[RFC] xfs: wait for the write of the superblock on unmount

[Mark Tinguely]

On 07/20/12 08:21, Mark Tinguely wrote:
> On 07/18/12 20:58, Dave Chinner wrote:
>> On Wed, Jul 18, 2012 at 12:33:58PM -0500, tinguely at sgi.com wrote:
>>> Sorry, I have been distracted away from this regression. This was
>>> previously
>>> titled "xfs: synchronously write the superblock on unmount".
>>>
>>> xfs_wait_buftarg() does not wait for the completion of the write of the
>>> uncached superblock. This write can race with the shutdown of the log
>>> and
>>> causes a panic if the write does not win the race.
>>>
>>> The log write of the superblock is important for possible recovery,
>>> but a
>>> second syncronous write of the same superblock seems redundant. Would
>>> just
>>> waiting for the iodone() of the log write before tearing down the log be
>>> enough?
>>
>> Yes. i.e. something like:
>>
>> /*
>> * The superblock buffer is uncached, so xfs_wait_buftarg()
>> * will not wait for it. Hence we need to explicitly wait
>> * for IO completion on the superblock to occur here.
>> */
>> error = xfs_buf_iowait(mp->m_sb_bp);
>> if (error)
>> AAAAIEEEE!
>>
>> This fix is also needed in xfs_quiesce_attr() for the freeze and
>> ro,remount cases as well.
>>
>> Cheers,
>>
>> Dave.
>>
>
> Thanks for the feedback.
>
> I chose the ugly lock to block on the superblock buffer because it has
> XBF_ASYNC set at that point and won't go through the complete() patch. I
> will look at turning the XBF_ASYNC flag off and wait on a xfs_buf_iowait().
>
> --Mark.

Looks like the xfsaild_push() sets the XBF_ASYNC flag on the superblock 
buffer. Waiting on the xfs_buf_iowait() won't work but have to use the 
less elegant lock.

--Mark.