On Friday 24 September 2010 17:54:23 Jeff Layton wrote:

On Fri, 24 Sep 2010 18:18:11 +0530 
"Aneesh Kumar K.V" [off-list ref] wrote:

quoted

@@ -2415,7 +2418,7 @@ int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_de
 	if (!inode)
 		return -ENOENT;
 
-	error = may_create(dir, new_dentry);
+	error = may_create(dir, new_dentry, S_ISDIR(inode->i_mode));

						^^^^ this is a little
						scary, but even if it's
						a directory, it'll get
						kicked out in a later
						check. Would it be
						clearer to move up the
						S_ISDIR() check in this
						function and then pass
						this in as false?

Ah, you mean this:

--- a/fs/namei.c
+++ b/fs/namei.c

@@ -2450,7 +2450,9 @@ int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_de
 	if (!inode)
 		return -ENOENT;
 
-	error = may_create(dir, new_dentry, S_ISDIR(inode->i_mode));
+	if (S_ISDIR(inode->i_mode))
+		return -EPERM;
+	error = may_create(dir, new_dentry, 0);
 	if (error)
 		return error;
 

@@ -2464,8 +2466,6 @@ int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_de
 		return -EPERM;
 	if (!dir->i_op->link)
 		return -EPERM;
-	if (S_ISDIR(inode->i_mode))
-		return -EPERM;
 
 	error = security_inode_link(old_dentry, dir, new_dentry);
 	if (error)

This is a clear improvement; I don't think it matters that user-space will
get -EPERM instead of -EXDEV when trying to hard-link a directory across
devices.

Thanks,
Andreas