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"
43 #include "xfs_utils.h"
44 #include "xfs_cksum.h"
45 #include "xfs_trace.h"
46 #include "xfs_icache.h"
53 xlog_clear_stale_blocks(
58 xlog_recover_check_summary(
61 #define xlog_recover_check_summary(log)
65 * This structure is used during recovery to record the buf log items which
66 * have been canceled and should not be replayed.
68 struct xfs_buf_cancel
{
72 struct list_head bc_list
;
76 * Sector aligned buffer routines for buffer create/read/write/access
80 * Verify the given count of basic blocks is valid number of blocks
81 * to specify for an operation involving the given XFS log buffer.
82 * Returns nonzero if the count is valid, 0 otherwise.
86 xlog_buf_bbcount_valid(
90 return bbcount
> 0 && bbcount
<= log
->l_logBBsize
;
94 * Allocate a buffer to hold log data. The buffer needs to be able
95 * to map to a range of nbblks basic blocks at any valid (basic
96 * block) offset within the log.
105 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
106 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
108 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
113 * We do log I/O in units of log sectors (a power-of-2
114 * multiple of the basic block size), so we round up the
115 * requested size to accommodate the basic blocks required
116 * for complete log sectors.
118 * In addition, the buffer may be used for a non-sector-
119 * aligned block offset, in which case an I/O of the
120 * requested size could extend beyond the end of the
121 * buffer. If the requested size is only 1 basic block it
122 * will never straddle a sector boundary, so this won't be
123 * an issue. Nor will this be a problem if the log I/O is
124 * done in basic blocks (sector size 1). But otherwise we
125 * extend the buffer by one extra log sector to ensure
126 * there's space to accommodate this possibility.
128 if (nbblks
> 1 && log
->l_sectBBsize
> 1)
129 nbblks
+= log
->l_sectBBsize
;
130 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
132 bp
= xfs_buf_get_uncached(log
->l_mp
->m_logdev_targp
, nbblks
, 0);
146 * Return the address of the start of the given block number's data
147 * in a log buffer. The buffer covers a log sector-aligned region.
156 xfs_daddr_t offset
= blk_no
& ((xfs_daddr_t
)log
->l_sectBBsize
- 1);
158 ASSERT(offset
+ nbblks
<= bp
->b_length
);
159 return bp
->b_addr
+ BBTOB(offset
);
164 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
175 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
176 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
178 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
182 blk_no
= round_down(blk_no
, log
->l_sectBBsize
);
183 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
186 ASSERT(nbblks
<= bp
->b_length
);
188 XFS_BUF_SET_ADDR(bp
, log
->l_logBBstart
+ blk_no
);
190 bp
->b_io_length
= nbblks
;
193 xfsbdstrat(log
->l_mp
, bp
);
194 error
= xfs_buf_iowait(bp
);
196 xfs_buf_ioerror_alert(bp
, __func__
);
210 error
= xlog_bread_noalign(log
, blk_no
, nbblks
, bp
);
214 *offset
= xlog_align(log
, blk_no
, nbblks
, bp
);
219 * Read at an offset into the buffer. Returns with the buffer in it's original
220 * state regardless of the result of the read.
225 xfs_daddr_t blk_no
, /* block to read from */
226 int nbblks
, /* blocks to read */
230 xfs_caddr_t orig_offset
= bp
->b_addr
;
231 int orig_len
= BBTOB(bp
->b_length
);
234 error
= xfs_buf_associate_memory(bp
, offset
, BBTOB(nbblks
));
238 error
= xlog_bread_noalign(log
, blk_no
, nbblks
, bp
);
240 /* must reset buffer pointer even on error */
241 error2
= xfs_buf_associate_memory(bp
, orig_offset
, orig_len
);
248 * Write out the buffer at the given block for the given number of blocks.
249 * The buffer is kept locked across the write and is returned locked.
250 * This can only be used for synchronous log writes.
261 if (!xlog_buf_bbcount_valid(log
, nbblks
)) {
262 xfs_warn(log
->l_mp
, "Invalid block length (0x%x) for buffer",
264 XFS_ERROR_REPORT(__func__
, XFS_ERRLEVEL_HIGH
, log
->l_mp
);
268 blk_no
= round_down(blk_no
, log
->l_sectBBsize
);
269 nbblks
= round_up(nbblks
, log
->l_sectBBsize
);
272 ASSERT(nbblks
<= bp
->b_length
);
274 XFS_BUF_SET_ADDR(bp
, log
->l_logBBstart
+ blk_no
);
275 XFS_BUF_ZEROFLAGS(bp
);
278 bp
->b_io_length
= nbblks
;
281 error
= xfs_bwrite(bp
);
283 xfs_buf_ioerror_alert(bp
, __func__
);
290 * dump debug superblock and log record information
293 xlog_header_check_dump(
295 xlog_rec_header_t
*head
)
297 xfs_debug(mp
, "%s: SB : uuid = %pU, fmt = %d\n",
298 __func__
, &mp
->m_sb
.sb_uuid
, XLOG_FMT
);
299 xfs_debug(mp
, " log : uuid = %pU, fmt = %d\n",
300 &head
->h_fs_uuid
, be32_to_cpu(head
->h_fmt
));
303 #define xlog_header_check_dump(mp, head)
307 * check log record header for recovery
310 xlog_header_check_recover(
312 xlog_rec_header_t
*head
)
314 ASSERT(head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
));
317 * IRIX doesn't write the h_fmt field and leaves it zeroed
318 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
319 * a dirty log created in IRIX.
321 if (unlikely(head
->h_fmt
!= cpu_to_be32(XLOG_FMT
))) {
323 "dirty log written in incompatible format - can't recover");
324 xlog_header_check_dump(mp
, head
);
325 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
326 XFS_ERRLEVEL_HIGH
, mp
);
327 return XFS_ERROR(EFSCORRUPTED
);
328 } else if (unlikely(!uuid_equal(&mp
->m_sb
.sb_uuid
, &head
->h_fs_uuid
))) {
330 "dirty log entry has mismatched uuid - can't recover");
331 xlog_header_check_dump(mp
, head
);
332 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
333 XFS_ERRLEVEL_HIGH
, mp
);
334 return XFS_ERROR(EFSCORRUPTED
);
340 * read the head block of the log and check the header
343 xlog_header_check_mount(
345 xlog_rec_header_t
*head
)
347 ASSERT(head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
));
349 if (uuid_is_nil(&head
->h_fs_uuid
)) {
351 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
352 * h_fs_uuid is nil, we assume this log was last mounted
353 * by IRIX and continue.
355 xfs_warn(mp
, "nil uuid in log - IRIX style log");
356 } else if (unlikely(!uuid_equal(&mp
->m_sb
.sb_uuid
, &head
->h_fs_uuid
))) {
357 xfs_warn(mp
, "log has mismatched uuid - can't recover");
358 xlog_header_check_dump(mp
, head
);
359 XFS_ERROR_REPORT("xlog_header_check_mount",
360 XFS_ERRLEVEL_HIGH
, mp
);
361 return XFS_ERROR(EFSCORRUPTED
);
372 * We're not going to bother about retrying
373 * this during recovery. One strike!
375 xfs_buf_ioerror_alert(bp
, __func__
);
376 xfs_force_shutdown(bp
->b_target
->bt_mount
,
377 SHUTDOWN_META_IO_ERROR
);
380 xfs_buf_ioend(bp
, 0);
384 * This routine finds (to an approximation) the first block in the physical
385 * log which contains the given cycle. It uses a binary search algorithm.
386 * Note that the algorithm can not be perfect because the disk will not
387 * necessarily be perfect.
390 xlog_find_cycle_start(
393 xfs_daddr_t first_blk
,
394 xfs_daddr_t
*last_blk
,
404 mid_blk
= BLK_AVG(first_blk
, end_blk
);
405 while (mid_blk
!= first_blk
&& mid_blk
!= end_blk
) {
406 error
= xlog_bread(log
, mid_blk
, 1, bp
, &offset
);
409 mid_cycle
= xlog_get_cycle(offset
);
410 if (mid_cycle
== cycle
)
411 end_blk
= mid_blk
; /* last_half_cycle == mid_cycle */
413 first_blk
= mid_blk
; /* first_half_cycle == mid_cycle */
414 mid_blk
= BLK_AVG(first_blk
, end_blk
);
416 ASSERT((mid_blk
== first_blk
&& mid_blk
+1 == end_blk
) ||
417 (mid_blk
== end_blk
&& mid_blk
-1 == first_blk
));
425 * Check that a range of blocks does not contain stop_on_cycle_no.
426 * Fill in *new_blk with the block offset where such a block is
427 * found, or with -1 (an invalid block number) if there is no such
428 * block in the range. The scan needs to occur from front to back
429 * and the pointer into the region must be updated since a later
430 * routine will need to perform another test.
433 xlog_find_verify_cycle(
435 xfs_daddr_t start_blk
,
437 uint stop_on_cycle_no
,
438 xfs_daddr_t
*new_blk
)
444 xfs_caddr_t buf
= NULL
;
448 * Greedily allocate a buffer big enough to handle the full
449 * range of basic blocks we'll be examining. If that fails,
450 * try a smaller size. We need to be able to read at least
451 * a log sector, or we're out of luck.
453 bufblks
= 1 << ffs(nbblks
);
454 while (bufblks
> log
->l_logBBsize
)
456 while (!(bp
= xlog_get_bp(log
, bufblks
))) {
458 if (bufblks
< log
->l_sectBBsize
)
462 for (i
= start_blk
; i
< start_blk
+ nbblks
; i
+= bufblks
) {
465 bcount
= min(bufblks
, (start_blk
+ nbblks
- i
));
467 error
= xlog_bread(log
, i
, bcount
, bp
, &buf
);
471 for (j
= 0; j
< bcount
; j
++) {
472 cycle
= xlog_get_cycle(buf
);
473 if (cycle
== stop_on_cycle_no
) {
490 * Potentially backup over partial log record write.
492 * In the typical case, last_blk is the number of the block directly after
493 * a good log record. Therefore, we subtract one to get the block number
494 * of the last block in the given buffer. extra_bblks contains the number
495 * of blocks we would have read on a previous read. This happens when the
496 * last log record is split over the end of the physical log.
498 * extra_bblks is the number of blocks potentially verified on a previous
499 * call to this routine.
502 xlog_find_verify_log_record(
504 xfs_daddr_t start_blk
,
505 xfs_daddr_t
*last_blk
,
510 xfs_caddr_t offset
= NULL
;
511 xlog_rec_header_t
*head
= NULL
;
514 int num_blks
= *last_blk
- start_blk
;
517 ASSERT(start_blk
!= 0 || *last_blk
!= start_blk
);
519 if (!(bp
= xlog_get_bp(log
, num_blks
))) {
520 if (!(bp
= xlog_get_bp(log
, 1)))
524 error
= xlog_bread(log
, start_blk
, num_blks
, bp
, &offset
);
527 offset
+= ((num_blks
- 1) << BBSHIFT
);
530 for (i
= (*last_blk
) - 1; i
>= 0; i
--) {
532 /* valid log record not found */
534 "Log inconsistent (didn't find previous header)");
536 error
= XFS_ERROR(EIO
);
541 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
546 head
= (xlog_rec_header_t
*)offset
;
548 if (head
->h_magicno
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
556 * We hit the beginning of the physical log & still no header. Return
557 * to caller. If caller can handle a return of -1, then this routine
558 * will be called again for the end of the physical log.
566 * We have the final block of the good log (the first block
567 * of the log record _before_ the head. So we check the uuid.
569 if ((error
= xlog_header_check_mount(log
->l_mp
, head
)))
573 * We may have found a log record header before we expected one.
574 * last_blk will be the 1st block # with a given cycle #. We may end
575 * up reading an entire log record. In this case, we don't want to
576 * reset last_blk. Only when last_blk points in the middle of a log
577 * record do we update last_blk.
579 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
580 uint h_size
= be32_to_cpu(head
->h_size
);
582 xhdrs
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
583 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
589 if (*last_blk
- i
+ extra_bblks
!=
590 BTOBB(be32_to_cpu(head
->h_len
)) + xhdrs
)
599 * Head is defined to be the point of the log where the next log write
600 * write could go. This means that incomplete LR writes at the end are
601 * eliminated when calculating the head. We aren't guaranteed that previous
602 * LR have complete transactions. We only know that a cycle number of
603 * current cycle number -1 won't be present in the log if we start writing
604 * from our current block number.
606 * last_blk contains the block number of the first block with a given
609 * Return: zero if normal, non-zero if error.
614 xfs_daddr_t
*return_head_blk
)
618 xfs_daddr_t new_blk
, first_blk
, start_blk
, last_blk
, head_blk
;
620 uint first_half_cycle
, last_half_cycle
;
622 int error
, log_bbnum
= log
->l_logBBsize
;
624 /* Is the end of the log device zeroed? */
625 if ((error
= xlog_find_zeroed(log
, &first_blk
)) == -1) {
626 *return_head_blk
= first_blk
;
628 /* Is the whole lot zeroed? */
630 /* Linux XFS shouldn't generate totally zeroed logs -
631 * mkfs etc write a dummy unmount record to a fresh
632 * log so we can store the uuid in there
634 xfs_warn(log
->l_mp
, "totally zeroed log");
639 xfs_warn(log
->l_mp
, "empty log check failed");
643 first_blk
= 0; /* get cycle # of 1st block */
644 bp
= xlog_get_bp(log
, 1);
648 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
652 first_half_cycle
= xlog_get_cycle(offset
);
654 last_blk
= head_blk
= log_bbnum
- 1; /* get cycle # of last block */
655 error
= xlog_bread(log
, last_blk
, 1, bp
, &offset
);
659 last_half_cycle
= xlog_get_cycle(offset
);
660 ASSERT(last_half_cycle
!= 0);
663 * If the 1st half cycle number is equal to the last half cycle number,
664 * then the entire log is stamped with the same cycle number. In this
665 * case, head_blk can't be set to zero (which makes sense). The below
666 * math doesn't work out properly with head_blk equal to zero. Instead,
667 * we set it to log_bbnum which is an invalid block number, but this
668 * value makes the math correct. If head_blk doesn't changed through
669 * all the tests below, *head_blk is set to zero at the very end rather
670 * than log_bbnum. In a sense, log_bbnum and zero are the same block
671 * in a circular file.
