On 10/27/12 1:47 PM, Nix wrote:

On 27 Oct 2012, Theodore Ts'o said:

quoted

On Sat, Oct 27, 2012 at 01:45:25PM +0100, Nix wrote:

quoted

Ah! it's turned on by journal_async_commit. OK, that alone argues
against use of journal_async_commit, tested or not, and I'd not have
turned it on if I'd noticed that.

(So, the combinations I'll be trying for effect on this bug are:

 journal_async_commit (as now)
 journal_checksum
 none

Can you also check and see whether the presence or absence of
"nobarrier" makes a difference?

Done. (Also checked the effect of your patches posted earlier this week:
no effect, I'm afraid, certainly not under the fails-even-on-3.6.1 test
I was carrying out, umount -l'ing /var as the very last thing I did
before /sbin/reboot -f.)

nobarrier makes a difference that I, at least, did not expect:

[no options]                    No corruption

nobarrier                       No corruption

          journal_checksum      Corruption
                                Corrupted transaction, journal aborted
                                
nobarrier,journal_checksum      Corruption
                                Corrupted transaction, journal aborted

          journal_async_commit  Corruption
                                Corrupted transaction, journal aborted

nobarrier,journal_async_commit  Corruption
                                No corrupted transaction or aborted journal

That's what we needed.  Woulda been great a few days ago ;)

In my testing journal_checksum is broken, and my bisection seems to
implicate

commit 119c0d4460b001e44b41dcf73dc6ee794b98bd31
Author: Theodore Ts'o [off-list ref]
Date:   Mon Feb 6 20:12:03 2012 -0500

    ext4: fold ext4_claim_inode into ext4_new_inode
    
as the culprit.  I haven't had time to look into why, yet.

-Eric

I didn't expect the last case at all, and it adequately explains why you
are mostly seeing corrupted journal messages in your tests but I was
not. It also explains why when I saw this for the first time I was able
to mount the resulting corrupted filesystem read-write and corrupt it
further before I noticed that anything was wrong.

It is also clear that journal_checksum and all that relies on it is
worse than useless right now, as Eric reported while I was testing this.
It should probably be marked CONFIG_BROKEN in future 3.[346].* stable
kernels, if CONFIG_BROKEN existed anymore, which it doesn't.

It's a shame journal_async_commit depends on a broken feature: it might
be notionally unsafe but on some of my systems (without nobarrier or
flashy caching controllers) it was associated with a noticeable speedup
of metadata-heavy workloads -- though that was way back in 2009...
however, "safety first" definitely applies in this case.