Merge branch 'develop' into statistics-sources-handle-overwrite

7 files changed, 51 insertions, 34 deletions

diff --git a/src/btree/bt_compact.c b/src/btree/bt_compact.c
index d8b3a638de3..1528d65b8c8 100644
--- a/src/btree/bt_compact.c
+++ b/src/btree/bt_compact.c
@@ -76,7 +76,7 @@ __wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
 	WT_CONNECTION_IMPL *conn;
 	WT_DECL_RET;
 	WT_REF *ref;
-	int block_manager_begin, skip;
+	int block_manager_begin, evict_reset, skip;
 
 	WT_UNUSED(cfg);
 
@@ -133,8 +133,9 @@ __wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
 	 * then let eviction continue;
 	 */
 	conn->compact_in_memory_pass = 1;
-	WT_ERR(__wt_evict_file_exclusive_on(session));
-	__wt_evict_file_exclusive_off(session);
+	WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
+	if (evict_reset)
+		__wt_evict_file_exclusive_off(session);
 
 	/* Start compaction. */
 	WT_ERR(bm->compact_start(bm, session));
diff --git a/src/btree/bt_page.c b/src/btree/bt_page.c
index 2f2ce4cf4f7..b5140beb792 100644
--- a/src/btree/bt_page.c
+++ b/src/btree/bt_page.c
@@ -129,15 +129,24 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 			    __evict_force_check(session, page, flags)) {
 				++force_attempts;
 				ret = __wt_page_release_evict(session, ref);
+				/* If forced eviction fails, stall. */
 				if (ret == EBUSY) {
-					/* If forced eviction fails, stall. */
 					ret = 0;
 					wait_cnt += 1000;
+					WT_STAT_FAST_CONN_INCR(session,
+					    page_forcible_evict_blocked);
+					break;
 				} else
 					WT_RET(ret);
-				WT_STAT_FAST_CONN_INCR(
-				    session, page_forcible_evict_blocked);
-				break;
+
+				/*
+				 * The result of a successful forced eviction
+				 * is a page-state transition (potentially to
+				 * an in-memory page we can use, or a restart
+				 * return for our caller), continue the outer
+				 * page-acquisition loop.
+				 */
+				continue;
 			}
 
 			/* Check if we need an autocommit transaction. */
diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index a75af03d8c8..d925eefc2fe 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -25,6 +25,7 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 	uint64_t internal_bytes, leaf_bytes;
 	uint64_t internal_pages, leaf_pages;
 	uint32_t flags;
+	int evict_reset;
 
 	btree = S2BT(session);
 
@@ -99,11 +100,11 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 		 * eviction to complete.
 		 */
 		btree->checkpointing = 1;
+		WT_FULL_BARRIER();
 
-		if (!F_ISSET(btree, WT_BTREE_NO_EVICTION)) {
-			WT_ERR(__wt_evict_file_exclusive_on(session));
+		WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
+		if (evict_reset)
 			__wt_evict_file_exclusive_off(session);
-		}
 
 		/* Write all dirty in-cache pages. */
 		flags |= WT_READ_NO_EVICT;
diff --git a/src/evict/evict_file.c b/src/evict/evict_file.c
index 910aef070ca..9e39fcc7a2c 100644
--- a/src/evict/evict_file.c
+++ b/src/evict/evict_file.c
@@ -15,21 +15,16 @@
 int
 __wt_evict_file(WT_SESSION_IMPL *session, int syncop)
 {
-	WT_BTREE *btree;
 	WT_DECL_RET;
 	WT_PAGE *page;
 	WT_REF *next_ref, *ref;
-	int eviction_enabled;
-
-	btree = S2BT(session);
-	eviction_enabled = !F_ISSET(btree, WT_BTREE_NO_EVICTION);
+	int evict_reset;
 
 	/*
 	 * We need exclusive access to the file -- disable ordinary eviction
 	 * and drain any blocks already queued.
 	 */
-	if (eviction_enabled)
-		WT_RET(__wt_evict_file_exclusive_on(session));
+	WT_RET(__wt_evict_file_exclusive_on(session, &evict_reset));
 
 	/* Make sure the oldest transaction ID is up-to-date. */
 	__wt_txn_update_oldest(session);
@@ -140,7 +135,7 @@ err:		/* On error, clear any left-over tree walk. */
 			    session, next_ref, WT_READ_NO_EVICT));
 	}
 
