Source code

Revision control

Other Tools

/* -*- 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 */
#ifndef mozilla_MultiWriterQueue_h_
#define mozilla_MultiWriterQueue_h_
#include <cstdint>
#include <utility>
#include "RollingNumber.h"
#include "mozilla/Atomics.h"
#include "mozilla/MemoryReporting.h"
#include "mozilla/Mutex.h"
#include "prthread.h"
namespace mozilla {
// Default reader locking strategy, using a mutex to ensure that concurrent
// PopAll calls won't overlap.
class MultiWriterQueueReaderLocking_Mutex {
: mMutex("MultiWriterQueueReaderLocking_Mutex") {}
void Lock() { mMutex.Lock(); };
void Unlock() { mMutex.Unlock(); };
Mutex mMutex;
// Reader non-locking strategy, trusting that PopAll will never be called
// concurrently (e.g., by only calling it from a specific thread).
class MultiWriterQueueReaderLocking_None {
#ifndef DEBUG
void Lock(){};
void Unlock(){};
// DEBUG-mode checks to catch concurrent misuses.
void Lock() { MOZ_ASSERT(mLocked.compareExchange(false, true)); };
void Unlock() { MOZ_ASSERT(mLocked.compareExchange(true, false)); };
Atomic<bool> mLocked{false};
static constexpr uint32_t MultiWriterQueueDefaultBufferSize = 8192;
// Multi-writer, single-reader queue of elements of type `T`.
// Elements are bunched together in buffers of `BufferSize` elements.
// This queue is heavily optimized for pushing. In most cases pushes will only
// cost a couple of atomic reads and a few non-atomic reads. Worst cases:
// - Once per buffer, a push will allocate or reuse a buffer for later pushes;
// - During the above new-buffer push, other pushes will be blocked.
// By default, popping is protected by mutex; it may be disabled if popping is
// guaranteed never to be concurrent.
// In any case, popping will never negatively impact pushes.
// (However, *not* popping will add runtime costs, as unread buffers will not
// be freed, or made available to future pushes; Push functions provide
// feedback as to when popping would be most efficient.)
template <typename T, uint32_t BufferSize = MultiWriterQueueDefaultBufferSize,
typename ReaderLocking = MultiWriterQueueReaderLocking_Mutex>
class MultiWriterQueue {
static_assert(BufferSize > 0, "0-sized MultiWriterQueue buffer");
// Constructor.
// Allocates the initial buffer that will receive the first `BufferSize`
// elements. Also allocates one reusable buffer, which will definitely be
// needed after the first `BufferSize` elements have been pushed.
// Ideally (if the reader can process each buffer quickly enough), there
// won't be a need for more buffer allocations.
: mBuffersCoverAtLeastUpTo(BufferSize - 1),
mMostRecentBuffer(new Buffer{}),
mReusableBuffers(new Buffer{}),
mAllocatedBuffersStats(2) {}
~MultiWriterQueue() {
auto DestroyList = [](Buffer* aBuffer) {
while (aBuffer) {
Buffer* older = aBuffer->Older();
delete aBuffer;
aBuffer = older;
// We need the index to be order-resistant to overflow, i.e., numbers before
// an overflow should test smaller-than numbers after the overflow.
// This is because we keep pushing elements with increasing Index, and this
// Index is used to find the appropriate buffer based on a range; and this
// need to work smoothly when crossing the overflow boundary.
using Index = RollingNumber<uint32_t>;
// Pushes indicate whether they have just reached the end of a buffer.
using DidReachEndOfBuffer = bool;
// Push new element and call aF on it.
// Element may be in just-created state, or recycled after a PopAll call.
// Atomically thread-safe; in the worst case some pushes may be blocked
// while a new buffer is created/reused for them.
// Returns whether that push reached the end of a buffer; useful if caller
// wants to trigger processing regularly at the most efficient time.
template <typename F>
DidReachEndOfBuffer PushF(F&& aF) {
// Atomically claim ownership of the next available element.
const Index index{mNextElementToWrite++};
// And now go and set that element.
for (;;) {
Index lastIndex{mBuffersCoverAtLeastUpTo};
if (MOZ_UNLIKELY(index == lastIndex)) {
// We have claimed the last element in the current head -> Allocate a
// new head in advance of more pushes. Make it point at the current
// most-recent buffer.
