how to use std::atomic<>
In the question above, obviously we can just use std::mutex to keep thread safety. I want to know when to use which one.
classs A
{
std::atomic<int> x;
public:
A()
{
x=0;
}
void Add()
{
x++;
}
void Sub()
{
x--;
}
};
and
std::mutex mtx;
classs A
{
int x;
public:
A()
{
x=0;
}
void Add()
{
std::lock_guard<std::mutex> guard(mtx);
x++;
}
void Sub()
{
std::lock_guard<std::mutex> guard(mtx);
x--;
}
};
As a rule of thumb, use std::atomic for POD types where the underlying specialisation will be able to use something clever like a bus lock on the CPU (which will give you no more overhead than a pipeline dump), or even a spin lock. On some systems, an int might already be atomic, so std::atomic<int> will specialise out effectively to an int.
Use std::mutex for non-POD types, bearing in mind that acquiring a mutex is at least an order of magnitude slower than a bus lock.
If you're still unsure, measure the performance.
Related
I'm reading C++ concurrency in action.
It introduces how to implement interrupting thread using std::condition_variable_any.
I try to understand the code more than a week, but I couldn't.
Below is the code and explanation in the book.
#include <condition_variable>
#include <future>
#include <iostream>
#include <thread>
class thread_interrupted : public std::exception {};
class interrupt_flag {
std::atomic<bool> flag;
std::condition_variable* thread_cond;
std::condition_variable_any* thread_cond_any;
std::mutex set_clear_mutex;
public:
interrupt_flag() : thread_cond(0), thread_cond_any(0) {}
void set() {
flag.store(true, std::memory_order_relaxed);
std::lock_guard<std::mutex> lk(set_clear_mutex);
if (thread_cond) {
thread_cond->notify_all();
} else if (thread_cond_any) {
thread_cond_any->notify_all();
}
}
bool is_set() const { return flag.load(std::memory_order_relaxed); }
template <typename Lockable>
void wait(std::condition_variable_any& cv, Lockable& lk);
};
thread_local static interrupt_flag this_thread_interrupt_flag;
void interruption_point() {
if (this_thread_interrupt_flag.is_set()) {
throw thread_interrupted();
}
}
template <typename Lockable>
void interrupt_flag::wait(std::condition_variable_any& cv, Lockable& lk) {
struct custom_lock {
interrupt_flag* self;
// (1) What is this lk for? Why is lk should be already locked when it is used in costume_lock constructor?
Lockable& lk;
custom_lock(interrupt_flag* self_, std::condition_variable_any& cond,
Lockable& lk_)
: self(self_), lk(lk_) {
self->set_clear_mutex.lock();
self->thread_cond_any = &cond;
}
void unlock() {
lk.unlock();
self->set_clear_mutex.unlock();
}
void lock() { std::lock(self->set_clear_mutex, lk); }
~custom_lock() {
self->thread_cond_any = 0;
self->set_clear_mutex.unlock();
}
};
custom_lock cl(this, cv, lk);
interruption_point();
cv.wait(cl);
interruption_point();
}
class interruptible_thread {
std::thread internal_thread;
interrupt_flag* flag;
public:
template <typename FunctionType>
interruptible_thread(FunctionType f) {
std::promise<interrupt_flag*> p;
internal_thread = std::thread([f, &p] {
p.set_value(&this_thread_interrupt_flag);
f();
});
flag = p.get_future().get();
}
void interrupt() {
if (flag) {
flag->set();
}
};
void join() { internal_thread.join(); };
void detach();
bool joinable() const;
};
template <typename Lockable>
void interruptible_wait(std::condition_variable_any& cv, Lockable& lk) {
this_thread_interrupt_flag.wait(cv, lk);
}
void foo() {
// (2) This is my implementation of how to use interruptible wait. Is it correct?
std::condition_variable_any cv;
std::mutex m;
std::unique_lock<std::mutex> lk(m);
try {
interruptible_wait(cv, lk);
} catch (...) {
std::cout << "interrupted" << std::endl;
}
}
int main() {
std::cout << "Hello" << std::endl;
interruptible_thread th(foo);
th.interrupt();
th.join();
}
Your custom lock type acquires the lock on the internal
set_clear_mutex when it’s constructed 1, and then sets the
thread_cond_any pointer to refer to the std:: condition_variable_any
passed in to the constructor 2.
The Lockable reference is stored for later; this must already be
locked. You can now check for an interruption without worrying about
races. If the interrupt flag is set at this point, it was set before
you acquired the lock on set_clear_mutex. When the condition variable
calls your unlock() function inside wait(), you unlock the Lockable
object and the internal set_clear_mutex 3.
This allows threads that are trying to interrupt you to acquire the
lock on set_clear_mutex and check the thread_cond_any pointer once
you’re inside the wait() call but not before. This is exactly what you
were after (but couldn’t manage) with std::condition_variable.
