Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge tag 'for-linus-v3.11-rc3' of git://oss.sgi.com/xfs/xfs
Pull xfs fix from Ben Myers:
"Fix for regression in commit cca9f93a52d2 ("xfs: don't do IO when
creating an new inode"), recovery causing filesystem corruption after
a crash"
* tag 'for-linus-v3.11-rc3' of git://oss.sgi.com/xfs/xfs:
xfs: di_flushiter considered harmful
+36
-11
+3
fs/xfs/xfs_dinode.h
+3
fs/xfs/xfs_dinode.h
···
39
39
* There is a very similar struct icdinode in xfs_inode which matches the
40
40
* layout of the first 96 bytes of this structure, but is kept in native
41
41
* format instead of big endian.
42
+
*
43
+
* Note: di_flushiter is only used by v1/2 inodes - it's effectively a zeroed
44
+
* padding field for v3 inodes.
42
45
*/
43
46
typedef struct xfs_dinode {
44
47
__be16 di_magic; /* inode magic # = XFS_DINODE_MAGIC */
+22
-9
fs/xfs/xfs_inode.c
+22
-9
fs/xfs/xfs_inode.c
···
896
896
to->di_projid_lo = cpu_to_be16(from->di_projid_lo);
897
897
to->di_projid_hi = cpu_to_be16(from->di_projid_hi);
898
898
memcpy(to->di_pad, from->di_pad, sizeof(to->di_pad));
899
-
to->di_flushiter = cpu_to_be16(from->di_flushiter);
900
899
to->di_atime.t_sec = cpu_to_be32(from->di_atime.t_sec);
901
900
to->di_atime.t_nsec = cpu_to_be32(from->di_atime.t_nsec);
902
901
to->di_mtime.t_sec = cpu_to_be32(from->di_mtime.t_sec);
···
923
924
to->di_lsn = cpu_to_be64(from->di_lsn);
924
925
memcpy(to->di_pad2, from->di_pad2, sizeof(to->di_pad2));
925
926
uuid_copy(&to->di_uuid, &from->di_uuid);
927
+
to->di_flushiter = 0;
928
+
} else {
929
+
to->di_flushiter = cpu_to_be16(from->di_flushiter);
926
930
}
927
931
}
928
932
···
1031
1029
/*
1032
1030
* Read the disk inode attributes into the in-core inode structure.
1033
1031
*
1034
-
* If we are initialising a new inode and we are not utilising the
1035
-
* XFS_MOUNT_IKEEP inode cluster mode, we can simple build the new inode core
1036
-
* with a random generation number. If we are keeping inodes around, we need to
1037
-
* read the inode cluster to get the existing generation number off disk.
1032
+
* For version 5 superblocks, if we are initialising a new inode and we are not
1033
+
* utilising the XFS_MOUNT_IKEEP inode cluster mode, we can simple build the new
1034
+
* inode core with a random generation number. If we are keeping inodes around,
1035
+
* we need to read the inode cluster to get the existing generation number off
1036
+
* disk. Further, if we are using version 4 superblocks (i.e. v1/v2 inode
1037
+
* format) then log recovery is dependent on the di_flushiter field being
1038
+
* initialised from the current on-disk value and hence we must also read the
1039
+
* inode off disk.
1038
1040
*/
1039
1041
int
1040
1042
xfs_iread(
···
1060
1054
1061
1055
/* shortcut IO on inode allocation if possible */
1062
1056
if ((iget_flags & XFS_IGET_CREATE) &&
1057
+
xfs_sb_version_hascrc(&mp->m_sb) &&
1063
1058
!(mp->m_flags & XFS_MOUNT_IKEEP)) {
1064
1059
/* initialise the on-disk inode core */
1065
1060
memset(&ip->i_d, 0, sizeof(ip->i_d));
···
2889
2882
__func__, ip->i_ino, ip->i_d.di_forkoff, ip);
2890
2883
goto corrupt_out;
2891
2884
}
2885
+
2892
2886
/*
2893
-
* bump the flush iteration count, used to detect flushes which
2894
-
* postdate a log record during recovery. This is redundant as we now
2895
-
* log every change and hence this can't happen. Still, it doesn't hurt.
2887
+
* Inode item log recovery for v1/v2 inodes are dependent on the
2888
+
* di_flushiter count for correct sequencing. We bump the flush
2889
+
* iteration count so we can detect flushes which postdate a log record
2890
+
* during recovery. This is redundant as we now log every change and
2891
+
* hence this can't happen but we need to still do it to ensure
2892
+
* backwards compatibility with old kernels that predate logging all
2893
+
* inode changes.
2896
2894
*/
2897
-
ip->i_d.di_flushiter++;
2895
+
if (ip->i_d.di_version < 3)
2896
+
ip->i_d.di_flushiter++;
2898
2897
2899
2898
/*
2900
2899
* Copy the dirty parts of the inode into the on-disk
+11
-2
fs/xfs/xfs_log_recover.c
+11
-2
fs/xfs/xfs_log_recover.c
···
2592
2592
goto error;
2593
2593
}
2594
2594
2595
-
/* Skip replay when the on disk inode is newer than the log one */
2596
-
if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2595
+
/*
2596
+
* di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
2597
+
* are transactional and if ordering is necessary we can determine that
2598
+
* more accurately by the LSN field in the V3 inode core. Don't trust
2599
+
* the inode versions we might be changing them here - use the
2600
+
* superblock flag to determine whether we need to look at di_flushiter
2601
+
* to skip replay when the on disk inode is newer than the log one
2602
+
*/
2603
+
if (!xfs_sb_version_hascrc(&mp->m_sb) &&
2604
+
dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2597
2605
/*
2598
2606
* Deal with the wrap case, DI_MAX_FLUSH is less
2599
2607
* than smaller numbers
···
2616
2608
goto error;
2617
2609
}
2618
2610
}
2611
+
2619
2612
/* Take the opportunity to reset the flush iteration count */
2620
2613
dicp->di_flushiter = 0;
2621
2614