c++ Schwarz counter with thread_local

Question

Can I use Schwarz counter (aka Nifty counter) idiom, with thread_local? (Assuming I replace all static with thread_local)

I need this (helper for java jni threads):

class ThisThread{
    JNIEnv* jni_env{nullptr};
public:
    JNIEnv* getEnv(){
        if (!jni_env){
            // Attach thread
            java_vm->GetEnv((void**)&jni_env, JNI_VERSION);
            java_vm->AttachCurrentThread(&jni_env, NULL);
        }

        return jni_env;
    }

    ~ThisThread(){
        if (!jni_env) return;
        // Deattach thread
        java_vm->DetachCurrentThread();
    }
};

static thread_local ThisThread this_thread;

To be constructed first, and destructed last in each thread. I may call this_thread->getEnv() from destructor/constructor of other static or thread_local object.

UPDATE

https://stackoverflow.com/a/30200992 - here, standard says that thread_local destructors called BEFORE static, and I need this one to be after.

Answer 1

I think the best solution is to implement the schwartz counter as normal, but implement the ThisThread class in terms of a thread_local static Impl.

Complete example with outputs:

// header file
#include <memory>
#include <mutex>
#include <iostream>
#include <thread>

std::mutex emit_mutex;

template<class...Ts>
void emit(Ts&&...ts)
{
    auto action = [](auto&&x) { std::cout << x; };
    auto lock = std::unique_lock<std::mutex>(emit_mutex);

    using expand = int[];
    expand{ 0,
        (action(std::forward<Ts>(ts)), 0)...
    };
}


struct ThisThread
{
    struct Impl
    {
        Impl()
        {
            emit("ThisThread created on thread ", std::this_thread::get_id(), '\n');
        }
        ~Impl()
        {
            emit("ThisThread destroyed on thread ", std::this_thread::get_id(), '\n');
        }
        void foo() 
        { 
            emit("foo on thread ", std::this_thread::get_id(), '\n');
        }
    };

    decltype(auto) foo() { return get_impl().foo(); }

private:
    static Impl& get_impl() { return impl_; }
    static thread_local Impl impl_;
};

struct ThisThreadInit
{

    ThisThreadInit();
    ~ThisThreadInit();

    static int initialised;
};

extern ThisThread& thisThread;
static ThisThreadInit thisThreadInit;



// cppfile

static std::aligned_storage_t<sizeof(ThisThread), alignof(ThisThread)> storage;
ThisThread& thisThread = *reinterpret_cast<ThisThread*>(std::addressof(storage));
int ThisThreadInit::initialised;
thread_local ThisThread::Impl ThisThread::impl_;

ThisThreadInit::ThisThreadInit()
{
    if (0 == initialised++)
    {
        new (std::addressof(storage)) ThisThread ();    
    }
}

ThisThreadInit::~ThisThreadInit()
{
    if (0 == --initialised)
    {
        thisThread.~ThisThread();
    }
}


// now use the object

#include <thread>

int main()
{
    thisThread.foo();

    auto t = std::thread([]{ thisThread.foo(); });
    t.join();
}

example output:

ThisThread created on thread 140475785611072
foo on thread 140475785611072
ThisThread created on thread 140475768067840
foo on thread 140475768067840
ThisThread destroyed on thread 140475768067840
ThisThread destroyed on thread 140475785611072

Answer 2

This does not answer how to make Schwarz counter for thread_local static 's (so I don't accept this as answer). But in the end, I came up with this platform-dependent(Linux/Android) solution.

#include <jni.h>
#include <cassert>
#include "JavaVM.h"

namespace jni_interface{

    class ThisThread{
        inline static thread_local pthread_key_t p_key;

        static void pthread_dstr(void *arg){
            if (!jni_env) return;
            java_vm->DetachCurrentThread();
            jni_env = nullptr;

            pthread_setspecific(p_key, NULL);
            pthread_key_delete(p_key);
        }

        static void register_dstr(void *arg){
            {
                const int res = pthread_key_create(&p_key, pthread_dstr);
                assert(res != EAGAIN);
                assert(res != ENOMEM);
                assert(res == 0);
            }
            {
                const int res = pthread_setspecific(p_key, arg);
                assert(res == 0);
            }
        }

        inline static thread_local JNIEnv* jni_env{nullptr};
    public:
        JNIEnv* getEnv(){
            if (!jni_env){
                assert(java_vm);
                java_vm->GetEnv((void**)&jni_env, JNI_VERSION);
                java_vm->AttachCurrentThread(&jni_env, NULL);       // safe to call in main thread

                register_dstr(jni_env);
            }

            return jni_env;
        }
    };

    static thread_local ThisThread this_thread;
}

Even if by some reason, pthread_dstr will be called before C++'s static thread_locals (or interleaved) [in other words ThisThread destroyed before last use], on next call to object (getEnv()) we kinda re-init/re-create it and register pthread_dstr for another round.

N.B. At total maximum we can have PTHREAD_DESTRUCTOR_ITERATIONS rounds, which is 4. But we always will end up on a second one, at worst case (if C++ thread_local implementation will use p_thread destructors [which will mean that OUR pthread_dstr may not be called last in the first round]).