673 if (first_half_cycle
== last_half_cycle
) {
675 * In this case we believe that the entire log should have
676 * cycle number last_half_cycle. We need to scan backwards
677 * from the end verifying that there are no holes still
678 * containing last_half_cycle - 1. If we find such a hole,
679 * then the start of that hole will be the new head. The
680 * simple case looks like
681 * x | x ... | x - 1 | x
682 * Another case that fits this picture would be
683 * x | x + 1 | x ... | x
684 * In this case the head really is somewhere at the end of the
685 * log, as one of the latest writes at the beginning was
688 * x | x + 1 | x ... | x - 1 | x
689 * This is really the combination of the above two cases, and
690 * the head has to end up at the start of the x-1 hole at the
693 * In the 256k log case, we will read from the beginning to the
694 * end of the log and search for cycle numbers equal to x-1.
695 * We don't worry about the x+1 blocks that we encounter,
696 * because we know that they cannot be the head since the log
699 head_blk
= log_bbnum
;
700 stop_on_cycle
= last_half_cycle
- 1;
703 * In this case we want to find the first block with cycle
704 * number matching last_half_cycle. We expect the log to be
706 * x + 1 ... | x ... | x
707 * The first block with cycle number x (last_half_cycle) will
708 * be where the new head belongs. First we do a binary search
709 * for the first occurrence of last_half_cycle. The binary
710 * search may not be totally accurate, so then we scan back
711 * from there looking for occurrences of last_half_cycle before
712 * us. If that backwards scan wraps around the beginning of
713 * the log, then we look for occurrences of last_half_cycle - 1
714 * at the end of the log. The cases we're looking for look
716 * v binary search stopped here
717 * x + 1 ... | x | x + 1 | x ... | x
718 * ^ but we want to locate this spot
720 * <---------> less than scan distance
721 * x + 1 ... | x ... | x - 1 | x
722 * ^ we want to locate this spot
724 stop_on_cycle
= last_half_cycle
;
725 if ((error
= xlog_find_cycle_start(log
, bp
, first_blk
,
726 &head_blk
, last_half_cycle
)))
731 * Now validate the answer. Scan back some number of maximum possible
732 * blocks and make sure each one has the expected cycle number. The
733 * maximum is determined by the total possible amount of buffering
734 * in the in-core log. The following number can be made tighter if
735 * we actually look at the block size of the filesystem.
737 num_scan_bblks
= XLOG_TOTAL_REC_SHIFT(log
);
738 if (head_blk
>= num_scan_bblks
) {
740 * We are guaranteed that the entire check can be performed
743 start_blk
= head_blk
- num_scan_bblks
;
744 if ((error
= xlog_find_verify_cycle(log
,
745 start_blk
, num_scan_bblks
,
746 stop_on_cycle
, &new_blk
)))
750 } else { /* need to read 2 parts of log */
752 * We are going to scan backwards in the log in two parts.
753 * First we scan the physical end of the log. In this part
754 * of the log, we are looking for blocks with cycle number
755 * last_half_cycle - 1.
756 * If we find one, then we know that the log starts there, as
757 * we've found a hole that didn't get written in going around
758 * the end of the physical log. The simple case for this is
759 * x + 1 ... | x ... | x - 1 | x
760 * <---------> less than scan distance
761 * If all of the blocks at the end of the log have cycle number
762 * last_half_cycle, then we check the blocks at the start of
763 * the log looking for occurrences of last_half_cycle. If we
764 * find one, then our current estimate for the location of the
765 * first occurrence of last_half_cycle is wrong and we move
766 * back to the hole we've found. This case looks like
767 * x + 1 ... | x | x + 1 | x ...
768 * ^ binary search stopped here
769 * Another case we need to handle that only occurs in 256k
771 * x + 1 ... | x ... | x+1 | x ...
772 * ^ binary search stops here
773 * In a 256k log, the scan at the end of the log will see the
774 * x + 1 blocks. We need to skip past those since that is
775 * certainly not the head of the log. By searching for
776 * last_half_cycle-1 we accomplish that.
778 ASSERT(head_blk
<= INT_MAX
&&
779 (xfs_daddr_t
) num_scan_bblks
>= head_blk
);
780 start_blk
= log_bbnum
- (num_scan_bblks
- head_blk
);
781 if ((error
= xlog_find_verify_cycle(log
, start_blk
,
782 num_scan_bblks
- (int)head_blk
,
783 (stop_on_cycle
- 1), &new_blk
)))
791 * Scan beginning of log now. The last part of the physical
792 * log is good. This scan needs to verify that it doesn't find
793 * the last_half_cycle.
796 ASSERT(head_blk
<= INT_MAX
);
797 if ((error
= xlog_find_verify_cycle(log
,
798 start_blk
, (int)head_blk
,
799 stop_on_cycle
, &new_blk
)))
807 * Now we need to make sure head_blk is not pointing to a block in
808 * the middle of a log record.
810 num_scan_bblks
= XLOG_REC_SHIFT(log
);
811 if (head_blk
>= num_scan_bblks
) {
812 start_blk
= head_blk
- num_scan_bblks
; /* don't read head_blk */
814 /* start ptr at last block ptr before head_blk */
815 if ((error
= xlog_find_verify_log_record(log
, start_blk
,
816 &head_blk
, 0)) == -1) {
817 error
= XFS_ERROR(EIO
);
823 ASSERT(head_blk
<= INT_MAX
);
824 if ((error
= xlog_find_verify_log_record(log
, start_blk
,
825 &head_blk
, 0)) == -1) {
826 /* We hit the beginning of the log during our search */
827 start_blk
= log_bbnum
- (num_scan_bblks
- head_blk
);
829 ASSERT(start_blk
<= INT_MAX
&&
830 (xfs_daddr_t
) log_bbnum
-start_blk
>= 0);
831 ASSERT(head_blk
<= INT_MAX
);
832 if ((error
= xlog_find_verify_log_record(log
,
834 (int)head_blk
)) == -1) {
835 error
= XFS_ERROR(EIO
);
839 if (new_blk
!= log_bbnum
)
846 if (head_blk
== log_bbnum
)
847 *return_head_blk
= 0;
849 *return_head_blk
= head_blk
;
851 * When returning here, we have a good block number. Bad block
852 * means that during a previous crash, we didn't have a clean break
853 * from cycle number N to cycle number N-1. In this case, we need
854 * to find the first block with cycle number N-1.
862 xfs_warn(log
->l_mp
, "failed to find log head");
867 * Find the sync block number or the tail of the log.
869 * This will be the block number of the last record to have its
870 * associated buffers synced to disk. Every log record header has
871 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
872 * to get a sync block number. The only concern is to figure out which
873 * log record header to believe.
875 * The following algorithm uses the log record header with the largest
876 * lsn. The entire log record does not need to be valid. We only care
877 * that the header is valid.
879 * We could speed up search by using current head_blk buffer, but it is not
885 xfs_daddr_t
*head_blk
,
886 xfs_daddr_t
*tail_blk
)
888 xlog_rec_header_t
*rhead
;
889 xlog_op_header_t
*op_head
;
890 xfs_caddr_t offset
= NULL
;
893 xfs_daddr_t umount_data_blk
;
894 xfs_daddr_t after_umount_blk
;
901 * Find previous log record
903 if ((error
= xlog_find_head(log
, head_blk
)))
906 bp
= xlog_get_bp(log
, 1);
909 if (*head_blk
== 0) { /* special case */
910 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
914 if (xlog_get_cycle(offset
) == 0) {
916 /* leave all other log inited values alone */
922 * Search backwards looking for log record header block
924 ASSERT(*head_blk
< INT_MAX
);
925 for (i
= (int)(*head_blk
) - 1; i
>= 0; i
--) {
926 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
930 if (*(__be32
*)offset
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
)) {
936 * If we haven't found the log record header block, start looking
937 * again from the end of the physical log. XXXmiken: There should be
938 * a check here to make sure we didn't search more than N blocks in
942 for (i
= log
->l_logBBsize
- 1; i
>= (int)(*head_blk
); i
--) {
943 error
= xlog_bread(log
, i
, 1, bp
, &offset
);
947 if (*(__be32
*)offset
==
948 cpu_to_be32(XLOG_HEADER_MAGIC_NUM
)) {
955 xfs_warn(log
->l_mp
, "%s: couldn't find sync record", __func__
);
957 return XFS_ERROR(EIO
);
960 /* find blk_no of tail of log */
961 rhead
= (xlog_rec_header_t
*)offset
;
962 *tail_blk
= BLOCK_LSN(be64_to_cpu(rhead
->h_tail_lsn
));
965 * Reset log values according to the state of the log when we
966 * crashed. In the case where head_blk == 0, we bump curr_cycle
967 * one because the next write starts a new cycle rather than
968 * continuing the cycle of the last good log record. At this
969 * point we have guaranteed that all partial log records have been
970 * accounted for. Therefore, we know that the last good log record
971 * written was complete and ended exactly on the end boundary
972 * of the physical log.
974 log
->l_prev_block
= i
;
975 log
->l_curr_block
= (int)*head_blk
;
976 log
->l_curr_cycle
= be32_to_cpu(rhead
->h_cycle
);
979 atomic64_set(&log
->l_tail_lsn
, be64_to_cpu(rhead
->h_tail_lsn
));
980 atomic64_set(&log
->l_last_sync_lsn
, be64_to_cpu(rhead
->h_lsn
));
981 xlog_assign_grant_head(&log
->l_reserve_head
.grant
, log
->l_curr_cycle
,
982 BBTOB(log
->l_curr_block
));
983 xlog_assign_grant_head(&log
->l_write_head
.grant
, log
->l_curr_cycle
,
984 BBTOB(log
->l_curr_block
));
987 * Look for unmount record. If we find it, then we know there
988 * was a clean unmount. Since 'i' could be the last block in
989 * the physical log, we convert to a log block before comparing
992 * Save the current tail lsn to use to pass to
993 * xlog_clear_stale_blocks() below. We won't want to clear the
994 * unmount record if there is one, so we pass the lsn of the
995 * unmount record rather than the block after it.
997 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
998 int h_size
= be32_to_cpu(rhead
->h_size
);
999 int h_version
= be32_to_cpu(rhead
->h_version
);
1001 if ((h_version
& XLOG_VERSION_2
) &&
1002 (h_size
> XLOG_HEADER_CYCLE_SIZE
)) {
1003 hblks
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
1004 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
1012 after_umount_blk
= (i
+ hblks
+ (int)
1013 BTOBB(be32_to_cpu(rhead
->h_len
))) % log
->l_logBBsize
;
1014 tail_lsn
= atomic64_read(&log
->l_tail_lsn
);
1015 if (*head_blk
== after_umount_blk
&&
1016 be32_to_cpu(rhead
->h_num_logops
) == 1) {
1017 umount_data_blk
= (i
+ hblks
) % log
->l_logBBsize
;
1018 error
= xlog_bread(log
, umount_data_blk
, 1, bp
, &offset
);
1022 op_head
= (xlog_op_header_t
*)offset
;
1023 if (op_head
->oh_flags
& XLOG_UNMOUNT_TRANS
) {
1025 * Set tail and last sync so that newly written
1026 * log records will point recovery to after the
1027 * current unmount record.
1029 xlog_assign_atomic_lsn(&log
->l_tail_lsn
,
1030 log
->l_curr_cycle
, after_umount_blk
);
1031 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
,
1032 log
->l_curr_cycle
, after_umount_blk
);
1033 *tail_blk
= after_umount_blk
;
1036 * Note that the unmount was clean. If the unmount
1037 * was not clean, we need to know this to rebuild the
1038 * superblock counters from the perag headers if we
1039 * have a filesystem using non-persistent counters.
1041 log
->l_mp
->m_flags
|= XFS_MOUNT_WAS_CLEAN
;
1046 * Make sure that there are no blocks in front of the head
1047 * with the same cycle number as the head. This can happen
1048 * because we allow multiple outstanding log writes concurrently,
1049 * and the later writes might make it out before earlier ones.
1051 * We use the lsn from before modifying it so that we'll never
1052 * overwrite the unmount record after a clean unmount.
1054 * Do this only if we are going to recover the filesystem
1056 * NOTE: This used to say "if (!readonly)"
1057 * However on Linux, we can & do recover a read-only filesystem.
1058 * We only skip recovery if NORECOVERY is specified on mount,
1059 * in which case we would not be here.
1061 * But... if the -device- itself is readonly, just skip this.
1062 * We can't recover this device anyway, so it won't matter.
1064 if (!xfs_readonly_buftarg(log
->l_mp
->m_logdev_targp
))
1065 error
= xlog_clear_stale_blocks(log
, tail_lsn
);
1071 xfs_warn(log
->l_mp
, "failed to locate log tail");
1076 * Is the log zeroed at all?
1078 * The last binary search should be changed to perform an X block read
1079 * once X becomes small enough. You can then search linearly through
1080 * the X blocks. This will cut down on the number of reads we need to do.
1082 * If the log is partially zeroed, this routine will pass back the blkno
1083 * of the first block with cycle number 0. It won't have a complete LR
1087 * 0 => the log is completely written to
1088 * -1 => use *blk_no as the first block of the log
1089 * >0 => error has occurred
1094 xfs_daddr_t
*blk_no
)
1098 uint first_cycle
, last_cycle
;
1099 xfs_daddr_t new_blk
, last_blk
, start_blk
;
1100 xfs_daddr_t num_scan_bblks
;
1101 int error
, log_bbnum
= log
->l_logBBsize
;
1105 /* check totally zeroed log */
1106 bp
= xlog_get_bp(log
, 1);
1109 error
= xlog_bread(log
, 0, 1, bp
, &offset
);
1113 first_cycle
= xlog_get_cycle(offset
);
1114 if (first_cycle
== 0) { /* completely zeroed log */
1120 /* check partially zeroed log */
1121 error
= xlog_bread(log
, log_bbnum
-1, 1, bp
, &offset
);
1125 last_cycle
= xlog_get_cycle(offset
);
1126 if (last_cycle
!= 0) { /* log completely written to */
1129 } else if (first_cycle
!= 1) {
1131 * If the cycle of the last block is zero, the cycle of
1132 * the first block must be 1. If it's not, maybe we're
1133 * not looking at a log... Bail out.
