C++ PlanExecutor::detachFromOperationContext方法代码示例

//.........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.........
        if (!shardingState->getVersion(nss.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
            // Version changed while retrieving a PlanExecutor. Terminate the operation,
            // signaling that mongos should retry.
            throw SendStaleConfigException(nss.ns(),
                                           "version changed during find command",
                                           shardingVersionAtStart,
                                           shardingState->getVersion(nss.ns()));
        }

        if (!collection) {
            // No collection. Just fill out curop indicating that there were zero results and
            // there is no ClientCursor id, and then return.
            const long long numResults = 0;
            const CursorId cursorId = 0;
            endQueryOp(txn, *exec, dbProfilingLevel, numResults, cursorId);
            appendCursorResponseObject(cursorId, nss.ns(), BSONArray(), &result);
            return true;
        }

        const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();

        // 4) If possible, register the execution plan inside a ClientCursor, and pin that
        // cursor. In this case, ownership of the PlanExecutor is transferred to the
        // ClientCursor, and 'exec' becomes null.
        //
        // First unregister the PlanExecutor so it can be re-registered with ClientCursor.
        exec->deregisterExec();

        // Create a ClientCursor containing this plan executor. We don't have to worry
        // about leaking it as it's inserted into a global map by its ctor.
        ClientCursor* cursor =
            new ClientCursor(collection->getCursorManager(),
                             exec.release(),
                             nss.ns(),
                             txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
                             pq.getOptions(),
                             pq.getFilter());
        CursorId cursorId = cursor->cursorid();
        ClientCursorPin ccPin(collection->getCursorManager(), cursorId);

        // On early return, get rid of the the cursor.
        ScopeGuard cursorFreer = MakeGuard(&ClientCursorPin::deleteUnderlying, ccPin);

        invariant(!exec);
        PlanExecutor* cursorExec = cursor->getExecutor();

        // 5) Stream query results, adding them to a BSONArray as we go.
        BSONArrayBuilder firstBatch;
        BSONObj obj;
        PlanExecutor::ExecState state;
        long long numResults = 0;
        while (!enoughForFirstBatch(pq, numResults, firstBatch.len()) &&
               PlanExecutor::ADVANCED == (state = cursorExec->getNext(&obj, NULL))) {
            // If adding this object will cause us to exceed the BSON size limit, then we stash
            // it for later.
            if (firstBatch.len() + obj.objsize() > BSONObjMaxUserSize && numResults > 0) {
                cursorExec->enqueue(obj);
                break;
            }

            // Add result to output buffer.
            firstBatch.append(obj);
            numResults++;
        }

        // Throw an assertion if query execution fails for any reason.
        if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
            const std::unique_ptr<PlanStageStats> stats(cursorExec->getStats());
            error() << "Plan executor error during find command: " << PlanExecutor::statestr(state)
                    << ", stats: " << Explain::statsToBSON(*stats);

            return appendCommandStatus(result,
                                       Status(ErrorCodes::OperationFailed,
                                              str::stream()
                                                  << "Executor error during find command: "
                                                  << WorkingSetCommon::toStatusString(obj)));
        }

        // 6) Set up the cursor for getMore.
        if (shouldSaveCursor(txn, collection, state, cursorExec)) {
            // State will be restored on getMore.
            cursorExec->saveState();
            cursorExec->detachFromOperationContext();

            cursor->setLeftoverMaxTimeMicros(CurOp::get(txn)->getRemainingMaxTimeMicros());
            cursor->setPos(numResults);
        } else {
            cursorId = 0;
        }

        // Fill out curop based on the results.
        endQueryOp(txn, *cursorExec, dbProfilingLevel, numResults, cursorId);

        // 7) Generate the response object to send to the client.
        appendCursorResponseObject(cursorId, nss.ns(), firstBatch.arr(), &result);
        if (cursorId) {
            cursorFreer.Dismiss();
        }
        return true;
    }

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

        std::unique_ptr<PlanExecutor> exec = std::move(statusWithPlanExecutor.getValue());

        if (!collection) {
            // No collection. Just fill out curop indicating that there were zero results and
            // there is no ClientCursor id, and then return.
            const long long numResults = 0;
            const CursorId cursorId = 0;
            endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, cursorId);
            appendCursorResponseObject(cursorId, nss.ns(), BSONArray(), &result);
            return true;
        }

        const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();

        // Stream query results, adding them to a BSONArray as we go.
        BSONArrayBuilder firstBatch;
        BSONObj obj;
        PlanExecutor::ExecState state = PlanExecutor::ADVANCED;
        long long numResults = 0;
        while (!FindCommon::enoughForFirstBatch(pq, numResults, firstBatch.len()) &&
               PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
            // If adding this object will cause us to exceed the BSON size limit, then we stash
            // it for later.
            if (firstBatch.len() + obj.objsize() > BSONObjMaxUserSize && numResults > 0) {
                exec->enqueue(obj);
                break;
            }

