tensorflow/core/distributed_runtime/eager/eager_service_impl.cc

/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "tensorflow/core/distributed_runtime/eager/eager_service_impl.h"

#include "absl/container/fixed_array.h"
#include "absl/memory/memory.h"
#include "absl/types/optional.h"
#include "tensorflow/c/c_api_internal.h"
#include "tensorflow/c/tf_status_helper.h"
#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/eager/context.h"
#include "tensorflow/core/common_runtime/eager/eager_operation.h"
#include "tensorflow/core/common_runtime/eager/execute.h"
#include "tensorflow/core/common_runtime/function.h"
#include "tensorflow/core/common_runtime/process_util.h"
#include "tensorflow/core/distributed_runtime/eager/cluster_function_library_runtime.h"
#include "tensorflow/core/distributed_runtime/eager/remote_mgr.h"
#include "tensorflow/core/distributed_runtime/eager/remote_tensor_handle.h"
#include "tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.h"
#include "tensorflow/core/distributed_runtime/server_lib.h"
#include "tensorflow/core/distributed_runtime/session_mgr.h"
#include "tensorflow/core/distributed_runtime/worker_cache.h"
#include "tensorflow/core/distributed_runtime/worker_cache_wrapper.h"
#include "tensorflow/core/distributed_runtime/worker_env.h"
#include "tensorflow/core/framework/rendezvous.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/gtl/cleanup.h"
#include "tensorflow/core/lib/random/random.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/lib/strings/stringprintf.h"
#include "tensorflow/core/platform/cpu_info.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/host_info.h"
#include "tensorflow/core/platform/mutex.h"
#include "tensorflow/core/profiler/lib/traceme.h"
#include "tensorflow/core/protobuf/error_codes.pb.h"

namespace tensorflow {
namespace eager {

namespace {
Status GetNumRetvals(tensorflow::EagerContext* context, const string& op_name,
                     const google::protobuf::Map<string, tensorflow::AttrValue>& attrs,
                     int* num_retvals) {
  const tensorflow::OpRegistrationData* op_reg_data = nullptr;
  auto status = tensorflow::OpRegistry::Global()->LookUp(op_name, &op_reg_data);
  if (errors::IsNotFound(status)) {
    status = context->FindFunctionOpData(op_name, &op_reg_data);
  }
  TF_RETURN_IF_ERROR(status);

  const tensorflow::OpDef& op_def = op_reg_data->op_def;

  for (const auto& output_arg : op_def.output_arg()) {
    if (!output_arg.number_attr().empty()) {
      auto iter = attrs.find(output_arg.number_attr());
      if (iter == attrs.end()) {
        return errors::InvalidArgument("Unable to find number_attr ",
                                       output_arg.number_attr(),
                                       " for Op: ", op_name);
      }
      *num_retvals += iter->second.i();
    } else if (!output_arg.type_list_attr().empty()) {
      auto iter = attrs.find(output_arg.type_list_attr());
      if (iter == attrs.end()) {
        return errors::InvalidArgument("Unable to find type_list_attr ",
                                       output_arg.type_list_attr(),
                                       " for Op: ", op_name);
      }
      *num_retvals += iter->second.list().type_size();
    } else {
      *num_retvals += 1;
    }
  }

  return Status::OK();
}
}  // namespace

Status EagerServiceImpl::CreateContext(const CreateContextRequest* request,
                                       CreateContextResponse* response) {
  {
    mutex_lock l(contexts_mu_);
    auto context_it = contexts_.find(request->context_id());
    if (context_it != contexts_.end()) {
      if (request->context_view_id() <
          context_it->second->Context()->GetContextViewId()) {
        return errors::InvalidArgument("EagerService:CreateContext failed. ",
                                       "Context id: <", request->context_id(),
                                       "> already exists.");
      } else {
        // For existing context with a stale context_view_id, close the old one
        // and recreate with new view id. This is likely due to the worker
        // disconnected and then reconnected after one or more cluster updates.
        context_it->second->Unref();
        contexts_.erase(context_it);
      }
    }
  }
  // make sure env_ , env_->rendezvous_mgr available
  if (env_ == nullptr || env_->rendezvous_mgr == nullptr) {
    return tensorflow::errors::Internal(
        "invalid eager env_ or env_->rendezvous_mgr.");
  }
  std::vector<DeviceAttributes> cluster_device_attributes;
  cluster_device_attributes.reserve(
      request->cluster_device_attributes().size());
  for (const auto& cluster_device : request->cluster_device_attributes()) {
    cluster_device_attributes.push_back(cluster_device);
  }

