WT-3012 Don't track the LSM Primary as part of dirty bytes in cache (#3136)

4 files changed, 61 insertions, 16 deletions

diff --git a/src/include/btree.h b/src/include/btree.h
index 713d46ae85f..0e0f7b4c40a 100644
--- a/src/include/btree.h
+++ b/src/include/btree.h
@@ -158,15 +158,16 @@ struct __wt_btree {
 #define	WT_BTREE_IGNORE_CACHE	0x000200 /* Cache-resident object */
 #define	WT_BTREE_IN_MEMORY	0x000400 /* Cache-resident object */
 #define	WT_BTREE_LOOKASIDE	0x000800 /* Look-aside table */
-#define	WT_BTREE_NO_CHECKPOINT	0x001000 /* Disable checkpoints */
-#define	WT_BTREE_NO_EVICTION	0x002000 /* Disable eviction */
-#define	WT_BTREE_NO_LOGGING	0x004000 /* Disable logging */
-#define	WT_BTREE_NO_RECONCILE	0x008000 /* Allow splits, even with no evict */
-#define	WT_BTREE_REBALANCE	0x010000 /* Handle is for rebalance */
-#define	WT_BTREE_SALVAGE	0x020000 /* Handle is for salvage */
-#define	WT_BTREE_SKIP_CKPT	0x040000 /* Handle skipped checkpoint */
-#define	WT_BTREE_UPGRADE	0x080000 /* Handle is for upgrade */
-#define	WT_BTREE_VERIFY		0x100000 /* Handle is for verify */
+#define	WT_BTREE_LSM_PRIMARY	0x001000 /* Handle is current LSM primary */
+#define	WT_BTREE_NO_CHECKPOINT	0x002000 /* Disable checkpoints */
+#define	WT_BTREE_NO_EVICTION	0x004000 /* Disable eviction */
+#define	WT_BTREE_NO_LOGGING	0x008000 /* Disable logging */
+#define	WT_BTREE_NO_RECONCILE	0x010000 /* Allow splits, even with no evict */
+#define	WT_BTREE_REBALANCE	0x020000 /* Handle is for rebalance */
+#define	WT_BTREE_SALVAGE	0x040000 /* Handle is for salvage */
+#define	WT_BTREE_SKIP_CKPT	0x080000 /* Handle skipped checkpoint */
+#define	WT_BTREE_UPGRADE	0x100000 /* Handle is for upgrade */
+#define	WT_BTREE_VERIFY		0x200000 /* Handle is for verify */
 	uint32_t flags;
 };
 
diff --git a/src/include/btree.i b/src/include/btree.i
index e48189d50ea..6fe2a101f48 100644
--- a/src/include/btree.i
+++ b/src/include/btree.i
@@ -107,7 +107,7 @@ __wt_cache_page_inmem_incr(WT_SESSION_IMPL *session, WT_PAGE *page, size_t size)
 		(void)__wt_atomic_addsize(&page->modify->bytes_dirty, size);
 		if (WT_PAGE_IS_INTERNAL(page))
 			(void)__wt_atomic_add64(&cache->bytes_dirty_intl, size);
-		else {
+		else if (!F_ISSET(btree, WT_BTREE_LSM_PRIMARY)) {
 			(void)__wt_atomic_add64(&cache->bytes_dirty_leaf, size);
 			(void)__wt_atomic_add64(&btree->bytes_dirty_leaf, size);
 		}
@@ -241,7 +241,7 @@ __wt_cache_page_byte_dirty_decr(
 	if (WT_PAGE_IS_INTERNAL(page))
 		__wt_cache_decr_check_uint64(session, &cache->bytes_dirty_intl,
 		    decr, "WT_CACHE.bytes_dirty_intl");
-	else {
+	else if (!F_ISSET(btree, WT_BTREE_LSM_PRIMARY)) {
 		__wt_cache_decr_check_uint64(session, &btree->bytes_dirty_leaf,
 		    decr, "WT_BTREE.bytes_dirty_leaf");
 		__wt_cache_decr_check_uint64(session, &cache->bytes_dirty_leaf,
@@ -300,8 +300,10 @@ __wt_cache_dirty_incr(WT_SESSION_IMPL *session, WT_PAGE *page)
 		(void)__wt_atomic_add64(&cache->bytes_dirty_intl, size);
 		(void)__wt_atomic_add64(&cache->pages_dirty_intl, 1);
 	} else {
-		(void)__wt_atomic_add64(&btree->bytes_dirty_leaf, size);
-		(void)__wt_atomic_add64(&cache->bytes_dirty_leaf, size);
+		if (!F_ISSET(btree, WT_BTREE_LSM_PRIMARY)) {
+			(void)__wt_atomic_add64(&btree->bytes_dirty_leaf, size);
+			(void)__wt_atomic_add64(&cache->bytes_dirty_leaf, size);
+		}
 		(void)__wt_atomic_add64(&cache->pages_dirty_leaf, 1);
 	}
 	(void)__wt_atomic_addsize(&page->modify->bytes_dirty, size);
@@ -394,7 +396,7 @@ __wt_cache_page_evict(WT_SESSION_IMPL *session, WT_PAGE *page)
 			__wt_cache_decr_zero_uint64(session,
 			    &cache->bytes_dirty_intl,
 			    modify->bytes_dirty, "WT_CACHE.bytes_dirty_intl");
-		else {
+		else if (!F_ISSET(btree, WT_BTREE_LSM_PRIMARY)) {
 			__wt_cache_decr_zero_uint64(session,
 			    &cache->bytes_dirty_leaf,
 			    modify->bytes_dirty, "WT_CACHE.bytes_dirty_leaf");
@@ -1543,6 +1545,48 @@ __wt_btree_lsm_over_size(WT_SESSION_IMPL *session, uint64_t maxsize)
 }
 
