
2734 * are transactional and if ordering is necessary we can determine that
2735 * more accurately by the LSN field in the V3 inode core. Don't trust
2736 * the inode versions we might be changing them here - use the
2737 * superblock flag to determine whether we need to look at di_flushiter
2738 * to skip replay when the on disk inode is newer than the log one
2740 if (!xfs_sb_version_hascrc(&mp
->m_sb
) &&
2741 dicp
->di_flushiter
< be16_to_cpu(dip
->di_flushiter
)) {
2743 * Deal with the wrap case, DI_MAX_FLUSH is less
2744 * than smaller numbers
2746 if (be16_to_cpu(dip
->di_flushiter
) == DI_MAX_FLUSH
&&
2747 dicp
->di_flushiter
< (DI_MAX_FLUSH
>> 1)) {
2750 trace_xfs_log_recover_inode_skip(log
, in_f
);
2756 /* Take the opportunity to reset the flush iteration count */
2757 dicp
->di_flushiter
= 0;
2759 if (unlikely(S_ISREG(dicp
->di_mode
))) {
2760 if ((dicp
->di_format
!= XFS_DINODE_FMT_EXTENTS
) &&
2761 (dicp
->di_format
!= XFS_DINODE_FMT_BTREE
)) {
2762 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2763 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2765 "%s: Bad regular inode log record, rec ptr 0x%p, "
2766 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2767 __func__
, item
, dip
, bp
, in_f
->ilf_ino
);
2768 error
= -EFSCORRUPTED
;
2771 } else if (unlikely(S_ISDIR(dicp
->di_mode
))) {
2772 if ((dicp
->di_format
!= XFS_DINODE_FMT_EXTENTS
) &&
2773 (dicp
->di_format
!= XFS_DINODE_FMT_BTREE
) &&
2774 (dicp
->di_format
!= XFS_DINODE_FMT_LOCAL
)) {
2775 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2776 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2778 "%s: Bad dir inode log record, rec ptr 0x%p, "
2779 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2780 __func__
, item
, dip
, bp
, in_f
->ilf_ino
);
2781 error
= -EFSCORRUPTED
;
2785 if (unlikely(dicp
->di_nextents
+ dicp
->di_anextents
> dicp
->di_nblocks
)){
2786 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2787 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2789 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2790 "dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2791 __func__
, item
, dip
, bp
, in_f
->ilf_ino
,
2792 dicp
->di_nextents
+ dicp
->di_anextents
,
2794 error
= -EFSCORRUPTED
;
2797 if (unlikely(dicp
->di_forkoff
> mp
->m_sb
.sb_inodesize
)) {
2798 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2799 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2801 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2802 "dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__
,
2803 item
, dip
, bp
, in_f
->ilf_ino
, dicp
->di_forkoff
);
2804 error
= -EFSCORRUPTED
;
2807 isize
= xfs_icdinode_size(dicp
->di_version
);
2808 if (unlikely(item
->ri_buf
[1].i_len
> isize
)) {
2809 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2810 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2812 "%s: Bad inode log record length %d, rec ptr 0x%p",
2813 __func__
, item
->ri_buf
[1].i_len
, item
);
2814 error
= -EFSCORRUPTED
;
2818 /* The core is in in-core format */
2819 xfs_dinode_to_disk(dip
, dicp
);
2821 /* the rest is in on-disk format */
2822 if (item
->ri_buf
[1].i_len
> isize
) {
2823 memcpy((char *)dip
+ isize
,
2824 item
->ri_buf
[1].i_addr
+ isize
,
2825 item
->ri_buf
[1].i_len
- isize
);
2828 fields
= in_f
->ilf_fields
;
2829 switch (fields
& (XFS_ILOG_DEV
| XFS_ILOG_UUID
)) {
2831 xfs_dinode_put_rdev(dip
, in_f
->ilf_u
.ilfu_rdev
);
2834 memcpy(XFS_DFORK_DPTR(dip
),
2835 &in_f
->ilf_u
.ilfu_uuid
,
2840 if (in_f
->ilf_size
== 2)
2841 goto out_owner_change
;
2842 len
= item
->ri_buf
[2].i_len
;
2843 src
= item
->ri_buf
[2].i_addr
;
2844 ASSERT(in_f
->ilf_size
<= 4);
2845 ASSERT((in_f
->ilf_size
== 3) || (fields
& XFS_ILOG_AFORK
));
2846 ASSERT(!(fields
& XFS_ILOG_DFORK
) ||
2847 (len
== in_f
->ilf_dsize
));
2849 switch (fields
& XFS_ILOG_DFORK
) {
2850 case XFS_ILOG_DDATA
:
2852 memcpy(XFS_DFORK_DPTR(dip
), src
, len
);
2855 case XFS_ILOG_DBROOT
:
2856 xfs_bmbt_to_bmdr(mp
, (struct xfs_btree_block
*)src
, len
,
2857 (xfs_bmdr_block_t
*)XFS_DFORK_DPTR(dip
),
2858 XFS_DFORK_DSIZE(dip
, mp
));
2863 * There are no data fork flags set.
2865 ASSERT((fields
& XFS_ILOG_DFORK
) == 0);
2870 * If we logged any attribute data, recover it. There may or
2871 * may not have been any other non-core data logged in this
2874 if (in_f
->ilf_fields
& XFS_ILOG_AFORK
) {
2875 if (in_f
->ilf_fields
& XFS_ILOG_DFORK
) {
2880 len
= item
->ri_buf
[attr_index
].i_len
;
2881 src
= item
->ri_buf
[attr_index
].i_addr
;
2882 ASSERT(len
== in_f
->ilf_asize
);
2884 switch (in_f
->ilf_fields
& XFS_ILOG_AFORK
) {
2885 case XFS_ILOG_ADATA
:
2887 dest
= XFS_DFORK_APTR(dip
);
2888 ASSERT(len
<= XFS_DFORK_ASIZE(dip
, mp
));
2889 memcpy(dest
, src
, len
);
2892 case XFS_ILOG_ABROOT
:
2893 dest
= XFS_DFORK_APTR(dip
);
2894 xfs_bmbt_to_bmdr(mp
, (struct xfs_btree_block
*)src
,
2895 len
, (xfs_bmdr_block_t
*)dest
,
2896 XFS_DFORK_ASIZE(dip
, mp
));
2900 xfs_warn(log
->l_mp
, "%s: Invalid flag", __func__
);
2908 if (in_f
->ilf_fields
& (XFS_ILOG_DOWNER
|XFS_ILOG_AOWNER
))
2909 error
= xfs_recover_inode_owner_change(mp
, dip
, in_f
,
2911 /* re-generate the checksum. */
2912 xfs_dinode_calc_crc(log
->l_mp
, dip
);
2914 ASSERT(bp
->b_target
->bt_mount
== mp
);
2915 bp
->b_iodone
= xlog_recover_iodone
;
2916 xfs_buf_delwri_queue(bp
, buffer_list
);
2927 * Recover QUOTAOFF records. We simply make a note of it in the xlog
2928 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2932 xlog_recover_quotaoff_pass1(
2934 struct xlog_recover_item
*item
)
2936 xfs_qoff_logformat_t
*qoff_f
= item
->ri_buf
[0].i_addr
;
2940 * The logitem format's flag tells us if this was user quotaoff,
2941 * group/project quotaoff or both.
