C++ Timestamp::isNull方法代码示例

void ReplicationRecoveryImpl::_applyToEndOfOplog(OperationContext* opCtx,
                                                 Timestamp oplogApplicationStartPoint,
                                                 Timestamp topOfOplog) {
    invariant(!oplogApplicationStartPoint.isNull());
    invariant(!topOfOplog.isNull());

    // Check if we have any unapplied ops in our oplog. It is important that this is done after
    // deleting the ragged end of the oplog.
    if (oplogApplicationStartPoint == topOfOplog) {
        log()
            << "No oplog entries to apply for recovery. appliedThrough is at the top of the oplog.";
        return;  // We've applied all the valid oplog we have.
    } else if (oplogApplicationStartPoint > topOfOplog) {
        severe() << "Applied op " << oplogApplicationStartPoint.toBSON()
                 << " not found. Top of oplog is " << topOfOplog.toBSON() << '.';
        fassertFailedNoTrace(40313);
    }

    log() << "Replaying stored operations from " << oplogApplicationStartPoint.toBSON()
          << " (exclusive) to " << topOfOplog.toBSON() << " (inclusive).";

    DBDirectClient db(opCtx);
    auto cursor = db.query(NamespaceString::kRsOplogNamespace.ns(),
                           QUERY("ts" << BSON("$gte" << oplogApplicationStartPoint)),
                           /*batchSize*/ 0,
                           /*skip*/ 0,
                           /*projection*/ nullptr,
                           QueryOption_OplogReplay);

    // Check that the first document matches our appliedThrough point then skip it since it's
    // already been applied.
    if (!cursor->more()) {
        // This should really be impossible because we check above that the top of the oplog is
        // strictly > appliedThrough. If this fails it represents a serious bug in either the
        // storage engine or query's implementation of OplogReplay.
        severe() << "Couldn't find any entries in the oplog >= "
                 << oplogApplicationStartPoint.toBSON() << " which should be impossible.";
        fassertFailedNoTrace(40293);
    }

    auto firstTimestampFound =
        fassertStatusOK(40291, OpTime::parseFromOplogEntry(cursor->nextSafe())).getTimestamp();
    if (firstTimestampFound != oplogApplicationStartPoint) {
        severe() << "Oplog entry at " << oplogApplicationStartPoint.toBSON()
                 << " is missing; actual entry found is " << firstTimestampFound.toBSON();
        fassertFailedNoTrace(40292);
    }

    // Apply remaining ops one at at time, but don't log them because they are already logged.
    UnreplicatedWritesBlock uwb(opCtx);

    while (cursor->more()) {
        auto entry = cursor->nextSafe();
        fassertStatusOK(40294,
                        SyncTail::syncApply(opCtx, entry, OplogApplication::Mode::kRecovering));
        _consistencyMarkers->setAppliedThrough(
            opCtx, fassertStatusOK(40295, OpTime::parseFromOplogEntry(entry)));
    }
}

void OplogBufferCollection::pushAllNonBlocking(OperationContext* txn,
        Batch::const_iterator begin,
        Batch::const_iterator end) {
    if (begin == end) {
        return;
    }
    size_t numDocs = std::distance(begin, end);
    Batch docsToInsert(numDocs);
    Timestamp ts;
    std::transform(begin, end, docsToInsert.begin(), [&ts](const Value& value) {
        auto pair = addIdToDocument(value);
        invariant(ts.isNull() || pair.second > ts);
        ts = pair.second;
        return pair.first;
    });

    stdx::lock_guard<stdx::mutex> lk(_mutex);
    auto status = _storageInterface->insertDocuments(txn, _nss, docsToInsert);
    fassertStatusOK(40161, status);

