mirror of
https://github.com/mozilla/gecko-dev.git
synced 2024-11-27 14:52:16 +00:00
76a7426f90
Without this, the necessary synchronization must be provided externally. This fixes the memory order in the following case of changing producer thread: - Thread A does SPSCQueue::Enqueue - non-atomic write into the ring buffer, at memory location X - mWriteIndex.load(relaxed) - mWriteIndex.store(release) - Producer thread is switched to B, no external memory order synchronization is provided, but thread B is guaranteed to run after thread A has finished its Enqueue task. - Thread B does SPSCQueue::Enqueue - mWriteIndex.load(relaxed) - mWriteIndex.store(release) - Thread C does SPSCQueue::Dequeue - mWriteIndex.load(acquire) - non-atomic read from the ring buffer, at memory location X In this scenario, there is no memory synchronization between threads A and B, and therefore the non-atomic read on C is a data race, and flagged as such by TSAN. A similar scenario can be applied to changing the consumer thread, if first A enqueues, then B dequeues, then C dequeues. However, since Dequeue doesn't necessarily (MoveOrCopy) do non-atomic writes to the ring buffer, and more importantly, since Enqueue doesn't do non-atomic reads from the ring buffer, this is less of a problem. Differential Revision: https://phabricator.services.mozilla.com/D190084
421 lines
15 KiB
C++
421 lines
15 KiB
C++
/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
|
|
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
|
|
/* This Source Code Form is subject to the terms of the Mozilla Public
|
|
* License, v. 2.0. If a copy of the MPL was not distributed with this
|
|
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
|
|
|
|
/* Single producer single consumer lock-free and wait-free queue. */
|
|
|
|
#ifndef mozilla_LockFreeQueue_h
|
|
#define mozilla_LockFreeQueue_h
|
|
|
|
#include "mozilla/Assertions.h"
|
|
#include "mozilla/Attributes.h"
|
|
#include "mozilla/PodOperations.h"
|
|
#include <algorithm>
|
|
#include <atomic>
|
|
#include <cstddef>
|
|
#include <limits>
|
|
#include <memory>
|
|
#include <thread>
|
|
#include <type_traits>
|
|
|
|
namespace mozilla {
|
|
|
|
namespace detail {
|
|
template <typename T, bool IsPod = std::is_trivial<T>::value>
|
|
struct MemoryOperations {
|
|
/**
|
|
* This allows zeroing (using memset) or default-constructing a number of
|
|
* elements calling the constructors if necessary.
|
|
*/
|
|
static void ConstructDefault(T* aDestination, size_t aCount);
|
|
/**
|
|
* This allows either moving (if T supports it) or copying a number of
|
|
* elements from a `aSource` pointer to a `aDestination` pointer.
|
|
* If it is safe to do so and this call copies, this uses PodCopy. Otherwise,
|
|
* constructors and destructors are called in a loop.
|
|
*/
|
|
static void MoveOrCopy(T* aDestination, T* aSource, size_t aCount);
|
|
};
|
|
|
|
template <typename T>
|
|
struct MemoryOperations<T, true> {
|
|
static void ConstructDefault(T* aDestination, size_t aCount) {
|
|
PodZero(aDestination, aCount);
|
|
}
|
|
static void MoveOrCopy(T* aDestination, T* aSource, size_t aCount) {
|
|
PodCopy(aDestination, aSource, aCount);
|
|
}
|
|
};
|
|
|
|
template <typename T>
|
|
struct MemoryOperations<T, false> {
|
|
static void ConstructDefault(T* aDestination, size_t aCount) {
|
|
for (size_t i = 0; i < aCount; i++) {
|
|
aDestination[i] = T();
|
|
}
|
|
}
|
|
static void MoveOrCopy(T* aDestination, T* aSource, size_t aCount) {
|
|
std::move(aSource, aSource + aCount, aDestination);
|
|
}
|
|
};
|
|
} // namespace detail
|
|
|
|
/**
|
|
* This data structure allows producing data from one thread, and consuming it
|
|
* on another thread, safely and without explicit synchronization.
