Swoole 源碼分析——Server模塊之Timer模塊與時間輪算法

前言

swoole 的 timer 模塊功能有三個：用戶定時任務、剔除空閒鏈接、更新 server 時間。timer 模塊的底層有兩種，一種是基於 alarm 信號，一種是基於 timefd。

timer 數據結構

timer 數據結構是 swTimer。其中 heap 是多個 swTimer_node 類型構成的一個數據堆，該數據堆按照下一次執行時間來排序，下次執行時間離當前時間越近，元素的位置越靠前；map 是 swTimer_node 類型的 map，其 key 是 swTimer_node 類型的 id，該數據結構能夠經過 id 快速查找對應的 swTimer_node 元素；num 是 swTimer_node 元素個數；use_pipe 標誌着 worker 進程中是否使用管道 pipe 來獲知 alarm 信號已觸發；fd 用於 timefd；_current_id 是當前最大 swTimer_node 的 id；_next_id 就是下一個新建的 swTimer_node 的 id 值，是 _current_id + 1；_next_msec 是下次檢查定時器的時間。

_swTimer_node 中 heap_node 是 _swTimer 中的數據堆元素；data 通常存儲 server；callback 是定時器觸發後須要執行的回調函數；exec_msec 是該元素應該執行的時間；id 是元素在 swTimer 中的 id；type 有三種：SW_TIMER_TYPE_KERNEL（server 內置定時函數）、SW_TIMER_TYPE_CORO（協程定時函數）、SW_TIMER_TYPE_PHP(PHP 定時函數)

struct _swTimer
{
    /*--------------timerfd & signal timer--------------*/
    swHeap *heap;
    swHashMap *map;
    int num;
    int use_pipe;
    int lasttime;
    int fd;
    long _next_id;
    long _current_id;
    long _next_msec;
    swPipe pipe;
    /*-----------------for EventTimer-------------------*/
    struct timeval basetime;
    /*--------------------------------------------------*/
    int (*set)(swTimer *timer, long exec_msec);
    swTimer_node* (*add)(swTimer *timer, int _msec, int persistent, void *data, swTimerCallback callback);
};

struct _swTimer_node
{
    swHeap_node *heap_node;
    void *data;
    swTimerCallback callback;
    int64_t exec_msec;
    uint32_t interval;
    long id;
    int type;                 //0 normal node 1 node for client_coro
    uint8_t remove;
};

Timer 定時器

swTimer_init 建立定時器

• 建立定時器須要給定一個間隔時間，每隔這個時間就要檢查 swTimer 中的 _swTimer_node 元素，若是時間已經超過了 _swTimer_node 元素的 exec_msec 時間，就要執行定時函數。
• swTimer_now 函數初始化 basetime：swTimer_now 函數能夠獲取當前時間，使用的是 clock_gettime 與 CLOCK_MONOTONIC 獲取絕對時間，或者使用 gettimeofday 函數
• 若是是 worker 進程，那麼調用 swSystemTimer_init 函數對定時器進行初始化；若是是 master 進程，那麼調用 swReactorTimer_init 進行初始化

int swTimer_now(struct timeval *time)
{
#if defined(SW_USE_MONOTONIC_TIME) && defined(CLOCK_MONOTONIC)
    struct timespec _now;
    if (clock_gettime(CLOCK_MONOTONIC, &_now) < 0)
    {
        swSysError("clock_gettime(CLOCK_MONOTONIC) failed.");
        return SW_ERR;
    }
    time->tv_sec = _now.tv_sec;
    time->tv_usec = _now.tv_nsec / 1000;
#else
    if (gettimeofday(time, NULL) < 0)
    {
        swSysError("gettimeofday() failed.");
        return SW_ERR;
    }
#endif
    return SW_OK;
}

int swTimer_init(long msec)
{
    if (swTimer_now(&SwooleG.timer.basetime) < 0)
    {
        return SW_ERR;
    }


    SwooleG.timer.heap = swHeap_new(1024, SW_MIN_HEAP);
    if (!SwooleG.timer.heap)
    {
        return SW_ERR;
    }

    SwooleG.timer.map = swHashMap_new(SW_HASHMAP_INIT_BUCKET_N, NULL);
    if (!SwooleG.timer.map)
    {
        swHeap_free(SwooleG.timer.heap);
        SwooleG.timer.heap = NULL;
        return SW_ERR;
    }

