#include "ili9341.h"
#include "cmsis_os.h"
#include "uart4.h"
#include <stdarg.h>
#include <stdio.h>

/*
 * ILI9341 TFT LCD 驱动
 * STM32H750 + SPI1 + DMA
 *
 * 本版本相对上一版的主要改动：
 *   1. DMA buffer 不再用裸指针指向硬编码地址，改用链接器 section，
 *      避免和其他默认分配（heap/栈/其他静态变量）物理重叠。
 *      —— 需要在 .ld 文件里配合添加一个 section，见文件末尾说明。
 *   2. DMA 完成/错误的等待方式从"HAL_GetTick() 忙等"改成
 *      "RTOS 信号量 + 超时"，真正让出 CPU，而不是空转。
 *   3. ISR 回调只做最少的事（释放信号量/记录错误码），所有
 *      恢复逻辑、日志打印都放在任务上下文里做，符合
 *      "中断里越短越好"的原则。
 *   4. 增加了 HAL_SPI_ErrorCallback，之前只处理了 TxCplt，
 *      一旦发生 SPI 总线错误（比如 CRC/Overrun/ModeFault），
 *      原来的代码只能傻等 100ms 超时，现在能立刻被唤醒并记录。
 *   5. 调试日志走 Print_DMA（非阻塞 UART DMA 发送），不占用
 *      本来就紧张的 SPI/DMA 时间片。
 */

extern SPI_HandleTypeDef hspi1;

static uint16_t lcd_width  = ILI9341_WIDTH;
static uint16_t lcd_height = ILI9341_HEIGHT;

/* LCD GPIO 控制 */
#define LCD_CS_LOW()      HAL_GPIO_WritePin(LCD_CS_PORT, LCD_CS_PIN, GPIO_PIN_RESET)
#define LCD_CS_HIGH()     HAL_GPIO_WritePin(LCD_CS_PORT, LCD_CS_PIN, GPIO_PIN_SET)

#define LCD_DC_CMD()      HAL_GPIO_WritePin(LCD_DC_PORT, LCD_DC_PIN, GPIO_PIN_RESET)
#define LCD_DC_DATA()     HAL_GPIO_WritePin(LCD_DC_PORT, LCD_DC_PIN, GPIO_PIN_SET)

#define LCD_RST_LOW()     HAL_GPIO_WritePin(LCD_RST_PORT, LCD_RST_PIN, GPIO_PIN_RESET)
#define LCD_RST_HIGH()    HAL_GPIO_WritePin(LCD_RST_PORT, LCD_RST_PIN, GPIO_PIN_SET)

#define LCD_BL_ON()       HAL_GPIO_WritePin(LCD_BL_PORT, LCD_BL_PIN, GPIO_PIN_SET)
#define LCD_BL_OFF()      HAL_GPIO_WritePin(LCD_BL_PORT, LCD_BL_PIN, GPIO_PIN_RESET)

/*
 * LVGL buffer 使用 20 行：
 * 320 * 20 * 2 = 12800 bytes
 */
#define LCD_DMA_MAX_PIXELS     (320U * 20U)
#define LCD_DMA_BUF_SIZE       (LCD_DMA_MAX_PIXELS * 2U)

/*
 * DMA buffer 放到链接器专属 section 里，而不是裸写绝对地址。
 * 需要在 .ld 文件的 MEMORY 区里确认 D2 SRAM（0x30000000 起）
 * 已经声明为一个可用 region（很多 CubeIDE 模板里叫 RAM_D2），
 * 并在 SECTIONS 里加：
 *
 *   .dma_buffer (NOLOAD) :
 *   {
 *     . = ALIGN(32);
 *     *(.dma_buffer)
 *     . = ALIGN(32);
 *   } >RAM_D2
 *
 * 这样链接器会知道这块内存已经被占用，不会再把 heap/栈/其他
 * 全局变量分配到这里，从根本上排除"覆盖冲突"的可能性。
 * 32 字节对齐是为了配合 D-Cache clean-by-address 的对齐要求。
 */
#if defined(__GNUC__)
#define LCD_DMA_BUF_ATTR __attribute__((section(".dma_buffer"), aligned(32)))
#else
#define LCD_DMA_BUF_ATTR
#warning "非 GNU 工具链，请自行确认 lcd_dma_buf 被放置在 DMA 可访问、且不会被其他数据覆盖的内存区域"
#endif