    _lastPushedTimestamp = ts;
    _count += numDocs;
    _size += std::accumulate(begin, end, 0U, [](const size_t& docSize, const Value& value) {
        return docSize + size_t(value.objsize());
    });
    _cvNoLongerEmpty.notify_all();
}

std::pair<BSONObj, Timestamp> OplogBufferCollection::addIdToDocument(const BSONObj& orig) {
    invariant(!orig.isEmpty());
    BSONObjBuilder bob;
    Timestamp ts = orig["ts"].timestamp();
    invariant(!ts.isNull());
    bob.append("_id", ts);
    bob.append(kOplogEntryFieldName, orig);
    return std::pair<BSONObj, Timestamp> {bob.obj(), ts};
}

void ReplicationRecoveryImpl::_recoverFromStableTimestamp(OperationContext* opCtx,
                                                          Timestamp stableTimestamp,
                                                          OpTime appliedThrough,
                                                          OpTime topOfOplog) {
    invariant(!stableTimestamp.isNull());
    invariant(!topOfOplog.isNull());
    const auto truncateAfterPoint = _consistencyMarkers->getOplogTruncateAfterPoint(opCtx);
    log() << "Recovering from stable timestamp: " << stableTimestamp
          << " (top of oplog: " << topOfOplog << ", appliedThrough: " << appliedThrough
          << ", TruncateAfter: " << truncateAfterPoint << ")";

    log() << "Starting recovery oplog application at the stable timestamp: " << stableTimestamp;
    _applyToEndOfOplog(opCtx, stableTimestamp, topOfOplog.getTimestamp());
}

//.........这里部分代码省略.........
        // If we don't cache the executor later, we are deleting it, so it must be deregistered.
        //
        // So, no matter what, deregister the executor.
        exec->deregisterExec();

        // Caller expects exceptions thrown in certain cases.
        if (PlanExecutor::FAILURE == state) {
            scoped_ptr<PlanStageStats> stats(exec->getStats());
            error() << "Plan executor error, stats: "
                    << Explain::statsToBSON(*stats);
            uasserted(17144, "Executor error: " + WorkingSetCommon::toStatusString(obj));
        }

        // TODO: Currently, chunk ranges are kept around until all ClientCursors created while the
        // chunk belonged on this node are gone. Separating chunk lifetime management from
        // ClientCursor should allow this check to go away.
        if (!shardingState.getVersion(nss.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
            // if the version changed during the query we might be missing some data and its safe to
            // send this as mongos can resend at this point
            throw SendStaleConfigException(nss.ns(), "version changed during initial query",
                                           shardingVersionAtStart,
                                           shardingState.getVersion(nss.ns()));
        }

        // Fill out curop based on query results. If we have a cursorid, we will fill out curop with
        // this cursorid later.
        long long ccId = 0;

        if (shouldSaveCursor(txn, collection, state, exec.get())) {
            // We won't use the executor until it's getMore'd.
            exec->saveState();

            // Allocate a new ClientCursor.  We don't have to worry about leaking it as it's
            // inserted into a global map by its ctor.
            ClientCursor* cc = new ClientCursor(collection->getCursorManager(),
                                                exec.release(),
                                                nss.ns(),
                                                pq.getOptions(),
                                                pq.getFilter());
            ccId = cc->cursorid();

            if (txn->getClient()->isInDirectClient()) {
                cc->setUnownedRecoveryUnit(txn->recoveryUnit());
            }
            else if (state == PlanExecutor::IS_EOF && pq.isTailable()) {
                // Don't stash the RU for tailable cursors at EOF, let them get a new RU on their
                // next getMore.
            }
            else {
                // We stash away the RecoveryUnit in the ClientCursor.  It's used for subsequent
                // getMore requests.  The calling OpCtx gets a fresh RecoveryUnit.
                txn->recoveryUnit()->abandonSnapshot();
                cc->setOwnedRecoveryUnit(txn->releaseRecoveryUnit());
                StorageEngine* storageEngine = getGlobalServiceContext()->getGlobalStorageEngine();
                invariant(txn->setRecoveryUnit(storageEngine->newRecoveryUnit(),
                                               OperationContext::kNotInUnitOfWork)
                          == OperationContext::kNotInUnitOfWork);
            }

            LOG(5) << "caching executor with cursorid " << ccId
                   << " after returning " << numResults << " results" << endl;

            // TODO document
            if (pq.isOplogReplay() && !slaveReadTill.isNull()) {
                cc->slaveReadTill(slaveReadTill);
            }