  auto* r = env_->rendezvous_mgr->Find(request->context_id());
  auto session_name =
      tensorflow::strings::StrCat("eager_", request->context_id());
  TF_RETURN_IF_ERROR(env_->session_mgr->CreateSession(
      session_name, request->server_def(), request->cluster_device_attributes(),
      true));
  int64 context_id = request->context_id();
  std::function<void()> session_destroyer = [this, context_id, session_name]() {
    env_->rendezvous_mgr->Cleanup(context_id);
    auto s = env_->session_mgr->DeleteSession(session_name);
    if (!s.ok()) {
      LOG(WARNING) << "Failed to destroy worker session '" << session_name
                   << "' due to " << s.error_message();
    }
  };

  std::shared_ptr<WorkerSession> worker_session;
  TF_RETURN_IF_ERROR(env_->session_mgr->WorkerSessionForSession(
      session_name, &worker_session));

  tensorflow::DeviceMgr* device_mgr = worker_session->device_mgr();

  // Initialize remote tensor communication based on worker session.
  TF_RETURN_IF_ERROR(r->Initialize(worker_session.get()));

  std::function<Rendezvous*(const int64)> rendezvous_creator =
      [worker_session, this](const int64 step_id) {
        auto* r = env_->rendezvous_mgr->Find(step_id);
        r->Initialize(worker_session.get()).IgnoreError();
        return r;
      };

  LOG(INFO) << "Creating " << (request->async() ? "async" : "sync")
            << " eager service context with rendezvous_id on host "
            << port::Hostname() << " " << worker_session->worker_name();
  SessionOptions opts;
  opts.config = request->server_def().default_session_config();
  tensorflow::EagerContext* ctx = new tensorflow::EagerContext(
      opts, tensorflow::ContextDevicePlacementPolicy::DEVICE_PLACEMENT_SILENT,
      tensorflow::ContextMirroringPolicy::MIRRORING_NONE, request->async(),
      device_mgr, false, r, GetDefaultCustomKernelCreator(),
      worker_session->cluster_flr());
  // Ownership will be transferred to the ServerContext, or else in an error
  // case ctx will be deleted by this unref.
  core::ScopedUnref unref_ctx(ctx);

  std::vector<string> remote_workers;
  worker_session->worker_cache()->ListWorkers(&remote_workers);
  remote_workers.erase(std::remove(remote_workers.begin(), remote_workers.end(),
                                   worker_session->worker_name()),
                       remote_workers.end());

  std::unique_ptr<tensorflow::eager::EagerClientCache> remote_eager_workers;
  TF_RETURN_IF_ERROR(worker_session->worker_cache()->GetEagerClientCache(
      &remote_eager_workers));
  DistributedFunctionLibraryRuntime* cluster_flr =
      eager::CreateClusterFLR(request->context_id(), ctx, worker_session.get());

  auto remote_mgr =
      absl::make_unique<tensorflow::eager::RemoteMgr>(/*is_master=*/false, ctx);
  Status s = ctx->InitializeRemoteWorker(
      std::move(remote_eager_workers), worker_session->remote_device_mgr(),
      remote_workers, request->context_id(), request->context_view_id(),
      std::move(rendezvous_creator), cluster_flr, std::move(remote_mgr),
      std::move(session_destroyer));
  if (!s.ok()) {
    VLOG(1) << "EagerContext::InitializeRemoteWorker failed with "
            << s.ToString();
    return s;
  }

  std::vector<DeviceAttributes> device_attributes;
  device_mgr->ListDeviceAttributes(&device_attributes);

  for (const auto& da : device_attributes) {
    *response->add_device_attributes() = da;
  }
  {
    mutex_lock l(contexts_mu_);
    auto context_it = contexts_.find(request->context_id());
    if (context_it != contexts_.end()) {
      return errors::InvalidArgument("EagerService:CreateContext failed. ",
                                     "Context id: <", request->context_id(),
                                     "> already exists.");
    }
    contexts_.emplace(request->context_id(),
                      new ServerContext(ctx, request->keep_alive_secs(), env_));
  }

  return Status::OK();
}

Status EagerServiceImpl::UpdateContext(const UpdateContextRequest* request,
                                       UpdateContextResponse* response) {
  // make sure env_ , env_->rendezvous_mgr available
  if (env_ == nullptr || env_->rendezvous_mgr == nullptr) {
    return tensorflow::errors::Internal(
        "invalid eager env_ or env_->rendezvous_mgr.");
  }