// This whole process is effectively guarded:
// - Later pushes will wait until mBuffersCoverAtLeastUpTo changes to
// one that can accept their claimed index.
// - Readers will stop until the last element is marked as valid.
Buffer* ourBuffer = mMostRecentBuffer;
Buffer* newBuffer = NewBuffer(ourBuffer, index + 1);
// Because we have claimed this very specific index, we should be the
// only one touching the most-recent buffer pointer.
MOZ_ASSERT(mMostRecentBuffer == ourBuffer);
// Just pivot the most-recent buffer pointer to our new buffer.
mMostRecentBuffer = newBuffer;
// Because we have claimed this very specific index, we should be the
// only one touching the buffer coverage watermark.
MOZ_ASSERT(mBuffersCoverAtLeastUpTo == lastIndex.Value());
// Update it to include the just-added most-recent buffer.
mBuffersCoverAtLeastUpTo = index.Value() + BufferSize;
// We know for sure that `ourBuffer` is the correct one for this index.
ourBuffer->SetAndValidateElement(aF, index);
// And indicate that we have reached the end of a buffer.
return true;
if (MOZ_UNLIKELY(index > lastIndex)) {
// We have claimed an element in a yet-unavailable buffer, wait for our
// target buffer to be created (see above).
while (Index(mBuffersCoverAtLeastUpTo) < index) {
// Then loop to examine the new situation.
// Here, we have claimed a number that is covered by current buffers.
// These buffers cannot be destroyed, because our buffer is not filled
// yet (we haven't written in it yet), therefore the reader thread will
// have to stop there (or before) and won't destroy our buffer or more
// recent ones.
MOZ_ASSERT(index < lastIndex);
Buffer* ourBuffer = mMostRecentBuffer;
// In rare situations, another thread may have had the time to create a
// new more-recent buffer, in which case we need to find our older buffer.
while (MOZ_UNLIKELY(index < ourBuffer->Origin())) {
// We assume that older buffers with still-invalid elements (e.g., the
// one we have just claimed) cannot be destroyed.
ourBuffer = ourBuffer->Older();
// Now we can set&validate the claimed element, and indicate that we have
// not reached the end of a buffer.
ourBuffer->SetAndValidateElement(aF, index);
return false;
// Push new element and assign it a value.
// Atomically thread-safe; in the worst case some pushes may be blocked
// while a new buffer is created/reused for them.
// Returns whether that push reached the end of a buffer; useful if caller
// wants to trigger processing regularly at the most efficient time.
DidReachEndOfBuffer Push(const T& aT) {
return PushF([&aT](T& aElement, Index) { aElement = aT; });
// Push new element and move-assign it a value.
// Atomically thread-safe; in the worst case some pushes may be blocked
// while a new buffer is created/reused for them.
// Returns whether that push reached the end of a buffer; useful if caller
// wants to trigger processing regularly at the most efficient time.
DidReachEndOfBuffer Push(T&& aT) {
return PushF([&aT](T& aElement, Index) { aElement = std::move(aT); });
// Pop all elements before the first invalid one, running aF on each of them
// in FIFO order.
// Thread-safety with other PopAll calls is controlled by the `Locking`
// template argument.
// Concurrent pushes are always allowed, because:
// - PopAll won't read elements until valid,
// - Pushes do not interfere with pop-related members -- except for
// mReusableBuffers, which is accessed atomically.
template <typename F>
void PopAll(F&& aF) {
// Destroy every second fully-read buffer.
// TODO: Research a better algorithm, probably based on stats.
bool destroy = false;
for (;;) {
Buffer* b = mOldestBuffer;
// The next element to pop must be in that oldest buffer.
MOZ_ASSERT(mNextElementToPop >= b->Origin());
MOZ_ASSERT(mNextElementToPop < b->Origin() + BufferSize);
// Start reading each element.
if (!b->ReadAndInvalidateAll(aF, mNextElementToPop)) {
// Found an invalid element, stop popping.
// Reached the end of this oldest buffer
MOZ_ASSERT(mNextElementToPop == b->Origin() + BufferSize);
// Delete this oldest buffer.
// Since the last element was valid, it must mean that there is a newer
// buffer.
