2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_dinode.h"
33 #include "xfs_inode.h"
34 #include "xfs_inode_item.h"
35 #include "xfs_alloc.h"
36 #include "xfs_ialloc.h"
37 #include "xfs_log_priv.h"
38 #include "xfs_buf_item.h"
39 #include "xfs_log_recover.h"
40 #include "xfs_extfree_item.h"
41 #include "xfs_trans_priv.h"
42 #include "xfs_quota.h"
44 #include "xfs_utils.h"
45 #include "xfs_trace.h"
47 STATIC
int xlog_find_zeroed(xlog_t
*, xfs_daddr_t
*);
48 STATIC
int xlog_clear_stale_blocks(xlog_t
*, xfs_lsn_t
);
50 STATIC
void xlog_recover_check_summary(xlog_t
*);
52 #define xlog_recover_check_summary(log)
56 * This structure is used during recovery to record the buf log items which
57 * have been canceled and should not be replayed.
59 struct xfs_buf_cancel
{
63 struct list_head bc_list
;
67 * Sector aligned buffer routines for buffer create/read/write/access
71 * Verify the given count of basic blocks is valid number of blocks
72 * to specify for an operation involving the given XFS log buffer.
73 * Returns nonzero if the count is valid, 0 otherwise.
77 xlog_buf_bbcount_valid(
81 return bbcount
> 0 && bbcount
<= log
->l_logBBsize
;
85 * Allocate a buffer to hold log data. The buffer needs to be able
86 * to map to a range of nbblks basic blocks at any valid (basic
87 * block) offset within the log.
96 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
97 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
99 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
104 * We do log I/O in units of log sectors (a power-of-2
105 * multiple of the basic block size), so we round up the
106 * requested size to accommodate the basic blocks required
107 * for complete log sectors.
109 * In addition, the buffer may be used for a non-sector-
110 * aligned block offset, in which case an I/O of the
111 * requested size could extend beyond the end of the
112 * buffer. If the requested size is only 1 basic block it
113 * will never straddle a sector boundary, so this won't be
114 * an issue. Nor will this be a problem if the log I/O is
115 * done in basic blocks (sector size 1). But otherwise we
116 * extend the buffer by one extra log sector to ensure
117 * there's space to accommodate this possibility.
119 if (nbblks
> 1 && log
->l_sectBBsize
> 1)
120 nbblks
+= log
->l_sectBBsize
;
121 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
123 bp
= xfs_buf_get_uncached(log
->l_mp
->m_logdev_targp
, BBTOB(nbblks
), 0);
137 * Return the address of the start of the given block number's data
138 * in a log buffer. The buffer covers a log sector-aligned region.
147 xfs_daddr_t offset
= blk_no
& ((xfs_daddr_t
)log
->l_sectBBsize
- 1);
149 ASSERT(BBTOB(offset
+ nbblks
) <= XFS_BUF_SIZE(bp
));
150 return bp
->b_addr
+ BBTOB(offset
);
155 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
166 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
167 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
169 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
173 blk_no
= round_down(blk_no
, log
->l_sectBBsize
);
174 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
177 ASSERT(BBTOB(nbblks
) <= XFS_BUF_SIZE(bp
));
179 XFS_BUF_SET_ADDR(bp
, log
->l_logBBstart
+ blk_no
);
181 XFS_BUF_SET_COUNT(bp
, BBTOB(nbblks
));
183 xfsbdstrat(log
->l_mp
, bp
);
184 error
= xfs_buf_iowait(bp
);
186 xfs_buf_ioerror_alert(bp
, __func__
);
200 error
= xlog_bread_noalign(log
, blk_no
, nbblks
, bp
);
204 *offset
= xlog_align(log
, blk_no
, nbblks
, bp
);
209 * Read at an offset into the buffer. Returns with the buffer in it's original
210 * state regardless of the result of the read.
215 xfs_daddr_t blk_no
, /* block to read from */
216 int nbblks
, /* blocks to read */
220 xfs_caddr_t orig_offset
= bp
->b_addr
;
221 int orig_len
= bp
->b_buffer_length
;
224 error
= xfs_buf_associate_memory(bp
, offset
, BBTOB(nbblks
));
228 error
= xlog_bread_noalign(log
, blk_no
, nbblks
, bp
);
230 /* must reset buffer pointer even on error */
231 error2
= xfs_buf_associate_memory(bp
, orig_offset
, orig_len
);
238 * Write out the buffer at the given block for the given number of blocks.
239 * The buffer is kept locked across the write and is returned locked.
240 * This can only be used for synchronous log writes.
251 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
252 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
254 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
258 blk_no
= round_down(blk_no
, log
->l_sectBBsize
);
259 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
262 ASSERT(BBTOB(nbblks
) <= XFS_BUF_SIZE(bp
));
264 XFS_BUF_SET_ADDR(bp
, log
->l_logBBstart
+ blk_no
);
265 XFS_BUF_ZEROFLAGS(bp
);
268 XFS_BUF_SET_COUNT(bp
, BBTOB(nbblks
));
270 error
= xfs_bwrite(bp
);
272 xfs_buf_ioerror_alert(bp
, __func__
);
279 * dump debug superblock and log record information
282 xlog_header_check_dump(
284 xlog_rec_header_t
*head
)
286 xfs_debug(mp
, "%s: SB : uuid = %pU, fmt = %d\n",
287 __func__
, &mp
->m_sb
.sb_uuid
, XLOG_FMT
);
288 xfs_debug(mp
, " log : uuid = %pU, fmt = %d\n",
289 &head
->h_fs_uuid
, be32_to_cpu(head
->h_fmt
));
292 #define xlog_header_check_dump(mp, head)
296 * check log record header for recovery
299 xlog_header_check_recover(
301 xlog_rec_header_t
*head
)
303 ASSERT(head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
));
306 * IRIX doesn't write the h_fmt field and leaves it zeroed
307 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
308 * a dirty log created in IRIX.
310 if (unlikely(head
->h_fmt
!= cpu_to_be32(XLOG_FMT
))) {
312 "dirty log written in incompatible format - can't recover");
313 xlog_header_check_dump(mp
, head
);
314 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
315 XFS_ERRLEVEL_HIGH
, mp
);
316 return XFS_ERROR(EFSCORRUPTED
);
317 } else if (unlikely(!uuid_equal(&mp
->m_sb
.sb_uuid
, &head
->h_fs_uuid
))) {
319 "dirty log entry has mismatched uuid - can't recover");
320 xlog_header_check_dump(mp
, head
);
321 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
322 XFS_ERRLEVEL_HIGH
, mp
);
323 return XFS_ERROR(EFSCORRUPTED
);
329 * read the head block of the log and check the header
332 xlog_header_check_mount(
334 xlog_rec_header_t
*head
)
336 ASSERT(head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
));
338 if (uuid_is_nil(&head
->h_fs_uuid
)) {
340 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
341 * h_fs_uuid is nil, we assume this log was last mounted
342 * by IRIX and continue.
344 xfs_warn(mp
, "nil uuid in log - IRIX style log");
345 } else if (unlikely(!uuid_equal(&mp
->m_sb
.sb_uuid
, &head
->h_fs_uuid
))) {
346 xfs_warn(mp
, "log has mismatched uuid - can't recover");
347 xlog_header_check_dump(mp
, head
);
348 XFS_ERROR_REPORT("xlog_header_check_mount",
349 XFS_ERRLEVEL_HIGH
, mp
);
350 return XFS_ERROR(EFSCORRUPTED
);
361 * We're not going to bother about retrying
362 * this during recovery. One strike!
364 xfs_buf_ioerror_alert(bp
, __func__
);
365 xfs_force_shutdown(bp
->b_target
->bt_mount
,
366 SHUTDOWN_META_IO_ERROR
);
369 xfs_buf_ioend(bp
, 0);
373 * This routine finds (to an approximation) the first block in the physical
374 * log which contains the given cycle. It uses a binary search algorithm.
375 * Note that the algorithm can not be perfect because the disk will not
376 * necessarily be perfect.
379 xlog_find_cycle_start(
382 xfs_daddr_t first_blk
,
383 xfs_daddr_t
*last_blk
,
393 mid_blk
= BLK_AVG(first_blk
, end_blk
);
394 while (mid_blk
!= first_blk
&& mid_blk
!= end_blk
) {
395 error
= xlog_bread(log
, mid_blk
, 1, bp
, &offset
);
398 mid_cycle
= xlog_get_cycle(offset
);
399 if (mid_cycle
== cycle
)
400 end_blk
= mid_blk
; /* last_half_cycle == mid_cycle */
402 first_blk
= mid_blk
; /* first_half_cycle == mid_cycle */
403 mid_blk
= BLK_AVG(first_blk
, end_blk
);
405 ASSERT((mid_blk
== first_blk
&& mid_blk
+1 == end_blk
) ||
406 (mid_blk
== end_blk
&& mid_blk
-1 == first_blk
));
414 * Check that a range of blocks does not contain stop_on_cycle_no.
415 * Fill in *new_blk with the block offset where such a block is
416 * found, or with -1 (an invalid block number) if there is no such
417 * block in the range. The scan needs to occur from front to back
418 * and the pointer into the region must be updated since a later
419 * routine will need to perform another test.
422 xlog_find_verify_cycle(
424 xfs_daddr_t start_blk
,
426 uint stop_on_cycle_no
,
427 xfs_daddr_t
*new_blk
)
433 xfs_caddr_t buf
= NULL
;
437 * Greedily allocate a buffer big enough to handle the full
438 * range of basic blocks we'll be examining. If that fails,
439 * try a smaller size. We need to be able to read at least
440 * a log sector, or we're out of luck.
442 bufblks
= 1 << ffs(nbblks
);
443 while (!(bp
= xlog_get_bp(log
, bufblks
))) {
445 if (bufblks
< log
->l_sectBBsize
)
449 for (i
= start_blk
; i
< start_blk
+ nbblks
; i
+= bufblks
) {
452 bcount
= min(bufblks
, (start_blk
+ nbblks
- i
));
454 error
= xlog_bread(log
, i
, bcount
, bp
, &buf
);
458 for (j
= 0; j
< bcount
; j
++) {
459 cycle
= xlog_get_cycle(buf
);
460 if (cycle
== stop_on_cycle_no
) {
477 * Potentially backup over partial log record write.
479 * In the typical case, last_blk is the number of the block directly after
480 * a good log record. Therefore, we subtract one to get the block number
481 * of the last block in the given buffer. extra_bblks contains the number
482 * of blocks we would have read on a previous read. This happens when the
483 * last log record is split over the end of the physical log.
485 * extra_bblks is the number of blocks potentially verified on a previous
486 * call to this routine.
489 xlog_find_verify_log_record(
491 xfs_daddr_t start_blk
,
492 xfs_daddr_t
*last_blk
,
497 xfs_caddr_t offset
= NULL
;
498 xlog_rec_header_t
*head
= NULL
;
501 int num_blks
= *last_blk
- start_blk
;
504 ASSERT(start_blk
!= 0 || *last_blk
!= start_blk
);
506 if (!(bp
= xlog_get_bp(log
, num_blks
))) {
507 if (!(bp
= xlog_get_bp(log
, 1)))
511 error
= xlog_bread(log
, start_blk
, num_blks
, bp
, &offset
);
514 offset
+= ((num_blks
- 1) << BBSHIFT
);
517 for (i
= (*last_blk
) - 1; i
>= 0; i
--) {
519 /* valid log record not found */
521 "Log inconsistent (didn't find previous header)");
523 error
= XFS_ERROR(EIO
);
528 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
533 head
= (xlog_rec_header_t
*)offset
;
535 if (head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
543 * We hit the beginning of the physical log & still no header. Return
544 * to caller. If caller can handle a return of -1, then this routine
545 * will be called again for the end of the physical log.
553 * We have the final block of the good log (the first block
554 * of the log record _before_ the head. So we check the uuid.
556 if ((error
= xlog_header_check_mount(log
->l_mp
, head
)))
560 * We may have found a log record header before we expected one.
561 * last_blk will be the 1st block # with a given cycle #. We may end
562 * up reading an entire log record. In this case, we don't want to
563 * reset last_blk. Only when last_blk points in the middle of a log
564 * record do we update last_blk.
566 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
567 uint h_size
= be32_to_cpu(head
->h_size
);
569 xhdrs
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
570 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
576 if (*last_blk
- i
+ extra_bblks
!=
577 BTOBB(be32_to_cpu(head
->h_len
)) + xhdrs
)
586 * Head is defined to be the point of the log where the next log write
587 * write could go. This means that incomplete LR writes at the end are
588 * eliminated when calculating the head. We aren't guaranteed that previous
589 * LR have complete transactions. We only know that a cycle number of
590 * current cycle number -1 won't be present in the log if we start writing
591 * from our current block number.
593 * last_blk contains the block number of the first block with a given
596 * Return: zero if normal, non-zero if error.
601 xfs_daddr_t
*return_head_blk
)
605 xfs_daddr_t new_blk
, first_blk
, start_blk
, last_blk
, head_blk
;
607 uint first_half_cycle
, last_half_cycle
;
609 int error
, log_bbnum
= log
->l_logBBsize
;
611 /* Is the end of the log device zeroed? */
612 if ((error
= xlog_find_zeroed(log
, &first_blk
)) == -1) {
613 *return_head_blk
= first_blk
;
615 /* Is the whole lot zeroed? */
617 /* Linux XFS shouldn't generate totally zeroed logs -
618 * mkfs etc write a dummy unmount record to a fresh
619 * log so we can store the uuid in there
621 xfs_warn(log
->l_mp
, "totally zeroed log");
626 xfs_warn(log
->l_mp
, "empty log check failed");
630 first_blk
= 0; /* get cycle # of 1st block */
631 bp
= xlog_get_bp(log
, 1);
635 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
639 first_half_cycle
= xlog_get_cycle(offset
);
641 last_blk
= head_blk
= log_bbnum
- 1; /* get cycle # of last block */
642 error
= xlog_bread(log
, last_blk
, 1, bp
, &offset
);
646 last_half_cycle
= xlog_get_cycle(offset
);
647 ASSERT(last_half_cycle
!= 0);
650 * If the 1st half cycle number is equal to the last half cycle number,
651 * then the entire log is stamped with the same cycle number. In this
652 * case, head_blk can't be set to zero (which makes sense). The below
653 * math doesn't work out properly with head_blk equal to zero. Instead,
654 * we set it to log_bbnum which is an invalid block number, but this
655 * value makes the math correct. If head_blk doesn't changed through
656 * all the tests below, *head_blk is set to zero at the very end rather
657 * than log_bbnum. In a sense, log_bbnum and zero are the same block
658 * in a circular file.
