Merge pull request #2429 from wiredtiger/WT-2193-backportmongodb-3.1.9mongodb-3.1.8mongodb-3.1.7mongodb-3.1.6mongodb-3.1.5mongodb-3.1.4mongodb-3.1.3mongodb-3.1.2mongodb-3.1.1mongodb-3.1.0mongodb-3.0.9mongodb-3.0.0

 WT-2193 Backport 3.0. Handle read-committed metadata checkpoints during snapshot transactions

5 files changed, 95 insertions, 6 deletions

diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index 939d8f808b1..5d5ad491145 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -22,16 +22,17 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 	WT_PAGE_MODIFY *mod;
 	WT_REF *walk;
 	WT_TXN *txn;
-	uint64_t internal_bytes, leaf_bytes;
-	uint64_t internal_pages, leaf_pages;
+	uint64_t internal_bytes, internal_pages, leaf_bytes, leaf_pages;
+	uint64_t saved_snap_min;
 	uint32_t flags;
 	bool evict_reset;
 
 	btree = S2BT(session);
 
-	flags = WT_READ_CACHE | WT_READ_NO_GEN;
 	walk = NULL;
 	txn = &session->txn;
+	saved_snap_min = WT_SESSION_TXN_STATE(session)->snap_min;
+	flags = WT_READ_CACHE | WT_READ_NO_GEN;
 
 	internal_bytes = leaf_bytes = 0;
 	internal_pages = leaf_pages = 0;
@@ -80,6 +81,19 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 		break;
 	case WT_SYNC_CHECKPOINT:
 		/*
+		 * If we are flushing a file at read-committed isolation, which
+		 * is of particular interest for flushing the metadata to make
+		 * schema-changing operation durable, get a transactional
+		 * snapshot now.
+		 *
+		 * All changes committed up to this point should be included.
+		 * We don't update the snapshot in between pages because (a)
+		 * the metadata shouldn't be that big, and (b) if we do ever
+		 */
+		if (txn->isolation == WT_ISO_READ_COMMITTED)
+			__wt_txn_get_snapshot(session);
+
+		/*
 		 * We cannot check the tree modified flag in the case of a
 		 * checkpoint, the checkpoint code has already cleared it.
 		 *
@@ -185,7 +199,12 @@ err:	/* On error, clear any left-over tree walk. */
 	if (walk != NULL)
 		WT_TRET(__wt_page_release(session, walk, flags));
 
-	if (txn->isolation == WT_ISO_READ_COMMITTED && session->ncursors == 0)
+	/*
+	 * If we got a snapshot in order to write pages, and there was no
+	 * snapshot active when we started, release it.
+	 */
+	if (txn->isolation == WT_ISO_READ_COMMITTED &&
+	    saved_snap_min == WT_TXN_NONE)
 		__wt_txn_release_snapshot(session);
 
 	if (btree->checkpointing) {
diff --git a/src/conn/conn_open.c b/src/conn/conn_open.c
index c86a6e039eb..826226a1975 100644
--- a/src/conn/conn_open.c
+++ b/src/conn/conn_open.c
@@ -122,6 +122,9 @@ __wt_connection_close(WT_CONNECTION_IMPL *conn)
 	/* Close open data handles. */
 	WT_TRET(__wt_conn_dhandle_discard(session));
 
+	/* Shut down metadata tracking, required before creating tables. */
+	WT_TRET(__wt_meta_track_destroy(session));
+
 	/*
 	 * Now that all data handles are closed, tell logging that a checkpoint
 	 * has completed then shut down the log manager (only after closing
@@ -254,6 +257,9 @@ __wt_connection_workers(WT_SESSION_IMPL *session, const char *cfg[])
 	 */
 	WT_RET(__wt_logmgr_open(session));
 
+	/* Initialize metadata tracking, required before creating tables. */
+	WT_RET(__wt_meta_track_init(session));
+
 	/* Start the optional checkpoint thread. */
 	WT_RET(__wt_checkpoint_server_create(session, cfg));
 
diff --git a/src/include/connection.h b/src/include/connection.h
index 6870fcb1640..691cca436dc 100644
--- a/src/include/connection.h
+++ b/src/include/connection.h
@@ -334,6 +334,8 @@ struct __wt_connection_impl {
 	uint32_t	 log_prealloc;	/* Log file pre-allocation */
 	uint32_t	 txn_logsync;	/* Log sync configuration */
 
