Merge pull request #50280 from nouiz:upstream_cuda_malloc_async_oom_deadlock

tensorflower-gardener · tensorflower-gardener · commit eb965f6965cc · 2021-06-17T07:10:34.000-07:00
PiperOrigin-RevId: 379949863
Change-Id: Icc5fc36016731b3ad6b5c93bd3c98711826c2ff7
diff --git a/tensorflow/core/common_runtime/gpu/gpu_cudamallocasync_allocator.cc b/tensorflow/core/common_runtime/gpu/gpu_cudamallocasync_allocator.cc
@@ -195,8 +195,7 @@ GpuCudaMallocAsyncAllocator::GpuCudaMallocAsyncAllocator(
 #endif  // TF_CUDA_MALLOC_ASYNC_SUPPORTED
 }
 
-GpuCudaMallocAsyncAllocator::~GpuCudaMallocAsyncAllocator() {
-}
+GpuCudaMallocAsyncAllocator::~GpuCudaMallocAsyncAllocator() {}
 
 void* GpuCudaMallocAsyncAllocator::AllocateRaw(size_t alignment,
                                                size_t num_bytes) {
@@ -215,7 +214,6 @@ void* GpuCudaMallocAsyncAllocator::AllocateRaw(size_t alignment,
                                   num_bytes, pool_, cuda_stream_)) {
     size_t free, total;
     cuMemGetInfo(&free, &total);
-    mutex_lock lock(lock_);
     LOG(ERROR) << Name() << " cuMemAllocAsync failed to allocate " << num_bytes
                << ": " << GetCudaErrorMessage(result)
                << "\n Free memory/Total memory: " << free << "/" << total;