660 if (first_half_cycle
== last_half_cycle
) {
662 * In this case we believe that the entire log should have
663 * cycle number last_half_cycle. We need to scan backwards
664 * from the end verifying that there are no holes still
665 * containing last_half_cycle - 1. If we find such a hole,
666 * then the start of that hole will be the new head. The
667 * simple case looks like
668 * x | x ... | x - 1 | x
669 * Another case that fits this picture would be
670 * x | x + 1 | x ... | x
671 * In this case the head really is somewhere at the end of the
672 * log, as one of the latest writes at the beginning was
675 * x | x + 1 | x ... | x - 1 | x
676 * This is really the combination of the above two cases, and
677 * the head has to end up at the start of the x-1 hole at the
680 * In the 256k log case, we will read from the beginning to the
681 * end of the log and search for cycle numbers equal to x-1.
682 * We don't worry about the x+1 blocks that we encounter,
683 * because we know that they cannot be the head since the log
686 head_blk
= log_bbnum
;
687 stop_on_cycle
= last_half_cycle
- 1;
690 * In this case we want to find the first block with cycle
691 * number matching last_half_cycle. We expect the log to be
693 * x + 1 ... | x ... | x
694 * The first block with cycle number x (last_half_cycle) will
695 * be where the new head belongs. First we do a binary search
696 * for the first occurrence of last_half_cycle. The binary
697 * search may not be totally accurate, so then we scan back
698 * from there looking for occurrences of last_half_cycle before
699 * us. If that backwards scan wraps around the beginning of
700 * the log, then we look for occurrences of last_half_cycle - 1
701 * at the end of the log. The cases we're looking for look
703 * v binary search stopped here
704 * x + 1 ... | x | x + 1 | x ... | x
705 * ^ but we want to locate this spot
707 * <---------> less than scan distance
708 * x + 1 ... | x ... | x - 1 | x
709 * ^ we want to locate this spot
711 stop_on_cycle
= last_half_cycle
;
712 if ((error
= xlog_find_cycle_start(log
, bp
, first_blk
,
713 &head_blk
, last_half_cycle
)))
718 * Now validate the answer. Scan back some number of maximum possible
719 * blocks and make sure each one has the expected cycle number. The
720 * maximum is determined by the total possible amount of buffering
721 * in the in-core log. The following number can be made tighter if
722 * we actually look at the block size of the filesystem.
724 num_scan_bblks
= XLOG_TOTAL_REC_SHIFT(log
);
725 if (head_blk
>= num_scan_bblks
) {
727 * We are guaranteed that the entire check can be performed
730 start_blk
= head_blk
- num_scan_bblks
;
731 if ((error
= xlog_find_verify_cycle(log
,
732 start_blk
, num_scan_bblks
,
733 stop_on_cycle
, &new_blk
)))
737 } else { /* need to read 2 parts of log */
739 * We are going to scan backwards in the log in two parts.
740 * First we scan the physical end of the log. In this part
741 * of the log, we are looking for blocks with cycle number
742 * last_half_cycle - 1.
743 * If we find one, then we know that the log starts there, as
744 * we've found a hole that didn't get written in going around
745 * the end of the physical log. The simple case for this is
746 * x + 1 ... | x ... | x - 1 | x
747 * <---------> less than scan distance
748 * If all of the blocks at the end of the log have cycle number
749 * last_half_cycle, then we check the blocks at the start of
750 * the log looking for occurrences of last_half_cycle. If we
751 * find one, then our current estimate for the location of the
752 * first occurrence of last_half_cycle is wrong and we move
753 * back to the hole we've found. This case looks like
754 * x + 1 ... | x | x + 1 | x ...
755 * ^ binary search stopped here
756 * Another case we need to handle that only occurs in 256k
758 * x + 1 ... | x ... | x+1 | x ...
759 * ^ binary search stops here
760 * In a 256k log, the scan at the end of the log will see the
761 * x + 1 blocks. We need to skip past those since that is
762 * certainly not the head of the log. By searching for
763 * last_half_cycle-1 we accomplish that.
765 ASSERT(head_blk
<= INT_MAX
&&
766 (xfs_daddr_t
) num_scan_bblks
>= head_blk
);
767 start_blk
= log_bbnum
- (num_scan_bblks
- head_blk
);
768 if ((error
= xlog_find_verify_cycle(log
, start_blk
,
769 num_scan_bblks
- (int)head_blk
,
770 (stop_on_cycle
- 1), &new_blk
)))
778 * Scan beginning of log now. The last part of the physical
779 * log is good. This scan needs to verify that it doesn't find
780 * the last_half_cycle.
783 ASSERT(head_blk
<= INT_MAX
);
784 if ((error
= xlog_find_verify_cycle(log
,
785 start_blk
, (int)head_blk
,
786 stop_on_cycle
, &new_blk
)))
794 * Now we need to make sure head_blk is not pointing to a block in
795 * the middle of a log record.
797 num_scan_bblks
= XLOG_REC_SHIFT(log
);
798 if (head_blk
>= num_scan_bblks
) {
799 start_blk
= head_blk
- num_scan_bblks
; /* don't read head_blk */
801 /* start ptr at last block ptr before head_blk */
802 if ((error
= xlog_find_verify_log_record(log
, start_blk
,
803 &head_blk
, 0)) == -1) {
804 error
= XFS_ERROR(EIO
);
810 ASSERT(head_blk
<= INT_MAX
);
811 if ((error
= xlog_find_verify_log_record(log
, start_blk
,
812 &head_blk
, 0)) == -1) {
813 /* We hit the beginning of the log during our search */
814 start_blk
= log_bbnum
- (num_scan_bblks
- head_blk
);
816 ASSERT(start_blk
<= INT_MAX
&&
817 (xfs_daddr_t
) log_bbnum
-start_blk
>= 0);
818 ASSERT(head_blk
<= INT_MAX
);
819 if ((error
= xlog_find_verify_log_record(log
,
821 (int)head_blk
)) == -1) {
822 error
= XFS_ERROR(EIO
);
826 if (new_blk
!= log_bbnum
)
833 if (head_blk
== log_bbnum
)
834 *return_head_blk
= 0;
836 *return_head_blk
= head_blk
;
838 * When returning here, we have a good block number. Bad block
839 * means that during a previous crash, we didn't have a clean break
840 * from cycle number N to cycle number N-1. In this case, we need
841 * to find the first block with cycle number N-1.
849 xfs_warn(log
->l_mp
, "failed to find log head");
854 * Find the sync block number or the tail of the log.
856 * This will be the block number of the last record to have its
857 * associated buffers synced to disk. Every log record header has
858 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
859 * to get a sync block number. The only concern is to figure out which
860 * log record header to believe.
862 * The following algorithm uses the log record header with the largest
863 * lsn. The entire log record does not need to be valid. We only care
864 * that the header is valid.
866 * We could speed up search by using current head_blk buffer, but it is not
872 xfs_daddr_t
*head_blk
,
873 xfs_daddr_t
*tail_blk
)
875 xlog_rec_header_t
*rhead
;
876 xlog_op_header_t
*op_head
;
877 xfs_caddr_t offset
= NULL
;
880 xfs_daddr_t umount_data_blk
;
881 xfs_daddr_t after_umount_blk
;
888 * Find previous log record
890 if ((error
= xlog_find_head(log
, head_blk
)))
893 bp
= xlog_get_bp(log
, 1);
896 if (*head_blk
== 0) { /* special case */
897 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
901 if (xlog_get_cycle(offset
) == 0) {
903 /* leave all other log inited values alone */
909 * Search backwards looking for log record header block
911 ASSERT(*head_blk
< INT_MAX
);
912 for (i
= (int)(*head_blk
) - 1; i
>= 0; i
--) {
913 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
917 if (*(__be32
*)offset
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
)) {
923 * If we haven't found the log record header block, start looking
924 * again from the end of the physical log. XXXmiken: There should be
925 * a check here to make sure we didn't search more than N blocks in
929 for (i
= log
->l_logBBsize
- 1; i
>= (int)(*head_blk
); i
--) {
930 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
934 if (*(__be32
*)offset
==
935 cpu_to_be32(XLOG_HEADER_MAGIC_NUM
)) {
942 xfs_warn(log
->l_mp
, "%s: couldn't find sync record", __func__
);
944 return XFS_ERROR(EIO
);
947 /* find blk_no of tail of log */
948 rhead
= (xlog_rec_header_t
*)offset
;
949 *tail_blk
= BLOCK_LSN(be64_to_cpu(rhead
->h_tail_lsn
));
952 * Reset log values according to the state of the log when we
953 * crashed. In the case where head_blk == 0, we bump curr_cycle
954 * one because the next write starts a new cycle rather than
955 * continuing the cycle of the last good log record. At this
956 * point we have guaranteed that all partial log records have been
957 * accounted for. Therefore, we know that the last good log record
958 * written was complete and ended exactly on the end boundary
959 * of the physical log.
961 log
->l_prev_block
= i
;
962 log
->l_curr_block
= (int)*head_blk
;
963 log
->l_curr_cycle
= be32_to_cpu(rhead
->h_cycle
);
966 atomic64_set(&log
->l_tail_lsn
, be64_to_cpu(rhead
->h_tail_lsn
));
967 atomic64_set(&log
->l_last_sync_lsn
, be64_to_cpu(rhead
->h_lsn
));
968 xlog_assign_grant_head(&log
->l_reserve_head
.grant
, log
->l_curr_cycle
,
969 BBTOB(log
->l_curr_block
));
970 xlog_assign_grant_head(&log
->l_write_head
.grant
, log
->l_curr_cycle
,
971 BBTOB(log
->l_curr_block
));
974 * Look for unmount record. If we find it, then we know there
975 * was a clean unmount. Since 'i' could be the last block in
976 * the physical log, we convert to a log block before comparing
979 * Save the current tail lsn to use to pass to
980 * xlog_clear_stale_blocks() below. We won't want to clear the
981 * unmount record if there is one, so we pass the lsn of the
982 * unmount record rather than the block after it.
984 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
985 int h_size
= be32_to_cpu(rhead
->h_size
);
986 int h_version
= be32_to_cpu(rhead
->h_version
);
988 if ((h_version
& XLOG_VERSION_2
) &&
989 (h_size
> XLOG_HEADER_CYCLE_SIZE
)) {
990 hblks
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
991 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
999 after_umount_blk
= (i
+ hblks
+ (int)
1000 BTOBB(be32_to_cpu(rhead
->h_len
))) % log
->l_logBBsize
;
1001 tail_lsn
= atomic64_read(&log
->l_tail_lsn
);
1002 if (*head_blk
== after_umount_blk
&&
1003 be32_to_cpu(rhead
->h_num_logops
) == 1) {
1004 umount_data_blk
= (i
+ hblks
) % log
->l_logBBsize
;
1005 error
= xlog_bread(log
, umount_data_blk
, 1, bp
, &offset
);
1009 op_head
= (xlog_op_header_t
*)offset
;
1010 if (op_head
->oh_flags
& XLOG_UNMOUNT_TRANS
) {
1012 * Set tail and last sync so that newly written
1013 * log records will point recovery to after the
1014 * current unmount record.
1016 xlog_assign_atomic_lsn(&log
->l_tail_lsn
,
1017 log
->l_curr_cycle
, after_umount_blk
);
1018 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
,
1019 log
->l_curr_cycle
, after_umount_blk
);
1020 *tail_blk
= after_umount_blk
;
1023 * Note that the unmount was clean. If the unmount
1024 * was not clean, we need to know this to rebuild the
1025 * superblock counters from the perag headers if we
1026 * have a filesystem using non-persistent counters.
1028 log
->l_mp
->m_flags
|= XFS_MOUNT_WAS_CLEAN
;
1033 * Make sure that there are no blocks in front of the head
1034 * with the same cycle number as the head. This can happen
1035 * because we allow multiple outstanding log writes concurrently,
1036 * and the later writes might make it out before earlier ones.
1038 * We use the lsn from before modifying it so that we'll never
1039 * overwrite the unmount record after a clean unmount.
1041 * Do this only if we are going to recover the filesystem
1043 * NOTE: This used to say "if (!readonly)"
1044 * However on Linux, we can & do recover a read-only filesystem.
1045 * We only skip recovery if NORECOVERY is specified on mount,
1046 * in which case we would not be here.
1048 * But... if the -device- itself is readonly, just skip this.
1049 * We can't recover this device anyway, so it won't matter.
1051 if (!xfs_readonly_buftarg(log
->l_mp
->m_logdev_targp
))
1052 error
= xlog_clear_stale_blocks(log
, tail_lsn
);
1058 xfs_warn(log
->l_mp
, "failed to locate log tail");
1063 * Is the log zeroed at all?
1065 * The last binary search should be changed to perform an X block read
1066 * once X becomes small enough. You can then search linearly through
1067 * the X blocks. This will cut down on the number of reads we need to do.
1069 * If the log is partially zeroed, this routine will pass back the blkno
1070 * of the first block with cycle number 0. It won't have a complete LR
1074 * 0 => the log is completely written to
1075 * -1 => use *blk_no as the first block of the log
1076 * >0 => error has occurred
1081 xfs_daddr_t
*blk_no
)
1085 uint first_cycle
, last_cycle
;
1086 xfs_daddr_t new_blk
, last_blk
, start_blk
;
1087 xfs_daddr_t num_scan_bblks
;
1088 int error
, log_bbnum
= log
->l_logBBsize
;
1092 /* check totally zeroed log */
1093 bp
= xlog_get_bp(log
, 1);
1096 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
1100 first_cycle
= xlog_get_cycle(offset
);
1101 if (first_cycle
== 0) { /* completely zeroed log */
1107 /* check partially zeroed log */
1108 error
= xlog_bread(log
, log_bbnum
-1, 1, bp
, &offset
);
1112 last_cycle
= xlog_get_cycle(offset
);
1113 if (last_cycle
!= 0) { /* log completely written to */
1116 } else if (first_cycle
!= 1) {
1118 * If the cycle of the last block is zero, the cycle of
1119 * the first block must be 1. If it's not, maybe we're
1120 * not looking at a log... Bail out.
