Re: ext3fs: umount+sync not enough to guarantee metadata-on-disk

[Mark Lord <lkml@rtr.ca>]

Jan Kara wrote:
>> Andrew Morton wrote:
>>> On Thu, 07 Jun 2007 12:11:58 -0400
>>> Chuck Ebbert <cebbert@redhat.com> wrote:
>>>
>>>> On 06/07/2007 11:41 AM, Andrew Morton wrote:
>>>>>>   mount /var/lib/mythtv -oremount,ro
>>>>>>   sync
>>>>>>   umount /var/lib/mythtv
>>>>> Did this succeed?  If the application is still truncating that file, the
>>>>> umount should have failed.
>>>> Shouldn't sync should wait for truncate to finish?
>>> I can't think of anything in there at present which would cause that to
>>> happen, and it's not immediately obvious how we _could_ make it happen - we
>>> have an inode which potentially has no dirty pages and which is itself
>>> clean.  The truncate can span multiple journal commits, so forcing a
>>> journal commit in sync() won't necessarily block behind the truncate.
>>>
>>> I guess we could ask sync to speculatively take and release every inode's
>>> i_mutex or something.  But even that would involve quite some hoop-jumping
>>> due to those infuriating spinlock-protected list_heads on the superblock.
>>>
>>> hmm.
>> Okay, I added more instrumentation and retested today.
>>
>> Good and Bad.
>> The umount does indeed fail while the massive unlink is happening,
>> so I can just loop on that a few times before giving up.
>>
>> But.. the earlier "remount,ro".. well.. I don't know what it does.
>> I did get it to lock up solid, though.. hung on the "remount,ro"
>> when issued during an unlink of a 15GB file.  The disk I/O eventually
>> completes, and drives go idle, but the system remains hung inside
>> the remount,ro call.
>>
>> Alt-sysrq-T was functioning, so I have some screen shots (.jpg) here:
>>
>>  http://rtr.ca/remount_ro/
>   Thanks for the traces.
> 
>> That's definitely a bug.
>   Yes. We have a nice lock inversion there. ext3_remount() is called
> with sb->s_lock held and waits for transaction to finish in
> journal_lock_updates(). On the other hand ext3_orphan_del() is called
> inside a transaction and tries to do lock_super()... Bad luck.
>

Peachy.  Do you have enough knowledge here to generate a fix for this?
Maybe just have the remount break out, releasing all locks, and then 
loop and retry (or return -EBUSY?) when this happens?

Cheers