1136 "Log inconsistent or not a log (last==0, first!=1)");
1137 return XFS_ERROR(EINVAL
);
1140 /* we have a partially zeroed log */
1141 last_blk
= log_bbnum
-1;
1142 if ((error
= xlog_find_cycle_start(log
, bp
, 0, &last_blk
, 0)))
1146 * Validate the answer. Because there is no way to guarantee that
1147 * the entire log is made up of log records which are the same size,
1148 * we scan over the defined maximum blocks. At this point, the maximum
1149 * is not chosen to mean anything special. XXXmiken
1151 num_scan_bblks
= XLOG_TOTAL_REC_SHIFT(log
);
1152 ASSERT(num_scan_bblks
<= INT_MAX
);
1154 if (last_blk
< num_scan_bblks
)
1155 num_scan_bblks
= last_blk
;
1156 start_blk
= last_blk
- num_scan_bblks
;
1159 * We search for any instances of cycle number 0 that occur before
1160 * our current estimate of the head. What we're trying to detect is
1161 * 1 ... | 0 | 1 | 0...
1162 * ^ binary search ends here
1164 if ((error
= xlog_find_verify_cycle(log
, start_blk
,
1165 (int)num_scan_bblks
, 0, &new_blk
)))
1171 * Potentially backup over partial log record write. We don't need
1172 * to search the end of the log because we know it is zero.
1174 if ((error
= xlog_find_verify_log_record(log
, start_blk
,
1175 &last_blk
, 0)) == -1) {
1176 error
= XFS_ERROR(EIO
);
1190 * These are simple subroutines used by xlog_clear_stale_blocks() below
1191 * to initialize a buffer full of empty log record headers and write
1192 * them into the log.
1203 xlog_rec_header_t
*recp
= (xlog_rec_header_t
*)buf
;
1205 memset(buf
, 0, BBSIZE
);
1206 recp
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1207 recp
->h_cycle
= cpu_to_be32(cycle
);
1208 recp
->h_version
= cpu_to_be32(
1209 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1210 recp
->h_lsn
= cpu_to_be64(xlog_assign_lsn(cycle
, block
));
1211 recp
->h_tail_lsn
= cpu_to_be64(xlog_assign_lsn(tail_cycle
, tail_block
));
1212 recp
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1213 memcpy(&recp
->h_fs_uuid
, &log
->l_mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1217 xlog_write_log_records(
1228 int sectbb
= log
->l_sectBBsize
;
1229 int end_block
= start_block
+ blocks
;
1235 * Greedily allocate a buffer big enough to handle the full
1236 * range of basic blocks to be written. If that fails, try
1237 * a smaller size. We need to be able to write at least a
1238 * log sector, or we're out of luck.
1240 bufblks
= 1 << ffs(blocks
);
1241 while (bufblks
> log
->l_logBBsize
)
1243 while (!(bp
= xlog_get_bp(log
, bufblks
))) {
1245 if (bufblks
< sectbb
)
1249 /* We may need to do a read at the start to fill in part of
1250 * the buffer in the starting sector not covered by the first
1253 balign
= round_down(start_block
, sectbb
);
1254 if (balign
!= start_block
) {
1255 error
= xlog_bread_noalign(log
, start_block
, 1, bp
);
1259 j
= start_block
- balign
;
1262 for (i
= start_block
; i
< end_block
; i
+= bufblks
) {
1263 int bcount
, endcount
;
1265 bcount
= min(bufblks
, end_block
- start_block
);
1266 endcount
= bcount
- j
;
1268 /* We may need to do a read at the end to fill in part of
1269 * the buffer in the final sector not covered by the write.
1270 * If this is the same sector as the above read, skip it.
1272 ealign
= round_down(end_block
, sectbb
);
1273 if (j
== 0 && (start_block
+ endcount
> ealign
)) {
1274 offset
= bp
->b_addr
+ BBTOB(ealign
- start_block
);
1275 error
= xlog_bread_offset(log
, ealign
, sectbb
,
1282 offset
= xlog_align(log
, start_block
, endcount
, bp
);
1283 for (; j
< endcount
; j
++) {
1284 xlog_add_record(log
, offset
, cycle
, i
+j
,
1285 tail_cycle
, tail_block
);
1288 error
= xlog_bwrite(log
, start_block
, endcount
, bp
);
1291 start_block
+= endcount
;
1301 * This routine is called to blow away any incomplete log writes out
1302 * in front of the log head. We do this so that we won't become confused
1303 * if we come up, write only a little bit more, and then crash again.
1304 * If we leave the partial log records out there, this situation could
1305 * cause us to think those partial writes are valid blocks since they
1306 * have the current cycle number. We get rid of them by overwriting them
1307 * with empty log records with the old cycle number rather than the
1310 * The tail lsn is passed in rather than taken from
1311 * the log so that we will not write over the unmount record after a
1312 * clean unmount in a 512 block log. Doing so would leave the log without
1313 * any valid log records in it until a new one was written. If we crashed
1314 * during that time we would not be able to recover.
1317 xlog_clear_stale_blocks(
1321 int tail_cycle
, head_cycle
;
1322 int tail_block
, head_block
;
1323 int tail_distance
, max_distance
;
1327 tail_cycle
= CYCLE_LSN(tail_lsn
);
1328 tail_block
= BLOCK_LSN(tail_lsn
);
1329 head_cycle
= log
->l_curr_cycle
;
1330 head_block
= log
->l_curr_block
;
1333 * Figure out the distance between the new head of the log
1334 * and the tail. We want to write over any blocks beyond the
1335 * head that we may have written just before the crash, but
1336 * we don't want to overwrite the tail of the log.
1338 if (head_cycle
== tail_cycle
) {
1340 * The tail is behind the head in the physical log,
1341 * so the distance from the head to the tail is the
1342 * distance from the head to the end of the log plus
1343 * the distance from the beginning of the log to the
1346 if (unlikely(head_block
< tail_block
|| head_block
>= log
->l_logBBsize
)) {
1347 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1348 XFS_ERRLEVEL_LOW
, log
->l_mp
);
1349 return XFS_ERROR(EFSCORRUPTED
);
1351 tail_distance
= tail_block
+ (log
->l_logBBsize
- head_block
);
1354 * The head is behind the tail in the physical log,
1355 * so the distance from the head to the tail is just
1356 * the tail block minus the head block.
1358 if (unlikely(head_block
>= tail_block
|| head_cycle
!= (tail_cycle
+ 1))){
1359 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1360 XFS_ERRLEVEL_LOW
, log
->l_mp
);
1361 return XFS_ERROR(EFSCORRUPTED
);
1363 tail_distance
= tail_block
- head_block
;
1367 * If the head is right up against the tail, we can't clear
1370 if (tail_distance
<= 0) {
1371 ASSERT(tail_distance
== 0);
1375 max_distance
= XLOG_TOTAL_REC_SHIFT(log
);
1377 * Take the smaller of the maximum amount of outstanding I/O
1378 * we could have and the distance to the tail to clear out.
1379 * We take the smaller so that we don't overwrite the tail and
1380 * we don't waste all day writing from the head to the tail
1383 max_distance
= MIN(max_distance
, tail_distance
);
1385 if ((head_block
+ max_distance
) <= log
->l_logBBsize
) {
1387 * We can stomp all the blocks we need to without
1388 * wrapping around the end of the log. Just do it
1389 * in a single write. Use the cycle number of the
1390 * current cycle minus one so that the log will look like:
1393 error
= xlog_write_log_records(log
, (head_cycle
- 1),
1394 head_block
, max_distance
, tail_cycle
,
1400 * We need to wrap around the end of the physical log in
1401 * order to clear all the blocks. Do it in two separate
1402 * I/Os. The first write should be from the head to the
1403 * end of the physical log, and it should use the current
1404 * cycle number minus one just like above.
1406 distance
= log
->l_logBBsize
- head_block
;
1407 error
= xlog_write_log_records(log
, (head_cycle
- 1),
1408 head_block
, distance
, tail_cycle
,
1415 * Now write the blocks at the start of the physical log.
1416 * This writes the remainder of the blocks we want to clear.
1417 * It uses the current cycle number since we're now on the
1418 * same cycle as the head so that we get:
1419 * n ... n ... | n - 1 ...
1420 * ^^^^^ blocks we're writing
1422 distance
= max_distance
- (log
->l_logBBsize
- head_block
);
1423 error
= xlog_write_log_records(log
, head_cycle
, 0, distance
,
1424 tail_cycle
, tail_block
);
1432 /******************************************************************************
1434 * Log recover routines
1436 ******************************************************************************
1439 STATIC xlog_recover_t
*
1440 xlog_recover_find_tid(
1441 struct hlist_head
*head
,
1444 xlog_recover_t
*trans
;
1446 hlist_for_each_entry(trans
, head
, r_list
) {
1447 if (trans
->r_log_tid
== tid
)
1454 xlog_recover_new_tid(
1455 struct hlist_head
*head
,
1459 xlog_recover_t
*trans
;
1461 trans
= kmem_zalloc(sizeof(xlog_recover_t
), KM_SLEEP
);
1462 trans
->r_log_tid
= tid
;
1464 INIT_LIST_HEAD(&trans
->r_itemq
);
1466 INIT_HLIST_NODE(&trans
->r_list
);
1467 hlist_add_head(&trans
->r_list
, head
);
1471 xlog_recover_add_item(
1472 struct list_head
*head
)
1474 xlog_recover_item_t
*item
;
1476 item
= kmem_zalloc(sizeof(xlog_recover_item_t
), KM_SLEEP
);
1477 INIT_LIST_HEAD(&item
->ri_list
);
1478 list_add_tail(&item
->ri_list
, head
);
1482 xlog_recover_add_to_cont_trans(
1484 struct xlog_recover
*trans
,
1488 xlog_recover_item_t
*item
;
1489 xfs_caddr_t ptr
, old_ptr
;
1492 if (list_empty(&trans
->r_itemq
)) {
1493 /* finish copying rest of trans header */
1494 xlog_recover_add_item(&trans
->r_itemq
);
1495 ptr
= (xfs_caddr_t
) &trans
->r_theader
+
1496 sizeof(xfs_trans_header_t
) - len
;
1497 memcpy(ptr
, dp
, len
); /* d, s, l */
1500 /* take the tail entry */
1501 item
= list_entry(trans
->r_itemq
.prev
, xlog_recover_item_t
, ri_list
);
1503 old_ptr
= item
->ri_buf
[item
->ri_cnt
-1].i_addr
;
1504 old_len
= item
->ri_buf
[item
->ri_cnt
-1].i_len
;
1506 ptr
= kmem_realloc(old_ptr
, len
+old_len
, old_len
, KM_SLEEP
);
1507 memcpy(&ptr
[old_len
], dp
, len
); /* d, s, l */
1508 item
->ri_buf
[item
->ri_cnt
-1].i_len
+= len
;
1509 item
->ri_buf
[item
->ri_cnt
-1].i_addr
= ptr
;
1510 trace_xfs_log_recover_item_add_cont(log
, trans
, item
, 0);
1515 * The next region to add is the start of a new region. It could be
1516 * a whole region or it could be the first part of a new region. Because
1517 * of this, the assumption here is that the type and size fields of all
1518 * format structures fit into the first 32 bits of the structure.
1520 * This works because all regions must be 32 bit aligned. Therefore, we
1521 * either have both fields or we have neither field. In the case we have
1522 * neither field, the data part of the region is zero length. We only have
1523 * a log_op_header and can throw away the header since a new one will appear
1524 * later. If we have at least 4 bytes, then we can determine how many regions
1525 * will appear in the current log item.
1528 xlog_recover_add_to_trans(
1530 struct xlog_recover
*trans
,
1534 xfs_inode_log_format_t
*in_f
; /* any will do */
1535 xlog_recover_item_t
*item
;
1540 if (list_empty(&trans
->r_itemq
)) {
1541 /* we need to catch log corruptions here */
1542 if (*(uint
*)dp
!= XFS_TRANS_HEADER_MAGIC
) {
1543 xfs_warn(log
->l_mp
, "%s: bad header magic number",
1546 return XFS_ERROR(EIO
);
1548 if (len
== sizeof(xfs_trans_header_t
))
1549 xlog_recover_add_item(&trans
->r_itemq
);
1550 memcpy(&trans
->r_theader
, dp
, len
); /* d, s, l */
1554 ptr
= kmem_alloc(len
, KM_SLEEP
);
1555 memcpy(ptr
, dp
, len
);
1556 in_f
= (xfs_inode_log_format_t
*)ptr
;
1558 /* take the tail entry */
1559 item
= list_entry(trans
->r_itemq
.prev
, xlog_recover_item_t
, ri_list
);
1560 if (item
->ri_total
!= 0 &&
1561 item
->ri_total
== item
->ri_cnt
) {
1562 /* tail item is in use, get a new one */
1563 xlog_recover_add_item(&trans
->r_itemq
);
1564 item
= list_entry(trans
->r_itemq
.prev
,
1565 xlog_recover_item_t
, ri_list
);
1568 if (item
->ri_total
== 0) { /* first region to be added */
1569 if (in_f
->ilf_size
== 0 ||
1570 in_f
->ilf_size
> XLOG_MAX_REGIONS_IN_ITEM
) {
1572 "bad number of regions (%d) in inode log format",
1575 return XFS_ERROR(EIO
);
1578 item
->ri_total
= in_f
->ilf_size
;
1580 kmem_zalloc(item
->ri_total
* sizeof(xfs_log_iovec_t
),
1583 ASSERT(item
->ri_total
> item
->ri_cnt
);
1584 /* Description region is ri_buf[0] */
1585 item
->ri_buf
[item
->ri_cnt
].i_addr
= ptr
;
1586 item
->ri_buf
[item
->ri_cnt
].i_len
= len
;
1588 trace_xfs_log_recover_item_add(log
, trans
, item
, 0);
1593 * Sort the log items in the transaction. Cancelled buffers need
1594 * to be put first so they are processed before any items that might
1595 * modify the buffers. If they are cancelled, then the modifications
1596 * don't need to be replayed.
