Re: [PATCH 2/4] XFS: Return case-insensitive match for dentry cache

["Barry Naujok" <bnaujok@xxxxxxx>]

On Mon, 21 Apr 2008 19:46:41 +1000, Christoph Hellwig <hch@xxxxxxxxxxxxx>  
wrote:

> > > +kmem_zone_t       *xfs_name_zone;
> >
> > What about just using kmalloc here?  We know the length of the name
> > anyway, so there is no point of allocating the maximum possible size.
> >
> > >    error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp, 0);
> > > -  if (error)
> > > +  if (error) {
> > > +          if (ci_match && *ci_match)
> > > +                  xfs_name_free(name->name);
> > >            goto out;
> > > +  }
> >
> > All the allocation and freeing for ci_match looks odd and error prone
> > to me.  I think the low-level directory code should never allocate
> > args->value unless it's explicitly asked for a CI match.  That way
> > there's only one place in xfs_ci_lookup to free it either.
>
> Also the low-level name duplication code could be factored out a little
> more ala:
>
> /*
>  * If a case-insensitive match, allocate a buffer and copy the actual
>  * name into the buffer. Return it via args->value.
>  */
> void xfs_copy_ci_name(struct xfs_da_args *args, const char *name, int  
> namelen)
> {
> 	if (args->return_ci_name && args->cmpresult == XFS_CMP_CASE) {
> 		args->valuelen = namelen;
> 		args->value = kmemdup(name, namelen, GFP_NOFS);
> 		/* error handled in higher layers */
> 	}
> }
>
> Instead of adding args->return_ci_name it might make sense to just
> replace the last four chars there with an unsigned short lookup_flags
> and just set bits in it.

There's already a "flags" field in xfs_da_args and only the low 16 bits
are currently used for attr operations.

We could use 5 of the high bits for these flags, or just add another
flags field (which are used for things other than lookup).