2943 if (qoff_f
->qf_flags
& XFS_UQUOTA_ACCT
)
2944 log
->l_quotaoffs_flag
|= XFS_DQ_USER
;
2945 if (qoff_f
->qf_flags
& XFS_PQUOTA_ACCT
)
2946 log
->l_quotaoffs_flag
|= XFS_DQ_PROJ
;
2947 if (qoff_f
->qf_flags
& XFS_GQUOTA_ACCT
)
2948 log
->l_quotaoffs_flag
|= XFS_DQ_GROUP
;
2954 * Recover a dquot record
2957 xlog_recover_dquot_pass2(
2959 struct list_head
*buffer_list
,
2960 struct xlog_recover_item
*item
,
2961 xfs_lsn_t current_lsn
)
2963 xfs_mount_t
*mp
= log
->l_mp
;
2965 struct xfs_disk_dquot
*ddq
, *recddq
;
2967 xfs_dq_logformat_t
*dq_f
;
2972 * Filesystems are required to send in quota flags at mount time.
2974 if (mp
->m_qflags
== 0)
2977 recddq
= item
->ri_buf
[1].i_addr
;
2978 if (recddq
== NULL
) {
2979 xfs_alert(log
->l_mp
, "NULL dquot in %s.", __func__
);
2982 if (item
->ri_buf
[1].i_len
< sizeof(xfs_disk_dquot_t
)) {
2983 xfs_alert(log
->l_mp
, "dquot too small (%d) in %s.",
2984 item
->ri_buf
[1].i_len
, __func__
);
2989 * This type of quotas was turned off, so ignore this record.
2991 type
= recddq
->d_flags
& (XFS_DQ_USER
| XFS_DQ_PROJ
| XFS_DQ_GROUP
);
2993 if (log
->l_quotaoffs_flag
& type
)
2997 * At this point we know that quota was _not_ turned off.
2998 * Since the mount flags are not indicating to us otherwise, this
2999 * must mean that quota is on, and the dquot needs to be replayed.
3000 * Remember that we may not have fully recovered the superblock yet,
3001 * so we can't do the usual trick of looking at the SB quota bits.
3003 * The other possibility, of course, is that the quota subsystem was
3004 * removed since the last mount - ENOSYS.
3006 dq_f
= item
->ri_buf
[0].i_addr
;
3008 error
= xfs_dqcheck(mp
, recddq
, dq_f
->qlf_id
, 0, XFS_QMOPT_DOWARN
,
3009 "xlog_recover_dquot_pass2 (log copy)");
3012 ASSERT(dq_f
->qlf_len
== 1);
3014 error
= xfs_trans_read_buf(mp
, NULL
, mp
->m_ddev_targp
, dq_f
->qlf_blkno
,
3015 XFS_FSB_TO_BB(mp
, dq_f
->qlf_len
), 0, &bp
,
3021 ddq
= (xfs_disk_dquot_t
*)xfs_buf_offset(bp
, dq_f
->qlf_boffset
);
3024 * At least the magic num portion should be on disk because this
3025 * was among a chunk of dquots created earlier, and we did some
3026 * minimal initialization then.
3028 error
= xfs_dqcheck(mp
, ddq
, dq_f
->qlf_id
, 0, XFS_QMOPT_DOWARN
,
3029 "xlog_recover_dquot_pass2");
3036 * If the dquot has an LSN in it, recover the dquot only if it's less
3037 * than the lsn of the transaction we are replaying.
3039 if (xfs_sb_version_hascrc(&mp
->m_sb
)) {
3040 struct xfs_dqblk
*dqb
= (struct xfs_dqblk
*)ddq
;
3041 xfs_lsn_t lsn
= be64_to_cpu(dqb
->dd_lsn
);
3043 if (lsn
&& lsn
!= -1 && XFS_LSN_CMP(lsn
, current_lsn
) >= 0) {
3048 memcpy(ddq
, recddq
, item
->ri_buf
[1].i_len
);
3049 if (xfs_sb_version_hascrc(&mp
->m_sb
)) {
3050 xfs_update_cksum((char *)ddq
, sizeof(struct xfs_dqblk
),
3054 ASSERT(dq_f
->qlf_size
== 2);
3055 ASSERT(bp
->b_target
->bt_mount
== mp
);
3056 bp
->b_iodone
= xlog_recover_iodone
;
3057 xfs_buf_delwri_queue(bp
, buffer_list
);
3065 * This routine is called to create an in-core extent free intent
3066 * item from the efi format structure which was logged on disk.
3067 * It allocates an in-core efi, copies the extents from the format
3068 * structure into it, and adds the efi to the AIL with the given
3072 xlog_recover_efi_pass2(
3074 struct xlog_recover_item
*item
,
3078 xfs_mount_t
*mp
= log
->l_mp
;
3079 xfs_efi_log_item_t
*efip
;
3080 xfs_efi_log_format_t
*efi_formatp
;
3082 efi_formatp
= item
->ri_buf
[0].i_addr
;
3084 efip
= xfs_efi_init(mp
, efi_formatp
->efi_nextents
);
3085 if ((error
= xfs_efi_copy_format(&(item
->ri_buf
[0]),
3086 &(efip
->efi_format
)))) {
3087 xfs_efi_item_free(efip
);
3090 atomic_set(&efip
->efi_next_extent
, efi_formatp
->efi_nextents
);
3092 spin_lock(&log
->l_ailp
->xa_lock
);
3094 * xfs_trans_ail_update() drops the AIL lock.
3096 xfs_trans_ail_update(log
->l_ailp
, &efip
->efi_item
, lsn
);
3102 * This routine is called when an efd format structure is found in
3103 * a committed transaction in the log. It's purpose is to cancel
3104 * the corresponding efi if it was still in the log. To do this
3105 * it searches the AIL for the efi with an id equal to that in the
3106 * efd format structure. If we find it, we remove the efi from the
3110 xlog_recover_efd_pass2(
3112 struct xlog_recover_item
*item
)
3114 xfs_efd_log_format_t
*efd_formatp
;
3115 xfs_efi_log_item_t
*efip
= NULL
;
3116 xfs_log_item_t
*lip
;
3118 struct xfs_ail_cursor cur
;
3119 struct xfs_ail
*ailp
= log
->l_ailp
;
3121 efd_formatp
= item
->ri_buf
[0].i_addr
;
3122 ASSERT((item
->ri_buf
[0].i_len
== (sizeof(xfs_efd_log_format_32_t
) +
3123 ((efd_formatp
->efd_nextents
- 1) * sizeof(xfs_extent_32_t
)))) ||
3124 (item
->ri_buf
[0].i_len
== (sizeof(xfs_efd_log_format_64_t
) +
3125 ((efd_formatp
->efd_nextents
- 1) * sizeof(xfs_extent_64_t
)))));
3126 efi_id
= efd_formatp
->efd_efi_id
;
3129 * Search for the efi with the id in the efd format structure
3132 spin_lock(&ailp
->xa_lock
);
3133 lip
= xfs_trans_ail_cursor_first(ailp
, &cur
, 0);
3134 while (lip
!= NULL
) {
3135 if (lip
->li_type
== XFS_LI_EFI
) {
3136 efip
= (xfs_efi_log_item_t
*)lip
;
3137 if (efip
->efi_format
.efi_id
== efi_id
) {
3139 * xfs_trans_ail_delete() drops the
3142 xfs_trans_ail_delete(ailp
, lip
,
3143 SHUTDOWN_CORRUPT_INCORE
);
3144 xfs_efi_item_free(efip
);
3145 spin_lock(&ailp
->xa_lock
);
3149 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3151 xfs_trans_ail_cursor_done(&cur
);
3152 spin_unlock(&ailp
->xa_lock
);
3158 * This routine is called when an inode create format structure is found in a
3159 * committed transaction in the log. It's purpose is to initialise the inodes
3160 * being allocated on disk. This requires us to get inode cluster buffers that
3161 * match the range to be intialised, stamped with inode templates and written
3162 * by delayed write so that subsequent modifications will hit the cached buffer
3163 * and only need writing out at the end of recovery.
