Use a split generation number to determine when memory replaced by a split can be freed.

11 files changed, 68 insertions, 71 deletions

diff --git a/src/btree/bt_cursor.c b/src/btree/bt_cursor.c
index ede7d733bfc..8a667392aeb 100644
--- a/src/btree/bt_cursor.c
+++ b/src/btree/bt_cursor.c
@@ -216,13 +216,14 @@ __cursor_valid(WT_CURSOR_BTREE *cbt, WT_UPDATE **updp)
 static inline int
 __cursor_col_search(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt)
 {
-	WT_BTREE *prev_btree;
+	uint64_t prev_split_gen;
 	WT_DECL_RET;
 
-	prev_btree = session->active_btree;
-	WT_PUBLISH(session->active_btree, cbt->btree);
+	if ((prev_split_gen = session->split_gen) == 0)
+		WT_PUBLISH(session->split_gen, S2C(session)->split_gen);
 	ret = __wt_col_search(session, cbt->iface.recno, NULL, cbt);
-	session->active_btree = prev_btree;
+	if (prev_split_gen == 0)
+		session->split_gen = 0;
 	return (ret);
 }
 
@@ -233,13 +234,14 @@ __cursor_col_search(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt)
 static inline int
 __cursor_row_search(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int insert)
 {
-	WT_BTREE *prev_btree;
+	uint64_t prev_split_gen;
 	WT_DECL_RET;
 
-	prev_btree = session->active_btree;
-	WT_PUBLISH(session->active_btree, cbt->btree);
+	if ((prev_split_gen = session->split_gen) == 0)
+		WT_PUBLISH(session->split_gen, S2C(session)->split_gen);
 	ret = __wt_row_search(session, &cbt->iface.key, NULL, cbt, insert);
-	session->active_btree = prev_btree;
+	if (prev_split_gen == 0)
+		session->split_gen = 0;
 	return (ret);
 }
 
diff --git a/src/btree/bt_walk.c b/src/btree/bt_walk.c
index f5a93210a26..ed9992b89ab 100644
--- a/src/btree/bt_walk.c
+++ b/src/btree/bt_walk.c
@@ -14,12 +14,13 @@
 int
 __wt_tree_walk(WT_SESSION_IMPL *session, WT_REF **refp, uint32_t flags)
 {
-	WT_BTREE *btree, *prev_tree;
+	WT_BTREE *btree;
 	WT_DECL_RET;
 	WT_PAGE *page;
 	WT_PAGE_INDEX *pindex;
 	WT_REF *couple, *ref;
 	WT_TXN_STATE *txn_state;
+	uint64_t prev_split_gen;
 	int descending, prev, skip;
 	uint32_t slot;
 
@@ -31,8 +32,8 @@ __wt_tree_walk(WT_SESSION_IMPL *session, WT_REF **refp, uint32_t flags)
 	 * may want to free.  Publish that the tree is active during this
 	 * window.
 	 */
-	prev_tree = session->active_btree;
-	WT_PUBLISH(session->active_btree, btree);
+	if ((prev_split_gen = session->split_gen) == 0)
+		WT_PUBLISH(session->split_gen, S2C(session)->split_gen);
 
 	/*
 	 * !!!
@@ -279,6 +280,7 @@ descend:		couple = ref;
 done:
 err:	if (txn_state != NULL)
 		txn_state->snap_min = WT_TXN_NONE;
-	session->active_btree = prev_tree;
+	if (prev_split_gen == 0)
+		session->split_gen = 0;
 	return (ret);
 }
diff --git a/src/btree/rec_split.c b/src/btree/rec_split.c
index 8296e19542d..6d80924896c 100644
--- a/src/btree/rec_split.c
+++ b/src/btree/rec_split.c
@@ -31,7 +31,8 @@ __safe_free(
 	 * We have swapped something in a page: if we don't have exclusive
 	 * access, check whether there are other threads in the same tree.
 	 */
-	if (!exclusive && __wt_btree_exclusive(session, S2BT(session)))
+	if (!exclusive &&
+	    __wt_oldest_split_gen(session) == S2C(session)->split_gen + 1)
 		exclusive = 1;
 
 	if (exclusive) {
diff --git a/src/conn/conn_handle.c b/src/conn/conn_handle.c
index a6c52678fbf..017658bafe4 100644
--- a/src/conn/conn_handle.c
+++ b/src/conn/conn_handle.c
@@ -28,6 +28,8 @@ __wt_connection_init(WT_CONNECTION_IMPL *conn)
 
 	TAILQ_INIT(&conn->lsmqh);		/* WT_LSM_TREE list */
 
+	conn->split_gen = 1;
+
 	/* Configuration. */
 	WT_RET(__wt_conn_config_init(session));
 
diff --git a/src/include/btree.i b/src/include/btree.i
index 7d71bfdb400..fb686b7bce7 100644
--- a/src/include/btree.i
+++ b/src/include/btree.i
@@ -937,27 +937,6 @@ __wt_page_hazard_check(WT_SESSION_IMPL *session, WT_PAGE *page)
 }
 
