[Offload] Do not do not re-initialize RPC server (llvm#218757)

Summary:
This was done for every executable that was loaded. In the current
OpenMP use this never caused issues, but we need to check if this was
already done so we don't leak the memory.

Author: jhuber6

offload/plugins-nextgen/common/src/RPC.cpp

@@ -207,32 +207,40 @@ RPCServerTy::isDeviceUsingRPC(plugin::GenericDeviceTy &Device,
 Error RPCServerTy::initDevice(plugin::GenericDeviceTy &Device,
                               plugin::GenericGlobalHandlerTy &Handler,
                               plugin::DeviceImageTy &Image) {
+  std::lock_guard<decltype(BufferMutex)> Lock(BufferMutex);
   uint64_t NumPorts =
       std::min(Device.requestedRPCPortCount(), rpc::MAX_PORT_COUNT);
-  auto RPCBufferOrErr = Device.allocate(
-      rpc::Server::allocation_size(Device.getRPCNumLanes(), NumPorts), nullptr,
-      TARGET_ALLOC_HOST);
-  if (!RPCBufferOrErr)
-    return RPCBufferOrErr.takeError();
-
-  void *RPCBuffer = *RPCBufferOrErr;
-  if (!RPCBuffer)
-    return plugin::Plugin::error(
-        error::ErrorCode::UNKNOWN,
-        "failed to initialize RPC server for device %d", Device.getDeviceId());
-
-  // The doorbell is used by AMDGPU targets to let the server thread be
-  // descheduled. It is optional and will be ignored if the fields are null.
-  rpc::Doorbell Doorbell{};
-  if (auto Err = Device.Plugin.initRPCDoorbell(Doorbell.value, Doorbell.mailbox,
-                                               Doorbell.event_id))
-    return Err;
-
-  auto *DoorbellPtr = reinterpret_cast<rpc::Doorbell *>(
-      static_cast<uint8_t *>(RPCBuffer) + rpc::Server::doorbell_offset());
-  std::memcpy(DoorbellPtr, &Doorbell, sizeof(rpc::Doorbell));
+  void *RPCBuffer = Buffers[Device.getDeviceId()];
+  if (!RPCBuffer) {
+    auto RPCBufferOrErr = Device.allocate(
+        rpc::Server::allocation_size(Device.getRPCNumLanes(), NumPorts),
+        nullptr, TARGET_ALLOC_HOST);
+    if (!RPCBufferOrErr)
+      return RPCBufferOrErr.takeError();
+
+    RPCBuffer = *RPCBufferOrErr;
+    if (!RPCBuffer)
+      return plugin::Plugin::error(
+          error::ErrorCode::UNKNOWN,
+          "failed to initialize RPC server for device %d",
+          Device.getDeviceId());
+
+    // The doorbell is used by AMDGPU targets to let the server thread be
+    // descheduled. It is optional and will be ignored if the fields are null.
+    rpc::Doorbell Doorbell{};
+    if (auto Err = Device.Plugin.initRPCDoorbell(
+            Doorbell.value, Doorbell.mailbox, Doorbell.event_id))
+      return Err;
+
+    auto *DoorbellPtr = reinterpret_cast<rpc::Doorbell *>(
+        static_cast<uint8_t *>(RPCBuffer) + rpc::Server::doorbell_offset());
+    std::memcpy(DoorbellPtr, &Doorbell, sizeof(rpc::Doorbell));
+
+    Buffers[Device.getDeviceId()] = RPCBuffer;
+    Devices[Device.getDeviceId()] = &Device;
+  }
 
-  // Get the address of the RPC client from the device.
+  // Each image has its own client that must point at the shared buffer.
   plugin::GlobalTy ClientGlobal("__llvm_rpc_client", sizeof(rpc::Client));
   if (auto Err =
           Handler.getGlobalMetadataFromDevice(Device, Image, ClientGlobal))
@@ -242,9 +250,6 @@ Error RPCServerTy::initDevice(plugin::GenericDeviceTy &Device,
   if (auto Err = Device.dataSubmit(ClientGlobal.getPtr(), &client,
                                    sizeof(rpc::Client), nullptr))
     return Err;
-  std::lock_guard<decltype(BufferMutex)> Lock(BufferMutex);
-  Buffers[Device.getDeviceId()] = RPCBuffer;
-  Devices[Device.getDeviceId()] = &Device;
 
   return Error::success();
 }