[bisected] xfs_repair refuses to run on cleanly mountable partition

[Eric Sandeen]

On 10/7/13 11:52 AM, Markus Trippelsdorf wrote:
> On 2013.10.07 at 10:54 -0500, Eric Sandeen wrote:
>> On 10/7/13 10:40 AM, Markus Trippelsdorf wrote:
>>> On 2013.10.07 at 10:36 -0500, Eric Sandeen wrote:
>>>> On 10/7/13 10:29 AM, Markus Trippelsdorf wrote:
>>>>> On 2013.10.07 at 10:21 -0500, Eric Sandeen wrote:
>>>>>> On 10/7/13 10:16 AM, Markus Trippelsdorf wrote:
>>>>>>> x4 ~ # xfs_repair -V
>>>>>>> xfs_repair version 3.2.0-alpha1
>>>>>>>
>>>>>>> x4 ~ # mount -o logbsize=256k /dev/sdc1 /mnt
>>>>>>> ...
>>>>>>> [ 6419.592649] XFS (sdc1): Mounting Filesystem
>>>>>>> [ 6419.642480] XFS (sdc1): Ending clean mount
>>>>>>>
>>>>>>> x4 ~ # xfs_info /dev/sdc1
>>>>>>> meta-data=/dev/sdc1              isize=256    agcount=4, agsize=61047552 blks
>>>>>>>          =                       sectsz=4096  attr=2, projid32bit=0
>>>>>>>          =                       crc=0
>>>>>>> data     =                       bsize=4096   blocks=244190208, imaxpct=25
>>>>>>>          =                       sunit=0      swidth=0 blks
>>>>>>> naming   =version 2              bsize=4096   ascii-ci=0
>>>>>>> log      =internal               bsize=4096   blocks=119233, version=2
>>>>>>>          =                       sectsz=4096  sunit=1 blks, lazy-count=1
>>>>>>> realtime =none                   extsz=4096   blocks=0, rtextents=0
>>>>>>>
>>>>>>> x4 ~ # umount /mnt
>>>>>>>
>>>>>>> x4 ~ # xfs_repair /dev/sdc1
>>>>>>> Phase 1 - find and verify superblock...
>>>>>>> Phase 2 - using internal log
>>>>>>>         - zero log...
>>>>>>> ERROR: The filesystem has valuable metadata changes in a log which needs to
>>>>>>> be replayed.  Mount the filesystem to replay the log, and unmount it before
>>>>>>> re-running xfs_repair.  If you are unable to mount the filesystem, then use
>>>>>>> the -L option to destroy the log and attempt a repair.
>>>>>>> Note that destroying the log may cause corruption -- please attempt a mount
>>>>>>> of the filesystem before doing this.
>>>>>>
>>>>>> What kernel are you running?  Does older xfs_repair behave differently?
>>>>>> (use xfs_repair -n if you test an old xfsprogs, to preserve this state
>>>>>> for debugging...)
>>>>>
>>>>> I'm running the latest git kernel 3.12.0-rc4. 
>>>>> "xfs_repair -n" runs fine even with xfsprogs 3.2.0-alpha1...
>>>>>
>>>>>> Perhaps copying out or dumping the log w/ xfs_logprint would also help, 
>>>>>> maybe start with:
>>>>>>
>>>>>> # xfs_logprint -t /dev/sdc1
>>>>> xfs_logprint:
>>>>>     data device: 0x821
>>>>>     log device: 0x821 daddr: 976760888 length: 953864
>>>>>
>>>>>     log tail: 53376 head: 53376 state: <CLEAN>
>>>>
>>>> Funky.
>>>>
>>>> How about an xfs_repair -v (for verbose).
>>> ...
>>>         - zero log..
>>> zero_log: head block 53048 tail block 49064
>>> ERROR: The filesystem has valuable metadata changes in a log which needs to
>>> ...
>>>
>>
>> Very strange.  Both xfs_logprint & xfs_repair should be using the same
>> function in libxfs for finding the head & tail.
>>
>> I asked off-list if you wanted to provide a metadump image I could look
>> at directly...
> 
> I've bisected this issue to the following commit from Dave:
> 
>  commit e0607266f23f82226f8aee502552d6ce25c4e6a5
>  Author: Dave Chinner <dchinner at redhat.com>
>  Date:   Fri Jun 7 10:25:47 2013 +1000
> 
>     xfsprogs: add crc format support to repair
> 
> 

Cool, thanks.

That commit added:

diff --git a/repair/phase2.c b/repair/phase2.c
index 2817fed..a62854e 100644
--- a/repair/phase2.c
+++ b/repair/phase2.c
@@ -64,6 +64,7 @@ zero_log(xfs_mount_t *mp)
                ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
        }
        log.l_sectbb_mask = (1 << log.l_sectbb_log) - 1;
+       log.l_sectBBsize = 1 << mp->m_sb.sb_logsectlog;
 
        if ((error = xlog_find_tail(&log, &head_blk, &tail_blk))) {
                do_warn(_("zero_log: cannot find log head/tail "

right before the call to xlog_find_tail, which is what found the dirty log.

those various things are:

        __uint8_t       sb_logsectlog;  /* log2 of the log sector size */
        uint            l_sectbb_log;   /* log2 of sector size in bbs */
        int             l_sectBBsize;   /* size of log sector in 512 byte chunks */

The hunk above sticks out as odd, because it was already set a different way about
12 lines prior:

        log.l_sectBBsize  = BTOBB(x.lbsize);

And "indeed" as Dave might say, ;) - l_sectBBsize is supposed to be in
512-byte units (i.e. 1 for 512, 8 for 4k), but it's coming out as 4096
because it's taking sb_logsectlog - describing byte units - and using it to get
something in sector units.

It still accidentally works for 512-byte sectors, because in in that case we set
sb_logsectlog to 0 (not 9, because - sure, why not!):

        if (lsectorsize != BBSIZE || sectorsize != BBSIZE) {
                sbp->sb_logsectlog = (__uint8_t)lsectorlog;
                sbp->sb_logsectsize = (__uint16_t)lsectorsize;
        } else {
                sbp->sb_logsectlog = 0;
                sbp->sb_logsectsize = 0;
        }



Anyway:

I bet if you remove "log.l_sectBBsize = 1 << mp->m_sb.sb_logsectlog;" from
around line 67 it'll fix it.

Want to try it?  Sorry for abusing your bandwidth in the meantime.  :)

If it works I'll send the patch.

Thanks,
-Eric