patch to fix symlink-related GPFs on Linux

[Magnus Ahltorp]

> kernel: xfs_devwrite
> kernel: xfs_message_receive opcode = 5
> kernel: xfs_message_installnode
> kernel: xfs_node_find
> kernel: new_xfs_node 0.537113399.480.30967
> kernel: xfs_node_find
> kernel: xfs_iget sb: 001f3fd4 node: 02e40134 newnode: 00f5f798

> The kernel appears to be spinning in this loop in __iget:
> 
> 	for (inode = h->inode; inode ; inode = inode->i_hash_next)
> 		if (inode->i_dev == sb->s_dev && inode->i_ino == nr)
> 			goto found_it;

This is probably caused by my inode->i_dev magic. Every xfs inode has
it's i_dev field set to the device of the cache file system. This
confuses the 2.0 VFS, but I don't know what to do about it. One option
is to skip the i_dev magic, and skip bmap for 2.0, and implement
readpage instead, but I like the simplicity of the bmap solution, so
I'd like to see another solution.

/Magnus
map at stacken.kth.se