radix-tree: fix several shrinking bugs with multiorder entries

Setting the indirect bit on the user data entry used to be unambiguous
because the tree walking code knew not to expect internal nodes in the
last level of the tree.  Multiorder entries can appear at any level of
the tree, and a leaf with the indirect bit set is indistinguishable from
a pointer to a node.

Introduce a special entry (RADIX_TREE_RETRY) which is neither a valid
user entry, nor a valid pointer to a node.  The radix_tree_deref_retry()
function continues to work the same way, but tree walking code can
distinguish it from a pointer to a node.

Also fix the condition for setting slot->parent to NULL; it does not
matter what height the tree is, it only matters whether slot is an
indirect pointer.  Move this code above the comment which is referring
to the assignment to root->rnode.

Also fix the condition for preventing the tree from shrinking to a
single entry if it's a multiorder entry.

Add a test-case to the test suite that checks that the tree goes back
down to its original height after an item is inserted & deleted from a
higher index in the tree.

Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Jan Kara <jack@suse.com>
Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 12 insertions, 11 deletions

diff --git a/lib/radix-tree.c b/lib/radix-tree.c
index f13ddbba8ace..a1ba41730071 100644
--- a/lib/radix-tree.c
+++ b/lib/radix-tree.c
@@ -80,6 +80,8 @@ static inline void *indirect_to_ptr(void *ptr)
 	return (void *)((unsigned long)ptr & ~RADIX_TREE_INDIRECT_PTR);
 }
 
+#define RADIX_TREE_RETRY	ptr_to_indirect(NULL)
+
 #ifdef CONFIG_RADIX_TREE_MULTIORDER
 /* Sibling slots point directly to another slot in the same node */
 static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
@@ -1443,6 +1445,14 @@ static inline void radix_tree_shrink(struct radix_tree_root *root)
 		slot = to_free->slots[0];
 		if (!slot)
 			break;
+		if (!radix_tree_is_indirect_ptr(slot) && (root->height > 1))
+			break;
+
+		if (radix_tree_is_indirect_ptr(slot)) {
+			slot = indirect_to_ptr(slot);
+			slot->parent = NULL;
+			slot = ptr_to_indirect(slot);
+		}
 
 		/*
 		 * We don't need rcu_assign_pointer(), since we are simply
@@ -1451,14 +1461,6 @@ static inline void radix_tree_shrink(struct radix_tree_root *root)
 		 * (to_free->slots[0]), it will be safe to dereference the new
 		 * one (root->rnode) as far as dependent read barriers go.
 		 */
-		if (root->height > 1) {
-			if (!radix_tree_is_indirect_ptr(slot))
-				break;
-
-			slot = indirect_to_ptr(slot);
-			slot->parent = NULL;
-			slot = ptr_to_indirect(slot);
-		}
 		root->rnode = slot;
 		root->height--;
 
@@ -1480,9 +1482,8 @@ static inline void radix_tree_shrink(struct radix_tree_root *root)
 		 * also results in a stale slot). So tag the slot as indirect
 		 * to force callers to retry.
 		 */
-		if (root->height == 0)
-			*((unsigned long *)&to_free->slots[0]) |=
-						RADIX_TREE_INDIRECT_PTR;
+		if (!radix_tree_is_indirect_ptr(slot))
+			to_free->slots[0] = RADIX_TREE_RETRY;
 
 		radix_tree_node_free(to_free);
 	}