1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 16 #ifndef TENSORFLOW_CORE_COMMON_RUNTIME_PROCESS_STATE_H_ 17 #define TENSORFLOW_CORE_COMMON_RUNTIME_PROCESS_STATE_H_ 18 19 #include <functional> 20 #include <map> 21 #include <unordered_map> 22 #include <vector> 23 24 #include "tensorflow/core/framework/allocator.h" 25 #include "tensorflow/core/framework/allocator_registry.h" 26 #include "tensorflow/core/platform/mutex.h" 27 #include "tensorflow/core/platform/thread_annotations.h" 28 #include "tensorflow/core/platform/types.h" 29 #include "tensorflow/core/protobuf/config.pb.h" 30 31 namespace tensorflow { 32 33 class Allocator; 34 class PoolAllocator; 35 36 // Singleton that manages per-process state, e.g. allocation of 37 // shared resources. 38 class ProcessState : public ProcessStateInterface { 39 public: 40 static ProcessState* singleton(); 41 42 // Descriptor for memory allocation attributes, used by optional 43 // runtime correctness analysis logic. 44 struct MemDesc { 45 enum MemLoc { CPU, GPU }; 46 MemLoc loc; 47 int dev_index; 48 bool gpu_registered; 49 bool nic_registered; MemDescMemDesc50 MemDesc() 51 : loc(CPU), 52 dev_index(0), 53 gpu_registered(false), 54 nic_registered(false) {} 55 string DebugString(); 56 }; 57 58 // If NUMA Allocators are desired, call this before calling any 59 // Allocator accessor. EnableNUMA()60 void EnableNUMA() { numa_enabled_ = true; } 61 62 // Returns what we know about the memory at ptr. 63 // If we know nothing, it's called CPU 0 with no other attributes. 64 MemDesc PtrType(const void* ptr); 65 66 // Returns the one CPUAllocator used for the given numa_node. 67 // Treats numa_node == kNUMANoAffinity as numa_node == 0. 68 Allocator* GetCPUAllocator(int numa_node); 69 70 // Registers alloc visitor for the CPU allocator(s). 71 // REQUIRES: must be called before GetCPUAllocator. 72 void AddCPUAllocVisitor(SubAllocator::Visitor v); 73 74 // Registers free visitor for the CPU allocator(s). 75 // REQUIRES: must be called before GetCPUAllocator. 76 void AddCPUFreeVisitor(SubAllocator::Visitor v); 77 78 typedef std::unordered_map<const void*, MemDesc> MDMap; 79 80 protected: 81 ProcessState(); 82 friend class GPUProcessState; 83 84 // If these flags need to be runtime configurable consider adding 85 // them to ConfigProto. 86 static const bool FLAGS_brain_mem_reg_gpu_dma = true; 87 static const bool FLAGS_brain_gpu_record_mem_types = false; 88 89 // Helper method for unit tests to reset the ProcessState singleton by 90 // cleaning up everything. Never use in production. 91 void TestOnlyReset(); 92 93 static ProcessState* instance_; 94 bool numa_enabled_; 95 96 mutex mu_; 97 98 // Indexed by numa_node. If we want numa-specific allocators AND a 99 // non-specific allocator, maybe should index by numa_node+1. 100 std::vector<Allocator*> cpu_allocators_ GUARDED_BY(mu_); 101 std::vector<SubAllocator::Visitor> cpu_alloc_visitors_ GUARDED_BY(mu_); 102 std::vector<SubAllocator::Visitor> cpu_free_visitors_ GUARDED_BY(mu_); 103 104 // Optional RecordingAllocators that wrap the corresponding 105 // Allocators for runtime attribute use analysis. 106 MDMap mem_desc_map_; 107 std::vector<Allocator*> cpu_al_ GUARDED_BY(mu_); 108 }; 109 110 namespace internal { 111 class RecordingAllocator : public Allocator { 112 public: RecordingAllocator(ProcessState::MDMap * mm,Allocator * a,ProcessState::MemDesc md,mutex * mu)113 RecordingAllocator(ProcessState::MDMap* mm, Allocator* a, 114 ProcessState::MemDesc md, mutex* mu) 115 : mm_(mm), a_(a), md_(md), mu_(mu) {} 116 Name()117 string Name() override { return a_->Name(); } AllocateRaw(size_t alignment,size_t num_bytes)118 void* AllocateRaw(size_t alignment, size_t num_bytes) override { 119 void* p = a_->AllocateRaw(alignment, num_bytes); 120 mutex_lock l(*mu_); 121 (*mm_)[p] = md_; 122 return p; 123 } DeallocateRaw(void * p)124 void DeallocateRaw(void* p) override { 125 mutex_lock l(*mu_); 126 auto iter = mm_->find(p); 127 mm_->erase(iter); 128 a_->DeallocateRaw(p); 129 } TracksAllocationSizes()130 bool TracksAllocationSizes() override { return a_->TracksAllocationSizes(); } RequestedSize(const void * p)131 size_t RequestedSize(const void* p) override { return a_->RequestedSize(p); } AllocatedSize(const void * p)132 size_t AllocatedSize(const void* p) override { return a_->AllocatedSize(p); } GetStats()133 absl::optional<AllocatorStats> GetStats() override { return a_->GetStats(); } ClearStats()134 void ClearStats() override { a_->ClearStats(); } 135 ProcessState::MDMap* mm_; // not owned 136 Allocator* a_; // not owned 137 ProcessState::MemDesc md_; 138 mutex* mu_; 139 }; 140 } // namespace internal 141 } // namespace tensorflow 142 #endif // TENSORFLOW_CORE_COMMON_RUNTIME_PROCESS_STATE_H_ 143