Don't automatically fail inserts if the write generation check fails: compare keys instead.

12 files changed, 129 insertions, 62 deletions

diff --git a/dist/serial.py b/dist/serial.py
index 8772c4bd6bc..46ebc86d0c2 100644
--- a/dist/serial.py
+++ b/dist/serial.py
@@ -20,6 +20,7 @@ Serial('col_append', [
 		SerialArg('uint32_t', 'write_gen'),
 		SerialArg('WT_INSERT_HEAD **', 'insheadp'),
 		SerialArg('WT_INSERT ***', 'ins_stack'),
+		SerialArg('WT_INSERT **', 'next_stack'),
 		SerialArg('WT_INSERT_HEAD **', 'new_inslist', 1),
 		SerialArg('WT_INSERT_HEAD *', 'new_inshead', 1),
 		SerialArg('WT_INSERT *', 'new_ins', 1),
@@ -31,6 +32,7 @@ Serial('insert', [
 		SerialArg('uint32_t', 'write_gen'),
 		SerialArg('WT_INSERT_HEAD **', 'inshead'),
 		SerialArg('WT_INSERT ***', 'ins_stack'),
+		SerialArg('WT_INSERT **', 'next_stack'),
 		SerialArg('WT_INSERT_HEAD **', 'new_inslist', 1),
 		SerialArg('WT_INSERT_HEAD *', 'new_inshead', 1),
 		SerialArg('WT_INSERT *', 'new_ins', 1),
diff --git a/src/btree/bt_curnext.c b/src/btree/bt_curnext.c
index d08f113fd90..6bfc4a8ad89 100644
--- a/src/btree/bt_curnext.c
+++ b/src/btree/bt_curnext.c
@@ -100,7 +100,7 @@ __cursor_fix_next(WT_CURSOR_BTREE *cbt, int newpage)
 new_page:	/* Check any insert list for a matching record. */
 		cbt->ins_head = WT_COL_UPDATE_SINGLE(cbt->page);
 		cbt->ins = __col_insert_search(
-		    cbt->ins_head, cbt->ins_stack, cbt->recno);
+		    cbt->ins_head, cbt->ins_stack, cbt->next_stack, cbt->recno);
 		if (cbt->ins != NULL &&
 		    cbt->recno != WT_INSERT_RECNO(cbt->ins))
 			cbt->ins = NULL;
diff --git a/src/btree/bt_curprev.c b/src/btree/bt_curprev.c
index 650f09f1a87..e836615c4bc 100644
--- a/src/btree/bt_curprev.c
+++ b/src/btree/bt_curprev.c
@@ -54,7 +54,8 @@ restart:
 			    session, cbt, cbt->ins_head, &key));
 		} else
 			cbt->ins = __col_insert_search(cbt->ins_head,
-			    cbt->ins_stack, WT_INSERT_RECNO(current));
+			    cbt->ins_stack, cbt->next_stack,
+			    WT_INSERT_RECNO(current));
 	}
 
