Change the sweep server to only operate on clean files. Track the maximum transaction ID seen in the checkpoint of a file so that we can be sure in sweep that all pages can be discarded (without dirtying anything in the tree).

Preparation work for SERVER-17587

10 files changed, 58 insertions, 37 deletions

diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index bc5d1051b1e..a6ad86c888f 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -117,8 +117,12 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 			mod = page->modify;
 
 			/* Skip clean pages. */
-			if (!__wt_page_is_modified(page))
+			if (!__wt_page_is_modified(page)) {
+				if (mod != NULL && TXNID_LT(
+				    btree->rec_max_txn, mod->rec_max_txn))
+					btree->rec_max_txn = mod->rec_max_txn;
 				continue;
+			}
 
 			/*
 			 * Write dirty pages, unless we can be sure they only
@@ -153,6 +157,9 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 				++leaf_pages;
 			}
 			WT_ERR(__wt_reconcile(session, walk, NULL, 0));
+
+			if (TXNID_LT(btree->rec_max_txn, mod->rec_max_txn))
+				btree->rec_max_txn = mod->rec_max_txn;
 		}
 		break;
 	}
diff --git a/src/conn/conn_dhandle.c b/src/conn/conn_dhandle.c
index 7756158594c..e28f18a6fa5 100644
--- a/src/conn/conn_dhandle.c
+++ b/src/conn/conn_dhandle.c
@@ -234,7 +234,7 @@ err:	WT_TRET(__wt_rwlock_destroy(session, &dhandle->rwlock));
  *	Sync and close the underlying btree handle.
  */
 int
-__wt_conn_btree_sync_and_close(WT_SESSION_IMPL *session, int force)
+__wt_conn_btree_sync_and_close(WT_SESSION_IMPL *session, int final, int force)
 {
 	WT_BTREE *btree;
 	WT_DATA_HANDLE *dhandle;
@@ -273,7 +273,7 @@ __wt_conn_btree_sync_and_close(WT_SESSION_IMPL *session, int force)
 	 */
 	if (!F_ISSET(btree,
 	    WT_BTREE_SALVAGE | WT_BTREE_UPGRADE | WT_BTREE_VERIFY))
-		WT_ERR(__wt_checkpoint_close(session, force));
+		WT_ERR(__wt_checkpoint_close(session, final, force));
 
 	if (dhandle->checkpoint == NULL)
 		--S2C(session)->open_btree_count;
@@ -392,7 +392,7 @@ __conn_btree_open(
 	 * in the tree that can block the close.
 	 */
 	if (F_ISSET(dhandle, WT_DHANDLE_OPEN))
-		WT_RET(__wt_conn_btree_sync_and_close(session, 0));
+		WT_RET(__wt_conn_btree_sync_and_close(session, 0, 0));
 
 	/* Discard any previous configuration, set up the new configuration. */
 	__conn_btree_config_clear(session);
@@ -424,7 +424,7 @@ __conn_btree_open(
 err:		F_CLR(btree, WT_BTREE_SPECIAL_FLAGS);
 		/* If the open failed, close the handle. */
 		if (F_ISSET(dhandle, WT_DHANDLE_OPEN))
-			WT_TRET(__wt_conn_btree_sync_and_close(session, 0));
+			WT_TRET(__wt_conn_btree_sync_and_close(session, 0, 0));
 	}
 
 	return (ret);
@@ -670,7 +670,7 @@ __wt_conn_dhandle_close_all(
 		if (F_ISSET(dhandle, WT_DHANDLE_OPEN)) {
 			if ((ret = __wt_meta_track_sub_on(session)) == 0)
 				ret = __wt_conn_btree_sync_and_close(
-				    session, force);
+				    session, 0, force);
 
 			/*
 			 * If the close succeeded, drop any locks it acquired.
@@ -732,7 +732,7 @@ __wt_conn_dhandle_discard_single(WT_SESSION_IMPL *session, int final)
 	dhandle = session->dhandle;
 
 	if (F_ISSET(dhandle, WT_DHANDLE_OPEN)) {
-		tret = __wt_conn_btree_sync_and_close(session, 0);
+		tret = __wt_conn_btree_sync_and_close(session, final, 0);
 		if (final && tret != 0) {
 			__wt_err(session, tret,
 			    "Final close of %s failed", dhandle->name);
diff --git a/src/conn/conn_sweep.c b/src/conn/conn_sweep.c
index 50812ccfa95..d788060ab17 100644
--- a/src/conn/conn_sweep.c
+++ b/src/conn/conn_sweep.c
@@ -15,6 +15,7 @@
 static int
 __sweep(WT_SESSION_IMPL *session)
 {
+	WT_BTREE *btree;
 	WT_CONNECTION_IMPL *conn;
 	WT_DATA_HANDLE *dhandle, *dhandle_next;
 	WT_DECL_RET;
@@ -63,10 +64,16 @@ __sweep(WT_SESSION_IMPL *session)
 		WT_RET(ret);
 		locked = 1;
 