1123 "Log inconsistent or not a log (last==0, first!=1)");
1124 return XFS_ERROR(EINVAL
);
1127 /* we have a partially zeroed log */
1128 last_blk
= log_bbnum
-1;
1129 if ((error
= xlog_find_cycle_start(log
, bp
, 0, &last_blk
, 0)))
1133 * Validate the answer. Because there is no way to guarantee that
1134 * the entire log is made up of log records which are the same size,
1135 * we scan over the defined maximum blocks. At this point, the maximum
1136 * is not chosen to mean anything special. XXXmiken
1138 num_scan_bblks
= XLOG_TOTAL_REC_SHIFT(log
);
1139 ASSERT(num_scan_bblks
<= INT_MAX
);
1141 if (last_blk
< num_scan_bblks
)
1142 num_scan_bblks
= last_blk
;
1143 start_blk
= last_blk
- num_scan_bblks
;
1146 * We search for any instances of cycle number 0 that occur before
1147 * our current estimate of the head. What we're trying to detect is
1148 * 1 ... | 0 | 1 | 0...
1149 * ^ binary search ends here
1151 if ((error
= xlog_find_verify_cycle(log
, start_blk
,
1152 (int)num_scan_bblks
, 0, &new_blk
)))
1158 * Potentially backup over partial log record write. We don't need
1159 * to search the end of the log because we know it is zero.
1161 if ((error
= xlog_find_verify_log_record(log
, start_blk
,
1162 &last_blk
, 0)) == -1) {
1163 error
= XFS_ERROR(EIO
);
1177 * These are simple subroutines used by xlog_clear_stale_blocks() below
1178 * to initialize a buffer full of empty log record headers and write
1179 * them into the log.
1190 xlog_rec_header_t
*recp
= (xlog_rec_header_t
*)buf
;
1192 memset(buf
, 0, BBSIZE
);
1193 recp
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1194 recp
->h_cycle
= cpu_to_be32(cycle
);
1195 recp
->h_version
= cpu_to_be32(
1196 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1197 recp
->h_lsn
= cpu_to_be64(xlog_assign_lsn(cycle
, block
));
1198 recp
->h_tail_lsn
= cpu_to_be64(xlog_assign_lsn(tail_cycle
, tail_block
));
1199 recp
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1200 memcpy(&recp
->h_fs_uuid
, &log
->l_mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1204 xlog_write_log_records(
1215 int sectbb
= log
->l_sectBBsize
;
1216 int end_block
= start_block
+ blocks
;
1222 * Greedily allocate a buffer big enough to handle the full
1223 * range of basic blocks to be written. If that fails, try
1224 * a smaller size. We need to be able to write at least a
1225 * log sector, or we're out of luck.
1227 bufblks
= 1 << ffs(blocks
);
1228 while (!(bp
= xlog_get_bp(log
, bufblks
))) {
1230 if (bufblks
< sectbb
)
1234 /* We may need to do a read at the start to fill in part of
1235 * the buffer in the starting sector not covered by the first
1238 balign
= round_down(start_block
, sectbb
);
1239 if (balign
!= start_block
) {
1240 error
= xlog_bread_noalign(log
, start_block
, 1, bp
);
1244 j
= start_block
- balign
;
1247 for (i
= start_block
; i
< end_block
; i
+= bufblks
) {
1248 int bcount
, endcount
;
1250 bcount
= min(bufblks
, end_block
- start_block
);
1251 endcount
= bcount
- j
;
1253 /* We may need to do a read at the end to fill in part of
1254 * the buffer in the final sector not covered by the write.
1255 * If this is the same sector as the above read, skip it.
1257 ealign
= round_down(end_block
, sectbb
);
1258 if (j
== 0 && (start_block
+ endcount
> ealign
)) {
1259 offset
= bp
->b_addr
+ BBTOB(ealign
- start_block
);
1260 error
= xlog_bread_offset(log
, ealign
, sectbb
,
1267 offset
= xlog_align(log
, start_block
, endcount
, bp
);
1268 for (; j
< endcount
; j
++) {
1269 xlog_add_record(log
, offset
, cycle
, i
+j
,
1270 tail_cycle
, tail_block
);
1273 error
= xlog_bwrite(log
, start_block
, endcount
, bp
);
1276 start_block
+= endcount
;
1286 * This routine is called to blow away any incomplete log writes out
1287 * in front of the log head. We do this so that we won't become confused
1288 * if we come up, write only a little bit more, and then crash again.
1289 * If we leave the partial log records out there, this situation could
1290 * cause us to think those partial writes are valid blocks since they
1291 * have the current cycle number. We get rid of them by overwriting them
1292 * with empty log records with the old cycle number rather than the
1295 * The tail lsn is passed in rather than taken from
1296 * the log so that we will not write over the unmount record after a
1297 * clean unmount in a 512 block log. Doing so would leave the log without
1298 * any valid log records in it until a new one was written. If we crashed
1299 * during that time we would not be able to recover.
1302 xlog_clear_stale_blocks(
1306 int tail_cycle
, head_cycle
;
1307 int tail_block
, head_block
;
1308 int tail_distance
, max_distance
;
1312 tail_cycle
= CYCLE_LSN(tail_lsn
);
1313 tail_block
= BLOCK_LSN(tail_lsn
);
1314 head_cycle
= log
->l_curr_cycle
;
1315 head_block
= log
->l_curr_block
;
1318 * Figure out the distance between the new head of the log
1319 * and the tail. We want to write over any blocks beyond the
1320 * head that we may have written just before the crash, but
1321 * we don't want to overwrite the tail of the log.
1323 if (head_cycle
== tail_cycle
) {
1325 * The tail is behind the head in the physical log,
1326 * so the distance from the head to the tail is the
1327 * distance from the head to the end of the log plus
1328 * the distance from the beginning of the log to the
1331 if (unlikely(head_block
< tail_block
|| head_block
>= log
->l_logBBsize
)) {
1332 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1333 XFS_ERRLEVEL_LOW
, log
->l_mp
);
1334 return XFS_ERROR(EFSCORRUPTED
);
1336 tail_distance
= tail_block
+ (log
->l_logBBsize
- head_block
);
1339 * The head is behind the tail in the physical log,
1340 * so the distance from the head to the tail is just
1341 * the tail block minus the head block.
1343 if (unlikely(head_block
>= tail_block
|| head_cycle
!= (tail_cycle
+ 1))){
1344 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1345 XFS_ERRLEVEL_LOW
, log
->l_mp
);
1346 return XFS_ERROR(EFSCORRUPTED
);
1348 tail_distance
= tail_block
- head_block
;
1352 * If the head is right up against the tail, we can't clear
1355 if (tail_distance
<= 0) {
1356 ASSERT(tail_distance
== 0);
1360 max_distance
= XLOG_TOTAL_REC_SHIFT(log
);
1362 * Take the smaller of the maximum amount of outstanding I/O
1363 * we could have and the distance to the tail to clear out.
1364 * We take the smaller so that we don't overwrite the tail and
1365 * we don't waste all day writing from the head to the tail
1368 max_distance
= MIN(max_distance
, tail_distance
);
1370 if ((head_block
+ max_distance
) <= log
->l_logBBsize
) {
1372 * We can stomp all the blocks we need to without
1373 * wrapping around the end of the log. Just do it
1374 * in a single write. Use the cycle number of the
1375 * current cycle minus one so that the log will look like:
1378 error
= xlog_write_log_records(log
, (head_cycle
- 1),
1379 head_block
, max_distance
, tail_cycle
,
1385 * We need to wrap around the end of the physical log in
1386 * order to clear all the blocks. Do it in two separate
1387 * I/Os. The first write should be from the head to the
1388 * end of the physical log, and it should use the current
1389 * cycle number minus one just like above.
1391 distance
= log
->l_logBBsize
- head_block
;
1392 error
= xlog_write_log_records(log
, (head_cycle
- 1),
1393 head_block
, distance
, tail_cycle
,
1400 * Now write the blocks at the start of the physical log.
1401 * This writes the remainder of the blocks we want to clear.
1402 * It uses the current cycle number since we're now on the
1403 * same cycle as the head so that we get:
1404 * n ... n ... | n - 1 ...
1405 * ^^^^^ blocks we're writing
1407 distance
= max_distance
- (log
->l_logBBsize
- head_block
);
1408 error
= xlog_write_log_records(log
, head_cycle
, 0, distance
,
1409 tail_cycle
, tail_block
);
1417 /******************************************************************************
1419 * Log recover routines
1421 ******************************************************************************
1424 STATIC xlog_recover_t
*
1425 xlog_recover_find_tid(
1426 struct hlist_head
*head
,
1429 xlog_recover_t
*trans
;
1430 struct hlist_node
*n
;
1432 hlist_for_each_entry(trans
, n
, head
, r_list
) {
1433 if (trans
->r_log_tid
== tid
)
1440 xlog_recover_new_tid(
1441 struct hlist_head
*head
,
1445 xlog_recover_t
*trans
;
1447 trans
= kmem_zalloc(sizeof(xlog_recover_t
), KM_SLEEP
);
1448 trans
->r_log_tid
= tid
;
1450 INIT_LIST_HEAD(&trans
->r_itemq
);
1452 INIT_HLIST_NODE(&trans
->r_list
);
1453 hlist_add_head(&trans
->r_list
, head
);
1457 xlog_recover_add_item(
1458 struct list_head
*head
)
1460 xlog_recover_item_t
*item
;
1462 item
= kmem_zalloc(sizeof(xlog_recover_item_t
), KM_SLEEP
);
1463 INIT_LIST_HEAD(&item
->ri_list
);
1464 list_add_tail(&item
->ri_list
, head
);
1468 xlog_recover_add_to_cont_trans(
1470 xlog_recover_t
*trans
,
1474 xlog_recover_item_t
*item
;
1475 xfs_caddr_t ptr
, old_ptr
;
1478 if (list_empty(&trans
->r_itemq
)) {
1479 /* finish copying rest of trans header */
1480 xlog_recover_add_item(&trans
->r_itemq
);
1481 ptr
= (xfs_caddr_t
) &trans
->r_theader
+
1482 sizeof(xfs_trans_header_t
) - len
;
1483 memcpy(ptr
, dp
, len
); /* d, s, l */
1486 /* take the tail entry */
1487 item
= list_entry(trans
->r_itemq
.prev
, xlog_recover_item_t
, ri_list
);
1489 old_ptr
= item
->ri_buf
[item
->ri_cnt
-1].i_addr
;
1490 old_len
= item
->ri_buf
[item
->ri_cnt
-1].i_len
;
1492 ptr
= kmem_realloc(old_ptr
, len
+old_len
, old_len
, KM_SLEEP
);
1493 memcpy(&ptr
[old_len
], dp
, len
); /* d, s, l */
1494 item
->ri_buf
[item
->ri_cnt
-1].i_len
+= len
;
1495 item
->ri_buf
[item
->ri_cnt
-1].i_addr
= ptr
;
1496 trace_xfs_log_recover_item_add_cont(log
, trans
, item
, 0);
1501 * The next region to add is the start of a new region. It could be
1502 * a whole region or it could be the first part of a new region. Because
1503 * of this, the assumption here is that the type and size fields of all
1504 * format structures fit into the first 32 bits of the structure.
1506 * This works because all regions must be 32 bit aligned. Therefore, we
1507 * either have both fields or we have neither field. In the case we have
1508 * neither field, the data part of the region is zero length. We only have
1509 * a log_op_header and can throw away the header since a new one will appear
1510 * later. If we have at least 4 bytes, then we can determine how many regions
1511 * will appear in the current log item.
1514 xlog_recover_add_to_trans(
1516 xlog_recover_t
*trans
,
1520 xfs_inode_log_format_t
*in_f
; /* any will do */
1521 xlog_recover_item_t
*item
;
1526 if (list_empty(&trans
->r_itemq
)) {
1527 /* we need to catch log corruptions here */
1528 if (*(uint
*)dp
!= XFS_TRANS_HEADER_MAGIC
) {
1529 xfs_warn(log
->l_mp
, "%s: bad header magic number",
1532 return XFS_ERROR(EIO
);
1534 if (len
== sizeof(xfs_trans_header_t
))
1535 xlog_recover_add_item(&trans
->r_itemq
);
1536 memcpy(&trans
->r_theader
, dp
, len
); /* d, s, l */
1540 ptr
= kmem_alloc(len
, KM_SLEEP
);
1541 memcpy(ptr
, dp
, len
);
1542 in_f
= (xfs_inode_log_format_t
*)ptr
;
1544 /* take the tail entry */
1545 item
= list_entry(trans
->r_itemq
.prev
, xlog_recover_item_t
, ri_list
);
1546 if (item
->ri_total
!= 0 &&
1547 item
->ri_total
== item
->ri_cnt
) {
1548 /* tail item is in use, get a new one */
1549 xlog_recover_add_item(&trans
->r_itemq
);
1550 item
= list_entry(trans
->r_itemq
.prev
,
1551 xlog_recover_item_t
, ri_list
);
1554 if (item
->ri_total
== 0) { /* first region to be added */
1555 if (in_f
->ilf_size
== 0 ||
1556 in_f
->ilf_size
> XLOG_MAX_REGIONS_IN_ITEM
) {
1558 "bad number of regions (%d) in inode log format",
1561 return XFS_ERROR(EIO
);
1564 item
->ri_total
= in_f
->ilf_size
;
1566 kmem_zalloc(item
->ri_total
* sizeof(xfs_log_iovec_t
),
1569 ASSERT(item
->ri_total
> item
->ri_cnt
);
1570 /* Description region is ri_buf[0] */
1571 item
->ri_buf
[item
->ri_cnt
].i_addr
= ptr
;
1572 item
->ri_buf
[item
->ri_cnt
].i_len
= len
;
1574 trace_xfs_log_recover_item_add(log
, trans
, item
, 0);
1579 * Sort the log items in the transaction. Cancelled buffers need
1580 * to be put first so they are processed before any items that might
1581 * modify the buffers. If they are cancelled, then the modifications
1582 * don't need to be replayed.
