mfbt/SPSCQueue.h

   1 /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
   2 /* vim: set ts=8 sts=2 et sw=2 tw=80: */
   3 /* This Source Code Form is subject to the terms of the Mozilla Public
   4  * License, v. 2.0. If a copy of the MPL was not distributed with this
   5  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
   6
   7 /* Single producer single consumer lock-free and wait-free queue. */
   8
   9 #ifndef mozilla_LockFreeQueue_h
  10 #define mozilla_LockFreeQueue_h
  11
  12 #include "mozilla/Assertions.h"
  13 #include "mozilla/Attributes.h"
  14 #include "mozilla/PodOperations.h"
  15 #include <algorithm>
  16 #include <atomic>
  17 #include <cstddef>
  18 #include <limits>
  19 #include <memory>
  20 #include <thread>
  21 #include <type_traits>
  22
  23 namespace mozilla {
  24
  25 namespace detail {
  26 template <typename T, bool IsPod = std::is_trivial<T>::value>
  27 struct MemoryOperations {
  28   /**
  29    * This allows zeroing (using memset) or default-constructing a number of
  30    * elements calling the constructors if necessary.
  31    */
  32   static void ConstructDefault(T* aDestination, size_t aCount);
  33   /**
  34    * This allows either moving (if T supports it) or copying a number of
  35    * elements from a `aSource` pointer to a `aDestination` pointer.
  36    * If it is safe to do so and this call copies, this uses PodCopy. Otherwise,
  37    * constructors and destructors are called in a loop.
  38    */
  39   static void MoveOrCopy(T* aDestination, T* aSource, size_t aCount);
  40 };
  41
  42 template <typename T>
  43 struct MemoryOperations<T, true> {
  44   static void ConstructDefault(T* aDestination, size_t aCount) {
  45     PodZero(aDestination, aCount);
  46   }
  47   static void MoveOrCopy(T* aDestination, T* aSource, size_t aCount) {
  48     PodCopy(aDestination, aSource, aCount);
  49   }
  50 };
  51
  52 template <typename T>
  53 struct MemoryOperations<T, false> {
  54   static void ConstructDefault(T* aDestination, size_t aCount) {
  55     for (size_t i = 0; i < aCount; i++) {
  56       aDestination[i] = T();
  57     }
  58   }
  59   static void MoveOrCopy(T* aDestination, T* aSource, size_t aCount) {
  60     std::move(aSource, aSource + aCount, aDestination);
  61   }
  62 };
  63 }  // namespace detail
  64
  65 /**
  66  * This data structure allows producing data from one thread, and consuming it
  67  * on another thread, safely and without explicit synchronization.
  68  *
  69  * The role for the producer and the consumer must be constant, i.e., the
  70  * producer should always be on one thread and the consumer should always be on
  71  * another thread.
  72  *
  73  * Some words about the inner workings of this class:
  74  * - Capacity is fixed. Only one allocation is performed, in the constructor.
  75  *   When reading and writing, the return value of the method allows checking if
  76  *   the ring buffer is empty or full.
  77  * - We always keep the read index at least one element ahead of the write
  78  *   index, so we can distinguish between an empty and a full ring buffer: an
  79  *   empty ring buffer is when the write index is at the same position as the
  80  *   read index. A full buffer is when the write index is exactly one position
  81  *   before the read index.
  82  * - We synchronize updates to the read index after having read the data, and
  83  *   the write index after having written the data. This means that the each
  84  *   thread can only touch a portion of the buffer that is not touched by the
  85  *   other thread.
  86  * - Callers are expected to provide buffers. When writing to the queue,
  87  *   elements are copied into the internal storage from the buffer passed in.
  88  *   When reading from the queue, the user is expected to provide a buffer.
  89  *   Because this is a ring buffer, data might not be contiguous in memory;
  90  *   providing an external buffer to copy into is an easy way to have linear
  91  *   data for further processing.
  92  */
  93 template <typename T>
  94 class SPSCRingBufferBase {
  95  public:
  96   /**
  97    * Constructor for a ring buffer.
  98    *
  99    * This performs an allocation on the heap, but is the only allocation that
 100    * will happen for the life time of a `SPSCRingBufferBase`.
