#include <winsock2.h>
#include <mswsock.h>
#include <stdlib.h>
#include <stdio.h>
#include <time.h>

#pragma comment(lib,"ws2_32")   // Standard socket API.
#pragma comment(lib,"mswsock")  // AcceptEx, TransmitFile, etc,.
#pragma comment(lib,"shlwapi")  // UrlUnescape.

#define WORKER_MAX_THREAD_COUNT 256
#define DATA_BUFSIZE    4096

enum
{
    COMPLETION_KEY_NONE = 0,
    COMPLETION_KEY_SHUTDOWN = 1,
    COMPLETION_KEY_USER_TASK = 2,
    COMPLETION_KEY_USER_DATA_LOAD = 3
};

static void dump_error(int err) {
    LPSTR messageBuffer = 0;
    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
        NULL, WSAGetLastError(), MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&messageBuffer, 0, NULL);
    fprintf(stderr, "--- [err: %d] %s\n", err, messageBuffer);
    LocalFree(messageBuffer);
}

static unsigned int __stdcall WorkerProc(void* IoPort) {
    fprintf(stderr, "- Event loop started [id: %d]\n", GetCurrentThreadId());

    for (;;)
    {
        BOOL Status = 0;
        DWORD NumTransferred = 0;
        ULONG_PTR CompKey = COMPLETION_KEY_NONE;
        LPOVERLAPPED pOver = 0;

        Status = GetQueuedCompletionStatus((HANDLE)IoPort, &NumTransferred, &CompKey, &pOver, INFINITE);
        if (Status == FALSE) {
            int err = WSAGetLastError();
            if (err != WSA_IO_PENDING) {
                dump_error(err);
            }

            continue;
        }

        //-
        if (COMPLETION_KEY_SHUTDOWN == CompKey)
            break;

        if (COMPLETION_KEY_USER_TASK == CompKey) {
            //fprintf(stderr, "- Processing [id: %d]\n", GetCurrentThreadId());
            time_t endTime = time(NULL) + 5;
            while (time(NULL) < endTime) {
                // do work here.
            }
            //fprintf(stderr, "- Done processing [id: %d]\n", GetCurrentThreadId());
        }

        if (COMPLETION_KEY_USER_DATA_LOAD == CompKey) {
            //fprintf(stderr, "- Loading [id: %d]\n", GetCurrentThreadId());
            time_t endTime = time(NULL) + 5;
            while (time(NULL) < endTime) {
                // do work here.
            }
        }
    }

    fprintf(stderr, "- Event loop done [id: %d]\n", GetCurrentThreadId());
    return 0;
}

static void hang_in_there(void) {
    fprintf(stderr, "*** Press enter to exit ...");
    fgetc(stdin);
}

int main(int argc, char **argv) {
    atexit(hang_in_there);

    WSADATA wsaData = { 0 };
    WSAStartup(MAKEWORD(2, 2), &wsaData);

    SYSTEM_INFO sysinfo;
    GetSystemInfo(&sysinfo);
    int numCPU = sysinfo.dwNumberOfProcessors;

    HANDLE IoPort = CreateIoCompletionPort(INVALID_HANDLE_VALUE, 0, 0, 0);
    HANDLE IoLoadPort = CreateIoCompletionPort(INVALID_HANDLE_VALUE, 0, 0, 0);

    HANDLE Workers[WORKER_MAX_THREAD_COUNT] = { 0 };
    unsigned int WorkerIds[WORKER_MAX_THREAD_COUNT] = { 0 };

    Workers[0] = (HANDLE)_beginthreadex(0, 0, WorkerProc, IoLoadPort, 0, WorkerIds);
    SetThreadAffinityMask(Workers[0], 0x01);

    for (size_t i = 1; i < numCPU; i++) {
        Workers[i] = (HANDLE)_beginthreadex(0, 0, WorkerProc, IoPort, 0, WorkerIds + i);
        SetThreadAffinityMask(Workers[i], 0xFE);
    }

    for(;;)
    {
        fgetc(stdin);
        PostQueuedCompletionStatus(IoPort, 0, COMPLETION_KEY_USER_TASK, 0);
        //PostQueuedCompletionStatus(IoLoadPort, 0, COMPLETION_KEY_USER_DATA_LOAD, 0);
    }

    for (size_t i = 0; i<numCPU; i++)
        PostQueuedCompletionStatus(IoPort, 0, COMPLETION_KEY_SHUTDOWN, 0);
    WaitForMultipleObjects(numCPU, Workers, TRUE, INFINITE);
    for (size_t i = 0; i<numCPU; i++)
        CloseHandle(Workers[i]);
    CloseHandle(IoPort);

    WSACleanup();
    return 0;
}