1585 xlog_recover_reorder_trans(
1587 xlog_recover_t
*trans
,
1590 xlog_recover_item_t
*item
, *n
;
1591 LIST_HEAD(sort_list
);
1593 list_splice_init(&trans
->r_itemq
, &sort_list
);
1594 list_for_each_entry_safe(item
, n
, &sort_list
, ri_list
) {
1595 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
1597 switch (ITEM_TYPE(item
)) {
1599 if (!(buf_f
->blf_flags
& XFS_BLF_CANCEL
)) {
1600 trace_xfs_log_recover_item_reorder_head(log
,
1602 list_move(&item
->ri_list
, &trans
->r_itemq
);
1607 case XFS_LI_QUOTAOFF
:
1610 trace_xfs_log_recover_item_reorder_tail(log
,
1612 list_move_tail(&item
->ri_list
, &trans
->r_itemq
);
1616 "%s: unrecognized type of log operation",
1619 return XFS_ERROR(EIO
);
1622 ASSERT(list_empty(&sort_list
));
1627 * Build up the table of buf cancel records so that we don't replay
1628 * cancelled data in the second pass. For buffer records that are
1629 * not cancel records, there is nothing to do here so we just return.
1631 * If we get a cancel record which is already in the table, this indicates
1632 * that the buffer was cancelled multiple times. In order to ensure
1633 * that during pass 2 we keep the record in the table until we reach its
1634 * last occurrence in the log, we keep a reference count in the cancel
1635 * record in the table to tell us how many times we expect to see this
1636 * record during the second pass.
1639 xlog_recover_buffer_pass1(
1641 xlog_recover_item_t
*item
)
1643 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
1644 struct list_head
*bucket
;
1645 struct xfs_buf_cancel
*bcp
;
1648 * If this isn't a cancel buffer item, then just return.
1650 if (!(buf_f
->blf_flags
& XFS_BLF_CANCEL
)) {
1651 trace_xfs_log_recover_buf_not_cancel(log
, buf_f
);
1656 * Insert an xfs_buf_cancel record into the hash table of them.
1657 * If there is already an identical record, bump its reference count.
1659 bucket
= XLOG_BUF_CANCEL_BUCKET(log
, buf_f
->blf_blkno
);
1660 list_for_each_entry(bcp
, bucket
, bc_list
) {
1661 if (bcp
->bc_blkno
== buf_f
->blf_blkno
&&
1662 bcp
->bc_len
== buf_f
->blf_len
) {
1664 trace_xfs_log_recover_buf_cancel_ref_inc(log
, buf_f
);
1669 bcp
= kmem_alloc(sizeof(struct xfs_buf_cancel
), KM_SLEEP
);
1670 bcp
->bc_blkno
= buf_f
->blf_blkno
;
1671 bcp
->bc_len
= buf_f
->blf_len
;
1672 bcp
->bc_refcount
= 1;
1673 list_add_tail(&bcp
->bc_list
, bucket
);
1675 trace_xfs_log_recover_buf_cancel_add(log
, buf_f
);
1680 * Check to see whether the buffer being recovered has a corresponding
1681 * entry in the buffer cancel record table. If it does then return 1
1682 * so that it will be cancelled, otherwise return 0. If the buffer is
1683 * actually a buffer cancel item (XFS_BLF_CANCEL is set), then decrement
1684 * the refcount on the entry in the table and remove it from the table
1685 * if this is the last reference.
1687 * We remove the cancel record from the table when we encounter its
1688 * last occurrence in the log so that if the same buffer is re-used
1689 * again after its last cancellation we actually replay the changes
1690 * made at that point.
1693 xlog_check_buffer_cancelled(
1699 struct list_head
*bucket
;
1700 struct xfs_buf_cancel
*bcp
;
1702 if (log
->l_buf_cancel_table
== NULL
) {
1704 * There is nothing in the table built in pass one,
1705 * so this buffer must not be cancelled.
1707 ASSERT(!(flags
& XFS_BLF_CANCEL
));
1712 * Search for an entry in the cancel table that matches our buffer.
1714 bucket
= XLOG_BUF_CANCEL_BUCKET(log
, blkno
);
1715 list_for_each_entry(bcp
, bucket
, bc_list
) {
1716 if (bcp
->bc_blkno
== blkno
&& bcp
->bc_len
== len
)
1721 * We didn't find a corresponding entry in the table, so return 0 so
1722 * that the buffer is NOT cancelled.
1724 ASSERT(!(flags
& XFS_BLF_CANCEL
));
1729 * We've go a match, so return 1 so that the recovery of this buffer
1730 * is cancelled. If this buffer is actually a buffer cancel log
1731 * item, then decrement the refcount on the one in the table and
1732 * remove it if this is the last reference.
1734 if (flags
& XFS_BLF_CANCEL
) {
1735 if (--bcp
->bc_refcount
== 0) {
1736 list_del(&bcp
->bc_list
);
1744 * Perform recovery for a buffer full of inodes. In these buffers, the only
1745 * data which should be recovered is that which corresponds to the
1746 * di_next_unlinked pointers in the on disk inode structures. The rest of the
1747 * data for the inodes is always logged through the inodes themselves rather
1748 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1750 * The only time when buffers full of inodes are fully recovered is when the
1751 * buffer is full of newly allocated inodes. In this case the buffer will
1752 * not be marked as an inode buffer and so will be sent to
1753 * xlog_recover_do_reg_buffer() below during recovery.
1756 xlog_recover_do_inode_buffer(
1757 struct xfs_mount
*mp
,
1758 xlog_recover_item_t
*item
,
1760 xfs_buf_log_format_t
*buf_f
)
1766 int reg_buf_offset
= 0;
1767 int reg_buf_bytes
= 0;
1768 int next_unlinked_offset
;
1770 xfs_agino_t
*logged_nextp
;
1771 xfs_agino_t
*buffer_nextp
;
1773 trace_xfs_log_recover_buf_inode_buf(mp
->m_log
, buf_f
);
1775 inodes_per_buf
= XFS_BUF_COUNT(bp
) >> mp
->m_sb
.sb_inodelog
;
1776 for (i
= 0; i
< inodes_per_buf
; i
++) {
1777 next_unlinked_offset
= (i
* mp
->m_sb
.sb_inodesize
) +
1778 offsetof(xfs_dinode_t
, di_next_unlinked
);
1780 while (next_unlinked_offset
>=
1781 (reg_buf_offset
+ reg_buf_bytes
)) {
1783 * The next di_next_unlinked field is beyond
1784 * the current logged region. Find the next
1785 * logged region that contains or is beyond
1786 * the current di_next_unlinked field.
1789 bit
= xfs_next_bit(buf_f
->blf_data_map
,
1790 buf_f
->blf_map_size
, bit
);
1793 * If there are no more logged regions in the
1794 * buffer, then we're done.
1799 nbits
= xfs_contig_bits(buf_f
->blf_data_map
,
1800 buf_f
->blf_map_size
, bit
);
1802 reg_buf_offset
= bit
<< XFS_BLF_SHIFT
;
1803 reg_buf_bytes
= nbits
<< XFS_BLF_SHIFT
;
1808 * If the current logged region starts after the current
1809 * di_next_unlinked field, then move on to the next
1810 * di_next_unlinked field.
1812 if (next_unlinked_offset
< reg_buf_offset
)
1815 ASSERT(item
->ri_buf
[item_index
].i_addr
!= NULL
);
1816 ASSERT((item
->ri_buf
[item_index
].i_len
% XFS_BLF_CHUNK
) == 0);
1817 ASSERT((reg_buf_offset
+ reg_buf_bytes
) <= XFS_BUF_COUNT(bp
));
1820 * The current logged region contains a copy of the
1821 * current di_next_unlinked field. Extract its value
1822 * and copy it to the buffer copy.
1824 logged_nextp
= item
->ri_buf
[item_index
].i_addr
+
1825 next_unlinked_offset
- reg_buf_offset
;
1826 if (unlikely(*logged_nextp
== 0)) {
1828 "Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1829 "Trying to replay bad (0) inode di_next_unlinked field.",
1831 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1832 XFS_ERRLEVEL_LOW
, mp
);
1833 return XFS_ERROR(EFSCORRUPTED
);
1836 buffer_nextp
= (xfs_agino_t
*)xfs_buf_offset(bp
,
1837 next_unlinked_offset
);
1838 *buffer_nextp
= *logged_nextp
;
1845 * Perform a 'normal' buffer recovery. Each logged region of the
1846 * buffer should be copied over the corresponding region in the
1847 * given buffer. The bitmap in the buf log format structure indicates
1848 * where to place the logged data.
1851 xlog_recover_do_reg_buffer(
1852 struct xfs_mount
*mp
,
1853 xlog_recover_item_t
*item
,
1855 xfs_buf_log_format_t
*buf_f
)
1862 trace_xfs_log_recover_buf_reg_buf(mp
->m_log
, buf_f
);
1865 i
= 1; /* 0 is the buf format structure */
1867 bit
= xfs_next_bit(buf_f
->blf_data_map
,
1868 buf_f
->blf_map_size
, bit
);
1871 nbits
= xfs_contig_bits(buf_f
->blf_data_map
,
1872 buf_f
->blf_map_size
, bit
);
1874 ASSERT(item
->ri_buf
[i
].i_addr
!= NULL
);
1875 ASSERT(item
->ri_buf
[i
].i_len
% XFS_BLF_CHUNK
== 0);
1876 ASSERT(XFS_BUF_COUNT(bp
) >=
1877 ((uint
)bit
<< XFS_BLF_SHIFT
)+(nbits
<<XFS_BLF_SHIFT
));
1880 * Do a sanity check if this is a dquot buffer. Just checking
1881 * the first dquot in the buffer should do. XXXThis is
1882 * probably a good thing to do for other buf types also.
1885 if (buf_f
->blf_flags
&
1886 (XFS_BLF_UDQUOT_BUF
|XFS_BLF_PDQUOT_BUF
|XFS_BLF_GDQUOT_BUF
)) {
1887 if (item
->ri_buf
[i
].i_addr
== NULL
) {
1889 "XFS: NULL dquot in %s.", __func__
);
1892 if (item
->ri_buf
[i
].i_len
< sizeof(xfs_disk_dquot_t
)) {
1894 "XFS: dquot too small (%d) in %s.",
1895 item
->ri_buf
[i
].i_len
, __func__
);
1898 error
= xfs_qm_dqcheck(mp
, item
->ri_buf
[i
].i_addr
,
1899 -1, 0, XFS_QMOPT_DOWARN
,
1900 "dquot_buf_recover");
1905 memcpy(xfs_buf_offset(bp
,
1906 (uint
)bit
<< XFS_BLF_SHIFT
), /* dest */
1907 item
->ri_buf
[i
].i_addr
, /* source */
1908 nbits
<<XFS_BLF_SHIFT
); /* length */
1914 /* Shouldn't be any more regions */
1915 ASSERT(i
== item
->ri_total
);
1919 * Do some primitive error checking on ondisk dquot data structures.
1923 struct xfs_mount
*mp
,
1924 xfs_disk_dquot_t
*ddq
,
1926 uint type
, /* used only when IO_dorepair is true */
1930 xfs_dqblk_t
*d
= (xfs_dqblk_t
*)ddq
;
1934 * We can encounter an uninitialized dquot buffer for 2 reasons:
1935 * 1. If we crash while deleting the quotainode(s), and those blks got
1936 * used for user data. This is because we take the path of regular
1937 * file deletion; however, the size field of quotainodes is never
1938 * updated, so all the tricks that we play in itruncate_finish
1939 * don't quite matter.
1941 * 2. We don't play the quota buffers when there's a quotaoff logitem.
1942 * But the allocation will be replayed so we'll end up with an
1943 * uninitialized quota block.
1945 * This is all fine; things are still consistent, and we haven't lost
1946 * any quota information. Just don't complain about bad dquot blks.
1948 if (ddq
->d_magic
!= cpu_to_be16(XFS_DQUOT_MAGIC
)) {
1949 if (flags
& XFS_QMOPT_DOWARN
)
1951 "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1952 str
, id
, be16_to_cpu(ddq
->d_magic
), XFS_DQUOT_MAGIC
);
1955 if (ddq
->d_version
!= XFS_DQUOT_VERSION
) {
1956 if (flags
& XFS_QMOPT_DOWARN
)
1958 "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1959 str
, id
, ddq
->d_version
, XFS_DQUOT_VERSION
);
1963 if (ddq
->d_flags
!= XFS_DQ_USER
&&
1964 ddq
->d_flags
!= XFS_DQ_PROJ
&&
1965 ddq
->d_flags
!= XFS_DQ_GROUP
) {
1966 if (flags
& XFS_QMOPT_DOWARN
)
1968 "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1969 str
, id
, ddq
->d_flags
);
1973 if (id
!= -1 && id
!= be32_to_cpu(ddq
->d_id
)) {
1974 if (flags
& XFS_QMOPT_DOWARN
)
1976 "%s : ondisk-dquot 0x%p, ID mismatch: "
1977 "0x%x expected, found id 0x%x",
1978 str
, ddq
, id
, be32_to_cpu(ddq
->d_id
));
1982 if (!errs
&& ddq
->d_id
) {
1983 if (ddq
->d_blk_softlimit
&&
1984 be64_to_cpu(ddq
->d_bcount
) >
1985 be64_to_cpu(ddq
->d_blk_softlimit
)) {
1986 if (!ddq
->d_btimer
) {
1987 if (flags
& XFS_QMOPT_DOWARN
)
1989 "%s : Dquot ID 0x%x (0x%p) BLK TIMER NOT STARTED",
1990 str
, (int)be32_to_cpu(ddq
->d_id
), ddq
);
1994 if (ddq
->d_ino_softlimit
&&
1995 be64_to_cpu(ddq
->d_icount
) >
1996 be64_to_cpu(ddq
->d_ino_softlimit
)) {
1997 if (!ddq
->d_itimer
) {
1998 if (flags
& XFS_QMOPT_DOWARN
)
2000 "%s : Dquot ID 0x%x (0x%p) INODE TIMER NOT STARTED",
2001 str
, (int)be32_to_cpu(ddq
->d_id
), ddq
);
2005 if (ddq
->d_rtb_softlimit
&&
2006 be64_to_cpu(ddq
->d_rtbcount
) >
2007 be64_to_cpu(ddq
->d_rtb_softlimit
)) {
2008 if (!ddq
->d_rtbtimer
) {
2009 if (flags
& XFS_QMOPT_DOWARN
)
2011 "%s : Dquot ID 0x%x (0x%p) RTBLK TIMER NOT STARTED",
2012 str
, (int)be32_to_cpu(ddq
->d_id
), ddq
);
2018 if (!errs
|| !(flags
& XFS_QMOPT_DQREPAIR
))
2021 if (flags
& XFS_QMOPT_DOWARN
)
2022 xfs_notice(mp
, "Re-initializing dquot ID 0x%x", id
);
2025 * Typically, a repair is only requested by quotacheck.
