StatusWith<executor::TaskExecutor::EventHandle> AsyncResultsMerger::nextEvent() {
stdx::lock_guard<stdx::mutex> lk(_mutex);
if (_lifecycleState != kAlive) {
// Can't schedule further network operations if the ARM is being killed.
return Status(ErrorCodes::IllegalOperation,
"nextEvent() called on a killed AsyncResultsMerger");
}
if (_currentEvent.isValid()) {
// We can't make a new event if there's still an unsignaled one, as every event must
// eventually be signaled.
return Status(ErrorCodes::IllegalOperation,
"nextEvent() called before an outstanding event was signaled");
}
// Schedule remote work on hosts for which we need more results.
for (size_t i = 0; i < _remotes.size(); ++i) {
auto& remote = _remotes[i];
// It is illegal to call this method if there is an error received from any shard.
invariant(remote.status.isOK());
if (!remote.hasNext() && !remote.exhausted() && !remote.cbHandle.isValid()) {
// If we already have established a cursor with this remote, and there is no outstanding
// request for which we have a valid callback handle, then schedule work to retrieve the
// next batch.
auto nextBatchStatus = askForNextBatch_inlock(i);
if (!nextBatchStatus.isOK()) {
return nextBatchStatus;
}
}
}
auto eventStatus = _executor->makeEvent();
if (!eventStatus.isOK()) {
return eventStatus;
}
auto eventToReturn = eventStatus.getValue();
_currentEvent = eventToReturn;
// It's possible that after we told the caller we had no ready results but before the call to
// this method, new results became available. In this case we have to signal the event right
// away so that the caller will not block.
signalCurrentEventIfReady_inlock();
return eventToReturn;
}
void AsyncResultsMerger::handleBatchResponse(
const executor::TaskExecutor::RemoteCommandCallbackArgs& cbData, size_t remoteIndex) {
stdx::lock_guard<stdx::mutex> lk(_mutex);
auto& remote = _remotes[remoteIndex];
// Clear the callback handle. This indicates that we are no longer waiting on a response from
// 'remote'.
remote.cbHandle = executor::TaskExecutor::CallbackHandle();
// If we're in the process of shutting down then there's no need to process the batch.
if (_lifecycleState != kAlive) {
invariant(_lifecycleState == kKillStarted);
// Make sure to wake up anyone waiting on '_currentEvent' if we're shutting down.
signalCurrentEventIfReady_inlock();
// Make a best effort to parse the response and retrieve the cursor id. We need the cursor
// id in order to issue a killCursors command against it.
if (cbData.response.isOK()) {
auto cursorResponse = parseCursorResponse(cbData.response.getValue().data, remote);
if (cursorResponse.isOK()) {
remote.cursorId = cursorResponse.getValue().getCursorId();
}
}
// If we're killed and we're not waiting on any more batches to come back, then we are ready
// to kill the cursors on the remote hosts and clean up this cursor. Schedule the
// killCursors command and signal that this cursor is safe now safe to destroy. We have to
// promise not to touch any members of this class because 'this' could become invalid as
// soon as we signal the event.
if (!haveOutstandingBatchRequests_inlock()) {
// If the event handle is invalid, then the executor is in the middle of shutting down,
// and we can't schedule any more work for it to complete.
if (_killCursorsScheduledEvent.isValid()) {
scheduleKillCursors_inlock();
_executor->signalEvent(_killCursorsScheduledEvent);
}
_lifecycleState = kKillComplete;
}
return;
}
// Early return from this point on signal anyone waiting on an event, if ready() is true.
