WT-2026 Allow forced eviction to split pages when BTREE_NO_EVICTION set (#2840)

Previously getting exclusive access for a checkpoint was blocking page splits, which allows pages to grow larger than the configured memory_page_max.

4 files changed, 26 insertions, 7 deletions

diff --git a/src/btree/bt_read.c b/src/btree/bt_read.c
index 94f3440d06e..3d396d5ae5b 100644
--- a/src/btree/bt_read.c
+++ b/src/btree/bt_read.c
@@ -548,10 +548,14 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 			 * if the page qualifies for forced eviction and update
 			 * the page's generation number. If eviction isn't being
 			 * done on this file, we're done.
+			 * In-memory split of large pages is allowed while
+			 * no_eviction is set on btree, whereas reconciliation
+			 * is not allowed.
 			 */
 			if (LF_ISSET(WT_READ_NO_EVICT) ||
 			    F_ISSET(session, WT_SESSION_NO_EVICTION) ||
-			    F_ISSET(btree, WT_BTREE_NO_EVICTION))
+			    (F_ISSET(btree, WT_BTREE_NO_EVICTION) &&
+			     !F_ISSET(btree, WT_BTREE_NO_RECONCILE)))
 				goto skip_evict;
 
 			/*
diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index c1a8ec3f73b..1371ed5ad49 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -128,7 +128,17 @@ __sync_file(WT_SESSION_IMPL *session, WT_CACHE_OP syncop)
 		 */
 		WT_PUBLISH(btree->checkpointing, WT_CKPT_PREPARE);
 
-		WT_ERR(__wt_evict_file_exclusive_on(session));
+		/*
+		 * Sync for checkpoint allows splits to happen while the queue
+		 * is being drained, but not reconciliation. We need to do this,
+		 * since draining the queue can take long enough for hot pages
+		 * to grow significantly larger than the configured maximum
+		 * size.
+		 */
+		F_SET(btree, WT_BTREE_NO_RECONCILE);
+		ret = __wt_evict_file_exclusive_on(session);
+		F_CLR(btree, WT_BTREE_NO_RECONCILE);
+		WT_ERR(ret);
 		__wt_evict_file_exclusive_off(session);
 
 		WT_PUBLISH(btree->checkpointing, WT_CKPT_RUNNING);
diff --git a/src/evict/evict_page.c b/src/evict/evict_page.c
index f5c900684a3..e661f27637e 100644
--- a/src/evict/evict_page.c
+++ b/src/evict/evict_page.c
@@ -467,6 +467,10 @@ __evict_review(
 		 */
 		if (LF_ISSET(WT_EVICT_INMEM_SPLIT))
 			return (__wt_split_insert(session, ref));
+
+		/* We are done if reconciliation is disabled. */
+		if (F_ISSET(S2BT(session), WT_BTREE_NO_RECONCILE))
+			return (EBUSY);
 	}
 
 	/* If the page is clean, we're done and we can evict. */
diff --git a/src/include/btree.h b/src/include/btree.h
index 96097115afd..a8b3cae3e55 100644
--- a/src/include/btree.h
+++ b/src/include/btree.h
@@ -156,11 +156,12 @@ struct __wt_btree {
 #define	WT_BTREE_NO_CHECKPOINT	0x00800	/* Disable checkpoints */
 #define	WT_BTREE_NO_EVICTION	0x01000	/* Disable eviction */
 #define	WT_BTREE_NO_LOGGING	0x02000	/* Disable logging */
-#define	WT_BTREE_REBALANCE	0x04000	/* Handle is for rebalance */
-#define	WT_BTREE_SALVAGE	0x08000	/* Handle is for salvage */
-#define	WT_BTREE_SKIP_CKPT	0x10000	/* Handle skipped checkpoint */
-#define	WT_BTREE_UPGRADE	0x20000	/* Handle is for upgrade */
-#define	WT_BTREE_VERIFY		0x40000	/* Handle is for verify */
+#define	WT_BTREE_NO_RECONCILE	0x04000 /* Allow splits, even with no evict */
+#define	WT_BTREE_REBALANCE	0x08000	/* Handle is for rebalance */
+#define	WT_BTREE_SALVAGE	0x10000	/* Handle is for salvage */
+#define	WT_BTREE_SKIP_CKPT	0x20000	/* Handle skipped checkpoint */
+#define	WT_BTREE_UPGRADE	0x40000	/* Handle is for upgrade */
+#define	WT_BTREE_VERIFY		0x80000	/* Handle is for verify */
 	uint32_t flags;
 };