+		/* Only sweep clean trees where all updates are visible. */
+		btree = dhandle->handle;
+		if (btree->modified ||
+		    !__wt_txn_visible_all(session, btree->rec_max_txn))
+			goto unlock;
+
 		/* If the handle is open, try to close it. */
 		if (F_ISSET(dhandle, WT_DHANDLE_OPEN)) {
-			WT_WITH_DHANDLE(session, dhandle,
-			    ret = __wt_conn_btree_sync_and_close(session, 0));
+			WT_WITH_DHANDLE(session, dhandle, ret =
+			    __wt_conn_btree_sync_and_close(session, 0, 0));
 			if (ret != 0)
 				goto unlock;
 
diff --git a/src/evict/evict_file.c b/src/evict/evict_file.c
index f546a5adae1..864c116a380 100644
--- a/src/evict/evict_file.c
+++ b/src/evict/evict_file.c
@@ -72,23 +72,17 @@ __wt_evict_file(WT_SESSION_IMPL *session, int syncop)
 		    WT_READ_CACHE | WT_READ_NO_EVICT));
 
 		switch (syncop) {
-		case WT_SYNC_DISCARD:
-			/*
-			 * Check that the page is clean: if we see a dirty page
-			 * (including a dirty parent page after evicting a
-			 * child), give up.  The higher level can try to
-			 * checkpoint, but during discard we aren't set up to
-			 * manage checkpoints.
-			 */
-			if (__wt_page_is_modified(page))
-				WT_ERR(EBUSY);
-			/* FALLTHROUGH */
 		case WT_SYNC_CLOSE:
 			/*
 			 * Evict the page.
 			 */
 			WT_ERR(__wt_evict(session, ref, 1));
 			break;
+		case WT_SYNC_DISCARD:
+			WT_ASSERT(session,
+			    __wt_page_can_evict(session, page, 0));
+			__wt_evict_page_clean_update(session, ref);
+			break;
 		case WT_SYNC_DISCARD_FORCE:
 			/*
 			 * Forced discard of the page, whether clean or dirty.
diff --git a/src/evict/evict_lru.c b/src/evict/evict_lru.c
index 2ebd699c579..2b5bd015223 100644
--- a/src/evict/evict_lru.c
+++ b/src/evict/evict_lru.c
@@ -1211,7 +1211,7 @@ __evict_walk_file(WT_SESSION_IMPL *session, u_int *slotp, uint32_t flags)
 		}
 
 fast:		/* If the page can't be evicted, give up. */
-		if (!__wt_page_can_evict(session, page, 0))
+		if (!__wt_page_can_evict(session, page, 1))
 			continue;
 
 		/*
diff --git a/src/include/btree.h b/src/include/btree.h
index f00a7ac9a8e..cc571124207 100644
--- a/src/include/btree.h
+++ b/src/include/btree.h
@@ -124,6 +124,7 @@ struct __wt_btree {
 	u_int	 block_header;		/* WT_PAGE_HEADER_BYTE_SIZE */
 
 	uint64_t checkpoint_gen;	/* Checkpoint generation */
+	uint64_t rec_max_txn;		/* Maximum txn seen (clean trees) */
 	uint64_t write_gen;		/* Write generation */
 
 	WT_REF  *evict_ref;		/* Eviction thread's location */
diff --git a/src/include/btree.i b/src/include/btree.i
index 3b18cb530e9..6c261f3768d 100644
--- a/src/include/btree.i
+++ b/src/include/btree.i
@@ -970,7 +970,7 @@ __wt_page_can_evict(WT_SESSION_IMPL *session, WT_PAGE *page, int check_splits)
 	 * a transaction value, once that's globally visible, we know we can
 	 * evict the created page.
 	 */
-	if (WT_PAGE_IS_INTERNAL(page) &&
+	if (check_splits && WT_PAGE_IS_INTERNAL(page) &&
 	    !__wt_txn_visible_all(session, mod->mod_split_txn))
 		return (0);
 
@@ -1013,7 +1013,7 @@ __wt_page_can_evict(WT_SESSION_IMPL *session, WT_PAGE *page, int check_splits)
 