ScopeGuard signaller = MakeGuard(&AsyncResultsMerger::signalCurrentEventIfReady_inlock, this);
StatusWith<CursorResponse> cursorResponseStatus(
cbData.response.isOK() ? parseCursorResponse(cbData.response.getValue().data, remote)
: cbData.response.getStatus());
if (!cursorResponseStatus.isOK()) {
auto shard = remote.getShard();
if (!shard) {
remote.status = Status(cursorResponseStatus.getStatus().code(),
str::stream() << "Could not find shard " << *remote.shardId
<< " containing host "
<< remote.getTargetHost().toString());
} else {
shard->updateReplSetMonitor(remote.getTargetHost(), cursorResponseStatus.getStatus());
// Retry initial cursor establishment if possible. Never retry getMores to avoid
// accidentally skipping results.
if (!remote.cursorId && remote.retryCount < kMaxNumFailedHostRetryAttempts &&
shard->isRetriableError(cursorResponseStatus.getStatus().code(),
Shard::RetryPolicy::kIdempotent)) {
invariant(remote.shardId);
LOG(1) << "Initial cursor establishment failed with retriable error and will be "
"retried"
<< causedBy(redact(cursorResponseStatus.getStatus()));
++remote.retryCount;
// Since we potentially updated the targeter that the last host it chose might be
// faulty, the call below may end up getting a different host.
remote.status = askForNextBatch_inlock(remoteIndex);
if (remote.status.isOK()) {
return;
}
// If we end up here, it means we failed to schedule the retry request, which is a
// more
// severe error that should not be retried. Just pass through to the error handling
// logic below.
} else {
remote.status = cursorResponseStatus.getStatus();
}
}
// Unreachable host errors are swallowed if the 'allowPartialResults' option is set. We
// remove the unreachable host entirely from consideration by marking it as exhausted.
if (_params.isAllowPartialResults) {
remote.status = Status::OK();
// Clear the results buffer and cursor id.
std::queue<BSONObj> emptyBuffer;
std::swap(remote.docBuffer, emptyBuffer);
remote.cursorId = 0;
}
return;
}
// Cursor id successfully established.
auto cursorResponse = std::move(cursorResponseStatus.getValue());
remote.cursorId = cursorResponse.getCursorId();
remote.initialCmdObj = boost::none;
for (const auto& obj : cursorResponse.getBatch()) {
// If there's a sort, we're expecting the remote node to give us back a sort key.
if (!_params.sort.isEmpty() &&
obj[ClusterClientCursorParams::kSortKeyField].type() != BSONType::Object) {
remote.status = Status(ErrorCodes::InternalError,
str::stream() << "Missing field '"
<< ClusterClientCursorParams::kSortKeyField
<< "' in document: "
<< obj);
return;
}
remote.docBuffer.push(obj);
++remote.fetchedCount;
}
// If we're doing a sorted merge, then we have to make sure to put this remote onto the
// merge queue.
if (!_params.sort.isEmpty() && !cursorResponse.getBatch().empty()) {
_mergeQueue.push(remoteIndex);
}
// If the cursor is tailable and we just received an empty batch, the next return value should
// be boost::none in order to indicate the end of the batch.
if (_params.isTailable && !remote.hasNext()) {
_eofNext = true;
}
// If even after receiving this batch we still don't have anything buffered (i.e. the batchSize
// was zero), then can schedule work to retrieve the next batch right away.
//
// We do not ask for the next batch if the cursor is tailable, as batches received from remote
// tailable cursors should be passed through to the client without asking for more batches.
if (!_params.isTailable && !remote.hasNext() && !remote.exhausted()) {
remote.status = askForNextBatch_inlock(remoteIndex);
if (!remote.status.isOK()) {
return;
}
}
// ScopeGuard requires dismiss on success, but we want waiter to be signalled on success as
// well as failure.
signaller.Dismiss();
signalCurrentEventIfReady_inlock();
}
void AsyncResultsMerger::handleBatchResponse(
const executor::TaskExecutor::RemoteCommandCallbackArgs& cbData,
OperationContext* opCtx,
size_t remoteIndex) {
stdx::lock_guard<stdx::mutex> lk(_mutex);
auto& remote = _remotes[remoteIndex];
// Clear the callback handle. This indicates that we are no longer waiting on a response from
// 'remote'.
remote.cbHandle = executor::TaskExecutor::CallbackHandle();
// If we're in the process of shutting down then there's no need to process the batch.
if (_lifecycleState != kAlive) {
invariant(_lifecycleState == kKillStarted);
// Make sure to wake up anyone waiting on '_currentEvent' if we're shutting down.
