[PATCH] namespace.c: fix race in mark_mounts_for_expiry()

This patch fixes a race found by Ram in mark_mounts_for_expiry() in
fs/namespace.c.

The bug can only be triggered with simultaneous exiting of a process having
a private namespace, and expiry of a mount from within that namespace.
It's practically impossible to trigger, and I haven't even tried.  But
still, a bug is a bug.

The race happens when put_namespace() is called by another task, while
mark_mounts_for_expiry() is between atomic_read() and get_namespace().  In
that case get_namespace() will be called on an already dead namespace with
unforeseeable results.

The solution was suggested by Al Viro, with his own words:

      Instead of screwing with atomic_read() in there, why don't we
      simply do the following:
      	a) atomic_dec_and_lock() in put_namespace()
      	b) __put_namespace() called without dropping lock
      	c) the first thing done by __put_namespace would be
      struct vfsmount *root = namespace->root;
      namespace->root = NULL;
      spin_unlock(...);
      ....
      umount_tree(root);
      ...
      	d) check in mark_... would be simply namespace && namespace->root.

      And we are all set; no screwing around with atomic_read(), no magic
      at all.  Dying namespace gets NULL ->root.
      All changes of ->root happen under spinlock.
      If under a spinlock we see non-NULL ->mnt_namespace, it won't be
      freed until we drop the lock (we will set ->mnt_namespace to NULL
      under that lock before we get to freeing namespace).
      If under a spinlock we see non-NULL ->mnt_namespace and
      ->mnt_namespace->root, we can grab a reference to namespace and be
      sure that it won't go away.

Signed-off-by: Miklos Szeredi <miklos@szeredi.hu>
Acked-by: Al Viro <viro@parcelfarce.linux.theplanet.co.uk>
Acked-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

fs/namespace.c

@@ -869,7 +869,7 @@ void mark_mounts_for_expiry(struct list_head *mounts)
 		/* don't do anything if the namespace is dead - all the
 		 * vfsmounts from it are going away anyway */
 		namespace = mnt->mnt_namespace;
-		if (!namespace || atomic_read(&namespace->count) <= 0)
+		if (!namespace || !namespace->root)
 			continue;
 		get_namespace(namespace);
 
@@ -1450,9 +1450,12 @@ void __init mnt_init(unsigned long mempages)
 
 void __put_namespace(struct namespace *namespace)
 {
+	struct vfsmount *root = namespace->root;
+	namespace->root = NULL;
+	spin_unlock(&vfsmount_lock);
 	down_write(&namespace->sem);
 	spin_lock(&vfsmount_lock);
-	umount_tree(namespace->root);
+	umount_tree(root);
 	spin_unlock(&vfsmount_lock);
 	up_write(&namespace->sem);
 	kfree(namespace);

include/linux/namespace.h

@@ -17,7 +17,8 @@ extern void __put_namespace(struct namespace *namespace);
 
 static inline void put_namespace(struct namespace *namespace)
 {
-	if (atomic_dec_and_test(&namespace->count))
+	if (atomic_dec_and_lock(&namespace->count, &vfsmount_lock))
+		/* releases vfsmount_lock */
 		__put_namespace(namespace);
 }