  // Find the context to update by the requested context_id
  ServerContext* server_context = nullptr;
  TF_RETURN_IF_ERROR(GetServerContext(request->context_id(), &server_context));
  core::ScopedUnref context_unref(server_context);

  tensorflow::EagerContext* ctx = server_context->Context();
  if (request->context_view_id() != ctx->GetContextViewId() + 1) {
    return errors::InvalidArgument(
        "EagerService:UpdateContext failed. Context id: <",
        request->context_id(), "> currently at view #", ctx->GetContextViewId(),
        " but received update request at view #", request->context_view_id(),
        ". View id should only be continuously incremented.");
  }
  ctx->ClearCaches();
  // TODO(b/143914772): Potential memory leak if rendezvous has pending
  // tensors for removed / replaced workers.

  std::vector<DeviceAttributes> cluster_device_attributes;
  cluster_device_attributes.reserve(
      request->cluster_device_attributes().size());
  for (const auto& cluster_device : request->cluster_device_attributes()) {
    cluster_device_attributes.push_back(cluster_device);
  }
  auto session_name =
      tensorflow::strings::StrCat("eager_", request->context_id());
  TF_RETURN_IF_ERROR(env_->session_mgr->UpdateSession(
      session_name, request->server_def(), request->cluster_device_attributes(),
      true));

  std::shared_ptr<WorkerSession> worker_session;
  TF_RETURN_IF_ERROR(env_->session_mgr->WorkerSessionForSession(
      session_name, &worker_session));

  tensorflow::DeviceMgr* device_mgr = worker_session->device_mgr();

  std::vector<string> remote_workers;
  worker_session->worker_cache()->ListWorkers(&remote_workers);
  remote_workers.erase(std::remove(remote_workers.begin(), remote_workers.end(),
                                   worker_session->worker_name()),
                       remote_workers.end());
  VLOG(1) << "On existing server " << worker_session->worker_name()
          << " updating remote workers";
  if (VLOG_IS_ON(2)) {
    for (const string& rw : remote_workers) {
      VLOG(2) << "Remote worker " << rw;
    }
  }

  std::unique_ptr<tensorflow::eager::EagerClientCache> remote_eager_workers;
  TF_RETURN_IF_ERROR(worker_session->worker_cache()->GetEagerClientCache(
      &remote_eager_workers));

  DistributedFunctionLibraryRuntime* cluster_flr =
      eager::CreateClusterFLR(request->context_id(), ctx, worker_session.get());

  Status s = ctx->UpdateRemoteWorker(
      device_mgr, std::move(remote_eager_workers),
      worker_session->remote_device_mgr(), remote_workers,
      request->context_id(), cluster_flr);
  if (!s.ok()) {
    VLOG(1) << "EagerContext::UpdateRemoteWorker failed with " << s.ToString();
    return s;
  }

  std::vector<DeviceAttributes> device_attributes;
  device_mgr->ListDeviceAttributes(&device_attributes);

  for (const auto& da : device_attributes) {
    *response->add_device_attributes() = da;
  }

  return Status::OK();
}

Status EagerServiceImpl::CreateMasterContext(
    const tensorflow::uint64 context_id, EagerContext* context) {
  {
    mutex_lock l(contexts_mu_);
    auto iter = contexts_.find(context_id);
    if (iter != contexts_.end()) {
      return errors::InvalidArgument(
          "EagerService:CreateMasterContext failed. ", "Context id: <",
          context_id, "> already exists.");
    }
  }
  ServerContext* server_context =
      ServerContext::CreateMasterContext(context, env_);
  mutex_lock l(contexts_mu_);
  contexts_.emplace(context_id, server_context);
  return Status::OK();
}

Status TensorHandleShape(TensorHandle* handle, TensorShapeProto* proto) {
  const tensorflow::Tensor* t = nullptr;

