Re: optimize DNAME_INLINE_LEN

[Andreas Dilger <adilger@turbolabs.com>]

On Dec 13, 2001  23:22 +0100, Manfred Spraul wrote:
> The dcache entries are allocated with SLAB_HWCACHE_ALIGN. For better
> memory usage, we should increase DNAME_INLINE_LEN so that sizeof(struct
> dentry) becomes a multiple of the L1 cache line size.
> 
> On i386 the DNAME_INLINE_LEN becomes 36 bytes instead of 16, which saves
> thousands of kmallocs for external file names. (12818 on my debug
> system, after updatedb)

If the dentries are already aligned on cachelines, I don't see any reason
NOT to do this.  Why waste all the memory for alignment when it can be
used for something else?

> The attached patch is preliminary, it doesn't compile with egcs-1.1.2.
> Which gcc version added support for unnamed structures?

Hmm, I thought it was gcc 3.0 that supported unnames structures.  For
sure gcc 2.95.2 does not, because unnamed structs are used in the NTFS
tools, and I couldn't compile them.

If you wanted to keep compatibility for older compilers (may be a good idea)
you could do something ugly like using a named struct like "du" and then:

#define d_inode du.du_inode
#define d_count du.du_count
.
.
.
#define d_fsdata du.du_fsdata



Alternately (also ugly) you could just define struct dentry the as now,
but have a fixed size declaration for d_iname, like:

#define DNAME_INLINE_MIN 16

	unsigned char d_iname[DNAME_INLINE_MIN];
	
and only set DNAME_INLINE_LEN afterwards like:

#define DNAME_INLINE_LEN \
	(DNAME_INLINE_MIN+L1_CACHE_BYTES - sizeof(struct dentry)%L1_CACHE_BYTES)

This _should_ work for all code that uses DNAME_INLINE_LEN, and if the
alignment doesn't change.  It will break horribly if you do sizeof(struct
dentry), declare a dentry on the stack, or if someone changes the alignment.

You can also do preprocessor macro tricks to get something like an unnamed
union in a struct, but it is also a bit ugly.

Cheers, Andreas
--
Andreas Dilger
http://sourceforge.net/projects/ext2resize/
http://www-mddsp.enel.ucalgary.ca/People/adilger/