|
|
*
|
|
* The role for the producer and the consumer must be constant, i.e., the
|
|
* producer should always be on one thread and the consumer should always be on
|
|
* another thread.
|
|
*
|
|
* Some words about the inner workings of this class:
|
|
* - Capacity is fixed. Only one allocation is performed, in the constructor.
|
|
* When reading and writing, the return value of the method allows checking if
|
|
* the ring buffer is empty or full.
|
|
* - We always keep the read index at least one element ahead of the write
|
|
* index, so we can distinguish between an empty and a full ring buffer: an
|
|
* empty ring buffer is when the write index is at the same position as the
|
|
* read index. A full buffer is when the write index is exactly one position
|
|
* before the read index.
|
|
* - We synchronize updates to the read index after having read the data, and
|
|
* the write index after having written the data. This means that the each
|
|
* thread can only touch a portion of the buffer that is not touched by the
|
|
* other thread.
|
|
* - Callers are expected to provide buffers. When writing to the queue,
|
|
* elements are copied into the internal storage from the buffer passed in.
|
|
* When reading from the queue, the user is expected to provide a buffer.
|
|
* Because this is a ring buffer, data might not be contiguous in memory;
|
|
* providing an external buffer to copy into is an easy way to have linear
|
|
* data for further processing.
|
|
*/
|
|
template <typename T>
|
|
class SPSCRingBufferBase {
|
|
public:
|
|
/**
|
|
* Constructor for a ring buffer.
|
|
*
|
|
* This performs an allocation on the heap, but is the only allocation that
|
|
* will happen for the life time of a `SPSCRingBufferBase`.
|
|
*
|
|
* @param Capacity The maximum number of element this ring buffer will hold.
|
|
*/
|
|
explicit SPSCRingBufferBase(int aCapacity)
|
|
: mReadIndex(0),
|
|
mWriteIndex(0),
|
|
/* One more element to distinguish from empty and full buffer. */
|
|
mCapacity(aCapacity + 1) {
|
|
MOZ_RELEASE_ASSERT(aCapacity != std::numeric_limits<int>::max());
|
|
MOZ_RELEASE_ASSERT(mCapacity > 0);
|
|
|
|
mData = std::make_unique<T[]>(StorageCapacity());
|
|
|
|
std::atomic_thread_fence(std::memory_order_seq_cst);
|
|
}
|
|
/**
|
|
* Push `aCount` zero or default constructed elements in the array.
|
|
*
|
|
* Only safely called on the producer thread.
|
|
*
|
|
* @param count The number of elements to enqueue.
|
|
* @return The number of element enqueued.
|
|
*/
|
|
[[nodiscard]] int EnqueueDefault(int aCount) {
|
|
return Enqueue(nullptr, aCount);
|
|
}
|
|
/**
|
|
* @brief Put an element in the queue.
|
|
*
|
|
* Only safely called on the producer thread.
|
|
*
|
|
* @param element The element to put in the queue.
|
|
*
|
|
* @return 1 if the element was inserted, 0 otherwise.
|
|
*/
|
|
[[nodiscard]] int Enqueue(T& aElement) { return Enqueue(&aElement, 1); }
|
|
/**
|
|
* Push `aCount` elements in the ring buffer.
|
|
*
|
|
* Only safely called on the producer thread.
|
|
*
|
|
* @param elements a pointer to a buffer containing at least `count` elements.
|
|
* If `elements` is nullptr, zero or default constructed elements are enqueud.
|
|
* @param count The number of elements to read from `elements`
|
|
* @return The number of elements successfully coped from `elements` and
|
|
* inserted into the ring buffer.