            // TODO document
            if (pq.isExhaust()) {
                curop.debug().exhaust = true;
            }

            cc->setPos(numResults);

            // If the query had a time limit, remaining time is "rolled over" to the cursor (for
            // use by future getmore ops).
            cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());

            endQueryOp(cc->getExecutor(), dbProfilingLevel, numResults, ccId, &curop);
        }
        else {
            LOG(5) << "Not caching executor but returning " << numResults << " results.\n";
            endQueryOp(exec.get(), dbProfilingLevel, numResults, ccId, &curop);
        }

        // Add the results from the query into the output buffer.
        result.appendData(bb.buf(), bb.len());
        bb.decouple();

        // Fill out the output buffer's header.
        QueryResult::View qr = result.header().view2ptr();
        qr.setCursorId(ccId);
        qr.setResultFlagsToOk();
        qr.msgdata().setOperation(opReply);
        qr.setStartingFrom(0);
        qr.setNReturned(numResults);

        // curop.debug().exhaust is set above.
        return curop.debug().exhaust ? nss.ns() : "";
    }

//.........这里部分代码省略.........
            }

            if (PlanExecutor::DEAD == state || PlanExecutor::FAILURE == state) {
                // Propagate this error to caller.
                if (PlanExecutor::FAILURE == state) {
                    scoped_ptr<PlanStageStats> stats(exec->getStats());
                    error() << "Plan executor error, stats: "
                            << Explain::statsToBSON(*stats);
                    uasserted(17406, "getMore executor error: " +
                              WorkingSetCommon::toStatusString(obj));
                }

                // In the old system tailable capped cursors would be killed off at the
                // cursorid level.  If a tailable capped cursor is nuked the cursorid
                // would vanish.
                //
                // In the new system they die and are cleaned up later (or time out).
                // So this is where we get to remove the cursorid.
                if (0 == numResults) {
                    resultFlags = ResultFlag_CursorNotFound;
                }
            }

            const bool shouldSaveCursor =
                    shouldSaveCursorGetMore(state, exec, isCursorTailable(cc));

            // In order to deregister a cursor, we need to be holding the DB + collection lock and
            // if the cursor is aggregation, we release these locks.
            if (cc->isAggCursor()) {
                invariant(NULL == ctx.get());
                unpinDBLock.reset(new Lock::DBLock(txn->lockState(), nss.db(), MODE_IS));
                unpinCollLock.reset(new Lock::CollectionLock(txn->lockState(), nss.ns(), MODE_IS));
            }

            // Our two possible ClientCursorPin cleanup paths are:
            // 1) If the cursor is not going to be saved, we call deleteUnderlying() on the pin.
            // 2) If the cursor is going to be saved, we simply let the pin go out of scope.  In
            //    this case, the pin's destructor will be invoked, which will call release() on the
            //    pin.  Because our ClientCursorPin is declared after our lock is declared, this
            //    will happen under the lock.
            if (!shouldSaveCursor) {
                ruSwapper.reset();
                ccPin.deleteUnderlying();

                // cc is now invalid, as is the executor
                cursorid = 0;
                cc = NULL;
                curop.debug().cursorExhausted = true;

                LOG(5) << "getMore NOT saving client cursor, ended with state "
                       << PlanExecutor::statestr(state)
                       << endl;
            }
            else {
                // Continue caching the ClientCursor.
                cc->incPos(numResults);
                exec->saveState();
                LOG(5) << "getMore saving client cursor ended with state "
                       << PlanExecutor::statestr(state)
                       << endl;

                if (PlanExecutor::IS_EOF == state && (queryOptions & QueryOption_CursorTailable)) {
                    if (!txn->getClient()->isInDirectClient()) {
                        // Don't stash the RU. Get a new one on the next getMore.
                        ruSwapper->dismiss();
                    }

                    if ((queryOptions & QueryOption_AwaitData)
                            && (numResults == 0)
                            && (pass < 1000)) {
                        // Bubble up to the AwaitData handling code in receivedGetMore which will
                        // try again.
                        return NULL;
                    }
                }