signalCurrentEventIfReady_inlock();
// If we're killed and we're not waiting on any more batches to come back, then we are ready
// to kill the cursors on the remote hosts and clean up this cursor. Schedule the
// killCursors command and signal that this cursor is safe now safe to destroy. We have to
// promise not to touch any members of this class because 'this' could become invalid as
// soon as we signal the event.
if (!haveOutstandingBatchRequests_inlock()) {
// If the event handle is invalid, then the executor is in the middle of shutting down,
// and we can't schedule any more work for it to complete.
if (_killCursorsScheduledEvent.isValid()) {
scheduleKillCursors_inlock(opCtx);
_executor->signalEvent(_killCursorsScheduledEvent);
}
_lifecycleState = kKillComplete;
}
return;
}
// Early return from this point on signal anyone waiting on an event, if ready() is true.
ScopeGuard signaller = MakeGuard(&AsyncResultsMerger::signalCurrentEventIfReady_inlock, this);
StatusWith<CursorResponse> cursorResponseStatus(
cbData.response.isOK() ? parseCursorResponse(cbData.response.data, remote)
: cbData.response.status);
if (!cursorResponseStatus.isOK()) {
auto shard = remote.getShard();
if (!shard) {
remote.status = Status(cursorResponseStatus.getStatus().code(),
str::stream() << "Could not find shard containing host "
<< remote.getTargetHost().toString());
} else {
shard->updateReplSetMonitor(remote.getTargetHost(), cursorResponseStatus.getStatus());
remote.status = cursorResponseStatus.getStatus();
}
// Unreachable host errors are swallowed if the 'allowPartialResults' option is set. We
// remove the unreachable host entirely from consideration by marking it as exhausted.
if (_params->isAllowPartialResults) {
remote.status = Status::OK();
// Clear the results buffer and cursor id.
std::queue<ClusterQueryResult> emptyBuffer;
std::swap(remote.docBuffer, emptyBuffer);
remote.cursorId = 0;
}
return;
}
// Response successfully received.
auto cursorResponse = std::move(cursorResponseStatus.getValue());
// Update the cursorId; it is sent as '0' when the cursor has been exhausted on the shard.
remote.cursorId = cursorResponse.getCursorId();
// Save the batch in the remote's buffer.
if (!addBatchToBuffer(remoteIndex, cursorResponse.getBatch())) {
return;
}
// If the cursor is tailable and we just received an empty batch, the next return value should
// be boost::none in order to indicate the end of the batch.
// (Note: tailable cursors are only valid on unsharded collections, so the end of the batch from
// one shard means the end of the overall batch).
if (_params->isTailable && !remote.hasNext()) {
_eofNext = true;
}
// If even after receiving this batch we still don't have anything buffered (i.e. the batchSize
// was zero), then can schedule work to retrieve the next batch right away.
//
// We do not ask for the next batch if the cursor is tailable, as batches received from remote
// tailable cursors should be passed through to the client without asking for more batches.
if (!_params->isTailable && !remote.hasNext() && !remote.exhausted()) {
remote.status = askForNextBatch_inlock(opCtx, remoteIndex);
if (!remote.status.isOK()) {
return;
}
}
// ScopeGuard requires dismiss on success, but we want waiter to be signalled on success as
// well as failure.
signaller.Dismiss();
signalCurrentEventIfReady_inlock();
}
void AsyncResultsMerger::handleBatchResponse(
const executor::TaskExecutor::RemoteCommandCallbackArgs& cbData, size_t remoteIndex) {
stdx::lock_guard<stdx::mutex> lk(_mutex);
auto& remote = _remotes[remoteIndex];
// Clear the callback handle. This indicates that we are no longer waiting on a response from
// 'remote'.
remote.cbHandle = executor::TaskExecutor::CallbackHandle();
// If we're in the process of shutting down then there's no need to process the batch.
if (_lifecycleState != kAlive) {
invariant(_lifecycleState == kKillStarted);
// Make sure to wake up anyone waiting on '_currentEvent' if we're shutting down.