 /*
- * __wt_btree_exclusive --
- *	Check if there is another thread active in this tree.
- */
-static inline int
-__wt_btree_exclusive(WT_SESSION_IMPL *session, WT_BTREE *btree)
-{
-	WT_CONNECTION_IMPL *conn;
-	WT_SESSION_IMPL *s;
-	uint32_t i, session_cnt;
-
-	conn = S2C(session);
-
-	/* No lock is required because the session array is fixed size. */
-	WT_ORDERED_READ(session_cnt, conn->session_cnt);
-	for (s = conn->sessions, i = 0; i < session_cnt; ++s, ++i)
-		if (s->active_btree == btree)
-			return (0);
-	return (1);
-}
-
-/*
  * __wt_skip_choose_depth --
  *	Randomly choose a depth for a skiplist insert.
  */
diff --git a/src/include/connection.h b/src/include/connection.h
index 245308e4a6c..f8ae7930aca 100644
--- a/src/include/connection.h
+++ b/src/include/connection.h
@@ -109,6 +109,8 @@ struct __wt_connection_impl {
 	pthread_t cache_evict_tid;	/* Eviction server thread ID */
 	int	  cache_evict_tid_set;	/* Eviction server thread ID set */
 
+	uint64_t  split_gen;		/* Generation number for splits */
+
 	WT_SPINLOCK dhandle_lock;	/* Locked: dhandle sweep */
 					/* Locked: data handle list */
 	SLIST_HEAD(__wt_dhandle_lh, __wt_data_handle) dhlh;
diff --git a/src/include/extern.h b/src/include/extern.h
index fd717e893ac..584b369abb3 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -1387,6 +1387,7 @@ extern int __wt_session_lock_checkpoint(WT_SESSION_IMPL *session,
     const char *checkpoint);
 extern void __wt_session_discard_btree( WT_SESSION_IMPL *session,
     WT_DATA_HANDLE_CACHE *dhandle_cache);
+extern uint64_t __wt_oldest_split_gen(WT_SESSION_IMPL *session);
 extern int __wt_session_fotxn_add(WT_SESSION_IMPL *session,
     void *p,
     size_t len);
diff --git a/src/include/session.h b/src/include/session.h
index 619fc970d80..133f8ab107e 100644
--- a/src/include/session.h
+++ b/src/include/session.h
@@ -129,21 +129,21 @@ struct __wt_session_impl {
 
 	/*
 	 * Sessions can "free" memory that may still be in use, and we use a
-	 * transactional generation to track it, that is, the session stores
-	 * a reference to the memory and a current transaction ID; when the
-	 * oldest transaction ID has moved beyond that point, the memory can
-	 * be discarded for real.
+	 * split generation number to track it, that is, the session stores a
+	 * reference to the memory and allocates a split generation; when no
+	 * session is reading from that split generation, the memory can be
+	 * discarded for real.
 	 */
-	WT_BTREE *active_btree;
 	struct __wt_fotxn {
-		WT_BTREE   *btree;	/* Tree that owned the memory */
-		uint64_t    txnid;	/* Transaction ID */
+		uint64_t    split_gen;	/* Tree generation */
 		void       *p;		/* Memory, length */
 		size_t	    len;
 	} *fotxn;			/* Free-on-transaction array */
 	size_t  fotxn_cnt;		/* Array entries */
 	size_t  fotxn_size;		/* Array size */
 
+	uint64_t split_gen;		/* Reading split generation */
+
 	/*
 	 * Hazard pointers.
 	 * The number of hazard pointers that can be in use grows dynamically.
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index d8d529a9d7c..3b793113e28 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -3210,9 +3210,9 @@ extern int wiredtiger_extension_terminate(WT_CONNECTION *connection);
 #define	WT_STAT_CONN_REC_PAGES_EVICTION			1081
 /*! reconciliation failed because an update could not be included */
 #define	WT_STAT_CONN_REC_SKIPPED_UPDATE			1082
-/*! split bytes awaiting free */
+/*! split bytes stashed */
 #define	WT_STAT_CONN_REC_SPLIT_STASHED_BYTES		1083
-/*! split objects awaiting free */
+/*! split objects stashed */
 #define	WT_STAT_CONN_REC_SPLIT_STASHED_OBJECTS		1084
 /*! pthread mutex shared lock read-lock calls */
 #define	WT_STAT_CONN_RWLOCK_READ			1085
diff --git a/src/session/session_misc.c b/src/session/session_misc.c
index 9aab0f8730e..c6f21fc9243 100644
--- a/src/session/session_misc.c
+++ b/src/session/session_misc.c
@@ -8,6 +8,29 @@
 #include "wt_internal.h"
 
