Re: [patch] fix up lock order reversal in writeback

[Eric Sandeen <sandeen@redhat.com>]

On 11/18/10 12:04 PM, Eric Sandeen wrote:
> On 11/18/10 11:10 AM, Andrew Morton wrote:
>> On Thu, 18 Nov 2010 08:55:18 -0600 Eric Sandeen <sandeen@redhat.com> wrote:
>>
>>>> Can we just delete writeback_inodes_sb_nr_if_idle() and
>>>> writeback_inodes_sb_if_idle()?  The changelog for 17bd55d037a02 is
>>>> pretty handwavy - do we know that deleting these things would make a
>>>> jot of difference?
>>>
>>> Really?  I thought it was pretty decent ;)
>>>
>>> Anyway, xfstests 204, "Test out ENOSPC flushing on small filesystems."
>>> shows the problem clearly, IIRC.  I should have included that in the
>>> changelog, I suppose, sorry.
>>
>> Your email didn't really impart any information :(
>>
>> I suppose I could accidentally delete those nasty little functions in a
>> drivers/parport patch then wait and see if anyone notices.
>>
> 
> Um, ok, then, to answer the question directly :
> 
> No, please don't delete those functions, it will break ENOSPC handling
> in ext4 as shown by xfstests regression test #204 ...

Further - 

What is going on here is that with delayed allocation, ext4 takes reservations
against free blocks based on the data blocks it must write out, and the
worst-case metadata that the writeout may take.  Getting writeback failing
with ENOSPC would be bad.

But then we wind up with a bunch of unflushed writes sitting on huge
metadata reservations, and start hitting ENOSPC due to that worst-case
reservation.  After a sync we have tons of free space again, because
the worst-case space reservations turned into usually best-case
reality.

That's what the function is used for; once we start filling up the
fs, we proactively flush data to free up the worst-case metadata
reservations.

Dropping it will put us back in the bad situation.

If there are other ideas to fix it, I'm all ears, but this worked.

-Eric