LCD_DMA_BUF_ATTR
static uint8_t lcd_dma_buf[LCD_DMA_BUF_SIZE];

/* ---- 状态与同步 ---- */
static volatile uint8_t  lcd_dma_busy       = 0;
static volatile uint32_t lcd_dma_start_tick = 0;
static volatile uint32_t lcd_spi_last_error = 0;

static osSemaphoreId_t lcd_dma_sem = NULL;

/* ---- 非阻塞调试日志：走 Print_DMA，忙的时候直接丢弃这条日志 ---- */
static char lcd_dbg_line[96];

static void LCD_DebugLogf(const char *fmt, ...)
{
    if (UART_IsBusy())
    {
        return; /* 调试串口正忙，宁可丢日志也不要阻塞关键路径 */
    }

    va_list args;
    va_start(args, fmt);
    vsnprintf(lcd_dbg_line, sizeof(lcd_dbg_line), fmt, args);
    va_end(args);

    Print_DMA(lcd_dbg_line);
}

static void LCD_DMA_EnsureSemaphore(void)
{
    if (lcd_dma_sem == NULL)
    {
        lcd_dma_sem = osSemaphoreNew(1, 0, NULL);
    }
}

/* ========================= 基础 SPI 写命令/数据（阻塞，保持不变） ========================= */

static void ILI9341_WriteCommand(uint8_t cmd)
{
    LCD_CS_LOW();
    LCD_DC_CMD();

    HAL_SPI_Transmit(&hspi1, &cmd, 1, HAL_MAX_DELAY);

    LCD_CS_HIGH();
}

static void ILI9341_WriteData(uint8_t *data, uint16_t size)
{
    LCD_CS_LOW();
    LCD_DC_DATA();

    HAL_SPI_Transmit(&hspi1, data, size, HAL_MAX_DELAY);

    LCD_CS_HIGH();
}

static void ILI9341_WriteData8(uint8_t data)
{
    LCD_CS_LOW();
    LCD_DC_DATA();

    HAL_SPI_Transmit(&hspi1, &data, 1, HAL_MAX_DELAY);

    LCD_CS_HIGH();
}

static void ILI9341_WriteData16(uint16_t data)
{
    uint8_t buf[2];

    buf[0] = (uint8_t)(data >> 8);
    buf[1] = (uint8_t)(data & 0xFF);

    LCD_CS_LOW();
    LCD_DC_DATA();

    HAL_SPI_Transmit(&hspi1, buf, 2, HAL_MAX_DELAY);

    LCD_CS_HIGH();
}

static void ILI9341_Reset(void)
{
    LCD_RST_HIGH();
    HAL_Delay(20);

    LCD_RST_LOW();
    HAL_Delay(20);

    LCD_RST_HIGH();
    HAL_Delay(120);
}

static void ILI9341_SetAddressWindow(uint16_t x0,
                                     uint16_t y0,
                                     uint16_t x1,
                                     uint16_t y1)
{
    uint8_t data[4];

    ILI9341_WriteCommand(0x2A);
    data[0] = (uint8_t)(x0 >> 8);
    data[1] = (uint8_t)(x0 & 0xFF);
    data[2] = (uint8_t)(x1 >> 8);
    data[3] = (uint8_t)(x1 & 0xFF);
    ILI9341_WriteData(data, 4);

    ILI9341_WriteCommand(0x2B);
    data[0] = (uint8_t)(y0 >> 8);
    data[1] = (uint8_t)(y0 & 0xFF);
    data[2] = (uint8_t)(y1 >> 8);
    data[3] = (uint8_t)(y1 & 0xFF);
    ILI9341_WriteData(data, 4);

    ILI9341_WriteCommand(0x2C);
}

/* ========================= 初始化 ========================= */

void ILI9341_Init(void)
{
    LCD_CS_HIGH();
    LCD_BL_ON();

    LCD_DMA_EnsureSemaphore();

    ILI9341_Reset();

    ILI9341_WriteCommand(0x01);
    HAL_Delay(120);

    ILI9341_WriteCommand(0xEF);
    {
        uint8_t data[] = {0x03, 0x80, 0x02};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xCF);
    {
        uint8_t data[] = {0x00, 0xC1, 0x30};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xED);
    {
        uint8_t data[] = {0x64, 0x03, 0x12, 0x81};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xE8);
    {
        uint8_t data[] = {0x85, 0x00, 0x78};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xCB);
    {
        uint8_t data[] = {0x39, 0x2C, 0x00, 0x34, 0x02};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xF7);
    ILI9341_WriteData8(0x20);