2028 ASSERT(flags
& XFS_QMOPT_DQREPAIR
);
2029 memset(d
, 0, sizeof(xfs_dqblk_t
));
2031 d
->dd_diskdq
.d_magic
= cpu_to_be16(XFS_DQUOT_MAGIC
);
2032 d
->dd_diskdq
.d_version
= XFS_DQUOT_VERSION
;
2033 d
->dd_diskdq
.d_flags
= type
;
2034 d
->dd_diskdq
.d_id
= cpu_to_be32(id
);
2040 * Perform a dquot buffer recovery.
2041 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2042 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2043 * Else, treat it as a regular buffer and do recovery.
2046 xlog_recover_do_dquot_buffer(
2049 xlog_recover_item_t
*item
,
2051 xfs_buf_log_format_t
*buf_f
)
2055 trace_xfs_log_recover_buf_dquot_buf(log
, buf_f
);
2058 * Filesystems are required to send in quota flags at mount time.
2060 if (mp
->m_qflags
== 0) {
2065 if (buf_f
->blf_flags
& XFS_BLF_UDQUOT_BUF
)
2066 type
|= XFS_DQ_USER
;
2067 if (buf_f
->blf_flags
& XFS_BLF_PDQUOT_BUF
)
2068 type
|= XFS_DQ_PROJ
;
2069 if (buf_f
->blf_flags
& XFS_BLF_GDQUOT_BUF
)
2070 type
|= XFS_DQ_GROUP
;
2072 * This type of quotas was turned off, so ignore this buffer
2074 if (log
->l_quotaoffs_flag
& type
)
2077 xlog_recover_do_reg_buffer(mp
, item
, bp
, buf_f
);
2081 * This routine replays a modification made to a buffer at runtime.
2082 * There are actually two types of buffer, regular and inode, which
2083 * are handled differently. Inode buffers are handled differently
2084 * in that we only recover a specific set of data from them, namely
2085 * the inode di_next_unlinked fields. This is because all other inode
2086 * data is actually logged via inode records and any data we replay
2087 * here which overlaps that may be stale.
2089 * When meta-data buffers are freed at run time we log a buffer item
2090 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2091 * of the buffer in the log should not be replayed at recovery time.
2092 * This is so that if the blocks covered by the buffer are reused for
2093 * file data before we crash we don't end up replaying old, freed
2094 * meta-data into a user's file.
2096 * To handle the cancellation of buffer log items, we make two passes
2097 * over the log during recovery. During the first we build a table of
2098 * those buffers which have been cancelled, and during the second we
2099 * only replay those buffers which do not have corresponding cancel
2100 * records in the table. See xlog_recover_do_buffer_pass[1,2] above
2101 * for more details on the implementation of the table of cancel records.
2104 xlog_recover_buffer_pass2(
2106 struct list_head
*buffer_list
,
2107 xlog_recover_item_t
*item
)
2109 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
2110 xfs_mount_t
*mp
= log
->l_mp
;
2116 * In this pass we only want to recover all the buffers which have
2117 * not been cancelled and are not cancellation buffers themselves.
2119 if (xlog_check_buffer_cancelled(log
, buf_f
->blf_blkno
,
2120 buf_f
->blf_len
, buf_f
->blf_flags
)) {
2121 trace_xfs_log_recover_buf_cancel(log
, buf_f
);
2125 trace_xfs_log_recover_buf_recover(log
, buf_f
);
2127 buf_flags
= XBF_LOCK
;
2128 if (!(buf_f
->blf_flags
& XFS_BLF_INODE_BUF
))
2129 buf_flags
|= XBF_MAPPED
;
2131 bp
= xfs_buf_read(mp
->m_ddev_targp
, buf_f
->blf_blkno
, buf_f
->blf_len
,
2134 return XFS_ERROR(ENOMEM
);
2135 error
= bp
->b_error
;
2137 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#1)");
2142 if (buf_f
->blf_flags
& XFS_BLF_INODE_BUF
) {
2143 error
= xlog_recover_do_inode_buffer(mp
, item
, bp
, buf_f
);
2144 } else if (buf_f
->blf_flags
&
2145 (XFS_BLF_UDQUOT_BUF
|XFS_BLF_PDQUOT_BUF
|XFS_BLF_GDQUOT_BUF
)) {
2146 xlog_recover_do_dquot_buffer(mp
, log
, item
, bp
, buf_f
);
2148 xlog_recover_do_reg_buffer(mp
, item
, bp
, buf_f
);
2151 return XFS_ERROR(error
);
2154 * Perform delayed write on the buffer. Asynchronous writes will be
2155 * slower when taking into account all the buffers to be flushed.
2157 * Also make sure that only inode buffers with good sizes stay in
2158 * the buffer cache. The kernel moves inodes in buffers of 1 block
2159 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger. The inode
2160 * buffers in the log can be a different size if the log was generated
2161 * by an older kernel using unclustered inode buffers or a newer kernel
2162 * running with a different inode cluster size. Regardless, if the
2163 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2164 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2165 * the buffer out of the buffer cache so that the buffer won't
2166 * overlap with future reads of those inodes.
2168 if (XFS_DINODE_MAGIC
==
2169 be16_to_cpu(*((__be16
*)xfs_buf_offset(bp
, 0))) &&
2170 (XFS_BUF_COUNT(bp
) != MAX(log
->l_mp
->m_sb
.sb_blocksize
,
2171 (__uint32_t
)XFS_INODE_CLUSTER_SIZE(log
->l_mp
)))) {
2173 error
= xfs_bwrite(bp
);
2175 ASSERT(bp
->b_target
->bt_mount
== mp
);
2176 bp
->b_iodone
= xlog_recover_iodone
;
2177 xfs_buf_delwri_queue(bp
, buffer_list
);
2185 xlog_recover_inode_pass2(
2187 struct list_head
*buffer_list
,
2188 xlog_recover_item_t
*item
)
2190 xfs_inode_log_format_t
*in_f
;
2191 xfs_mount_t
*mp
= log
->l_mp
;
2200 xfs_icdinode_t
*dicp
;
2203 if (item
->ri_buf
[0].i_len
== sizeof(xfs_inode_log_format_t
)) {
2204 in_f
= item
->ri_buf
[0].i_addr
;
2206 in_f
= kmem_alloc(sizeof(xfs_inode_log_format_t
), KM_SLEEP
);
2208 error
= xfs_inode_item_format_convert(&item
->ri_buf
[0], in_f
);
2214 * Inode buffers can be freed, look out for it,
2215 * and do not replay the inode.
2217 if (xlog_check_buffer_cancelled(log
, in_f
->ilf_blkno
,
2218 in_f
->ilf_len
, 0)) {
2220 trace_xfs_log_recover_inode_cancel(log
, in_f
);
2223 trace_xfs_log_recover_inode_recover(log
, in_f
);
2225 bp
= xfs_buf_read(mp
->m_ddev_targp
, in_f
->ilf_blkno
, in_f
->ilf_len
,
2231 error
= bp
->b_error
;
2233 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#2)");
2237 ASSERT(in_f
->ilf_fields
& XFS_ILOG_CORE
);
2238 dip
= (xfs_dinode_t
*)xfs_buf_offset(bp
, in_f
->ilf_boffset
);
2241 * Make sure the place we're flushing out to really looks
2244 if (unlikely(dip
->di_magic
!= cpu_to_be16(XFS_DINODE_MAGIC
))) {
2247 "%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2248 __func__
, dip
, bp
, in_f
->ilf_ino
);
2249 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2250 XFS_ERRLEVEL_LOW
, mp
);
2251 error
= EFSCORRUPTED
;
2254 dicp
= item
->ri_buf
[1].i_addr
;
2255 if (unlikely(dicp
->di_magic
!= XFS_DINODE_MAGIC
)) {
2258 "%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2259 __func__
, item
, in_f
->ilf_ino
);
2260 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2261 XFS_ERRLEVEL_LOW
, mp
);
2262 error
= EFSCORRUPTED
;
2266 /* Skip replay when the on disk inode is newer than the log one */
2267 if (dicp
->di_flushiter
< be16_to_cpu(dip
->di_flushiter
)) {
2269 * Deal with the wrap case, DI_MAX_FLUSH is less
2270 * than smaller numbers
2272 if (be16_to_cpu(dip
->di_flushiter
) == DI_MAX_FLUSH
&&
2273 dicp
->di_flushiter
< (DI_MAX_FLUSH
>> 1)) {
2277 trace_xfs_log_recover_inode_skip(log
, in_f
);
2282 /* Take the opportunity to reset the flush iteration count */
2283 dicp
->di_flushiter
= 0;
2285 if (unlikely(S_ISREG(dicp
->di_mode
))) {
2286 if ((dicp
->di_format
!= XFS_DINODE_FMT_EXTENTS
) &&
2287 (dicp
->di_format
!= XFS_DINODE_FMT_BTREE
)) {
2288 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2289 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2292 "%s: Bad regular inode log record, rec ptr 0x%p, "
2293 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2294 __func__
, item
, dip
, bp
, in_f
->ilf_ino
);
2295 error
= EFSCORRUPTED
;
2298 } else if (unlikely(S_ISDIR(dicp
->di_mode
))) {
2299 if ((dicp
->di_format
!= XFS_DINODE_FMT_EXTENTS
) &&
2300 (dicp
->di_format
!= XFS_DINODE_FMT_BTREE
) &&
2301 (dicp
->di_format
!= XFS_DINODE_FMT_LOCAL
)) {
2302 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2303 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2306 "%s: Bad dir inode log record, rec ptr 0x%p, "
2307 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2308 __func__
, item
, dip
, bp
, in_f
->ilf_ino
);
2309 error
= EFSCORRUPTED
;
2313 if (unlikely(dicp
->di_nextents
+ dicp
->di_anextents
> dicp
->di_nblocks
)){
2314 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2315 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2318 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2319 "dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2320 __func__
, item
, dip
, bp
, in_f
->ilf_ino
,
2321 dicp
->di_nextents
+ dicp
->di_anextents
,
2323 error
= EFSCORRUPTED
;
2326 if (unlikely(dicp
->di_forkoff
> mp
->m_sb
.sb_inodesize
)) {
2327 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2328 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2331 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2332 "dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__
,
2333 item
, dip
, bp
, in_f
->ilf_ino
, dicp
->di_forkoff
);
2334 error
= EFSCORRUPTED
;
2337 if (unlikely(item
->ri_buf
[1].i_len
> sizeof(struct xfs_icdinode
))) {
2338 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2339 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2342 "%s: Bad inode log record length %d, rec ptr 0x%p",
2343 __func__
, item
->ri_buf
[1].i_len
, item
);
2344 error
= EFSCORRUPTED
;
2348 /* The core is in in-core format */
2349 xfs_dinode_to_disk(dip
, item
->ri_buf
[1].i_addr
);
2351 /* the rest is in on-disk format */
2352 if (item
->ri_buf
[1].i_len
> sizeof(struct xfs_icdinode
)) {
2353 memcpy((xfs_caddr_t
) dip
+ sizeof(struct xfs_icdinode
),
2354 item
->ri_buf
[1].i_addr
+ sizeof(struct xfs_icdinode
),
2355 item
->ri_buf
[1].i_len
- sizeof(struct xfs_icdinode
));
2358 fields
= in_f
->ilf_fields
;
2359 switch (fields
& (XFS_ILOG_DEV
| XFS_ILOG_UUID
)) {
2361 xfs_dinode_put_rdev(dip
, in_f
->ilf_u
.ilfu_rdev
);
2364 memcpy(XFS_DFORK_DPTR(dip
),
2365 &in_f
->ilf_u
.ilfu_uuid
,
2370 if (in_f
->ilf_size
== 2)
2371 goto write_inode_buffer
;
2372 len
= item
->ri_buf
[2].i_len
;
2373 src
= item
->ri_buf
[2].i_addr
;
2374 ASSERT(in_f
->ilf_size
<= 4);
2375 ASSERT((in_f
->ilf_size
== 3) || (fields
& XFS_ILOG_AFORK
));
2376 ASSERT(!(fields
& XFS_ILOG_DFORK
) ||
2377 (len
== in_f
->ilf_dsize
));
2379 switch (fields
& XFS_ILOG_DFORK
) {
2380 case XFS_ILOG_DDATA
:
2382 memcpy(XFS_DFORK_DPTR(dip
), src
, len
);
2385 case XFS_ILOG_DBROOT
:
2386 xfs_bmbt_to_bmdr(mp
, (struct xfs_btree_block
*)src
, len
,
2387 (xfs_bmdr_block_t
*)XFS_DFORK_DPTR(dip
),
2388 XFS_DFORK_DSIZE(dip
, mp
));
2393 * There are no data fork flags set.