            // Add result to output buffer.
            firstBatch.append(obj);
            numResults++;
        }

        // Throw an assertion if query execution fails for any reason.
        if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
            const std::unique_ptr<PlanStageStats> stats(exec->getStats());
            error() << "Plan executor error during find command: " << PlanExecutor::statestr(state)
                    << ", stats: " << Explain::statsToBSON(*stats);

            return appendCommandStatus(result,
                                       Status(ErrorCodes::OperationFailed,
                                              str::stream()
                                                  << "Executor error during find command: "
                                                  << 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)) {
            // Version changed while retrieving a PlanExecutor. Terminate the operation,
            // signaling that mongos should retry.
            throw SendStaleConfigException(nss.ns(),
                                           "version changed during find command",
                                           shardingVersionAtStart,
                                           shardingState->getVersion(nss.ns()));
        }

        // Set up the cursor for getMore.
        CursorId cursorId = 0;
        if (shouldSaveCursor(txn, collection, state, exec.get())) {
            // Register the execution plan inside a ClientCursor. Ownership of the PlanExecutor is
            // transferred to the ClientCursor.
            //
            // First unregister the PlanExecutor so it can be re-registered with ClientCursor.
            exec->deregisterExec();

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

            invariant(!exec);
            PlanExecutor* cursorExec = cursor->getExecutor();

            // State will be restored on getMore.
            cursorExec->saveState();
            cursorExec->detachFromOperationContext();

            cursor->setLeftoverMaxTimeMicros(CurOp::get(txn)->getRemainingMaxTimeMicros());
            cursor->setPos(numResults);

            // Fill out curop based on the results.
            endQueryOp(txn, collection, *cursorExec, dbProfilingLevel, numResults, cursorId);
        } else {
            endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, cursorId);
        }

        // Generate the response object to send to the client.
        appendCursorResponseObject(cursorId, nss.ns(), firstBatch.arr(), &result);
        return true;
    }

//.........这里部分代码省略.........
            getExecutorFind(txn, collection, nss, std::move(cq), PlanExecutor::YIELD_AUTO);
        if (!statusWithPlanExecutor.isOK()) {
            return appendCommandStatus(result, statusWithPlanExecutor.getStatus());
        }

        std::unique_ptr<PlanExecutor> exec = std::move(statusWithPlanExecutor.getValue());

        {
            stdx::lock_guard<Client>(*txn->getClient());
            CurOp::get(txn)->setPlanSummary_inlock(Explain::getPlanSummary(exec.get()));
        }

        if (!collection) {
            // No collection. Just fill out curop indicating that there were zero results and
            // there is no ClientCursor id, and then return.
            const long long numResults = 0;
            const CursorId cursorId = 0;
            endQueryOp(txn, collection, *exec, numResults, cursorId);
            appendCursorResponseObject(cursorId, nss.ns(), BSONArray(), &result);
            return true;
        }

        const QueryRequest& originalQR = exec->getCanonicalQuery()->getQueryRequest();

        // Stream query results, adding them to a BSONArray as we go.
        CursorResponseBuilder firstBatch(/*isInitialResponse*/ true, &result);
        BSONObj obj;
        PlanExecutor::ExecState state = PlanExecutor::ADVANCED;
        long long numResults = 0;
        while (!FindCommon::enoughForFirstBatch(originalQR, numResults) &&
               PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
            // If we can't fit this result inside the current batch, then we stash it for later.
            if (!FindCommon::haveSpaceForNext(obj, numResults, firstBatch.bytesUsed())) {
                exec->enqueue(obj);
                break;
            }

            // Add result to output buffer.
            firstBatch.append(obj);
            numResults++;
        }

        // Throw an assertion if query execution fails for any reason.
        if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
            firstBatch.abandon();
            error() << "Plan executor error during find command: " << PlanExecutor::statestr(state)
                    << ", stats: " << Explain::getWinningPlanStats(exec.get());

            return appendCommandStatus(result,
                                       Status(ErrorCodes::OperationFailed,
                                              str::stream()
                                                  << "Executor error during find command: "
                                                  << 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);

        // Set up the cursor for getMore.
        CursorId cursorId = 0;
        if (shouldSaveCursor(txn, collection, state, exec.get())) {
            // Register the execution plan inside a ClientCursor. Ownership of the PlanExecutor is
            // transferred to the ClientCursor.
            //
            // First unregister the PlanExecutor so it can be re-registered with ClientCursor.
            exec->deregisterExec();