3166 xlog_recover_do_icreate_pass2(
3168 struct list_head
*buffer_list
,
3169 xlog_recover_item_t
*item
)
3171 struct xfs_mount
*mp
= log
->l_mp
;
3172 struct xfs_icreate_log
*icl
;
3173 xfs_agnumber_t agno
;
3174 xfs_agblock_t agbno
;
3177 xfs_agblock_t length
;
3179 icl
= (struct xfs_icreate_log
*)item
->ri_buf
[0].i_addr
;
3180 if (icl
->icl_type
!= XFS_LI_ICREATE
) {
3181 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad type");
3185 if (icl
->icl_size
!= 1) {
3186 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad icl size");
3190 agno
= be32_to_cpu(icl
->icl_ag
);
3191 if (agno
>= mp
->m_sb
.sb_agcount
) {
3192 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad agno");
3195 agbno
= be32_to_cpu(icl
->icl_agbno
);
3196 if (!agbno
|| agbno
== NULLAGBLOCK
|| agbno
>= mp
->m_sb
.sb_agblocks
) {
3197 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad agbno");
3200 isize
= be32_to_cpu(icl
->icl_isize
);
3201 if (isize
!= mp
->m_sb
.sb_inodesize
) {
3202 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad isize");
3205 count
= be32_to_cpu(icl
->icl_count
);
3207 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad count");
3210 length
= be32_to_cpu(icl
->icl_length
);
3211 if (!length
|| length
>= mp
->m_sb
.sb_agblocks
) {
3212 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad length");
3216 /* existing allocation is fixed value */
3217 ASSERT(count
== mp
->m_ialloc_inos
);
3218 ASSERT(length
== mp
->m_ialloc_blks
);
3219 if (count
!= mp
->m_ialloc_inos
||
3220 length
!= mp
->m_ialloc_blks
) {
3221 xfs_warn(log
->l_mp
, "xlog_recover_do_icreate_trans: bad count 2");
3226 * Inode buffers can be freed. Do not replay the inode initialisation as
3227 * we could be overwriting something written after this inode buffer was
3230 * XXX: we need to iterate all buffers and only init those that are not
3231 * cancelled. I think that a more fine grained factoring of
3232 * xfs_ialloc_inode_init may be appropriate here to enable this to be
3235 if (xlog_check_buffer_cancelled(log
,
3236 XFS_AGB_TO_DADDR(mp
, agno
, agbno
), length
, 0))
3239 xfs_ialloc_inode_init(mp
, NULL
, buffer_list
, agno
, agbno
, length
,
3240 be32_to_cpu(icl
->icl_gen
));
3245 * Free up any resources allocated by the transaction
3247 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
3250 xlog_recover_free_trans(
3251 struct xlog_recover
*trans
)
3253 xlog_recover_item_t
*item
, *n
;
3256 list_for_each_entry_safe(item
, n
, &trans
->r_itemq
, ri_list
) {
3257 /* Free the regions in the item. */
3258 list_del(&item
->ri_list
);
3259 for (i
= 0; i
< item
->ri_cnt
; i
++)
3260 kmem_free(item
->ri_buf
[i
].i_addr
);
3261 /* Free the item itself */
3262 kmem_free(item
->ri_buf
);
3265 /* Free the transaction recover structure */
3270 xlog_recover_buffer_ra_pass2(
3272 struct xlog_recover_item
*item
)
3274 struct xfs_buf_log_format
*buf_f
= item
->ri_buf
[0].i_addr
;
3275 struct xfs_mount
*mp
= log
->l_mp
;
3277 if (xlog_peek_buffer_cancelled(log
, buf_f
->blf_blkno
,
3278 buf_f
->blf_len
, buf_f
->blf_flags
)) {
3282 xfs_buf_readahead(mp
->m_ddev_targp
, buf_f
->blf_blkno
,
3283 buf_f
->blf_len
, NULL
);
3287 xlog_recover_inode_ra_pass2(
3289 struct xlog_recover_item
*item
)
3291 struct xfs_inode_log_format ilf_buf
;
3292 struct xfs_inode_log_format
*ilfp
;
3293 struct xfs_mount
*mp
= log
->l_mp
;
3296 if (item
->ri_buf
[0].i_len
== sizeof(struct xfs_inode_log_format
)) {
3297 ilfp
= item
->ri_buf
[0].i_addr
;
3300 memset(ilfp
, 0, sizeof(*ilfp
));
3301 error
= xfs_inode_item_format_convert(&item
->ri_buf
[0], ilfp
);
3306 if (xlog_peek_buffer_cancelled(log
, ilfp
->ilf_blkno
, ilfp
->ilf_len
, 0))
3309 xfs_buf_readahead(mp
->m_ddev_targp
, ilfp
->ilf_blkno
,
3310 ilfp
->ilf_len
, &xfs_inode_buf_ra_ops
);
3314 xlog_recover_dquot_ra_pass2(
3316 struct xlog_recover_item
*item
)
3318 struct xfs_mount
*mp
= log
->l_mp
;
3319 struct xfs_disk_dquot
*recddq
;
3320 struct xfs_dq_logformat
*dq_f
;
3324 if (mp
->m_qflags
== 0)
3327 recddq
= item
->ri_buf
[1].i_addr
;
3330 if (item
->ri_buf
[1].i_len
< sizeof(struct xfs_disk_dquot
))
3333 type
= recddq
->d_flags
& (XFS_DQ_USER
| XFS_DQ_PROJ
| XFS_DQ_GROUP
);
3335 if (log
->l_quotaoffs_flag
& type
)
3338 dq_f
= item
->ri_buf
[0].i_addr
;
3340 ASSERT(dq_f
->qlf_len
== 1);
3342 xfs_buf_readahead(mp
->m_ddev_targp
, dq_f
->qlf_blkno
,
3343 XFS_FSB_TO_BB(mp
, dq_f
->qlf_len
), NULL
);
3347 xlog_recover_ra_pass2(
3349 struct xlog_recover_item
*item
)
3351 switch (ITEM_TYPE(item
)) {
3353 xlog_recover_buffer_ra_pass2(log
, item
);
3356 xlog_recover_inode_ra_pass2(log
, item
);
3359 xlog_recover_dquot_ra_pass2(log
, item
);
3363 case XFS_LI_QUOTAOFF
:
3370 xlog_recover_commit_pass1(
3372 struct xlog_recover
*trans
,
3373 struct xlog_recover_item
*item
)
3375 trace_xfs_log_recover_item_recover(log
, trans
, item
, XLOG_RECOVER_PASS1
);
3377 switch (ITEM_TYPE(item
)) {
3379 return xlog_recover_buffer_pass1(log
, item
);
3380 case XFS_LI_QUOTAOFF
:
3381 return xlog_recover_quotaoff_pass1(log
, item
);
3386 case XFS_LI_ICREATE
:
3387 /* nothing to do in pass 1 */
3390 xfs_warn(log
->l_mp
, "%s: invalid item type (%d)",
3391 __func__
, ITEM_TYPE(item