    ILI9341_WriteCommand(0xEA);
    {
        uint8_t data[] = {0x00, 0x00};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xC0);
    ILI9341_WriteData8(0x23);

    ILI9341_WriteCommand(0xC1);
    ILI9341_WriteData8(0x10);

    ILI9341_WriteCommand(0xC5);
    {
        uint8_t data[] = {0x3E, 0x28};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xC7);
    ILI9341_WriteData8(0x86);

    ILI9341_WriteCommand(0x36);
    ILI9341_WriteData8(0x48);

    ILI9341_WriteCommand(0x3A);
    ILI9341_WriteData8(0x55);

    ILI9341_WriteCommand(0xB1);
    {
        uint8_t data[] = {0x00, 0x18};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xB6);
    {
        uint8_t data[] = {0x08, 0x82, 0x27};
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xF2);
    ILI9341_WriteData8(0x00);

    ILI9341_WriteCommand(0x26);
    ILI9341_WriteData8(0x01);

    ILI9341_WriteCommand(0xE0);
    {
        uint8_t data[] = {
            0x0F, 0x31, 0x2B, 0x0C, 0x0E,
            0x08, 0x4E, 0xF1, 0x37, 0x07,
            0x10, 0x03, 0x0E, 0x09, 0x00
        };
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0xE1);
    {
        uint8_t data[] = {
            0x00, 0x0E, 0x14, 0x03, 0x11,
            0x07, 0x31, 0xC1, 0x48, 0x08,
            0x0F, 0x0C, 0x31, 0x36, 0x0F
        };
        ILI9341_WriteData(data, sizeof(data));
    }

    ILI9341_WriteCommand(0x11);
    HAL_Delay(120);

    ILI9341_WriteCommand(0x29);
    HAL_Delay(20);

    ILI9341_SetRotation(0);
}

/* ========================= 方向设置 ========================= */

void ILI9341_SetRotation(uint8_t rotation)
{
    uint8_t madctl;

    rotation %= 4;

    switch (rotation)
    {
        case 0:
            madctl = 0x48;
            lcd_width = 240;
            lcd_height = 320;
            break;

        case 1:
            madctl = 0x28;
            lcd_width = 320;
            lcd_height = 240;
            break;

        case 2:
            madctl = 0x88;
            lcd_width = 240;
            lcd_height = 320;
            break;

        default:
            madctl = 0xE8;
            lcd_width = 320;
            lcd_height = 240;
            break;
    }

    ILI9341_WriteCommand(0x36);
    ILI9341_WriteData8(madctl);
}

/* ========================= 基础绘图（阻塞，保持不变） ========================= */

void ILI9341_DrawPixel(uint16_t x, uint16_t y, uint16_t color)
{
    if (x >= lcd_width || y >= lcd_height)
    {
        return;
    }

    ILI9341_SetAddressWindow(x, y, x, y);
    ILI9341_WriteData16(color);
}

void ILI9341_FillRect(uint16_t x,
                      uint16_t y,
                      uint16_t w,
                      uint16_t h,
                      uint16_t color)
{
    uint32_t pixels;
    uint8_t buffer[128];
    uint16_t chunk_pixels;

    if (x >= lcd_width || y >= lcd_height)
    {
        return;
    }

    if ((x + w) > lcd_width)
    {
        w = lcd_width - x;
    }

    if ((y + h) > lcd_height)
    {
        h = lcd_height - y;
    }

    pixels = (uint32_t)w * h;

    for (uint16_t i = 0; i < sizeof(buffer); i += 2)
    {
        buffer[i]     = (uint8_t)(color >> 8);
        buffer[i + 1] = (uint8_t)(color & 0xFF);
    }

    ILI9341_SetAddressWindow(x, y, x + w - 1, y + h - 1);

    LCD_CS_LOW();
    LCD_DC_DATA();

    while (pixels > 0)
    {
        chunk_pixels = pixels > 64U ? 64U : (uint16_t)pixels;

