Re: [PATCH] VFS: Cut down inode->i_op->xyz accesses in path walking

[Christoph Hellwig <hch@lst.de>]

On Fri, Jul 22, 2011 at 08:55:48PM -0700, Linus Torvalds wrote:
> If something sets dentry->d_inode without going through __d_instantiate, 
> that misses it. I'm looking at d_obtain_alias(), and wondering, for 
> example.

As pointed out in my last mail d_obtain_alias will absolute need setting
up these flags as well.

> This is a very minimal version of the previous patch: those *two* places 
> where we now test the new DCACHE_OP_xyz bits are where we really see the 
> dereferencing of the inode->i_op in the normal path walking. HOWEVER, the 
> "inode_permission()" function also shows up pretty clearly on profiles, 
> and the hottest part of that is the testing of "inode->i_op->permission".
> 
> However, with Al's other changes, the old special-case "exec_permission()" 
> no longer exists. The whole reason it existed before was that the execute 
> permission checking on directory entries is/was special from a performance 
> standpoint. So it might make sense to resurrect it, and then add the
> DCACHE_OP_PERMISSION case like in my older patch.

I don't think we need to re-add exec_permission for it.  By checking
DCACHE_OP_PERMISSION instead of inode->i_op->permission you'll always
got directly to generic_permission in inode_permission, and in there
acl_permission_check -> check_acl should simply do the right thing with
your your patch to take the ACL cache checking into common code.