WT-2725 Enforce visibility constraints before evicting a page. (#2875)

Otherwise there is a race between when we add a page to the queue
and when the page is pulled off the queue - where the updates on
the page may have changed, thus violating transactional constraints.

4 files changed, 20 insertions, 24 deletions

diff --git a/src/btree/bt_split.c b/src/btree/bt_split.c
index 39c10133542..28fcb60e95d 100644
--- a/src/btree/bt_split.c
+++ b/src/btree/bt_split.c
@@ -511,7 +511,8 @@ __split_deepen(WT_SESSION_IMPL *session, WT_PAGE *parent)
 		 * array, a thread might see a freed WT_REF.  Set the eviction
 		 * transaction requirement for the newly created internal pages.
 		 */
-		child->modify->mod_split_txn = __wt_txn_id_alloc(session, false);
+		child->modify->mod_split_txn =
+		    __wt_txn_id_alloc(session, false);
 
 		/*
 		 * The newly allocated child's page index references the same
@@ -865,7 +866,7 @@ __split_parent_lock(
 		}
 		/*
 		 * If a checkpoint is running and we fail to lock the parent
-		 * page, give up immmediately to avoid deadlock.
+		 * page, give up immediately to avoid deadlock.
 		 */
 		if (S2BT(session)->checkpointing)
 			return (EBUSY);
diff --git a/src/evict/evict_lru.c b/src/evict/evict_lru.c
index dc64a55091d..42966d636f6 100644
--- a/src/evict/evict_lru.c
+++ b/src/evict/evict_lru.c
@@ -1277,25 +1277,6 @@ fast:		/* If the page can't be evicted, give up. */
 		    !__wt_txn_visible_all(session, mod->rec_max_txn))
 			continue;
 
-		/*
-		 * If the oldest transaction hasn't changed since the last time
-		 * this page was written, it's unlikely that we can make
-		 * progress.  Similarly, if the most recent update on the page
-		 * is not yet globally visible, eviction will fail.  These
-		 * heuristics attempt to avoid repeated attempts to evict the
-		 * same page.
-		 *
-		 * That said, if eviction is stuck, or we are helping with
-		 * forced eviction, try anyway: maybe a transaction that was
-		 * running last time we wrote the page has since rolled back,
-		 * or we can help get the checkpoint completed sooner.
-		 */
-		if (modified && !LF_ISSET(
-		    WT_EVICT_PASS_AGGRESSIVE | WT_EVICT_PASS_WOULD_BLOCK) &&
-		    (mod->disk_snap_min == S2C(session)->txn_global.oldest_id ||
-		    !__wt_txn_visible_all(session, mod->update_txn)))
-			continue;
-
 		WT_ASSERT(session, evict->ref == NULL);
 		__evict_init_candidate(session, evict, ref);
 		++evict;
diff --git a/src/include/btree.i b/src/include/btree.i
index 6dff3156d09..8086df92087 100644
--- a/src/include/btree.i
+++ b/src/include/btree.i
@@ -1073,6 +1073,7 @@ __wt_page_can_evict(WT_SESSION_IMPL *session,
 	WT_BTREE *btree;
 	WT_PAGE_MODIFY *mod;
 	WT_TXN_GLOBAL *txn_global;
+	bool modified;
 
 	if (inmem_splitp != NULL)
 		*inmem_splitp = false;
@@ -1109,6 +1110,8 @@ __wt_page_can_evict(WT_SESSION_IMPL *session,
 	    !__wt_txn_visible_all(session, mod->mod_split_txn))
 		return (false);
 
+	modified = __wt_page_is_modified(page);
+
 	/*
 	 * If the file is being checkpointed, we can't evict dirty pages:
 	 * if we write a page and free the previous version of the page, that
@@ -1116,8 +1119,7 @@ __wt_page_can_evict(WT_SESSION_IMPL *session,
 	 * been written in the checkpoint, leaving the checkpoint inconsistent.
 	 */
 	if (btree->checkpointing &&
-	    (__wt_page_is_modified(page) ||
-	    mod->rec_result == WT_PM_REC_MULTIBLOCK)) {
+	    (modified || mod->rec_result == WT_PM_REC_MULTIBLOCK)) {
 		WT_STAT_FAST_CONN_INCR(session, cache_eviction_checkpoint);
 		WT_STAT_FAST_DATA_INCR(session, cache_eviction_checkpoint);
 		return (false);
@@ -1133,6 +1135,18 @@ __wt_page_can_evict(WT_SESSION_IMPL *session,
 			return (false);
 	}
 
+	/*
+	 * If the oldest transaction hasn't changed since the last time this
+	 * page was written, it's unlikely that we can make progress.
+	 * Similarly, if the most recent update on the page is not yet globally
+	 * visible, eviction must fail.  These checks also aim to avoid
+	 * repeated attempts to evict the same page.
+	 */
+	if (modified &&
+	    (mod->disk_snap_min == __wt_txn_oldest_id(session) ||
+	    !__wt_txn_visible_all(session, mod->update_txn)))
+		return (false);
+
 	return (true);
 }
 
diff --git a/src/meta/meta_track.c b/src/meta/meta_track.c
index b8d7aa930ea..6c8296a9bdd 100644
--- a/src/meta/meta_track.c
+++ b/src/meta/meta_track.c
@@ -487,7 +487,7 @@ __wt_meta_track_handle_lock(WT_SESSION_IMPL *session, bool created)
 
 /*
  * __wt_meta_track_init --
- *	Intialize metadata tracking.
+ *	Initialize metadata tracking.
  */
 int
 __wt_meta_track_init(WT_SESSION_IMPL *session)