Once wait() has finished waiting (either because it was notified or
because of a spurious wake), it will call your lock() function, which
again acquires the lock on the internal set_clear_mutex and the lock
on the Lockable object 4. You can now check again for interruptions
that happened during the wait() call before clearing the
thread_cond_any pointer in your custom_lock destructor 5, where you
also unlock the set_clear_mutex.
First, I couldn't understand what is the purpose of Lockabel& lk in mark (1) and why it is already locked in constructor of custom_lock. (It could be locked in the very custom_lock constructor. )
Second there is no example in this book of how to use interruptible wait, so foo() {} in mark (2) is my guess implementation of how to use it. Is it correct way of using it ?
You need a mutex-like object (lk in your foo function) to call the interruptiple waiting just as you would need it for the plain std::condition_variable::wait function.
What's problematic (I also read the book and I have doubts about this example) is that the flag member points to a memory location inside the other thread which could finish right before calling flag->set(). In this specific example the thread only exists after we set the flag so that is okay, but otherwise this approach is limited in my opinion (correct me if I am wrong).
Along with the main thread, i have one more thread that receives data to write them in a file.
std::queue<std::vector<int>> dataQueue;
std::mutex mutex;
void setData(const std::vector<int>& data) {
std::lock_guard<std::mutex> lock(mutex);
dataQueue.push(data);
}
void write(const std::string& fileName) {
std::unique_ptr<std::ostream> ofs = std::unique_ptr<std::ostream>(new zstr::ofstream(fileName));
while (store) {
std::lock_guard<std::mutex> lock(mutex);
while (!dataQueue.empty()) {
std::vector<int>& data= dataQueue.front();
ofs->write(reinterpret_cast<char*>(data.data()), sizeof(data[0])*data.size());
dataQueue.pop();
}
}
}
}
setData is used by the main thread and write is actually the writing thread. I use std::lock_quard to avoid memory conflict but when locking on the writing thread, it slows down the main thread as it has to wait for the Queue to be unlocked. But i guess i can avoid this as the threads never act on the same element of the queue at the same time.
So i would like to do it lock-free but i don't really understand how i should implement it. I mean, how can i do it without locking anything ? moreover, if the writing thread is faster than the main thread, the queue might be empty most of the time, so it should somehow waits for new data instead of looping infinitly to check for non empty queue.
EDIT: I changed simple std::lock_guard by std::cond_variable so that it could wait when the queue is empty. But the main thread can still be blocked as , when cvQeue.wait(.) is resolved, it reacquire the lock. moreover, what if the main thread does cvQueue.notify_one() but the writing thread is not waiting ?
std::queue<std::vector<int>> dataQueue;
std::mutex mutex;
std::condition_variable cvQueue;
void setData(const std::vector<int>& data) {
std::unique_lock<std::mutex> lock(mutex);
dataQueue.push(data);
cvQueue.notify_one();
}
void write(const std::string& fileName) {
std::unique_ptr<std::ostream> ofs = std::unique_ptr<std::ostream>(new zstr::ofstream(fileName));
while (store) {
std::lock_guard<std::mutex> lock(mutex);
while (!dataQueue.empty()) {
std::unique_lock<std::mutex> lock(mutex);
cvQueue.wait(lock);
ofs->write(reinterpret_cast<char*>(data.data()), sizeof(data[0])*data.size());
dataQueue.pop();
}
}
}
}
If you only have two threads, than you could use a lock-free single-producer-single-consumer (SPSC) queue.
A bounded version can be found here: https://github.com/rigtor/SPSCQueue
Dmitry Vyukov presented an unbounded version here: http://www.1024cores.net/home/lock-free-algorithms/queues/unbounded-spsc-queue (You should note though, that this code should be adapted to use atomics.)
Regarding a blocking pop operation - this is something that lock-free data structures do not provide since such an operation is obviously not lock-free. However, it should be relatively straight forward to adapt the linked implementations in such a way, that a push operation notifies a condition variable if the queue was empty before the push.
i guess i have something that met my needs. I did a LockFreeQueue that uses std::atomic. I can thus manage the state of the head/tail of the queue atomically.
template<typename T>
class LockFreeQueue {
public:
void push(const T& newElement) {
fifo.push(newElement);
tail.fetch_add(1);
cvQueue.notify_one();
}
void pop() {
size_t oldTail = tail.load();
size_t oldHead = head.load();
if (oldTail == oldHead) {
return;
}
fifo.pop();
head.store(++oldHead);
}
bool isEmpty() {
return head.load() == tail.load();
}
T& getFront() {
return fifo.front();
}
void waitForNewElements() {
if (tail.load() == head.load()) {
std::mutex m;
std::unique_lock<std::mutex> lock(m);
cvQueue.wait_for(lock, std::chrono::milliseconds(TIMEOUT_VALUE));
}
}
private:
std::queue<T> fifo;
std::atomic<size_t> head = { 0 };
std::atomic<size_t> tail = { 0 };
std::condition_variable cvQueue;
};
LockFreeQueue<std::vector<int>> dataQueue;
std::atomic<bool> store(true);
void setData(const std::vector<int>& data) {
dataQueue.push(data);
// do other things
}
void write(const std::string& fileName) {
std::unique_ptr<std::ostream> ofs = std::unique_ptr<std::ostream>(new zstr::ofstream(fileName));
while (store.load()) {
dataQueue.waitForNewElements();
while (!dataQueue.isEmpty()) {
std::vector<int>& data= dataQueue.getFront();
ofs->write(reinterpret_cast<char*>(data.data()), sizeof(data[0])*data.size());
dataQueue.pop();
}
}
}
}
I still have one lock in waitForNewElements but it is not locking the whole process as it is waiting for things to do. But the big improvement is that the producer can push while the consumer pop. It is only forbidden when LockFreQueue::tail and LockFreeQueue::head are the same. Meaning that the queue is empty and it enters the waiting state.
