// tun_if.c - Cross-platform TUN interface
#include "config_parser.h"
#include "tun_if.h"
#include "etcp.h"
#include "packet_dump.h"
#include "../lib/debug_config.h"
#include "../lib/u_async.h"
#include "../lib/ll_queue.h"
#include "../lib/memory_pool.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include "../lib/mem.h"

#ifdef _WIN32
#include <windows.h>
#endif

// ===================================================================
// Linux-only: TUN read callback (uasync) — определена ПЕРЕД tun_init
// ===================================================================
#ifndef _WIN32
// Internal read callback - called by uasync when data available on TUN
static void tun_read_callback(int fd, void* user_arg)
{
    (void)fd;
    struct tun_if* tun = (struct tun_if*)user_arg;
    uint8_t buffer[TUN_MAX_PACKET_SIZE];

    // Read from TUN device
    ssize_t nread = tun_platform_read(tun, buffer, sizeof(buffer));
    if (nread < 0) {
        DEBUG_ERROR(DEBUG_CATEGORY_TUN, "Failed to read from TUN device %s", tun->ifname);
        tun->read_errors++;
        return;
    }
    if (nread == 0) return;

    // Allocate packet data (+1 for prefix byte)
    uint8_t* packet_data = u_malloc(nread + 1);
    if (!packet_data) {
        DEBUG_ERROR(DEBUG_CATEGORY_TUN, "Failed to allocate packet data (size=%zd)", nread + 1);
        tun->read_errors++;
        return;
    }
    packet_data[0] = 0;  // Prefix byte
    memcpy(packet_data + 1, buffer, nread);

    struct ll_entry* pkt = queue_entry_new_from_pool(tun->pool);
    if (!pkt) {
        DEBUG_ERROR(DEBUG_CATEGORY_TUN, "Failed to allocate struct ll_entry from pool");
        u_free(packet_data);
        tun->read_errors++;
        return;
    }

    // Initialize entry
    pkt->dgram = packet_data;
    pkt->len = nread + 1;

    // Add to output queue (TUN → routing)
    if (queue_data_put(tun->output_queue, pkt, 0) != 0) {
        DEBUG_ERROR(DEBUG_CATEGORY_TUN, "Failed to add packet to output queue");
        u_free(packet_data);
        memory_pool_free(tun->pool, pkt);
        tun->read_errors++;
        return;
    }

    // Update statistics
    tun->bytes_read += nread + 1;
    tun->packets_read++;
    DEBUG_DEBUG(DEBUG_CATEGORY_TUN, "Read %zd bytes from TUN %s", nread + 1, tun->ifname);
    if (debug_should_output(DEBUG_LEVEL_DEBUG, DEBUG_CATEGORY_TUN)) {
        DEBUG_DEBUG(DEBUG_CATEGORY_TUN, "[TUN->] %s", dump_ip_packet_to_buffer(packet_data + 1, nread));
    }
}
#endif

// ===================================================================
// Callback: routing → TUN (вызывается из main thread)
// ===================================================================
static void tun_input_queue_callback(struct ll_queue* q, void* arg)
{
    (void)q;
    struct tun_if* tun = (struct tun_if*)arg;
    struct ll_entry* pkt = queue_data_get(tun->input_queue);
    if (!pkt) return;

    if (pkt->dgram && pkt->len > 1) {
        if (pkt->len > TUN_MAX_PACKET_SIZE + 1) {
            DEBUG_WARN(DEBUG_CATEGORY_TUN, "Packet too large: %zu bytes", pkt->len);
        } else {
            // Skip prefix byte, write from second byte
            ssize_t n = tun_platform_write(tun, pkt->dgram + 1, pkt->len - 1);
            if (n < 0) {
                DEBUG_ERROR(DEBUG_CATEGORY_TUN, "tun_platform_write failed");
                tun->write_errors++;
            } else {
                tun->bytes_written += n;
                tun->packets_written++;
                if (debug_should_output(DEBUG_LEVEL_DEBUG, DEBUG_CATEGORY_TUN)) {
                    DEBUG_DEBUG(DEBUG_CATEGORY_TUN, "[->TUN] %s", dump_ip_packet_to_buffer(pkt->dgram + 1, pkt->len - 1));
                }
            }
        }
        queue_dgram_free(pkt);
    }
    queue_entry_free(pkt);

    queue_resume_callback(tun->input_queue);
}

// ===================================================================
// Init
// ===================================================================
struct tun_if* tun_init(struct UASYNC* ua, struct utun_config* config)
{
    if (!ua || !config) return NULL;