        HAL_SPI_Transmit(&hspi1,
                         buffer,
                         chunk_pixels * 2U,
                         HAL_MAX_DELAY);

        pixels -= chunk_pixels;
    }

    LCD_CS_HIGH();
}

void ILI9341_FillScreen(uint16_t color)
{
    ILI9341_FillRect(0, 0, lcd_width, lcd_height, color);
}

uint16_t ILI9341_Color565(uint8_t r, uint8_t g, uint8_t b)
{
    return ((r & 0xF8U) << 8) |
           ((g & 0xFCU) << 3) |
           (b >> 3);
}

/* ========================= LVGL 阻塞区域刷新（保持不变） ========================= */

void ILI9341_WriteArea(uint16_t x,
                       uint16_t y,
                       uint16_t w,
                       uint16_t h,
                       const uint16_t *pixels)
{
    uint32_t total_pixels;
    uint32_t sent_pixels = 0;

    static uint8_t tx_buf[512];

    if (w == 0 || h == 0 || pixels == NULL)
    {
        return;
    }

    total_pixels = (uint32_t)w * h;

    ILI9341_SetAddressWindow(x, y, x + w - 1, y + h - 1);

    LCD_CS_LOW();
    LCD_DC_DATA();

    while (sent_pixels < total_pixels)
    {
        uint32_t remain = total_pixels - sent_pixels;
        uint32_t chunk_pixels = remain > 256U ? 256U : remain;

        for (uint32_t i = 0; i < chunk_pixels; i++)
        {
            uint16_t color = pixels[sent_pixels + i];

            tx_buf[i * 2U]     = (uint8_t)(color >> 8);
            tx_buf[i * 2U + 1] = (uint8_t)(color & 0xFF);
        }

        HAL_SPI_Transmit(&hspi1,
                         tx_buf,
                         chunk_pixels * 2U,
                         HAL_MAX_DELAY);

        sent_pixels += chunk_pixels;
    }

    LCD_CS_HIGH();
}

/* ========================= LVGL DMA 区域刷新：重写版 ========================= */

static void LCD_CleanDCache(void *addr, uint32_t size)
{
#if (__DCACHE_PRESENT == 1U)
    if ((SCB->CCR & SCB_CCR_DC_Msk) != 0U)
    {
        uint32_t start_addr = (uint32_t)addr & ~31U;
        uint32_t end_addr = ((uint32_t)addr + size + 31U) & ~31U;

        SCB_CleanDCache_by_Addr((uint32_t *)start_addr, end_addr - start_addr);
    }
#else
    (void)addr;
    (void)size;
#endif
}

static void ILI9341_SPI_WaitIdle(void)
{
    uint32_t tick_start;

#if defined(SPI_FLAG_TXC)
    tick_start = HAL_GetTick();

    while (__HAL_SPI_GET_FLAG(&hspi1, SPI_FLAG_TXC) == RESET)
    {
        if ((HAL_GetTick() - tick_start) > 20U)
        {
            break;
        }
    }
#elif defined(SPI_FLAG_EOT)
    tick_start = HAL_GetTick();

    while (__HAL_SPI_GET_FLAG(&hspi1, SPI_FLAG_EOT) == RESET)
    {
        if ((HAL_GetTick() - tick_start) > 20U)
        {
            break;
        }
    }
#else
    for (volatile uint32_t i = 0; i < 1000U; i++)
    {
        __NOP();
    }
#endif

    __DSB();
}

void ILI9341_WriteArea_DMA(uint16_t x,
                           uint16_t y,
                           uint16_t w,
                           uint16_t h,
                           const uint16_t *pixels,
                           ILI9341_DmaDoneCallback done_cb,
                           void *user_data)
{
    uint32_t total_pixels;
    uint32_t byte_len;

    if (w == 0 || h == 0 || pixels == NULL)
    {
        if (done_cb != NULL)
        {
            done_cb(user_data);
        }
        return;
    }

    total_pixels = (uint32_t)w * h;

    if (total_pixels > LCD_DMA_MAX_PIXELS)
    {
        /* 超出单次 DMA buffer 容量，退化为阻塞传输 */
        ILI9341_WriteArea(x, y, w, h, pixels);

        if (done_cb != NULL)
        {
            done_cb(user_data);
        }

        return;
    }

    LCD_DMA_EnsureSemaphore();

    if (lcd_dma_busy)
    {
        /*
         * 正常情况下，LVGL 在上一次 flush 没有调用 flush_ready
         * 之前不会发起新的 flush，所以走到这里说明上一次传输
         * 大概率已经出了问题（信号量没有被正确释放）。
         * 这里做一次强制恢复，而不是继续等待，避免连锁卡死。
         */
        LCD_DebugLogf("[LCD] WARN busy-reenter, force recover\r\n");

