SERVER-22027 Sharding should not retry killed operations

14 files changed, 73 insertions, 88 deletions

diff --git a/src/mongo/base/error_codes.err b/src/mongo/base/error_codes.err
index faed0dbb14a..43fdefa1c50 100644
--- a/src/mongo/base/error_codes.err
+++ b/src/mongo/base/error_codes.err
@@ -160,6 +160,7 @@ error_code("CannotGrowDocumentInCappedNamespace", 10003)
 error_code("DuplicateKey", 11000)
 error_code("InterruptedAtShutdown", 11600)
 error_code("Interrupted", 11601)
+error_code("InterruptedDueToReplStateChange", 11602)
 error_code("OutOfDiskSpace", 14031 )
 error_code("KeyTooLong", 17280);
 error_code("BackgroundOperationInProgressForDatabase", 12586);
@@ -173,7 +174,10 @@ error_code("PrepareConfigsFailed", 13104);
 
 # SERVER-21428 explains the extra changes needed if error_class components change.
 error_class("NetworkError", ["HostUnreachable", "HostNotFound", "NetworkTimeout"])
-error_class("Interruption", ["Interrupted", "InterruptedAtShutdown", "ExceededTimeLimit"])
+error_class("Interruption", ["Interrupted",
+                             "InterruptedAtShutdown",
+                             "InterruptedDueToReplStateChange",
+                             "ExceededTimeLimit"])
 error_class("NotMasterError", ["NotMaster", "NotMasterNoSlaveOk"])
 error_class("StaleShardingError",
             ["RecvStaleConfig", "SendStaleConfig", "StaleShardVersion", "StaleEpoch"])
diff --git a/src/mongo/db/operation_context.cpp b/src/mongo/db/operation_context.cpp
index 1b79dadcb50..1434ee5ddc7 100644
--- a/src/mongo/db/operation_context.cpp
+++ b/src/mongo/db/operation_context.cpp
@@ -46,12 +46,13 @@ Client* OperationContext::getClient() const {
     return _client;
 }
 
-void OperationContext::markKilled() {
-    _killPending.store(1);
+void OperationContext::markKilled(ErrorCodes::Error killCode) {
+    invariant(killCode != ErrorCodes::OK);
+    _killCode.compareAndSwap(ErrorCodes::OK, killCode);
 }
 
-bool OperationContext::isKillPending() const {
-    return _killPending.loadRelaxed();
+ErrorCodes::Error OperationContext::getKillStatus() const {
+    return _killCode.loadRelaxed();
 }
 
 }  // namespace mongo
diff --git a/src/mongo/db/operation_context.h b/src/mongo/db/operation_context.h
index a146a8f7f14..3a7e3165dda 100644
--- a/src/mongo/db/operation_context.h
+++ b/src/mongo/db/operation_context.h
@@ -77,7 +77,6 @@ public:
      */
     virtual RecoveryUnit* recoveryUnit() const = 0;
 
-
     /**
      * Returns the RecoveryUnit (same return value as recoveryUnit()) but the caller takes
      * ownership of the returned RecoveryUnit, and the OperationContext instance relinquishes
@@ -183,23 +182,33 @@ public:
     virtual bool writesAreReplicated() const = 0;
 
     /**
-     * Marks this operation as killed.
+     * Marks this operation as killed so that subsequent calls to checkForInterrupt and
+     * checkForInterruptNoAssert by the thread executing the operation will start returning the
+     * specified error code.
      *
-     * Subsequent calls to checkForInterrupt and checkForInterruptNoAssert by the thread
-     * executing the operation will indicate that the operation has been killed.
+     * If multiple threads kill the same operation with different codes, only the first code will
+     * be preserved.
      *
-     * May be called by any thread that has locked the Client owning this operation context,
-     * or by the thread executing on behalf of this operation context.
+     * May be called by any thread that has locked the Client owning this operation context.
      */
-    void markKilled();
+    void markKilled(ErrorCodes::Error killCode = ErrorCodes::Interrupted);
 