                // Possibly note slave's position in the oplog.
                if ((queryOptions & QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
                    cc->slaveReadTill(slaveReadTill);
                }

                exhaust = (queryOptions & QueryOption_Exhaust);

                // If the getmore had a time limit, remaining time is "rolled over" back to the
                // cursor (for use by future getmore ops).
                cc->setLeftoverMaxTimeMicros( curop.getRemainingMaxTimeMicros() );
            }
        }

        QueryResult::View qr = bb.buf();
        qr.msgdata().setLen(bb.len());
        qr.msgdata().setOperation(opReply);
        qr.setResultFlags(resultFlags);
        qr.setCursorId(cursorid);
        qr.setStartingFrom(startingResult);
        qr.setNReturned(numResults);
        bb.decouple();
        LOG(5) << "getMore returned " << numResults << " results\n";
        return qr;
    }

//.........这里部分代码省略.........
            // Must get the version before we call generateBatch in case a write comes in after
            // that call and before we call wait on the notifier.
            notifierVersion = notifier->getVersion();
        }

        PlanExecutor* exec = cc->getExecutor();
        exec->reattachToOperationContext(txn);
        exec->restoreState();
        PlanExecutor::ExecState state;

        generateBatch(ntoreturn, cc, &bb, &numResults, &slaveReadTill, &state);

        // If this is an await data cursor, and we hit EOF without generating any results, then
        // we block waiting for new data to arrive.
        if (isCursorAwaitData(cc) && state == PlanExecutor::IS_EOF && numResults == 0) {
            // Save the PlanExecutor and drop our locks.
            exec->saveState();
            ctx.reset();

            // Block waiting for data for up to 1 second.
            Seconds timeout(1);
            notifier->wait(notifierVersion, timeout);
            notifier.reset();

            // Set expected latency to match wait time. This makes sure the logs aren't spammed
            // by awaitData queries that exceed slowms due to blocking on the CappedInsertNotifier.
            curop.setExpectedLatencyMs(durationCount<Milliseconds>(timeout));

            // Reacquiring locks.
            ctx = make_unique<AutoGetCollectionForRead>(txn, nss);
            exec->restoreState();

            // We woke up because either the timed_wait expired, or there was more data. Either
            // way, attempt to generate another batch of results.
            generateBatch(ntoreturn, cc, &bb, &numResults, &slaveReadTill, &state);
        }

        // We have to do this before re-acquiring locks in the agg case because
        // shouldSaveCursorGetMore() can make a network call for agg cursors.
        //
        // TODO: Getting rid of PlanExecutor::isEOF() in favor of PlanExecutor::IS_EOF would mean
        // that this network operation is no longer necessary.
        const bool shouldSaveCursor = shouldSaveCursorGetMore(state, exec, isCursorTailable(cc));

        // In order to deregister a cursor, we need to be holding the DB + collection lock and
        // if the cursor is aggregation, we release these locks.
        if (cc->isAggCursor()) {
            invariant(NULL == ctx.get());
            unpinDBLock = make_unique<Lock::DBLock>(txn->lockState(), nss.db(), MODE_IS);
            unpinCollLock = make_unique<Lock::CollectionLock>(txn->lockState(), nss.ns(), MODE_IS);
        }

        // Our two possible ClientCursorPin cleanup paths are:
        // 1) If the cursor is not going to be saved, we call deleteUnderlying() on the pin.
        // 2) If the cursor is going to be saved, we simply let the pin go out of scope.  In
        //    this case, the pin's destructor will be invoked, which will call release() on the
        //    pin.  Because our ClientCursorPin is declared after our lock is declared, this
        //    will happen under the lock.
        if (!shouldSaveCursor) {
            ccPin.deleteUnderlying();

            // cc is now invalid, as is the executor
            cursorid = 0;
            cc = NULL;
            curop.debug().cursorExhausted = true;

            LOG(5) << "getMore NOT saving client cursor, ended with state "
                   << PlanExecutor::statestr(state) << endl;
        } else {
            // Continue caching the ClientCursor.
            cc->incPos(numResults);
            exec->saveState();
            exec->detachFromOperationContext();
            LOG(5) << "getMore saving client cursor ended with state "
                   << PlanExecutor::statestr(state) << endl;