  // TODO(nareshmodi): This call makes async calls sync calls. Fix this.
  TF_RETURN_IF_ERROR(handle->Tensor(&t));

  t->shape().AsProto(proto);

  return Status::OK();
}

Status EagerServiceImpl::ExecuteOp(const Operation& operation,
                                   EagerContext* eager_context,
                                   EagerExecutor* eager_executor,
                                   QueueResponse* queue_response) {
  std::unique_ptr<tensorflow::EagerOperation> op;
  const char* name = operation.name().c_str();  // Shorthand
  const tensorflow::AttrTypeMap* types;
  bool is_function = false;
  TF_RETURN_IF_ERROR(tensorflow::AttrTypeMapForOp(name, &types, &is_function));
  if (is_function && !eager_context->FindFunctionByName(name)) {
    return errors::NotFound(
        "'", name,
        "' is neither a type of a primitive operation nor a name "
        "of a function registered in binary running on ",
        port::Hostname(),
        ". One possible root cause is the client and server binaries are not "
        "built with the same version. Please make sure the operation or "
        "function is registered in the binary running in this process.");
  }
  absl::optional<tensorflow::EagerRemoteFunctionParams> remote_func_params =
      absl::nullopt;
  if (operation.is_function()) {
    if (operation.is_component_function()) {
      remote_func_params = {operation.id(), operation.func_step_id()};
    } else {
      remote_func_params = {operation.id(), absl::nullopt};
    }
  }
  op.reset(new tensorflow::EagerOperation(eager_context, name, is_function,
                                          types, eager_executor,
                                          remote_func_params));

  TF_RETURN_IF_ERROR(op->SetDeviceName(operation.device().c_str()));

  {
    profiler::TraceMe activity("EagerService:RemoteTensorHandleInternal",
                               profiler::TraceMeLevel::kVerbose);
    for (const auto& remote_handle : operation.inputs()) {
      tensorflow::TensorHandle* handle;
      TF_RETURN_IF_ERROR(
          eager_context->RemoteMgr()->DeserializeRemoteTensorHandle(
              remote_handle, &handle));
      op->AddInput(handle);
      // Unref handle since it has a ref as an input now.
      handle->Unref();
    }
  }

  for (const auto& attr : operation.attrs()) {
    op->MutableAttrs()->Set(attr.first, attr.second);
  }

  int num_retvals = 0;
  // TODO(nareshmodi): Consider caching this.
  TF_RETURN_IF_ERROR(GetNumRetvals(eager_context, operation.name(),
                                   operation.attrs(), &num_retvals));

  absl::FixedArray<tensorflow::TensorHandle*> retvals(num_retvals);
  VLOG(3) << "ServerContext: Calling EagerExecute for op " << operation.id();
  TF_RETURN_IF_ERROR(EagerExecute(op.get(), retvals.data(), &num_retvals));

  eager_context->RemoteMgr()->AddOperationOutputs(
      absl::MakeSpan(retvals.data(), num_retvals), operation.id());

  for (int i = 0; i < num_retvals; i++) {
    TF_RETURN_IF_ERROR(
        TensorHandleShape(retvals[i], queue_response->add_shape()));
  }

  return Status::OK();
}

Status EagerServiceImpl::Enqueue(const EnqueueRequest* request,
                                 EnqueueResponse* response, uint64 stream_id) {
  profiler::TraceMe activity(
      [&] {
        return absl::StrCat("EagerService:Enqueue:", request->DebugString());
      },
      profiler::TraceMeLevel::kInfo);
  ServerContext* context = nullptr;
  TF_RETURN_IF_ERROR(GetServerContext(request->context_id(), &context));
  core::ScopedUnref context_unref(context);

  EagerExecutor& executor =
      stream_id == kInvalidStreamId
          ? context->Context()->Executor()
          : context->Context()->RemoteMgr()->GetOrCreateExecutorForStream(
                stream_id);
  Status s;
  for (const auto& item : request->queue()) {
    auto* queue_response = response->add_queue_response();
    if (item.has_operation()) {
      s = ExecuteOp(item.operation(), context->Context(), &executor,
                    queue_response);
    } else if (item.has_handle_to_decref()) {
      auto handle_to_decref = absl::make_unique<RemoteTensorHandleInternal>(
          item.handle_to_decref());
      auto node = absl::make_unique<ClientTensorHandleDeleteNode>(
          context, std::move(handle_to_decref));
      s = context->Context()->Executor().AddOrExecute(std::move(node));
    } else if (item.has_send_tensor()) {
      s = SendTensor(item.send_tensor(), context->Context());
    } else if (item.has_register_function()) {
      s = RegisterFunction(item.register_function(), context->Context());
    } else {
      s = CleanupFunction(item.cleanup_function());
    }

    if (!s.ok()) {
      if (stream_id != kInvalidStreamId) {
        // TODO(b/138847548): Cleanup the executor when StreamCall is deleted.
        context->Context()->RemoteMgr()->DeleteExecutorForStream(stream_id);
      }
      return s;
    }
  }

  return Status::OK();
}

Status EagerServiceImpl::WaitQueueDone(const WaitQueueDoneRequest* request,
                                       WaitQueueDoneResponse* response) {
  ServerContext* context = nullptr;
  TF_RETURN_IF_ERROR(GetServerContext(request->context_id(), &context));
  core::ScopedUnref context_unref(context);

  if (request->op_id_size() > 0) {
    return errors::Unimplemented(
        "EagerServiceImpl::WaitQueueDone is not "
        "implemented for particular op IDs.");
  }
  return context->Context()->Executor().WaitForAllPendingNodes();
}

