sequence_lock.h (7692B)
1 // 2 // Copyright 2020 The Abseil Authors. 3 // 4 // Licensed under the Apache License, Version 2.0 (the "License"); 5 // you may not use this file except in compliance with the License. 6 // You may obtain a copy of the License at 7 // 8 // https://www.apache.org/licenses/LICENSE-2.0 9 // 10 // Unless required by applicable law or agreed to in writing, software 11 // distributed under the License is distributed on an "AS IS" BASIS, 12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 // See the License for the specific language governing permissions and 14 // limitations under the License. 15 16 #ifndef ABSL_FLAGS_INTERNAL_SEQUENCE_LOCK_H_ 17 #define ABSL_FLAGS_INTERNAL_SEQUENCE_LOCK_H_ 18 19 #include <stddef.h> 20 #include <stdint.h> 21 22 #include <atomic> 23 #include <cassert> 24 #include <cstring> 25 26 #include "absl/base/optimization.h" 27 28 namespace absl { 29 ABSL_NAMESPACE_BEGIN 30 namespace flags_internal { 31 32 // Align 'x' up to the nearest 'align' bytes. 33 inline constexpr size_t AlignUp(size_t x, size_t align) { 34 return align * ((x + align - 1) / align); 35 } 36 37 // A SequenceLock implements lock-free reads. A sequence counter is incremented 38 // before and after each write, and readers access the counter before and after 39 // accessing the protected data. If the counter is verified to not change during 40 // the access, and the sequence counter value was even, then the reader knows 41 // that the read was race-free and valid. Otherwise, the reader must fall back 42 // to a Mutex-based code path. 43 // 44 // This particular SequenceLock starts in an "uninitialized" state in which 45 // TryRead() returns false. It must be enabled by calling MarkInitialized(). 46 // This serves as a marker that the associated flag value has not yet been 47 // initialized and a slow path needs to be taken. 48 // 49 // The memory reads and writes protected by this lock must use the provided 50 // `TryRead()` and `Write()` functions. These functions behave similarly to 51 // `memcpy()`, with one oddity: the protected data must be an array of 52 // `std::atomic<uint64>`. This is to comply with the C++ standard, which 53 // considers data races on non-atomic objects to be undefined behavior. See "Can 54 // Seqlocks Get Along With Programming Language Memory Models?"[1] by Hans J. 55 // Boehm for more details. 56 // 57 // [1] https://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf 58 class SequenceLock { 59 public: 60 constexpr SequenceLock() : lock_(kUninitialized) {} 61 62 // Mark that this lock is ready for use. 63 void MarkInitialized() { 64 assert(lock_.load(std::memory_order_relaxed) == kUninitialized); 65 lock_.store(0, std::memory_order_release); 66 } 67 68 // Copy "size" bytes of data from "src" to "dst", protected as a read-side 69 // critical section of the sequence lock. 70 // 71 // Unlike traditional sequence lock implementations which loop until getting a 72 // clean read, this implementation returns false in the case of concurrent 73 // calls to `Write`. In such a case, the caller should fall back to a 74 // locking-based slow path. 75 // 76 // Returns false if the sequence lock was not yet marked as initialized. 77 // 78 // NOTE: If this returns false, "dst" may be overwritten with undefined 79 // (potentially uninitialized) data. 80 bool TryRead(void* dst, const std::atomic<uint64_t>* src, size_t size) const { 81 // Acquire barrier ensures that no loads done by f() are reordered 82 // above the first load of the sequence counter. 83 int64_t seq_before = lock_.load(std::memory_order_acquire); 84 if (ABSL_PREDICT_FALSE(seq_before & 1) == 1) return false; 85 RelaxedCopyFromAtomic(dst, src, size); 86 // Another acquire fence ensures that the load of 'lock_' below is 87 // strictly ordered after the RelaxedCopyToAtomic call above. 