    SwooleG.timer._current_id = -1;
    SwooleG.timer._next_msec = msec;
    SwooleG.timer._next_id = 1;
    SwooleG.timer.add = swTimer_add;

    if (swIsTaskWorker())
    {
        swSystemTimer_init(msec, SwooleG.use_timer_pipe);
    }
    else
    {
        swReactorTimer_init(msec);
    }

    return SW_OK;
}

swReactorTimer_init 初始化

對於 master 進程，只須要設置 main_reactor 的超時時間便可，當發生超時事件以後，main_reactor 會調用 onTimeout 函數；或者一個事件循環最後，會調用 onFinish 函數；這兩個函數都會最終調用 swTimer_select，來篩選那些已經到了執行時間的元素。

static int swReactorTimer_init(long exec_msec)
{
    SwooleG.main_reactor->check_timer = SW_TRUE;
    SwooleG.main_reactor->timeout_msec = exec_msec;
    SwooleG.timer.set = swReactorTimer_set;
    SwooleG.timer.fd = -1;
    return SW_OK;
}

static int swReactorEpoll_wait(swReactor *reactor, struct timeval *timeo)
{
    ...
    
    if (reactor->timeout_msec == 0)
    {
        if (timeo == NULL)
        {
            reactor->timeout_msec = -1;
        }
        else
        {
            reactor->timeout_msec = timeo->tv_sec * 1000 + timeo->tv_usec / 1000;
        }
    }
    
    while (reactor->running > 0)
    {
        msec = reactor->timeout_msec;
        n = epoll_wait(epoll_fd, events, max_event_num, msec);
        if (n < 0)
        {
           ...
        }
        else if (n == 0)
        {
            if (reactor->onTimeout != NULL)
            {
                reactor->onTimeout(reactor);
            }
            continue;
        }
        
        ...
        
        if (reactor->onFinish != NULL)
        {
            reactor->onFinish(reactor);
        }
        
        ...
    }
    
    ...
}

static void swReactor_onTimeout(swReactor *reactor)
{
    swReactor_onTimeout_and_Finish(reactor);

    if (reactor->disable_accept)
    {
        reactor->enable_accept(reactor);
        reactor->disable_accept = 0;
    }
}

static void swReactor_onFinish(swReactor *reactor)
{
    //check signal
    if (reactor->singal_no)
    {
        swSignal_callback(reactor->singal_no);
        reactor->singal_no = 0;
    }
    swReactor_onTimeout_and_Finish(reactor);
}

static void swReactor_onTimeout_and_Finish(swReactor *reactor)
{
    if (reactor->check_timer)
    {
        swTimer_select(&SwooleG.timer);
    }
    
    ...
}

swSystemTimer_init 初始化

• 對於 worker 進程來講，因爲定時任務比較多並且複雜，就不能簡單使用 reactor 超時來實現功能。
• swSystemTimer_init 採用 SIGALRM 鬧鐘信號或者 timefd 來觸發中斷 reactor 的等待。
• 對於 timefd 來講，須要使用 timerfd_settime 系統調用來設置超時時間，而後將 timefd 加入 worker 的 reactor 監控中，將其當作文件描述符來監控。當其就緒時，會調用 swTimer_select 執行定時函數。
• 對於普通 SIGALRM 信號來講，將 timer->pipe 放入 reactor 的監控中，使用 setitimer 來定時觸發 SIGALRM 信號，設置信號處理函數。信號處理函數中，會向 timer->pipe 寫入數據，進而觸發 swTimer_select 執行定時函數。

int swSystemTimer_init(int interval, int use_pipe)
{
    swTimer *timer = &SwooleG.timer;
    timer->lasttime = interval;

#ifndef HAVE_TIMERFD
    SwooleG.use_timerfd = 0;
#endif

    if (SwooleG.use_timerfd)
    {
        if (swSystemTimer_timerfd_set(timer, interval) < 0)
        {
            return SW_ERR;
        }
        timer->use_pipe = 0;
    }
    else
    {
        if (use_pipe)
        {
            if (swPipeNotify_auto(&timer->pipe, 0, 0) < 0)
            {
                return SW_ERR;
            }
            timer->fd = timer->pipe.getFd(&timer->pipe, 0);
            timer->use_pipe = 1;
        }
        else
        {
            timer->fd = 1;
            timer->use_pipe = 0;
        }

        if (swSystemTimer_signal_set(timer, interval) < 0)
        {
            return SW_ERR;
        }
        swSignal_add(SIGALRM, swSystemTimer_signal_handler);
    }