));
3398 xlog_recover_commit_pass2(
3400 struct xlog_recover
*trans
,
3401 struct list_head
*buffer_list
,
3402 struct xlog_recover_item
*item
)
3404 trace_xfs_log_recover_item_recover(log
, trans
, item
, XLOG_RECOVER_PASS2
);
3406 switch (ITEM_TYPE(item
)) {
3408 return xlog_recover_buffer_pass2(log
, buffer_list
, item
,
3411 return xlog_recover_inode_pass2(log
, buffer_list
, item
,
3414 return xlog_recover_efi_pass2(log
, item
, trans
->r_lsn
);
3416 return xlog_recover_efd_pass2(log
, item
);
3418 return xlog_recover_dquot_pass2(log
, buffer_list
, item
,
3420 case XFS_LI_ICREATE
:
3421 return xlog_recover_do_icreate_pass2(log
, buffer_list
, item
);
3422 case XFS_LI_QUOTAOFF
:
3423 /* nothing to do in pass2 */
3426 xfs_warn(log
->l_mp
, "%s: invalid item type (%d)",
3427 __func__
, ITEM_TYPE(item
));
3434 xlog_recover_items_pass2(
3436 struct xlog_recover
*trans
,
3437 struct list_head
*buffer_list
,
3438 struct list_head
*item_list
)
3440 struct xlog_recover_item
*item
;
3443 list_for_each_entry(item
, item_list
, ri_list
) {
3444 error
= xlog_recover_commit_pass2(log
, trans
,
3454 * Perform the transaction.
3456 * If the transaction modifies a buffer or inode, do it now. Otherwise,
3457 * EFIs and EFDs get queued up by adding entries into the AIL for them.
3460 xlog_recover_commit_trans(
3462 struct xlog_recover
*trans
,
3467 int items_queued
= 0;
3468 struct xlog_recover_item
*item
;
3469 struct xlog_recover_item
*next
;
3470 LIST_HEAD (buffer_list
);
3471 LIST_HEAD (ra_list
);
3472 LIST_HEAD (done_list
);
3474 #define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
3476 hlist_del(&trans
->r_list
);
3478 error
= xlog_recover_reorder_trans(log
, trans
, pass
);
3482 list_for_each_entry_safe(item
, next
, &trans
->r_itemq
, ri_list
) {
3484 case XLOG_RECOVER_PASS1
:
3485 error
= xlog_recover_commit_pass1(log
, trans
, item
);
3487 case XLOG_RECOVER_PASS2
:
3488 xlog_recover_ra_pass2(log
, item
);
3489 list_move_tail(&item
->ri_list
, &ra_list
);
3491 if (items_queued
>= XLOG_RECOVER_COMMIT_QUEUE_MAX
) {
3492 error
= xlog_recover_items_pass2(log
, trans
,
3493 &buffer_list
, &ra_list
);
3494 list_splice_tail_init(&ra_list
, &done_list
);
3508 if (!list_empty(&ra_list
)) {
3510 error
= xlog_recover_items_pass2(log
, trans
,
3511 &buffer_list
, &ra_list
);
3512 list_splice_tail_init(&ra_list
, &done_list
);
3515 if (!list_empty(&done_list
))
3516 list_splice_init(&done_list
, &trans
->r_itemq
);
3518 xlog_recover_free_trans(trans
);
3520 error2
= xfs_buf_delwri_submit(&buffer_list
);
3521 return error
? error
: error2
;
3525 xlog_recover_unmount_trans(
3528 /* Do nothing now */
3529 xfs_warn(log
->l_mp
, "%s: Unmount LR", __func__
);
3534 * There are two valid states of the r_state field. 0 indicates that the
3535 * transaction structure is in a normal state. We have either seen the
3536 * start of the transaction or the last operation we added was not a partial
3537 * operation. If the last operation we added to the transaction was a
3538 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
3540 * NOTE: skip LRs with 0 data length.
3543 xlog_recover_process_data(
3545 struct hlist_head rhash
[],
3546 struct xlog_rec_header
*rhead
,
3552 xlog_op_header_t
*ohead
;
3553 xlog_recover_t
*trans
;
3559 lp
= dp
+ be32_to_cpu(rhead
->h_len
);
3560 num_logops
= be32_to_cpu(rhead
->h_num_logops
);
3562 /* check the log format matches our own - else we can't recover */
3563 if (xlog_header_check_recover(log
->l_mp
, rhead
))
3566 while ((dp
< lp
) && num_logops
) {
3567 ASSERT(dp
+ sizeof(xlog_op_header_t
) <= lp
);
3568 ohead
= (xlog_op_header_t
*)dp
;
3569 dp
+= sizeof(xlog_op_header_t
);
3570 if (ohead
->oh_clientid
!= XFS_TRANSACTION
&&
3571 ohead
->oh_clientid
!= XFS_LOG
) {
3572 xfs_warn(log
->l_mp
, "%s: bad clientid 0x%x",
3573 __func__
, ohead
->oh_clientid
);
3577 tid
= be32_to_cpu(ohead
->oh_tid
);
3578 hash
= XLOG_RHASH(tid
);
3579 trans
= xlog_recover_find_tid(&rhash
[hash
], tid
);
3580 if (trans
== NULL
) { /* not found; add new tid */
3581 if (ohead
->oh_flags
& XLOG_START_TRANS
)
3582 xlog_recover_new_tid(&rhash
[hash
], tid
,
3583 be64_to_cpu(rhead
->h_lsn
));
3585 if (dp
+ be32_to_cpu(ohead
->oh_len
) > lp
) {
3586 xfs_warn(log
->l_mp
, "%s: bad length 0x%x",
3587 __func__
, be32_to_cpu(ohead
->oh_len
));
3591 flags
= ohead
->oh_flags
& ~XLOG_END_TRANS
;
3592 if (flags
& XLOG_WAS_CONT_TRANS
)
3593 flags
&= ~XLOG_CONTINUE_TRANS
;
3595 case XLOG_COMMIT_TRANS
:
3596 error
= xlog_recover_commit_trans(log
,
3599 case XLOG_UNMOUNT_TRANS
:
3600 error
= xlog_recover_unmount_trans(log
);
3602 case XLOG_WAS_CONT_TRANS
:
3603 error
= xlog_recover_add_to_cont_trans(log
,
3605 be32_to_cpu(ohead
->oh_len
));
3607 case XLOG_START_TRANS
:
3608 xfs_warn(log
->l_mp
, "%s: bad transaction",
3614 case XLOG_CONTINUE_TRANS
:
3615 error
= xlog_recover_add_to_trans(log
, trans
,
3616 dp
, be32_to_cpu(ohead
->oh_len
));
3619 xfs_warn(log
->l_mp
, "%s: bad flag 0x%x",
3626 xlog_recover_free_trans(trans
);
3630 dp
+= be32_to_cpu(ohead
->oh_len
);
3637 * Process an extent free intent item that was recovered from
3638 * the log. We need to free the extents that it describes.
