#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#include <stdio.h>
#include <mmsystem.h>
#pragma comment(lib, "winmm.lib")

DWORD g_tls = 0;

DWORD WINAPI Thread(PVOID p)
{
    ULONGLONG count = 0;
    g_tls = TlsAlloc();
    TlsSetValue(g_tls, &count);
    while(InterlockedCompareExchange((LONG*)p, 0, 0))
    {
        Sleep(500);
    }
    DWORD time = timeGetTime(), endTime;
    SleepEx(0, TRUE);
    endTime = timeGetTime();
    printf("Thread dispatched %I64u apcs, took %lu ms\n", count, endTime - time);
    TlsFree(g_tls);
    return 0;
}

VOID NTAPI Routine(PVOID, PVOID, PVOID)
{
    ULONGLONG* pLong = (ULONGLONG*)TlsGetValue(g_tls);
    ++(*pLong);
}

int main()
{
    LONG loopFlag = 1;
    HANDLE hThread = CreateThread(NULL, 0, &Thread, &loopFlag, 0, NULL);
    Sleep(250);
    NTSTATUS stat;
    ULONGLONG count = 0;
    ULONG a = 0, b = 0, c = 0;
    // just in case some caching of similar argumented apc's
    while(NT_SUCCESS(stat = NtQueueApcThread(hThread, &Routine, UlongToPtr(a), UlongToPtr(b), UlongToPtr(c))))
    {
        ++count;
        ++a;
        b += (a / ULONG_MAX);
        c += (b / ULONG_MAX);
    }
    printf("Queued up %I64u APCs, fail status was %#8x\n", count, stat);
    InterlockedExchange(&loopFlag, 0);
    WaitForSingleObject(hThread, INFINITE);
    CloseHandle(hThread);

    return 0;
}