[Top] [All Lists]

Re: Reduce XFS footprint (was Re: TAKE - remove a function xfs added to

To: "Ralf G. R. Bergs" <rabe@xxxxxxxxxxxxxx>
Subject: Re: Reduce XFS footprint (was Re: TAKE - remove a function xfs added to filemap.c
From: Stephen Lord <lord@xxxxxxx>
Date: Sat, 26 Jan 2002 13:59:15 -0600
Cc: Linux XFS Mailing List <linux-xfs@xxxxxxxxxxx>
References: <E16UW6R-0007K6-00@xxxxxxxxxxxxxxxxxxxxxxxxxxxx>
Sender: owner-linux-xfs@xxxxxxxxxxx
User-agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:0.9.7) Gecko/20011226
Ralf G. R. Bergs wrote:

then it gets a bit smaller. There are definitely other code paths you cannot get to in there
which we could cut out - just a matter of finding them.

That reminds me of another question I'm asking myself since a while: It seems as if you're developing XFS for Linux from scratch instead of "just" porting it from SGI. Is that right? Or why do you seem to change even fundamental code pieces? Or does porting to Linux mean you have to rewrite it because it's so different from IRIX?

I would say 95% of the code is untouched, except for the endian conversion work which was neccessary for a little endian implementation. The interfaces between the vfs and the filesystem and between the filesystem and the block layer are where all the changes have happened. We did need to do a lot of work for data caching since the Irix model and the linux model were not at all close. The vfs layer has some major
differences too.

You could try editing xfs_types.h and  changing

#define XFS_BIG_FILES           1


#define XFS_BIG_FILES           0

I probably can't do that since as you know I use hardware RAIDs so I probably need "big" filesystems, and I also need "big" files.

The limits in xfs if you turn these off are:

Max file size 2^40 bytes
Max filesystem size is 2 Tbytes I think (not sure without some research).

It does compile and run, and you save about 45K in code size by doing it,
so not too much. You can also edit xfs_macros.h and define WANT_SPACE_C
to 1 (replace the existing definition). This turns a lot of inline macros into function calls - it will be slower, but it is another 10K less code - again, not
very much.

So  and smp build (in 2.5) went from

  text    data     bss     dec     hex filename
520994    4896    2080  527970   80e62 fs/xfs/xfs.o

  text    data     bss     dec     hex filename
487334    4896    2080  494310   78ae6 fs/xfs/xfs.o

with these two build changes.


<Prev in Thread] Current Thread [Next in Thread>