        HAL_SPI_Abort(&hspi1);
        LCD_CS_HIGH();
        lcd_dma_busy = 0;
    }

    lcd_dma_busy = 1;
    lcd_spi_last_error = 0;

    /* 组包到 DMA buffer */
    for (uint32_t i = 0; i < total_pixels; i++)
    {
        uint16_t color = pixels[i];

        lcd_dma_buf[i * 2U]     = (uint8_t)(color >> 8);
        lcd_dma_buf[i * 2U + 1] = (uint8_t)(color & 0xFF);
    }

    byte_len = total_pixels * 2U;

    LCD_CleanDCache(lcd_dma_buf, byte_len);

    ILI9341_SetAddressWindow(x, y, x + w - 1U, y + h - 1U);

    lcd_dma_start_tick = HAL_GetTick();

    LCD_CS_LOW();
    LCD_DC_DATA();

    if (HAL_SPI_Transmit_DMA(&hspi1, lcd_dma_buf, (uint16_t)byte_len) != HAL_OK)
    {
        LCD_CS_HIGH();

        lcd_dma_busy = 0;

        LCD_DebugLogf("[LCD] ERR DMA start failed\r\n");

        if (done_cb != NULL)
        {
            done_cb(user_data);
        }

        return;
    }

    /*
     * 真正的关键改动：这里不再用 HAL_GetTick() 忙等，而是让任务
     * 通过信号量挂起，把 CPU 让给调度器；DMA 完成中断
     * （HAL_SPI_TxCpltCallback）或错误中断（HAL_SPI_ErrorCallback）
     * 会释放这个信号量把任务唤醒。100ms 超时依旧作为兜底。
     */
    if (osSemaphoreAcquire(lcd_dma_sem, 100) != osOK)
    {
        LCD_DebugLogf("[LCD] ERR DMA timeout, aborting\r\n");
        HAL_SPI_Abort(&hspi1);
    }
    else if (lcd_spi_last_error != 0U)
    {
        LCD_DebugLogf("[LCD] ERR SPI err=0x%08lX\r\n",
                     (unsigned long)lcd_spi_last_error);
        lcd_spi_last_error = 0U;
    }

    ILI9341_SPI_WaitIdle();

    LCD_CS_HIGH();

    lcd_dma_busy = 0;

    if (done_cb != NULL)
    {
        done_cb(user_data);
    }
}

void ILI9341_DMA_Watchdog(uint32_t timeout_ms)
{
    if (!lcd_dma_busy)
    {
        return;
    }

    uint32_t elapsed = HAL_GetTick() - lcd_dma_start_tick;

    if (elapsed > timeout_ms)
    {
        LCD_DebugLogf("[LCD] WDG force recover after %lums\r\n",
                     (unsigned long)elapsed);

        HAL_SPI_Abort(&hspi1);
        LCD_CS_HIGH();

        lcd_dma_busy = 0;

        /* 如果此时正好有任务卡在 WriteArea_DMA 里等信号量，唤醒它 */
        if (lcd_dma_sem != NULL)
        {
            osSemaphoreRelease(lcd_dma_sem);
        }
    }
}

/*
 * SPI1 TX DMA 完成回调（ISR 上下文）。
 * 只做最必要的事：释放信号量。不打印日志、不调用 LVGL。
 */
void HAL_SPI_TxCpltCallback(SPI_HandleTypeDef *hspi)
{
    if (hspi->Instance == SPI1)
    {
        if (lcd_dma_sem != NULL)
        {
            osSemaphoreRelease(lcd_dma_sem);
        }
    }
}

/*
 * SPI 错误回调（ISR 上下文）。
 * 之前的版本没有实现这个回调，一旦发生总线错误（Overrun/
 * ModeFault/CRC 等），只能傻等 100ms 超时才能恢复。
 * 现在记录错误码并立即释放信号量唤醒等待的任务，
 * 真正的日志打印和处理放到任务上下文（WriteArea_DMA 里）去做。
 */
void HAL_SPI_ErrorCallback(SPI_HandleTypeDef *hspi)
{
    if (hspi->Instance == SPI1)
    {
        lcd_spi_last_error = hspi->ErrorCode;

        if (lcd_dma_sem != NULL)
        {
            osSemaphoreRelease(lcd_dma_sem);
        }
    }
}