The more I read the more confused I become... I would have thought it trivial to find a formally correct MPSC queue implemented in C++.
Every time I find another stab at it, further research seems to suggest there are issues such as ABA or other subtle race conditions.
Many talk of the need for garbage collection. This is something I want to avoid.
Is there an accepted correct open-source implementation out there?
You may want to check disruptor; it's available in C++ here: http://lmax-exchange.github.io/disruptor/
You can also find explanation how it works here on stackoverflow Basically it's circular buffer with no locking, optimized for passing FIFO messages between threads in a fixed-size slots.
Here are two implementations which I found useful: Lock-free Multi-producer Multi-consumer Queue on Ring Buffer @ NatSys Lab. Blog and
Yet another implementation of a lock-free circular array queue
@ CodeProject
NOTE: the code below is incorrect, I leave it only as an example how tricky these things can be.
If you don't like the complexity of google version, here is something similar from me - it's much simpler, but I leave it as an exercise to the reader to make it work (it's part of larger project, not portable at the moment). The whole idea is to maintain cirtular buffer for data and a small set of counters to identify slots for writing/written and reading/read. Since each counter is in its own cache line, and (normally) each is only atomically updated once in the live of a message, they can all be read without any synchronisation. There is one potential contention point between writing threads in post_done, it's required for FIFO guarantee. Counters (head_, wrtn_, rdng_, tail_) were selected to ensure correctness and FIFO, so dropping FIFO would also require change of counters (and that might be difficult to do without sacrifying correctness). It is possible to slightly improve performance for scenarios with one consumer, but I would not bother - you would have to undo it if other use cases with multiple readers are found.
On my machine latency looks like following (percentile on left, mean within this percentile on right, unit is microsecond, measured by rdtsc):
total=1000000 samples, avg=0.24us
50%=0.214us, avg=0.093us
90%=0.23us, avg=0.151us
99%=0.322us, avg=0.159us
99.9%=15.566us, avg=0.173us
These results are for single polling consumer, i.e. worker thread calling wheel.read() in tight loop and checking if not empty (scroll to bottom for example). Waiting consumers (much lower CPU utilization) would wait on event (one of acquire... functions), this adds about 1-2us to average latency due to context switch.
Since there is verly little contention on read, consumers scale very well with number of worker threads, e.g. for 3 threads on my machine:
total=1500000 samples, avg=0.07us
50%=0us, avg=0us
90%=0.155us, avg=0.016us
99%=0.361us, avg=0.038us
99.9%=8.723us, avg=0.044us
Patches will be welcome :)
// Copyright (c) 2011-2012, Bronislaw (Bronek) Kozicki
//
// Distributed under the Boost Software License, Version 1.0. (See accompanying
// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
#pragma once
#include <core/api.hxx>
#include <core/wheel/exception.hxx>
#include <boost/noncopyable.hpp>
#include <boost/type_traits.hpp>
#include <boost/lexical_cast.hpp>
#include <typeinfo>
namespace core { namespace wheel
{
struct bad_size : core::exception
{
template<typename T> explicit bad_size(const T&, size_t m)
: core::exception(std::string("Slot capacity exceeded, sizeof(")
+ typeid(T).name()
+ ") = "
+ boost::lexical_cast<std::string>(sizeof(T))
+ ", capacity = "
+ boost::lexical_cast<std::string>(m)
)
{}
};
// inspired by Disruptor
template <typename Header>
class wheel : boost::noncopyable
{
__declspec(align(64))
struct slot_detail
{
// slot write: (memory barrier in wheel) > post_done > (memory barrier in wheel)
// slot read: (memory barrier in wheel) > read_done > (memory barrier in wheel)
// done writing or reading, must update wrtn_ or tail_ in wheel, as appropriate
template <bool Writing>
void done(wheel* w)
{
if (Writing)
w->post_done(sequence);
else
w->read_done();
}
// cache line for sequence number and header
long long sequence;
Header header;
// there is no such thing as data type with variable size, but we need it to avoid thrashing
// cache - so we invent one. The memory is reserved in runtime and we simply go beyond last element.
// This is well into UB territory! Using template parameter for this is not good, since it
// results in this small implementation detail leaking to all possible user interfaces.