signalCurrentEventIfReady_inlock();
// If we're killed and we're not waiting on any more batches to come back, then we are ready
// to kill the cursors on the remote hosts and clean up this cursor. Schedule the
// killCursors command and signal that this cursor is safe now safe to destroy. We have to
// promise not to touch any members of this class because 'this' could become invalid as
// soon as we signal the event.
if (!haveOutstandingBatchRequests_inlock()) {
// If the event handle is invalid, then the executor is in the middle of shutting down,
// and we can't schedule any more work for it to complete.
if (_killCursorsScheduledEvent.isValid()) {
scheduleKillCursors_inlock();
_executor->signalEvent(_killCursorsScheduledEvent);
}
_lifecycleState = kKillComplete;
}
return;
}
// Early return from this point on signal anyone waiting on an event, if ready() is true.
ScopeGuard signaller = MakeGuard(&AsyncResultsMerger::signalCurrentEventIfReady_inlock, this);
if (!cbData.response.isOK()) {
remote.status = cbData.response.getStatus();
// If we failed to retrieve the batch because we couldn't contact the remote, we notify that
// targeter that the host is unreachable. The caller can then retry on a new host.
if (remote.status == ErrorCodes::HostUnreachable && remote.shardId) {
auto shard = _params.shardRegistry->getShard(_params.txn, *remote.shardId);
if (!shard) {
remote.status =
Status(ErrorCodes::HostUnreachable,
str::stream() << "Could not find shard " << *remote.shardId
<< " containing host " << remote.hostAndPort.toString());
} else {
shard->getTargeter()->markHostUnreachable(remote.hostAndPort);
}
}
return;
}
auto getMoreParseStatus = CursorResponse::parseFromBSON(cbData.response.getValue().data);
if (!getMoreParseStatus.isOK()) {
remote.status = getMoreParseStatus.getStatus();
return;
}
auto cursorResponse = getMoreParseStatus.getValue();
// If we have a cursor established, and we get a non-zero cursorid that is not equal to the
// established cursorid, we will fail the operation.
if (remote.cursorId && cursorResponse.cursorId != 0 &&
*remote.cursorId != cursorResponse.cursorId) {
remote.status = Status(ErrorCodes::BadValue,
str::stream() << "Expected cursorid " << *remote.cursorId
<< " but received " << cursorResponse.cursorId);
return;
}
remote.cursorId = cursorResponse.cursorId;
remote.cmdObj = boost::none;
for (const auto& obj : cursorResponse.batch) {
// If there's a sort, we're expecting the remote node to give us back a sort key.
if (!_params.sort.isEmpty() &&
obj[ClusterClientCursorParams::kSortKeyField].type() != BSONType::Object) {
remote.status = Status(ErrorCodes::InternalError,
str::stream() << "Missing field '"
<< ClusterClientCursorParams::kSortKeyField
<< "' in document: " << obj);
return;
}
remote.docBuffer.push(obj);
++remote.fetchedCount;
}
// If we're doing a sorted merge, then we have to make sure to put this remote onto the
// merge queue.
if (!_params.sort.isEmpty() && !cursorResponse.batch.empty()) {
_mergeQueue.push(remoteIndex);
}
// If the cursor is tailable and we just received an empty batch, the next return value should
// be boost::none in order to indicate the end of the batch.
if (_params.isTailable && !remote.hasNext()) {
_eofNext = true;
}
// If even after receiving this batch we still don't have anything buffered (i.e. the batchSize
// was zero), then can schedule work to retrieve the next batch right away.
//
// We do not ask for the next batch if the cursor is tailable, as batches received from remote
// tailable cursors should be passed through to the client without asking for more batches.
if (!_params.isTailable && !remote.hasNext() && !remote.exhausted()) {
auto nextBatchStatus = askForNextBatch_inlock(remoteIndex);
if (!nextBatchStatus.isOK()) {
remote.status = nextBatchStatus;
return;
}
}
// ScopeGuard requires dismiss on success, but we want waiter to be signalled on success as
// well as failure.
signaller.Dismiss();
signalCurrentEventIfReady_inlock();
}