88 std::atomic_thread_fence(std::memory_order_acquire); 89 int64_t seq_after = lock_.load(std::memory_order_relaxed); 90 return ABSL_PREDICT_TRUE(seq_before == seq_after); 91 } 92 93 // Copy "size" bytes from "src" to "dst" as a write-side critical section 94 // of the sequence lock. Any concurrent readers will be forced to retry 95 // until they get a read that does not conflict with this write. 96 // 97 // This call must be externally synchronized against other calls to Write, 98 // but may proceed concurrently with reads. 99 void Write(std::atomic<uint64_t>* dst, const void* src, size_t size) { 100 // We can use relaxed instructions to increment the counter since we 101 // are extenally synchronized. The std::atomic_thread_fence below 102 // ensures that the counter updates don't get interleaved with the 103 // copy to the data. 104 int64_t orig_seq = lock_.load(std::memory_order_relaxed); 105 assert((orig_seq & 1) == 0); // Must be initially unlocked. 106 lock_.store(orig_seq + 1, std::memory_order_relaxed); 107 108 // We put a release fence between update to lock_ and writes to shared data. 109 // Thus all stores to shared data are effectively release operations and 110 // update to lock_ above cannot be re-ordered past any of them. Note that 111 // this barrier is not for the fetch_add above. A release barrier for the 112 // fetch_add would be before it, not after. 113 std::atomic_thread_fence(std::memory_order_release); 114 RelaxedCopyToAtomic(dst, src, size); 115 // "Release" semantics ensure that none of the writes done by 116 // RelaxedCopyToAtomic() can be reordered after the following modification. 117 lock_.store(orig_seq + 2, std::memory_order_release); 118 } 119 120 // Return the number of times that Write() has been called. 121 // 122 // REQUIRES: This must be externally synchronized against concurrent calls to 123 // `Write()` or `IncrementModificationCount()`. 124 // REQUIRES: `MarkInitialized()` must have been previously called. 125 int64_t ModificationCount() const { 126 int64_t val = lock_.load(std::memory_order_relaxed); 127 assert(val != kUninitialized && (val & 1) == 0); 128 return val / 2; 129 } 130 131 // REQUIRES: This must be externally synchronized against concurrent calls to 132 // `Write()` or `ModificationCount()`. 133 // REQUIRES: `MarkInitialized()` must have been previously called. 134 void IncrementModificationCount() { 135 int64_t val = lock_.load(std::memory_order_relaxed); 136 assert(val != kUninitialized); 137 lock_.store(val + 2, std::memory_order_relaxed); 138 } 139 140 private: 141 // Perform the equivalent of "memcpy(dst, src, size)", but using relaxed 142 // atomics. 143 static void RelaxedCopyFromAtomic(void* dst, const std::atomic<uint64_t>* src, 144 size_t size) { 145 char* dst_byte = static_cast<char*>(dst); 146 while (size >= sizeof(uint64_t)) { 147 uint64_t word = src->load(std::memory_order_relaxed); 148 std::memcpy(dst_byte, &word, sizeof(word)); 149 dst_byte += sizeof(word); 150 src++; 151 size -= sizeof(word); 152 } 153 if (size > 0) { 154 uint64_t word = src->load(std::memory_order_relaxed); 155 std::memcpy(dst_byte, &word, size); 156 } 157 } 158 159 // Perform the equivalent of "memcpy(dst, src, size)", but using relaxed 160 // atomics. 161 static void RelaxedCopyToAtomic(std::atomic<uint64_t>* dst, const void* src, 162 size_t size) { 163 const char* src_byte = static_cast<const char*>(src); 164 while (size >= sizeof(uint64_t)) { 165 uint64_t word; 166 std::memcpy(&word, src_byte, sizeof(word)); 167 dst->store(word, std::memory_order_relaxed); 168 src_byte += sizeof(word); 169 dst++; 170 size -= sizeof(word); 171 } 172 if (size > 0) { 173 uint64_t word = 0; 174 std::memcpy(&word, src_byte, size); 175 dst->store(word, std::memory_order_relaxed); 176 } 177 } 178 179 static constexpr int64_t kUninitialized = -1; 180 std::atomic<int64_t> lock_; 181 }; 182 183 } // namespace flags_internal 184 ABSL_NAMESPACE_END 185 } // namespace absl 186 187 #endif // ABSL_FLAGS_INTERNAL_SEQUENCE_LOCK_H_