SERVER-31933 Add prepareSnapshot method to avoid doing this implicitly in _getNextOptimes mutex

5 files changed, 17 insertions, 1 deletions

diff --git a/src/mongo/db/repl/oplog.cpp b/src/mongo/db/repl/oplog.cpp
index dcc54adf696..a2535fdb44e 100644
--- a/src/mongo/db/repl/oplog.cpp
+++ b/src/mongo/db/repl/oplog.cpp
@@ -158,6 +158,8 @@ void _getNextOpTimes(OperationContext* opCtx,
         term = replCoord->getTerm();
     }
 
+    // Allow the storage engine to start the transaction outside the critical section.
+    opCtx->recoveryUnit()->prepareSnapshot();
     stdx::lock_guard<stdx::mutex> lk(newOpMutex);
 
     auto ts = LogicalClock::get(opCtx)->reserveTicks(count).asTimestamp();
diff --git a/src/mongo/db/storage/recovery_unit.h b/src/mongo/db/storage/recovery_unit.h
index fe9ed473163..5ac712b4887 100644
--- a/src/mongo/db/storage/recovery_unit.h
+++ b/src/mongo/db/storage/recovery_unit.h
@@ -85,6 +85,14 @@ public:
     virtual void abandonSnapshot() = 0;
 
     /**
+     * Informs the RecoveryUnit that a snapshot will be needed soon, if one was not already
+     * established. This specifically allows the storage engine to preallocate any required
+     * transaction resources while minimizing the critical section between generating a new
+     * timestamp and setting it using setTimestamp.
+     */
+    virtual void prepareSnapshot() {}
+
+    /**
      * Informs this RecoveryUnit that all future reads through it should be from a snapshot
      * marked as Majority Committed. Snapshots should still be separately acquired and newer
      * committed snapshots should be used if available whenever implementations would normally
diff --git a/src/mongo/db/storage/wiredtiger/wiredtiger_record_store.cpp b/src/mongo/db/storage/wiredtiger/wiredtiger_record_store.cpp
index 232ecdcdbb7..336b7d9d486 100644
--- a/src/mongo/db/storage/wiredtiger/wiredtiger_record_store.cpp
+++ b/src/mongo/db/storage/wiredtiger/wiredtiger_record_store.cpp
@@ -1675,7 +1675,7 @@ void WiredTigerRecordStore::cappedTruncateAfter(OperationContext* opCtx,
 
 Status WiredTigerRecordStore::oplogDiskLocRegister(OperationContext* opCtx,
                                                    const Timestamp& opTime) {
-    // This starts a new transaction and gives it a timestamp.
+    // This labels the current transaction with a timestamp.
     // This is required for oplog visibility to work correctly, as WiredTiger uses the transaction
     // list to determine where there are holes in the oplog.
     return opCtx->recoveryUnit()->setTimestamp(SnapshotName(opTime));
diff --git a/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.cpp b/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.cpp
index 8be67c449a5..54db183fe36 100644
--- a/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.cpp
+++ b/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.cpp
@@ -178,6 +178,11 @@ void WiredTigerRecoveryUnit::abandonSnapshot() {
     _areWriteUnitOfWorksBanned = false;
 }
 
+void WiredTigerRecoveryUnit::prepareSnapshot() {
+    // Begin a new transaction, if one is not already started.
+    getSession();
+}
+
 void* WiredTigerRecoveryUnit::writingPtr(void* data, size_t len) {
     // This API should not be used for anything other than the MMAP V1 storage engine
     MONGO_UNREACHABLE;
diff --git a/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.h b/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.h
index 397d17a142b..a1040952dc3 100644
--- a/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.h
+++ b/src/mongo/db/storage/wiredtiger/wiredtiger_recovery_unit.h
@@ -70,6 +70,7 @@ public:
     void registerChange(Change* change) override;
 
     void abandonSnapshot() override;
+    void prepareSnapshot() override;
 
     Status setReadFromMajorityCommittedSnapshot() override;
     bool isReadingFromMajorityCommittedSnapshot() const override {