     /**
-     * Returns true if markKilled has been called on this operation context.
+     * Returns the code passed to markKilled if this operation context has been killed previously
+     * or ErrorCodes::OK otherwise.
      *
-     * May be called by any thread that has locked the Client owning this operation context,
-     * or by the thread executing on behalf of this operation context.
+     * May be called by any thread that has locked the Client owning this operation context, or
+     * without lock by the thread executing on behalf of this operation context.
+     */
+    ErrorCodes::Error getKillStatus() const;
+
+    /**
+     * Shortcut method, which checks whether getKillStatus returns a non-OK value. Has the same
+     * concurrency rules as getKillStatus.
      */
-    bool isKillPending() const;
+    bool isKillPending() const {
+        return getKillStatus() != ErrorCodes::OK;
+    }
 
 protected:
     OperationContext(Client* client, unsigned int opId, Locker* locker);
@@ -211,11 +220,14 @@ private:
     Client* const _client;
     const unsigned int _opId;
 
-    // The lifetime of locker is managed by subclasses of OperationContext, so it is not
-    // safe to access _locker in the destructor of OperationContext.
+    // Not owned.
     Locker* const _locker;
 
-    AtomicInt32 _killPending{0};
+    // Follows the values of ErrorCodes::Error. The default value is 0 (OK), which means the
+    // operation is not killed. If killed, it will contain a specific code. This value changes only
+    // once from OK to some kill code.
+    AtomicWord<ErrorCodes::Error> _killCode{ErrorCodes::OK};
+
     WriteConcernOptions _writeConcern;
 };
 
diff --git a/src/mongo/db/operation_context_impl.cpp b/src/mongo/db/operation_context_impl.cpp
index f2c0166876f..b7733958bc2 100644
--- a/src/mongo/db/operation_context_impl.cpp
+++ b/src/mongo/db/operation_context_impl.cpp
@@ -140,6 +140,7 @@ uint64_t OperationContextImpl::getRemainingMaxTimeMicros() const {
 MONGO_FP_DECLARE(checkForInterruptFail);
 
 namespace {
+
 // Helper function for checkForInterrupt fail point.  Decides whether the operation currently
 // being run by the given Client meet the (probabilistic) conditions for interruption as
 // specified in the fail point info.
@@ -194,8 +195,9 @@ Status OperationContextImpl::checkForInterruptNoAssert() {
         }
     }
 
-    if (isKillPending()) {
-        return Status(ErrorCodes::Interrupted, "operation was interrupted");
+    const auto killStatus = getKillStatus();
+    if (killStatus != ErrorCodes::OK) {
+        return Status(killStatus, "operation was interrupted");
     }
 
     return Status::OK();
diff --git a/src/mongo/db/repl/replication_coordinator_external_state_impl.cpp b/src/mongo/db/repl/replication_coordinator_external_state_impl.cpp
index c42b80a605d..52b0fee436c 100644
--- a/src/mongo/db/repl/replication_coordinator_external_state_impl.cpp
+++ b/src/mongo/db/repl/replication_coordinator_external_state_impl.cpp
@@ -350,8 +350,8 @@ void ReplicationCoordinatorExternalStateImpl::closeConnections() {
 }
 
 void ReplicationCoordinatorExternalStateImpl::killAllUserOperations(OperationContext* txn) {
-    ServiceContext* environment = getGlobalServiceContext();
-    environment->killAllUserOperations(txn);
+    ServiceContext* environment = txn->getServiceContext();
+    environment->killAllUserOperations(txn, ErrorCodes::InterruptedDueToReplStateChange);
 }
 
 void ReplicationCoordinatorExternalStateImpl::clearShardingState() {
diff --git a/src/mongo/db/service_context.h b/src/mongo/db/service_context.h
index 6f9aa36a22f..4e63046726f 100644
--- a/src/mongo/db/service_context.h
+++ b/src/mongo/db/service_context.h
@@ -284,7 +284,7 @@ public:
      * Kills all operations that have a Client that is associated with an incoming user
      * connection, except for the one associated with txn.
      */
-    virtual void killAllUserOperations(const OperationContext* txn) = 0;
+    virtual void killAllUserOperations(const OperationContext* txn, ErrorCodes::Error killCode) = 0;
 
     /**
      * Registers a listener to be notified each time an op is killed.
diff --git a/src/mongo/db/service_context_d.cpp b/src/mongo/db/service_context_d.cpp
index 3e5fd2d5be7..470dcb927a5 100644
--- a/src/mongo/db/service_context_d.cpp
+++ b/src/mongo/db/service_context_d.cpp
@@ -220,21 +220,9 @@ bool ServiceContextMongoD::getKillAllOperations() {
     return _globalKill;
 }
 