|
|
*/
|
|
[[nodiscard]] int Enqueue(T* aElements, int aCount) {
|
|
#ifdef DEBUG
|
|
AssertCorrectThread(mProducerId);
|
|
#endif
|
|
|
|
int rdIdx = mReadIndex.load(std::memory_order_acquire);
|
|
int wrIdx = mWriteIndex.load(std::memory_order_relaxed);
|
|
|
|
if (IsFull(rdIdx, wrIdx)) {
|
|
return 0;
|
|
}
|
|
|
|
int toWrite = std::min(AvailableWriteInternal(rdIdx, wrIdx), aCount);
|
|
|
|
/* First part, from the write index to the end of the array. */
|
|
int firstPart = std::min(StorageCapacity() - wrIdx, toWrite);
|
|
/* Second part, from the beginning of the array */
|
|
int secondPart = toWrite - firstPart;
|
|
|
|
if (aElements) {
|
|
detail::MemoryOperations<T>::MoveOrCopy(mData.get() + wrIdx, aElements,
|
|
firstPart);
|
|
detail::MemoryOperations<T>::MoveOrCopy(
|
|
mData.get(), aElements + firstPart, secondPart);
|
|
} else {
|
|
detail::MemoryOperations<T>::ConstructDefault(mData.get() + wrIdx,
|
|
firstPart);
|
|
detail::MemoryOperations<T>::ConstructDefault(mData.get(), secondPart);
|
|
}
|
|
|
|
mWriteIndex.store(IncrementIndex(wrIdx, toWrite),
|
|
std::memory_order_release);
|
|
|
|
return toWrite;
|
|
}
|
|
/**
|
|
* Retrieve at most `count` elements from the ring buffer, and copy them to
|
|
* `elements`, if non-null.
|
|
*
|
|
* Only safely called on the consumer side.
|
|
*
|
|
* @param elements A pointer to a buffer with space for at least `count`
|
|
* elements. If `elements` is `nullptr`, `count` element will be discarded.
|
|
* @param count The maximum number of elements to Dequeue.
|
|
* @return The number of elements written to `elements`.
|
|
*/
|
|
[[nodiscard]] int Dequeue(T* elements, int count) {
|
|
#ifdef DEBUG
|
|
AssertCorrectThread(mConsumerId);
|
|
#endif
|
|
|
|
int wrIdx = mWriteIndex.load(std::memory_order_acquire);
|
|
int rdIdx = mReadIndex.load(std::memory_order_relaxed);
|
|
|
|
if (IsEmpty(rdIdx, wrIdx)) {
|
|
return 0;
|
|
}
|
|
|
|
int toRead = std::min(AvailableReadInternal(rdIdx, wrIdx), count);
|
|
|
|
int firstPart = std::min(StorageCapacity() - rdIdx, toRead);
|
|
int secondPart = toRead - firstPart;
|
|
|
|
if (elements) {
|
|
detail::MemoryOperations<T>::MoveOrCopy(elements, mData.get() + rdIdx,
|
|
firstPart);
|
|
detail::MemoryOperations<T>::MoveOrCopy(elements + firstPart, mData.get(),
|
|
secondPart);
|
|
}
|
|
|
|
mReadIndex.store(IncrementIndex(rdIdx, toRead), std::memory_order_release);
|
|
|
|
return toRead;
|
|
}
|
|
/**
|
|
* Get the number of available elements for consuming.
|
|
*
|
|
* This can be less than the actual number of elements in the queue, since the
|
|
* mWriteIndex is updated at the very end of the Enqueue method on the
|
|
* producer thread, but consequently always returns a number of elements such
|
|
* that a call to Dequeue return this number of elements.
|
|
*
|
|
* @return The number of available elements for reading.
|
|
*/
|
|
int AvailableRead() const {
|
|
return AvailableReadInternal(mReadIndex.load(std::memory_order_relaxed),
|
|
mWriteIndex.load(std::memory_order_relaxed));
|
|
}
|
|
/**
|
|
* Get the number of available elements for writing.