2395 ASSERT((fields
& XFS_ILOG_DFORK
) == 0);
2400 * If we logged any attribute data, recover it. There may or
2401 * may not have been any other non-core data logged in this
2404 if (in_f
->ilf_fields
& XFS_ILOG_AFORK
) {
2405 if (in_f
->ilf_fields
& XFS_ILOG_DFORK
) {
2410 len
= item
->ri_buf
[attr_index
].i_len
;
2411 src
= item
->ri_buf
[attr_index
].i_addr
;
2412 ASSERT(len
== in_f
->ilf_asize
);
2414 switch (in_f
->ilf_fields
& XFS_ILOG_AFORK
) {
2415 case XFS_ILOG_ADATA
:
2417 dest
= XFS_DFORK_APTR(dip
);
2418 ASSERT(len
<= XFS_DFORK_ASIZE(dip
, mp
));
2419 memcpy(dest
, src
, len
);
2422 case XFS_ILOG_ABROOT
:
2423 dest
= XFS_DFORK_APTR(dip
);
2424 xfs_bmbt_to_bmdr(mp
, (struct xfs_btree_block
*)src
,
2425 len
, (xfs_bmdr_block_t
*)dest
,
2426 XFS_DFORK_ASIZE(dip
, mp
));
2430 xfs_warn(log
->l_mp
, "%s: Invalid flag", __func__
);
2439 ASSERT(bp
->b_target
->bt_mount
== mp
);
2440 bp
->b_iodone
= xlog_recover_iodone
;
2441 xfs_buf_delwri_queue(bp
, buffer_list
);
2446 return XFS_ERROR(error
);
2450 * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2451 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2455 xlog_recover_quotaoff_pass1(
2457 xlog_recover_item_t
*item
)
2459 xfs_qoff_logformat_t
*qoff_f
= item
->ri_buf
[0].i_addr
;
2463 * The logitem format's flag tells us if this was user quotaoff,
2464 * group/project quotaoff or both.
2466 if (qoff_f
->qf_flags
& XFS_UQUOTA_ACCT
)
2467 log
->l_quotaoffs_flag
|= XFS_DQ_USER
;
2468 if (qoff_f
->qf_flags
& XFS_PQUOTA_ACCT
)
2469 log
->l_quotaoffs_flag
|= XFS_DQ_PROJ
;
2470 if (qoff_f
->qf_flags
& XFS_GQUOTA_ACCT
)
2471 log
->l_quotaoffs_flag
|= XFS_DQ_GROUP
;
2477 * Recover a dquot record
2480 xlog_recover_dquot_pass2(
2482 struct list_head
*buffer_list
,
2483 xlog_recover_item_t
*item
)
2485 xfs_mount_t
*mp
= log
->l_mp
;
2487 struct xfs_disk_dquot
*ddq
, *recddq
;
2489 xfs_dq_logformat_t
*dq_f
;
2494 * Filesystems are required to send in quota flags at mount time.
2496 if (mp
->m_qflags
== 0)
2499 recddq
= item
->ri_buf
[1].i_addr
;
2500 if (recddq
== NULL
) {
2501 xfs_alert(log
->l_mp
, "NULL dquot in %s.", __func__
);
2502 return XFS_ERROR(EIO
);
2504 if (item
->ri_buf
[1].i_len
< sizeof(xfs_disk_dquot_t
)) {
2505 xfs_alert(log
->l_mp
, "dquot too small (%d) in %s.",
2506 item
->ri_buf
[1].i_len
, __func__
);
2507 return XFS_ERROR(EIO
);
2511 * This type of quotas was turned off, so ignore this record.
2513 type
= recddq
->d_flags
& (XFS_DQ_USER
| XFS_DQ_PROJ
| XFS_DQ_GROUP
);
2515 if (log
->l_quotaoffs_flag
& type
)
2519 * At this point we know that quota was _not_ turned off.
2520 * Since the mount flags are not indicating to us otherwise, this
2521 * must mean that quota is on, and the dquot needs to be replayed.
2522 * Remember that we may not have fully recovered the superblock yet,
2523 * so we can't do the usual trick of looking at the SB quota bits.
2525 * The other possibility, of course, is that the quota subsystem was
2526 * removed since the last mount - ENOSYS.
2528 dq_f
= item
->ri_buf
[0].i_addr
;
2530 error
= xfs_qm_dqcheck(mp
, recddq
, dq_f
->qlf_id
, 0, XFS_QMOPT_DOWARN
,
2531 "xlog_recover_dquot_pass2 (log copy)");
2533 return XFS_ERROR(EIO
);
2534 ASSERT(dq_f
->qlf_len
== 1);
2536 error
= xfs_read_buf(mp
, mp
->m_ddev_targp
,
2538 XFS_FSB_TO_BB(mp
, dq_f
->qlf_len
),
2541 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#3)");
2545 ddq
= (xfs_disk_dquot_t
*)xfs_buf_offset(bp
, dq_f
->qlf_boffset
);
2548 * At least the magic num portion should be on disk because this
2549 * was among a chunk of dquots created earlier, and we did some
2550 * minimal initialization then.
2552 error
= xfs_qm_dqcheck(mp
, ddq
, dq_f
->qlf_id
, 0, XFS_QMOPT_DOWARN
,
2553 "xlog_recover_dquot_pass2");
2556 return XFS_ERROR(EIO
);
2559 memcpy(ddq
, recddq
, item
->ri_buf
[1].i_len
);
2561 ASSERT(dq_f
->qlf_size
== 2);
2562 ASSERT(bp
->b_target
->bt_mount
== mp
);
2563 bp
->b_iodone
= xlog_recover_iodone
;
2564 xfs_buf_delwri_queue(bp
, buffer_list
);
2571 * This routine is called to create an in-core extent free intent
2572 * item from the efi format structure which was logged on disk.
2573 * It allocates an in-core efi, copies the extents from the format
2574 * structure into it, and adds the efi to the AIL with the given
2578 xlog_recover_efi_pass2(
2580 xlog_recover_item_t
*item
,
2584 xfs_mount_t
*mp
= log
->l_mp
;
2585 xfs_efi_log_item_t
*efip
;
2586 xfs_efi_log_format_t
*efi_formatp
;
2588 efi_formatp
= item
->ri_buf
[0].i_addr
;
2590 efip
= xfs_efi_init(mp
, efi_formatp
->efi_nextents
);
2591 if ((error
= xfs_efi_copy_format(&(item
->ri_buf
[0]),
2592 &(efip
->efi_format
)))) {
2593 xfs_efi_item_free(efip
);
2596 atomic_set(&efip
->efi_next_extent
, efi_formatp
->efi_nextents
);
2598 spin_lock(&log
->l_ailp
->xa_lock
);
2600 * xfs_trans_ail_update() drops the AIL lock.
2602 xfs_trans_ail_update(log
->l_ailp
, &efip
->efi_item
, lsn
);
2608 * This routine is called when an efd format structure is found in
2609 * a committed transaction in the log. It's purpose is to cancel
2610 * the corresponding efi if it was still in the log. To do this
2611 * it searches the AIL for the efi with an id equal to that in the
2612 * efd format structure. If we find it, we remove the efi from the
2616 xlog_recover_efd_pass2(
2618 xlog_recover_item_t
*item
)
2620 xfs_efd_log_format_t
*efd_formatp
;
2621 xfs_efi_log_item_t
*efip
= NULL
;
2622 xfs_log_item_t
*lip
;
2624 struct xfs_ail_cursor cur
;
2625 struct xfs_ail
*ailp
= log
->l_ailp
;
2627 efd_formatp
= item
->ri_buf
[0].i_addr
;
2628 ASSERT((item
->ri_buf
[0].i_len
== (sizeof(xfs_efd_log_format_32_t
) +
2629 ((efd_formatp
->efd_nextents
- 1) * sizeof(xfs_extent_32_t
)))) ||
2630 (item
->ri_buf
[0].i_len
== (sizeof(xfs_efd_log_format_64_t
) +
2631 ((efd_formatp
->efd_nextents
- 1) * sizeof(xfs_extent_64_t
)))));
2632 efi_id
= efd_formatp
->efd_efi_id
;
2635 * Search for the efi with the id in the efd format structure
2638 spin_lock(&ailp
->xa_lock
);
2639 lip
= xfs_trans_ail_cursor_first(ailp
, &cur
, 0);
2640 while (lip
!= NULL
) {
2641 if (lip
->li_type
== XFS_LI_EFI
) {
2642 efip
= (xfs_efi_log_item_t
*)lip
;
2643 if (efip
->efi_format
.efi_id
== efi_id
) {
2645 * xfs_trans_ail_delete() drops the
2648 xfs_trans_ail_delete(ailp
, lip
,
2649 SHUTDOWN_CORRUPT_INCORE
);
2650 xfs_efi_item_free(efip
);
2651 spin_lock(&ailp
->xa_lock
);
2655 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
2657 xfs_trans_ail_cursor_done(ailp
, &cur
);
2658 spin_unlock(&ailp
->xa_lock
);
2664 * Free up any resources allocated by the transaction
2666 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2669 xlog_recover_free_trans(
2670 struct xlog_recover
*trans
)
2672 xlog_recover_item_t
*item
, *n
;
2675 list_for_each_entry_safe(item
, n
, &trans
->r_itemq
, ri_list
) {
2676 /* Free the regions in the item. */
2677 list_del(&item
->ri_list
);
2678 for (i
= 0; i
< item
->ri_cnt
; i
++)
2679 kmem_free(item
->ri_buf
[i
].i_addr
);
2680 /* Free the item itself */
2681 kmem_free(item
->ri_buf
);
2684 /* Free the transaction recover structure */
2689 xlog_recover_commit_pass1(
2691 struct xlog_recover
*trans
,
2692 xlog_recover_item_t
*item
)
2694 trace_xfs_log_recover_item_recover(log
, trans
, item
, XLOG_RECOVER_PASS1
);
2696 switch (ITEM_TYPE(item
)) {
2698 return xlog_recover_buffer_pass1(log
, item
);
2699 case XFS_LI_QUOTAOFF
:
2700 return xlog_recover_quotaoff_pass1(log
, item
);
2705 /* nothing to do in pass 1 */
2708 xfs_warn(log
->l_mp
, "%s: invalid item type (%d)",
2709 __func__
, ITEM_TYPE(item
));
2711 return XFS_ERROR(EIO
);
2716 xlog_recover_commit_pass2(
2718 struct xlog_recover
*trans
,
2719 struct list_head
*buffer_list
,
2720 xlog_recover_item_t
*item
)
2722 trace_xfs_log_recover_item_recover(log
, trans
, item
, XLOG_RECOVER_PASS2
);
2724 switch (ITEM_TYPE(item
)) {
2726 return xlog_recover_buffer_pass2(log
, buffer_list
, item
);
2728 return xlog_recover_inode_pass2(log
, buffer_list
, item
);
2730 return xlog_recover_efi_pass2(log
, item
, trans
->r_lsn
);
2732 return xlog_recover_efd_pass2(log
, item
);
2734 return xlog_recover_dquot_pass2(log
, buffer_list
, item
);
2735 case XFS_LI_QUOTAOFF
:
2736 /* nothing to do in pass2 */
2739 xfs_warn(log
->l_mp
, "%s: invalid item type (%d)",
2740 __func__
, ITEM_TYPE(item
));
2742 return XFS_ERROR(EIO
);
2747 * Perform the transaction.
2749 * If the transaction modifies a buffer or inode, do it now. Otherwise,
2750 * EFIs and EFDs get queued up by adding entries into the AIL for them.
2753 xlog_recover_commit_trans(
2755 struct xlog_recover
*trans
,
2758 int error
= 0, error2
;
2759 xlog_recover_item_t
*item
;
2760 LIST_HEAD (buffer_list
);
2762 hlist_del(&trans
->r_list
);
2764 error
= xlog_recover_reorder_trans(log
, trans
, pass
);
2768 list_for_each_entry(item
, &trans
->r_itemq
, ri_list
) {
2770 case XLOG_RECOVER_PASS1
:
2771 error
= xlog_recover_commit_pass1(log
, trans
, item
);
2773 case XLOG_RECOVER_PASS2
:
2774 error
= xlog_recover_commit_pass2(log
, trans
,
2775 &buffer_list
, item
);
2785 xlog_recover_free_trans(trans
);
2788 error2
= xfs_buf_delwri_submit(&buffer_list
);
2789 return error
? error
: error2
;
2793 xlog_recover_unmount_trans(
2795 xlog_recover_t
*trans
)
2797 /* Do nothing now */
2798 xfs_warn(log
->l_mp
, "%s: Unmount LR", __func__
);
2803 * There are two valid states of the r_state field. 0 indicates that the
2804 * transaction structure is in a normal state. We have either seen the
2805 * start of the transaction or the last operation we added was not a partial
2806 * operation. If the last operation we added to the transaction was a
2807 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2809 * NOTE: skip LRs with 0 data length.
