Looking to confirm issue and seek advice on fix for inode btree fragmentation

[Dave Chinner]

On Thu, Aug 13, 2015 at 05:35:31PM -0600, Pippin Wallace wrote:
> I am writing to confirm the issue I think I have and get advice on the best
> short term and long term “fixes”.
> 
> 
> 
> In reading the following threads and others I am pretty sure I have inode
> btree fragmentation and I have included detailed output below to help
> confirm this.

Make a new filesystem with a current xfsprogs 3.2.4, and it will use
the new free inode btree for allocation that was designed to avoid
this problem.

> Long term fix is upgrade kernel to >= 3.16, xfsprogs >=3.2.1 and rebuild fs
> with new finobt structure.
> 
>         # mkfs.xfs -m crc=1,finobt=1 <dev>

i.e. there are the default mkfs options in xfsprogs >= 3.2.3.

It's really the only viable fix, especially given that you can't use
inode64. And, realistically, it's still going to cause problems
unless you ensure that you have smaller-than-default AGs (e.g. 250GB
AGs) so that there are multiple inode btrees in the 32-bit inode
number region of the filesystem.(*)

(*) which is actually the first 2TB of the filesystem with CRC
enabled filesystems because the inode size is 512 bytes...

> *KERNEL OPTIONS: Which of these kernels should I choose for the best XFS
> support for this issue?*

One your distro vendor supports; the newer the better.

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com