1599 xlog_recover_reorder_trans(
1601 struct xlog_recover
*trans
,
1604 xlog_recover_item_t
*item
, *n
;
1605 LIST_HEAD(sort_list
);
1607 list_splice_init(&trans
->r_itemq
, &sort_list
);
1608 list_for_each_entry_safe(item
, n
, &sort_list
, ri_list
) {
1609 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
1611 switch (ITEM_TYPE(item
)) {
1613 if (!(buf_f
->blf_flags
& XFS_BLF_CANCEL
)) {
1614 trace_xfs_log_recover_item_reorder_head(log
,
1616 list_move(&item
->ri_list
, &trans
->r_itemq
);
1621 case XFS_LI_QUOTAOFF
:
1624 trace_xfs_log_recover_item_reorder_tail(log
,
1626 list_move_tail(&item
->ri_list
, &trans
->r_itemq
);
1630 "%s: unrecognized type of log operation",
1633 return XFS_ERROR(EIO
);
1636 ASSERT(list_empty(&sort_list
));
1641 * Build up the table of buf cancel records so that we don't replay
1642 * cancelled data in the second pass. For buffer records that are
1643 * not cancel records, there is nothing to do here so we just return.
1645 * If we get a cancel record which is already in the table, this indicates
1646 * that the buffer was cancelled multiple times. In order to ensure
1647 * that during pass 2 we keep the record in the table until we reach its
1648 * last occurrence in the log, we keep a reference count in the cancel
1649 * record in the table to tell us how many times we expect to see this
1650 * record during the second pass.
1653 xlog_recover_buffer_pass1(
1655 struct xlog_recover_item
*item
)
1657 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
1658 struct list_head
*bucket
;
1659 struct xfs_buf_cancel
*bcp
;
1662 * If this isn't a cancel buffer item, then just return.
1664 if (!(buf_f
->blf_flags
& XFS_BLF_CANCEL
)) {
1665 trace_xfs_log_recover_buf_not_cancel(log
, buf_f
);
1670 * Insert an xfs_buf_cancel record into the hash table of them.
1671 * If there is already an identical record, bump its reference count.
1673 bucket
= XLOG_BUF_CANCEL_BUCKET(log
, buf_f
->blf_blkno
);
1674 list_for_each_entry(bcp
, bucket
, bc_list
) {
1675 if (bcp
->bc_blkno
== buf_f
->blf_blkno
&&
1676 bcp
->bc_len
== buf_f
->blf_len
) {
1678 trace_xfs_log_recover_buf_cancel_ref_inc(log
, buf_f
);
1683 bcp
= kmem_alloc(sizeof(struct xfs_buf_cancel
), KM_SLEEP
);
1684 bcp
->bc_blkno
= buf_f
->blf_blkno
;
1685 bcp
->bc_len
= buf_f
->blf_len
;
1686 bcp
->bc_refcount
= 1;
1687 list_add_tail(&bcp
->bc_list
, bucket
);
1689 trace_xfs_log_recover_buf_cancel_add(log
, buf_f
);
1694 * Check to see whether the buffer being recovered has a corresponding
1695 * entry in the buffer cancel record table. If it does then return 1
1696 * so that it will be cancelled, otherwise return 0. If the buffer is
1697 * actually a buffer cancel item (XFS_BLF_CANCEL is set), then decrement
1698 * the refcount on the entry in the table and remove it from the table
1699 * if this is the last reference.
1701 * We remove the cancel record from the table when we encounter its
1702 * last occurrence in the log so that if the same buffer is re-used
1703 * again after its last cancellation we actually replay the changes
1704 * made at that point.
1707 xlog_check_buffer_cancelled(
1713 struct list_head
*bucket
;
1714 struct xfs_buf_cancel
*bcp
;
1716 if (log
->l_buf_cancel_table
== NULL
) {
1718 * There is nothing in the table built in pass one,
1719 * so this buffer must not be cancelled.
1721 ASSERT(!(flags
& XFS_BLF_CANCEL
));
1726 * Search for an entry in the cancel table that matches our buffer.
1728 bucket
= XLOG_BUF_CANCEL_BUCKET(log
, blkno
);
1729 list_for_each_entry(bcp
, bucket
, bc_list
) {
1730 if (bcp
->bc_blkno
== blkno
&& bcp
->bc_len
== len
)
1735 * We didn't find a corresponding entry in the table, so return 0 so
1736 * that the buffer is NOT cancelled.
1738 ASSERT(!(flags
& XFS_BLF_CANCEL
));
1743 * We've go a match, so return 1 so that the recovery of this buffer
1744 * is cancelled. If this buffer is actually a buffer cancel log
1745 * item, then decrement the refcount on the one in the table and
1746 * remove it if this is the last reference.
1748 if (flags
& XFS_BLF_CANCEL
) {
1749 if (--bcp
->bc_refcount
== 0) {
1750 list_del(&bcp
->bc_list
);
1758 * Perform recovery for a buffer full of inodes. In these buffers, the only
1759 * data which should be recovered is that which corresponds to the
1760 * di_next_unlinked pointers in the on disk inode structures. The rest of the
1761 * data for the inodes is always logged through the inodes themselves rather
1762 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1764 * The only time when buffers full of inodes are fully recovered is when the
1765 * buffer is full of newly allocated inodes. In this case the buffer will
1766 * not be marked as an inode buffer and so will be sent to
1767 * xlog_recover_do_reg_buffer() below during recovery.
1770 xlog_recover_do_inode_buffer(
1771 struct xfs_mount
*mp
,
1772 xlog_recover_item_t
*item
,
1774 xfs_buf_log_format_t
*buf_f
)
1780 int reg_buf_offset
= 0;
1781 int reg_buf_bytes
= 0;
1782 int next_unlinked_offset
;
1784 xfs_agino_t
*logged_nextp
;
1785 xfs_agino_t
*buffer_nextp
;
1787 trace_xfs_log_recover_buf_inode_buf(mp
->m_log
, buf_f
);
1789 inodes_per_buf
= BBTOB(bp
->b_io_length
) >> mp
->m_sb
.sb_inodelog
;
1790 for (i
= 0; i
< inodes_per_buf
; i
++) {
1791 next_unlinked_offset
= (i
* mp
->m_sb
.sb_inodesize
) +
1792 offsetof(xfs_dinode_t
, di_next_unlinked
);
1794 while (next_unlinked_offset
>=
1795 (reg_buf_offset
+ reg_buf_bytes
)) {
1797 * The next di_next_unlinked field is beyond
1798 * the current logged region. Find the next
1799 * logged region that contains or is beyond
1800 * the current di_next_unlinked field.
1803 bit
= xfs_next_bit(buf_f
->blf_data_map
,
1804 buf_f
->blf_map_size
, bit
);
1807 * If there are no more logged regions in the
1808 * buffer, then we're done.
1813 nbits
= xfs_contig_bits(buf_f
->blf_data_map
,
1814 buf_f
->blf_map_size
, bit
);
1816 reg_buf_offset
= bit
<< XFS_BLF_SHIFT
;
1817 reg_buf_bytes
= nbits
<< XFS_BLF_SHIFT
;
1822 * If the current logged region starts after the current
1823 * di_next_unlinked field, then move on to the next
1824 * di_next_unlinked field.
1826 if (next_unlinked_offset
< reg_buf_offset
)
1829 ASSERT(item
->ri_buf
[item_index
].i_addr
!= NULL
);
1830 ASSERT((item
->ri_buf
[item_index
].i_len
% XFS_BLF_CHUNK
) == 0);
1831 ASSERT((reg_buf_offset
+ reg_buf_bytes
) <=
1832 BBTOB(bp
->b_io_length
));
1835 * The current logged region contains a copy of the
1836 * current di_next_unlinked field. Extract its value
1837 * and copy it to the buffer copy.
1839 logged_nextp
= item
->ri_buf
[item_index
].i_addr
+
1840 next_unlinked_offset
- reg_buf_offset
;
1841 if (unlikely(*logged_nextp
== 0)) {
1843 "Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1844 "Trying to replay bad (0) inode di_next_unlinked field.",
1846 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1847 XFS_ERRLEVEL_LOW
, mp
);
1848 return XFS_ERROR(EFSCORRUPTED
);
1851 buffer_nextp
= (xfs_agino_t
*)xfs_buf_offset(bp
,
1852 next_unlinked_offset
);
1853 *buffer_nextp
= *logged_nextp
;
1860 * Perform a 'normal' buffer recovery. Each logged region of the
1861 * buffer should be copied over the corresponding region in the
1862 * given buffer. The bitmap in the buf log format structure indicates
1863 * where to place the logged data.
1866 xlog_recover_do_reg_buffer(
1867 struct xfs_mount
*mp
,
1868 xlog_recover_item_t
*item
,
1870 xfs_buf_log_format_t
*buf_f
)
1877 trace_xfs_log_recover_buf_reg_buf(mp
->m_log
, buf_f
);
1880 i
= 1; /* 0 is the buf format structure */
1882 bit
= xfs_next_bit(buf_f
->blf_data_map
,
1883 buf_f
->blf_map_size
, bit
);
1886 nbits
= xfs_contig_bits(buf_f
->blf_data_map
,
1887 buf_f
->blf_map_size
, bit
);
1889 ASSERT(item
->ri_buf
[i
].i_addr
!= NULL
);
1890 ASSERT(item
->ri_buf
[i
].i_len
% XFS_BLF_CHUNK
== 0);
1891 ASSERT(BBTOB(bp
->b_io_length
) >=
1892 ((uint
)bit
<< XFS_BLF_SHIFT
) + (nbits
<< XFS_BLF_SHIFT
));
1895 * Do a sanity check if this is a dquot buffer. Just checking
1896 * the first dquot in the buffer should do. XXXThis is
1897 * probably a good thing to do for other buf types also.
1900 if (buf_f
->blf_flags
&
1901 (XFS_BLF_UDQUOT_BUF
|XFS_BLF_PDQUOT_BUF
|XFS_BLF_GDQUOT_BUF
)) {
1902 if (item
->ri_buf
[i
].i_addr
== NULL
) {
1904 "XFS: NULL dquot in %s.", __func__
);
1907 if (item
->ri_buf
[i
].i_len
< sizeof(xfs_disk_dquot_t
)) {
1909 "XFS: dquot too small (%d) in %s.",
1910 item
->ri_buf
[i
].i_len
, __func__
);
1913 error
= xfs_qm_dqcheck(mp
, item
->ri_buf
[i
].i_addr
,
1914 -1, 0, XFS_QMOPT_DOWARN
,
1915 "dquot_buf_recover");
1920 memcpy(xfs_buf_offset(bp
,
1921 (uint
)bit
<< XFS_BLF_SHIFT
), /* dest */
1922 item
->ri_buf
[i
].i_addr
, /* source */
1923 nbits
<<XFS_BLF_SHIFT
); /* length */
1929 /* Shouldn't be any more regions */
1930 ASSERT(i
== item
->ri_total
);
1934 * Do some primitive error checking on ondisk dquot data structures.
1938 struct xfs_mount
*mp
,
1939 xfs_disk_dquot_t
*ddq
,
1941 uint type
, /* used only when IO_dorepair is true */
1945 xfs_dqblk_t
*d
= (xfs_dqblk_t
*)ddq
;
1949 * We can encounter an uninitialized dquot buffer for 2 reasons:
1950 * 1. If we crash while deleting the quotainode(s), and those blks got
1951 * used for user data. This is because we take the path of regular
1952 * file deletion; however, the size field of quotainodes is never
1953 * updated, so all the tricks that we play in itruncate_finish
1954 * don't quite matter.
1956 * 2. We don't play the quota buffers when there's a quotaoff logitem.
1957 * But the allocation will be replayed so we'll end up with an
1958 * uninitialized quota block.
1960 * This is all fine; things are still consistent, and we haven't lost
1961 * any quota information. Just don't complain about bad dquot blks.
1963 if (ddq
->d_magic
!= cpu_to_be16(XFS_DQUOT_MAGIC
)) {
1964 if (flags
& XFS_QMOPT_DOWARN
)
1966 "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1967 str
, id
, be16_to_cpu(ddq
->d_magic
), XFS_DQUOT_MAGIC
);
1970 if (ddq
->d_version
!= XFS_DQUOT_VERSION
) {
1971 if (flags
& XFS_QMOPT_DOWARN
)
1973 "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1974 str
, id
, ddq
->d_version
, XFS_DQUOT_VERSION
);
1978 if (ddq
->d_flags
!= XFS_DQ_USER
&&
1979 ddq
->d_flags
!= XFS_DQ_PROJ
&&
1980 ddq
->d_flags
!= XFS_DQ_GROUP
) {
1981 if (flags
& XFS_QMOPT_DOWARN
)
1983 "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1984 str
, id
, ddq
->d_flags
);
1988 if (id
!= -1 && id
!= be32_to_cpu(ddq
->d_id
)) {
1989 if (flags
& XFS_QMOPT_DOWARN
)
1991 "%s : ondisk-dquot 0x%p, ID mismatch: "
1992 "0x%x expected, found id 0x%x",
1993 str
, ddq
, id
, be32_to_cpu(ddq
->d_id
));
1997 if (!errs
&& ddq
->d_id
) {
1998 if (ddq
->d_blk_softlimit
&&
1999 be64_to_cpu(ddq
->d_bcount
) >
2000 be64_to_cpu(ddq
->d_blk_softlimit
)) {
2001 if (!ddq
->d_btimer
) {
2002 if (flags
& XFS_QMOPT_DOWARN
)
2004 "%s : Dquot ID 0x%x (0x%p) BLK TIMER NOT STARTED",
2005 str
, (int)be32_to_cpu(ddq
->d_id
), ddq
);
2009 if (ddq
->d_ino_softlimit
&&
2010 be64_to_cpu(ddq
->d_icount
) >
2011 be64_to_cpu(ddq
->d_ino_softlimit
)) {
2012 if (!ddq
->d_itimer
) {
2013 if (flags
& XFS_QMOPT_DOWARN
)
2015 "%s : Dquot ID 0x%x (0x%p) INODE TIMER NOT STARTED",
2016 str
, (int)be32_to_cpu(ddq
->d_id
), ddq
);
2020 if (ddq
->d_rtb_softlimit
&&
2021 be64_to_cpu(ddq
->d_rtbcount
) >
2022 be64_to_cpu(ddq
->d_rtb_softlimit
)) {
2023 if (!ddq
->d_rtbtimer
) {
2024 if (flags
& XFS_QMOPT_DOWARN
)
2026 "%s : Dquot ID 0x%x (0x%p) RTBLK TIMER NOT STARTED",
2027 str
, (int)be32_to_cpu(ddq
->d_id
), ddq
);
2033 if (!errs
|| !(flags
& XFS_QMOPT_DQREPAIR
))
2036 if (flags
& XFS_QMOPT_DOWARN
)
2037 xfs_notice(mp
, "Re-initializing dquot ID 0x%x", id
);
2040 * Typically, a repair is only requested by quotacheck.