 	/*
@@ -83,6 +84,7 @@ restart:
 	if (ins == NULL || ins == current)
 		for (; i >= 0; i--) {
 			cbt->ins_stack[i] = NULL;
+			cbt->next_stack[i] = NULL;
 			ins = cbt->ins_head->head[i];
 			if (ins != NULL && ins != current)
 				break;
@@ -96,12 +98,14 @@ restart:
 		 */
 		if (ins == NULL) {
 			cbt->ins_stack[0] = NULL;
+			cbt->next_stack[0] = NULL;
 			goto restart;
 		}
 		if (ins->next[i] != current)		/* Stay at this level */
 			ins = ins->next[i];
 		else {					/* Drop down a level */
 			cbt->ins_stack[i] = &ins->next[i];
+			cbt->next_stack[i] = ins->next[i];
 			--i;
 		}
 	}
@@ -231,7 +235,7 @@ __cursor_fix_prev(WT_CURSOR_BTREE *cbt, int newpage)
 new_page:	/* Check any insert list for a matching record. */
 		cbt->ins_head = WT_COL_UPDATE_SINGLE(cbt->page);
 		cbt->ins = __col_insert_search(
-		    cbt->ins_head, cbt->ins_stack, cbt->recno);
+		    cbt->ins_head, cbt->ins_stack, cbt->next_stack, cbt->recno);
 		if (cbt->ins != NULL &&
 		    cbt->recno != WT_INSERT_RECNO(cbt->ins))
 			cbt->ins = NULL;
diff --git a/src/btree/col_modify.c b/src/btree/col_modify.c
index 7035ccdc74b..fbe7898dc40 100644
--- a/src/btree/col_modify.c
+++ b/src/btree/col_modify.c
@@ -134,8 +134,10 @@ __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int op)
 		if (*inshead == NULL) {
 			new_inshead_size = sizeof(WT_INSERT_HEAD);
 			WT_ERR(__wt_calloc_def(session, 1, &new_inshead));
-			for (i = 0; i < WT_SKIP_MAXDEPTH; i++)
+			for (i = 0; i < WT_SKIP_MAXDEPTH; i++) {
 				cbt->ins_stack[i] = &new_inshead->head[i];
+				cbt->next_stack[i] = NULL;
+			}
 			cbt->ins_head = new_inshead;
 		}
 
@@ -164,7 +166,8 @@ __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int op)
 			ins_copy = ins;
 
 			WT_ERR(__wt_col_append_serial(session,
-			    page, cbt->write_gen, inshead, cbt->ins_stack,
+			    page, cbt->write_gen, inshead,
+			    cbt->ins_stack, cbt->next_stack,
 			    &new_inslist, new_inslist_size,
 			    &new_inshead, new_inshead_size,
 			    &ins, ins_size, skipdepth));
@@ -173,8 +176,8 @@ __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int op)
 			cbt->recno = WT_INSERT_RECNO(ins_copy);
 		} else
 			WT_ERR(__wt_insert_serial(session,
-			    page, cbt->write_gen,
-			    inshead, cbt->ins_stack,
+			    page, cbt->write_gen, inshead,
+			    cbt->ins_stack, cbt->next_stack,
 			    &new_inslist, new_inslist_size,
 			    &new_inshead, new_inshead_size,
 			    &ins, ins_size, skipdepth));
@@ -230,7 +233,7 @@ int
 __wt_col_append_serial_func(WT_SESSION_IMPL *session, void *args)
 {
 	WT_BTREE *btree;
-	WT_INSERT *ins, *new_ins, ***ins_stack;
+	WT_INSERT *ins, *new_ins, ***ins_stack, **next_stack;
 	WT_INSERT_HEAD *inshead, **insheadp, **new_inslist, *new_inshead;
 	WT_PAGE *page;
 	uint64_t recno;
@@ -239,8 +242,9 @@ __wt_col_append_serial_func(WT_SESSION_IMPL *session, void *args)
 
 	btree = session->btree;
 
-	__wt_col_append_unpack(args, &page, &write_gen, &insheadp,
-	    &ins_stack, &new_inslist, &new_inshead, &new_ins, &skipdepth);
+	__wt_col_append_unpack(args,
+	    &page, &write_gen, &insheadp, &ins_stack, &next_stack,
+	    &new_inslist, &new_inshead, &new_ins, &skipdepth);
 
 	/* Check the page's write-generation. */
 	WT_RET(__wt_page_write_gen_check(session, page, write_gen));
@@ -259,7 +263,7 @@ __wt_col_append_serial_func(WT_SESSION_IMPL *session, void *args)
 	if ((recno = WT_INSERT_RECNO(new_ins)) == 0)
 		recno = WT_INSERT_RECNO(new_ins) = ++btree->last_recno;
 
-	ins = __col_insert_search(inshead, ins_stack, recno);
+	ins = __col_insert_search(inshead, ins_stack, next_stack, recno);
 