 101    *
 102    * @param Capacity The maximum number of element this ring buffer will hold.
 103    */
 104   explicit SPSCRingBufferBase(int aCapacity)
 105       : mReadIndex(0),
 106         mWriteIndex(0)
 107         /* One more element to distinguish from empty and full buffer. */
 108         ,
 109         mCapacity(aCapacity + 1) {
 110     MOZ_ASSERT(StorageCapacity() < std::numeric_limits<int>::max() / 2,
 111                "buffer too large for the type of index used.");
 112     MOZ_ASSERT(mCapacity > 0 && aCapacity != std::numeric_limits<int>::max());
 113
 114     mData = std::make_unique<T[]>(StorageCapacity());
 115
 116     std::atomic_thread_fence(std::memory_order_seq_cst);
 117   }
 118   /**
 119    * Push `aCount` zero or default constructed elements in the array.
 120    *
 121    * Only safely called on the producer thread.
 122    *
 123    * @param count The number of elements to enqueue.
 124    * @return The number of element enqueued.
 125    */
 126   [[nodiscard]] int EnqueueDefault(int aCount) {
 127     return Enqueue(nullptr, aCount);
 128   }
 129   /**
 130    * @brief Put an element in the queue.
 131    *
 132    * Only safely called on the producer thread.
 133    *
 134    * @param element The element to put in the queue.
 135    *
 136    * @return 1 if the element was inserted, 0 otherwise.
 137    */
 138   [[nodiscard]] int Enqueue(T& aElement) { return Enqueue(&aElement, 1); }
 139   /**
 140    * Push `aCount` elements in the ring buffer.
 141    *
 142    * Only safely called on the producer thread.
 143    *
 144    * @param elements a pointer to a buffer containing at least `count` elements.
 145    * If `elements` is nullptr, zero or default constructed elements are enqueud.
 146    * @param count The number of elements to read from `elements`
 147    * @return The number of elements successfully coped from `elements` and
 148    * inserted into the ring buffer.
 149    */
 150   [[nodiscard]] int Enqueue(T* aElements, int aCount) {
 151 #ifdef DEBUG
 152     AssertCorrectThread(mProducerId);
 153 #endif
 154
 155     int rdIdx = mReadIndex.load(std::memory_order_acquire);
 156     int wrIdx = mWriteIndex.load(std::memory_order_relaxed);
 157
 158     if (IsFull(rdIdx, wrIdx)) {
 159       return 0;
 160     }
 161
 162     int toWrite = std::min(AvailableWriteInternal(rdIdx, wrIdx), aCount);
 163
 164     /* First part, from the write index to the end of the array. */
 165     int firstPart = std::min(StorageCapacity() - wrIdx, toWrite);
 166     /* Second part, from the beginning of the array */
 167     int secondPart = toWrite - firstPart;
 168
 169     if (aElements) {
 170       detail::MemoryOperations<T>::MoveOrCopy(mData.get() + wrIdx, aElements,
 171                                               firstPart);
 172       detail::MemoryOperations<T>::MoveOrCopy(
 173           mData.get(), aElements + firstPart, secondPart);
 174     } else {
 175       detail::MemoryOperations<T>::ConstructDefault(mData.get() + wrIdx,
 176                                                     firstPart);
 177       detail::MemoryOperations<T>::ConstructDefault(mData.get(), secondPart);
 178     }
 179
 180     mWriteIndex.store(IncrementIndex(wrIdx, toWrite),
 181                       std::memory_order_release);
 182
 183     return toWrite;
 184   }
 185   /**
 186    * Retrieve at most `count` elements from the ring buffer, and copy them to
 187    * `elements`, if non-null.
 188    *
 189    * Only safely called on the consumer side.
 190    *
 191    * @param elements A pointer to a buffer with space for at least `count`
 192    * elements. If `elements` is `nullptr`, `count` element will be discarded.
 193    * @param count The maximum number of elements to Dequeue.
 194    * @return The number of elements written to `elements`.