2812 xlog_recover_process_data(
2814 struct hlist_head rhash
[],
2815 xlog_rec_header_t
*rhead
,
2821 xlog_op_header_t
*ohead
;
2822 xlog_recover_t
*trans
;
2828 lp
= dp
+ be32_to_cpu(rhead
->h_len
);
2829 num_logops
= be32_to_cpu(rhead
->h_num_logops
);
2831 /* check the log format matches our own - else we can't recover */
2832 if (xlog_header_check_recover(log
->l_mp
, rhead
))
2833 return (XFS_ERROR(EIO
));
2835 while ((dp
< lp
) && num_logops
) {
2836 ASSERT(dp
+ sizeof(xlog_op_header_t
) <= lp
);
2837 ohead
= (xlog_op_header_t
*)dp
;
2838 dp
+= sizeof(xlog_op_header_t
);
2839 if (ohead
->oh_clientid
!= XFS_TRANSACTION
&&
2840 ohead
->oh_clientid
!= XFS_LOG
) {
2841 xfs_warn(log
->l_mp
, "%s: bad clientid 0x%x",
2842 __func__
, ohead
->oh_clientid
);
2844 return (XFS_ERROR(EIO
));
2846 tid
= be32_to_cpu(ohead
->oh_tid
);
2847 hash
= XLOG_RHASH(tid
);
2848 trans
= xlog_recover_find_tid(&rhash
[hash
], tid
);
2849 if (trans
== NULL
) { /* not found; add new tid */
2850 if (ohead
->oh_flags
& XLOG_START_TRANS
)
2851 xlog_recover_new_tid(&rhash
[hash
], tid
,
2852 be64_to_cpu(rhead
->h_lsn
));
2854 if (dp
+ be32_to_cpu(ohead
->oh_len
) > lp
) {
2855 xfs_warn(log
->l_mp
, "%s: bad length 0x%x",
2856 __func__
, be32_to_cpu(ohead
->oh_len
));
2858 return (XFS_ERROR(EIO
));
2860 flags
= ohead
->oh_flags
& ~XLOG_END_TRANS
;
2861 if (flags
& XLOG_WAS_CONT_TRANS
)
2862 flags
&= ~XLOG_CONTINUE_TRANS
;
2864 case XLOG_COMMIT_TRANS
:
2865 error
= xlog_recover_commit_trans(log
,
2868 case XLOG_UNMOUNT_TRANS
:
2869 error
= xlog_recover_unmount_trans(log
, trans
);
2871 case XLOG_WAS_CONT_TRANS
:
2872 error
= xlog_recover_add_to_cont_trans(log
,
2874 be32_to_cpu(ohead
->oh_len
));
2876 case XLOG_START_TRANS
:
2877 xfs_warn(log
->l_mp
, "%s: bad transaction",
2880 error
= XFS_ERROR(EIO
);
2883 case XLOG_CONTINUE_TRANS
:
2884 error
= xlog_recover_add_to_trans(log
, trans
,
2885 dp
, be32_to_cpu(ohead
->oh_len
));
2888 xfs_warn(log
->l_mp
, "%s: bad flag 0x%x",
2891 error
= XFS_ERROR(EIO
);
2897 dp
+= be32_to_cpu(ohead
->oh_len
);
2904 * Process an extent free intent item that was recovered from
2905 * the log. We need to free the extents that it describes.
2908 xlog_recover_process_efi(
2910 xfs_efi_log_item_t
*efip
)
2912 xfs_efd_log_item_t
*efdp
;
2917 xfs_fsblock_t startblock_fsb
;
2919 ASSERT(!test_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
));
2922 * First check the validity of the extents described by the
2923 * EFI. If any are bad, then assume that all are bad and
2924 * just toss the EFI.
2926 for (i
= 0; i
< efip
->efi_format
.efi_nextents
; i
++) {
2927 extp
= &(efip
->efi_format
.efi_extents
[i
]);
2928 startblock_fsb
= XFS_BB_TO_FSB(mp
,
2929 XFS_FSB_TO_DADDR(mp
, extp
->ext_start
));
2930 if ((startblock_fsb
== 0) ||
2931 (extp
->ext_len
== 0) ||
2932 (startblock_fsb
>= mp
->m_sb
.sb_dblocks
) ||
2933 (extp
->ext_len
>= mp
->m_sb
.sb_agblocks
)) {
2935 * This will pull the EFI from the AIL and
2936 * free the memory associated with it.
2938 xfs_efi_release(efip
, efip
->efi_format
.efi_nextents
);
2939 return XFS_ERROR(EIO
);
2943 tp
= xfs_trans_alloc(mp
, 0);
2944 error
= xfs_trans_reserve(tp
, 0, XFS_ITRUNCATE_LOG_RES(mp
), 0, 0, 0);
2947 efdp
= xfs_trans_get_efd(tp
, efip
, efip
->efi_format
.efi_nextents
);
2949 for (i
= 0; i
< efip
->efi_format
.efi_nextents
; i
++) {
2950 extp
= &(efip
->efi_format
.efi_extents
[i
]);
2951 error
= xfs_free_extent(tp
, extp
->ext_start
, extp
->ext_len
);
2954 xfs_trans_log_efd_extent(tp
, efdp
, extp
->ext_start
,
2958 set_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
);
2959 error
= xfs_trans_commit(tp
, 0);
2963 xfs_trans_cancel(tp
, XFS_TRANS_ABORT
);
2968 * When this is called, all of the EFIs which did not have
2969 * corresponding EFDs should be in the AIL. What we do now
2970 * is free the extents associated with each one.
2972 * Since we process the EFIs in normal transactions, they
2973 * will be removed at some point after the commit. This prevents
2974 * us from just walking down the list processing each one.
2975 * We'll use a flag in the EFI to skip those that we've already
2976 * processed and use the AIL iteration mechanism's generation
2977 * count to try to speed this up at least a bit.
2979 * When we start, we know that the EFIs are the only things in
2980 * the AIL. As we process them, however, other items are added
2981 * to the AIL. Since everything added to the AIL must come after
2982 * everything already in the AIL, we stop processing as soon as
2983 * we see something other than an EFI in the AIL.
2986 xlog_recover_process_efis(
2989 xfs_log_item_t
*lip
;
2990 xfs_efi_log_item_t
*efip
;
2992 struct xfs_ail_cursor cur
;
2993 struct xfs_ail
*ailp
;
2996 spin_lock(&ailp
->xa_lock
);
2997 lip
= xfs_trans_ail_cursor_first(ailp
, &cur
, 0);
2998 while (lip
!= NULL
) {
3000 * We're done when we see something other than an EFI.
3001 * There should be no EFIs left in the AIL now.
3003 if (lip
->li_type
!= XFS_LI_EFI
) {
3005 for (; lip
; lip
= xfs_trans_ail_cursor_next(ailp
, &cur
))
3006 ASSERT(lip
->li_type
!= XFS_LI_EFI
);
3012 * Skip EFIs that we've already processed.
3014 efip
= (xfs_efi_log_item_t
*)lip
;
3015 if (test_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
)) {
3016 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3020 spin_unlock(&ailp
->xa_lock
);
3021 error
= xlog_recover_process_efi(log
->l_mp
, efip
);
3022 spin_lock(&ailp
->xa_lock
);
3025 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3028 xfs_trans_ail_cursor_done(ailp
, &cur
);
3029 spin_unlock(&ailp
->xa_lock
);
3034 * This routine performs a transaction to null out a bad inode pointer
3035 * in an agi unlinked inode hash bucket.
3038 xlog_recover_clear_agi_bucket(
3040 xfs_agnumber_t agno
,
3049 tp
= xfs_trans_alloc(mp
, XFS_TRANS_CLEAR_AGI_BUCKET
);
3050 error
= xfs_trans_reserve(tp
, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp
),
3055 error
= xfs_read_agi(mp
, tp
, agno
, &agibp
);
3059 agi
= XFS_BUF_TO_AGI(agibp
);
3060 agi
->agi_unlinked
[bucket
] = cpu_to_be32(NULLAGINO
);
3061 offset
= offsetof(xfs_agi_t
, agi_unlinked
) +
3062 (sizeof(xfs_agino_t
) * bucket
);
3063 xfs_trans_log_buf(tp
, agibp
, offset
,
3064 (offset
+ sizeof(xfs_agino_t
) - 1));
3066 error
= xfs_trans_commit(tp
, 0);
3072 xfs_trans_cancel(tp
, XFS_TRANS_ABORT
);
3074 xfs_warn(mp
, "%s: failed to clear agi %d. Continuing.", __func__
, agno
);
3079 xlog_recover_process_one_iunlink(
3080 struct xfs_mount
*mp
,
3081 xfs_agnumber_t agno
,
3085 struct xfs_buf
*ibp
;
3086 struct xfs_dinode
*dip
;
3087 struct xfs_inode
*ip
;
3091 ino
= XFS_AGINO_TO_INO(mp
, agno
, agino
);
3092 error
= xfs_iget(mp
, NULL
, ino
, 0, 0, &ip
);
3097 * Get the on disk inode to find the next inode in the bucket.
3099 error
= xfs_itobp(mp
, NULL
, ip
, &dip
, &ibp
, XBF_LOCK
);
3103 ASSERT(ip
->i_d
.di_nlink
== 0);
3104 ASSERT(ip
->i_d
.di_mode
!= 0);
3106 /* setup for the next pass */
3107 agino
= be32_to_cpu(dip
->di_next_unlinked
);
3111 * Prevent any DMAPI event from being sent when the reference on
3112 * the inode is dropped.
3114 ip
->i_d
.di_dmevmask
= 0;
3123 * We can't read in the inode this bucket points to, or this inode
3124 * is messed up. Just ditch this bucket of inodes. We will lose
3125 * some inodes and space, but at least we won't hang.
3127 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3128 * clear the inode pointer in the bucket.
3130 xlog_recover_clear_agi_bucket(mp
, agno
, bucket
);
3135 * xlog_iunlink_recover
3137 * This is called during recovery to process any inodes which
3138 * we unlinked but not freed when the system crashed. These
3139 * inodes will be on the lists in the AGI blocks. What we do
3140 * here is scan all the AGIs and fully truncate and free any
3141 * inodes found on the lists. Each inode is removed from the
3142 * lists when it has been fully truncated and is freed. The
3143 * freeing of the inode and its removal from the list must be
3147 xlog_recover_process_iunlinks(
3151 xfs_agnumber_t agno
;
3162 * Prevent any DMAPI event from being sent while in this function.
3164 mp_dmevmask
= mp
->m_dmevmask
;
3167 for (agno
= 0; agno
< mp
->m_sb
.sb_agcount
; agno
++) {
3169 * Find the agi for this ag.
3171 error
= xfs_read_agi(mp
, NULL
, agno
, &agibp
);
3174 * AGI is b0rked. Don't process it.
3176 * We should probably mark the filesystem as corrupt
3177 * after we've recovered all the ag's we can....
3182 * Unlock the buffer so that it can be acquired in the normal
3183 * course of the transaction to truncate and free each inode.
3184 * Because we are not racing with anyone else here for the AGI
3185 * buffer, we don't even need to hold it locked to read the
3186 * initial unlinked bucket entries out of the buffer. We keep
3187 * buffer reference though, so that it stays pinned in memory
3188 * while we need the buffer.
3190 agi
= XFS_BUF_TO_AGI(agibp
);
3191 xfs_buf_unlock(agibp
);
3193 for (bucket
= 0; bucket
< XFS_AGI_UNLINKED_BUCKETS
; bucket
++) {
3194 agino
= be32_to_cpu(agi
->agi_unlinked
[bucket
]);
3195 while (agino
!= NULLAGINO
) {
3196 agino
= xlog_recover_process_one_iunlink(mp
,
3197 agno
, agino
, bucket
);
3200 xfs_buf_rele(agibp
);
3203 mp
->m_dmevmask
= mp_dmevmask
;
3209 xlog_pack_data_checksum(
3211 xlog_in_core_t
*iclog
,
3218 up
= (__be32
*)iclog
->ic_datap
;
3219 /* divide length by 4 to get # words */
3220 for (i
= 0; i
< (size
>> 2); i
++) {
3221 chksum
^= be32_to_cpu(*up
);
3224 iclog
->ic_header
.h_chksum
= cpu_to_be32(chksum
);
3227 #define xlog_pack_data_checksum(log, iclog, size)
3231 * Stamp cycle number in every block
3236 xlog_in_core_t
*iclog
,
3240 int size
= iclog
->ic_offset
+ roundoff
;
3244 xlog_pack_data_checksum(log
, iclog
, size
);
3246 cycle_lsn
= CYCLE_LSN_DISK(iclog
->ic_header
.h_lsn
);
3248 dp
= iclog
->ic_datap
;
3249 for (i
= 0; i
< BTOBB(size
) &&
3250 i
< (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
); i
++) {
3251 iclog
->ic_header
.h_cycle_data
[i
] = *(__be32
*)dp
;
3252 *(__be32
*)dp
= cycle_lsn
;
3256 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
3257 xlog_in_core_2_t
*xhdr
= iclog
->ic_data
;
3259 for ( ; i
< BTOBB(size
); i
++) {
3260 j
= i
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3261 k
= i
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3262 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
] = *(__be32
*)dp
;
3263 *(__be32
*)dp
= cycle_lsn
;
3267 for (i
= 1; i
< log
->l_iclog_heads
; i
++) {
3268 xhdr
[i
].hic_xheader
.xh_cycle
= cycle_lsn
;
3275 xlog_rec_header_t
*rhead
,
3281 for (i
= 0; i
< BTOBB(be32_to_cpu(rhead
->h_len
)) &&
3282 i
< (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
); i
++) {
3283 *(__be32
*)dp
= *(__be32
*)&rhead
->h_cycle_data
[i
];
3287 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
3288 xlog_in_core_2_t
*xhdr
= (xlog_in_core_2_t
*)rhead
;
3289 for ( ; i
< BTOBB(be32_to_cpu(rhead
->h_len
)); i
++) {
3290 j
= i
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3291 k
= i
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3292 *(__be32
*)dp
= xhdr
[j
].hic_xheader
.xh_cycle_data
[k
];
3299 xlog_valid_rec_header(
3301 xlog_rec_header_t
*rhead
,
3306 if (unlikely(rhead
->h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))) {
3307 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3308 XFS_ERRLEVEL_LOW
, log
->l_mp
);
3309 return XFS_ERROR(EFSCORRUPTED
);
3312 (!rhead
->h_version
||
3313 (be32_to_cpu(rhead
->h_version
) & (~XLOG_VERSION_OKBITS
))))) {
3314 xfs_warn(log
->l_mp
, "%s: unrecognised log version (%d).",
3315 __func__
, be32_to_cpu(rhead
->h_version
));
3316 return XFS_ERROR(EIO
);
3319 /* LR body must have data or it wouldn't have been written */
3320 hlen
= be32_to_cpu(rhead
->h_len
);
3321 if (unlikely( hlen
<= 0 || hlen
> INT_MAX
)) {
3322 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3323 XFS_ERRLEVEL_LOW
, log
->l_mp
);
3324 return XFS_ERROR(EFSCORRUPTED
);
3326 if (unlikely( blkno
> log
->l_logBBsize
|| blkno
> INT_MAX
)) {
3327 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3328 XFS_ERRLEVEL_LOW
, log
->l_mp
);
3329 return XFS_ERROR(EFSCORRUPTED
);
3335 * Read the log from tail to head and process the log records found.