    if (timer->fd > 1)
    {
        SwooleG.main_reactor->setHandle(SwooleG.main_reactor, SW_FD_TIMER, swSystemTimer_event_handler);
        SwooleG.main_reactor->add(SwooleG.main_reactor, SwooleG.timer.fd, SW_FD_TIMER);
    }
    timer->set = swSystemTimer_set;
    return SW_OK;
}

swSystemTimer_timerfd_set 設置 timefd

• 該函數目的是使用 timerfd_settime 系統調用，該系統調用須要 timefd 和 itimerspec 類型對象
• timefd 能夠由 timerfd_create 系統函數建立
• itimerspec 對象須要當前時間和 interval 間隔時間共同設置。it_value 是首次超時時間，須要填寫當前時間，並加上要超時的時間，值得注意的是 tv_nsec 加上去後必定要判斷是否超出1000000000（若是超過要秒加一），不然會設置失敗；it_interval 是後續週期性超時時間。

static int swSystemTimer_timerfd_set(swTimer *timer, long interval)
{

    struct timeval now;
    int sec = interval / 1000;
    int msec = (((float) interval / 1000) - sec) * 1000;

    if (gettimeofday(&now, NULL) < 0)
    {
        swWarn("gettimeofday() failed. Error: %s[%d]", strerror(errno), errno);
        return SW_ERR;
    }

    struct itimerspec timer_set;
    bzero(&timer_set, sizeof(timer_set));

    if (interval < 0)
    {
        if (timer->fd == 0)
        {
            return SW_OK;
        }
    }
    else
    {
        timer_set.it_interval.tv_sec = sec;
        timer_set.it_interval.tv_nsec = msec * 1000 * 1000;

        timer_set.it_value.tv_sec = now.tv_sec + sec;
        timer_set.it_value.tv_nsec = (now.tv_usec * 1000) + timer_set.it_interval.tv_nsec;

        if (timer_set.it_value.tv_nsec > 1e9)
        {
            timer_set.it_value.tv_nsec = timer_set.it_value.tv_nsec - 1e9;
            timer_set.it_value.tv_sec += 1;
        }

        if (timer->fd == 0)
        {
            timer->fd = timerfd_create(CLOCK_REALTIME, TFD_NONBLOCK | TFD_CLOEXEC);
            if (timer->fd < 0)
            {
                swWarn("timerfd_create() failed. Error: %s[%d]", strerror(errno), errno);
                return SW_ERR;
            }
        }
    }

    if (timerfd_settime(timer->fd, TFD_TIMER_ABSTIME, &timer_set, NULL) == -1)
    {
        swWarn("timerfd_settime() failed. Error: %s[%d]", strerror(errno), errno);
        return SW_ERR;
    }
    return SW_OK;
#else
    swWarn("kernel not support timerfd.");
    return SW_ERR;
#endif
}

swSystemTimer_signal_set 設置信號超時時間

• setitimer 是一個比較經常使用的函數，可用來實現延時和定時的功能。

  • ITIMER_REAL：以系統真實的時間來計算，它送出 SIGALRM 信號。
  • ITIMER_VIRTUAL：以該進程在用戶態下花費的時間來計算，它送出 SIGVTALRM 信號。
  • ITIMER_PROF：以該進程在用戶態下和內核態下所費的時間來計算，它送出 SIGPROF 信號。
  • it_interval 爲計時間隔，it_value 爲延時時長，也就是距離現有時間第一次延遲觸發的相對時間，而不是絕對時間。（因此我認爲代碼中 gettimeofday 函數是多餘的，並不須要獲取當前時間）

*/
static int swSystemTimer_signal_set(swTimer *timer, long interval)
{
    struct itimerval timer_set;
    int sec = interval / 1000;
    int msec = (((float) interval / 1000) - sec) * 1000;

    struct timeval now;
    if (gettimeofday(&now, NULL) < 0)
    {
        swWarn("gettimeofday() failed. Error: %s[%d]", strerror(errno), errno);
        return SW_ERR;
    }
    bzero(&timer_set, sizeof(timer_set));

    if (interval > 0)
    {
        timer_set.it_interval.tv_sec = sec;
        timer_set.it_interval.tv_usec = msec * 1000;

        timer_set.it_value.tv_sec = sec;
        timer_set.it_value.tv_usec = timer_set.it_interval.tv_usec;

        if (timer_set.it_value.tv_usec > 1e6)
        {
            timer_set.it_value.tv_usec = timer_set.it_value.tv_usec - 1e6;
            timer_set.it_value.tv_sec += 1;
        }
    }

