Re: BTRFS constantly reports "No space left on device" even with a huge unallocated space

[Chris Murphy <lists@colorremedies.com>]

On Thu, Sep 1, 2016 at 11:12 AM, Jeff Mahoney <jeffm@suse.com> wrote:
> On 9/1/16 1:04 PM, Austin S. Hemmelgarn wrote:
>> On 2016-09-01 12:34, Ronan Arraes Jardim Chagas wrote:
>>> Em Qui, 2016-09-01 às 09:21 -0400, Austin S. Hemmelgarn escreveu:
>>>> Yes, you can just run `btrfs quota disable /` and it should
>>>> work.  This
>>>> ironically reiterates that one of the bigger problems with BTRFS is
>>>> that
>>>> distros are enabling unstable and known broken features by default
>>>> on
>>>> install.  I was pretty much dumbfounded when I first learned that
>>>> OpenSUSE is enabling BTRFS qgroups by default since they are known
>>>> to
>>>> not work reliably and cause all kinds of issues.
>>>
>>> Thanks Austin! I executed the command and now I get:
>>>
>>> btrfs qgroup show /
>>> ERROR: can't perform the search - No such file or directory
>>> ERROR: can't list qgroups: No such file or directory
>>>
>>> as expected. Now I will wait for +- 1 week to see if the problem will
>>> occur and, if not, I will send an e-mail to openSUSE factory mailing
>>> list to start a discussion if it is better to not enable qgroups by
>>> default.
>> I have a feeling that you'll probably have no issues.
>>
>> As far as having qgroups enabled by default, I think the reasoning is to
>> emulate having separate filesystems with their own space limits.  I can
>
> It's not.  We use qgroups because that's the only way we can track how
> much space each subvolume is using, regardless of whether anyone wants
> to do enforcement.  When it's working properly, snapper can make use of
> that information to make informed decisions on how much space will
> actually be released when removing old snapshots.
>
>> entirely understand this use case, and TBH it's about the only use case
>> I'd consider quota groups for (per-user subvolumes for home directories
>> are great, but there are numerous perfectly legitimate reasons to have
>> very large amounts of data in your home directory for very short periods
>> of time, so I wouldn't personally use qgroups there).  The problem
>> arises from the fact that it doesn't _look_ like separate filesystems
>> (single entry in df, all the mounts point at the same device, etc), and
>
> On SUSE-based kernels, the inodes on different subvolumes report the
> anonymous device associated with the subvolume.
>
> That said, I have a WIP that creates (and auto-tears down) vfsmounts for
> each subvolume.  It's not all the way to a working df that would use the
> qgroup information to report space usage, but it's a start.


Jeff, I'm a little bit irritated because I initially suspected in this
thread that this was an opensuse issue. That I questioned the kernel
as the source is really beside the point. You didn't even recognize
this might be quota related based on what was going on, because you
bounced him back to this list when I suggested he take the issue to
the opensuse-factory list.

What Ronan was reporting was behavior that no one on this list has
ever previously reported. And upstream does not have quotas enabled by
default so there is no reason why any regular testers here would have
come across this.

So now we come full circle and I have to call this a misfeature that's
trying to make up for another one, which is neurotic levels of
snapshots taken by snapper out of the box. There is no good goddamn
reason for it to take 100 read only snapshots in two fucking days.
It's any wonder why the results are pathological.

-- 
Chris Murphy