3336 * Handle the two cases where the tail and head are in the same cycle
3337 * and where the active portion of the log wraps around the end of
3338 * the physical log separately. The pass parameter is passed through
3339 * to the routines called to process the data and is not looked at
3343 xlog_do_recovery_pass(
3345 xfs_daddr_t head_blk
,
3346 xfs_daddr_t tail_blk
,
3349 xlog_rec_header_t
*rhead
;
3352 xfs_buf_t
*hbp
, *dbp
;
3353 int error
= 0, h_size
;
3354 int bblks
, split_bblks
;
3355 int hblks
, split_hblks
, wrapped_hblks
;
3356 struct hlist_head rhash
[XLOG_RHASH_SIZE
];
3358 ASSERT(head_blk
!= tail_blk
);
3361 * Read the header of the tail block and get the iclog buffer size from
3362 * h_size. Use this to tell how many sectors make up the log header.
3364 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
3366 * When using variable length iclogs, read first sector of
3367 * iclog header and extract the header size from it. Get a
3368 * new hbp that is the correct size.
3370 hbp
= xlog_get_bp(log
, 1);
3374 error
= xlog_bread(log
, tail_blk
, 1, hbp
, &offset
);
3378 rhead
= (xlog_rec_header_t
*)offset
;
3379 error
= xlog_valid_rec_header(log
, rhead
, tail_blk
);
3382 h_size
= be32_to_cpu(rhead
->h_size
);
3383 if ((be32_to_cpu(rhead
->h_version
) & XLOG_VERSION_2
) &&
3384 (h_size
> XLOG_HEADER_CYCLE_SIZE
)) {
3385 hblks
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
3386 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
3389 hbp
= xlog_get_bp(log
, hblks
);
3394 ASSERT(log
->l_sectBBsize
== 1);
3396 hbp
= xlog_get_bp(log
, 1);
3397 h_size
= XLOG_BIG_RECORD_BSIZE
;
3402 dbp
= xlog_get_bp(log
, BTOBB(h_size
));
3408 memset(rhash
, 0, sizeof(rhash
));
3409 if (tail_blk
<= head_blk
) {
3410 for (blk_no
= tail_blk
; blk_no
< head_blk
; ) {
3411 error
= xlog_bread(log
, blk_no
, hblks
, hbp
, &offset
);
3415 rhead
= (xlog_rec_header_t
*)offset
;
3416 error
= xlog_valid_rec_header(log
, rhead
, blk_no
);
3420 /* blocks in data section */
3421 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
3422 error
= xlog_bread(log
, blk_no
+ hblks
, bblks
, dbp
,
3427 xlog_unpack_data(rhead
, offset
, log
);
3428 if ((error
= xlog_recover_process_data(log
,
3429 rhash
, rhead
, offset
, pass
)))
3431 blk_no
+= bblks
+ hblks
;
3435 * Perform recovery around the end of the physical log.
3436 * When the head is not on the same cycle number as the tail,
3437 * we can't do a sequential recovery as above.
3440 while (blk_no
< log
->l_logBBsize
) {
3442 * Check for header wrapping around physical end-of-log
3444 offset
= hbp
->b_addr
;
3447 if (blk_no
+ hblks
<= log
->l_logBBsize
) {
3448 /* Read header in one read */
3449 error
= xlog_bread(log
, blk_no
, hblks
, hbp
,
3454 /* This LR is split across physical log end */
3455 if (blk_no
!= log
->l_logBBsize
) {
3456 /* some data before physical log end */
3457 ASSERT(blk_no
<= INT_MAX
);
3458 split_hblks
= log
->l_logBBsize
- (int)blk_no
;
3459 ASSERT(split_hblks
> 0);
3460 error
= xlog_bread(log
, blk_no
,
3468 * Note: this black magic still works with
3469 * large sector sizes (non-512) only because:
3470 * - we increased the buffer size originally
3471 * by 1 sector giving us enough extra space
3472 * for the second read;
3473 * - the log start is guaranteed to be sector
3475 * - we read the log end (LR header start)
3476 * _first_, then the log start (LR header end)
3477 * - order is important.
3479 wrapped_hblks
= hblks
- split_hblks
;
3480 error
= xlog_bread_offset(log
, 0,
3482 offset
+ BBTOB(split_hblks
));
3486 rhead
= (xlog_rec_header_t
*)offset
;
3487 error
= xlog_valid_rec_header(log
, rhead
,
3488 split_hblks
? blk_no
: 0);
3492 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
3495 /* Read in data for log record */
3496 if (blk_no
+ bblks
<= log
->l_logBBsize
) {
3497 error
= xlog_bread(log
, blk_no
, bblks
, dbp
,
3502 /* This log record is split across the
3503 * physical end of log */
3504 offset
= dbp
->b_addr
;
3506 if (blk_no
!= log
->l_logBBsize
) {
3507 /* some data is before the physical
3509 ASSERT(!wrapped_hblks
);
3510 ASSERT(blk_no
<= INT_MAX
);
3512 log
->l_logBBsize
- (int)blk_no
;
3513 ASSERT(split_bblks
> 0);
3514 error
= xlog_bread(log
, blk_no
,
3522 * Note: this black magic still works with
3523 * large sector sizes (non-512) only because:
3524 * - we increased the buffer size originally
3525 * by 1 sector giving us enough extra space
3526 * for the second read;
3527 * - the log start is guaranteed to be sector
3529 * - we read the log end (LR header start)
3530 * _first_, then the log start (LR header end)
3531 * - order is important.
3533 error
= xlog_bread_offset(log
, 0,
3534 bblks
- split_bblks
, hbp
,
3535 offset
+ BBTOB(split_bblks
));
3539 xlog_unpack_data(rhead
, offset
, log
);
3540 if ((error
= xlog_recover_process_data(log
, rhash
,
3541 rhead
, offset
, pass
)))
3546 ASSERT(blk_no
>= log
->l_logBBsize
);
3547 blk_no
-= log
->l_logBBsize
;
3549 /* read first part of physical log */
3550 while (blk_no
< head_blk
) {
3551 error
= xlog_bread(log
, blk_no
, hblks
, hbp
, &offset
);
3555 rhead
= (xlog_rec_header_t
*)offset
;
3556 error
= xlog_valid_rec_header(log
, rhead
, blk_no
);
3560 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
3561 error
= xlog_bread(log
, blk_no
+hblks
, bblks
, dbp
,
3566 xlog_unpack_data(rhead
, offset
, log
);
3567 if ((error
= xlog_recover_process_data(log
, rhash
,
3568 rhead
, offset
, pass
)))
3570 blk_no
+= bblks
+ hblks
;
3582 * Do the recovery of the log. We actually do this in two phases.
3583 * The two passes are necessary in order to implement the function
3584 * of cancelling a record written into the log. The first pass
3585 * determines those things which have been cancelled, and the
3586 * second pass replays log items normally except for those which
3587 * have been cancelled. The handling of the replay and cancellations
3588 * takes place in the log item type specific routines.
3590 * The table of items which have cancel records in the log is allocated
3591 * and freed at this level, since only here do we know when all of
3592 * the log recovery has been completed.
3595 xlog_do_log_recovery(
3597 xfs_daddr_t head_blk
,
3598 xfs_daddr_t tail_blk
)
3602 ASSERT(head_blk
!= tail_blk
);
3605 * First do a pass to find all of the cancelled buf log items.
3606 * Store them in the buf_cancel_table for use in the second pass.
3608 log
->l_buf_cancel_table
= kmem_zalloc(XLOG_BC_TABLE_SIZE
*
3609 sizeof(struct list_head
),
3611 for (i
= 0; i
< XLOG_BC_TABLE_SIZE
; i
++)
3612 INIT_LIST_HEAD(&log
->l_buf_cancel_table
[i
]);
3614 error
= xlog_do_recovery_pass(log
, head_blk
, tail_blk
,
3615 XLOG_RECOVER_PASS1
);
3617 kmem_free(log
->l_buf_cancel_table
);
3618 log
->l_buf_cancel_table
= NULL
;
3622 * Then do a second pass to actually recover the items in the log.
3623 * When it is complete free the table of buf cancel items.
3625 error
= xlog_do_recovery_pass(log
, head_blk
, tail_blk
,
3626 XLOG_RECOVER_PASS2
);
3631 for (i
= 0; i
< XLOG_BC_TABLE_SIZE
; i
++)
3632 ASSERT(list_empty(&log
->l_buf_cancel_table
[i
]));
3636 kmem_free(log
->l_buf_cancel_table
);
3637 log
->l_buf_cancel_table
= NULL
;
3643 * Do the actual recovery
3648 xfs_daddr_t head_blk
,
3649 xfs_daddr_t tail_blk
)
3656 * First replay the images in the log.
3658 error
= xlog_do_log_recovery(log
, head_blk
, tail_blk
);
3663 * If IO errors happened during recovery, bail out.
3665 if (XFS_FORCED_SHUTDOWN(log
->l_mp
)) {
3670 * We now update the tail_lsn since much of the recovery has completed
3671 * and there may be space available to use. If there were no extent
3672 * or iunlinks, we can free up the entire log and set the tail_lsn to
3673 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3674 * lsn of the last known good LR on disk. If there are extent frees
3675 * or iunlinks they will have some entries in the AIL; so we look at
3676 * the AIL to determine how to set the tail_lsn.
3678 xlog_assign_tail_lsn(log
->l_mp
);
3681 * Now that we've finished replaying all buffer and inode
3682 * updates, re-read in the superblock.
3684 bp
= xfs_getsb(log
->l_mp
, 0);
3686 ASSERT(!(XFS_BUF_ISWRITE(bp
)));
3688 XFS_BUF_UNASYNC(bp
);
3689 xfsbdstrat(log
->l_mp
, bp
);
3690 error
= xfs_buf_iowait(bp
);
3692 xfs_buf_ioerror_alert(bp
, __func__
);
3698 /* Convert superblock from on-disk format */
3699 sbp
= &log
->l_mp
->m_sb
;
3700 xfs_sb_from_disk(log
->l_mp
, XFS_BUF_TO_SBP(bp
));
3701 ASSERT(sbp
->sb_magicnum
== XFS_SB_MAGIC
);
3702 ASSERT(xfs_sb_good_version(sbp
));
3705 /* We've re-read the superblock so re-initialize per-cpu counters */
3706 xfs_icsb_reinit_counters(log
->l_mp
);
3708 xlog_recover_check_summary(log
);
3710 /* Normal transactions can now occur */
3711 log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
3716 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3718 * Return error or zero.
3724 xfs_daddr_t head_blk
, tail_blk
;
3727 /* find the tail of the log */
3728 if ((error
= xlog_find_tail(log
, &head_blk
, &tail_blk
)))
3731 if (tail_blk
!= head_blk
) {
3732 /* There used to be a comment here:
3734 * disallow recovery on read-only mounts. note -- mount
3735 * checks for ENOSPC and turns it into an intelligent
3737 * ...but this is no longer true. Now, unless you specify
3738 * NORECOVERY (in which case this function would never be
3739 * called), we just go ahead and recover. We do this all
3740 * under the vfs layer, so we can get away with it unless
3741 * the device itself is read-only, in which case we fail.
3743 if ((error
= xfs_dev_is_read_only(log
->l_mp
, "recovery"))) {
3747 xfs_notice(log
->l_mp
, "Starting recovery (logdev: %s)",
3748 log
->l_mp
->m_logname
? log
->l_mp
->m_logname
3751 error
= xlog_do_recover(log
, head_blk
, tail_blk
);
3752 log
->l_flags
|= XLOG_RECOVERY_NEEDED
;
3758 * In the first part of recovery we replay inodes and buffers and build
3759 * up the list of extent free items which need to be processed. Here
3760 * we process the extent free items and clean up the on disk unlinked
3761 * inode lists. This is separated from the first part of recovery so
3762 * that the root and real-time bitmap inodes can be read in from disk in
3763 * between the two stages. This is necessary so that we can free space
3764 * in the real-time portion of the file system.
3767 xlog_recover_finish(
3771 * Now we're ready to do the transactions needed for the
3772 * rest of recovery. Start with completing all the extent
3773 * free intent records and then process the unlinked inode
3774 * lists. At this point, we essentially run in normal mode
3775 * except that we're still performing recovery actions
3776 * rather than accepting new requests.
3778 if (log
->l_flags
& XLOG_RECOVERY_NEEDED
) {
3780 error
= xlog_recover_process_efis(log
);
3782 xfs_alert(log
->l_mp
, "Failed to recover EFIs");
3786 * Sync the log to get all the EFIs out of the AIL.
3787 * This isn't absolutely necessary, but it helps in
3788 * case the unlink transactions would have problems
3789 * pushing the EFIs out of the way.
3791 xfs_log_force(log
->l_mp
, XFS_LOG_SYNC
);
3793 xlog_recover_process_iunlinks(log
);
3795 xlog_recover_check_summary(log
);
3797 xfs_notice(log
->l_mp
, "Ending recovery (logdev: %s)",
3798 log
->l_mp
->m_logname
? log
->l_mp
->m_logname
3800 log
->l_flags
&= ~XLOG_RECOVERY_NEEDED
;
3802 xfs_info(log
->l_mp
, "Ending clean mount");
3810 * Read all of the agf and agi counters and check that they
3811 * are consistent with the superblock counters.
3814 xlog_recover_check_summary(
3821 xfs_agnumber_t agno
;
3822 __uint64_t freeblks
;
3832 for (agno
= 0; agno
< mp
->m_sb
.sb_agcount
; agno
++) {
3833 error
= xfs_read_agf(mp
, NULL
, agno
, 0, &agfbp
);
3835 xfs_alert(mp
, "%s agf read failed agno %d error %d",
3836 __func__
, agno
, error
);
3838 agfp
= XFS_BUF_TO_AGF(agfbp
);
3839 freeblks
+= be32_to_cpu(agfp
->agf_freeblks
) +
3840 be32_to_cpu(agfp
->agf_flcount
);
3841 xfs_buf_relse(agfbp
);
3844 error
= xfs_read_agi(mp
, NULL
, agno
, &agibp
);
3846 xfs_alert(mp
, "%s agi read failed agno %d error %d",
3847 __func__
, agno
, error
);
3849 struct xfs_agi
*agi
= XFS_BUF_TO_AGI(agibp
);
3851 itotal
+= be32_to_cpu(agi
->agi_count
);
3852 ifree
+= be32_to_cpu(agi
->agi_freecount
);
3853 xfs_buf_relse(agibp
);