From 0dae18bfa7ecba8820f032a0a53efa6cc02e4b0c Mon Sep 17 00:00:00 2001
From: Matthew Russotto <matthew.russotto@mongodb.com>
Date: Wed, 18 May 2022 14:36:27 -0400
Subject: SERVER-65671 Use _termShadow in updateTerm and processReplSetMetadata

---
 src/mongo/db/repl/replication_coordinator_impl.cpp | 26 +++++++++++++++++-----
 src/mongo/db/repl/replication_coordinator_impl.h   |  7 ++++++
 2 files changed, 27 insertions(+), 6 deletions(-)

(limited to 'src/mongo/db/repl')

diff --git a/src/mongo/db/repl/replication_coordinator_impl.cpp b/src/mongo/db/repl/replication_coordinator_impl.cpp
index ca7b9e692c7..5c1e3eb8778 100644
--- a/src/mongo/db/repl/replication_coordinator_impl.cpp
+++ b/src/mongo/db/repl/replication_coordinator_impl.cpp
@@ -3369,13 +3369,15 @@ void ReplicationCoordinatorImpl::processReplSetGetConfig(BSONObjBuilder* result,
 void ReplicationCoordinatorImpl::processReplSetMetadata(const rpc::ReplSetMetadata& replMetadata) {
     EventHandle evh;
 
-    {
-        stdx::lock_guard<Latch> lock(_mutex);
-        evh = _processReplSetMetadata_inlock(replMetadata);
-    }
+    if (_needToUpdateTerm(replMetadata.getTerm())) {
+        {
+            stdx::lock_guard<Latch> lock(_mutex);
+            evh = _processReplSetMetadata_inlock(replMetadata);
+        }
 
-    if (evh) {
-        _replExecutor->waitForEvent(evh);
+        if (evh) {
+            _replExecutor->waitForEvent(evh);
+        }
     }
 }
 
@@ -3386,6 +3388,9 @@ void ReplicationCoordinatorImpl::cancelAndRescheduleElectionTimeout() {
 
 EventHandle ReplicationCoordinatorImpl::_processReplSetMetadata_inlock(
     const rpc::ReplSetMetadata& replMetadata) {
+    // Note that public method processReplSetMetadata() above depends on this method not needing
+    // to do anything when the term is up to date.  If that changes, be sure to update that
+    // method as well.
     return _updateTerm_inlock(replMetadata.getTerm());
 }
 
@@ -5902,6 +5907,11 @@ Status ReplicationCoordinatorImpl::updateTerm(OperationContext* opCtx, long long
 
     // Check we haven't acquired any lock, because potential stepdown needs global lock.
     dassert(!opCtx->lockState()->isLocked() || opCtx->lockState()->isNoop());
+
+    // If the term is already up to date, we can skip the update and the mutex acquisition.
+    if (!_needToUpdateTerm(term))
+        return Status::OK();
+
     TopologyCoordinator::UpdateTermResult updateTermResult;
     EventHandle finishEvh;
 
@@ -5923,6 +5933,10 @@ Status ReplicationCoordinatorImpl::updateTerm(OperationContext* opCtx, long long
     return Status::OK();
 }
 
+bool ReplicationCoordinatorImpl::_needToUpdateTerm(long long term) {
+    return term > _termShadow.load();
+}
+
 EventHandle ReplicationCoordinatorImpl::_updateTerm_inlock(
     long long term, TopologyCoordinator::UpdateTermResult* updateTermResult) {
 
diff --git a/src/mongo/db/repl/replication_coordinator_impl.h b/src/mongo/db/repl/replication_coordinator_impl.h
index 7c6a70868e0..a6dc8fe9066 100644
--- a/src/mongo/db/repl/replication_coordinator_impl.h
+++ b/src/mongo/db/repl/replication_coordinator_impl.h
@@ -1434,6 +1434,13 @@ private:
      */
     void _updateLastCommittedOpTimeAndWallTime(WithLock lk);
 
+    /** Terms only increase, so if an incoming term is less than or equal to our
+     * current term (_termShadow), there is no need to take the mutex and call _updateTerm_inlock.
+     * Since _termShadow may be lagged, this may return true when the term does not need to be
+     * updated, which is harmless because _updateTerm_inlock will do nothing in that case.
+     */
+    bool _needToUpdateTerm(long long term);
+
     /**
      * Callback that attempts to set the current term in topology coordinator and
      * relinquishes primary if the term actually changes and we are primary.
-- 
cgit v1.2.1