同步线程安全的API，用于公开缓存的数据

Question

We offer a package which interfaces with a bunch of other packages who's APIs are not thread safe. Our package's API is entirely message based and therefore asynchronous to allow thread safety for users of our package. Our package therefore wraps a bunch of non-thread safe packages and offers a thread-safe API. That means that users of our package can interface with our package from any thread.

We would like to add synchronous APIs to our package while maintaining thread safety. I've done some research and have come up with two possible patterns to do this which I will share below. I'm not entirely happy with either approach so am wondering if the community may have more suggestions for patterns we can use. Note that the code below is for design and illustration purposes (c++ pseudocode) so is not meant to compile.

Approach 1 - Package users dependency inject data access classes to our package. We access these classes using run time type inference.

// Define an interface class for all data accessor classes
class DataAccessor
{

}

// Some random data
class FooData
{
    int foo;
    int bar;
}

// A concrete data accessor
class FooDataAccessor : public DataAccessor
{
public:
    FooData getFooData()
    {
        FooData fooDataCopy;
        {
            //Locks cachedFooData in this scope
            ScopedCriticalSection _(dataCriticalSection);
            fooDataCopy.foo = cachedFooData.foo;
            fooDataCopy.bar = cachedFooData.bar;
        }
        return fooDataCopy;
    }
    void setFooData(FooData& fooData)
    {
        //Locks cachedFooData in this scope
        ScopedCriticalSection _(dataCriticalSection);
        cachedFooData.foo = dooData.foo;
        cachedFooData.bar = dooData.bar;
    }

private:
    FooData cachedFooData;
    CriticalSection dataCriticalSection; //Use this for locking cachedFooData to set the data.
}

class OurPackage
{
    OurPackage(std::vector<DataAccessor*>); //constructor which is injected the data accessors so that our package customers can own their lifecycle.
}

// How package users would inject the data accessors into our package, then later access the data
int main()
{
    std::vector<DataAccessor*> dataAccessors;
    //The package customer now populates the data Accessors with the instances they need.
    dataAccessors.push_back(new FooDataAccessor());        

    auto package = new OurPackage(dataAccessors);

    // How package users access the data, assume FooDataAccessor was at the front
    auto fooAccessor = dataAccessors.front();
    if (fooAccessor)
    {
        FooData data = fooAccessor->getFooData();
    }
}

// In OurPackage, set the actual data in caches
for (DataAccessor* dataAccessor : dataAccessors)
{
    //Use RTTI to find the instance we want
    if (auto dataAccessorTypeWeWant = dynamic_cast<DataAccessorTypeWeWant*>(dataAccessor) != nullptr)
    {
        //Set the data on dataAccessorTypeWeWant 
        //For example, set FooData
        FooData fooData;
        fooData.foo = 1;
        fooData.bar = 2;
        dataAccessorTypeWeWant.setFooData(fooData);
        break;
    }
}

2 - Use a singleton pattern instead

If the data access caches are singletons instead, package users don't have to manage the lifecycle of these classes and don't need to worry about passing pointers to instances of the data access caches around their application. This has all the pitfalls of singletons though.

Answer 1

Whatever patter that you chose you should use an Atomic Type that is founded into the library <atomic> , functionality available since C++11. Whit this type you can create threadsafe variables, for example:

// Some random data
class FooData
{
    std::atomic<int>  foo;
    std::atomic<int>  bar;
}

I share you a description of this library from CPlusPlus :

Atomic types are types that encapsulate a value whose access is guaranteed to not cause data races and can be used to synchronize memory accesses among different threads.

Answer 2

This is a sample. If copy by value is what you need.

#include <vector>
#include <iostream>
#include <thread>
#include <atomic>
#include <mutex>

// Undefine to see data corruptions
#define USING_LOCK

std::atomic<int> atomic_i;

class FooData {
public:
    FooData() :foo(atomic_i.fetch_add(1, std::memory_order_relaxed)), bar(foo) {}
    ~FooData() { if (foo != bar) { std::cout << "Data corrupted!\n"; } }
private:
    int foo;
    int bar;
};

class FooDataAccessor {
public:
    FooData getFooData() {
#ifdef USING_LOCK
        std::lock_guard<std::mutex> l(_lock);
#endif // USING_LOCK
        return cachedFooData;
    }
    void setFooData(const FooData& fooData) {
#ifdef USING_LOCK
        std::lock_guard<std::mutex> l(_lock);
#endif // USING_LOCK
        cachedFooData = fooData;
    }

private:
    FooData cachedFooData;
#ifdef USING_LOCK
    std::mutex _lock;
#endif // USING_LOCK
};

void f(FooDataAccessor* accessor) {
    for (size_t i = 0; i < 1000; i++) {
        accessor->getFooData();
        accessor->setFooData(FooData());
    }
}

int main() {
    FooDataAccessor accessor;
    std::vector<std::thread> v;
    for (size_t i = 0; i < 5; i++) {
        v.emplace_back(f, &accessor);
    }
    for (auto& t : v) {
        t.join();
    }
}