 	/* If we find the record number, there's been a race. */
 	if (ins != NULL && WT_INSERT_RECNO(ins) == recno)
diff --git a/src/btree/col_srch.c b/src/btree/col_srch.c
index c2fbf80c2e2..ffcca036cab 100644
--- a/src/btree/col_srch.c
+++ b/src/btree/col_srch.c
@@ -117,8 +117,8 @@ __wt_col_search(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int is_modify)
 	 * For that reason, don't set the cursor's WT_INSERT_HEAD/WT_INSERT pair
 	 * until we know we have a useful entry.
 	 */
-	if ((ins =
-	    __col_insert_search(ins_head, cbt->ins_stack, recno)) != NULL)
+	if ((ins = __col_insert_search(
+	    ins_head, cbt->ins_stack, cbt->next_stack, recno)) != NULL)
 		if (recno == WT_INSERT_RECNO(ins)) {
 			cbt->ins_head = ins_head;
 			cbt->ins = ins;
@@ -135,8 +135,8 @@ past_end:
 	 * past the end of the table.
 	 */
 	cbt->ins_head = WT_COL_APPEND(page);
-	if ((cbt->ins =
-	    __col_insert_search(cbt->ins_head, cbt->ins_stack, recno)) == NULL)
+	if ((cbt->ins = __col_insert_search(
+	    cbt->ins_head, cbt->ins_stack, cbt->next_stack, recno)) == NULL)
 		cbt->compare = -1;
 	else {
 		cbt->recno = WT_INSERT_RECNO(cbt->ins);
diff --git a/src/btree/row_modify.c b/src/btree/row_modify.c
index 2ad6947d6de..64123bc53e5 100644
--- a/src/btree/row_modify.c
+++ b/src/btree/row_modify.c
@@ -121,8 +121,10 @@ __wt_row_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int is_remove)
 		if (*inshead == NULL) {
 			new_inshead_size = sizeof(WT_INSERT_HEAD);
 			WT_ERR(__wt_calloc_def(session, 1, &new_inshead));
-			for (i = 0; i < WT_SKIP_MAXDEPTH; i++)
+			for (i = 0; i < WT_SKIP_MAXDEPTH; i++) {
 				cbt->ins_stack[i] = &new_inshead->head[i];
+				cbt->next_stack[i] = NULL;
+			}
 			cbt->ins_head = new_inshead;
 		}
 
@@ -144,7 +146,7 @@ __wt_row_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int is_remove)
 
 		/* Insert the WT_INSERT structure. */
 		WT_ERR(__wt_insert_serial(session, page, cbt->write_gen,
-		    inshead, cbt->ins_stack,
+		    inshead, cbt->ins_stack, cbt->next_stack,
 		    &new_inslist, new_inslist_size,
 		    &new_inshead, new_inshead_size,
 		    &ins, ins_size, skipdepth));
@@ -205,17 +207,38 @@ __wt_row_insert_alloc(WT_SESSION_IMPL *session,
 int
 __wt_insert_serial_func(WT_SESSION_IMPL *session, void *args)
 {
-	WT_INSERT *new_ins, ***ins_stack;
+	WT_INSERT *new_ins, ***ins_stack, **next_stack;
 	WT_INSERT_HEAD *inshead, **insheadp, **new_inslist, *new_inshead;
 	WT_PAGE *page;
 	uint32_t write_gen;
 	u_int i, skipdepth;
 
 	__wt_insert_unpack(args, &page, &write_gen, &insheadp,
-	    &ins_stack, &new_inslist, &new_inshead, &new_ins, &skipdepth);
+	    &ins_stack, &next_stack,
+	    &new_inslist, &new_inshead, &new_ins, &skipdepth);
 
-	/* Check the page's write-generation. */
-	WT_RET(__wt_page_write_gen_check(session, page, write_gen));
+	if ((inshead = *insheadp) == NULL)
+		inshead = new_inshead;
+
+	/*
+	 * Check the page's write-generation: if that fails, check whether we
+	 * are still in the expected position, and no item has been added where
+	 * our insert belongs.
+	 */
+	if (page->modify->write_gen + 1 == page->modify->disk_gen)
+		return (WT_RESTART);
+
+	if (page->modify->write_gen != write_gen) {
+		for (i = 0; i < skipdepth; i++) {
+			if (ins_stack[i] == NULL ||
+			    *ins_stack[i] != next_stack[i])
+				return (WT_RESTART);
+			if (next_stack[i] == NULL &&
+			    inshead->tail[i] != NULL &&
+			    ins_stack[i] != &inshead->tail[i]->next[i])
+				return (WT_RESTART);
+		}
+	}
 