3641 xlog_recover_process_efi(
3643 xfs_efi_log_item_t
*efip
)
3645 xfs_efd_log_item_t
*efdp
;
3650 xfs_fsblock_t startblock_fsb
;
3652 ASSERT(!test_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
));
3655 * First check the validity of the extents described by the
3656 * EFI. If any are bad, then assume that all are bad and
3657 * just toss the EFI.
3659 for (i
= 0; i
< efip
->efi_format
.efi_nextents
; i
++) {
3660 extp
= &(efip
->efi_format
.efi_extents
[i
]);
3661 startblock_fsb
= XFS_BB_TO_FSB(mp
,
3662 XFS_FSB_TO_DADDR(mp
, extp
->ext_start
));
3663 if ((startblock_fsb
== 0) ||
3664 (extp
->ext_len
== 0) ||
3665 (startblock_fsb
>= mp
->m_sb
.sb_dblocks
) ||
3666 (extp
->ext_len
>= mp
->m_sb
.sb_agblocks
)) {
3668 * This will pull the EFI from the AIL and
3669 * free the memory associated with it.
3671 set_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
);
3672 xfs_efi_release(efip
, efip
->efi_format
.efi_nextents
);
3677 tp
= xfs_trans_alloc(mp
, 0);
3678 error
= xfs_trans_reserve(tp
, &M_RES(mp
)->tr_itruncate
, 0, 0);
3681 efdp
= xfs_trans_get_efd(tp
, efip
, efip
->efi_format
.efi_nextents
);
3683 for (i
= 0; i
< efip
->efi_format
.efi_nextents
; i
++) {
3684 extp
= &(efip
->efi_format
.efi_extents
[i
]);
3685 error
= xfs_free_extent(tp
, extp
->ext_start
, extp
->ext_len
);
3688 xfs_trans_log_efd_extent(tp
, efdp
, extp
->ext_start
,
3692 set_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
);
3693 error
= xfs_trans_commit(tp
, 0);
3697 xfs_trans_cancel(tp
, XFS_TRANS_ABORT
);
3702 * When this is called, all of the EFIs which did not have
3703 * corresponding EFDs should be in the AIL. What we do now
3704 * is free the extents associated with each one.
3706 * Since we process the EFIs in normal transactions, they
3707 * will be removed at some point after the commit. This prevents
3708 * us from just walking down the list processing each one.
3709 * We'll use a flag in the EFI to skip those that we've already
3710 * processed and use the AIL iteration mechanism's generation
3711 * count to try to speed this up at least a bit.
3713 * When we start, we know that the EFIs are the only things in
3714 * the AIL. As we process them, however, other items are added
3715 * to the AIL. Since everything added to the AIL must come after
3716 * everything already in the AIL, we stop processing as soon as
3717 * we see something other than an EFI in the AIL.
3720 xlog_recover_process_efis(
3723 xfs_log_item_t
*lip
;
3724 xfs_efi_log_item_t
*efip
;
3726 struct xfs_ail_cursor cur
;
3727 struct xfs_ail
*ailp
;
3730 spin_lock(&ailp
->xa_lock
);
3731 lip
= xfs_trans_ail_cursor_first(ailp
, &cur
, 0);
3732 while (lip
!= NULL
) {
3734 * We're done when we see something other than an EFI.
3735 * There should be no EFIs left in the AIL now.
3737 if (lip
->li_type
!= XFS_LI_EFI
) {
3739 for (; lip
; lip
= xfs_trans_ail_cursor_next(ailp
, &cur
))
3740 ASSERT(lip
->li_type
!= XFS_LI_EFI
);
3746 * Skip EFIs that we've already processed.
3748 efip
= (xfs_efi_log_item_t
*)lip
;
3749 if (test_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
)) {
3750 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3754 spin_unlock(&ailp
->xa_lock
);
3755 error
= xlog_recover_process_efi(log
->l_mp
, efip
);
3756 spin_lock(&ailp
->xa_lock
);
3759 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3762 xfs_trans_ail_cursor_done(&cur
);
3763 spin_unlock(&ailp
->xa_lock
);
3768 * This routine performs a transaction to null out a bad inode pointer
3769 * in an agi unlinked inode hash bucket.
3772 xlog_recover_clear_agi_bucket(
3774 xfs_agnumber_t agno
,
3783 tp
= xfs_trans_alloc(mp
, XFS_TRANS_CLEAR_AGI_BUCKET
);
3784 error
= xfs_trans_reserve(tp
, &M_RES(mp
)->tr_clearagi
, 0, 0);
3788 error
= xfs_read_agi(mp
, tp
, agno
, &agibp
);
3792 agi
= XFS_BUF_TO_AGI(agibp
);
3793 agi
->agi_unlinked
[bucket
] = cpu_to_be32(NULLAGINO
);
3794 offset
= offsetof(xfs_agi_t
, agi_unlinked
) +
3795 (sizeof(xfs_agino_t
) * bucket
);
3796 xfs_trans_log_buf(tp
, agibp
, offset
,
3797 (offset
+ sizeof(xfs_agino_t
) - 1));
3799 error
= xfs_trans_commit(tp
, 0);
3805 xfs_trans_cancel(tp
, XFS_TRANS_ABORT
);
3807 xfs_warn(mp
, "%s: failed to clear agi %d. Continuing.", __func__
, agno
);
3812 xlog_recover_process_one_iunlink(
3813 struct xfs_mount
*mp
,
3814 xfs_agnumber_t agno
,
3818 struct xfs_buf
*ibp
;
3819 struct xfs_dinode
*dip
;
3820 struct xfs_inode
*ip
;
3824 ino
= XFS_AGINO_TO_INO(mp
, agno
, agino
);
3825 error
= xfs_iget(mp
, NULL
, ino
, 0, 0, &ip
);
3830 * Get the on disk inode to find the next inode in the bucket.
3832 error
= xfs_imap_to_bp(mp
, NULL
, &ip
->i_imap
, &dip
, &ibp
, 0, 0);
3836 ASSERT(ip
->i_d
.di_nlink
== 0);
3837 ASSERT(ip
->i_d
.di_mode
!= 0);
3839 /* setup for the next pass */
3840 agino
= be32_to_cpu(dip
->di_next_unlinked
);
3844 * Prevent any DMAPI event from being sent when the reference on
3845 * the inode is dropped.
