Re: memory overflow or undeflow in free space tree / space_info?

["Holger Hoffstätte" <holger@applied-asynchrony.com>]

On Fri, 29 Jul 2016 17:03:43 -0400, Josef Bacik wrote:

> On 07/29/2016 03:14 PM, Omar Sandoval wrote:
>> On Fri, Jul 29, 2016 at 12:11:53PM -0700, Omar Sandoval wrote:
>>> On Fri, Jul 29, 2016 at 08:40:26PM +0200, Stefan Priebe - Profihost AG wrote:
>>>> Dear list,
>>>>
>>>> i'm seeing btrfs no space messages frequently on big filesystems (> 30TB).
>>>>
>>>> In all cases i'm getting a trace like this one a space_info warning.
>>>> (since commit [1]). Could someone please be so kind and help me
>>>> debugging / fixing this bug? I'm using space_cache=v2 on all those systems.
>>>
>>> Hm, so I think this indicates a bug in space accounting somewhere else
>>> rather than the free space tree itself. I haven't debugged one of these
>>> issues before, I'll see if I can reproduce it. Cc'ing Josef, too.
>>
>> I should've asked, what sort of filesystem activity triggers this?
>>
> 
> Chris just fixed this I think, try his next branch from his git tree
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs.git
> 
> and see if it still happens.  Thanks,
> 
> Josef

Hi Josef,

can you say which patch you have in mind? The tree in question
doesn't have any of Chandra's pagesize/sectorsize patches (carefully
patched around, for stability and LTS patchability) so I hope it's
not the recent commit

8b8b08cb "fix delalloc accounting after copy_from_user faults"

because that would be too fiddly (at least for me) to backport
correctly.

The only other patch I just found missing and which looks like it
could/should (I think?) work on top of the 4.4.x pagesize-based
calculations in file.c is:

a2af23b7 "__btrfs_buffered_write: Pass valid file offset when
releasing delalloc space"

Would that make sense? Neither I nor any other users of that tree
have observed weird space-info underflows so far (and I use my
fs daily), so it's definitely something peculiar Stefan is doing
with his weird compressed rsync-inplace workload. Odd sector offsets
causing slowly creeping space_info underflow sounds to me like it
just might be the problem.

thanks,
Holger