__declspec(align(8))
char data[8];
};
// use this as a storage space for slot_detail, to guarantee 64 byte alignment
_declspec(align(64))
struct slot_block { long long padding[8]; };
public:
// wrap slot data to outside world
template <bool Writable>
class slot
{
template<typename> friend class wheel;
slot& operator=(const slot&); // moveable but non-assignable
// may only be constructed by wheel
slot(slot_detail* impl, wheel<Header>* w, size_t c)
: slot_(impl) , wheel_(w) , capacity_(c)
{}
public:
slot(slot&& s)
: slot_(s.slot_) , wheel_(s.wheel_) , capacity_(s.capacity_)
{
s.slot_ = NULL;
}
~slot()
{
if (slot_)
{
slot_->done<Writable>(wheel_);
}
}
// slot accessors - use Header to store information on what type is actually stored in data
bool empty() const { return !slot_; }
long long sequence() const { return slot_->sequence; }
Header& header() { return slot_->header; }
char* data() { return slot_->data; }
template <typename T> T& cast()
{
static_assert(boost::is_pod<T>::value, "Data type must be POD");
if (sizeof(T) > capacity_)
throw bad_size(T(), capacity_);
if (empty())
throw no_data();
return *((T*) data());
}
private:
slot_detail* slot_;
wheel<Header>* wheel_;
const size_t capacity_;
};
private:
// dynamic size of slot, with extra capacity, expressed in 64 byte blocks
static size_t sizeof_slot(size_t s)
{
size_t m = sizeof(slot_detail);
// add capacity less 8 bytes already within sizeof(slot_detail)
m += max(8, s) - 8;
// round up to 64 bytes, i.e. alignment of slot_detail
size_t r = m & ~(unsigned int)63;
if (r < m)
r += 64;
r /= 64;
return r;
}
// calculate actual slot capacity back from number of 64 byte blocks
static size_t slot_capacity(size_t s)
{
return s*64 - sizeof(slot_detail) + 8;
}
// round up to power of 2
static size_t round_size(size_t s)
{
// enfore minimum size
if (s <= min_size)
return min_size;
// find rounded value
--s;
size_t r = 1;
while (s)
{
s >>= 1;
r <<= 1;
};
return r;
}
slot_detail& at(long long sequence)
{
// find index from sequence number and return slot at found index of the wheel
return *((slot_detail*) &wheel_[(sequence & (size_ - 1)) * blocks_]);
}
public:
wheel(size_t capacity, size_t size)
: head_(0) , wrtn_(0) , rdng_(0) , tail_(0) , event_()
, blocks_(sizeof_slot(capacity)) , capacity_(slot_capacity(blocks_)) , size_(round_size(size))
{
static_assert(boost::is_pod<Header>::value, "Header type must be POD");
static_assert(sizeof(slot_block) == 64, "This was unexpected");
wheel_ = new slot_block[size_ * blocks_];
// all slots must be initialised to 0
memset(wheel_, 0, size_ * 64 * blocks_);
active_ = 1;
}
~wheel()
{
stop();
delete[] wheel_;
}
// all accessors needed
size_t capacity() const { return capacity_; } // capacity of a single slot
size_t size() const { return size_; } // number of slots available
size_t queue() const { return (size_t)head_ - (size_t)tail_; }
bool active() const { return active_ == 1; }
// enough to call it just once, to fine tune slot capacity
template <typename T>
void check() const
{
static_assert(boost::is_pod<T>::value, "Data type must be POD");
if (sizeof(T) > capacity_)
throw bad_size(T(), capacity_);
}
// stop the wheel - safe to execute many times
size_t stop()
{
InterlockedExchange(&active_, 0);
// must wait for current read to complete
while (rdng_ != tail_)
Sleep(10);
return size_t(head_ - tail_);
}
// return first available slot for write
slot<true> post()
{
if (!active_)
throw stopped();
// the only memory barrier on head seq. number we need, if not overflowing
long long h = InterlockedIncrement64(&head_);
while(h - (long long) size_ > tail_)
{
if (InterlockedDecrement64(&head_) == h - 1)
throw overflowing();
// protection against case of race condition when we are overflowing
// and two or more threads try to post and two or more messages are read,
// all at the same time. If this happens we must re-try, otherwise we
// could have skipped a sequence number - causing infinite wait in post_done
Sleep(0);
h = InterlockedIncrement64(&head_);
}
slot_detail& r = at(h);
r.sequence = h;
// wrap in writeable slot
return slot<true>(&r, this, capacity_);
}
// return first available slot for write, nothrow variant
slot<true> post(std::nothrow_t)
{
if (!active_)
return slot<true>(NULL, this, capacity_);
// the only memory barrier on head seq. number we need, if not overflowing
long long h = InterlockedIncrement64(&head_);
while(h - (long long) size_ > tail_)
{
if (InterlockedDecrement64(&head_) == h - 1)
return slot<true>(NULL, this, capacity_);
// must retry if race condition described above
Sleep(0);
h = InterlockedIncrement64(&head_);
}
slot_detail& r = at(h);
r.sequence = h;
// wrap in writeable slot
return slot<true>(&r, this, capacity_);
}
// read first available slot for read
slot<false> read()
{
slot_detail* r = NULL;
// compare rdng_ and wrtn_ early to avoid unnecessary memory barrier
if (active_ && rdng_ < wrtn_)
{
// the only memory barrier on reading seq. number we need
const long long h = InterlockedIncrement64(&rdng_);
