Correct data race in GpuAgent::GetXgmiBlit.

Threads may race against xgmi_peer_list_ when dynamically assigning
peers to sdma engines.

Change-Id: I300c10f0cfa0ff7d6a5515364070a0895e2f4644


[ROCm/ROCR-Runtime commit: bb4dfbba1e]

projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_gpu_agent.h

@@ -400,6 +400,9 @@ class GpuAgent : public GpuAgentInt {
   // List of agents connected via xGMI
   std::vector<const core::Agent*> xgmi_peer_list_;
 
+  // Protects xgmi_peer_list_
+  KernelMutex xgmi_peer_list_lock_;
+
   // @brief AQL queues for cache management and blit compute usage.
   enum QueueEnum {
     QueueUtility,   // Cache management and device to {host,device} blit compute

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -1483,6 +1483,8 @@ lazy_ptr<core::Blit>& GpuAgent::GetXgmiBlit(const core::Agent& dst_agent) {
   uint32_t xgmi_engine_cnt = properties_.NumSdmaXgmiEngines;
   assert((xgmi_engine_cnt > 0) && ("Illegal condition, should not happen"));
 
+  ScopedAcquire<KernelMutex> lock(&xgmi_peer_list_lock_);
+
   for (uint32_t idx = 0; idx < xgmi_peer_list_.size(); idx++) {
     uint64_t dst_handle = dst_agent.public_handle().handle;
     uint64_t peer_handle = xgmi_peer_list_[idx]->public_handle().handle;