-bool ServiceContextMongoD::_killOperationsAssociatedWithClientAndOpId_inlock(Client* client,
-                                                                             unsigned int opId) {
-    OperationContext* opCtx = client->getOperationContext();
-    if (!opCtx) {
-        return false;
-    }
-    if (opCtx->getOpID() != opId) {
-        return false;
-    }
-    _killOperation_inlock(opCtx);
-    return true;
-}
-
-void ServiceContextMongoD::_killOperation_inlock(OperationContext* opCtx) {
-    opCtx->markKilled();
+void ServiceContextMongoD::_killOperation_inlock(OperationContext* opCtx,
+                                                 ErrorCodes::Error killCode) {
+    opCtx->markKilled(killCode);
 
     for (const auto listener : _killOpListeners) {
         try {
@@ -248,8 +236,10 @@ void ServiceContextMongoD::_killOperation_inlock(OperationContext* opCtx) {
 bool ServiceContextMongoD::killOperation(unsigned int opId) {
     for (LockedClientsCursor cursor(this); Client* client = cursor.next();) {
         stdx::lock_guard<Client> lk(*client);
-        bool found = _killOperationsAssociatedWithClientAndOpId_inlock(client, opId);
-        if (found) {
+
+        OperationContext* opCtx = client->getOperationContext();
+        if (opCtx && opCtx->getOpID() == opId) {
+            _killOperation_inlock(opCtx, ErrorCodes::Interrupted);
             return true;
         }
     }
@@ -257,7 +247,8 @@ bool ServiceContextMongoD::killOperation(unsigned int opId) {
     return false;
 }
 
-void ServiceContextMongoD::killAllUserOperations(const OperationContext* txn) {
+void ServiceContextMongoD::killAllUserOperations(const OperationContext* txn,
+                                                 ErrorCodes::Error killCode) {
     for (LockedClientsCursor cursor(this); Client* client = cursor.next();) {
         if (!client->isFromUserConnection()) {
             // Don't kill system operations.
@@ -266,16 +257,11 @@ void ServiceContextMongoD::killAllUserOperations(const OperationContext* txn) {
 
         stdx::lock_guard<Client> lk(*client);
         OperationContext* toKill = client->getOperationContext();
-        if (!toKill) {
-            continue;
-        }
 
-        if (toKill->getOpID() == txn->getOpID()) {
-            // Don't kill ourself.
-            continue;
+        // Don't kill ourself.
+        if (toKill && toKill->getOpID() != txn->getOpID()) {
+            _killOperation_inlock(toKill, killCode);
         }
-
-        _killOperation_inlock(toKill);
     }
 }
 
diff --git a/src/mongo/db/service_context_d.h b/src/mongo/db/service_context_d.h
index 0c560ff17f4..dc834197e9e 100644
--- a/src/mongo/db/service_context_d.h
+++ b/src/mongo/db/service_context_d.h
@@ -67,7 +67,7 @@ public:
 
     bool killOperation(unsigned int opId) override;
 
-    void killAllUserOperations(const OperationContext* txn) override;
+    void killAllUserOperations(const OperationContext* txn, ErrorCodes::Error killCode) override;
 
     void registerKillOpListener(KillOpListenerInterface* listener) override;
 
@@ -79,22 +79,11 @@ private:
     std::unique_ptr<OperationContext> _newOpCtx(Client* client) override;
 
     /**
-     * Kills the active operation on "client" if that operation is associated with operation id
-     * "opId".
-     *
-     * Returns true if an operation was killed.
-     *
-     * Must only be called by a thread owning both this service context's mutex and the
-     * client's.
-     */
-    bool _killOperationsAssociatedWithClientAndOpId_inlock(Client* client, unsigned int opId);
-
-    /**
      * Kills the given operation.
      *
      * Caller must own the service context's _mutex.
      */
-    void _killOperation_inlock(OperationContext* opCtx);
+    void _killOperation_inlock(OperationContext* opCtx, ErrorCodes::Error killCode);
 
     bool _globalKill;
 
diff --git a/src/mongo/db/service_context_noop.cpp b/src/mongo/db/service_context_noop.cpp
index 65184906442..fd94b35db89 100644
--- a/src/mongo/db/service_context_noop.cpp
+++ b/src/mongo/db/service_context_noop.cpp
@@ -79,7 +79,8 @@ bool ServiceContextNoop::killOperation(unsigned int opId) {
     return false;
 }
 