 	/*
 	 * Publish: First, point the new WT_INSERT item's skiplist references
@@ -223,10 +246,9 @@ __wt_insert_serial_func(WT_SESSION_IMPL *session, void *args)
 	 * update the skiplist elements that reference the new WT_INSERT item,
 	 * this ensures the list is never inconsistent.
 	 */
-	if ((inshead = *insheadp) == NULL)
-		inshead = new_inshead;
 	for (i = 0; i < skipdepth; i++)
 		new_ins->next[i] = *ins_stack[i];
+
 	WT_WRITE_BARRIER();
 	for (i = 0; i < skipdepth; i++) {
 		if (inshead->tail[i] == NULL ||
diff --git a/src/btree/row_srch.c b/src/btree/row_srch.c
index 885979bffbf..f005db24a1e 100644
--- a/src/btree/row_srch.c
+++ b/src/btree/row_srch.c
@@ -16,27 +16,39 @@ __wt_search_insert(WT_SESSION_IMPL *session,
     WT_CURSOR_BTREE *cbt, WT_INSERT_HEAD *inshead, WT_ITEM *srch_key)
 {
 	WT_BTREE *btree;
-	WT_INSERT **insp, *ret_ins;
+	WT_INSERT **insp, *last_ins, *ret_ins;
 	WT_ITEM insert_key;
 	int cmp, i;
 
+	btree = session->btree;
+
 	/* If there's no insert chain to search, we're done. */
 	if ((ret_ins = WT_SKIP_LAST(inshead)) == NULL) {
 		cbt->ins = NULL;
+		cbt->next_stack[0] = NULL;
 		return (0);
 	}
 
-	btree = session->btree;
-
 	/* Fast-path appends. */
 	insert_key.data = WT_INSERT_KEY(ret_ins);
 	insert_key.size = WT_INSERT_KEY_SIZE(ret_ins);
 	(void)WT_BTREE_CMP(session, btree, srch_key, &insert_key, cmp);
 	if (cmp >= 0) {
-		for (i = WT_SKIP_MAXDEPTH - 1; i >= 0; i--)
-			cbt->ins_stack[i] = (inshead->tail[i] != NULL) ?
-			    &inshead->tail[i]->next[i] :
-			    &inshead->head[i];
+		/*
+		 * XXX We may race with another appending thread.
+		 *
+		 * To catch that case, rely on the atomic pointer read above
+		 * and set the next stack to NULL here.  If we have raced with
+		 * another thread, one of the next pointers will not be NULL by
+		 * the time they are checked against the next stack inside the
+		 * serialized insert function.
+		 */
+		for (i = WT_SKIP_MAXDEPTH - 1; i >= 0; i--) {
+			cbt->ins_stack[i] = (i == 0) ? &ret_ins->next[0] :
+			    (inshead->tail[i] != NULL) ?
+			    &inshead->tail[i]->next[i] : &inshead->head[i];
+			cbt->next_stack[i] = NULL;
+		}
 		cbt->compare = -cmp;
 		cbt->ins = ret_ins;
 		return (0);
@@ -46,9 +58,10 @@ __wt_search_insert(WT_SESSION_IMPL *session,
 	 * The insert list is a skip list: start at the highest skip level, then
 	 * go as far as possible at each level before stepping down to the next.
 	 */
-	ret_ins = NULL;
+	last_ins = ret_ins = NULL;
 	for (i = WT_SKIP_MAXDEPTH - 1, insp = &inshead->head[i]; i >= 0;) {
-		if (*insp == NULL) {
+		if ((ret_ins = *insp) == NULL) {
+			cbt->next_stack[i] = NULL;
 			cbt->ins_stack[i--] = insp--;
 			continue;
 		}
@@ -57,8 +70,8 @@ __wt_search_insert(WT_SESSION_IMPL *session,
 		 * Comparisons may be repeated as we drop down skiplist levels;
 		 * don't repeat comparisons, they might be expensive.
 		 */
-		if (ret_ins != *insp) {
-			ret_ins = *insp;
+		if (ret_ins != last_ins) {
+			last_ins = ret_ins;
 			insert_key.data = WT_INSERT_KEY(ret_ins);
 			insert_key.size = WT_INSERT_KEY_SIZE(ret_ins);
 			WT_RET(WT_BTREE_CMP(
@@ -68,10 +81,14 @@ __wt_search_insert(WT_SESSION_IMPL *session,
 		if (cmp > 0)		/* Keep going at this level */
 			insp = &ret_ins->next[i];
 		else if (cmp == 0)
-			for (; i >= 0; i--)
+			for (; i >= 0; i--) {
+				cbt->next_stack[i] = ret_ins->next[i];
 				cbt->ins_stack[i] = &ret_ins->next[i];
-		else			/* Drop down a level */
+			}
+		else {			/* Drop down a level */
+			cbt->next_stack[i] = ret_ins;
 			cbt->ins_stack[i--] = insp--;
+		}
 	}
 