+	WT_SESSION_IMPL *meta_ckpt_session;/* Metadata checkpoint session */
+
 	WT_SESSION_IMPL *sweep_session;	/* Handle sweep session */
 	wt_thread_t	 sweep_tid;	/* Handle sweep thread */
 	int		 sweep_tid_set;	/* Handle sweep thread set */
diff --git a/src/include/extern.h b/src/include/extern.h
index d30bb916e12..00f97c3d818 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -443,6 +443,8 @@ extern int __wt_meta_track_update(WT_SESSION_IMPL *session, const char *key);
 extern int __wt_meta_track_fileop( WT_SESSION_IMPL *session, const char *olduri, const char *newuri);
 extern int __wt_meta_track_drop( WT_SESSION_IMPL *session, const char *filename);
 extern int __wt_meta_track_handle_lock(WT_SESSION_IMPL *session, bool created);
+extern int __wt_meta_track_init(WT_SESSION_IMPL *session);
+extern int __wt_meta_track_destroy(WT_SESSION_IMPL *session);
 extern int __wt_turtle_init(WT_SESSION_IMPL *session);
 extern int __wt_turtle_read(WT_SESSION_IMPL *session, const char *key, char **valuep);
 extern int __wt_turtle_update( WT_SESSION_IMPL *session, const char *key, const char *value);
diff --git a/src/meta/meta_track.c b/src/meta/meta_track.c
index ff1173585b4..b8d7aa930ea 100644
--- a/src/meta/meta_track.c
+++ b/src/meta/meta_track.c
@@ -261,6 +261,7 @@ __wt_meta_track_off(WT_SESSION_IMPL *session, bool need_sync, bool unroll)
 {
 	WT_DECL_RET;
 	WT_META_TRACK *trk, *trk_orig;
+	WT_SESSION_IMPL *ckpt_session;
 
 	WT_ASSERT(session,
 	    WT_META_TRACKING(session) && session->meta_track_nest > 0);
@@ -304,8 +305,18 @@ __wt_meta_track_off(WT_SESSION_IMPL *session, bool need_sync, bool unroll)
 			session, false, WT_TXN_LOG_CKPT_SYNC, NULL));
 		WT_RET(ret);
 	} else {
-		WT_WITH_DHANDLE(session, session->meta_dhandle,
-		    ret = __wt_checkpoint(session, NULL));
+		WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
+		ckpt_session = S2C(session)->meta_ckpt_session;
+		/*
+		 * If this operation is part of a running transaction, that
+		 * should be included in the checkpoint.
+		 */
+		ckpt_session->txn.id = session->txn.id;
+		F_SET(ckpt_session, WT_SESSION_LOCKED_SCHEMA);
+		WT_WITH_DHANDLE(ckpt_session, session->meta_dhandle, ret =
+		    __wt_checkpoint(ckpt_session, NULL));
+		F_CLR(ckpt_session, WT_SESSION_LOCKED_SCHEMA);
+		ckpt_session->txn.id = WT_TXN_NONE;
 		WT_RET(ret);
 		WT_WITH_DHANDLE(session, session->meta_dhandle,
 		    ret = __wt_checkpoint_sync(session, NULL));
@@ -473,3 +484,52 @@ __wt_meta_track_handle_lock(WT_SESSION_IMPL *session, bool created)
 	trk->created = created;
 	return (0);
 }
+
+/*
+ * __wt_meta_track_init --
+ *	Intialize metadata tracking.
+ */
+int
+__wt_meta_track_init(WT_SESSION_IMPL *session)
+{
+	WT_CONNECTION_IMPL *conn;
+
+	conn = S2C(session);
+	if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED)) {
+		WT_RET(__wt_open_internal_session(conn,
+		    "metadata-ckpt", false, false,
+		    &conn->meta_ckpt_session));
+
+		/*
+		 * Sessions default to read-committed isolation, we rely on
+		 * that for the correctness of metadata checkpoints.
+		 */
+		WT_ASSERT(session, conn->meta_ckpt_session->txn.isolation ==
+		    WT_ISO_READ_COMMITTED);
+	}
+
+	return (0);
+}
+
+/*
+ * __wt_meta_track_destroy --
+ *	Release resources allocated for metadata tracking.
+ */
+int
+__wt_meta_track_destroy(WT_SESSION_IMPL *session)
+{
+	WT_CONNECTION_IMPL *conn;
+	WT_DECL_RET;
+	WT_SESSION *wt_session;
+
+	conn = S2C(session);
+
+	/* Close the session used for metadata checkpoints. */
+	if (conn->meta_ckpt_session != NULL) {
+		wt_session = &conn->meta_ckpt_session->iface;
+		WT_TRET(wt_session->close(wt_session, NULL));
+		conn->meta_ckpt_session = NULL;
+	}
+
+	return (ret);
+}