    if (setitimer(ITIMER_REAL, &timer_set, NULL) < 0)
    {
        swWarn("setitimer() failed. Error: %s[%d]", strerror(errno), errno);
        return SW_ERR;
    }
    return SW_OK;
}

swSystemTimer_signal_handler 超時信號處理函數

swSystemTimer_signal_handler 函數是 SIGALARM 信號的處理函數，該函數被觸發說明 epoll_wait 函數被鬧鐘信號中斷，此時本函數向 timer.pipe 寫入數據，而後即返回。reactor 會檢測到 timer.pipe 的寫就緒，進而調用對應的回調函數 swSystemTimer_event_handler

void swSystemTimer_signal_handler(int sig)
{
    SwooleG.signal_alarm = 1;
    uint64_t flag = 1;

    if (SwooleG.timer.use_pipe)
    {
        SwooleG.timer.pipe.write(&SwooleG.timer.pipe, &flag, sizeof(flag));
    }
}

swSystemTimer_event_handler 寫就緒回調函數

寫就緒回調函數多是由 timer.pipe 的寫就緒觸發，也多是 timefd 的寫就緒觸發，不管哪一個都會調用 swTimer_select 函數執行對應的定時函數。

int swSystemTimer_event_handler(swReactor *reactor, swEvent *event)
{
    uint64_t exp;
    swTimer *timer = &SwooleG.timer;

    if (read(timer->fd, &exp, sizeof(uint64_t)) != sizeof(uint64_t))
    {
        return SW_ERR;
    }
    SwooleG.signal_alarm = 0;
    return swTimer_select(timer);
}

swTimer_add 添加元素

• swTimer_add 用於添加定時函數元素。本函數邏輯比較簡單，新建一個 swTimer_node 對象，初始化賦值以後加入到 timer->heap 中，程序會自動根據其 exec_msec 進行有小到大的排序，而後再更新 timer->map 哈希表。
• 值得注意的是，當新添加的定時函數須要執行的時間小於當前 timer 下次執行時間的時候，咱們須要調用 timer->set 函數更新 time 的間隔時間。在 master 進程中，這個 set 函數是 swReactorTimer_set，用於設置 reactor 的超時時間；在 worker 進程中，set 函數是 swSystemTimer_set，用於更新 timerfd_settime 或 setitimer 函數。

static swTimer_node* swTimer_add(swTimer *timer, int _msec, int interval, void *data, swTimerCallback callback)
{
    swTimer_node *tnode = sw_malloc(sizeof(swTimer_node));
    if (!tnode)
    {
        swSysError("malloc(%ld) failed.", sizeof(swTimer_node));
        return NULL;
    }

    int64_t now_msec = swTimer_get_relative_msec();
    if (now_msec < 0)
    {
        sw_free(tnode);
        return NULL;
    }

    tnode->data = data;
    tnode->type = SW_TIMER_TYPE_KERNEL;
    tnode->exec_msec = now_msec + _msec;
    tnode->interval = interval ? _msec : 0;
    tnode->remove = 0;
    tnode->callback = callback;

    if (timer->_next_msec < 0 || timer->_next_msec > _msec)
    {
        timer->set(timer, _msec);
        timer->_next_msec = _msec;
    }

    tnode->id = timer->_next_id++;
    if (unlikely(tnode->id < 0))
    {
        tnode->id = 1;
        timer->_next_id = 2;
    }
    timer->num++;

    tnode->heap_node = swHeap_push(timer->heap, tnode->exec_msec, tnode);
    if (tnode->heap_node == NULL)
    {
        sw_free(tnode);
        return NULL;
    }
    swHashMap_add_int(timer->map, tnode->id, tnode);
    return tnode;
}

static int swSystemTimer_set(swTimer *timer, long new_interval)
{
    if (new_interval == current_interval)
    {
        return SW_OK;
    }
    current_interval = new_interval;
    if (SwooleG.use_timerfd)
    {
        return swSystemTimer_timerfd_set(timer, new_interval);
    }
    else
    {
        return swSystemTimer_signal_set(timer, new_interval);
    }
}

swTimer_del 刪除元素

int swTimer_del(swTimer *timer, swTimer_node *tnode)
{
    if (tnode->remove)
    {
        return SW_FALSE;
    }
    if (SwooleG.timer._current_id > 0 && tnode->id == SwooleG.timer._current_id)
    {
        tnode->remove = 1;
        return SW_TRUE;
    }
    if (swHashMap_del_int(timer->map, tnode->id) < 0)
    {
        return SW_ERR;
    }
    if (tnode->heap_node)
    {
        //remove from min-heap
        swHeap_remove(timer->heap, tnode->heap_node);
        sw_free(tnode->heap_node);
    }
    sw_free(tnode);
    timer->num --;
    return SW_TRUE;
}