 195    */
 196   [[nodiscard]] int Dequeue(T* elements, int count) {
 197 #ifdef DEBUG
 198     AssertCorrectThread(mConsumerId);
 199 #endif
 200
 201     int wrIdx = mWriteIndex.load(std::memory_order_acquire);
 202     int rdIdx = mReadIndex.load(std::memory_order_relaxed);
 203
 204     if (IsEmpty(rdIdx, wrIdx)) {
 205       return 0;
 206     }
 207
 208     int toRead = std::min(AvailableReadInternal(rdIdx, wrIdx), count);
 209
 210     int firstPart = std::min(StorageCapacity() - rdIdx, toRead);
 211     int secondPart = toRead - firstPart;
 212
 213     if (elements) {
 214       detail::MemoryOperations<T>::MoveOrCopy(elements, mData.get() + rdIdx,
 215                                               firstPart);
 216       detail::MemoryOperations<T>::MoveOrCopy(elements + firstPart, mData.get(),
 217                                               secondPart);
 218     }
 219
 220     mReadIndex.store(IncrementIndex(rdIdx, toRead), std::memory_order_release);
 221
 222     return toRead;
 223   }
 224   /**
 225    * Get the number of available elements for consuming.
 226    *
 227    * This can be less than the actual number of elements in the queue, since the
 228    * mWriteIndex is updated at the very end of the Enqueue method on the
 229    * producer thread, but consequently always returns a number of elements such
 230    * that a call to Dequeue return this number of elements.
 231    *
 232    * @return The number of available elements for reading.
 233    */
 234   int AvailableRead() const {
 235     return AvailableReadInternal(mReadIndex.load(std::memory_order_relaxed),
 236                                  mWriteIndex.load(std::memory_order_relaxed));
 237   }
 238   /**
 239    * Get the number of available elements for writing.
 240    *
 241    * This can be less than than the actual number of slots that are available,
 242    * because mReadIndex is updated at the very end of the Deque method. It
 243    * always returns a number such that a call to Enqueue with this number will
 244    * succeed in enqueuing this number of elements.
 245    *
 246    * @return The number of empty slots in the buffer, available for writing.
 247    */
 248   int AvailableWrite() const {
 249     return AvailableWriteInternal(mReadIndex.load(std::memory_order_relaxed),
 250                                   mWriteIndex.load(std::memory_order_relaxed));
 251   }
 252   /**
 253    * Get the total Capacity, for this ring buffer.
 254    *
 255    * Can be called safely on any thread.
 256    *
 257    * @return The maximum Capacity of this ring buffer.
 258    */
 259   int Capacity() const { return StorageCapacity() - 1; }
 260   /**
 261    * Reset the consumer and producer thread identifier, in case the threads are
 262    * being changed. This has to be externally synchronized. This is no-op when
 263    * asserts are disabled.
 264    */
 265   void ResetThreadIds() {
 266     ResetProducerThreadId();
 267     ResetConsumerThreadId();
 268   }
 269
 270   void ResetConsumerThreadId() {
 271 #ifdef DEBUG
 272     mConsumerId = std::thread::id();
 273 #endif
 274   }
 275
 276   void ResetProducerThreadId() {
 277 #ifdef DEBUG
 278     mProducerId = std::thread::id();
 279 #endif
 280   }
 281
 282  private:
 283   /** Return true if the ring buffer is empty.
 284    *
 285    * This can be called from the consumer or the producer thread.
 286    *
 287    * @param aReadIndex the read index to consider
 288    * @param writeIndex the write index to consider
 289    * @return true if the ring buffer is empty, false otherwise.
 290    **/
 291   bool IsEmpty(int aReadIndex, int aWriteIndex) const {
 292     return aWriteIndex == aReadIndex;
 293   }
 294   /** Return true if the ring buffer is full.
 295    *
 296    * This happens if the write index is exactly one element behind the read
 297    * index.
 298    *
 299    * This can be called from the consummer or the producer thread.
 300    *
 301    * @param aReadIndex the read index to consider
 302    * @param writeIndex the write index to consider
 303    * @return true if the ring buffer is full, false otherwise.