 	/*
diff --git a/src/include/cache.i b/src/include/cache.i
index 4312304d968..5ecc0c11701 100644
--- a/src/include/cache.i
+++ b/src/include/cache.i
@@ -55,7 +55,7 @@ __wt_cache_full_check(WT_SESSION_IMPL *session)
 	for (wake = 0;; wake = (wake + 1) % 100) {
 		__wt_eviction_check(session, &lockout, wake == 0);
 		if (!lockout || F_ISSET(session,
-                    WT_SESSION_NO_CACHE_CHECK | WT_SESSION_SCHEMA_LOCKED))
+		    WT_SESSION_NO_CACHE_CHECK | WT_SESSION_SCHEMA_LOCKED))
 			return (0);
 		if ((btree = session->btree) != NULL && F_ISSET(btree,
 		    WT_BTREE_BULK | WT_BTREE_NO_CACHE | WT_BTREE_NO_EVICTION))
diff --git a/src/include/column.i b/src/include/column.i
index 83b2783dbf0..0cf83a6d2e9 100644
--- a/src/include/column.i
+++ b/src/include/column.i
@@ -58,8 +58,8 @@ __col_insert_search_match(WT_INSERT_HEAD *inshead, uint64_t recno)
  *	Search a column-store insert list, creating a skiplist stack as we go.
  */
 static inline WT_INSERT *
-__col_insert_search(
-    WT_INSERT_HEAD *inshead, WT_INSERT ***ins_stack, uint64_t recno)
+__col_insert_search(WT_INSERT_HEAD *inshead,
+    WT_INSERT ***ins_stack, WT_INSERT **next_stack, uint64_t recno)
 {
 	WT_INSERT **insp, *ret_ins;
 	uint64_t ins_recno;
@@ -71,9 +71,12 @@ __col_insert_search(
 
 	/* Fast path appends. */
 	if (recno >= WT_INSERT_RECNO(ret_ins)) {
-		for (i = 0; i < WT_SKIP_MAXDEPTH; i++)
-			ins_stack[i] = (inshead->tail[i] != NULL) ?
+		for (i = 0; i < WT_SKIP_MAXDEPTH; i++) {
+			ins_stack[i] = (i == 0) ? &ret_ins->next[0] :
+			    (inshead->tail[i] != NULL) ?
 			    &inshead->tail[i]->next[i] : &inshead->head[i];
+			next_stack[i] = NULL;
+		}
 		return (ret_ins);
 	}
 
@@ -82,22 +85,26 @@ __col_insert_search(
 	 * go as far as possible at each level before stepping down to the next.
 	 */
 	for (i = WT_SKIP_MAXDEPTH - 1, insp = &inshead->head[i]; i >= 0; ) {
-		if (*insp == NULL) {
+		if ((ret_ins = *insp) == NULL) {
+			next_stack[i] = NULL;
 			ins_stack[i--] = insp--;
 			continue;
 		}
 
