SERVER-33355 Support readConcern level snapshot reads on secondaries

11 files changed, 32 insertions, 50 deletions

diff --git a/src/mongo/client/embedded/service_entry_point_embedded.cpp b/src/mongo/client/embedded/service_entry_point_embedded.cpp
index 2eb715adc14..2a0e2ecb3b3 100644
--- a/src/mongo/client/embedded/service_entry_point_embedded.cpp
+++ b/src/mongo/client/embedded/service_entry_point_embedded.cpp
@@ -93,9 +93,7 @@ public:
 
     void waitForReadConcern(OperationContext*,
                             const CommandInvocation*,
-                            const std::string&,
-                            const OpMsgRequest&,
-                            const BSONObj&) const override {}
+                            const OpMsgRequest&) const override {}
 
     void waitForWriteConcern(OperationContext* opCtx,
                              const std::string& commandName,
diff --git a/src/mongo/db/commands/get_last_error.cpp b/src/mongo/db/commands/get_last_error.cpp
index c7673052e8b..900434db737 100644
--- a/src/mongo/db/commands/get_last_error.cpp
+++ b/src/mongo/db/commands/get_last_error.cpp
@@ -236,10 +236,7 @@ public:
         // Validate write concern no matter what, this matches 2.4 behavior
         //
         if (status.isOK()) {
-            // Ensure options are valid for this host. Since getLastError doesn't do writes itself,
-            // treat it as if these are admin database writes, which need to be replicated so we do
-            // the strictest checks write concern checks.
-            status = validateWriteConcern(opCtx, writeConcern, NamespaceString::kAdminDb);
+            status = validateWriteConcern(opCtx, writeConcern);
         }
 
         if (!status.isOK()) {
diff --git a/src/mongo/db/service_entry_point_common.cpp b/src/mongo/db/service_entry_point_common.cpp
index 753add5bb6d..52623b256a9 100644
--- a/src/mongo/db/service_entry_point_common.cpp
+++ b/src/mongo/db/service_entry_point_common.cpp
@@ -403,21 +403,13 @@ bool runCommandImpl(OperationContext* opCtx,
         bytesToReserve = 0;
 #endif
 
-    // run expects non-const bsonobj
-    BSONObj cmd = request.body;
-
-    // run expects const db std::string (can't bind to temporary)
-    const std::string db = request.getDatabase().toString();
-
     CommandReplyBuilder crb(replyBuilder->getInPlaceReplyBuilder(bytesToReserve));
 
-    behaviors.waitForReadConcern(opCtx, invocation, db, request, cmd);
-
     if (!invocation->supportsWriteConcern()) {
-        behaviors.uassertCommandDoesNotSpecifyWriteConcern(cmd);
+        behaviors.uassertCommandDoesNotSpecifyWriteConcern(request.body);
         invocation->run(opCtx, &crb);
     } else {
-        auto wcResult = uassertStatusOK(extractWriteConcern(opCtx, cmd, db));
+        auto wcResult = uassertStatusOK(extractWriteConcern(opCtx, request.body));
 
         auto lastOpBeforeRun = repl::ReplClientInfo::forClient(opCtx->getClient()).getLastOp();
 
@@ -647,12 +639,6 @@ void execCommandDatabase(OperationContext* opCtx,
                     "readConcernLevel snapshot requires a txnNumber",
                     opCtx->getTxnNumber());
 
-            // TODO SERVER-33355: Remove once readConcern level snapshot is supported on
-            // secondaries.
-            uassert(ErrorCodes::InvalidOptions,
-                    "readConcern level snapshot only supported on primaries",
-                    iAmPrimary);
-
             opCtx->lockState()->setSharedLocksShouldTwoPhaseLock(true);
         }
 
@@ -693,6 +679,8 @@ void execCommandDatabase(OperationContext* opCtx,
             rpc::TrackingMetadata::get(opCtx).setIsLogged(true);
         }
 