 	/*
 	 * If the page was recently split in-memory, don't force it out: we
-	 * hope eviction will find it first.
+	 * hope an eviction thread will find it first.
 	 */
 	if (check_splits &&
 	    !__wt_txn_visible_all(session, mod->inmem_split_txn))
diff --git a/src/include/extern.h b/src/include/extern.h
index e9c37e62bb6..6ac926b494c 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -225,7 +225,7 @@ extern int __wt_checkpoint_server_create(WT_SESSION_IMPL *session, const char *c
 extern int __wt_checkpoint_server_destroy(WT_SESSION_IMPL *session);
 extern int __wt_checkpoint_signal(WT_SESSION_IMPL *session, wt_off_t logsize);
 extern int __wt_conn_dhandle_find(WT_SESSION_IMPL *session, const char *name, const char *ckpt, uint32_t flags);
-extern int __wt_conn_btree_sync_and_close(WT_SESSION_IMPL *session, int force);
+extern int __wt_conn_btree_sync_and_close(WT_SESSION_IMPL *session, int final, int force);
 extern int __wt_conn_btree_get(WT_SESSION_IMPL *session, const char *name, const char *ckpt, const char *cfg[], uint32_t flags);
 extern int __wt_conn_btree_apply(WT_SESSION_IMPL *session, int apply_checkpoints, const char *uri, int (*func)(WT_SESSION_IMPL *, const char *[]), const char *cfg[]);
 extern int __wt_conn_btree_apply_single_ckpt(WT_SESSION_IMPL *session, const char *uri, int (*func)(WT_SESSION_IMPL *, const char *[]), const char *cfg[]);
@@ -672,7 +672,7 @@ extern int __wt_checkpoint_list(WT_SESSION_IMPL *session, const char *cfg[]);
 extern int __wt_txn_checkpoint(WT_SESSION_IMPL *session, const char *cfg[]);
 extern int __wt_checkpoint(WT_SESSION_IMPL *session, const char *cfg[]);
 extern int __wt_checkpoint_sync(WT_SESSION_IMPL *session, const char *cfg[]);
-extern int __wt_checkpoint_close(WT_SESSION_IMPL *session, int force);
+extern int __wt_checkpoint_close(WT_SESSION_IMPL *session, int final, int force);
 extern uint64_t __wt_ext_transaction_id(WT_EXTENSION_API *wt_api, WT_SESSION *wt_session);
 extern int __wt_ext_transaction_isolation_level( WT_EXTENSION_API *wt_api, WT_SESSION *wt_session);
 extern int __wt_ext_transaction_notify( WT_EXTENSION_API *wt_api, WT_SESSION *wt_session, WT_TXN_NOTIFY *notify);
diff --git a/src/session/session_dhandle.c b/src/session/session_dhandle.c
index 833d098efeb..0825f783ca3 100644
--- a/src/session/session_dhandle.c
+++ b/src/session/session_dhandle.c
@@ -168,7 +168,7 @@ __wt_session_release_btree(WT_SESSION_IMPL *session)
 		WT_ASSERT(session, F_ISSET(dhandle, WT_DHANDLE_EXCLUSIVE));
 		F_CLR(dhandle, WT_DHANDLE_DISCARD);
 
-		WT_TRET(__wt_conn_btree_sync_and_close(session, 0));
+		WT_TRET(__wt_conn_btree_sync_and_close(session, 0, 0));
 	}
 
 	if (F_ISSET(dhandle, WT_DHANDLE_EXCLUSIVE))
diff --git a/src/txn/txn_ckpt.c b/src/txn/txn_ckpt.c
index fa77d2b5fa5..c910b1cbc90 100644
--- a/src/txn/txn_ckpt.c
+++ b/src/txn/txn_ckpt.c
@@ -1090,27 +1090,39 @@ __wt_checkpoint_sync(WT_SESSION_IMPL *session, const char *cfg[])
  *	Checkpoint a single file as part of closing the handle.
  */
 int
-__wt_checkpoint_close(WT_SESSION_IMPL *session, int force)
+__wt_checkpoint_close(WT_SESSION_IMPL *session, int final, int force)
 {
-	WT_DECL_RET;
+	WT_BTREE *btree;
+
+	btree = S2BT(session);
 
 	/* Handle forced discard (when dropping a file). */
 	if (force)
 		return (__wt_cache_op(session, NULL, WT_SYNC_DISCARD_FORCE));
 
-	/* If closing an unmodified file, try to evict its pages. */
-	if (!S2BT(session)->modified) {
-		ret = __wt_cache_op(session, NULL, WT_SYNC_DISCARD);
-		if (ret != EBUSY)
-			return (ret);
+	/*
+	 * If closing an unmodified file, check that no update is required
+	 * for active readers.
+	 */
+	if (!btree->modified && !F_ISSET(btree, WT_BTREE_BULK)) {
+		__wt_txn_update_oldest(session);
+		return (__wt_txn_visible_all(session, btree->rec_max_txn) ?
+		    __wt_cache_op(session, NULL, WT_SYNC_DISCARD) : EBUSY);
 	}
 
 	/*
-	 * If closing a modified file, or closing an unmodified file was blocked
-	 * for any reason, checkpoint the file and optionally flush the writes
-	 * (the checkpoint call will discard the blocks, there's no additional
-	 * step needed).
+	 * If closing a modified file, checkpoint the file and optionally flush
+	 * the writes (the checkpoint call will discard the blocks, there's no
+	 * additional step needed).
+	 *
+	 * We should already have the schema lock unless we're finishing a bulk
+	 * load -- the only other paths to closing files (sweep and LSM) have
+	 * already checked for read-only trees.
 	 */
+	if (!final)
+		WT_ASSERT(session, F_ISSET(session, WT_SESSION_SCHEMA_LOCKED) ||
+		    F_ISSET(btree, WT_BTREE_BULK));
+
 	WT_RET(__checkpoint_worker(session, NULL, 0));
 	if (F_ISSET(S2C(session), WT_CONN_CKPT_SYNC))
 		WT_RET(__wt_checkpoint_sync(session, NULL));