Re: balance disables nodatacow

[Kyle Gates <kylegates@gmail.com>]

On Mon, Jul 30, 2012 at 11:58 PM, Liu Bo <liubo2009@cn.fujitsu.com> wrote:
> On 07/31/2012 12:35 PM, Kyle Gates wrote:
>
>> On Mon, Jul 30, 2012 at 9:00 PM, Liu Bo <liubo2009@cn.fujitsu.com> wrote:
>>> On 07/31/2012 03:55 AM, Kyle Gates wrote:
>>>
>>>> I have a 3 disk raid1 filesystem mounted with nodatacow. I have a
>>>> folder in said filesystem with the 'C' NOCOW & 'Z' Not_Compressed
>>>> flags set for good measure. I then copy in a large file and proceed to
>>>> make random modifications. Filefrag shows no additional extents
>>>> created, good so far. A big thank you to the those devs who got that
>>>> working.
>>>> However, after a balance run (which crashed btw), modifications to the
>>>> file result in increasing extent counts, ie. no longer obeying
>>>> nodatacow mount option nor the 'C' flag.
>>>> I realize that a balance will cow the data but it should not disable
>>>> the ability to make nodatacow inline changes to a file.
>>>>
>>>> I'm currently running 3.5.0 mainline Ubuntu kernel with the latest
>>>> master of btrfs-progs from a ppa.
>>>>
>>>
>>> hmm, I want to know if your 3.5.0 mainline ubuntu kernel has the commit:
>>>
>>> Btrfs: make btrfs's allocation smoothly with preallocation
>>>
>>> This commit may help you out of trouble.
>>> Please let us know the result.
>>>
>>> thanks,
>>> liubo
>>>
>>>> Thank you.
>>>> Kyle Gates
>> I didn't have that commit before but I do now:
>> Linux home 3.5.0-999-generic #201207290405 SMP Sun Jul 29 08:13:26 UTC
>> 2012 i686 athlon i386 GNU/Linux
>>
>> One of my disks is having some issues though. I got this while running a
>> balance which is now hung:
>>
>
>
> oops, I know where goes wrong.  I should have noticed it, sorry Kyle.
>
> We've not merged the following patch into mainline:
>
> Btrfs: allow delayed refs to be merged V2
> (link: https://patchwork.kernel.org/patch/1197821/)
>
> So could you please apply the patch and test it again?
>
> thanks,
> liubo

I have yet to try a kernel compile so I'll just wait until that patch
is merged, sorry.

On a related note...
Setting the 'C' NOCOW flag on a file will only be obeyed if the
filesystem is mounted with 'nodatasum'
David sent a patch for the simplest case back on 5/15/2012 but I don't
think it got merged.

A nice enhancement would be when nodatacow is enabled, checksums would
be selectively cleared as data extents are modified.

Thanks,
Kyle