 /*
+ * __wt_oldest_split_gen --
+ *	Calculate the oldest active split generation.
+ */
+uint64_t
+__wt_oldest_split_gen(WT_SESSION_IMPL *session)
+{
+	WT_CONNECTION_IMPL *conn;
+	WT_SESSION_IMPL *s;
+	uint64_t gen, oldest;
+	u_int i, session_cnt;
+
+	conn = S2C(session);
+	WT_ORDERED_READ(session_cnt, conn->session_cnt);
+	for (i = 0, s = conn->sessions, oldest = conn->split_gen + 1;
+	    i < session_cnt;
+	    i++, s++)
+		if (((gen = s->split_gen) != 0) && gen < oldest)
+			oldest = gen;
+
+	return (oldest);
+}
+
+/*
  * __wt_session_fotxn_add --
  *	Add a new entry into the session's free-on-transaction generation list.
  */
@@ -16,22 +39,14 @@ __wt_session_fotxn_add(WT_SESSION_IMPL *session, void *p, size_t len)
 {
 	WT_FOTXN *fotxn;
 
-	/*
-	 * Make sure the current thread has a transaction pinned so that
-	 * we don't immediately free the memory we are stashing.
-	 */
-	WT_ASSERT(session,
-	    WT_SESSION_TXN_STATE(session)->snap_min != WT_TXN_NONE);
+	WT_ASSERT(session, p != NULL);
 
 	/* Grow the list as necessary. */
 	WT_RET(__wt_realloc_def(session,
 	    &session->fotxn_size, session->fotxn_cnt + 1, &session->fotxn));
 
 	fotxn = session->fotxn + session->fotxn_cnt++;
-	fotxn->btree = S2BT(session);
-	fotxn->txnid = __wt_txn_current_id(session) + 1;
-	WT_ASSERT(session, !__wt_txn_visible_all(session, fotxn->txnid));
-	WT_ASSERT(session, fotxn->p == NULL);
+	fotxn->split_gen = WT_ATOMIC_ADD(S2C(session)->split_gen, 1);
 	fotxn->p = p;
 	fotxn->len = len;
 
@@ -53,26 +68,19 @@ __wt_session_fotxn_add(WT_SESSION_IMPL *session, void *p, size_t len)
 void
 __wt_session_fotxn_discard(WT_SESSION_IMPL *session)
 {
-	WT_BTREE *prev_btree;
 	WT_FOTXN *fotxn;
+	uint64_t oldest;
 	size_t i;
 
-	/* The last known tree that wasn't busy. */
-	prev_btree = NULL;
-
-	/* Bump the oldest transaction ID. */
-	__wt_txn_update_oldest(session);
+	/* Get the oldest split generation. */
+	oldest = __wt_oldest_split_gen(session);
 
 	for (i = 0, fotxn = session->fotxn;
 	    i < session->fotxn_cnt;
 	    ++i, ++fotxn) {
 		if (fotxn->p == NULL)
 			continue;
-		else if (fotxn->btree == prev_btree)
-			;
-		else if (__wt_btree_exclusive(session, fotxn->btree))
-			prev_btree = fotxn->btree;
-		else if (!__wt_txn_visible_all(session, fotxn->txnid))
+		else if (fotxn->split_gen >= oldest)
 			break;
 		/*
 		 * It's a bad thing if another thread is in this memory
@@ -90,12 +98,12 @@ __wt_session_fotxn_discard(WT_SESSION_IMPL *session)
 	 * If there are enough free slots at the beginning of the list, shuffle
 	 * everything down.
 	 */
-	if ((i > 100 || i == session->fotxn_cnt) &&
-	    (session->fotxn_cnt -= i) > 0) {
-		memmove(session->fotxn, session->fotxn + i,
-		    session->fotxn_cnt * sizeof(session->fotxn[0]));
+	if (i > 100 || i == session->fotxn_cnt) {
+		if ((session->fotxn_cnt -= i) > 0)
+			memmove(session->fotxn, fotxn,
+			    session->fotxn_cnt * sizeof(*fotxn));
 		memset(session->fotxn + session->fotxn_cnt, 0,
-		    i * sizeof(session->fotxn[0]));
+		    i * sizeof(*fotxn));
 	}
 }
 
diff --git a/src/support/stat.c b/src/support/stat.c
index ed22a214dd9..f1eb8236c0f 100644
--- a/src/support/stat.c
+++ b/src/support/stat.c
@@ -413,8 +413,8 @@ __wt_stat_init_connection_stats(WT_CONNECTION_STATS *stats)
 	    "page reconciliation calls for eviction";
 	stats->rec_skipped_update.desc =
 	    "reconciliation failed because an update could not be included";
-	stats->rec_split_stashed_bytes.desc = "split bytes awaiting free";
-	stats->rec_split_stashed_objects.desc = "split objects awaiting free";
+	stats->rec_split_stashed_bytes.desc = "split bytes stashed";
+	stats->rec_split_stashed_objects.desc = "split objects stashed";
 	stats->rwlock_read.desc = "pthread mutex shared lock read-lock calls";
 	stats->rwlock_write.desc =
 	    "pthread mutex shared lock write-lock calls";