WT-2346 Protect LSM flushes from racing with checkpoints.

Internal flush operations used to rely on the schema lock to avoid
racing with checkpoints.  That is no longer sufficient, use the
checkpoint lock instead but also switch on "no wait" mode to avoid
blocking for long periods while checkpoints are running.

1 files changed, 17 insertions, 4 deletions

diff --git a/src/lsm/lsm_work_unit.c b/src/lsm/lsm_work_unit.c
index 4faa25967ad..dbaf8602846 100644
--- a/src/lsm/lsm_work_unit.c
+++ b/src/lsm/lsm_work_unit.c
@@ -334,14 +334,27 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
 	/*
 	 * Turn on metadata tracking to ensure the checkpoint gets the
 	 * necessary handle locks.
+	 *
+	 * Ensure that we don't race with a running checkpoint: the checkpoint
+	 * lock protects against us racing with an application checkpoint in
+	 * this chunk.  Don't wait for it, though: checkpoints can take a long
+	 * time, and our checkpoint operation should be very quick.
 	 */
 	WT_ERR(__wt_meta_track_on(session));
-	WT_WITH_SCHEMA_LOCK(session, ret,
-	    ret = __wt_schema_worker(
-	    session, chunk->uri, __wt_checkpoint, NULL, NULL, 0));
+	F_SET(session, WT_SESSION_LOCK_NO_WAIT);
+	WT_WITH_CHECKPOINT_LOCK(session, ret,
+	    WT_WITH_SCHEMA_LOCK(session, ret,
+		ret = __wt_schema_worker(
+		session, chunk->uri, __wt_checkpoint, NULL, NULL, 0)));
 	WT_TRET(__wt_meta_track_off(session, false, ret != 0));
-	if (ret != 0)
+	F_CLR(session, WT_SESSION_LOCK_NO_WAIT);
+	if (ret != 0) {
+		if (ret == EBUSY) {
+			ret = 0;
+			goto err;
+		}
 		WT_ERR_MSG(session, ret, "LSM checkpoint");
+	}
 
 	/* Now the file is written, get the chunk size. */
 	WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));