On Tue, Oct 11, 2016 at 01:34:18PM +0800, Ian Kent wrote:
> + path = file->f_path;
> +
> /*
> * An empty directory in an autofs file system is always a
> * mount point. The daemon must have failed to mount this
> @@ -123,7 +126,7 @@ static int autofs4_dir_open(struct inode *inode, struct
> file *file)
> * it.
> */
> spin_lock(&sbi->lookup_lock);
> - if (!d_mountpoint(dentry) && simple_empty(dentry)) {
> + if (!path_is_mountpoint(&path) && simple_empty(dentry)) {
Why not &file->f_path, provided that you constify that thing properly?
> + if (rcu_walk) {
> + if (!path_is_mountpoint_rcu(path))
> + return -EISDIR;
> + } else {
> + if (!path_is_mountpoint(path))
> + return -EISDIR;
IDGI. What's the point of _having_ the _rcu() variant, anyway? Here you
are probably paying more in terms of i-cache footprint/branch prediction
than you win on not doing that rcu_read_lock()/rcu_read_unlock()...
_rcu variants make sense when non-RCU case does something you can't do
under RCU; here your path_is_mountpoint() is pretty close to being
rcu_read_lock()+path_is_mountpoint_rcu()+rcu_read_unlock() anyway...