2043 ASSERT(flags
& XFS_QMOPT_DQREPAIR
);
2044 memset(d
, 0, sizeof(xfs_dqblk_t
));
2046 d
->dd_diskdq
.d_magic
= cpu_to_be16(XFS_DQUOT_MAGIC
);
2047 d
->dd_diskdq
.d_version
= XFS_DQUOT_VERSION
;
2048 d
->dd_diskdq
.d_flags
= type
;
2049 d
->dd_diskdq
.d_id
= cpu_to_be32(id
);
2055 * Perform a dquot buffer recovery.
2056 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2057 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2058 * Else, treat it as a regular buffer and do recovery.
2061 xlog_recover_do_dquot_buffer(
2062 struct xfs_mount
*mp
,
2064 struct xlog_recover_item
*item
,
2066 struct xfs_buf_log_format
*buf_f
)
2070 trace_xfs_log_recover_buf_dquot_buf(log
, buf_f
);
2073 * Filesystems are required to send in quota flags at mount time.
2075 if (mp
->m_qflags
== 0) {
2080 if (buf_f
->blf_flags
& XFS_BLF_UDQUOT_BUF
)
2081 type
|= XFS_DQ_USER
;
2082 if (buf_f
->blf_flags
& XFS_BLF_PDQUOT_BUF
)
2083 type
|= XFS_DQ_PROJ
;
2084 if (buf_f
->blf_flags
& XFS_BLF_GDQUOT_BUF
)
2085 type
|= XFS_DQ_GROUP
;
2087 * This type of quotas was turned off, so ignore this buffer
2089 if (log
->l_quotaoffs_flag
& type
)
2092 xlog_recover_do_reg_buffer(mp
, item
, bp
, buf_f
);
2096 * This routine replays a modification made to a buffer at runtime.
2097 * There are actually two types of buffer, regular and inode, which
2098 * are handled differently. Inode buffers are handled differently
2099 * in that we only recover a specific set of data from them, namely
2100 * the inode di_next_unlinked fields. This is because all other inode
2101 * data is actually logged via inode records and any data we replay
2102 * here which overlaps that may be stale.
2104 * When meta-data buffers are freed at run time we log a buffer item
2105 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2106 * of the buffer in the log should not be replayed at recovery time.
2107 * This is so that if the blocks covered by the buffer are reused for
2108 * file data before we crash we don't end up replaying old, freed
2109 * meta-data into a user's file.
2111 * To handle the cancellation of buffer log items, we make two passes
2112 * over the log during recovery. During the first we build a table of
2113 * those buffers which have been cancelled, and during the second we
2114 * only replay those buffers which do not have corresponding cancel
2115 * records in the table. See xlog_recover_do_buffer_pass[1,2] above
2116 * for more details on the implementation of the table of cancel records.
2119 xlog_recover_buffer_pass2(
2121 struct list_head
*buffer_list
,
2122 struct xlog_recover_item
*item
)
2124 xfs_buf_log_format_t
*buf_f
= item
->ri_buf
[0].i_addr
;
2125 xfs_mount_t
*mp
= log
->l_mp
;
2131 * In this pass we only want to recover all the buffers which have
2132 * not been cancelled and are not cancellation buffers themselves.
2134 if (xlog_check_buffer_cancelled(log
, buf_f
->blf_blkno
,
2135 buf_f
->blf_len
, buf_f
->blf_flags
)) {
2136 trace_xfs_log_recover_buf_cancel(log
, buf_f
);
2140 trace_xfs_log_recover_buf_recover(log
, buf_f
);
2143 if (buf_f
->blf_flags
& XFS_BLF_INODE_BUF
)
2144 buf_flags
|= XBF_UNMAPPED
;
2146 bp
= xfs_buf_read(mp
->m_ddev_targp
, buf_f
->blf_blkno
, buf_f
->blf_len
,
2149 return XFS_ERROR(ENOMEM
);
2150 error
= bp
->b_error
;
2152 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#1)");
2157 if (buf_f
->blf_flags
& XFS_BLF_INODE_BUF
) {
2158 error
= xlog_recover_do_inode_buffer(mp
, item
, bp
, buf_f
);
2159 } else if (buf_f
->blf_flags
&
2160 (XFS_BLF_UDQUOT_BUF
|XFS_BLF_PDQUOT_BUF
|XFS_BLF_GDQUOT_BUF
)) {
2161 xlog_recover_do_dquot_buffer(mp
, log
, item
, bp
, buf_f
);
2163 xlog_recover_do_reg_buffer(mp
, item
, bp
, buf_f
);
2166 return XFS_ERROR(error
);
2169 * Perform delayed write on the buffer. Asynchronous writes will be
2170 * slower when taking into account all the buffers to be flushed.
2172 * Also make sure that only inode buffers with good sizes stay in
2173 * the buffer cache. The kernel moves inodes in buffers of 1 block
2174 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger. The inode
2175 * buffers in the log can be a different size if the log was generated
2176 * by an older kernel using unclustered inode buffers or a newer kernel
2177 * running with a different inode cluster size. Regardless, if the
2178 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2179 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2180 * the buffer out of the buffer cache so that the buffer won't
2181 * overlap with future reads of those inodes.
2183 if (XFS_DINODE_MAGIC
==
2184 be16_to_cpu(*((__be16
*)xfs_buf_offset(bp
, 0))) &&
2185 (BBTOB(bp
->b_io_length
) != MAX(log
->l_mp
->m_sb
.sb_blocksize
,
2186 (__uint32_t
)XFS_INODE_CLUSTER_SIZE(log
->l_mp
)))) {
2188 error
= xfs_bwrite(bp
);
2190 ASSERT(bp
->b_target
->bt_mount
== mp
);
2191 bp
->b_iodone
= xlog_recover_iodone
;
2192 xfs_buf_delwri_queue(bp
, buffer_list
);
2200 xlog_recover_inode_pass2(
2202 struct list_head
*buffer_list
,
2203 struct xlog_recover_item
*item
)
2205 xfs_inode_log_format_t
*in_f
;
2206 xfs_mount_t
*mp
= log
->l_mp
;
2215 xfs_icdinode_t
*dicp
;
2218 if (item
->ri_buf
[0].i_len
== sizeof(xfs_inode_log_format_t
)) {
2219 in_f
= item
->ri_buf
[0].i_addr
;
2221 in_f
= kmem_alloc(sizeof(xfs_inode_log_format_t
), KM_SLEEP
);
2223 error
= xfs_inode_item_format_convert(&item
->ri_buf
[0], in_f
);
2229 * Inode buffers can be freed, look out for it,
2230 * and do not replay the inode.
2232 if (xlog_check_buffer_cancelled(log
, in_f
->ilf_blkno
,
2233 in_f
->ilf_len
, 0)) {
2235 trace_xfs_log_recover_inode_cancel(log
, in_f
);
2238 trace_xfs_log_recover_inode_recover(log
, in_f
);
2240 bp
= xfs_buf_read(mp
->m_ddev_targp
, in_f
->ilf_blkno
, in_f
->ilf_len
, 0,
2246 error
= bp
->b_error
;
2248 xfs_buf_ioerror_alert(bp
, "xlog_recover_do..(read#2)");
2252 ASSERT(in_f
->ilf_fields
& XFS_ILOG_CORE
);
2253 dip
= (xfs_dinode_t
*)xfs_buf_offset(bp
, in_f
->ilf_boffset
);
2256 * Make sure the place we're flushing out to really looks
2259 if (unlikely(dip
->di_magic
!= cpu_to_be16(XFS_DINODE_MAGIC
))) {
2262 "%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2263 __func__
, dip
, bp
, in_f
->ilf_ino
);
2264 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2265 XFS_ERRLEVEL_LOW
, mp
);
2266 error
= EFSCORRUPTED
;
2269 dicp
= item
->ri_buf
[1].i_addr
;
2270 if (unlikely(dicp
->di_magic
!= XFS_DINODE_MAGIC
)) {
2273 "%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2274 __func__
, item
, in_f
->ilf_ino
);
2275 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2276 XFS_ERRLEVEL_LOW
, mp
);
2277 error
= EFSCORRUPTED
;
2281 /* Skip replay when the on disk inode is newer than the log one */
2282 if (dicp
->di_flushiter
< be16_to_cpu(dip
->di_flushiter
)) {
2284 * Deal with the wrap case, DI_MAX_FLUSH is less
2285 * than smaller numbers
2287 if (be16_to_cpu(dip
->di_flushiter
) == DI_MAX_FLUSH
&&
2288 dicp
->di_flushiter
< (DI_MAX_FLUSH
>> 1)) {
2292 trace_xfs_log_recover_inode_skip(log
, in_f
);
2297 /* Take the opportunity to reset the flush iteration count */
2298 dicp
->di_flushiter
= 0;
2300 if (unlikely(S_ISREG(dicp
->di_mode
))) {
2301 if ((dicp
->di_format
!= XFS_DINODE_FMT_EXTENTS
) &&
2302 (dicp
->di_format
!= XFS_DINODE_FMT_BTREE
)) {
2303 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2304 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2307 "%s: Bad regular inode log record, rec ptr 0x%p, "
2308 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2309 __func__
, item
, dip
, bp
, in_f
->ilf_ino
);
2310 error
= EFSCORRUPTED
;
2313 } else if (unlikely(S_ISDIR(dicp
->di_mode
))) {
2314 if ((dicp
->di_format
!= XFS_DINODE_FMT_EXTENTS
) &&
2315 (dicp
->di_format
!= XFS_DINODE_FMT_BTREE
) &&
2316 (dicp
->di_format
!= XFS_DINODE_FMT_LOCAL
)) {
2317 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2318 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2321 "%s: Bad dir inode log record, rec ptr 0x%p, "
2322 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2323 __func__
, item
, dip
, bp
, in_f
->ilf_ino
);
2324 error
= EFSCORRUPTED
;
2328 if (unlikely(dicp
->di_nextents
+ dicp
->di_anextents
> dicp
->di_nblocks
)){
2329 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2330 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2333 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2334 "dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2335 __func__
, item
, dip
, bp
, in_f
->ilf_ino
,
2336 dicp
->di_nextents
+ dicp
->di_anextents
,
2338 error
= EFSCORRUPTED
;
2341 if (unlikely(dicp
->di_forkoff
> mp
->m_sb
.sb_inodesize
)) {
2342 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2343 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2346 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2347 "dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__
,
2348 item
, dip
, bp
, in_f
->ilf_ino
, dicp
->di_forkoff
);
2349 error
= EFSCORRUPTED
;
2352 if (unlikely(item
->ri_buf
[1].i_len
> sizeof(struct xfs_icdinode
))) {
2353 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2354 XFS_ERRLEVEL_LOW
, mp
, dicp
);
2357 "%s: Bad inode log record length %d, rec ptr 0x%p",
2358 __func__
, item
->ri_buf
[1].i_len
, item
);
2359 error
= EFSCORRUPTED
;
2363 /* The core is in in-core format */
2364 xfs_dinode_to_disk(dip
, item
->ri_buf
[1].i_addr
);
2366 /* the rest is in on-disk format */
2367 if (item
->ri_buf
[1].i_len
> sizeof(struct xfs_icdinode
)) {
2368 memcpy((xfs_caddr_t
) dip
+ sizeof(struct xfs_icdinode
),
2369 item
->ri_buf
[1].i_addr
+ sizeof(struct xfs_icdinode
),
2370 item
->ri_buf
[1].i_len
- sizeof(struct xfs_icdinode
));
2373 fields
= in_f
->ilf_fields
;
2374 switch (fields
& (XFS_ILOG_DEV
| XFS_ILOG_UUID
)) {
2376 xfs_dinode_put_rdev(dip
, in_f
->ilf_u
.ilfu_rdev
);
2379 memcpy(XFS_DFORK_DPTR(dip
),
2380 &in_f
->ilf_u
.ilfu_uuid
,
2385 if (in_f
->ilf_size
== 2)
2386 goto write_inode_buffer
;
2387 len
= item
->ri_buf
[2].i_len
;
2388 src
= item
->ri_buf
[2].i_addr
;
2389 ASSERT(in_f
->ilf_size
<= 4);
2390 ASSERT((in_f
->ilf_size
== 3) || (fields
& XFS_ILOG_AFORK
));
2391 ASSERT(!(fields
& XFS_ILOG_DFORK
) ||
2392 (len
== in_f
->ilf_dsize
));
2394 switch (fields
& XFS_ILOG_DFORK
) {
2395 case XFS_ILOG_DDATA
:
2397 memcpy(XFS_DFORK_DPTR(dip
), src
, len
);
2400 case XFS_ILOG_DBROOT
:
2401 xfs_bmbt_to_bmdr(mp
, (struct xfs_btree_block
*)src
, len
,
2402 (xfs_bmdr_block_t
*)XFS_DFORK_DPTR(dip
),
2403 XFS_DFORK_DSIZE(dip
, mp
));
2408 * There are no data fork flags set.