    int test_mode = config->global.tun_test_mode;

    struct tun_if* tun = u_calloc(1, sizeof(struct tun_if));
    if (!tun) return NULL;

    tun->ua = ua;
    tun->test_mode = test_mode;
    tun->fd = -1;

    const char* name = config->global.tun_ifname;
    if (name && name[0]) strncpy(tun->ifname, name, sizeof(tun->ifname)-1);

    char ip_str[64];
    if (config->global.tun_ip.family == AF_INET) {
        snprintf(ip_str, sizeof(ip_str), "%s/32", ip_to_str(&config->global.tun_ip.addr.v4, AF_INET).str);
    } else {
        u_free(tun);
        return NULL;
    }

    int mtu = config->global.mtu > 0 ? config->global.mtu : TUN_MTU_DEFAULT;

    if (!test_mode) {
        if (tun_platform_init(tun, tun->ifname, ip_str, mtu) != 0) {
            u_free(tun);
            return NULL;
        }
    } else if (tun->ifname[0] == '\0') {
        strncpy(tun->ifname, "tun_test", sizeof(tun->ifname)-1);
    }

    tun->pool = memory_pool_init(sizeof(struct ll_entry));
    if (!tun->pool) goto fail;

    tun->output_queue = queue_new(ua, 0, "TUN output");
    tun->input_queue  = queue_new(ua, 0, "TUN input");
    if (!tun->output_queue || !tun->input_queue) goto fail;

    queue_set_callback(tun->input_queue, tun_input_queue_callback, tun);

    if (!test_mode) {
        int poll_fd = tun_platform_get_poll_fd(tun);

        if (poll_fd >= 0) {
#ifndef _WIN32
            // ==================== LINUX ====================
            tun->socket_id = uasync_add_socket(ua, poll_fd, tun_read_callback, NULL, NULL, tun);
            if (!tun->socket_id) goto fail;
#endif
        } else {
#ifdef _WIN32
            // ==================== WINDOWS ====================
            tun->running = 1;
            tun->stop_event = CreateEvent(NULL, TRUE, FALSE, NULL);
            if (!tun->stop_event) goto fail;

            tun->read_thread = CreateThread(NULL, 0, tun_read_thread_proc, tun, 0, NULL);
            if (!tun->read_thread) {
                CloseHandle(tun->stop_event);
                goto fail;
            }
#endif
        }
    }

    DEBUG_INFO(DEBUG_CATEGORY_TUN, "TUN %s initialized (%s mode)", tun->ifname,
               test_mode ? "TEST" : "REAL");
    return tun;

fail:
#ifdef _WIN32
    if (tun->stop_event) CloseHandle(tun->stop_event);
    if (tun->read_thread) {
        tun->running = 0;
        WaitForSingleObject(tun->read_thread, INFINITE);
        CloseHandle(tun->read_thread);
    }
#endif
    if (tun->output_queue) queue_free(tun->output_queue);
    if (tun->input_queue)  queue_free(tun->input_queue);
    if (tun->pool) memory_pool_destroy(tun->pool);
    if (!test_mode) tun_platform_cleanup(tun);
    u_free(tun);
    return NULL;
}

// ===================================================================
// Остальные функции (без изменений)
// ===================================================================
void tun_close(struct tun_if* tun)
{
    if (!tun) return;

#ifdef _WIN32
    if (tun->read_thread) {
        tun->running = 0;
        SetEvent(tun->stop_event);
        WaitForSingleObject(tun->read_thread, INFINITE);
        CloseHandle(tun->read_thread);
        CloseHandle(tun->stop_event);
    }
#endif

    if (tun->socket_id && tun->ua) {
        uasync_remove_socket(tun->ua, tun->socket_id);
    }

    struct ll_entry* pkt;
    while ((pkt = queue_data_get(tun->input_queue)) != NULL) {
        queue_dgram_free(pkt);
        queue_entry_free(pkt);
    }
    queue_free(tun->input_queue);

    while ((pkt = queue_data_get(tun->output_queue)) != NULL) {
        queue_dgram_free(pkt);
        queue_entry_free(pkt);
    }
    queue_free(tun->output_queue);

    memory_pool_destroy(tun->pool);
    if (!tun->test_mode) tun_platform_cleanup(tun);
    u_free(tun);
}

ssize_t tun_write(struct tun_if* tun, const uint8_t* buf, size_t len)
{
    if (!tun || !buf || len <= 1) return -1;

    if (tun->test_mode) {
        // Skip prefix byte for test mode injection
        return tun_inject_packet(tun, buf + 1, len - 1) == 0 ? (ssize_t)(len - 1) : -1;
    }

