eCryptfs: Prevent lower dentry from going negative during unlink

When calling vfs_unlink() on the lower dentry, d_delete() turns the
dentry into a negative dentry when the d_count is 1.  This eventually
caused a NULL pointer deref when a read() or write() was done and the
negative dentry's d_inode was dereferenced in
ecryptfs_read_update_atime() or ecryptfs_getxattr().

Placing mutt's tmpdir in an eCryptfs mount is what initially triggered
the oops and I was able to reproduce it with the following sequence:

open("/tmp/upper/foo", O_RDWR|O_CREAT|O_EXCL|O_NOFOLLOW, 0600) = 3
link("/tmp/upper/foo", "/tmp/upper/bar") = 0
unlink("/tmp/upper/foo")                = 0
open("/tmp/upper/bar", O_RDWR|O_CREAT|O_NOFOLLOW, 0600) = 4
unlink("/tmp/upper/bar")                = 0
write(4, "eCryptfs test\n"..., 14 <unfinished ...>
+++ killed by SIGKILL +++

https://bugs.launchpad.net/ecryptfs/+bug/387073

Reported-by: Loïc Minier <loic.minier@canonical.com>
Cc: Serge Hallyn <serue@us.ibm.com>
Cc: Dave Kleikamp <shaggy@linux.vnet.ibm.com>
Cc: ecryptfs-devel@lists.launchpad.net
Cc: stable <stable@kernel.org>
Signed-off-by: Tyler Hicks <tyhicks@linux.vnet.ibm.com>

fs/ecryptfs/inode.c

@@ -476,6 +476,7 @@ static int ecryptfs_unlink(struct inode *dir, struct dentry *dentry)
 	struct inode *lower_dir_inode = ecryptfs_inode_to_lower(dir);
 	struct dentry *lower_dir_dentry;
 
+	dget(lower_dentry);
 	lower_dir_dentry = lock_parent(lower_dentry);
 	rc = vfs_unlink(lower_dir_inode, lower_dentry);
 	if (rc) {
@@ -489,6 +490,7 @@ static int ecryptfs_unlink(struct inode *dir, struct dentry *dentry)
 	d_drop(dentry);
 out_unlock:
 	unlock_dir(lower_dir_dentry);
+	dput(lower_dentry);
 	return rc;
 }