On Thu, Sep 06, 2001 at 01:51:06PM -0500, Steve Lord wrote:
> > On Wed, Sep 05, 2001 at 10:36:24AM -0500, Steve Lord wrote:
> > > This is not a raid5 thing, it is a filesystem size issue, once you get
> > > above 1 Tbyte in filesystem size then xfs inode numbers (which are really
> > > a disk address) can take more than 32 bits. Since lots of linux code,
> > > including NFS, does not cope with this, we need to change things in xfs
> >
> > Since 2.4.5 or so mainstream 2.4 has the fh_to_dentry/dentry_to_fh super
> > block interfaces. They are currently only used by reiserfs to handle
> > their equivalent of 64bit inodes; but XFS could use them too to at least
> > avoid problems with NFS for the big file systems.
> >
> > -Andi
>
> Hmm, it does not solve the problem of stat calls etc which return a 32
> bit inode to user space. We also have an application issue with XFS on
> Irix which means we need to do the solution where we force inodes into
> 32 bits by placement within the filesystem.
It requires LFS aware applications yes and for those a glibc change, but
luckily no recompile.
[I guess as soon as 2.5 started a new stat syscall should be reserved for that]
If you have a nice way to make the lower 32bit of your inodes unique that
would be of course more compatible. I just wanted to point out that it is
not required for NFS.
-Andi
|