/**
 * @file cpu_load.cpp
 * @brief CPU 负载率测量实现（1.5.61，度量定义与并发协议见 cpu_load.h）。
 *
 * 热路径(beginWork/endWork/isrEnter/isrExit/tick)放进 .ramfunc——它们会被包进
 * EXTI0(Sync0, preempt=0) 等延迟敏感的 ISR 包装，与 EthercatSync::onSync0Exti
 * 同款 section 属性，避免 flash 等待态给 Sync0 入口添抖动。1s 一次的窗口
 * 结算(含 64 位除法)留在普通 flash 函数里。
 */
#include "sys/cpu_load.h"

#include "RTE_Components.h"
#include CMSIS_device_header

volatile uint32_t CpuLoad::s_depth    = 0U;
volatile uint32_t CpuLoad::s_stamp    = 0U;
volatile uint32_t CpuLoad::s_busy     = 0U;
volatile uint32_t CpuLoad::s_winStart = 0U;
volatile uint16_t CpuLoad::s_loadPerm = 0U;
volatile uint32_t CpuLoad::s_seq      = 0U;

/** @brief 区间并集的共享进入点。0→1 者负责盖时间戳；STREX 失败方重试时深度
 *         已被抢占者修正，不会再重复盖戳。 */
__attribute__((always_inline)) static inline void cpuLoadEnter()
{
    uint32_t d;
    do {
        d = __LDREXW(&CpuLoad::s_depth);
        if (d == 0U) {
            CpuLoad::s_stamp = DWT->CYCCNT;
            __DMB();
        }
    } while (__STREXW(d + 1U, &CpuLoad::s_depth) != 0U);
    __DMB();
}

/** @brief 区间并集的共享退出点。1→0 者把区间累进 s_busy（同样走独占访问，
 *         防"ISR 恰好插在主循环退出的读-改-写之间"丢累加）。 */
__attribute__((always_inline)) static inline void cpuLoadLeave()
{
    const uint32_t end   = DWT->CYCCNT;
    const uint32_t start = CpuLoad::s_stamp;
    uint32_t d;
    do {
        d = __LDREXW(&CpuLoad::s_depth);
    } while (__STREXW(d - 1U, &CpuLoad::s_depth) != 0U);
    if (d == 1U) {
        const uint32_t delta = end - start;
        uint32_t b;
        do {
            b = __LDREXW(&CpuLoad::s_busy);
        } while (__STREXW(b + delta, &CpuLoad::s_busy) != 0U);
    }
}

void CpuLoad::init()
{
    /* DWT 已由 EscLan9252::init 使能；这里幂等兜底（万一 EtherCAT 初始化被裁剪）。 */
    CoreDebug->DEMCR |= CoreDebug_DEMCR_TRCENA_Msk;
    DWT->CTRL |= DWT_CTRL_CYCCNTENA_Msk;

    s_depth    = 0U;
    s_busy     = 0U;
    s_winStart = DWT->CYCCNT;
    s_loadPerm = 0U;
    s_seq      = 0U;
}

__attribute__((section(".ramfunc"), noinline)) void CpuLoad::beginWork()
{
    cpuLoadEnter();
}

__attribute__((section(".ramfunc"), noinline)) void CpuLoad::isrEnter()
{
    cpuLoadEnter();
}

__attribute__((section(".ramfunc"), noinline)) void CpuLoad::isrExit()
{
    cpuLoadLeave();
}

/**
 * @brief 窗口结算（普通 flash 函数，1s 才进一次）。
 * @note  s_busy 用独占访问换出：窗口边界上 ISR 的在途累加要么算进本窗口、
 *        要么算进下窗口，不丢。
 */
static void cpuLoadRollWindow(uint32_t now)
{
    const uint32_t elapsed = now - CpuLoad::s_winStart; /* 调用方已保证 >= 1s */
    uint32_t busy;
    do {
        busy = __LDREXW(&CpuLoad::s_busy);
    } while (__STREXW(0U, &CpuLoad::s_busy) != 0U);
    CpuLoad::s_winStart = now;

    uint32_t perm = (uint32_t)(((uint64_t)busy * 1000ULL) / (uint64_t)elapsed);
    if (perm > 1000U) {
        perm = 1000U;
    }
    CpuLoad::s_loadPerm = (uint16_t)perm;
    CpuLoad::s_seq++;
}

__attribute__((section(".ramfunc"), noinline)) void CpuLoad::endWork()
{
    cpuLoadLeave();
}

__attribute__((section(".ramfunc"), noinline)) void CpuLoad::tick()
{
    const uint32_t now = DWT->CYCCNT;
    if ((uint32_t)(now - s_winStart) >= SystemCoreClock) {
        cpuLoadRollWindow(now);
    }
}

/* C 侧挂钩：供 ecatappl.c / dc_app.c 计入 EtherCAT 过程数据路径。
 * 与 beginWork/endWork 共用深度计数器，嵌套不重不漏。 */
extern "C" void CpuLoad_Enter(void) { CpuLoad::beginWork(); }
extern "C" void CpuLoad_Leave(void) { CpuLoad::endWork(); }