-void ServiceContextNoop::killAllUserOperations(const OperationContext* txn) {}
+void ServiceContextNoop::killAllUserOperations(const OperationContext* txn,
+                                               ErrorCodes::Error killCode) {}
 
 void ServiceContextNoop::registerKillOpListener(KillOpListenerInterface* listener) {}
 
diff --git a/src/mongo/db/service_context_noop.h b/src/mongo/db/service_context_noop.h
index 0e74b61b7a2..62e6a169896 100644
--- a/src/mongo/db/service_context_noop.h
+++ b/src/mongo/db/service_context_noop.h
@@ -49,7 +49,7 @@ public:
 
     bool killOperation(unsigned int opId) override;
 
-    void killAllUserOperations(const OperationContext* txn) override;
+    void killAllUserOperations(const OperationContext* txn, ErrorCodes::Error killCode) override;
 
     void setKillAllOperations() override;
 
diff --git a/src/mongo/s/catalog/replset/catalog_manager_replica_set.cpp b/src/mongo/s/catalog/replset/catalog_manager_replica_set.cpp
index 5943cc6e255..5686476e1a3 100644
--- a/src/mongo/s/catalog/replset/catalog_manager_replica_set.cpp
+++ b/src/mongo/s/catalog/replset/catalog_manager_replica_set.cpp
@@ -389,8 +389,6 @@ StatusWith<OpTimePair<DatabaseType>> CatalogManagerReplicaSet::_fetchDatabaseMet
 
 StatusWith<OpTimePair<CollectionType>> CatalogManagerReplicaSet::getCollection(
     OperationContext* txn, const std::string& collNs) {
-    auto configShard = grid.shardRegistry()->getShard(txn, "config");
-
     auto statusFind = _exhaustiveFindOnConfig(txn,
                                               kConfigReadSelector,
                                               NamespaceString(CollectionType::ConfigNS),
diff --git a/src/mongo/s/catalog/replset/catalog_manager_replica_set_write_retry_test.cpp b/src/mongo/s/catalog/replset/catalog_manager_replica_set_write_retry_test.cpp
index b21ff63961f..9bf44eaa09c 100644
--- a/src/mongo/s/catalog/replset/catalog_manager_replica_set_write_retry_test.cpp
+++ b/src/mongo/s/catalog/replset/catalog_manager_replica_set_write_retry_test.cpp
@@ -93,7 +93,7 @@ TEST_F(InsertRetryTest, RetryOnInterruptedAndNetworkErrorSuccess) {
     onCommand([&](const RemoteCommandRequest& request) {
         ASSERT_EQ(request.target, kTestHosts[0]);
         configTargeter()->setFindHostReturnValue({kTestHosts[1]});
-        return Status(ErrorCodes::Interrupted, "Interruption");
+        return Status(ErrorCodes::InterruptedDueToReplStateChange, "Interruption");
     });
 
     onCommand([&](const RemoteCommandRequest& request) {
@@ -271,7 +271,7 @@ TEST_F(UpdateRetryTest, OperationInterruptedDueToPrimaryStepDown) {
 
         auto writeErrDetail = stdx::make_unique<WriteErrorDetail>();
         writeErrDetail->setIndex(0);
-        writeErrDetail->setErrCode(ErrorCodes::Interrupted);
+        writeErrDetail->setErrCode(ErrorCodes::InterruptedDueToReplStateChange);
         writeErrDetail->setErrMessage("Operation interrupted");
         response.addToErrDetails(writeErrDetail.release());
 
diff --git a/src/mongo/s/client/shard_registry.cpp b/src/mongo/s/client/shard_registry.cpp
index 55121f42209..e5b91dda83a 100644
--- a/src/mongo/s/client/shard_registry.cpp
+++ b/src/mongo/s/client/shard_registry.cpp
@@ -126,6 +126,7 @@ const ShardRegistry::ErrorCodesSet ShardRegistry::kNotMasterErrors{ErrorCodes::N
 const ShardRegistry::ErrorCodesSet ShardRegistry::kAllRetriableErrors{
     ErrorCodes::NotMaster,
     ErrorCodes::NotMasterNoSlaveOk,
+    ErrorCodes::NotMasterOrSecondary,
     // If write concern failed to be satisfied on the remote server, this most probably means that
     // some of the secondary nodes were unreachable or otherwise unresponsive, so the call is safe
     // to be retried if idempotency can be guaranteed.
@@ -133,10 +134,7 @@ const ShardRegistry::ErrorCodesSet ShardRegistry::kAllRetriableErrors{
     ErrorCodes::HostUnreachable,
     ErrorCodes::HostNotFound,
     ErrorCodes::NetworkTimeout,
-    // This set includes interrupted because replica set step down kills all server operations
-    // before it closes connections so it may happen that the caller actually receives the
-    // interruption.
-    ErrorCodes::Interrupted};
+    ErrorCodes::InterruptedDueToReplStateChange};
 