Status EagerServiceImpl::KeepAlive(const KeepAliveRequest* request,
                                   KeepAliveResponse* response) {
  ServerContext* context = nullptr;
  TF_RETURN_IF_ERROR(GetServerContext(request->context_id(), &context));
  core::ScopedUnref context_unref(context);

  tensorflow::EagerContext* ctx = context->Context();
  response->set_context_view_id(ctx->GetContextViewId());
  return Status::OK();
}

Status EagerServiceImpl::CloseContext(const CloseContextRequest* request,
                                      CloseContextResponse* response) {
  VLOG(1) << "Executing EagerService::CloseContext for context "
          << request->context_id();
  ServerContext* context = nullptr;
  if (!GetServerContext(request->context_id(), &context).ok()) {
    // Swallow the error here.
    return Status::OK();
  }
  core::ScopedUnref context_unref(context);

  if (request->context_view_id() < context->Context()->GetContextViewId()) {
    // Swallow the error here.
    LOG(INFO) << "Ignoring CloseContext request with a stale context_view_id "
              << request->context_view_id() << "  for context_id "
              << request->context_id() << ". The current context_view_id is "
              << context->Context()->GetContextViewId() << ".";
    return Status::OK();
  }

  mutex_lock l(contexts_mu_);
  contexts_.erase(request->context_id());

  // GetServerContext returns a newly Reffed copy of ServerContext, which is
  // unreffed by context_unref. Additionally, we need to unref it one time since
  // we are releasing it from the map.
  context->Unref();

  return Status::OK();
}

Status EagerServiceImpl::RegisterFunction(
    const RegisterFunctionOp& register_function, EagerContext* eager_context) {
  // If the function is a component of a multi-device function, we only need to
  // register it locally.
  return eager_context->AddFunctionDef(
      register_function.function_def(), register_function.library(),
      register_function.is_component_function());
}

Status EagerServiceImpl::CleanupFunction(
    const CleanupFunctionOp& cleanup_function) {
  env_->rendezvous_mgr->Cleanup(cleanup_function.step_id());
  return Status::OK();
}

Status EagerServiceImpl::SendTensor(const SendTensorOp& send_tensor,
                                    EagerContext* eager_context) {
  tensorflow::gtl::InlinedVector<tensorflow::TensorHandle*, 2> tensors;
  for (const auto& tensor_proto : send_tensor.tensors()) {
    Tensor tensor;
    if (!tensor.FromProto(tensor_proto)) {
      return errors::InvalidArgument("Unable to parse tensor proto");
    }

    TensorHandle* tensor_handle = nullptr;
    TF_RETURN_IF_ERROR(TensorHandle::CreateLocalHandle(tensor, &tensor_handle));
    TensorHandle* copied_handle = nullptr;
    Device* device;
    TF_RETURN_IF_ERROR(eager_context->FindDeviceFromName(
        send_tensor.device_name().c_str(), &device));
    TF_RETURN_IF_ERROR(EagerCopyToDevice(tensor_handle, eager_context,
                                         &eager_context->Executor(), device,
                                         false, &copied_handle));
    tensors.push_back(copied_handle);
    tensor_handle->Unref();
  }

  eager_context->RemoteMgr()->AddOperationOutputs(tensors, send_tensor.op_id());

  return Status::OK();
}

tensorflow::Status EagerServiceImpl::GetServerContext(
    uint64 context_id, ServerContext** server_context) {
  mutex_lock l(contexts_mu_);
  auto iter = contexts_.find(context_id);
  if (iter == contexts_.end()) {
    *server_context = nullptr;
    return errors::InvalidArgument(strings::Printf(
        "Unable to find a context_id matching the specified one "
        "(%llu). Perhaps the worker was restarted, or the context was GC'd?",
        context_id));
  }

  *server_context = iter->second;
  (*server_context)->Ref();

  (*server_context)->RecordAccess();

  return Status::OK();
}

}  // namespace eager
}  // namespace tensorflow