MOZ_ASSERT(mNextElementToPop == b->Newer()->Origin());
StopUsing(b, destroy);
destroy = !destroy;
// We will loop and start reading the now-oldest buffer.
// Size of all buffers (used, or recyclable), excluding external data.
size_t ShallowSizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
return mAllocatedBuffersStats.Count() * sizeof(Buffer);
struct CountAndWatermark {
int mCount;
int mWatermark;
CountAndWatermark LiveBuffersStats() const { return mLiveBuffersStats.Get(); }
CountAndWatermark ReusableBuffersStats() const {
return mReusableBuffersStats.Get();
CountAndWatermark AllocatedBuffersStats() const {
return mAllocatedBuffersStats.Get();
// Structure containing the element to be stored, and a validity-marker.
class BufferedElement {
// Run aF on an invalid element, and mark it as valid.
template <typename F>
void SetAndValidate(F&& aF, Index aIndex) {
aF(mT, aIndex);
mValid = true;
// Run aF on a valid element and mark it as invalid, return true.
// Return false if element was invalid.
template <typename F>
bool ReadAndInvalidate(F&& aF) {
if (!mValid) {
return false;
mValid = false;
return true;
T mT;
// mValid should be atomically changed to true *after* mT has been written,
// so that the reader can only see valid data.
// ReleaseAcquire, because when set to `true`, we want the just-written mT
// to be visible to the thread reading this `true`; and when set to `false`,
// we want the previous reads to have completed.
Atomic<bool, ReleaseAcquire> mValid{false};
// Buffer contains a sequence of BufferedElements starting at a specific
// index, and it points to the next-older buffer (if any).
class Buffer {
// Constructor of the very first buffer.
Buffer() : mOlder(nullptr), mNewer(nullptr), mOrigin(0) {}
// Constructor of later buffers.
Buffer(Buffer* aOlder, Index aOrigin)
: mOlder(aOlder), mNewer(nullptr), mOrigin(aOrigin) {
aOlder->mNewer = this;
Buffer* Older() const { return mOlder; }
void SetOlder(Buffer* aOlder) { mOlder = aOlder; }
Buffer* Newer() const { return mNewer; }
void SetNewer(Buffer* aNewer) { mNewer = aNewer; }
Index Origin() const { return mOrigin; }
void SetOrigin(Index aOrigin) { mOrigin = aOrigin; }
// Run aF on a yet-invalid element.
// Not thread-safe by itself, but nothing else should write this element,
// and reader won't access it until after it becomes valid.
template <typename F>
void SetAndValidateElement(F&& aF, Index aIndex) {
MOZ_ASSERT(aIndex >= Origin());
MOZ_ASSERT(aIndex < Origin() + BufferSize);
mElements[aIndex - Origin()].SetAndValidate(aF, aIndex);
using DidReadLastElement = bool;
// Read all valid elements starting at aIndex, marking them invalid and
// updating aIndex.
// Returns true if we ended up reading the last element in this buffer.
// Accessing the validity bit is thread-safe (as it's atomic), but once
// an element is valid, the reading itself is not thread-safe and should be
// guarded.
template <typename F>
DidReadLastElement ReadAndInvalidateAll(F&& aF, Index& aIndex) {
MOZ_ASSERT(aIndex >= Origin());
MOZ_ASSERT(aIndex < Origin() + BufferSize);
for (; aIndex < Origin() + BufferSize; ++aIndex) {
if (!mElements[aIndex - Origin()].ReadAndInvalidate(aF)) {
// Found an invalid element, stop here. (aIndex will not be updated
// past it, so we will start from here next time.)
return false;
return true;
Buffer* mOlder;
Buffer* mNewer;
Index mOrigin;
BufferedElement mElements[BufferSize];
// Reuse a buffer, or create a new one.
// All buffered elements will be invalid.
Buffer* NewBuffer(Buffer* aOlder, Index aOrigin) {
for (;;) {
Buffer* head = mReusableBuffers;
if (!head) {
Buffer* buffer = new Buffer(aOlder, aOrigin);
return buffer;
Buffer* older = head->Older();
// Try to pivot the reusable-buffer pointer from the current head to the
// next buffer in line.
if (mReusableBuffers.compareExchange(head, older)) {
// Success! The reusable-buffer pointer now points at the older buffer,
// so we can recycle this ex-head.
