radix_tree_tag_get() is not as safe as the docs make out [ver #2]

radix_tree_tag_get() is not safe to use concurrently with radix_tree_tag_set()
or radix_tree_tag_clear().  The problem is that the double tag_get() in
radix_tree_tag_get():

		if (!tag_get(node, tag, offset))
			saw_unset_tag = 1;
		if (height == 1) {
			int ret = tag_get(node, tag, offset);

may see the value change due to the action of set/clear.  RCU is no protection
against this as no pointers are being changed, no nodes are being replaced
according to a COW protocol - set/clear alter the node directly.

The documentation in linux/radix-tree.h, however, says that
radix_tree_tag_get() is an exception to the rule that "any function modifying
the tree or tags (...) must exclude other modifications, and exclude any
functions reading the tree".

The problem is that the next statement in radix_tree_tag_get() checks that the
tag doesn't vary over time:

			BUG_ON(ret && saw_unset_tag);

This has been seen happening in FS-Cache:

	https://www.redhat.com/archives/linux-cachefs/2010-April/msg00013.html

To this end, remove the BUG_ON() from radix_tree_tag_get() and note in various
comments that the value of the tag may change whilst the RCU read lock is held,
and thus that the return value of radix_tree_tag_get() may not be relied upon
unless radix_tree_tag_set/clear() and radix_tree_delete() are excluded from
running concurrently with it.

Reported-by: Romain DEGEZ <romain.degez@smartjog.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

2 files changed, 13 insertions, 6 deletions

diff --git a/include/linux/radix-tree.h b/include/linux/radix-tree.h
index c5da7491809..55ca73cf25e 100644
--- a/include/linux/radix-tree.h
+++ b/include/linux/radix-tree.h
@@ -121,6 +121,13 @@ do {									\
  * (Note, rcu_assign_pointer and rcu_dereference are not needed to control
  * access to data items when inserting into or looking up from the radix tree)
  *
+ * Note that the value returned by radix_tree_tag_get() may not be relied upon
+ * if only the RCU read lock is held.  Functions to set/clear tags and to
+ * delete nodes running concurrently with it may affect its result such that
+ * two consecutive reads in the same locked section may return different
+ * values.  If reliability is required, modification functions must also be
+ * excluded from concurrency.
+ *
  * radix_tree_tagged is able to be called without locking or RCU.
  */
 
diff --git a/lib/radix-tree.c b/lib/radix-tree.c
index 0871582aa29..2a087e0f986 100644
--- a/lib/radix-tree.c
+++ b/lib/radix-tree.c
@@ -555,6 +555,10 @@ EXPORT_SYMBOL(radix_tree_tag_clear);
  *
  *  0: tag not present or not set
  *  1: tag set
+ *
+ * Note that the return value of this function may not be relied on, even if
+ * the RCU lock is held, unless tag modification and node deletion are excluded
+ * from concurrency.
  */
 int radix_tree_tag_get(struct radix_tree_root *root,
 			unsigned long index, unsigned int tag)
@@ -595,12 +599,8 @@ int radix_tree_tag_get(struct radix_tree_root *root,
 		 */
 		if (!tag_get(node, tag, offset))
 			saw_unset_tag = 1;
-		if (height == 1) {
-			int ret = tag_get(node, tag, offset);
-
-			BUG_ON(ret && saw_unset_tag);
-			return !!ret;
-		}
+		if (height == 1)
+			return !!tag_get(node, tag, offset);
 		node = rcu_dereference_raw(node->slots[offset]);
 		shift -= RADIX_TREE_MAP_SHIFT;
 		height--;