-	if (eviction_enabled)
+	if (evict_reset)
 		__wt_evict_file_exclusive_off(session);
 
 	return (ret);
diff --git a/src/evict/evict_lru.c b/src/evict/evict_lru.c
index 83a9aa5c8c5..78c34f0267e 100644
--- a/src/evict/evict_lru.c
+++ b/src/evict/evict_lru.c
@@ -653,7 +653,7 @@ __wt_evict_page(WT_SESSION_IMPL *session, WT_REF *ref)
  *	blocks queued for eviction.
  */
 int
-__wt_evict_file_exclusive_on(WT_SESSION_IMPL *session)
+__wt_evict_file_exclusive_on(WT_SESSION_IMPL *session, int *evict_resetp)
 {
 	WT_BTREE *btree;
 	WT_CACHE *cache;
@@ -664,6 +664,15 @@ __wt_evict_file_exclusive_on(WT_SESSION_IMPL *session)
 	cache = S2C(session)->cache;
 
 	/*
+	 * If the file isn't evictable, there's no work to do.
+	 */
+	if (F_ISSET(btree, WT_BTREE_NO_EVICTION)) {
+		*evict_resetp = 0;
+		return (0);
+	}
+	*evict_resetp = 1;
+
+	/*
 	 * Hold the walk lock to set the "no eviction" flag: no new pages from
 	 * the file will be queued for eviction after this point.
 	 */
@@ -1072,34 +1081,37 @@ __evict_walk_file(WT_SESSION_IMPL *session, u_int *slotp, uint32_t flags)
 	WT_PAGE_MODIFY *mod;
 	uint64_t pages_walked;
 	uint32_t walk_flags;
-	int internal_pages, modified, restarts;
+	int enough, internal_pages, modified, restarts;
 
 	btree = S2BT(session);
 	cache = S2C(session)->cache;
 	start = cache->evict + *slotp;
 	end = WT_MIN(start + WT_EVICT_WALK_PER_FILE,
 	    cache->evict + cache->evict_slots);
+	enough = internal_pages = restarts = 0;
 
 	walk_flags =
 	    WT_READ_CACHE | WT_READ_NO_EVICT | WT_READ_NO_GEN | WT_READ_NO_WAIT;
 
 	/*
 	 * Get some more eviction candidate pages.
+	 *
+	 * !!! Take care terminating this loop.
+	 *
+	 * Don't make an extra call to __wt_tree_walk after we hit the end of a
+	 * tree: that will leave a page pinned, which may prevent any work from
+	 * being done.
+	 *
+	 * Once we hit the page limit, do one more step through the walk in
+	 * case we are appending and only the last page in the file is live.
 	 */
-	for (evict = start, pages_walked = 0, internal_pages = restarts = 0;
-	    evict < end && pages_walked < WT_EVICT_MAX_PER_FILE &&
-	    (ret == 0 || ret == WT_NOTFOUND);
+	for (evict = start, pages_walked = 0;
+	    evict < end && !enough && (ret == 0 || ret == WT_NOTFOUND);
 	    ret = __wt_tree_walk(
 	    session, &btree->evict_ref, &pages_walked, walk_flags)) {
+		enough = (pages_walked > WT_EVICT_MAX_PER_FILE);
 		if (btree->evict_ref == NULL) {
-			/*
-			 * Take care with terminating this loop.
-			 *
-			 * Don't make an extra call to __wt_tree_walk: that will
-			 * leave a page pinned, which may prevent any work from
-			 * being done.
-			 */
-			if (++restarts == 2)
+			if (++restarts == 2 || enough)
 				break;
 			continue;
 		}
diff --git a/src/include/extern.h b/src/include/extern.h
index 23bb36623e5..5d3ee5bc8f8 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -297,7 +297,7 @@ extern int __wt_evict_server_wake(WT_SESSION_IMPL *session);
 extern int __wt_evict_create(WT_SESSION_IMPL *session);
 extern int __wt_evict_destroy(WT_SESSION_IMPL *session);
 extern int __wt_evict_page(WT_SESSION_IMPL *session, WT_REF *ref);
-extern int __wt_evict_file_exclusive_on(WT_SESSION_IMPL *session);
+extern int __wt_evict_file_exclusive_on(WT_SESSION_IMPL *session, int *evict_resetp);
 extern void __wt_evict_file_exclusive_off(WT_SESSION_IMPL *session);
 extern int __wt_evict_lru_page(WT_SESSION_IMPL *session, int is_server);
 extern int __wt_cache_wait(WT_SESSION_IMPL *session, int full);
diff --git a/src/log/log_slot.c b/src/log/log_slot.c
index ff61afb698c..8dcb2f9f165 100644
--- a/src/log/log_slot.c
+++ b/src/log/log_slot.c
@@ -203,8 +203,7 @@ retry:
 		 * churn is used to change how long we pause before closing
 		 * the slot - which leads to more consolidation and less churn.
 		 */
-		if (++switch_fails % SLOT_POOL == 0 &&
-		    switch_fails != 0 && slot->slot_churn < 5)
+		if (++switch_fails % SLOT_POOL == 0 && slot->slot_churn < 5)
 			++slot->slot_churn;
 		__wt_yield();
 		goto retry;