|
|
*
|
|
* This can be less than than the actual number of slots that are available,
|
|
* because mReadIndex is updated at the very end of the Deque method. It
|
|
* always returns a number such that a call to Enqueue with this number will
|
|
* succeed in enqueuing this number of elements.
|
|
*
|
|
* @return The number of empty slots in the buffer, available for writing.
|
|
*/
|
|
int AvailableWrite() const {
|
|
return AvailableWriteInternal(mReadIndex.load(std::memory_order_relaxed),
|
|
mWriteIndex.load(std::memory_order_relaxed));
|
|
}
|
|
/**
|
|
* Get the total Capacity, for this ring buffer.
|
|
*
|
|
* Can be called safely on any thread.
|
|
*
|
|
* @return The maximum Capacity of this ring buffer.
|
|
*/
|
|
int Capacity() const { return StorageCapacity() - 1; }
|
|
|
|
/**
|
|
* Reset the consumer thread id to the current thread. The caller must
|
|
* guarantee that the last call to Dequeue() on the previous consumer thread
|
|
* has completed, and subsequent calls to Dequeue() will only happen on the
|
|
* current thread.
|
|
*/
|
|
void ResetConsumerThreadId() {
|
|
#ifdef DEBUG
|
|
mConsumerId = std::this_thread::get_id();
|
|
#endif
|
|
|
|
// When changing consumer from thread A to B, the last Dequeue on A (synced
|
|
// by mReadIndex.store with memory_order_release) must be picked up by B
|
|
// through an acquire operation.
|
|
std::ignore = mReadIndex.load(std::memory_order_acquire);
|
|
}
|
|
|
|
/**
|
|
* Reset the producer thread id to the current thread. The caller must
|
|
* guarantee that the last call to Enqueue() on the previous consumer thread
|
|
* has completed, and subsequent calls to Dequeue() will only happen on the
|
|
* current thread.
|
|
*/
|
|
void ResetProducerThreadId() {
|
|
#ifdef DEBUG
|
|
mProducerId = std::this_thread::get_id();
|
|
#endif
|
|
|
|
// When changing producer from thread A to B, the last Enqueue on A (synced
|
|
// by mWriteIndex.store with memory_order_release) must be picked up by B
|
|
// through an acquire operation.
|
|
std::ignore = mWriteIndex.load(std::memory_order_acquire);
|
|
}
|
|
|
|
private:
|
|
/** Return true if the ring buffer is empty.
|
|
*
|
|
* This can be called from the consumer or the producer thread.
|
|
*
|
|
* @param aReadIndex the read index to consider
|
|
* @param writeIndex the write index to consider
|
|
* @return true if the ring buffer is empty, false otherwise.
|
|
**/
|
|
bool IsEmpty(int aReadIndex, int aWriteIndex) const {
|
|
return aWriteIndex == aReadIndex;
|
|
}
|
|
/** Return true if the ring buffer is full.
|
|
*
|
|
* This happens if the write index is exactly one element behind the read
|
|
* index.
|
|
*
|
|
* This can be called from the consummer or the producer thread.
|
|
*
|
|
* @param aReadIndex the read index to consider
|
|
* @param writeIndex the write index to consider
|
|
* @return true if the ring buffer is full, false otherwise.
|
|
**/
|
|
bool IsFull(int aReadIndex, int aWriteIndex) const {
|
|
return (aWriteIndex + 1) % StorageCapacity() == aReadIndex;
|
|
}
|
|
/**
|
|
* Return the size of the storage. It is one more than the number of elements
|
|
* that can be stored in the buffer.
|
|
*
|
|
* This can be called from any thread.
|
|
*
|
|
* @return the number of elements that can be stored in the buffer.
|
|
*/
|
|
int StorageCapacity() const { return mCapacity; }
|
|
/**
|
|
* Returns the number of elements available for reading.
|
|
*
|
|
* This can be called from the consummer or producer thread, but see the
|
|
* comment in `AvailableRead`.