 ShardRegistry::ShardRegistry(std::unique_ptr<RemoteCommandTargeterFactory> targeterFactory,
                              std::unique_ptr<executor::TaskExecutorPool> executorPool,
@@ -782,7 +780,8 @@ StatusWith<ShardRegistry::CommandResponse> ShardRegistry::_runCommandWithMetadat
 void ShardRegistry::updateReplSetMonitor(const std::shared_ptr<RemoteCommandTargeter>& targeter,
                                          const HostAndPort& remoteHost,
                                          const Status& remoteCommandStatus) {
-    if (ErrorCodes::isNotMasterError(remoteCommandStatus.code())) {
+    if (ErrorCodes::isNotMasterError(remoteCommandStatus.code()) ||
+        (remoteCommandStatus == ErrorCodes::InterruptedDueToReplStateChange)) {
         targeter->markHostNotMaster(remoteHost);
     } else if (ErrorCodes::isNetworkError(remoteCommandStatus.code())) {
         targeter->markHostUnreachable(remoteHost);
diff --git a/src/mongo/s/query/async_results_merger.cpp b/src/mongo/s/query/async_results_merger.cpp
index 8b23528a04a..c82de6a3bbe 100644
--- a/src/mongo/s/query/async_results_merger.cpp
+++ b/src/mongo/s/query/async_results_merger.cpp
@@ -51,15 +51,6 @@ namespace {
 // Maximum number of retries for network and replication notMaster errors (per host).
 const int kMaxNumFailedHostRetryAttempts = 3;
 
-/**
- * Returns whether a particular error code returned from the initial cursor establishment should
- * be retried.
- */
-bool isPerShardRetriableError(ErrorCodes::Error err) {
-    return (ShardRegistry::kAllRetriableErrors.count(err) ||
-            err == ErrorCodes::NotMasterOrSecondary);
-}
-
 }  // namespace
 
 AsyncResultsMerger::AsyncResultsMerger(executor::TaskExecutor* executor,
@@ -438,8 +429,7 @@ void AsyncResultsMerger::handleBatchResponse(
     if (!cursorResponseStatus.isOK()) {
         // Notify the shard registry of the failure.
         if (remote.shardId) {
-            // TODO: Pass down an OperationContext* to use here.
-            auto shard = grid.shardRegistry()->getShard(nullptr, *remote.shardId);
+            auto shard = grid.shardRegistry()->getShardNoReload(*remote.shardId);
             if (!shard) {
                 remote.status = Status(cursorResponseStatus.getStatus().code(),
                                        str::stream() << "Could not find shard " << *remote.shardId
@@ -453,7 +443,10 @@ void AsyncResultsMerger::handleBatchResponse(
 
         // If the error is retriable, schedule another request.
         if (!remote.cursorId && remote.retryCount < kMaxNumFailedHostRetryAttempts &&
-            isPerShardRetriableError(cursorResponseStatus.getStatus().code())) {
+            ShardRegistry::kAllRetriableErrors.count(cursorResponseStatus.getStatus().code())) {
+            LOG(1) << "Initial cursor establishment failed with retriable error and will be retried"
+                   << causedBy(cursorResponseStatus.getStatus());
+
             ++remote.retryCount;
 
             // Since we potentially updated the targeter that the last host it chose might be
@@ -641,13 +634,13 @@ Status AsyncResultsMerger::RemoteCursorData::resolveShardIdToHostAndPort(
     invariant(shardId);
     invariant(!cursorId);
 
-    // TODO: Pass down an OperationContext* to use here.
-    const auto shard = grid.shardRegistry()->getShard(nullptr, *shardId);
+    const auto shard = grid.shardRegistry()->getShardNoReload(*shardId);
     if (!shard) {
         return Status(ErrorCodes::ShardNotFound,
                       str::stream() << "Could not find shard " << *shardId);
     }
 
+    // TODO: Pass down an OperationContext* to use here.
     auto findHostStatus = shard->getTargeter()->findHost(
         readPref, RemoteCommandTargeter::selectFindHostMaxWaitTime(nullptr));
     if (!findHostStatus.isOK()) {