            // Possibly note slave's position in the oplog.
            if ((cc->queryOptions() & QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
                cc->slaveReadTill(slaveReadTill);
            }

            *exhaust = cc->queryOptions() & QueryOption_Exhaust;

            // If the getmore had a time limit, remaining time is "rolled over" back to the
            // cursor (for use by future getmore ops).
            cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
        }
    }

    QueryResult::View qr = bb.buf();
    qr.msgdata().setLen(bb.len());
    qr.msgdata().setOperation(opReply);
    qr.setResultFlags(resultFlags);
    qr.setCursorId(cursorid);
    qr.setStartingFrom(startingResult);
    qr.setNReturned(numResults);
    bb.decouple();
    LOG(5) << "getMore returned " << numResults << " results\n";
    return qr;
}

//.........这里部分代码省略.........

        // Count the result.
        ++numResults;

        // Possibly note slave's position in the oplog.
        if (pq.isOplogReplay()) {
            BSONElement e = obj["ts"];
            if (Date == e.type() || bsonTimestamp == e.type()) {
                slaveReadTill = e.timestamp();
            }
        }

        if (FindCommon::enoughForFirstBatch(pq, numResults)) {
            LOG(5) << "Enough for first batch, wantMore=" << pq.wantMore()
                   << " ntoreturn=" << pq.getNToReturn().value_or(0) << " numResults=" << numResults
                   << endl;
            break;
        }
    }

    // If we cache the executor later, we want to deregister it as it receives notifications
    // anyway by virtue of being cached.
    //
    // If we don't cache the executor later, we are deleting it, so it must be deregistered.
    //
    // So, no matter what, deregister the executor.
    exec->deregisterExec();

    // Caller expects exceptions thrown in certain cases.
    if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
        error() << "Plan executor error during find: " << PlanExecutor::statestr(state)
                << ", stats: " << Explain::getWinningPlanStats(exec.get());
        uasserted(17144, "Executor error: " + WorkingSetCommon::toStatusString(obj));
    }

    // Before saving the cursor, ensure that whatever plan we established happened with the expected
    // collection version
    auto css = CollectionShardingState::get(txn, nss);
    css->checkShardVersionOrThrow(txn);

    // Fill out curop based on query results. If we have a cursorid, we will fill out curop with
    // this cursorid later.
    long long ccId = 0;

    if (shouldSaveCursor(txn, collection, state, exec.get())) {
        // We won't use the executor until it's getMore'd.
        exec->saveState();
        exec->detachFromOperationContext();

        // Allocate a new ClientCursor.  We don't have to worry about leaking it as it's
        // inserted into a global map by its ctor.
        ClientCursor* cc =
            new ClientCursor(collection->getCursorManager(),
                             exec.release(),
                             nss.ns(),
                             txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
                             pq.getOptions(),
                             pq.getFilter());
        ccId = cc->cursorid();

        LOG(5) << "caching executor with cursorid " << ccId << " after returning " << numResults
               << " results" << endl;

        // TODO document
        if (pq.isOplogReplay() && !slaveReadTill.isNull()) {
            cc->slaveReadTill(slaveReadTill);
        }

        // TODO document
        if (pq.isExhaust()) {
            curop.debug().exhaust = true;
        }

        cc->setPos(numResults);

        // If the query had a time limit, remaining time is "rolled over" to the cursor (for
        // use by future getmore ops).
        cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());

        endQueryOp(txn, collection, *cc->getExecutor(), dbProfilingLevel, numResults, ccId);
    } else {
        LOG(5) << "Not caching executor but returning " << numResults << " results.\n";
        endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, ccId);
    }

    // Add the results from the query into the output buffer.
    result.appendData(bb.buf(), bb.len());
    bb.decouple();

    // Fill out the output buffer's header.
    QueryResult::View qr = result.header().view2ptr();
    qr.setCursorId(ccId);
    qr.setResultFlagsToOk();
    qr.msgdata().setOperation(opReply);
    qr.setStartingFrom(0);
    qr.setNReturned(numResults);