            // Create a ClientCursor containing this plan executor. We don't have to worry about
            // leaking it as it's inserted into a global map by its ctor.
            ClientCursor* cursor =
                new ClientCursor(collection->getCursorManager(),
                                 exec.release(),
                                 nss.ns(),
                                 txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
                                 originalQR.getOptions(),
                                 cmdObj.getOwned());
            cursorId = cursor->cursorid();

            invariant(!exec);
            PlanExecutor* cursorExec = cursor->getExecutor();

            // State will be restored on getMore.
            cursorExec->saveState();
            cursorExec->detachFromOperationContext();

            cursor->setLeftoverMaxTimeMicros(txn->getRemainingMaxTimeMicros());
            cursor->setPos(numResults);

            // Fill out curop based on the results.
            endQueryOp(txn, collection, *cursorExec, numResults, cursorId);
        } else {
            endQueryOp(txn, collection, *exec, numResults, cursorId);
        }

        // Generate the response object to send to the client.
        firstBatch.done(cursorId, nss.ns());
        return true;
    }

//.........这里部分代码省略.........
            if (isCursorAwaitData(cursor)) {
                Seconds awaitDataTimeout(1);
                CurOp::get(txn)->setMaxTimeMicros(durationCount<Microseconds>(awaitDataTimeout));
            } else {
                CurOp::get(txn)->setMaxTimeMicros(cursor->getLeftoverMaxTimeMicros());
            }
        }
        txn->checkForInterrupt();  // May trigger maxTimeAlwaysTimeOut fail point.

        if (cursor->isAggCursor()) {
            // Agg cursors handle their own locking internally.
            ctx.reset();  // unlocks
        }

        PlanExecutor* exec = cursor->getExecutor();
        exec->reattachToOperationContext(txn);
        exec->restoreState();

        // If we're tailing a capped collection, retrieve a monotonically increasing insert
        // counter.
        uint64_t lastInsertCount = 0;
        if (isCursorAwaitData(cursor)) {
            invariant(ctx->getCollection()->isCapped());
            lastInsertCount = ctx->getCollection()->getCappedInsertNotifier()->getCount();
        }

        CursorId respondWithId = 0;
        BSONArrayBuilder nextBatch;
        BSONObj obj;
        PlanExecutor::ExecState state;
        long long numResults = 0;
        Status batchStatus = generateBatch(cursor, request, &nextBatch, &state, &numResults);
        if (!batchStatus.isOK()) {
            return appendCommandStatus(result, batchStatus);
        }

        // 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(cursor) && state == PlanExecutor::IS_EOF && numResults == 0) {
            // Retrieve the notifier which we will wait on until new data arrives. We make sure
            // to do this in the lock because once we drop the lock it is possible for the
            // collection to become invalid. The notifier itself will outlive the collection if
            // the collection is dropped, as we keep a shared_ptr to it.
            auto notifier = ctx->getCollection()->getCappedInsertNotifier();

            // Save the PlanExecutor and drop our locks.
            exec->saveState();
            ctx.reset();

            // Block waiting for data.
            Microseconds timeout(CurOp::get(txn)->getRemainingMaxTimeMicros());
            notifier->waitForInsert(lastInsertCount, timeout);
            notifier.reset();

            ctx.reset(new AutoGetCollectionForRead(txn, request.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.
            batchStatus = generateBatch(cursor, request, &nextBatch, &state, &numResults);
            if (!batchStatus.isOK()) {
                return appendCommandStatus(result, batchStatus);
            }
        }

        if (shouldSaveCursorGetMore(state, exec, isCursorTailable(cursor))) {
            respondWithId = request.cursorid;

            exec->saveState();
            exec->detachFromOperationContext();

            // If maxTimeMS was set directly on the getMore rather than being rolled over
            // from a previous find, then don't roll remaining micros over to the next
            // getMore.
            if (!hasOwnMaxTime) {
                cursor->setLeftoverMaxTimeMicros(CurOp::get(txn)->getRemainingMaxTimeMicros());
            }

            cursor->incPos(numResults);
        } else {
            CurOp::get(txn)->debug().cursorExhausted = true;
        }

        appendGetMoreResponseObject(respondWithId, request.nss.ns(), nextBatch.arr(), &result);

        if (respondWithId) {
            cursorFreer.Dismiss();

            // If we are operating on an aggregation cursor, then we dropped our collection lock
            // earlier and need to reacquire it in order to clean up our ClientCursorPin.
            if (cursor->isAggCursor()) {
                invariant(NULL == ctx.get());
                unpinDBLock.reset(new Lock::DBLock(txn->lockState(), request.nss.db(), MODE_IS));
                unpinCollLock.reset(
                    new Lock::CollectionLock(txn->lockState(), request.nss.ns(), MODE_IS));
            }
        }

        return true;
    }