2410 ASSERT((fields
& XFS_ILOG_DFORK
) == 0);
2415 * If we logged any attribute data, recover it. There may or
2416 * may not have been any other non-core data logged in this
2419 if (in_f
->ilf_fields
& XFS_ILOG_AFORK
) {
2420 if (in_f
->ilf_fields
& XFS_ILOG_DFORK
) {
2425 len
= item
->ri_buf
[attr_index
].i_len
;
2426 src
= item
->ri_buf
[attr_index
].i_addr
;
2427 ASSERT(len
== in_f
->ilf_asize
);
2429 switch (in_f
->ilf_fields
& XFS_ILOG_AFORK
) {
2430 case XFS_ILOG_ADATA
:
2432 dest
= XFS_DFORK_APTR(dip
);
2433 ASSERT(len
<= XFS_DFORK_ASIZE(dip
, mp
));
2434 memcpy(dest
, src
, len
);
2437 case XFS_ILOG_ABROOT
:
2438 dest
= XFS_DFORK_APTR(dip
);
2439 xfs_bmbt_to_bmdr(mp
, (struct xfs_btree_block
*)src
,
2440 len
, (xfs_bmdr_block_t
*)dest
,
2441 XFS_DFORK_ASIZE(dip
, mp
));
2445 xfs_warn(log
->l_mp
, "%s: Invalid flag", __func__
);
2454 ASSERT(bp
->b_target
->bt_mount
== mp
);
2455 bp
->b_iodone
= xlog_recover_iodone
;
2456 xfs_buf_delwri_queue(bp
, buffer_list
);
2461 return XFS_ERROR(error
);
2465 * Recover QUOTAOFF records. We simply make a note of it in the xlog
2466 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2470 xlog_recover_quotaoff_pass1(
2472 struct xlog_recover_item
*item
)
2474 xfs_qoff_logformat_t
*qoff_f
= item
->ri_buf
[0].i_addr
;
2478 * The logitem format's flag tells us if this was user quotaoff,
2479 * group/project quotaoff or both.
2481 if (qoff_f
->qf_flags
& XFS_UQUOTA_ACCT
)
2482 log
->l_quotaoffs_flag
|= XFS_DQ_USER
;
2483 if (qoff_f
->qf_flags
& XFS_PQUOTA_ACCT
)
2484 log
->l_quotaoffs_flag
|= XFS_DQ_PROJ
;
2485 if (qoff_f
->qf_flags
& XFS_GQUOTA_ACCT
)
2486 log
->l_quotaoffs_flag
|= XFS_DQ_GROUP
;
2492 * Recover a dquot record
2495 xlog_recover_dquot_pass2(
2497 struct list_head
*buffer_list
,
2498 struct xlog_recover_item
*item
)
2500 xfs_mount_t
*mp
= log
->l_mp
;
2502 struct xfs_disk_dquot
*ddq
, *recddq
;
2504 xfs_dq_logformat_t
*dq_f
;
2509 * Filesystems are required to send in quota flags at mount time.
2511 if (mp
->m_qflags
== 0)
2514 recddq
= item
->ri_buf
[1].i_addr
;
2515 if (recddq
== NULL
) {
2516 xfs_alert(log
->l_mp
, "NULL dquot in %s.", __func__
);
2517 return XFS_ERROR(EIO
);
2519 if (item
->ri_buf
[1].i_len
< sizeof(xfs_disk_dquot_t
)) {
2520 xfs_alert(log
->l_mp
, "dquot too small (%d) in %s.",
2521 item
->ri_buf
[1].i_len
, __func__
);
2522 return XFS_ERROR(EIO
);
2526 * This type of quotas was turned off, so ignore this record.
2528 type
= recddq
->d_flags
& (XFS_DQ_USER
| XFS_DQ_PROJ
| XFS_DQ_GROUP
);
2530 if (log
->l_quotaoffs_flag
& type
)
2534 * At this point we know that quota was _not_ turned off.
2535 * Since the mount flags are not indicating to us otherwise, this
2536 * must mean that quota is on, and the dquot needs to be replayed.
2537 * Remember that we may not have fully recovered the superblock yet,
2538 * so we can't do the usual trick of looking at the SB quota bits.
2540 * The other possibility, of course, is that the quota subsystem was
2541 * removed since the last mount - ENOSYS.
2543 dq_f
= item
->ri_buf
[0].i_addr
;
2545 error
= xfs_qm_dqcheck(mp
, recddq
, dq_f
->qlf_id
, 0, XFS_QMOPT_DOWARN
,
2546 "xlog_recover_dquot_pass2 (log copy)");
2548 return XFS_ERROR(EIO
);
2549 ASSERT(dq_f
->qlf_len
== 1);
2551 error
= xfs_trans_read_buf(mp
, NULL
, mp
->m_ddev_targp
, dq_f
->qlf_blkno
,
2552 XFS_FSB_TO_BB(mp
, dq_f
->qlf_len
), 0, &bp
,
2558 ddq
= (xfs_disk_dquot_t
*)xfs_buf_offset(bp
, dq_f
->qlf_boffset
);
2561 * At least the magic num portion should be on disk because this
2562 * was among a chunk of dquots created earlier, and we did some
2563 * minimal initialization then.
2565 error
= xfs_qm_dqcheck(mp
, ddq
, dq_f
->qlf_id
, 0, XFS_QMOPT_DOWARN
,
2566 "xlog_recover_dquot_pass2");
2569 return XFS_ERROR(EIO
);
2572 memcpy(ddq
, recddq
, item
->ri_buf
[1].i_len
);
2574 ASSERT(dq_f
->qlf_size
== 2);
2575 ASSERT(bp
->b_target
->bt_mount
== mp
);
2576 bp
->b_iodone
= xlog_recover_iodone
;
2577 xfs_buf_delwri_queue(bp
, buffer_list
);
2584 * This routine is called to create an in-core extent free intent
2585 * item from the efi format structure which was logged on disk.
2586 * It allocates an in-core efi, copies the extents from the format
2587 * structure into it, and adds the efi to the AIL with the given
2591 xlog_recover_efi_pass2(
2593 struct xlog_recover_item
*item
,
2597 xfs_mount_t
*mp
= log
->l_mp
;
2598 xfs_efi_log_item_t
*efip
;
2599 xfs_efi_log_format_t
*efi_formatp
;
2601 efi_formatp
= item
->ri_buf
[0].i_addr
;
2603 efip
= xfs_efi_init(mp
, efi_formatp
->efi_nextents
);
2604 if ((error
= xfs_efi_copy_format(&(item
->ri_buf
[0]),
2605 &(efip
->efi_format
)))) {
2606 xfs_efi_item_free(efip
);
2609 atomic_set(&efip
->efi_next_extent
, efi_formatp
->efi_nextents
);
2611 spin_lock(&log
->l_ailp
->xa_lock
);
2613 * xfs_trans_ail_update() drops the AIL lock.
2615 xfs_trans_ail_update(log
->l_ailp
, &efip
->efi_item
, lsn
);
2621 * This routine is called when an efd format structure is found in
2622 * a committed transaction in the log. It's purpose is to cancel
2623 * the corresponding efi if it was still in the log. To do this
2624 * it searches the AIL for the efi with an id equal to that in the
2625 * efd format structure. If we find it, we remove the efi from the
2629 xlog_recover_efd_pass2(
2631 struct xlog_recover_item
*item
)
2633 xfs_efd_log_format_t
*efd_formatp
;
2634 xfs_efi_log_item_t
*efip
= NULL
;
2635 xfs_log_item_t
*lip
;
2637 struct xfs_ail_cursor cur
;
2638 struct xfs_ail
*ailp
= log
->l_ailp
;
2640 efd_formatp
= item
->ri_buf
[0].i_addr
;
2641 ASSERT((item
->ri_buf
[0].i_len
== (sizeof(xfs_efd_log_format_32_t
) +
2642 ((efd_formatp
->efd_nextents
- 1) * sizeof(xfs_extent_32_t
)))) ||
2643 (item
->ri_buf
[0].i_len
== (sizeof(xfs_efd_log_format_64_t
) +
2644 ((efd_formatp
->efd_nextents
- 1) * sizeof(xfs_extent_64_t
)))));
2645 efi_id
= efd_formatp
->efd_efi_id
;
2648 * Search for the efi with the id in the efd format structure
2651 spin_lock(&ailp
->xa_lock
);
2652 lip
= xfs_trans_ail_cursor_first(ailp
, &cur
, 0);
2653 while (lip
!= NULL
) {
2654 if (lip
->li_type
== XFS_LI_EFI
) {
2655 efip
= (xfs_efi_log_item_t
*)lip
;
2656 if (efip
->efi_format
.efi_id
== efi_id
) {
2658 * xfs_trans_ail_delete() drops the
2661 xfs_trans_ail_delete(ailp
, lip
,
2662 SHUTDOWN_CORRUPT_INCORE
);
2663 xfs_efi_item_free(efip
);
2664 spin_lock(&ailp
->xa_lock
);
2668 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
2670 xfs_trans_ail_cursor_done(ailp
, &cur
);
2671 spin_unlock(&ailp
->xa_lock
);
2677 * Free up any resources allocated by the transaction
2679 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2682 xlog_recover_free_trans(
2683 struct xlog_recover
*trans
)
2685 xlog_recover_item_t
*item
, *n
;
2688 list_for_each_entry_safe(item
, n
, &trans
->r_itemq
, ri_list
) {
2689 /* Free the regions in the item. */
2690 list_del(&item
->ri_list
);
2691 for (i
= 0; i
< item
->ri_cnt
; i
++)
2692 kmem_free(item
->ri_buf
[i
].i_addr
);
2693 /* Free the item itself */
2694 kmem_free(item
->ri_buf
);
2697 /* Free the transaction recover structure */
2702 xlog_recover_commit_pass1(
2704 struct xlog_recover
*trans
,
2705 struct xlog_recover_item
*item
)
2707 trace_xfs_log_recover_item_recover(log
, trans
, item
, XLOG_RECOVER_PASS1
);
2709 switch (ITEM_TYPE(item
)) {
2711 return xlog_recover_buffer_pass1(log
, item
);
2712 case XFS_LI_QUOTAOFF
:
2713 return xlog_recover_quotaoff_pass1(log
, item
);
2718 /* nothing to do in pass 1 */
2721 xfs_warn(log
->l_mp
, "%s: invalid item type (%d)",
2722 __func__
, ITEM_TYPE(item
));
2724 return XFS_ERROR(EIO
);
2729 xlog_recover_commit_pass2(
2731 struct xlog_recover
*trans
,
2732 struct list_head
*buffer_list
,
2733 struct xlog_recover_item
*item
)
2735 trace_xfs_log_recover_item_recover(log
, trans
, item
, XLOG_RECOVER_PASS2
);
2737 switch (ITEM_TYPE(item
)) {
2739 return xlog_recover_buffer_pass2(log
, buffer_list
, item
);
2741 return xlog_recover_inode_pass2(log
, buffer_list
, item
);
2743 return xlog_recover_efi_pass2(log
, item
, trans
->r_lsn
);
2745 return xlog_recover_efd_pass2(log
, item
);
2747 return xlog_recover_dquot_pass2(log
, buffer_list
, item
);
2748 case XFS_LI_QUOTAOFF
:
2749 /* nothing to do in pass2 */
2752 xfs_warn(log
->l_mp
, "%s: invalid item type (%d)",
2753 __func__
, ITEM_TYPE(item
));
2755 return XFS_ERROR(EIO
);
2760 * Perform the transaction.
2762 * If the transaction modifies a buffer or inode, do it now. Otherwise,
2763 * EFIs and EFDs get queued up by adding entries into the AIL for them.
2766 xlog_recover_commit_trans(
2768 struct xlog_recover
*trans
,
2771 int error
= 0, error2
;
2772 xlog_recover_item_t
*item
;
2773 LIST_HEAD (buffer_list
);
2775 hlist_del(&trans
->r_list
);
2777 error
= xlog_recover_reorder_trans(log
, trans
, pass
);
2781 list_for_each_entry(item
, &trans
->r_itemq
, ri_list
) {
2783 case XLOG_RECOVER_PASS1
:
2784 error
= xlog_recover_commit_pass1(log
, trans
, item
);
2786 case XLOG_RECOVER_PASS2
:
2787 error
= xlog_recover_commit_pass2(log
, trans
,
2788 &buffer_list
, item
);
2798 xlog_recover_free_trans(trans
);
2801 error2
= xfs_buf_delwri_submit(&buffer_list
);
2802 return error
? error
: error2
;
2806 xlog_recover_unmount_trans(
2808 struct xlog_recover
*trans
)
2810 /* Do nothing now */
2811 xfs_warn(log
->l_mp
, "%s: Unmount LR", __func__
);
2816 * There are two valid states of the r_state field. 0 indicates that the
2817 * transaction structure is in a normal state. We have either seen the
2818 * start of the transaction or the last operation we added was not a partial
2819 * operation. If the last operation we added to the transaction was a
2820 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2822 * NOTE: skip LRs with 0 data length.
2825 xlog_recover_process_data(
2827 struct hlist_head rhash
[],
2828 struct xlog_rec_header
*rhead
,
2834 xlog_op_header_t
*ohead
;
2835 xlog_recover_t
*trans
;
2841 lp
= dp
+ be32_to_cpu(rhead
->h_len
);
2842 num_logops
= be32_to_cpu(rhead
->h_num_logops
);
2844 /* check the log format matches our own - else we can't recover */
2845 if (xlog_header_check_recover(log
->l_mp
, rhead
))
2846 return (XFS_ERROR(EIO
));
2848 while ((dp
< lp
) && num_logops
) {
2849 ASSERT(dp
+ sizeof(xlog_op_header_t
) <= lp
);
2850 ohead
= (xlog_op_header_t
*)dp
;
2851 dp
+= sizeof(xlog_op_header_t
);
2852 if (ohead
->oh_clientid
!= XFS_TRANSACTION
&&
2853 ohead
->oh_clientid
!= XFS_LOG
) {
2854 xfs_warn(log
->l_mp
, "%s: bad clientid 0x%x",
2855 __func__
, ohead
->oh_clientid
);
2857 return (XFS_ERROR(EIO
));
2859 tid
= be32_to_cpu(ohead
->oh_tid
);
2860 hash
= XLOG_RHASH(tid
);
2861 trans
= xlog_recover_find_tid(&rhash
[hash
], tid
);
2862 if (trans
== NULL
) { /* not found; add new tid */
2863 if (ohead
->oh_flags
& XLOG_START_TRANS
)
2864 xlog_recover_new_tid(&rhash
[hash
], tid
,
2865 be64_to_cpu(rhead
->h_lsn
));
2867 if (dp
+ be32_to_cpu(ohead
->oh_len
) > lp
) {
2868 xfs_warn(log
->l_mp
, "%s: bad length 0x%x",
2869 __func__
, be32_to_cpu(ohead
->oh_len
));
2871 return (XFS_ERROR(EIO
));
2873 flags
= ohead
->oh_flags
& ~XLOG_END_TRANS
;
2874 if (flags
& XLOG_WAS_CONT_TRANS
)
2875 flags
&= ~XLOG_CONTINUE_TRANS
;
2877 case XLOG_COMMIT_TRANS
:
2878 error
= xlog_recover_commit_trans(log
,
2881 case XLOG_UNMOUNT_TRANS
:
2882 error
= xlog_recover_unmount_trans(log
, trans
);
2884 case XLOG_WAS_CONT_TRANS
:
2885 error
= xlog_recover_add_to_cont_trans(log
,
2887 be32_to_cpu(ohead
->oh_len
));
2889 case XLOG_START_TRANS
:
2890 xfs_warn(log
->l_mp
, "%s: bad transaction",
2893 error
= XFS_ERROR(EIO
);
2896 case XLOG_CONTINUE_TRANS
:
2897 error
= xlog_recover_add_to_trans(log
, trans
,
2898 dp
, be32_to_cpu(ohead
->oh_len
));
2901 xfs_warn(log
->l_mp
, "%s: bad flag 0x%x",
2904 error
= XFS_ERROR(EIO
);
2910 dp
+= be32_to_cpu(ohead
->oh_len
);
2917 * Process an extent free intent item that was recovered from
2918 * the log. We need to free the extents that it describes.