swTimer_select 篩選定時函數

• swTimer_select 函數的篩選原理是從 timer->heap 中不斷 pop 出定時元素，比較它們的 exec_msec 是否超過了當前時間，若是超過了時間，就執行對應的定時函數；若是沒有超過，因爲 timer->heap 是排序事後的數據堆，所以當前定時元素以後的都不會超過當前時間，也就是尚未到執行的時間。
• 若是當前的定時元素超過了當前時間，說明該元素應該執行定時函數。設置 timer->_current_id 爲當前的 id 後，執行 tnode->callback 回調函數；若是當前定時元素不是一次執行的任務，而是須要每隔一段時間定時的任務，就要再次將元素放入 timer->heap 中；若是當前定時元素是一次執行的任務，就要將元素從 timer->map、timer->map 中刪除
• 循環結束後，tnode 就是下一個要執行的定時元素，咱們須要調用 timer->set 函數設置鬧鐘信號（worker 進程）或者 reactor 超時時間（master 進程）。

int swTimer_select(swTimer *timer)
{
    int64_t now_msec = swTimer_get_relative_msec();
    if (now_msec < 0)
    {
        return SW_ERR;
    }

    swTimer_node *tnode = NULL;
    swHeap_node *tmp;
    long timer_id;

    while ((tmp = swHeap_top(timer->heap)))
    {
        tnode = tmp->data;
        if (tnode->exec_msec > now_msec)
        {
            break;
        }

        timer_id = timer->_current_id = tnode->id;
        if (!tnode->remove)
        {
            tnode->callback(timer, tnode);
        }
        timer->_current_id = -1;

        //persistent timer
        if (tnode->interval > 0 && !tnode->remove)
        {
            while (tnode->exec_msec <= now_msec)
            {
                tnode->exec_msec += tnode->interval;
            }
            swHeap_change_priority(timer->heap, tnode->exec_msec, tmp);
            continue;
        }

        timer->num--;
        swHeap_pop(timer->heap);
        swHashMap_del_int(timer->map, timer_id);
        sw_free(tnode);
    }

    if (!tnode || !tmp)
    {
        timer->_next_msec = -1;
        timer->set(timer, -1);
    }
    else
    {
        timer->set(timer, tnode->exec_msec - now_msec);
    }
    return SW_OK;
}

Timer 定時器的使用

master 進程 swServer_start_proxy

timer 模塊在 master 進程中最重要的做用是每隔一秒更新 serv->gs->now 的值。除此以外，當 reactor 線程調度 worker 進程時，若是一段時間內沒有任何空閒的 worker 進程空閒，timer 模塊還負責寫入錯誤日誌。

static int swServer_start_proxy(swServer *serv)
{
    ...
    if (swTimer_init(1000) < 0)
    {
        return SW_ERR;
    }
    
    if (SwooleG.timer.add(&SwooleG.timer, 1000, 1, serv, swServer_master_onTimer) == NULL)
    {
        return SW_ERR;
    }
    ...
}

void swServer_master_onTimer(swTimer *timer, swTimer_node *tnode)
{
    swServer *serv = (swServer *) tnode->data;
    swServer_update_time(serv);
    if (serv->scheduler_warning && serv->warning_time < serv->gs->now)
    {
        serv->scheduler_warning = 0;
        serv->warning_time = serv->gs->now;
        swoole_error_log(SW_LOG_WARNING, SW_ERROR_SERVER_NO_IDLE_WORKER, "No idle worker is available.");
    }

    if (serv->hooks[SW_SERVER_HOOK_MASTER_TIMER])
    {
        swServer_call_hook(serv, SW_SERVER_HOOK_MASTER_TIMER, serv);
    }
}

void swServer_update_time(swServer *serv)
{
    time_t now = time(NULL);
    if (now < 0)
    {
        swWarn("get time failed. Error: %s[%d]", strerror(errno), errno);
    }
    else
    {
        serv->gs->now = now;
    }
}

worker 進程超時中止

worker 進程將要中止時，並不會馬上中止，而是會等待事件循環結束後中止，這時爲了防止 worker 進程不退出，還設置了 30s 的延遲，超過 30s 就會中止該進程。

static void swWorker_stop()
{
    swWorker *worker = SwooleWG.worker;
    swServer *serv = SwooleG.serv;
    worker->status = SW_WORKER_BUSY;
    