// check if this slot has been written, step back if not
if (h > wrtn_)
InterlockedDecrement64(&rdng_);
else
r = &at(h);
}
// wrap in readable slot
return slot<false>(r , this, capacity_);
}
// waiting for new post, to be used by non-polling clients
void acquire()
{
event_.acquire();
}
bool try_acquire()
{
return event_.try_acquire();
}
bool try_acquire(unsigned long timeout)
{
return event_.try_acquire(timeout);
}
void release()
{}
private:
void post_done(long long sequence)
{
const long long t = sequence - 1;
// the only memory barrier on written seq. number we need
while(InterlockedCompareExchange64(&wrtn_, sequence, t) != t)
Sleep(0);
// this is outside of critical path for polling clients
event_.set();
}
void read_done()
{
// the only memory barrier on tail seq. number we need
InterlockedIncrement64(&tail_);
}
// each in its own cache line
// head_ - wrtn_ = no. of messages being written at this moment
// rdng_ - tail_ = no. of messages being read at the moment
// head_ - tail_ = no. of messages to read (including those being written and read)
// wrtn_ - rdng_ = no. of messages to read (excluding those being written or read)
__declspec(align(64)) volatile long long head_; // currently writing or written
__declspec(align(64)) volatile long long wrtn_; // written
__declspec(align(64)) volatile long long rdng_; // currently reading or read
__declspec(align(64)) volatile long long tail_; // read
__declspec(align(64)) volatile long active_; // flag switched to 0 when stopped
__declspec(align(64))
api::event event_; // set when new message is posted
const size_t blocks_; // number of 64-byte blocks in a single slot_detail
const size_t capacity_; // capacity of data() section per single slot. Initialisation depends on blocks_
const size_t size_; // number of slots available, always power of 2
slot_block* wheel_;
};
}}
Here is what polling consumer worker thread may look like:
while (wheel.active())
{
core::wheel::wheel<int>::slot<false> slot = wheel.read();
if (!slot.empty())
{
Data& d = slot.cast<Data>();
// do work
}
// uncomment below for waiting consumer, saving CPU cycles
// else
// wheel.try_acquire(10);
}
Edited added consumer example
The most suitable implementation depends on the desired properties of a queue. Should it be unbounded or a bounded one is fine? Should it be linearizable, or less strict requirements would be fine? How strong FIFO guarantees you need? Are you willing to pay the cost of reverting the list by the consumer (there exists a very simple implementation where the consumer grabs the tail of a single-linked list, thus getting at once all items put by producers till the moment)? Should it guarantee that no thread is ever blocked, or tiny chances to get some thread blocked are ok? And etc.
Some useful links:
Is multiple-producer, single-consumer possible in a lockfree setting?
http://www.1024cores.net/home/lock-free-algorithms/queues
http://www.1024cores.net/home/lock-free-algorithms/queues/intrusive-mpsc-node-based-queue
https://groups.google.com/group/comp.programming.threads/browse_frm/thread/33f79c75146582f3
Hope that helps.
Below is the technique I used for my Cooperative Multi-tasking / Multi-threading library (MACE) http://bytemaster.github.com/mace/. It has the benefit of being lock-free except for when the queue is empty.
struct task {
boost::function<void()> func;
task* next;
};
boost::mutex task_ready_mutex;
boost::condition_variable task_ready;
boost::atomic<task*> task_in_queue;
// this can be called from any thread
void thread::post_task( task* t ) {
// atomically post the task to the queue.
task* stale_head = task_in_queue.load(boost::memory_order_relaxed);
do { t->next = stale_head;
} while( !task_in_queue.compare_exchange_weak( stale_head, t, boost::memory_order_release ) );
// Because only one thread can post the 'first task', only that thread will attempt
// to aquire the lock and therefore there should be no contention on this lock except
// when *this thread is about to block on a wait condition.
if( !stale_head ) {
boost::unique_lock<boost::mutex> lock(task_ready_mutex);
task_ready.notify_one();
}
}
// this is the consumer thread.
void process_tasks() {
while( !done ) {
// this will atomically pop everything that has been posted so far.
pending = task_in_queue.exchange(0,boost::memory_order_consume);
// pending is a linked list in 'reverse post order', so process them
// from tail to head if you want to maintain order.
if( !pending ) { // lock scope
boost::unique_lock<boost::mutex> lock(task_ready_mutex);
// check one last time while holding the lock before blocking.
if( !task_in_queue ) task_ready.wait( lock );
}
}
I'm guessing no such thing exists - and if it does, it either isn't portable or isn't open source.
Conceptually, you are trying to control two pointers simultaneously: the tail pointer and the tail->next pointer. That can't generally be done with just lock-free primitives.
© 2022 - 2024 — McMap. All rights reserved.