2921 xlog_recover_process_efi(
2923 xfs_efi_log_item_t
*efip
)
2925 xfs_efd_log_item_t
*efdp
;
2930 xfs_fsblock_t startblock_fsb
;
2932 ASSERT(!test_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
));
2935 * First check the validity of the extents described by the
2936 * EFI. If any are bad, then assume that all are bad and
2937 * just toss the EFI.
2939 for (i
= 0; i
< efip
->efi_format
.efi_nextents
; i
++) {
2940 extp
= &(efip
->efi_format
.efi_extents
[i
]);
2941 startblock_fsb
= XFS_BB_TO_FSB(mp
,
2942 XFS_FSB_TO_DADDR(mp
, extp
->ext_start
));
2943 if ((startblock_fsb
== 0) ||
2944 (extp
->ext_len
== 0) ||
2945 (startblock_fsb
>= mp
->m_sb
.sb_dblocks
) ||
2946 (extp
->ext_len
>= mp
->m_sb
.sb_agblocks
)) {
2948 * This will pull the EFI from the AIL and
2949 * free the memory associated with it.
2951 xfs_efi_release(efip
, efip
->efi_format
.efi_nextents
);
2952 return XFS_ERROR(EIO
);
2956 tp
= xfs_trans_alloc(mp
, 0);
2957 error
= xfs_trans_reserve(tp
, 0, XFS_ITRUNCATE_LOG_RES(mp
), 0, 0, 0);
2960 efdp
= xfs_trans_get_efd(tp
, efip
, efip
->efi_format
.efi_nextents
);
2962 for (i
= 0; i
< efip
->efi_format
.efi_nextents
; i
++) {
2963 extp
= &(efip
->efi_format
.efi_extents
[i
]);
2964 error
= xfs_free_extent(tp
, extp
->ext_start
, extp
->ext_len
);
2967 xfs_trans_log_efd_extent(tp
, efdp
, extp
->ext_start
,
2971 set_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
);
2972 error
= xfs_trans_commit(tp
, 0);
2976 xfs_trans_cancel(tp
, XFS_TRANS_ABORT
);
2981 * When this is called, all of the EFIs which did not have
2982 * corresponding EFDs should be in the AIL. What we do now
2983 * is free the extents associated with each one.
2985 * Since we process the EFIs in normal transactions, they
2986 * will be removed at some point after the commit. This prevents
2987 * us from just walking down the list processing each one.
2988 * We'll use a flag in the EFI to skip those that we've already
2989 * processed and use the AIL iteration mechanism's generation
2990 * count to try to speed this up at least a bit.
2992 * When we start, we know that the EFIs are the only things in
2993 * the AIL. As we process them, however, other items are added
2994 * to the AIL. Since everything added to the AIL must come after
2995 * everything already in the AIL, we stop processing as soon as
2996 * we see something other than an EFI in the AIL.
2999 xlog_recover_process_efis(
3002 xfs_log_item_t
*lip
;
3003 xfs_efi_log_item_t
*efip
;
3005 struct xfs_ail_cursor cur
;
3006 struct xfs_ail
*ailp
;
3009 spin_lock(&ailp
->xa_lock
);
3010 lip
= xfs_trans_ail_cursor_first(ailp
, &cur
, 0);
3011 while (lip
!= NULL
) {
3013 * We're done when we see something other than an EFI.
3014 * There should be no EFIs left in the AIL now.
3016 if (lip
->li_type
!= XFS_LI_EFI
) {
3018 for (; lip
; lip
= xfs_trans_ail_cursor_next(ailp
, &cur
))
3019 ASSERT(lip
->li_type
!= XFS_LI_EFI
);
3025 * Skip EFIs that we've already processed.
3027 efip
= (xfs_efi_log_item_t
*)lip
;
3028 if (test_bit(XFS_EFI_RECOVERED
, &efip
->efi_flags
)) {
3029 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3033 spin_unlock(&ailp
->xa_lock
);
3034 error
= xlog_recover_process_efi(log
->l_mp
, efip
);
3035 spin_lock(&ailp
->xa_lock
);
3038 lip
= xfs_trans_ail_cursor_next(ailp
, &cur
);
3041 xfs_trans_ail_cursor_done(ailp
, &cur
);
3042 spin_unlock(&ailp
->xa_lock
);
3047 * This routine performs a transaction to null out a bad inode pointer
3048 * in an agi unlinked inode hash bucket.
3051 xlog_recover_clear_agi_bucket(
3053 xfs_agnumber_t agno
,
3062 tp
= xfs_trans_alloc(mp
, XFS_TRANS_CLEAR_AGI_BUCKET
);
3063 error
= xfs_trans_reserve(tp
, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp
),
3068 error
= xfs_read_agi(mp
, tp
, agno
, &agibp
);
3072 agi
= XFS_BUF_TO_AGI(agibp
);
3073 agi
->agi_unlinked
[bucket
] = cpu_to_be32(NULLAGINO
);
3074 offset
= offsetof(xfs_agi_t
, agi_unlinked
) +
3075 (sizeof(xfs_agino_t
) * bucket
);
3076 xfs_trans_log_buf(tp
, agibp
, offset
,
3077 (offset
+ sizeof(xfs_agino_t
) - 1));
3079 error
= xfs_trans_commit(tp
, 0);
3085 xfs_trans_cancel(tp
, XFS_TRANS_ABORT
);
3087 xfs_warn(mp
, "%s: failed to clear agi %d. Continuing.", __func__
, agno
);
3092 xlog_recover_process_one_iunlink(
3093 struct xfs_mount
*mp
,
3094 xfs_agnumber_t agno
,
3098 struct xfs_buf
*ibp
;
3099 struct xfs_dinode
*dip
;
3100 struct xfs_inode
*ip
;
3104 ino
= XFS_AGINO_TO_INO(mp
, agno
, agino
);
3105 error
= xfs_iget(mp
, NULL
, ino
, 0, 0, &ip
);
3110 * Get the on disk inode to find the next inode in the bucket.
3112 error
= xfs_imap_to_bp(mp
, NULL
, &ip
->i_imap
, &dip
, &ibp
, 0, 0);
3116 ASSERT(ip
->i_d
.di_nlink
== 0);
3117 ASSERT(ip
->i_d
.di_mode
!= 0);
3119 /* setup for the next pass */
3120 agino
= be32_to_cpu(dip
->di_next_unlinked
);
3124 * Prevent any DMAPI event from being sent when the reference on
3125 * the inode is dropped.
3127 ip
->i_d
.di_dmevmask
= 0;
3136 * We can't read in the inode this bucket points to, or this inode
3137 * is messed up. Just ditch this bucket of inodes. We will lose
3138 * some inodes and space, but at least we won't hang.
3140 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3141 * clear the inode pointer in the bucket.
3143 xlog_recover_clear_agi_bucket(mp
, agno
, bucket
);
3148 * xlog_iunlink_recover
3150 * This is called during recovery to process any inodes which
3151 * we unlinked but not freed when the system crashed. These
3152 * inodes will be on the lists in the AGI blocks. What we do
3153 * here is scan all the AGIs and fully truncate and free any
3154 * inodes found on the lists. Each inode is removed from the
3155 * lists when it has been fully truncated and is freed. The
3156 * freeing of the inode and its removal from the list must be
3160 xlog_recover_process_iunlinks(
3164 xfs_agnumber_t agno
;
3175 * Prevent any DMAPI event from being sent while in this function.
3177 mp_dmevmask
= mp
->m_dmevmask
;
3180 for (agno
= 0; agno
< mp
->m_sb
.sb_agcount
; agno
++) {
3182 * Find the agi for this ag.
3184 error
= xfs_read_agi(mp
, NULL
, agno
, &agibp
);
3187 * AGI is b0rked. Don't process it.
3189 * We should probably mark the filesystem as corrupt
3190 * after we've recovered all the ag's we can....
3195 * Unlock the buffer so that it can be acquired in the normal
3196 * course of the transaction to truncate and free each inode.
3197 * Because we are not racing with anyone else here for the AGI
3198 * buffer, we don't even need to hold it locked to read the
3199 * initial unlinked bucket entries out of the buffer. We keep
3200 * buffer reference though, so that it stays pinned in memory
3201 * while we need the buffer.
3203 agi
= XFS_BUF_TO_AGI(agibp
);
3204 xfs_buf_unlock(agibp
);
3206 for (bucket
= 0; bucket
< XFS_AGI_UNLINKED_BUCKETS
; bucket
++) {
3207 agino
= be32_to_cpu(agi
->agi_unlinked
[bucket
]);
3208 while (agino
!= NULLAGINO
) {
3209 agino
= xlog_recover_process_one_iunlink(mp
,
3210 agno
, agino
, bucket
);
3213 xfs_buf_rele(agibp
);
3216 mp
->m_dmevmask
= mp_dmevmask
;
3220 * Upack the log buffer data and crc check it. If the check fails, issue a
3221 * warning if and only if the CRC in the header is non-zero. This makes the
3222 * check an advisory warning, and the zero CRC check will prevent failure
3223 * warnings from being emitted when upgrading the kernel from one that does not
3224 * add CRCs by default.
3226 * When filesystems are CRC enabled, this CRC mismatch becomes a fatal log
3227 * corruption failure
3230 xlog_unpack_data_crc(
3231 struct xlog_rec_header
*rhead
,
3237 crc
= xlog_cksum(log
, rhead
, dp
, be32_to_cpu(rhead
->h_len
));
3238 if (crc
!= rhead
->h_crc
) {
3239 if (rhead
->h_crc
|| xfs_sb_version_hascrc(&log
->l_mp
->m_sb
)) {
3240 xfs_alert(log
->l_mp
,
3241 "log record CRC mismatch: found 0x%x, expected 0x%x.\n",
3242 le32_to_cpu(rhead
->h_crc
),
3244 xfs_hex_dump(dp
, 32);
3248 * If we've detected a log record corruption, then we can't
3249 * recover past this point. Abort recovery if we are enforcing
3250 * CRC protection by punting an error back up the stack.
3252 if (xfs_sb_version_hascrc(&log
->l_mp
->m_sb
))
3253 return EFSCORRUPTED
;
3261 struct xlog_rec_header
*rhead
,
3268 error
= xlog_unpack_data_crc(rhead
, dp
, log
);
3272 for (i
= 0; i
< BTOBB(be32_to_cpu(rhead
->h_len
)) &&
3273 i
< (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
); i
++) {
3274 *(__be32
*)dp
= *(__be32
*)&rhead
->h_cycle_data
[i
];
3278 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
3279 xlog_in_core_2_t
*xhdr
= (xlog_in_core_2_t
*)rhead
;
3280 for ( ; i
< BTOBB(be32_to_cpu(rhead
->h_len
)); i
++) {
3281 j
= i
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3282 k
= i
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3283 *(__be32
*)dp
= xhdr
[j
].hic_xheader
.xh_cycle_data
[k
];
3292 xlog_valid_rec_header(
3294 struct xlog_rec_header
*rhead
,
3299 if (unlikely(rhead
->h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))) {
3300 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3301 XFS_ERRLEVEL_LOW
, log
->l_mp
);
3302 return XFS_ERROR(EFSCORRUPTED
);
3305 (!rhead
->h_version
||
3306 (be32_to_cpu(rhead
->h_version
) & (~XLOG_VERSION_OKBITS
))))) {
3307 xfs_warn(log
->l_mp
, "%s: unrecognised log version (%d).",
3308 __func__
, be32_to_cpu(rhead
->h_version
));
3309 return XFS_ERROR(EIO
);
3312 /* LR body must have data or it wouldn't have been written */
3313 hlen
= be32_to_cpu(rhead
->h_len
);
3314 if (unlikely( hlen
<= 0 || hlen
> INT_MAX
)) {
3315 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3316 XFS_ERRLEVEL_LOW
, log
->l_mp
);
3317 return XFS_ERROR(EFSCORRUPTED
);
3319 if (unlikely( blkno
> log
->l_logBBsize
|| blkno
> INT_MAX
)) {
3320 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3321 XFS_ERRLEVEL_LOW
, log
->l_mp
);
3322 return XFS_ERROR(EFSCORRUPTED
);
3328 * Read the log from tail to head and process the log records found.
3329 * Handle the two cases where the tail and head are in the same cycle
3330 * and where the active portion of the log wraps around the end of
3331 * the physical log separately. The pass parameter is passed through
3332 * to the routines called to process the data and is not looked at
3336 xlog_do_recovery_pass(
3338 xfs_daddr_t head_blk
,
3339 xfs_daddr_t tail_blk
,
3342 xlog_rec_header_t
*rhead
;
3345 xfs_buf_t
*hbp
, *dbp
;
3346 int error
= 0, h_size
;
3347 int bblks
, split_bblks
;
3348 int hblks
, split_hblks
, wrapped_hblks
;
3349 struct hlist_head rhash
[XLOG_RHASH_SIZE
];
3351 ASSERT(head_blk
!= tail_blk
);
3354 * Read the header of the tail block and get the iclog buffer size from
3355 * h_size. Use this to tell how many sectors make up the log header.
