Import wiredtiger-wiredtiger-mongodb-3.2.0-rc1-210-gf95877a.tar.gz from wiredtiger branch mongodb-3.2

13 files changed, 138 insertions, 33 deletions

diff --git a/src/third_party/wiredtiger/src/block/block_open.c b/src/third_party/wiredtiger/src/block/block_open.c
index e1f3e6de0fe..e97e86d9aaa 100644
--- a/src/third_party/wiredtiger/src/block/block_open.c
+++ b/src/third_party/wiredtiger/src/block/block_open.c
@@ -422,9 +422,13 @@ int
 __wt_block_manager_size(
     WT_SESSION_IMPL *session, const char *filename, WT_DSRC_STATS *stats)
 {
+	WT_DECL_RET;
 	wt_off_t filesize;
 
-	WT_RET(__wt_filesize_name(session, filename, &filesize));
+	ret = __wt_filesize_name(session, filename, &filesize);
+	if (ret != 0)
+		WT_RET_MSG(session, ret, "%s: file size", filename);
+
 	stats->block_size = filesize;
 
 	return (0);
diff --git a/src/third_party/wiredtiger/src/btree/bt_split.c b/src/third_party/wiredtiger/src/btree/bt_split.c
index 2145d6ac014..c0c739d68ad 100644
--- a/src/third_party/wiredtiger/src/btree/bt_split.c
+++ b/src/third_party/wiredtiger/src/btree/bt_split.c
@@ -1031,10 +1031,13 @@ __split_parent(WT_SESSION_IMPL *session, WT_REF *ref,
 
 	/*
 	 * If the entire (sub)tree is empty, give up: we can't leave an empty
-	 * internal page.
+	 * internal page.  Mark it to be evicted soon and clean up any
+	 * references that have changed state.
 	 */
-	if (result_entries == 0)
-		return (0);
+	if (result_entries == 0) {
+		__wt_page_evict_soon(parent);
+		goto err;
+	}
 
 	/*
 	 * Allocate and initialize a new page index array for the parent, then
@@ -1226,6 +1229,12 @@ err:	/*
 				next_ref->state = WT_REF_DELETED;
 		}
 
+		/* If we gave up on a reverse split, unlock the child. */
+		if (ref_new == NULL) {
+			WT_ASSERT(session, ref->state == WT_REF_LOCKED);
+			ref->state = WT_REF_DELETED;
+		}
+
 		__wt_free_ref_index(session, NULL, alloc_index, false);
 	}
 
diff --git a/src/third_party/wiredtiger/src/btree/bt_sync.c b/src/third_party/wiredtiger/src/btree/bt_sync.c
index 237d900c3d1..110f9f8fb0c 100644
--- a/src/third_party/wiredtiger/src/btree/bt_sync.c
+++ b/src/third_party/wiredtiger/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.
 		 *
@@ -174,7 +188,12 @@ 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 != WT_CKPT_OFF) {
diff --git a/src/third_party/wiredtiger/src/conn/conn_open.c b/src/third_party/wiredtiger/src/conn/conn_open.c
index 04815c8e152..4fe1db1c524 100644
--- a/src/third_party/wiredtiger/src/conn/conn_open.c
+++ b/src/third_party/wiredtiger/src/conn/conn_open.c
@@ -126,6 +126,9 @@ __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
@@ -252,6 +255,9 @@ __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));
+
 	/* Create the lookaside table. */
 	WT_RET(__wt_las_create(session));
 
diff --git a/src/third_party/wiredtiger/src/evict/evict_lru.c b/src/third_party/wiredtiger/src/evict/evict_lru.c
index c28b89b81ce..18335d6fb5e 100644
--- a/src/third_party/wiredtiger/src/evict/evict_lru.c
+++ b/src/third_party/wiredtiger/src/evict/evict_lru.c
@@ -159,7 +159,7 @@ __evict_server(void *arg)
 	WT_DECL_RET;
 	WT_SESSION_IMPL *session;
 #ifdef HAVE_DIAGNOSTIC
-	struct timespec now, stuck_ts;
+	struct timespec now, stuck_ts = { 0, 0 };
 #endif
 	u_int spins;
 
diff --git a/src/third_party/wiredtiger/src/evict/evict_page.c b/src/third_party/wiredtiger/src/evict/evict_page.c
index 2c0312a6f60..e49098e90db 100644
--- a/src/third_party/wiredtiger/src/evict/evict_page.c
+++ b/src/third_party/wiredtiger/src/evict/evict_page.c
@@ -394,7 +394,9 @@ __evict_review(
 	/*
 	 * Retrieve the modified state of the page. This must happen after the
 	 * check for evictable internal pages otherwise there is a race where a
-	 * page could be marked modified whilst performing the check.
+	 * page could be marked modified due to a child being transitioned to
+	 * WT_REF_DISK after the modified check and before we visited the ref
+	 * while walking the parent index.
 	 */
 	modified = __wt_page_is_modified(page);
 
diff --git a/src/third_party/wiredtiger/src/include/connection.h b/src/third_party/wiredtiger/src/include/connection.h
index 03b8174b7e1..a585e08ef1a 100644
--- a/src/third_party/wiredtiger/src/include/connection.h
+++ b/src/third_party/wiredtiger/src/include/connection.h
@@ -363,6 +363,8 @@ 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/third_party/wiredtiger/src/include/extern.h b/src/third_party/wiredtiger/src/include/extern.h
index 3dd479acc0a..4c934f95c2b 100644
--- a/src/third_party/wiredtiger/src/include/extern.h
+++ b/src/third_party/wiredtiger/src/include/extern.h
@@ -456,6 +456,8 @@ 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/third_party/wiredtiger/src/lsm/lsm_tree.c b/src/third_party/wiredtiger/src/lsm/lsm_tree.c
index bd54202de5e..7371f3a407c 100644
--- a/src/third_party/wiredtiger/src/lsm/lsm_tree.c
+++ b/src/third_party/wiredtiger/src/lsm/lsm_tree.c
@@ -213,6 +213,7 @@ int
 __wt_lsm_tree_set_chunk_size(
     WT_SESSION_IMPL *session, WT_LSM_CHUNK *chunk)
 {
+	WT_DECL_RET;
 	wt_off_t size;
 	const char *filename;
 