3847 ip
->i_d
.di_dmevmask
= 0;
3856 * We can't read in the inode this bucket points to, or this inode
3857 * is messed up. Just ditch this bucket of inodes. We will lose
3858 * some inodes and space, but at least we won't hang.
3860 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3861 * clear the inode pointer in the bucket.
3863 xlog_recover_clear_agi_bucket(mp
, agno
, bucket
);
3868 * xlog_iunlink_recover
3870 * This is called during recovery to process any inodes which
3871 * we unlinked but not freed when the system crashed. These
3872 * inodes will be on the lists in the AGI blocks. What we do
3873 * here is scan all the AGIs and fully truncate and free any
3874 * inodes found on the lists. Each inode is removed from the
3875 * lists when it has been fully truncated and is freed. The
3876 * freeing of the inode and its removal from the list must be
3880 xlog_recover_process_iunlinks(
3884 xfs_agnumber_t agno
;
3895 * Prevent any DMAPI event from being sent while in this function.
3897 mp_dmevmask
= mp
->m_dmevmask
;
3900 for (agno
= 0; agno
< mp
->m_sb
.sb_agcount
; agno
++) {
3902 * Find the agi for this ag.
3904 error
= xfs_read_agi(mp
, NULL
, agno
, &agibp
);
3907 * AGI is b0rked. Don't process it.
3909 * We should probably mark the filesystem as corrupt
3910 * after we've recovered all the ag's we can....
3915 * Unlock the buffer so that it can be acquired in the normal
3916 * course of the transaction to truncate and free each inode.
3917 * Because we are not racing with anyone else here for the AGI
3918 * buffer, we don't even need to hold it locked to read the
3919 * initial unlinked bucket entries out of the buffer. We keep
3920 * buffer reference though, so that it stays pinned in memory
3921 * while we need the buffer.
3923 agi
= XFS_BUF_TO_AGI(agibp
);
3924 xfs_buf_unlock(agibp
);
3926 for (bucket
= 0; bucket
< XFS_AGI_UNLINKED_BUCKETS
; bucket
++) {
3927 agino
= be32_to_cpu(agi
->agi_unlinked
[bucket
]);
3928 while (agino
!= NULLAGINO
) {
3929 agino
= xlog_recover_process_one_iunlink(mp
,
3930 agno
, agino
, bucket
);
3933 xfs_buf_rele(agibp
);
3936 mp
->m_dmevmask
= mp_dmevmask
;
3940 * Upack the log buffer data and crc check it. If the check fails, issue a
3941 * warning if and only if the CRC in the header is non-zero. This makes the
3942 * check an advisory warning, and the zero CRC check will prevent failure
3943 * warnings from being emitted when upgrading the kernel from one that does not
3944 * add CRCs by default.
3946 * When filesystems are CRC enabled, this CRC mismatch becomes a fatal log
3947 * corruption failure
3950 xlog_unpack_data_crc(
3951 struct xlog_rec_header
*rhead
,
3957 crc
= xlog_cksum(log
, rhead
, dp
, be32_to_cpu(rhead
->h_len
));
3958 if (crc
!= rhead
->h_crc
) {
3959 if (rhead
->h_crc
|| xfs_sb_version_hascrc(&log
->l_mp
->m_sb
)) {
3960 xfs_alert(log
->l_mp
,
3961 "log record CRC mismatch: found 0x%x, expected 0x%x.",
3962 le32_to_cpu(rhead
->h_crc
),
3964 xfs_hex_dump(dp
, 32);
3968 * If we've detected a log record corruption, then we can't
3969 * recover past this point. Abort recovery if we are enforcing
3970 * CRC protection by punting an error back up the stack.
3972 if (xfs_sb_version_hascrc(&log
->l_mp
->m_sb
))
3973 return -EFSCORRUPTED
;
3981 struct xlog_rec_header
*rhead
,
3988 error
= xlog_unpack_data_crc(rhead
, dp
, log
);
3992 for (i
= 0; i
< BTOBB(be32_to_cpu(rhead
->h_len
)) &&
3993 i
< (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
); i
++) {
3994 *(__be32
*)dp
= *(__be32
*)&rhead
->h_cycle_data
[i
];
3998 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
3999 xlog_in_core_2_t
*xhdr
= (xlog_in_core_2_t
*)rhead
;
4000 for ( ; i
< BTOBB(be32_to_cpu(rhead
->h_len
)); i
++) {
4001 j
= i
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
4002 k
= i
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
4003 *(__be32
*)dp
= xhdr
[j
].hic_xheader
.xh_cycle_data
[k
];
4012 xlog_valid_rec_header(
4014 struct xlog_rec_header
*rhead
,
4019 if (unlikely(rhead
->h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))) {
4020 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
4021 XFS_ERRLEVEL_LOW
, log
->l_mp
);
4022 return -EFSCORRUPTED
;
4025 (!rhead
->h_version
||
4026 (be32_to_cpu(rhead
->h_version
) & (~XLOG_VERSION_OKBITS
))))) {
4027 xfs_warn(log
->l_mp
, "%s: unrecognised log version (%d).",
4028 __func__
, be32_to_cpu(rhead
->h_version
));
4032 /* LR body must have data or it wouldn't have been written */
4033 hlen
= be32_to_cpu(rhead
->h_len
);
4034 if (unlikely( hlen
<= 0 || hlen
> INT_MAX
)) {
4035 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
4036 XFS_ERRLEVEL_LOW
, log
->l_mp
);
4037 return -EFSCORRUPTED
;
4039 if (unlikely( blkno
> log
->l_logBBsize
|| blkno
> INT_MAX
)) {
4040 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
4041 XFS_ERRLEVEL_LOW
, log
->l_mp
);
4042 return -EFSCORRUPTED
;
4048 * Read the log from tail to head and process the log records found.
4049 * Handle the two cases where the tail and head are in the same cycle
4050 * and where the active portion of the log wraps around the end of
4051 * the physical log separately. The pass parameter is passed through
4052 * to the routines called to process the data and is not looked at
4056 xlog_do_recovery_pass(
4058 xfs_daddr_t head_blk
,
4059 xfs_daddr_t tail_blk
,
4062 xlog_rec_header_t
*rhead
;
4065 xfs_buf_t
*hbp
, *dbp
;
4066 int error
= 0, h_size
;
4067 int bblks
, split_bblks
;
4068 int hblks
, split_hblks
, wrapped_hblks
;
4069 struct hlist_head rhash
[XLOG_RHASH_SIZE
];
4071 ASSERT(head_blk
!= tail_blk
);
4074 * Read the header of the tail block and get the iclog buffer size from
4075 * h_size. Use this to tell how many sectors make up the log header.
4077 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
4079 * When using variable length iclogs, read first sector of
4080 * iclog header and extract the header size from it. Get a
4081 * new hbp that is the correct size.
