[PATCH] BUG on fsync/fdatasync with Ext3 data=journal

[Christopher Chan]

Andrew Morton wrote:
> Seiji Kihara <kihara.seiji at lab.ntt.co.jp> wrote:
> 
>>We found that fsync and fdatasync syscalls sometimes don't sync
>>data in an ext3 file system under the following conditions.
>>
>>1. Kernel version is 2.6.6 or later (including 2.6.8.1 and 2.6.9-rc2).
>>2. Ext3's journalling mode is "data=journal".
>>
>>The problem was occurred since 2.6.5-bk1, which includes the patch
>>"[PATCH] ext3 fsync() and fdatasync() speedup".  We found that the
>>problem was solved by deleting the part of the patch which
>>modifies ext3_sync_file().  Maybe, i_state is not correctly set to
>>I_DIRTY when the related page cache is dirty (is it true?)
> 

I have a few qmail (about the heaviest fsync using mta software around) 
boxes that have their queues on ext3.

On a 2.6.7 kernel, these guys are guaranteed to crash within hours if I 
used data=journal for the fs on which the qmail queues are. I say this 
because I ran two of them with data=journal mode and they crashed once 
or more a day. Another one which stayed with ordered had no problems 
during the same period.

Going back to ordered meant that they ran stable for days (weeks now).

The only thing I could get from the logs is:

---------------------------

Aug 17 05:58:22 mta1-7 kernel: Assertion failure in 
__journal_drop_transaction() at fs/jbd/checkpoint.c:613: "transaction->t
_forget == NULL"
Aug 17 05:58:22 mta1-7 kernel: ------------[ cut here ]------------
Aug 17 05:58:22 mta1-7 kernel: kernel BUG at fs/jbd/checkpoint.c:613!
Aug 17 05:58:22 mta1-7 kernel: invalid operand: 0000 [#1]
Aug 17 05:58:22 mta1-7 kernel: SMP
Aug 17 05:58:22 mta1-7 kernel: Modules linked in: nfs lockd sunrpc e1000 
e100 mii usbcore
Aug 17 05:58:22 mta1-7 kernel: CPU:    0
Aug 17 05:58:22 mta1-7 kernel: EIP:    0060:[<c0193db0>]    Not tainted
Aug 17 05:58:22 mta1-7 kernel: EFLAGS: 00010202   (2.6.7)