+        behaviors.waitForReadConcern(opCtx, invocation.get(), request);
+
         sessionTxnState.unstashTransactionResources();
         retval = runCommandImpl(
             opCtx, invocation.get(), request, replyBuilder, startOperationTime, behaviors);
diff --git a/src/mongo/db/service_entry_point_common.h b/src/mongo/db/service_entry_point_common.h
index a7e7155881a..f90fa0660af 100644
--- a/src/mongo/db/service_entry_point_common.h
+++ b/src/mongo/db/service_entry_point_common.h
@@ -62,9 +62,7 @@ struct ServiceEntryPointCommon {
         virtual bool lockedForWriting() const = 0;
         virtual void waitForReadConcern(OperationContext* opCtx,
                                         const CommandInvocation* invocation,
-                                        const std::string& db,
-                                        const OpMsgRequest& request,
-                                        const BSONObj& cmdObj) const = 0;
+                                        const OpMsgRequest& request) const = 0;
         virtual void waitForWriteConcern(OperationContext* opCtx,
                                          const std::string& commandName,
                                          const repl::OpTime& lastOpBeforeRun,
diff --git a/src/mongo/db/service_entry_point_mongod.cpp b/src/mongo/db/service_entry_point_mongod.cpp
index a442f40ee27..eb281981054 100644
--- a/src/mongo/db/service_entry_point_mongod.cpp
+++ b/src/mongo/db/service_entry_point_mongod.cpp
@@ -52,16 +52,15 @@ public:
 
     void waitForReadConcern(OperationContext* opCtx,
                             const CommandInvocation* invocation,
-                            const std::string& db,
-                            const OpMsgRequest& request,
-                            const BSONObj& cmdObj) const override {
+                            const OpMsgRequest& request) const override {
         Status rcStatus = mongo::waitForReadConcern(
             opCtx, repl::ReadConcernArgs::get(opCtx), invocation->allowsAfterClusterTime());
+
         if (!rcStatus.isOK()) {
             if (rcStatus == ErrorCodes::ExceededTimeLimit) {
                 const int debugLevel =
                     serverGlobalParams.clusterRole == ClusterRole::ConfigServer ? 0 : 2;
-                LOG(debugLevel) << "Command on database " << db
+                LOG(debugLevel) << "Command on database " << request.getDatabase()
                                 << " timed out waiting for read concern to be satisfied. Command: "
                                 << redact(ServiceEntryPointCommon::getRedactedCopyForLogging(
                                        invocation->definition(), request.body));
diff --git a/src/mongo/db/write_concern.cpp b/src/mongo/db/write_concern.cpp
index 0668e4c54ad..2237cdcbdf0 100644
--- a/src/mongo/db/write_concern.cpp
+++ b/src/mongo/db/write_concern.cpp
@@ -67,12 +67,11 @@ bool commandSpecifiesWriteConcern(const BSONObj& cmdObj) {
 }
 
 StatusWith<WriteConcernOptions> extractWriteConcern(OperationContext* opCtx,
-                                                    const BSONObj& cmdObj,
-                                                    const std::string& dbName) {
+                                                    const BSONObj& cmdObj) {
     // The default write concern if empty is {w:1}. Specifying {w:0} is/was allowed, but is
     // interpreted identically to {w:1}.
     auto wcResult = WriteConcernOptions::extractWCFromCommand(
-        cmdObj, dbName, repl::ReplicationCoordinator::get(opCtx)->getGetLastErrorDefault());
+        cmdObj, repl::ReplicationCoordinator::get(opCtx)->getGetLastErrorDefault());
     if (!wcResult.isOK()) {
         return wcResult.getStatus();
     }
@@ -88,7 +87,7 @@ StatusWith<WriteConcernOptions> extractWriteConcern(OperationContext* opCtx,
                 WriteConcernOptions::kMajority, WriteConcernOptions::SyncMode::UNSET, Seconds(30)};
         }
     } else {
-        Status wcStatus = validateWriteConcern(opCtx, writeConcern, dbName);
+        Status wcStatus = validateWriteConcern(opCtx, writeConcern);
         if (!wcStatus.isOK()) {
             return wcStatus;
         }
@@ -97,9 +96,7 @@ StatusWith<WriteConcernOptions> extractWriteConcern(OperationContext* opCtx,
     return writeConcern;
 }
 
-Status validateWriteConcern(OperationContext* opCtx,
-                            const WriteConcernOptions& writeConcern,
-                            StringData dbName) {
+Status validateWriteConcern(OperationContext* opCtx, const WriteConcernOptions& writeConcern) {
     if (writeConcern.syncMode == WriteConcernOptions::SyncMode::JOURNAL &&
         !opCtx->getServiceContext()->getGlobalStorageEngine()->isDurable()) {
         return Status(ErrorCodes::BadValue,
diff --git a/src/mongo/db/write_concern.h b/src/mongo/db/write_concern.h
index 60c31e6ff26..6a3021c969c 100644
--- a/src/mongo/db/write_concern.h
+++ b/src/mongo/db/write_concern.h
@@ -52,16 +52,12 @@ bool commandSpecifiesWriteConcern(const BSONObj& cmdObj);
  * Verifies that the writeConcern is of type Object (BSON type) and
  * that the resulting writeConcern is valid for this particular host.
  */
-StatusWith<WriteConcernOptions> extractWriteConcern(OperationContext* opCtx,
-                                                    const BSONObj& cmdObj,
-                                                    const std::string& dbName);
+StatusWith<WriteConcernOptions> extractWriteConcern(OperationContext* opCtx, const BSONObj& cmdObj);
 
