Re: [PATCH 2/3] Fix problems on multi-TB filesystem and file

[Andreas Dilger <adilger@clusterfs.com>]

On Jan 13, 2006  12:28 -0800, Andrew Morton wrote:
> So now we're proposing to repeat the sector_t problem with a bunch of new
> fields.  Fortunately we're less likely to be putting these particular
> fields into printk statements but I note that CIFS (at least) has a couple
> such statements and with your patch they're now generating warnings (and
> potential runtime bugs).
> 
> On the other hand, for a fairly fat .config which has 17 filesystems in
> .vmlinux:
> 
>    text    data     bss     dec     hex filename
> 4633032 1011304  248288 5892624  59ea10 vmlinux		CONFIG_LSF=y
> 4633680 1011304  248288 5893272  59ec98 vmlinux		CONFIG_LSF=n
> 
> It's probably less 0.5 kbytes for usual embedded .config.
> I just don't think the benefit of CONFIG_LSF outweighs its costs.

We were originally going to use CONFIG_LBD, but there were some complaints
that "sector_t" isn't the right variable to use for this, even though they
are remarkably close.  That would at least remove one config change.

I don't think the cost is in the vmlinux itself, but rather that having a
long long for i_blocks is overkill for any but the very largest systems
(Lustre has been running fine w/o this, at the expense of some accuracy
for the i_blocks count on many-TB files).  Growing struct inode for these
0.0000001% of systems is probably harmful for small systems, given how
many inodes are used in a system.

Two options exist IMHO:
- remove the new CONFIG_* parameters and stick with CONFIG_LBD (this could
  still use a separate type from sector_t if desired) to reduce the amount
  of testing combinations needed
- make the new CONFIG_* default to on and allow it to be disabled with
  CONFIG_BASE_SMALL

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