4083 hbp
= xlog_get_bp(log
, 1);
4087 error
= xlog_bread(log
, tail_blk
, 1, hbp
, &offset
);
4091 rhead
= (xlog_rec_header_t
*)offset
;
4092 error
= xlog_valid_rec_header(log
, rhead
, tail_blk
);
4095 h_size
= be32_to_cpu(rhead
->h_size
);
4096 if ((be32_to_cpu(rhead
->h_version
) & XLOG_VERSION_2
) &&
4097 (h_size
> XLOG_HEADER_CYCLE_SIZE
)) {
4098 hblks
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
4099 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
4102 hbp
= xlog_get_bp(log
, hblks
);
4107 ASSERT(log
->l_sectBBsize
== 1);
4109 hbp
= xlog_get_bp(log
, 1);
4110 h_size
= XLOG_BIG_RECORD_BSIZE
;
4115 dbp
= xlog_get_bp(log
, BTOBB(h_size
));
4121 memset(rhash
, 0, sizeof(rhash
));
4122 if (tail_blk
<= head_blk
) {
4123 for (blk_no
= tail_blk
; blk_no
< head_blk
; ) {
4124 error
= xlog_bread(log
, blk_no
, hblks
, hbp
, &offset
);
4128 rhead
= (xlog_rec_header_t
*)offset
;
4129 error
= xlog_valid_rec_header(log
, rhead
, blk_no
);
4133 /* blocks in data section */
4134 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
4135 error
= xlog_bread(log
, blk_no
+ hblks
, bblks
, dbp
,
4140 error
= xlog_unpack_data(rhead
, offset
, log
);
4144 error
= xlog_recover_process_data(log
,
4145 rhash
, rhead
, offset
, pass
);
4148 blk_no
+= bblks
+ hblks
;
4152 * Perform recovery around the end of the physical log.
4153 * When the head is not on the same cycle number as the tail,
4154 * we can't do a sequential recovery as above.
4157 while (blk_no
< log
->l_logBBsize
) {
4159 * Check for header wrapping around physical end-of-log
4161 offset
= hbp
->b_addr
;
4164 if (blk_no
+ hblks
<= log
->l_logBBsize
) {
4165 /* Read header in one read */
4166 error
= xlog_bread(log
, blk_no
, hblks
, hbp
,
4171 /* This LR is split across physical log end */
4172 if (blk_no
!= log
->l_logBBsize
) {
4173 /* some data before physical log end */
4174 ASSERT(blk_no
<= INT_MAX
);
4175 split_hblks
= log
->l_logBBsize
- (int)blk_no
;
4176 ASSERT(split_hblks
> 0);
4177 error
= xlog_bread(log
, blk_no
,
4185 * Note: this black magic still works with
4186 * large sector sizes (non-512) only because:
4187 * - we increased the buffer size originally
4188 * by 1 sector giving us enough extra space
4189 * for the second read;
4190 * - the log start is guaranteed to be sector
4192 * - we read the log end (LR header start)
4193 * _first_, then the log start (LR header end)
4194 * - order is important.
4196 wrapped_hblks
= hblks
- split_hblks
;
4197 error
= xlog_bread_offset(log
, 0,
4199 offset
+ BBTOB(split_hblks
));
4203 rhead
= (xlog_rec_header_t
*)offset
;
4204 error
= xlog_valid_rec_header(log
, rhead
,
4205 split_hblks
? blk_no
: 0);
4209 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
4212 /* Read in data for log record */
4213 if (blk_no
+ bblks
<= log
->l_logBBsize
) {
4214 error
= xlog_bread(log
, blk_no
, bblks
, dbp
,
4219 /* This log record is split across the
4220 * physical end of log */
4221 offset
= dbp
->b_addr
;
4223 if (blk_no
!= log
->l_logBBsize
) {
4224 /* some data is before the physical
4226 ASSERT(!wrapped_hblks
);
4227 ASSERT(blk_no
<= INT_MAX
);
4229 log
->l_logBBsize
- (int)blk_no
;
4230 ASSERT(split_bblks
> 0);
4231 error
= xlog_bread(log
, blk_no
,
4239 * Note: this black magic still works with
4240 * large sector sizes (non-512) only because:
4241 * - we increased the buffer size originally
4242 * by 1 sector giving us enough extra space
4243 * for the second read;
4244 * - the log start is guaranteed to be sector
4246 * - we read the log end (LR header start)
4247 * _first_, then the log start (LR header end)
4248 * - order is important.
4250 error
= xlog_bread_offset(log
, 0,
4251 bblks
- split_bblks
, dbp
,
4252 offset
+ BBTOB(split_bblks
));
4257 error
= xlog_unpack_data(rhead
, offset
, log
);
4261 error
= xlog_recover_process_data(log
, rhash
,
4262 rhead
, offset
, pass
);
4268 ASSERT(blk_no
>= log
->l_logBBsize
);
4269 blk_no
-= log
->l_logBBsize
;
4271 /* read first part of physical log */
4272 while (blk_no
< head_blk
) {
4273 error
= xlog_bread(log
, blk_no
, hblks
, hbp
, &offset
);
4277 rhead
= (xlog_rec_header_t
*)offset
;
4278 error
= xlog_valid_rec_header(log
, rhead
, blk_no
);
4282 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
4283 error
= xlog_bread(log
, blk_no
+hblks
, bblks
, dbp
,
4288 error
= xlog_unpack_data(rhead
, offset
, log
);
4292 error
= xlog_recover_process_data(log
, rhash
,
4293 rhead
, offset
, pass
);
4296 blk_no
+= bblks
+ hblks
;
4308 * Do the recovery of the log. We actually do this in two phases.
4309 * The two passes are necessary in order to implement the function
4310 * of cancelling a record written into the log. The first pass
4311 * determines those things which have been cancelled, and the
4312 * second pass replays log items normally except for those which
4313 * have been cancelled. The handling of the replay and cancellations
4314 * takes place in the log item type specific routines.
4316 * The table of items which have cancel records in the log is allocated
4317 * and freed at this level, since only here do we know when all of
4318 * the log recovery has been completed.
4321 xlog_do_log_recovery(
4323 xfs_daddr_t head_blk
,
4324 xfs_daddr_t tail_blk
)
4328 ASSERT(head_blk
!= tail_blk
);
4331 * First do a pass to find all of the cancelled buf log items.
4332 * Store them in the buf_cancel_table for use in the second pass.
4334 log
->l_buf_cancel_table
= kmem_zalloc(XLOG_BC_TABLE_SIZE
*
4335 sizeof(struct list_head
),
4337 for (i
= 0; i
< XLOG_BC_TABLE_SIZE
; i
++)
4338 INIT_LIST_HEAD(&log
->l_buf_cancel_table
[i
]);
4340 error
= xlog_do_recovery_pass(log
, head_blk
, tail_blk
,
4341 XLOG_RECOVER_PASS1
);
4343 kmem_free(log
->l_buf_cancel_table
);
4344 log
->l_buf_cancel_table
= NULL
;
4348 * Then do a second pass to actually recover the items in the log.
4349 * When it is complete free the table of buf cancel items.
4351 error
= xlog_do_recovery_pass(log
, head_blk
, tail_blk
,
4352 XLOG_RECOVER_PASS2
);
4357 for (i
= 0; i
< XLOG_BC_TABLE_SIZE
; i
++)
4358 ASSERT(list_empty(&log
->l_buf_cancel_table
[i
]));
4362 kmem_free(log
->l_buf_cancel_table
);
4363 log
->l_buf_cancel_table
= NULL
;
4369 * Do the actual recovery
4374 xfs_daddr_t head_blk
,
4375 xfs_daddr_t tail_blk
)
4382 * First replay the images in the log.