3357 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
3359 * When using variable length iclogs, read first sector of
3360 * iclog header and extract the header size from it. Get a
3361 * new hbp that is the correct size.
3363 hbp
= xlog_get_bp(log
, 1);
3367 error
= xlog_bread(log
, tail_blk
, 1, hbp
, &offset
);
3371 rhead
= (xlog_rec_header_t
*)offset
;
3372 error
= xlog_valid_rec_header(log
, rhead
, tail_blk
);
3375 h_size
= be32_to_cpu(rhead
->h_size
);
3376 if ((be32_to_cpu(rhead
->h_version
) & XLOG_VERSION_2
) &&
3377 (h_size
> XLOG_HEADER_CYCLE_SIZE
)) {
3378 hblks
= h_size
/ XLOG_HEADER_CYCLE_SIZE
;
3379 if (h_size
% XLOG_HEADER_CYCLE_SIZE
)
3382 hbp
= xlog_get_bp(log
, hblks
);
3387 ASSERT(log
->l_sectBBsize
== 1);
3389 hbp
= xlog_get_bp(log
, 1);
3390 h_size
= XLOG_BIG_RECORD_BSIZE
;
3395 dbp
= xlog_get_bp(log
, BTOBB(h_size
));
3401 memset(rhash
, 0, sizeof(rhash
));
3402 if (tail_blk
<= head_blk
) {
3403 for (blk_no
= tail_blk
; blk_no
< head_blk
; ) {
3404 error
= xlog_bread(log
, blk_no
, hblks
, hbp
, &offset
);
3408 rhead
= (xlog_rec_header_t
*)offset
;
3409 error
= xlog_valid_rec_header(log
, rhead
, blk_no
);
3413 /* blocks in data section */
3414 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
3415 error
= xlog_bread(log
, blk_no
+ hblks
, bblks
, dbp
,
3420 error
= xlog_unpack_data(rhead
, offset
, log
);
3424 error
= xlog_recover_process_data(log
,
3425 rhash
, rhead
, offset
, pass
);
3428 blk_no
+= bblks
+ hblks
;
3432 * Perform recovery around the end of the physical log.
3433 * When the head is not on the same cycle number as the tail,
3434 * we can't do a sequential recovery as above.
3437 while (blk_no
< log
->l_logBBsize
) {
3439 * Check for header wrapping around physical end-of-log
3441 offset
= hbp
->b_addr
;
3444 if (blk_no
+ hblks
<= log
->l_logBBsize
) {
3445 /* Read header in one read */
3446 error
= xlog_bread(log
, blk_no
, hblks
, hbp
,
3451 /* This LR is split across physical log end */
3452 if (blk_no
!= log
->l_logBBsize
) {
3453 /* some data before physical log end */
3454 ASSERT(blk_no
<= INT_MAX
);
3455 split_hblks
= log
->l_logBBsize
- (int)blk_no
;
3456 ASSERT(split_hblks
> 0);
3457 error
= xlog_bread(log
, blk_no
,
3465 * Note: this black magic still works with
3466 * large sector sizes (non-512) only because:
3467 * - we increased the buffer size originally
3468 * by 1 sector giving us enough extra space
3469 * for the second read;
3470 * - the log start is guaranteed to be sector
3472 * - we read the log end (LR header start)
3473 * _first_, then the log start (LR header end)
3474 * - order is important.
3476 wrapped_hblks
= hblks
- split_hblks
;
3477 error
= xlog_bread_offset(log
, 0,
3479 offset
+ BBTOB(split_hblks
));
3483 rhead
= (xlog_rec_header_t
*)offset
;
3484 error
= xlog_valid_rec_header(log
, rhead
,
3485 split_hblks
? blk_no
: 0);
3489 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
3492 /* Read in data for log record */
3493 if (blk_no
+ bblks
<= log
->l_logBBsize
) {
3494 error
= xlog_bread(log
, blk_no
, bblks
, dbp
,
3499 /* This log record is split across the
3500 * physical end of log */
3501 offset
= dbp
->b_addr
;
3503 if (blk_no
!= log
->l_logBBsize
) {
3504 /* some data is before the physical
3506 ASSERT(!wrapped_hblks
);
3507 ASSERT(blk_no
<= INT_MAX
);
3509 log
->l_logBBsize
- (int)blk_no
;
3510 ASSERT(split_bblks
> 0);
3511 error
= xlog_bread(log
, blk_no
,
3519 * Note: this black magic still works with
3520 * large sector sizes (non-512) only because:
3521 * - we increased the buffer size originally
3522 * by 1 sector giving us enough extra space
3523 * for the second read;
3524 * - the log start is guaranteed to be sector
3526 * - we read the log end (LR header start)
3527 * _first_, then the log start (LR header end)
3528 * - order is important.
3530 error
= xlog_bread_offset(log
, 0,
3531 bblks
- split_bblks
, dbp
,
3532 offset
+ BBTOB(split_bblks
));
3537 error
= xlog_unpack_data(rhead
, offset
, log
);
3541 error
= xlog_recover_process_data(log
, rhash
,
3542 rhead
, offset
, pass
);
3548 ASSERT(blk_no
>= log
->l_logBBsize
);
3549 blk_no
-= log
->l_logBBsize
;
3551 /* read first part of physical log */
3552 while (blk_no
< head_blk
) {
3553 error
= xlog_bread(log
, blk_no
, hblks
, hbp
, &offset
);
3557 rhead
= (xlog_rec_header_t
*)offset
;
3558 error
= xlog_valid_rec_header(log
, rhead
, blk_no
);
3562 bblks
= (int)BTOBB(be32_to_cpu(rhead
->h_len
));
3563 error
= xlog_bread(log
, blk_no
+hblks
, bblks
, dbp
,
3568 error
= xlog_unpack_data(rhead
, offset
, log
);
3572 error
= xlog_recover_process_data(log
, rhash
,
3573 rhead
, offset
, pass
);
3576 blk_no
+= bblks
+ hblks
;
3588 * Do the recovery of the log. We actually do this in two phases.
3589 * The two passes are necessary in order to implement the function
3590 * of cancelling a record written into the log. The first pass
3591 * determines those things which have been cancelled, and the
3592 * second pass replays log items normally except for those which
3593 * have been cancelled. The handling of the replay and cancellations
3594 * takes place in the log item type specific routines.
3596 * The table of items which have cancel records in the log is allocated
3597 * and freed at this level, since only here do we know when all of
3598 * the log recovery has been completed.
3601 xlog_do_log_recovery(
3603 xfs_daddr_t head_blk
,
3604 xfs_daddr_t tail_blk
)
3608 ASSERT(head_blk
!= tail_blk
);
3611 * First do a pass to find all of the cancelled buf log items.
3612 * Store them in the buf_cancel_table for use in the second pass.
3614 log
->l_buf_cancel_table
= kmem_zalloc(XLOG_BC_TABLE_SIZE
*
3615 sizeof(struct list_head
),
3617 for (i
= 0; i
< XLOG_BC_TABLE_SIZE
; i
++)
3618 INIT_LIST_HEAD(&log
->l_buf_cancel_table
[i
]);
3620 error
= xlog_do_recovery_pass(log
, head_blk
, tail_blk
,
3621 XLOG_RECOVER_PASS1
);
3623 kmem_free(log
->l_buf_cancel_table
);
3624 log
->l_buf_cancel_table
= NULL
;
3628 * Then do a second pass to actually recover the items in the log.
3629 * When it is complete free the table of buf cancel items.
3631 error
= xlog_do_recovery_pass(log
, head_blk
, tail_blk
,
3632 XLOG_RECOVER_PASS2
);
3637 for (i
= 0; i
< XLOG_BC_TABLE_SIZE
; i
++)
3638 ASSERT(list_empty(&log
->l_buf_cancel_table
[i
]));
3642 kmem_free(log
->l_buf_cancel_table
);
3643 log
->l_buf_cancel_table
= NULL
;
3649 * Do the actual recovery
3654 xfs_daddr_t head_blk
,
3655 xfs_daddr_t tail_blk
)
3662 * First replay the images in the log.
3664 error
= xlog_do_log_recovery(log
, head_blk
, tail_blk
);
3669 * If IO errors happened during recovery, bail out.
3671 if (XFS_FORCED_SHUTDOWN(log
->l_mp
)) {
3676 * We now update the tail_lsn since much of the recovery has completed
3677 * and there may be space available to use. If there were no extent
3678 * or iunlinks, we can free up the entire log and set the tail_lsn to
3679 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3680 * lsn of the last known good LR on disk. If there are extent frees
3681 * or iunlinks they will have some entries in the AIL; so we look at
3682 * the AIL to determine how to set the tail_lsn.
3684 xlog_assign_tail_lsn(log
->l_mp
);
3687 * Now that we've finished replaying all buffer and inode
3688 * updates, re-read in the superblock and reverify it.
3690 bp
= xfs_getsb(log
->l_mp
, 0);
3692 ASSERT(!(XFS_BUF_ISWRITE(bp
)));
3694 XFS_BUF_UNASYNC(bp
);
3695 bp
->b_ops
= &xfs_sb_buf_ops
;
3696 xfsbdstrat(log
->l_mp
, bp
);
3697 error
= xfs_buf_iowait(bp
);
3699 xfs_buf_ioerror_alert(bp
, __func__
);
3705 /* Convert superblock from on-disk format */
3706 sbp
= &log
->l_mp
->m_sb
;
3707 xfs_sb_from_disk(sbp
, XFS_BUF_TO_SBP(bp
));
3708 ASSERT(sbp
->sb_magicnum
== XFS_SB_MAGIC
);
3709 ASSERT(xfs_sb_good_version(sbp
));
3712 /* We've re-read the superblock so re-initialize per-cpu counters */
3713 xfs_icsb_reinit_counters(log
->l_mp
);
3715 xlog_recover_check_summary(log
);
3717 /* Normal transactions can now occur */
3718 log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
3723 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3725 * Return error or zero.
3731 xfs_daddr_t head_blk
, tail_blk
;
3734 /* find the tail of the log */
3735 if ((error
= xlog_find_tail(log
, &head_blk
, &tail_blk
)))
3738 if (tail_blk
!= head_blk
) {
3739 /* There used to be a comment here:
3741 * disallow recovery on read-only mounts. note -- mount
3742 * checks for ENOSPC and turns it into an intelligent
3744 * ...but this is no longer true. Now, unless you specify
3745 * NORECOVERY (in which case this function would never be
3746 * called), we just go ahead and recover. We do this all
3747 * under the vfs layer, so we can get away with it unless
3748 * the device itself is read-only, in which case we fail.
3750 if ((error
= xfs_dev_is_read_only(log
->l_mp
, "recovery"))) {
3754 xfs_notice(log
->l_mp
, "Starting recovery (logdev: %s)",
3755 log
->l_mp
->m_logname
? log
->l_mp
->m_logname
3758 error
= xlog_do_recover(log
, head_blk
, tail_blk
);
3759 log
->l_flags
|= XLOG_RECOVERY_NEEDED
;
3765 * In the first part of recovery we replay inodes and buffers and build
3766 * up the list of extent free items which need to be processed. Here
3767 * we process the extent free items and clean up the on disk unlinked
3768 * inode lists. This is separated from the first part of recovery so
3769 * that the root and real-time bitmap inodes can be read in from disk in
3770 * between the two stages. This is necessary so that we can free space
3771 * in the real-time portion of the file system.
3774 xlog_recover_finish(
3778 * Now we're ready to do the transactions needed for the
3779 * rest of recovery. Start with completing all the extent
3780 * free intent records and then process the unlinked inode
3781 * lists. At this point, we essentially run in normal mode
3782 * except that we're still performing recovery actions
3783 * rather than accepting new requests.
3785 if (log
->l_flags
& XLOG_RECOVERY_NEEDED
) {
3787 error
= xlog_recover_process_efis(log
);
3789 xfs_alert(log
->l_mp
, "Failed to recover EFIs");
3793 * Sync the log to get all the EFIs out of the AIL.
3794 * This isn't absolutely necessary, but it helps in
3795 * case the unlink transactions would have problems
3796 * pushing the EFIs out of the way.
3798 xfs_log_force(log
->l_mp
, XFS_LOG_SYNC
);
3800 xlog_recover_process_iunlinks(log
);
3802 xlog_recover_check_summary(log
);
3804 xfs_notice(log
->l_mp
, "Ending recovery (logdev: %s)",
3805 log
->l_mp
->m_logname
? log
->l_mp
->m_logname
3807 log
->l_flags
&= ~XLOG_RECOVERY_NEEDED
;
3809 xfs_info(log
->l_mp
, "Ending clean mount");
3817 * Read all of the agf and agi counters and check that they
3818 * are consistent with the superblock counters.
3821 xlog_recover_check_summary(
3828 xfs_agnumber_t agno
;
3829 __uint64_t freeblks
;
3839 for (agno
= 0; agno
< mp
->m_sb
.sb_agcount
; agno
++) {
3840 error
= xfs_read_agf(mp
, NULL
, agno
, 0, &agfbp
);
3842 xfs_alert(mp
, "%s agf read failed agno %d error %d",
3843 __func__
, agno
, error
);
3845 agfp
= XFS_BUF_TO_AGF(agfbp
);
3846 freeblks
+= be32_to_cpu(agfp
->agf_freeblks
) +
3847 be32_to_cpu(agfp
->agf_flcount
);
3848 xfs_buf_relse(agfbp
);
3851 error
= xfs_read_agi(mp
, NULL
, agno
, &agibp
);
3853 xfs_alert(mp
, "%s agi read failed agno %d error %d",
3854 __func__
, agno
, error
);
3856 struct xfs_agi
*agi
= XFS_BUF_TO_AGI(agibp
);
3858 itotal
+= be32_to_cpu(agi
->agi_count
);
3859 ifree
+= be32_to_cpu(agi
->agi_freecount
);
3860 xfs_buf_relse(agibp
);