    // Skip prefix byte, write from second byte
    ssize_t n = tun_platform_write(tun, buf + 1, len - 1);
    if (n > 0) {
        tun->bytes_written += n;
        tun->packets_written++;
        if (debug_should_output(DEBUG_LEVEL_DEBUG, DEBUG_CATEGORY_TUN)) {
            DEBUG_DEBUG(DEBUG_CATEGORY_TUN, "[->TUN] %s", dump_ip_packet_to_buffer(buf + 1, len - 1));
        }
    } else {
        tun->write_errors++;
    }
    return n;
}

struct ll_queue* tun_get_input_queue(struct tun_if* tun)  { return tun ? tun->input_queue : NULL; }
struct ll_queue* tun_get_output_queue(struct tun_if* tun) { return tun ? tun->output_queue : NULL; }
int tun_is_test_mode(struct tun_if* tun) { return tun ? tun->test_mode : -1; }

// ===================================================================
// Callback для обработки пакетов из TUN в main thread
// Вызывается через uasync_post из read thread (Windows)
// ===================================================================
void tun_packet_handler(void* arg) {
    struct tun_packet_data* pd = (struct tun_packet_data*)arg;
    if (!pd || !pd->tun || !pd->entry) {
        DEBUG_ERROR(DEBUG_CATEGORY_TUN, "ARG ZERO");
        if (pd) u_free(pd);
        return;
    }

    struct tun_if* tun = pd->tun;
    struct ll_entry* entry = pd->entry;

    // Освободить временную структуру
    u_free(pd);

    if (debug_should_output(DEBUG_LEVEL_DEBUG, DEBUG_CATEGORY_TUN)) {
        DEBUG_DEBUG(DEBUG_CATEGORY_TUN, "[TUNh->] %s", dump_ip_packet_to_buffer(entry->dgram, entry->len));
    }

    // Положить в очередь (уже в main thread)
    int ok = queue_data_put(tun->output_queue, entry, 0);
    if (ok != 0) {
        DEBUG_ERROR(DEBUG_CATEGORY_TUN, "Put error");
        queue_entry_free(entry);
        tun->read_errors++;
    }
    // Статистика уже обновлена в read thread
}

int tun_inject_packet(struct tun_if* tun, const uint8_t* buf, size_t len)
{
    if (!tun || !buf || len == 0 || len > TUN_MAX_PACKET_SIZE) return -1;

    // Allocate with +1 for prefix byte
    uint8_t* data = u_malloc(len + 1);
    if (!data) { tun->read_errors++; return -1; }
    data[0] = 0;  // Prefix byte
    memcpy(data + 1, buf, len);

    struct ll_entry* pkt = queue_entry_new_from_pool(tun->pool);
    if (!pkt) { u_free(data); tun->read_errors++; return -1; }

    pkt->dgram = data;
    pkt->len = len + 1;

    int ret = queue_data_put(tun->output_queue, pkt, 0);

    if (ret != 0) {
        u_free(data);
        memory_pool_free(tun->pool, pkt);
        tun->read_errors++;
        return -1;
    }

    tun->bytes_read += len + 1;
    tun->packets_read++;
    return 0;
}

ssize_t tun_read_packet(struct tun_if* tun, uint8_t* buf, size_t len)
{
    if (!tun || !buf || len == 0) return -1;

    struct ll_entry* pkt = queue_data_get(tun->input_queue);
    if (!pkt) return 0;

    if (pkt->len > len) {
        queue_data_put(tun->input_queue, pkt, 0);
        return -1;
    }

    // Add prefix byte at start
    buf[0] = 0;
    memcpy(buf + 1, pkt->dgram, pkt->len);
    ssize_t ret = pkt->len + 1;

    queue_dgram_free(pkt);
    queue_entry_free(pkt);

    tun->bytes_written += ret;
    tun->packets_written++;
    if (debug_should_output(DEBUG_LEVEL_DEBUG, DEBUG_CATEGORY_TUN)) {
        DEBUG_DEBUG(DEBUG_CATEGORY_TUN, "[TUN->] %s", dump_ip_packet_to_buffer(buf + 1, pkt->len));
    }
    return ret;
}