SERVER-7199 Bump minvalid correctly when batch goes past minvalid

3 files changed, 44 insertions, 16 deletions

diff --git a/src/mongo/db/repl/rs_initialsync.cpp b/src/mongo/db/repl/rs_initialsync.cpp
index 3a56066e091..bc98c903167 100644
--- a/src/mongo/db/repl/rs_initialsync.cpp
+++ b/src/mongo/db/repl/rs_initialsync.cpp
@@ -295,7 +295,7 @@ namespace mongo {
         // apply startingTS..mvoptime portion of the oplog
         {
             try {
-                syncer.oplogApplication(lastOp, minValid);
+                minValid = syncer.oplogApplication(lastOp, minValid);
             }
             catch (const DBException&) {
                 log() << "replSet initial sync failed during oplog application phase" << rsLog;
diff --git a/src/mongo/db/repl/rs_sync.cpp b/src/mongo/db/repl/rs_sync.cpp
index 7c0a5557122..bd7d24c6742 100644
--- a/src/mongo/db/repl/rs_sync.cpp
+++ b/src/mongo/db/repl/rs_sync.cpp
@@ -240,18 +240,17 @@ namespace replset {
 
     InitialSync::~InitialSync() {}
 
-    void SyncTail::oplogApplySegment(const BSONObj& applyGTEObj, const BSONObj& minValidObj,
+    BSONObj SyncTail::oplogApplySegment(const BSONObj& applyGTEObj, const BSONObj& minValidObj,
                                      MultiSyncApplyFunc func) {
         OpTime applyGTE = applyGTEObj["ts"]._opTime();
         OpTime minValid = minValidObj["ts"]._opTime();
 
-        // if there were no writes during the initial sync, there will be nothing in the queue so
-        // just go live
-        if (minValid == applyGTE) {
-            return;
-        }
+        // We have to keep track of the last op applied to the data, because there's no other easy
+        // way of getting this data synchronously.  Batches may go past minValidObj, so we need to
+        // know to bump minValid past minValidObj.
+        BSONObj lastOp = applyGTEObj;
+        OpTime ts = applyGTE;
 
-        OpTime ts;
         time_t start = time(0);
         unsigned long long n = 0, lastN = 0;
 
@@ -281,17 +280,19 @@ namespace replset {
             }
 
             // we want to keep a record of the last op applied, to compare with minvalid
-            const BSONObj& lastOp = ops.getDeque().back();
+            lastOp = ops.getDeque().back();
             OpTime tempTs = lastOp["ts"]._opTime();
             applyOpsToOplog(&ops.getDeque());
 
             ts = tempTs;
         }
+
+        return lastOp;
     }
 
     /* initial oplog application, during initial sync, after cloning.
     */
-    void InitialSync::oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj) {
+    BSONObj InitialSync::oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj) {
         if (replSetForceInitialSyncFailure > 0) {
             log() << "replSet test code invoked, forced InitialSync failure: " << replSetForceInitialSyncFailure << rsLog;
             replSetForceInitialSyncFailure--;
@@ -302,11 +303,11 @@ namespace replset {
         syncApply(applyGTEObj);
         _logOpObjRS(applyGTEObj);
 
-        oplogApplySegment(applyGTEObj, minValidObj, multiInitialSyncApply);
+        return oplogApplySegment(applyGTEObj, minValidObj, multiInitialSyncApply);
     }
 
-    void SyncTail::oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj) {
-        oplogApplySegment(applyGTEObj, minValidObj, multiSyncApply);
+    BSONObj SyncTail::oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj) {
+        return oplogApplySegment(applyGTEObj, minValidObj, multiSyncApply);
     }
 
     /* tail an oplog.  ok to return, will be re-called. */
diff --git a/src/mongo/db/repl/rs_sync.h b/src/mongo/db/repl/rs_sync.h
index aa3a5c9a2cb..1e52380b458 100644
--- a/src/mongo/db/repl/rs_sync.h
+++ b/src/mongo/db/repl/rs_sync.h
@@ -39,9 +39,31 @@ namespace replset {
         SyncTail(BackgroundSyncInterface *q);
         virtual ~SyncTail();
         virtual bool syncApply(const BSONObj &o, bool convertUpdateToUpsert = false);
-        void oplogApplySegment(const BSONObj& applyGTEObj, const BSONObj& minValidObj,
+
+        /**
+         * Apply ops from applyGTEObj's ts to at least minValidObj's ts.  Note that, due to
+         * batching, this may end up applying ops beyond minValidObj's ts.
+         *
+         * @param applyGTEObj the op to start replicating at.  This is actually not used except in
+         *                    comparision to minValidObj: the background sync thread keeps its own
+         *                    record of where we're synced to and starts providing ops from that
+         *                    point.
+         * @param minValidObj the op to finish syncing at.  This function cannot return (other than
+         *                    fatally erroring out) without applying at least this op.
+         * @param func        whether this should use initial sync logic (recloning docs) or
+         *                    "normal" logic.
+         * @return BSONObj    the op that was synced to.  This may be greater than minValidObj, as a
+         *                    single batch might blow right by minvalid. If applyGTEObj is the same
+         *                    op as minValidObj, this will be applyGTEObj.
+         */
+        BSONObj oplogApplySegment(const BSONObj& applyGTEObj, const BSONObj& minValidObj,
                                MultiSyncApplyFunc func);
-        virtual void oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj);
+
+        /**
+         * Runs oplogApplySegment without allowing recloning documents.
+         */
+        virtual BSONObj oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj);
+
         void oplogApplication();
         bool peek(BSONObj* obj);
 
@@ -107,7 +129,12 @@ namespace replset {
     public:
         virtual ~InitialSync();
         InitialSync(BackgroundSyncInterface *q);
-        void oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj);
+
+        /**
+         * Creates the initial oplog entry: applies applyGTEObj and writes it to the oplog.  Then
+         * this runs oplogApplySegment allowing recloning documents.
+         */
+        BSONObj oplogApplication(const BSONObj& applyGTEObj, const BSONObj& minValidObj);
     };
 
     // TODO: move hbmsg into an error-keeping class (SERVER-4444)