fix potential issue in MT context

fserver/csrc/public.hpp

@@ -55,12 +55,11 @@ void RequestHandler(const AFTensorMeta& req_meta, AFTensorServer* server) {
     std::lock_guard<std::mutex> lock(mu_);
     meta_map_[handler_counter_] = req_meta;
 
-    q_[req_meta.sender_rank].emplace_back(handler_counter_,
-                                          std::move(tensors),
+    q_[req_meta.sender_rank].emplace_back(handler_counter_, std::move(tensors),
                                           keys);
     q_signal_.fetch_or(1 << req_meta.sender_rank);
+    ++handler_counter_;
   }
-  ++handler_counter_;
 }
 
 std::vector<ServerDataBatch> get_batch() {
@@ -180,13 +179,13 @@ void init() {
 
   worker_mask_ = (1 << num_worker_) - 1;
   q_.resize(num_worker_);
-  q_signal_.store(0);;
+  q_signal_.store(0);
   ps::StartPS(0, role_,  group_size_ * node_rank_ + gpu_ + offset, true);
   if (role_ == Node::WORKER) {
-    fworker_ = new AFTensorWorker(instance_id_ );
+    fworker_ = new AFTensorWorker(instance_id_);
     barrier(true, true);
   } else if (role_ == Node::SERVER) {
-    fserver_ = new AFTensorServer(instance_id_ );
+    fserver_ = new AFTensorServer(instance_id_);
     fserver_->SetRequestHandle(RequestHandler);
     ps::RegisterExitCallback([]() { delete fserver_; });
     barrier(true, true);

include/ps/internal/assign_op.h

@@ -27,22 +27,22 @@ template <typename T>
 inline void AssignFunc(const T& lhs, AssignOp op, T* rhs) {
   switch (op) {
     case ASSIGN:
-      *right = left;
+      *rhs = lhs;
       break;
     case PLUS:
-      *right += left;
+      *rhs += lhs;
       break;
     case MINUS:
-      *right -= left;
+      *rhs -= lhs;
       break;
     case TIMES:
-      *right *= left;
+      *rhs *= lhs;
       break;
     case DIVIDE:
-      *right /= left;
+      *rhs /= lhs;
       break;
     default:
-      LOG(FATAL) << "use AssignOpInt..";
+      PS_LOG(FATAL) << "use AssignOpInt..";
   }
 }
 
@@ -54,28 +54,28 @@ template <typename T>
 inline void AssignFuncInt(const T& lhs, AssignOp op, T* rhs) {
   switch (op) {
     case ASSIGN:
-      *right = left;
+      *rhs = lhs;
       break;
     case PLUS:
-      *right += left;
+      *rhs += lhs;
       break;
     case MINUS:
-      *right -= left;
+      *rhs -= lhs;
       break;
     case TIMES:
-      *right *= left;
+      *rhs *= lhs;
       break;
     case DIVIDE:
-      *right /= left;
+      *rhs /= lhs;
       break;
     case AND:
-      *right &= left;
+      *rhs &= lhs;
       break;
     case OR:
-      *right |= left;
+      *rhs |= lhs;
       break;
     case XOR:
-      *right ^= left;
+      *rhs ^= lhs;
       break;
   }
 }

include/ps/internal/backend.h

@@ -109,6 +109,9 @@ class Backend {
     static Backend* backend_impl = nullptr;
     if (backend_impl == nullptr) {
       std::unique_lock<std::mutex> lock(backends_mutex_);
+      if (backend_impl) {
+        return backend_impl;
+      }
       std::string backend_type = "GPU";
       backend_type = Environment::Get()->find("STEPMESH_BAKCEND", backend_type);
       PS_CHECK_NE(backends_.find(backend_type), backends_.end())

include/ps/internal/postoffice.h

@@ -155,7 +155,7 @@ class Postoffice {
 
   /**
    * \brief convert a server group's rank into a instance id with the
-   * provded instance offset from that group
+   * provided instance offset from that group
    * \param rank the server group rank
    * \param instance_idx the offset of the instance in the group
    */