Fix __wt_conn_btree_apply to acquire handles correctly,

pass WT_BTREE_SNAPSHOT_OP from checkpoint code.

6 files changed, 40 insertions, 22 deletions

diff --git a/src/btree/bt_handle.c b/src/btree/bt_handle.c
index 3cc126b1c89..5b9a1fef052 100644
--- a/src/btree/bt_handle.c
+++ b/src/btree/bt_handle.c
@@ -133,8 +133,10 @@ __wt_btree_close(WT_SESSION_IMPL *session)
 	__wt_btree_huffman_close(session);
 
 	/* Snapshot lock. */
-	if (btree->snaplock != NULL)
+	if (btree->snaplock != NULL) {
 		(void)__wt_rwlock_destroy(session, btree->snaplock);
+		btree->snaplock = NULL;
+	}
 
 	/* Free allocated memory. */
 	__wt_free(session, btree->key_format);
diff --git a/src/conn/conn_btree.c b/src/conn/conn_btree.c
index 90331c2adca..b68448be4f6 100644
--- a/src/conn/conn_btree.c
+++ b/src/conn/conn_btree.c
@@ -300,8 +300,15 @@ __wt_conn_btree_apply(WT_SESSION_IMPL *session,
 
 	__wt_spin_lock(session, &conn->spinlock);
 	TAILQ_FOREACH(btree, &conn->btqh, q)
-		if (btree->snapshot == NULL &&
+		if (F_ISSET(btree, WT_BTREE_OPEN) &&
+		    btree->snapshot == NULL &&
 		    strcmp(btree->name, WT_METADATA_URI) != 0) {
+			/*
+			 * We have the connection spinlock, which prevents
+			 * handles being opened or closed, so there is no need
+			 * for additional handle locking here, or pulling every
+			 * tree into this session's handle cache.
+			 */
 			session->btree = btree;
 			WT_ERR(func(session, cfg));
 		}
diff --git a/src/include/txn.i b/src/include/txn.i
index 37395341750..da9f5af9716 100644
--- a/src/include/txn.i
+++ b/src/include/txn.i
@@ -125,7 +125,7 @@ __wt_txn_update_check(WT_SESSION_IMPL *session, WT_UPDATE *upd)
 	WT_TXN *txn;
 
 	txn = &session->txn;
-	if (txn->isolation == TXN_ISO_SNAPSHOT)
+	if (F_ISSET(txn, TXN_RUNNING) && txn->isolation == TXN_ISO_SNAPSHOT)
 		while (upd != NULL && !__wt_txn_visible(session, upd->txnid)) {
 			if (upd->txnid != WT_TXN_ABORTED)
 				return (WT_DEADLOCK);
diff --git a/src/txn/txn.c b/src/txn/txn.c
index 36149170239..8244a093cf0 100644
--- a/src/txn/txn.c
+++ b/src/txn/txn.c
@@ -193,15 +193,11 @@ int
 __wt_txn_checkpoint(WT_SESSION_IMPL *session, const char *cfg[])
 {
 	WT_CONFIG_ITEM cval;
-	WT_CURSOR *cursor;
 	WT_DECL_RET;
 	WT_TXN_GLOBAL *txn_global;
 	const char *snapshot;
 	const char *txn_cfg[] = { "isolation=snapshot", NULL };
 
-	cursor = NULL;
-	txn_global = &S2C(session)->txn_global;
-
 	if ((ret = __wt_config_gets(
 	    session, cfg, "snapshot", &cval)) != 0 && ret != WT_NOTFOUND)
 		WT_RET(ret);
@@ -212,12 +208,14 @@ __wt_txn_checkpoint(WT_SESSION_IMPL *session, const char *cfg[])
 
 	/* Only one checkpoint can be active at a time. */
 	__wt_writelock(session, S2C(session)->ckpt_rwlock);
-
 	WT_ERR(__wt_txn_begin(session, txn_cfg));
 
 	/* Prevent eviction from evicting anything newer than this. */
+	txn_global = &S2C(session)->txn_global;
 	txn_global->ckpt_txnid = session->txn.snap_min;
 
+	WT_ERR(__wt_meta_track_on(session));
+
 	/*
 	 * If we're doing an ordinary unnamed checkpoint, we only need to flush
 	 * open files.	If we're creating a named snapshot, we need to walk the
@@ -225,16 +223,27 @@ __wt_txn_checkpoint(WT_SESSION_IMPL *session, const char *cfg[])
 	 */
 	WT_TRET((snapshot == NULL) ?
 	    __wt_conn_btree_apply(session, __wt_snapshot, cfg) :
-	    __wt_meta_btree_apply(session, __wt_snapshot, cfg, 0));
+	    __wt_meta_btree_apply(session,
+		__wt_snapshot, cfg, WT_BTREE_SNAPSHOT_OP));
 
-	if (cursor != NULL)
-		WT_TRET(cursor->close(cursor));
-
-	txn_global->ckpt_txnid = WT_TXN_NONE;
-
-	WT_TRET(__txn_release(session));
+	/*
+	 * XXX Rolling back the changes here is problematic.
+	 *
+	 * If we unroll here, we need a way to roll back changes to the avail
+	 * list for each tree that was successfully synced before the error
+	 * occurred.  Otherwise, the next time we try this operation, we will
+	 * try to free an old snapshot again.
+	 *
+	 * OTOH, if we commit the changes after a failure, we have partially
+	 * overwritten the checkpoint, so what ends up on disk is not
+	 * consistent.
+	 */
+	WT_TRET(__wt_meta_track_off(session, ret != 0));
 
-err:	__wt_rwunlock(session, S2C(session)->ckpt_rwlock);
+err:	txn_global->ckpt_txnid = WT_TXN_NONE;
+	if (F_ISSET(&session->txn, TXN_RUNNING))
+		WT_TRET(__txn_release(session));
+	__wt_rwunlock(session, S2C(session)->ckpt_rwlock);
 	__wt_free(session, snapshot);
 	return (ret);
 }
diff --git a/test/format/wts_ops.c b/test/format/wts_ops.c
index 0ec003d186a..7e28f029ab7 100644
--- a/test/format/wts_ops.c
+++ b/test/format/wts_ops.c
@@ -171,10 +171,10 @@ ops(void *arg)
 					    WT_TABLENAME, sync_name);
 				sync_drop = 0;
 			} else {
-				if ((ret = session->sync(
-				    session, WT_TABLENAME, sync_name)) != 0)
-					die(ret, "session.sync: %s: %s",
-					    WT_TABLENAME, sync_name);
+				if ((ret = session->checkpoint(
+				    session, sync_name)) != 0)
+					die(ret, "session.checkpoint: %s",
+					    sync_name);
 				sync_drop = 1;
 			}
 
diff --git a/test/thread/t.c b/test/thread/t.c
index 554c624073b..0f40e0424e4 100644
--- a/test/thread/t.c
+++ b/test/thread/t.c
@@ -163,8 +163,8 @@ wt_shutdown(void)
 	if ((ret = session->verify(session, FNAME, NULL)) != 0)
 		die("session.verify", ret);
 
-	if ((ret = session->sync(session, FNAME, NULL)) != 0)
-		die("session.sync", ret);
+	if ((ret = session->checkpoint(session, NULL)) != 0)
+		die("session.checkpoint", ret);
 
 	if ((ret = conn->close(conn, NULL)) != 0)
 		die("conn.close", ret);