-		ret_ins = *insp;
 		ins_recno = WT_INSERT_RECNO(ret_ins);
 		cmp = (recno == ins_recno) ? 0 : (recno < ins_recno) ? -1 : 1;
 
 		if (cmp > 0)			/* Keep going at this level */
-			insp = &(*insp)->next[i];
+			insp = &ret_ins->next[i];
 		else if (cmp == 0)		/* Exact match: return */
-			for (; i >= 0; i--)
+			for (; i >= 0; i--) {
+				next_stack[i] = ret_ins->next[i];
 				ins_stack[i] = &ret_ins->next[i];
-		else				/* Drop down a level */
+			}
+		else {				/* Drop down a level */
+			next_stack[i] = ret_ins;
 			ins_stack[i--] = insp--;
+		}
 	}
 	return (ret_ins);
 }
diff --git a/src/include/cursor.h b/src/include/cursor.h
index 3afc5a8ea7c..85c61076b75 100644
--- a/src/include/cursor.h
+++ b/src/include/cursor.h
@@ -36,6 +36,9 @@ struct __wt_cursor_btree {
 					/* Search stack */
 	WT_INSERT	**ins_stack[WT_SKIP_MAXDEPTH];
 
+					/* Next item(s) found during search */
+	WT_INSERT	*next_stack[WT_SKIP_MAXDEPTH];
+
 	uint64_t recno;			/* Record number */
 	uint32_t write_gen;		/* Saved leaf page's write generation */
 
diff --git a/src/include/serial_funcs.i b/src/include/serial_funcs.i
index bbddf47042b..4b39b877088 100644
--- a/src/include/serial_funcs.i
+++ b/src/include/serial_funcs.i
@@ -5,6 +5,7 @@ typedef struct {
 	uint32_t write_gen;
 	WT_INSERT_HEAD **insheadp;
 	WT_INSERT ***ins_stack;
+	WT_INSERT **next_stack;
 	WT_INSERT_HEAD **new_inslist;
 	size_t new_inslist_size;
 	int new_inslist_taken;
@@ -20,10 +21,10 @@ typedef struct {
 static inline int
 __wt_col_append_serial(
 	WT_SESSION_IMPL *session, WT_PAGE *page, uint32_t write_gen,
-	WT_INSERT_HEAD **insheadp, WT_INSERT ***ins_stack, WT_INSERT_HEAD
-	***new_inslistp, size_t new_inslist_size, WT_INSERT_HEAD
-	**new_insheadp, size_t new_inshead_size, WT_INSERT **new_insp, size_t
-	new_ins_size, u_int skipdepth)
+	WT_INSERT_HEAD **insheadp, WT_INSERT ***ins_stack, WT_INSERT
+	**next_stack, WT_INSERT_HEAD ***new_inslistp, size_t new_inslist_size,
+	WT_INSERT_HEAD **new_insheadp, size_t new_inshead_size, WT_INSERT
+	**new_insp, size_t new_ins_size, u_int skipdepth)
 {
 	__wt_col_append_args _args, *args = &_args;
 	WT_DECL_RET;
@@ -36,6 +37,8 @@ __wt_col_append_serial(
 
 	args->ins_stack = ins_stack;
 
+	args->next_stack = next_stack;
+
 	if (new_inslistp == NULL)
 		args->new_inslist = NULL;
 	else {
@@ -81,9 +84,9 @@ __wt_col_append_serial(
 static inline void
 __wt_col_append_unpack(
     void *untyped_args, WT_PAGE **pagep, uint32_t *write_genp,
-    WT_INSERT_HEAD ***insheadpp, WT_INSERT ****ins_stackp, WT_INSERT_HEAD
-    ***new_inslistp, WT_INSERT_HEAD **new_insheadp, WT_INSERT **new_insp,
-    u_int *skipdepthp)
+    WT_INSERT_HEAD ***insheadpp, WT_INSERT ****ins_stackp, WT_INSERT
+    ***next_stackp, WT_INSERT_HEAD ***new_inslistp, WT_INSERT_HEAD
+    **new_insheadp, WT_INSERT **new_insp, u_int *skipdepthp)
 {
 	__wt_col_append_args *args = (__wt_col_append_args *)untyped_args;
 
