Change WT_RESTART into an "expected" failure, similar to WT_NOTFOUND

when walking the cache, requires the caller to release the held page.

3 files changed, 23 insertions, 15 deletions

diff --git a/src/btree/bt_walk.c b/src/btree/bt_walk.c
index 7abbc994938..d1dd0b45d62 100644
--- a/src/btree/bt_walk.c
+++ b/src/btree/bt_walk.c
@@ -160,7 +160,7 @@ __tree_walk_read(WT_SESSION_IMPL *session, WT_REF *ref, int *skipp)
 #undef	PAGE_SWAP
 #define	PAGE_SWAP(session, couple, page, ref, flags, ret) do {		\
 	if (((ret) = __wt_page_swap(					\
-	    session, couple, page, ref, 0, flags)) == WT_RESTART)	\
+	    session, couple, page, ref, flags)) == WT_RESTART)		\
 		goto restart;						\
 } while (0)
 
@@ -278,12 +278,13 @@ ascend:	/*
 			 * Can this page-swap function return restart because of
 			 * a page split?  I don't think so (this is an internal
 			 * page that's pinned in memory), but I'm not 100% sure.
+			 * Note we're not handling a return of WT_RESTART.
 			 */
 			if (WT_PAGE_IS_ROOT(page))
 				WT_RET(__wt_page_release(session, couple));
 			else
 				WT_RET(__wt_page_swap(session, couple, page,
-				    __wt_page_ref(session, page), 1, flags));
+				    __wt_page_ref(session, page), flags));
 
 			*pagep = page;
 			return (0);
diff --git a/src/btree/col_srch.c b/src/btree/col_srch.c
index c8c842961e9..69850b0edd2 100644
--- a/src/btree/col_srch.c
+++ b/src/btree/col_srch.c
@@ -92,16 +92,17 @@ descend:	WT_ASSERT(session, ref != NULL);
 		 * while we're waiting for it, restart the search, otherwise
 		 * return on error.
 		 */
-		if ((ret =
-		    __wt_page_swap(session, page, page, ref, 1, 0)) == 0) {
+		if ((ret = __wt_page_swap(session, page, page, ref, 0)) == 0) {
 			page = ref->page;
 			continue;
 		}
 		/*
-		 * Restart is returned if we find a page that's been split;
-		 * restart the search from the top of the tree.
+		 * Restart is returned if we find a page that's been split; the
+		 * held page isn't discarded when restart is returned, discard
+		 * it and restart the search from the top of the tree.
 		 */
-		if (ret == WT_RESTART)
+		if (ret == WT_RESTART &&
+		    (ret = __wt_page_release(session, page)) == 0)
 			goto restart;
 		return (ret);
 	}
diff --git a/src/btree/row_srch.c b/src/btree/row_srch.c
index 93e589f6809..f10bba16b4a 100644
--- a/src/btree/row_srch.c
+++ b/src/btree/row_srch.c
@@ -320,16 +320,17 @@ descend:	WT_ASSERT(session, ref != NULL);
 		 * return on error, the swap call ensures we're holding nothing
 		 * on failure.
 		 */
-		if ((ret =
-		    __wt_page_swap(session, page, page, ref, 1, 0)) == 0) {
+		if ((ret = __wt_page_swap(session, page, page, ref, 0)) == 0) {
 			page = ref->page;
 			continue;
 		}
 		/*
-		 * Restart is returned if we find a page that's been split;
-		 * restart the search from the top of the tree.
+		 * Restart is returned if we find a page that's been split; the
+		 * held page isn't discarded when restart is returned, discard
+		 * it and restart the search from the top of the tree.
 		 */
-		if (ret == WT_RESTART)
+		if (ret == WT_RESTART &&
+		    (ret = __wt_page_release(session, page)) == 0)
 			goto restart;
 		return (ret);
 	}
@@ -488,12 +489,17 @@ restart:
 		 * Swap the parent page for the child page; return on error,
 		 * the swap function ensures we're holding nothing on failure.
 		 */
-		if ((ret =
-		    __wt_page_swap(session, page, page, ref, 1, 0)) == 0) {
+		if ((ret = __wt_page_swap(session, page, page, ref, 0)) == 0) {
 			page = ref->page;
 			continue;
 		}
-		if (ret == WT_RESTART)
+		/*
+		 * Restart is returned if we find a page that's been split; the
+		 * held page isn't discarded when restart is returned, discard
+		 * it and restart the search from the top of the tree.
+		 */
+		if (ret == WT_RESTART &&
+		    (ret = __wt_page_release(session, page)) == 0)
 			goto restart;
 		return (ret);
 	}