 /*
+ * __wt_btree_lsm_switch_primary --
+ *      Switch a btree handle to/from the current primary chunk of an LSM tree.
+ */
+static inline void
+__wt_btree_lsm_switch_primary(WT_SESSION_IMPL *session, bool on)
+{
+	WT_BTREE *btree;
+	WT_CACHE *cache;
+	WT_PAGE *child, *root;
+	WT_PAGE_INDEX *pindex;
+	WT_REF *first;
+	size_t size;
+
+	btree = S2BT(session);
+	cache = S2C(session)->cache;
+	root = btree->root.page;
+
+       if (on) {
+	       F_SET(btree, WT_BTREE_LSM_PRIMARY);
+	       F_SET(btree, WT_BTREE_NO_EVICTION);
+       } else {
+		pindex = WT_INTL_INDEX_GET_SAFE(root);
+		first = pindex->index[0];
+
+	/*
+	 * We're reaching down into the page without a hazard pointer, but
+	 * that's OK because we know that no-eviction is set and so the page
+	 * cannot disappear.
+	 */
+		child = first->page;
+		if (!__wt_page_is_modified(child))
+			return;
+		size = child->modify->bytes_dirty;
+		(void)__wt_atomic_add64(&btree->bytes_dirty_leaf, size);
+		(void)__wt_atomic_add64(&cache->bytes_dirty_leaf, size);
+
+		F_CLR(btree, WT_BTREE_LSM_PRIMARY);
+		F_CLR(btree, WT_BTREE_NO_EVICTION);
+	}
+}
+
+/*
  * __wt_split_descent_race --
  *	Return if we raced with an internal page split when descending the tree.
  */
diff --git a/src/lsm/lsm_cursor.c b/src/lsm/lsm_cursor.c
index 698fed77e77..c40f0f64b33 100644
--- a/src/lsm/lsm_cursor.c
+++ b/src/lsm/lsm_cursor.c
@@ -700,7 +700,7 @@ retry:	if (F_ISSET(clsm, WT_CLSM_MERGE)) {
 		if (btree->bulk_load_ok) {
 			btree->bulk_load_ok = false;
 			WT_WITH_BTREE(session, btree,
-			    __wt_btree_evictable(session, false));
+			    __wt_btree_lsm_switch_primary(session, true));
 		}
 	}
 
diff --git a/src/lsm/lsm_work_unit.c b/src/lsm/lsm_work_unit.c
index 917104031fc..dfb8690f1ed 100644
--- a/src/lsm/lsm_work_unit.c
+++ b/src/lsm/lsm_work_unit.c
@@ -383,7 +383,7 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
 	 * forced eviction.
 	 */
 	WT_ERR(__wt_session_get_btree(session, chunk->uri, NULL, NULL, 0));
-	__wt_btree_evictable(session, true);
+	__wt_btree_lsm_switch_primary(session, false);
 	WT_ERR(__wt_session_release_btree(session));
 
 	/* Make sure we aren't pinning a transaction ID. */