@@ -220,7 +221,9 @@ __wt_lsm_tree_set_chunk_size(
 	if (!WT_PREFIX_SKIP(filename, "file:"))
 		WT_RET_MSG(session, EINVAL,
 		    "Expected a 'file:' URI: %s", chunk->uri);
-	WT_RET(__wt_filesize_name(session, filename, &size));
+	ret = __wt_filesize_name(session, filename, &size);
+	if (ret != 0)
+		WT_RET_MSG(session, ret, "%s: file size", filename);
 
 	chunk->size = (uint64_t)size;
 
diff --git a/src/third_party/wiredtiger/src/meta/meta_track.c b/src/third_party/wiredtiger/src/meta/meta_track.c
index bc96a35efc7..1a773ef67c4 100644
--- a/src/third_party/wiredtiger/src/meta/meta_track.c
+++ b/src/third_party/wiredtiger/src/meta/meta_track.c
@@ -231,6 +231,7 @@ __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);
@@ -275,9 +276,18 @@ __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,
-		    WT_WITH_TXN_ISOLATION(session, WT_ISO_READ_COMMITTED,
-			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));
@@ -458,3 +468,52 @@ __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, WT_SESSION_NO_DATA_HANDLES,
+		    &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);
+}
diff --git a/src/third_party/wiredtiger/src/os_posix/os_filesize.c b/src/third_party/wiredtiger/src/os_posix/os_filesize.c
index b01fc91514b..09174ffcd90 100644
--- a/src/third_party/wiredtiger/src/os_posix/os_filesize.c
+++ b/src/third_party/wiredtiger/src/os_posix/os_filesize.c
@@ -47,10 +47,9 @@ __wt_filesize_name(
 
 	__wt_free(session, path);
 
-	if (ret == 0) {
+	if (ret == 0)
 		*sizep = sb.st_size;
-		return (0);
-	}
 
-	WT_RET_MSG(session, ret, "%s: fstat", filename);
+	/* Some callers expect failure, so don't log an error message. */
+	return (ret);
 }
diff --git a/src/third_party/wiredtiger/src/os_win/os_filesize.c b/src/third_party/wiredtiger/src/os_win/os_filesize.c
index dfeadc31fc4..c51303d7215 100644
--- a/src/third_party/wiredtiger/src/os_win/os_filesize.c
+++ b/src/third_party/wiredtiger/src/os_win/os_filesize.c
@@ -47,11 +47,10 @@ __wt_filesize_name(
 
 	__wt_free(session, path);
 
-	if (ret != 0) {
+	if (ret != 0)
 		*sizep =
 		    ((int64_t)data.nFileSizeHigh << 32) | data.nFileSizeLow;
-		return (0);
-	}
 
-	WT_RET_MSG(session, __wt_errno(), "%s: GetFileAttributesEx", filename);
+	/* Some callers expect failure, so don't log an error message. */
+	return (ret);
 }
diff --git a/src/third_party/wiredtiger/src/schema/schema_stat.c b/src/third_party/wiredtiger/src/schema/schema_stat.c
index d14b81d389f..355abf8f343 100644
--- a/src/third_party/wiredtiger/src/schema/schema_stat.c
+++ b/src/third_party/wiredtiger/src/schema/schema_stat.c
@@ -90,18 +90,19 @@ __curstat_size_only(WT_SESSION_IMPL *session,
 	WT_ERR(__wt_buf_fmt(
 	    session, &namebuf, "%s.wt", uri + strlen("table:")));
 	/*
-	 * Get the size of the underlying file.  There is nothing stopping a
-	 * race with schema level table operations (for example drop) if there
-	 * is a race there will be an error message generated.
+	 * Get the size of the underlying file. This will fail for anything
+	 * other than simple tables (LSM for example) and will fail if there
+	 * are concurrent schema level operations (for example drop). That is
+	 * fine - failing here results in falling back to the slow path of
+	 * opening the handle.
 	 */
-	WT_ERR(__wt_filesize_name(session, namebuf.data, &filesize));
-
-	/* Setup and populate the statistics structure */
-	__wt_stat_dsrc_init_single(&cst->u.dsrc_stats);
-	cst->u.dsrc_stats.block_size = filesize;
-	__wt_curstat_dsrc_final(cst);
-
-	*was_fast = true;
+	if (__wt_filesize_name(session, namebuf.data, &filesize) == 0) {
+		/* Setup and populate the statistics structure */
+		__wt_stat_dsrc_init_single(&cst->u.dsrc_stats);
+		cst->u.dsrc_stats.block_size = filesize;
+		__wt_curstat_dsrc_final(cst);
+		*was_fast = true;
+	}
 
 err:	__wt_free(session, tableconf);
 	__wt_buf_free(session, &namebuf);