    // curop.debug().exhaust is set above.
    return curop.debug().exhaust ? nss.ns() : "";
}

void ReplicationRecoveryImpl::_applyToEndOfOplog(OperationContext* opCtx,
                                                 const Timestamp& oplogApplicationStartPoint,
                                                 const Timestamp& topOfOplog) {
    invariant(!oplogApplicationStartPoint.isNull());
    invariant(!topOfOplog.isNull());

    // Check if we have any unapplied ops in our oplog. It is important that this is done after
    // deleting the ragged end of the oplog.
    if (oplogApplicationStartPoint == topOfOplog) {
        log() << "No oplog entries to apply for recovery. Start point is at the top of the oplog.";
        return;  // We've applied all the valid oplog we have.
    } else if (oplogApplicationStartPoint > topOfOplog) {
        severe() << "Applied op " << oplogApplicationStartPoint.toBSON()
                 << " not found. Top of oplog is " << topOfOplog.toBSON() << '.';
        fassertFailedNoTrace(40313);
    }

    log() << "Replaying stored operations from " << oplogApplicationStartPoint.toBSON()
          << " (exclusive) to " << topOfOplog.toBSON() << " (inclusive).";

    OplogBufferLocalOplog oplogBuffer(oplogApplicationStartPoint);
    oplogBuffer.startup(opCtx);

    RecoveryOplogApplierStats stats;

    auto writerPool = OplogApplier::makeWriterPool();
    OplogApplier::Options options;
    options.allowNamespaceNotFoundErrorsOnCrudOps = true;
    options.skipWritesToOplog = true;
    // During replication recovery, the stableTimestampForRecovery refers to the stable timestamp
    // from which we replay the oplog.
    // For startup recovery, this will be the recovery timestamp, which is the stable timestamp that
    // the storage engine recovered to on startup. For rollback recovery, this will be the last
    // stable timestamp, returned when we call recoverToStableTimestamp.
    // We keep track of this for prepared transactions so that when we apply a commitTransaction
    // oplog entry, we can check if it occurs before or after the stable timestamp and decide
    // whether the operations would have already been reflected in the data.
    options.stableTimestampForRecovery = oplogApplicationStartPoint;
    OplogApplierImpl oplogApplier(nullptr,
                                  &oplogBuffer,
                                  &stats,
                                  nullptr,
                                  _consistencyMarkers,
                                  _storageInterface,
                                  options,
                                  writerPool.get());

    OplogApplier::BatchLimits batchLimits;
    batchLimits.bytes = OplogApplier::calculateBatchLimitBytes(opCtx, _storageInterface);
    batchLimits.ops = OplogApplier::getBatchLimitOperations();

    OpTime applyThroughOpTime;
    OplogApplier::Operations batch;
    while (
        !(batch = fassert(50763, oplogApplier.getNextApplierBatch(opCtx, batchLimits))).empty()) {
        applyThroughOpTime = uassertStatusOK(oplogApplier.multiApply(opCtx, std::move(batch)));
    }
    stats.complete(applyThroughOpTime);
    invariant(oplogBuffer.isEmpty(),
              str::stream() << "Oplog buffer not empty after applying operations. Last operation "
                               "applied with optime: "
                            << applyThroughOpTime.toBSON());
    invariant(applyThroughOpTime.getTimestamp() == topOfOplog,
              str::stream() << "Did not apply to top of oplog. Applied through: "
                            << applyThroughOpTime.toString()
                            << ". Top of oplog: "
                            << topOfOplog.toString());
    oplogBuffer.shutdown(opCtx);

    // We may crash before setting appliedThrough. If we have a stable checkpoint, we will recover
    // to that checkpoint at a replication consistent point, and applying the oplog is safe.
    // If we don't have a stable checkpoint, then we must be in startup recovery, and not rollback
    // recovery, because we only roll back to a stable timestamp when we have a stable checkpoint.
    // Startup recovery from an unstable checkpoint only ever applies a single batch and it is safe
    // to replay the batch from any point.
    _consistencyMarkers->setAppliedThrough(opCtx, applyThroughOpTime);
}