Import wiredtiger: 4d5794b93795d52af97dc150d81b637442b89c5d from branch mongodb-3.6

ref: bc82f0f038..4d5794b937
for: 3.6.4

WT-1228       Improve performance of WT_SESSION::open_cursor
WT-3805       Avoid reading lookaside pages in truncate fast path
WT-3815       Cursor caching: measure and tune performance
WT-3829       WiredTiger metadata can be logically inconsistent.
WT-3848       Enhance new prepare transaction API to enforce post conditions
WT-3850       Implement WT_SESSSION::prepare_transaction
WT-3868       Bi-weekly WT codebase lint
WT-3901       Corruption of operation tracking log files
WT-3904       Reconsider error path in log server thread
WT-3905       Save the timestamp used for a checkpoint
WT-3906       Respect stable_timestamp in WT_CONNECTION::close
WT-3911       Ignore lookaside data on shutdown
WT-3912       fast-delete pages should re-instantiate the delete transaction's timestamp.
WT-3923       __wt_txn_context_prepare_check() requires API initialization
WT-3925       Fix test format operation selection code
WT-3926       Allow read_timestamp to be set after begin_transaction
WT-3927       Disable truncate testing with LSM
WT-3930       Set the recovery timestamp even if recovery doesn't run
WT-3932       WiredTiger memory allocation failure in js_test
WT-3933       test/format failure illegal WT_REF.state rolling back deleted page
WT-3934       LSM chunk checkpoints can race with system checkpoints
WT-3935       Enable cursor caching by default in WiredTiger.
WT-3936       Add multi-threaded tests for prepare_transaction()
WT-3938       Reduce memory usage with many tables and sessions
WT-3939       test_txn14.test_txn14.test_log_flush timeout
WT-3940       s_export issue detected during WiredTiger release execution
WT-3942       Update test_compact02 to handle being halted by eviction pressure.
WT-3945       Support libwiredtiger.so checking in s_export
WT-3946       Truncate segfault with a NULL start cursor
WT-3947       Allow wiredtiger_open configuration to disable cursor caching
WT-3948       Data handle loop may terminate without applying operations
WT-3949       Buffer overflow in WT_CURSOR::modify for string values
WT-3952       page-delete update list traversed after it has been discarded.
WT-3953       test/format can attempt to set an illegal prepare timestamp
WT-3954       test/format: prepared operations evicted before commit
WT-3961       The all_committed timestamp should be less than any in-flight transaction
WT-3964       Stop wrapping schema operations in a transaction
WT-3967       Fix long test for cursor cache sweep

1 files changed, 97 insertions, 1 deletions

diff --git a/src/third_party/wiredtiger/src/cache/cache_las.c b/src/third_party/wiredtiger/src/cache/cache_las.c
index 1f0b9c4b285..569a0247e7b 100644
--- a/src/third_party/wiredtiger/src/cache/cache_las.c
+++ b/src/third_party/wiredtiger/src/cache/cache_las.c
@@ -354,6 +354,102 @@ __wt_las_cursor_close(
 }
 
 /*
+ * __wt_las_page_skip_locked --
+ *	 Check if we can skip reading a page with lookaside entries, where
+ * the page is already locked.
+ */
+bool
+__wt_las_page_skip_locked(WT_SESSION_IMPL *session, WT_REF *ref)
+{
+	WT_TXN *txn;
+
+	txn = &session->txn;
+
+	/*
+	 * Skip lookaside pages if reading without a timestamp and all the
+	 * updates in lookaside are in the past.
+	 *
+	 * Lookaside eviction preferentially chooses the newest updates when
+	 * creating page images with no stable timestamp. If a stable timestamp
+	 * has been set, we have to visit the page because eviction chooses old
+	 * version of records in that case.
+	 *
+	 * One case where we may need to visit the page is if lookaside eviction
+	 * is active in tree 2 when a checkpoint has started and is working its
+	 * way through tree 1. In that case, lookaside may have created a page
+	 * image with updates in the future of the checkpoint.
+	 *
+	 * We also need to instantiate a lookaside page if this is an update
+	 * operation in progress.
+	 */
+	if (ref->page_las->invalid)
+		return (false);
+
+	if (F_ISSET(txn, WT_TXN_UPDATE))
+		return (false);
+
+	if (!F_ISSET(txn, WT_TXN_HAS_SNAPSHOT))
+		return (false);
+
+	if (WT_TXNID_LE(txn->snap_min, ref->page_las->las_max_txn))
+		return (false);
+
+	if (!F_ISSET(txn, WT_TXN_HAS_TS_READ) && ref->page_las->las_skew_newest)
+		return (true);
+
+#ifdef HAVE_TIMESTAMPS
+	/*
+	 * Skip lookaside pages if reading as of a timestamp, we evicted new
+	 * versions of data and all the updates are in the past.
+	 */
+	if (F_ISSET(&session->txn, WT_TXN_HAS_TS_READ) &&
+	    ref->page_las->las_skew_newest &&
+	    __wt_timestamp_cmp(
+	    &ref->page_las->onpage_timestamp, &session->txn.read_timestamp) < 0)
+		return (true);
+
+	/*
+	 * Skip lookaside pages if reading as of a timestamp, we evicted old
+	 * versions of data and all the updates are in the future.
+	 */
+	if (F_ISSET(&session->txn, WT_TXN_HAS_TS_READ) &&
+	    !ref->page_las->las_skew_newest &&
+	    __wt_timestamp_cmp(
+	    &ref->page_las->min_timestamp, &session->txn.read_timestamp) > 0)
+		return (true);
+#endif
+
+	return (false);
+}
+
+/*
+ * __wt_las_page_skip --
+ *	 Check if we can skip reading a page with lookaside entries, where the
+ * page needs to be locked before checking.
+ */
+bool
+__wt_las_page_skip(WT_SESSION_IMPL *session, WT_REF *ref)
+{
+	uint32_t previous_state;
+	bool skip;
+
+	if ((previous_state = ref->state) != WT_REF_LIMBO &&
+	    previous_state != WT_REF_LOOKASIDE)
+		return (false);
+
+	if (!__wt_atomic_casv32(&ref->state, previous_state, WT_REF_LOCKED))
+		return (false);
+
+	skip = __wt_las_page_skip_locked(session, ref);
+
+	/* Restore the state and push the change. */
+	ref->state = previous_state;
+	WT_FULL_BARRIER();
+
+	return (skip);
+}
+
+/*
  * __las_remove_block --
  *	Remove all records for a given page from the lookaside store.
  */
@@ -709,7 +805,7 @@ __wt_las_cursor_position(WT_CURSOR *cursor, uint32_t btree_id, uint64_t pageid)
 			 * Because of the special visibility rules for
 			 * lookaside, a new block can appear in between our
 			 * search and the block of interest.  Keep trying while
-			 * we have a key lower that we expect.
+			 * we have a key lower than we expect.
 			 *
 			 * There may be no block of lookaside entries if they
 			 * have been removed by