时间轮

老早之前就听说时间轮算法特别高效，Linux内核都用的它，这两天抽空实现了遍……嗯，被差一bug搞死(~￣▽￣~) 啊哈

网上扣来的图，原理好懂：轮子里的每格代表一小段时间（精度），连起来就能表示时间点了（我去年买了个表），格子内含链表，中存回调函数；时间指针每次转动一格，指向某格时，取出链表里的回调函数依次执行，后清空链表，等待下一次转动。

加入节点逻辑也简单：在轮子可表示的时间范围内（格子数*格子精度），配合当前时针位置，对格子总数取余，即得本节点需放哪个格子。

进一步为扩大时间轮的表示范围，使用分级方式，跟时分秒一样，上一级转一圈，下一级动一格。

对吧，很容易理解……然后coding是另一码事(╯‵□′)╯︵┴─┴

首先，最重要的，数据结构：用了环形链表，方便增删。

struct NodeLink {NodeLink* prev;NodeLink* next;NodeLink() { prev = next = this; } //circle
};
struct stWheel {NodeLink* slots; //每个slot维护的node链表为一个环，slot->next为node链表中第一个节点，prev为node的最后一个节点const uint32 size;uint32 slotIdx;stWheel(uint32 n) : size(n), slotIdx(0){ slots = new NodeLink[size]; }~stWheel() {if (slots) {for (uint32 j = 0; j < size; ++j) {NodeLink* link = (slots + j)->next;while (link != slots + j) {TimerNode* node = (TimerNode*)link;link = node->link.next;delete node;}}delete[]slots;}}
};

具体时间节点的数据结构如下：

struct TimerNode {Pool_Obj_Define(TimerNode, 32) //内存池声明，不含数据NodeLink link; //must in the head
    uint32 timeDead;uint32 interval; //间隔多久int loop;        //总共循环多久std::function<void()> func;
};

TimeNode里保存上下关系，stWheel的NodeLink辅助用的，环状链表的头，没实际数据，用以记录首尾TimeNode。

核心代码如下：

——增删节点——

void CTimerMgr::_AddTimerNode(uint32 milseconds, TimerNode* node) {NodeLink* slot = NULL;uint32 tickCnt = milseconds / TIME_TICK_LEN;if (tickCnt < WHEEL_CAP[0]) {uint32 index = (_wheels[0]->slotIdx + tickCnt) & (WHEEL_SIZE[0] - 1); //2的N次幂位操作取余slot = _wheels[0]->slots + index;} else {for (int i = 1; i < WHEEL_NUM; ++i) {if (tickCnt < WHEEL_CAP[i]) {uint32 preCap = WHEEL_CAP[i - 1]; //上一级总容量即为本级的一格容量uint32 index = (_wheels[i]->slotIdx + tickCnt / preCap - 1) & (WHEEL_SIZE[i] - 1); //勿忘-1slot = _wheels[i]->slots + index;break;}}}NodeLink* link = &(node->link);link->prev = slot->prev; //插入格子的prev位置(尾节点)link->prev->next = link;link->next = slot;slot->prev = link;
}
void CTimerMgr::RemoveTimer(TimerNode* node) {LOG_TRACK("node[%p], timeDead[%lld]", node, node->timeDead);NodeLink* link = &(node->link);if (link->prev) {link->prev->next = link->next;}if (link->next) {link->next->prev = link->prev;}link->prev = link->next = NULL;delete node;
}

——轮子启动——

void CTimerMgr::CheckTimerList(const uint32 timenow) {uint32 tickCnt = timenow > _checkTime ? (timenow - _checkTime) / TIME_TICK_LEN : 0;//if (tickCnt) Printf();for (uint32 i = 0; i < tickCnt; ++i) { //扫过的slot均超时stWheel* wheel = _wheels[0];NodeLink* slot = wheel->slots + wheel->slotIdx;NodeLink* link = slot->next;slot->next = slot->prev = slot; //清空当前格子while (link != slot) {            //环形链表遍历TimerNode* node = (TimerNode*)link;link = node->link.next; //得放在前面，后续函数调用，可能会更改node的链接关系
            AddToReadyNode(node);}if (++(wheel->slotIdx) >= wheel->size) {wheel->slotIdx = 0;Cascade(1, timenow); //跳级
        }_checkTime += TIME_TICK_LEN;}DoTimeOutCallBack();
}
uint32 CTimerMgr::Cascade(uint32 wheelIdx, const uint32 timenow) {if (wheelIdx < 1 || wheelIdx >= WHEEL_NUM) {return 0;}int casCnt = 0;stWheel* wheel = _wheels[wheelIdx];NodeLink* slot = wheel->slots + wheel->slotIdx;NodeLink* link = slot->next;slot->next = slot->prev = slot; //清空当前格子while (link != slot) {TimerNode* node = (TimerNode*)link;link = node->link.next;if (node->timeDead <= timenow) {AddToReadyNode(node);} else {_AddTimerNode(node->timeDead - timenow, node); //本级精度下已超时，精度提升，重新加一遍++casCnt;LOG_TRACK("wheelIdx[%u], link[%p], milseconds[%u]", wheelIdx, link, node->timeDead - timenow);}}if (++(wheel->slotIdx) >= wheel->size) {wheel->slotIdx = 0;casCnt += Cascade(++wheelIdx, timenow);}return casCnt;
}