@@ -91,6 +94,7 @@ __wt_col_append_unpack(
 	*write_genp = args->write_gen;
 	*insheadpp = args->insheadp;
 	*ins_stackp = args->ins_stack;
+	*next_stackp = args->next_stack;
 	*new_inslistp = args->new_inslist;
 	*new_insheadp = args->new_inshead;
 	*new_insp = args->new_ins;
@@ -138,6 +142,7 @@ typedef struct {
 	uint32_t write_gen;
 	WT_INSERT_HEAD **inshead;
 	WT_INSERT ***ins_stack;
+	WT_INSERT **next_stack;
 	WT_INSERT_HEAD **new_inslist;
 	size_t new_inslist_size;
 	int new_inslist_taken;
@@ -153,10 +158,10 @@ typedef struct {
 static inline int
 __wt_insert_serial(
 	WT_SESSION_IMPL *session, WT_PAGE *page, uint32_t write_gen,
-	WT_INSERT_HEAD **inshead, WT_INSERT ***ins_stack, WT_INSERT_HEAD
-	***new_inslistp, size_t new_inslist_size, WT_INSERT_HEAD
-	**new_insheadp, size_t new_inshead_size, WT_INSERT **new_insp, size_t
-	new_ins_size, u_int skipdepth)
+	WT_INSERT_HEAD **inshead, WT_INSERT ***ins_stack, WT_INSERT
+	**next_stack, WT_INSERT_HEAD ***new_inslistp, size_t new_inslist_size,
+	WT_INSERT_HEAD **new_insheadp, size_t new_inshead_size, WT_INSERT
+	**new_insp, size_t new_ins_size, u_int skipdepth)
 {
 	__wt_insert_args _args, *args = &_args;
 	WT_DECL_RET;
@@ -169,6 +174,8 @@ __wt_insert_serial(
 
 	args->ins_stack = ins_stack;
 
+	args->next_stack = next_stack;
+
 	if (new_inslistp == NULL)
 		args->new_inslist = NULL;
 	else {
@@ -214,9 +221,9 @@ __wt_insert_serial(
 static inline void
 __wt_insert_unpack(
     void *untyped_args, WT_PAGE **pagep, uint32_t *write_genp,
-    WT_INSERT_HEAD ***insheadp, WT_INSERT ****ins_stackp, WT_INSERT_HEAD
-    ***new_inslistp, WT_INSERT_HEAD **new_insheadp, WT_INSERT **new_insp,
-    u_int *skipdepthp)
+    WT_INSERT_HEAD ***insheadp, WT_INSERT ****ins_stackp, WT_INSERT
+    ***next_stackp, WT_INSERT_HEAD ***new_inslistp, WT_INSERT_HEAD
+    **new_insheadp, WT_INSERT **new_insp, u_int *skipdepthp)
 {
 	__wt_insert_args *args = (__wt_insert_args *)untyped_args;
 
@@ -224,6 +231,7 @@ __wt_insert_unpack(
 	*write_genp = args->write_gen;
 	*insheadp = args->inshead;
 	*ins_stackp = args->ins_stack;
+	*next_stackp = args->next_stack;
 	*new_inslistp = args->new_inslist;
 	*new_insheadp = args->new_inshead;
 	*new_insp = args->new_ins;
diff --git a/src/lsm/lsm_worker.c b/src/lsm/lsm_worker.c
index 429c38d39ed..647d79a3897 100644
--- a/src/lsm/lsm_worker.c
+++ b/src/lsm/lsm_worker.c
@@ -186,11 +186,11 @@ __wt_lsm_checkpoint_worker(void *arg)
 			}
 
 			++j;
-			__wt_spin_lock(session, &lsm_tree->lock);
+			__wt_writelock(session, lsm_tree->rwlock);
 			F_SET(chunk, WT_LSM_CHUNK_ONDISK);
 			++lsm_tree->dsk_gen;
 			ret = __wt_lsm_meta_write(session, lsm_tree);
-			__wt_spin_unlock(session, &lsm_tree->lock);
+			__wt_rwunlock(session, lsm_tree->rwlock);
 
 			if (ret != 0) {
 				(void)__wt_err(session, ret,