 304    **/
 305   bool IsFull(int aReadIndex, int aWriteIndex) const {
 306     return (aWriteIndex + 1) % StorageCapacity() == aReadIndex;
 307   }
 308   /**
 309    * Return the size of the storage. It is one more than the number of elements
 310    * that can be stored in the buffer.
 311    *
 312    * This can be called from any thread.
 313    *
 314    * @return the number of elements that can be stored in the buffer.
 315    */
 316   int StorageCapacity() const { return mCapacity; }
 317   /**
 318    * Returns the number of elements available for reading.
 319    *
 320    * This can be called from the consummer or producer thread, but see the
 321    * comment in `AvailableRead`.
 322    *
 323    * @return the number of available elements for reading.
 324    */
 325   int AvailableReadInternal(int aReadIndex, int aWriteIndex) const {
 326     if (aWriteIndex >= aReadIndex) {
 327       return aWriteIndex - aReadIndex;
 328     } else {
 329       return aWriteIndex + StorageCapacity() - aReadIndex;
 330     }
 331   }
 332   /**
 333    * Returns the number of empty elements, available for writing.
 334    *
 335    * This can be called from the consummer or producer thread, but see the
 336    * comment in `AvailableWrite`.
 337    *
 338    * @return the number of elements that can be written into the array.
 339    */
 340   int AvailableWriteInternal(int aReadIndex, int aWriteIndex) const {
 341     /* We subtract one element here to always keep at least one sample
 342      * free in the buffer, to distinguish between full and empty array. */
 343     int rv = aReadIndex - aWriteIndex - 1;
 344     if (aWriteIndex >= aReadIndex) {
 345       rv += StorageCapacity();
 346     }
 347     return rv;
 348   }
 349   /**
 350    * Increments an index, wrapping it around the storage.
 351    *
 352    * Incrementing `mWriteIndex` can be done on the producer thread.
 353    * Incrementing `mReadIndex` can be done on the consummer thread.
 354    *
 355    * @param index a reference to the index to increment.
 356    * @param increment the number by which `index` is incremented.
 357    * @return the new index.
 358    */
 359   int IncrementIndex(int aIndex, int aIncrement) const {
 360     MOZ_ASSERT(aIncrement >= 0 && aIncrement < StorageCapacity() &&
 361                aIndex < StorageCapacity());
 362     return (aIndex + aIncrement) % StorageCapacity();
 363   }
 364   /**
 365    * @brief This allows checking that Enqueue (resp. Dequeue) are always
 366    * called by the right thread.
 367    *
 368    * The role of the thread are assigned the first time they call Enqueue or
 369    * Dequeue, and cannot change, except when ResetThreadIds is called..
 370    *
 371    * @param id the id of the thread that has called the calling method first.
 372    */
 373 #ifdef DEBUG
 374   static void AssertCorrectThread(std::thread::id& aId) {
 375     if (aId == std::thread::id()) {
 376       aId = std::this_thread::get_id();
 377       return;
 378     }
 379     MOZ_ASSERT(aId == std::this_thread::get_id());
 380   }
 381 #endif
 382   /** Index at which the oldest element is. */
 383   std::atomic<int> mReadIndex;
 384   /** Index at which to write new elements. `mWriteIndex` is always at
 385    * least one element ahead of `mReadIndex`. */
 386   std::atomic<int> mWriteIndex;
 387   /** Maximum number of elements that can be stored in the ring buffer. */
 388   const int mCapacity;
 389   /** Data storage, of size `mCapacity + 1` */
 390   std::unique_ptr<T[]> mData;
 391 #ifdef DEBUG
 392   /** The id of the only thread that is allowed to read from the queue. */
 393   mutable std::thread::id mConsumerId;
 394   /** The id of the only thread that is allowed to write from the queue. */
 395   mutable std::thread::id mProducerId;
 396 #endif
 397 };
 398
 399 /**
 400  * Instantiation of the `SPSCRingBufferBase` type. This is safe to use
 401  * from two threads, one producer, one consumer (that never change role),
 402  * without explicit synchronization.
 403  */
 404 template <typename T>
 405 using SPSCQueue = SPSCRingBufferBase<T>;
 406
 407 }  // namespace mozilla
 408
 409 #endif  // mozilla_LockFreeQueue_h