The thing that i'm not very satisfied at is cvQueue.wait_for(lock, TIMEOUT_VALUE). I wanted to do a simple cvQueue.wait(lock), but the problem is that when it comes to end the thread, I do store.store(false) in the main thread. So if the writing thread is waiting it will never end without a timeout. So, I set a big enough timeout so that most of the time the condition_variable is resolved by the lock, and when the thread ends it is resolved by the timeout.
If you feel that something must be wrong or must be improved, feel free to comment.
I have the following situation
std::mutex m;
void t() {
//lock the mutex m here
}
main() {
//create thread t here
//lock the mutex m here
}
I would like the thread t() to acquire the mutex before main() does, how can I obtain this behaviour using the threading functions provided by C++11?
Putting simply an std::lock_guard inside main() and t() would not work because it can take a bit before the thread is spawned, an so the mutex can be locked by main().
Regarding the conditional variable that Sneftel mentioned in the comment section, and a somewhat similar solution to the one provided by Angew:
One possible solution:
std::condition_variable cv;
std::mutex m;
bool threadIsReady = false; //bool should be fine in this case
void t() {
std::unique_lock<std::mutex> g(m);
threadIsReady = true;
cv.notify_one();
}
int main() {
std::thread th(t);
//if main locks the mutex first, it will have to wait until threadIsReady becomes true
//if main locks the mutex later, wait will do nothing since threadIsReady would have already been true
std::unique_lock<std::mutex> g(m);
cv.wait(g, [] {return threadIsReady; });
}
Here's a quick & dirty way to achieve this effect:
std::atomic<bool> threadIsReady{false};
void t()
{
std::lock_guard<std::mutex> g(m);
threadIsReady = true;
}
main()
{
std::thread th(t);
while (!threadIsReady) {}
std::lock_guard<std::mutex> g(m);
}
In this code is std::swap thread safe so it can be called from two execution threads at the same time or do I need use atomic_compare_exchange_weak() instead of swap()?
How do I know if this will work on all CPUs? I am happy if it just works on Intel CPUs.
#include <utility>
class resource {
int x = 0;
};
class foo
{
public:
foo() : p{new resource{}}
{ }
foo(const foo& other) : p{new resource{*(other.p)}}
{ }
foo(foo&& other) : p{other.p}
{
other.p = nullptr;
}
foo& operator=(foo other)
{
swap(*this, other);
return *this;
}
virtual ~foo()
{
delete p;
}
friend void swap(foo& first, foo& second)
{
using std::swap;
swap(first.p, second.p);
}
private:
resource* p;
};
I understand it is overkill to swap a pointer, but this migth be good pracise.
is std::swap thread safe so it can be called from two execution threads at the same time
std::swap is thread-safe as long as different threads pass different objects into it. Otherwise a race condition arises.
I would like to wrap the following code with clang thread annotations:
std::mutex mutex;
int counter = 0; // should be accessed while the mutex is locked
std::unique_lock<std::mutex> lock(mutex, std::try_to_lock);
if (lock) {
++counter;
}
I want both to use a RAII lock (std::unique_guard) and I also would like to add thread annotations to this code.
class __attribute__((capability("mutex"))) Mutex {
public:
Mutex() = default;
std::mutex &getInternalMutex() { return m_mutex; }
private:
std::mutex m_mutex{};
};
class __attribute__((scoped_lockable)) TryLock {
public:
TryLock(Mutex &mutex)
// PROBLEM HERE: __attribute__((acquire_capability(mutex)))
// is not suitable (lock may fail) but
// __attribute__((try_acquire_capability(true, mutex)))
// cannot be used neither (requires to be used in a method
// returning whether the lock succeed or not)
: m_try_lock(mutex.getInternalMutex(), std::try_to_lock) {}
~TryLock() __attribute__((release_capability)) = default;
bool isLocked() const {
return !!m_try_lock;
}
private:
std::unique_lock<std::mutex> m_try_lock;
};
clang only provides try_acquire_capability which is not suitable here (as it should be used for a function returning a boolean indicating if the lock succeeded or not).
What would be the correct way to annotate this lock?