4384 error
= xlog_do_log_recovery(log
, head_blk
, tail_blk
);
4389 * If IO errors happened during recovery, bail out.
4391 if (XFS_FORCED_SHUTDOWN(log
->l_mp
)) {
4396 * We now update the tail_lsn since much of the recovery has completed
4397 * and there may be space available to use. If there were no extent
4398 * or iunlinks, we can free up the entire log and set the tail_lsn to
4399 * be the last_sync_lsn. This was set in xlog_find_tail to be the
4400 * lsn of the last known good LR on disk. If there are extent frees
4401 * or iunlinks they will have some entries in the AIL; so we look at
4402 * the AIL to determine how to set the tail_lsn.
4404 xlog_assign_tail_lsn(log
->l_mp
);
4407 * Now that we've finished replaying all buffer and inode
4408 * updates, re-read in the superblock and reverify it.
4410 bp
= xfs_getsb(log
->l_mp
, 0);
4412 ASSERT(!(XFS_BUF_ISWRITE(bp
)));
4414 XFS_BUF_UNASYNC(bp
);
4415 bp
->b_ops
= &xfs_sb_buf_ops
;
4417 if (XFS_FORCED_SHUTDOWN(log
->l_mp
)) {
4422 xfs_buf_iorequest(bp
);
4423 error
= xfs_buf_iowait(bp
);
4425 xfs_buf_ioerror_alert(bp
, __func__
);
4431 /* Convert superblock from on-disk format */
4432 sbp
= &log
->l_mp
->m_sb
;
4433 xfs_sb_from_disk(sbp
, XFS_BUF_TO_SBP(bp
));
4434 ASSERT(sbp
->sb_magicnum
== XFS_SB_MAGIC
);
4435 ASSERT(xfs_sb_good_version(sbp
));
4438 /* We've re-read the superblock so re-initialize per-cpu counters */
4439 xfs_icsb_reinit_counters(log
->l_mp
);
4441 xlog_recover_check_summary(log
);
4443 /* Normal transactions can now occur */
4444 log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
4449 * Perform recovery and re-initialize some log variables in xlog_find_tail.
4451 * Return error or zero.
4457 xfs_daddr_t head_blk
, tail_blk
;
4460 /* find the tail of the log */
4461 if ((error
= xlog_find_tail(log
, &head_blk
, &tail_blk
)))
4464 if (tail_blk
!= head_blk
) {
4465 /* There used to be a comment here:
4467 * disallow recovery on read-only mounts. note -- mount
4468 * checks for ENOSPC and turns it into an intelligent
4470 * ...but this is no longer true. Now, unless you specify
4471 * NORECOVERY (in which case this function would never be
4472 * called), we just go ahead and recover. We do this all
4473 * under the vfs layer, so we can get away with it unless
4474 * the device itself is read-only, in which case we fail.
4476 if ((error
= xfs_dev_is_read_only(log
->l_mp
, "recovery"))) {
4481 * Version 5 superblock log feature mask validation. We know the
4482 * log is dirty so check if there are any unknown log features
4483 * in what we need to recover. If there are unknown features
4484 * (e.g. unsupported transactions, then simply reject the
4485 * attempt at recovery before touching anything.
4487 if (XFS_SB_VERSION_NUM(&log
->l_mp
->m_sb
) == XFS_SB_VERSION_5
&&
4488 xfs_sb_has_incompat_log_feature(&log
->l_mp
->m_sb
,
4489 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN
)) {
4491 "Superblock has unknown incompatible log features (0x%x) enabled.\n"
4492 "The log can not be fully and/or safely recovered by this kernel.\n"
4493 "Please recover the log on a kernel that supports the unknown features.",
4494 (log
->l_mp
->m_sb
.sb_features_log_incompat
&
4495 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN
));
4499 xfs_notice(log
->l_mp
, "Starting recovery (logdev: %s)",
4500 log
->l_mp
->m_logname
? log
->l_mp
->m_logname
4503 error
= xlog_do_recover(log
, head_blk
, tail_blk
);
4504 log
->l_flags
|= XLOG_RECOVERY_NEEDED
;
4510 * In the first part of recovery we replay inodes and buffers and build
4511 * up the list of extent free items which need to be processed. Here
4512 * we process the extent free items and clean up the on disk unlinked
4513 * inode lists. This is separated from the first part of recovery so
4514 * that the root and real-time bitmap inodes can be read in from disk in
4515 * between the two stages. This is necessary so that we can free space
4516 * in the real-time portion of the file system.
4519 xlog_recover_finish(
4523 * Now we're ready to do the transactions needed for the
4524 * rest of recovery. Start with completing all the extent
4525 * free intent records and then process the unlinked inode
4526 * lists. At this point, we essentially run in normal mode
4527 * except that we're still performing recovery actions
4528 * rather than accepting new requests.
4530 if (log
->l_flags
& XLOG_RECOVERY_NEEDED
) {
4532 error
= xlog_recover_process_efis(log
);
4534 xfs_alert(log
->l_mp
, "Failed to recover EFIs");
4538 * Sync the log to get all the EFIs out of the AIL.
4539 * This isn't absolutely necessary, but it helps in
4540 * case the unlink transactions would have problems
4541 * pushing the EFIs out of the way.
4543 xfs_log_force(log
->l_mp
, XFS_LOG_SYNC
);
4545 xlog_recover_process_iunlinks(log
);
4547 xlog_recover_check_summary(log
);
4549 xfs_notice(log
->l_mp
, "Ending recovery (logdev: %s)",
4550 log
->l_mp
->m_logname
? log
->l_mp
->m_logname
4552 log
->l_flags
&= ~XLOG_RECOVERY_NEEDED
;
4554 xfs_info(log
->l_mp
, "Ending clean mount");
4562 * Read all of the agf and agi counters and check that they
4563 * are consistent with the superblock counters.
4566 xlog_recover_check_summary(
4573 xfs_agnumber_t agno
;
4574 __uint64_t freeblks
;
4584 for (agno
= 0; agno
< mp
->m_sb
.sb_agcount
; agno
++) {
4585 error
= xfs_read_agf(mp
, NULL
, agno
, 0, &agfbp
);
4587 xfs_alert(mp
, "%s agf read failed agno %d error %d",
4588 __func__
, agno
, error
);
4590 agfp
= XFS_BUF_TO_AGF(agfbp
);
4591 freeblks
+= be32_to_cpu(agfp
->agf_freeblks
) +
4592 be32_to_cpu(agfp
->agf_flcount
);
4593 xfs_buf_relse(agfbp
);
4596 error
= xfs_read_agi(mp
, NULL
, agno
, &agibp
);
4598 xfs_alert(mp
, "%s agi read failed agno %d error %d",
4599 __func__
, agno
, error
);
4601 struct xfs_agi
*agi
= XFS_BUF_TO_AGI(agibp
);
4603 itotal
+= be32_to_cpu(agi
->agi_count
);
4604 ifree
+= be32_to_cpu(agi
->agi_freecount
);
4605 xfs_buf_relse(agibp
);