 /**
- * Verifies that a WriteConcern is valid for this particular host and database.
+ * Verifies that a WriteConcern is valid for this particular host.
  */
-Status validateWriteConcern(OperationContext* opCtx,
-                            const WriteConcernOptions& writeConcern,
-                            StringData dbName);
+Status validateWriteConcern(OperationContext* opCtx, const WriteConcernOptions& writeConcern);
 
 struct WriteConcernResult {
     WriteConcernResult() {
diff --git a/src/mongo/db/write_concern_options.cpp b/src/mongo/db/write_concern_options.cpp
index 51b8484a51b..26f5778be37 100644
--- a/src/mongo/db/write_concern_options.cpp
+++ b/src/mongo/db/write_concern_options.cpp
@@ -151,7 +151,7 @@ Status WriteConcernOptions::parse(const BSONObj& obj) {
 }
 
 StatusWith<WriteConcernOptions> WriteConcernOptions::extractWCFromCommand(
-    const BSONObj& cmdObj, const std::string& dbName, const WriteConcernOptions& defaultWC) {
+    const BSONObj& cmdObj, const WriteConcernOptions& defaultWC) {
     WriteConcernOptions writeConcern = defaultWC;
     writeConcern.usedDefault = true;
     if (writeConcern.wNumNodes == 0 && writeConcern.wMode.empty()) {
diff --git a/src/mongo/db/write_concern_options.h b/src/mongo/db/write_concern_options.h
index d0a3be810e3..a809a2a1de5 100644
--- a/src/mongo/db/write_concern_options.h
+++ b/src/mongo/db/write_concern_options.h
@@ -69,9 +69,7 @@ public:
      * Verifies that the writeConcern is of type Object (BSON type).
      */
     static StatusWith<WriteConcernOptions> extractWCFromCommand(
-        const BSONObj& cmdObj,
-        const std::string& dbName,
-        const WriteConcernOptions& defaultWC = WriteConcernOptions());
+        const BSONObj& cmdObj, const WriteConcernOptions& defaultWC = WriteConcernOptions());
 
     /**
      * Return true if the server needs to wait for other secondary nodes to satisfy this
diff --git a/src/mongo/s/commands/strategy.cpp b/src/mongo/s/commands/strategy.cpp
index f51fdabe7c0..c9e9276764d 100644
--- a/src/mongo/s/commands/strategy.cpp
+++ b/src/mongo/s/commands/strategy.cpp
@@ -151,7 +151,7 @@ void execCommandClient(OperationContext* opCtx,
         appendRequiredFieldsToResponse(opCtx, &body);
     });
 
-    const auto dbname = request.getDatabase().toString();
+    const auto dbname = request.getDatabase();
     uassert(ErrorCodes::IllegalOperation,
             "Can't use 'local' database through mongos",
             dbname != NamespaceString::kLocalDb);
@@ -191,7 +191,7 @@ void execCommandClient(OperationContext* opCtx,
     }
 
     StatusWith<WriteConcernOptions> wcResult =
-        WriteConcernOptions::extractWCFromCommand(request.body, dbname);
+        WriteConcernOptions::extractWCFromCommand(request.body);
     if (!wcResult.isOK()) {
         auto body = result->getBodyBuilder();
         CommandHelpers::appendCommandStatus(body, wcResult.getStatus());
diff --git a/src/mongo/shell/replsettest.js b/src/mongo/shell/replsettest.js
index c11edd7c4c8..d848aa761d1 100644
--- a/src/mongo/shell/replsettest.js
+++ b/src/mongo/shell/replsettest.js
@@ -338,6 +338,17 @@ var ReplSetTest = function(opts) {
     }
 
     /**
+     * Returns the {readConcern: majority} OpTime for the host. Throws if not available.
+     */
+    this.getReadConcernMajorityOpTimeOrThrow = function(conn) {
+        const majorityOpTime = _getReadConcernMajorityOpTime(conn);
+        if (friendlyEqual(majorityOpTime, {ts: Timestamp(0, 0), t: NumberLong(0)})) {
+            throw new Error("readConcern majority optime not available");
+        }
+        return majorityOpTime;
+    };
+
+    /**
      * Returns the last durable OpTime for the host if running with journaling.
      * Returns the last applied OpTime otherwise.
      */