Bug 120622

Workaround: Tell the installer not to format the partition you're
going to install onto, switch to a virtual terminal and do something
similar to the following:

mkfs.xfs -i size=512 /dev/hda5

this will set the inode size to 512 bytes, and should be enough for
SELinux to perform well. Ideally, there'd be a configuration in the
installer to manually set FS parameters (such as log and inode size).

Am currently doing an install like this of FC3t1.

Have reported a bug for Anaconda, related to allowing customisation of
mkfs parameters:
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=129644

Regardless of the Anaconda issue, I believe that the default should be 512 
bytes. 
 
Note that this will also improve performance for recent versions of Samba, 
ACLs, and user XATTRs. 

Russell, can you please have a look whether the default in xfsprogs 2.7.11-1 
changed in the meantime?

It hasn't changed.  I've spoken to some of the XFS developers about this (they 
are based in Melbourne so I meet them periodically).  As of our last 
conversation they reported that it would be difficult for them to get such a 
change included and suggested that I get the distributions to include patches 
instead.

Created attachment 127452 [details]
Patch to provide this functionality.

The attached patch gives a default Inode size of 512 bytes.  This will
significantly improve performance for SE Linux machines (default install) and
for Samba servers.

Looks like Jeff isn't xfsprogs maintainer any longer or he unfortunately doesn't 
use Bugzilla. Florian, could you apply Russell's patch for the next xfsprogs 
rebuild, please?

Committed

Does Committed mean that it is 512 by default?

# mkfs.xfs /dev/appsvg/vmwarelv 
meta-data=/dev/appsvg/vmwarelv   isize=256    agcount=16, agsize=327680 blks
         =                       sectsz=512   attr=0
data     =                       bsize=4096   blocks=5242880, imaxpct=25
         =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096  
log      =internal log           bsize=4096   blocks=2560, version=1
         =                       sectsz=512   sunit=0 blks
realtime =none                   extsz=65536  blocks=0, rtextents=0
[root@echelon opt]# cat /etc/redhat-release 
Fedora Core release 6 (Zod)


According to that it is still 256 on FC6?

# mkfs.xfs -i size=512 /dev/appsvg/vmwarelv -f
meta-data=/dev/appsvg/vmwarelv   isize=512    agcount=16, agsize=327680 blks
         =                       sectsz=512   attr=0
data     =                       bsize=4096   blocks=5242880, imaxpct=25
         =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096  
log      =internal log           bsize=4096   blocks=2560, version=1
         =                       sectsz=512   sunit=0 blks
realtime =none                   extsz=65536  blocks=0, rtextents=0

Just a note, anaconda is now using "attr2" for installs from FC6 and beyond, so
space is more efficiently shared between extents & attributes in the inode. 
With that, selinux & attrs should still be pretty efficient in 256 bytes,
without the extra space overhead of the larger inodes, and the 512-byte default
patch has been removed.  Note that attr2 is still not the default in mkfs.xfs,
though - it can be set at mkfs or mount time.