// We will be the newest; newer-pointer should already be null.
return head;
// Failure, someone else must have touched the list, loop to try again.
// Discard a fully-read buffer.
// If aDestroy is true, delete it.
// If aDestroy is false, move the buffer to a reusable-buffer stack.
void StopUsing(Buffer* aBuffer, bool aDestroy) {
// We should only stop using the oldest buffer.
// The newest buffer should not be modified here.
MOZ_ASSERT(aBuffer->Newer()->Older() == aBuffer);
// Detach from the second-oldest buffer.
// Make the second-oldest buffer the now-oldest buffer.
mOldestBuffer = aBuffer->Newer();
if (aDestroy) {
delete aBuffer;
} else {
// The recycling stack only uses mOlder; mNewer is not needed.
// Make the given buffer the new head of reusable buffers.
for (;;) {
Buffer* head = mReusableBuffers;
if (mReusableBuffers.compareExchange(head, aBuffer)) {
// Index of the next element to write. Modified when an element index is
// claimed for a push. If the last element of a buffer is claimed, that push
// will be responsible for adding a new head buffer.
// Relaxed, because there is no synchronization based on this variable, each
// thread just needs to get a different value, and will then write different
// things (which themselves have some atomic validation before they may be
// read elsewhere, independent of this `mNextElementToWrite`.)
Atomic<Index::ValueType, Relaxed> mNextElementToWrite{0};
// Index that a live recent buffer reaches. If a push claims a lesser-or-
// equal number, the corresponding buffer is guaranteed to still be alive:
// - It will have been created before this index was updated,
// - It will not be destroyed until all its values have been written,
// including the one that just claimed a position within it.
// Also, the push that claims this exact number is responsible for adding the
// next buffer and updating this value accordingly.
// ReleaseAcquire, because when set to a certain value, the just-created
// buffer covering the new range must be visible to readers.
Atomic<Index::ValueType, ReleaseAcquire> mBuffersCoverAtLeastUpTo;
// Pointer to the most recent buffer. Never null.
// This is the most recent of a deque of yet-unread buffers.
// Only modified when adding a new head buffer.
// ReleaseAcquire, because when modified, the just-created new buffer must be
// visible to readers.
Atomic<Buffer*, ReleaseAcquire> mMostRecentBuffer;
// Stack of reusable buffers.
// ReleaseAcquire, because when modified, the just-added buffer must be
// visible to readers.
Atomic<Buffer*, ReleaseAcquire> mReusableBuffers;
// Template-provided locking mechanism to protect PopAll()-only member
// variables below.
ReaderLocking mReaderLocking;
// Pointer to the oldest buffer, which contains the new element to be popped.
// Never null.
Buffer* mOldestBuffer;
// Index of the next element to be popped.
Index mNextElementToPop{0};
// Stats.
class AtomicCountAndWatermark {
explicit AtomicCountAndWatermark(int aCount)
: mCount(aCount), mWatermark(aCount) {}
int Count() const { return int(mCount); }
CountAndWatermark Get() const {
return CountAndWatermark{int(mCount), int(mWatermark)};
int operator++() {
int count = int(++mCount);
// Update watermark.
for (;;) {
int watermark = int(mWatermark);
if (watermark >= count) {
// printf("++[%p] -=> %d-%d\n", this, count, watermark);
if (mWatermark.compareExchange(watermark, count)) {
// printf("++[%p] -x> %d-(was %d now %d)\n", this, count, watermark,
// count);
return count;
int operator--() {
int count = int(--mCount);
// printf("--[%p] -> %d\n", this, count);
return count;
// Relaxed, as these are just gathering stats, so consistency is not
// critical.
Atomic<int, Relaxed> mCount;
Atomic<int, Relaxed> mWatermark;
// All buffers in the mMostRecentBuffer deque.
AtomicCountAndWatermark mLiveBuffersStats;
// All buffers in the mReusableBuffers stack.
AtomicCountAndWatermark mReusableBuffersStats;
// All allocated buffers (sum of above).
AtomicCountAndWatermark mAllocatedBuffersStats;
} // namespace mozilla
#endif // mozilla_MultiWriterQueue_h_