|
|
*
|
|
* @return the number of available elements for reading.
|
|
*/
|
|
int AvailableReadInternal(int aReadIndex, int aWriteIndex) const {
|
|
if (aWriteIndex >= aReadIndex) {
|
|
return aWriteIndex - aReadIndex;
|
|
} else {
|
|
return aWriteIndex + StorageCapacity() - aReadIndex;
|
|
}
|
|
}
|
|
/**
|
|
* Returns the number of empty elements, available for writing.
|
|
*
|
|
* This can be called from the consummer or producer thread, but see the
|
|
* comment in `AvailableWrite`.
|
|
*
|
|
* @return the number of elements that can be written into the array.
|
|
*/
|
|
int AvailableWriteInternal(int aReadIndex, int aWriteIndex) const {
|
|
/* We subtract one element here to always keep at least one sample
|
|
* free in the buffer, to distinguish between full and empty array. */
|
|
int rv = aReadIndex - aWriteIndex - 1;
|
|
if (aWriteIndex >= aReadIndex) {
|
|
rv += StorageCapacity();
|
|
}
|
|
return rv;
|
|
}
|
|
/**
|
|
* Increments an index, wrapping it around the storage.
|
|
*
|
|
* Incrementing `mWriteIndex` can be done on the producer thread.
|
|
* Incrementing `mReadIndex` can be done on the consummer thread.
|
|
*
|
|
* @param index a reference to the index to increment.
|
|
* @param increment the number by which `index` is incremented.
|
|
* @return the new index.
|
|
*/
|
|
int IncrementIndex(int aIndex, int aIncrement) const {
|
|
MOZ_ASSERT(aIncrement >= 0 && aIncrement < StorageCapacity() &&
|
|
aIndex < StorageCapacity());
|
|
return (aIndex + aIncrement) % StorageCapacity();
|
|
}
|
|
/**
|
|
* @brief This allows checking that Enqueue (resp. Dequeue) are always
|
|
* called by the right thread.
|
|
*
|
|
* The role of the thread are assigned the first time they call Enqueue or
|
|
* Dequeue, and cannot change, except by a ResetThreadId method.
|
|
*
|
|
* @param id the id of the thread that has called the calling method first.
|
|
*/
|
|
#ifdef DEBUG
|
|
static void AssertCorrectThread(std::thread::id& aId) {
|
|
if (aId == std::thread::id()) {
|
|
aId = std::this_thread::get_id();
|
|
return;
|
|
}
|
|
MOZ_ASSERT(aId == std::this_thread::get_id());
|
|
}
|
|
#endif
|
|
/** Index at which the oldest element is. */
|
|
std::atomic<int> mReadIndex;
|
|
/** Index at which to write new elements. `mWriteIndex` is always at
|
|
* least one element ahead of `mReadIndex`. */
|
|
std::atomic<int> mWriteIndex;
|
|
/** Maximum number of elements that can be stored in the ring buffer. */
|
|
const int mCapacity;
|
|
/** Data storage, of size `mCapacity + 1` */
|
|
std::unique_ptr<T[]> mData;
|
|
#ifdef DEBUG
|
|
/** The id of the only thread that is allowed to read from the queue. */
|
|
mutable std::thread::id mConsumerId;
|
|
/** The id of the only thread that is allowed to write from the queue. */
|
|
mutable std::thread::id mProducerId;
|
|
#endif
|
|
};
|
|
|
|
/**
|
|
* Instantiation of the `SPSCRingBufferBase` type. This is safe to use
|
|
* from two threads, one producer, one consumer (that never change role),
|
|
* without explicit synchronization.
|
|
*/
|
|
template <typename T>
|
|
using SPSCQueue = SPSCRingBufferBase<T>;
|
|
|
|
} // namespace mozilla
|
|
|
|
#endif // mozilla_LockFreeQueue_h
|