Re: stat64 for over 2TB file returned invalid st_blocks

[Andreas Dilger <adilger@clusterfs.com>]

On Dec 03, 2005  22:00 +0900, Takashi Sato wrote:
> Andreas Dilger wrote:
> >Actually, it should probably be "sector_t", because it isn't really
> >possible to have a file with more blocks than the size of the block
> >device.  This avoids memory overhead for small systems that have no
> >need for it in a very highly-used struct.  It may be for some network
> >filesystems that support gigantic non-sparse files they would need to
> >enable CONFIG_LBD in order to get support for this.
> 
> I think sector_t is ok for local filesystem as you said.  However,
> on NFS, there may be over 2TB file on server side, and inode.i_blocks
> for over 2TB file will become invalid on client side in case CONFIG_LBD
> is disabled.

I don't know the exact specs of NFS v2 and v3, but I doubt they can have
single files larger than 2TB.  Even if they could then this is not a
very common situation and if someone is running in such an environment
then they can easily enable CONFIG_LBD (or make e.g. CONFIG_NFS_V4 have
a dependency to enable this if it is important enough).  What I'd rather
avoid is needless growth of heavily-used structures for rather uncommon
cases (at the current time at least, this can be re-examined later).

It might also be possible to have a separate CONFIG_LSF (or whatever)
that enables support for large single files, maybe enabled by default
with CONFIG_LBD and also configurable separately for clients of network
filesystems with large single files.  Someone who cares more about the
proliferation of configuration options than I can decide whether it
makes sense to keep these as separate options.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.