    ...

    try_to_exit: SwooleWG.wait_exit = 1;
    if (SwooleG.timer.fd == 0)
    {
        swTimer_init(serv->max_wait_time * 1000);
    }
    SwooleG.timer.add(&SwooleG.timer, serv->max_wait_time * 1000, 0, NULL, swWorker_onTimeout);

    swWorker_try_to_exit();
}

static void swWorker_onTimeout(swTimer *timer, swTimer_node *tnode)
{
    SwooleG.running = 0;
    SwooleG.main_reactor->running = 0;
    swoole_error_log(SW_LOG_WARNING, SW_ERROR_SERVER_WORKER_EXIT_TIMEOUT, "worker exit timeout, forced to terminate.");
}

swoole_timer_tick 添加定時任務

timer 模塊另外一個很是重要的功能是添加定時任務，通常是使用 swoole_timer_tick 函數、swoole_timer_after 函數、swoole_server->tick 函數、swoole_server->after 函數：

PHP_FUNCTION(swoole_timer_tick)
{
    long after_ms;
    zval *callback;
    zval *param = NULL;

    if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "lz|z", &after_ms, &callback, &param) == FAILURE)
    {
        return;
    }

    long timer_id = php_swoole_add_timer(after_ms, callback, param, 1 TSRMLS_CC);
    if (timer_id < 0)
    {
        RETURN_FALSE;
    }
    else
    {
        RETURN_LONG(timer_id);
    }
}

PHP_FUNCTION(swoole_timer_after)
{
    long after_ms;
    zval *callback;
    zval *param = NULL;

    if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "lz|z", &after_ms, &callback, &param) == FAILURE)
    {
        return;
    }

    long timer_id = php_swoole_add_timer(after_ms, callback, param, 0 TSRMLS_CC);
    if (timer_id < 0)
    {
        RETURN_FALSE;
    }
    else
    {
        RETURN_LONG(timer_id);
    }
}

php_swoole_add_timer 函數

本函數主要調用 SwooleG.timer.add 函數將添加新的定時任務，值得注意的是 swTimer_callback 類型的對象 cb 和兩個回調函數 php_swoole_onInterval、php_swoole_onTimeout，真正的回調函數存放在了 swTimer_callback 對象中，若是用戶有參數設置，也會放入 cb->data 中。

long php_swoole_add_timer(int ms, zval *callback, zval *param, int persistent TSRMLS_DC)
{
    char *func_name = NULL;

    if (!swIsTaskWorker())
    {
        php_swoole_check_reactor();
    }

    php_swoole_check_timer(ms);
    swTimer_callback *cb = emalloc(sizeof(swTimer_callback));

    cb->data = &cb->_data;
    cb->callback = &cb->_callback;
    memcpy(cb->callback, callback, sizeof(zval));
    if (param)
    {
        memcpy(cb->data, param, sizeof(zval));
    }
    else
    {
        cb->data = NULL;
    }

    swTimerCallback timer_func;
    if (persistent)
    {
        cb->type = SW_TIMER_TICK;
        timer_func = php_swoole_onInterval;
    }
    else
    {
        cb->type = SW_TIMER_AFTER;
        timer_func = php_swoole_onTimeout;
    }

    sw_zval_add_ref(&cb->callback);
    if (cb->data)
    {
        sw_zval_add_ref(&cb->data);
    }

    swTimer_node *tnode = SwooleG.timer.add(&SwooleG.timer, ms, persistent, cb, timer_func);
    {
        tnode->type = SW_TIMER_TYPE_PHP;
        return tnode->id;
    }
}

void php_swoole_check_timer(int msec)
{
    if (unlikely(SwooleG.timer.fd == 0))
    {
        swTimer_init(msec);
    }
}

php_swoole_onInterval 函數

本函數主要調用 cb->callback，若是有用戶參數，還要將 cb->data 放入調用函數中。

void php_swoole_onInterval(swTimer *timer, swTimer_node *tnode)
{
    zval *retval = NULL;
    int argc = 1;

    zval *ztimer_id;

    swTimer_callback *cb = tnode->data;

    SW_MAKE_STD_ZVAL(ztimer_id);
    ZVAL_LONG(ztimer_id, tnode->id);

    {
        zval **args[2];
        if (cb->data)
        {
            argc = 2;
            sw_zval_add_ref(&cb->data);
            args[1] = &cb->data;
        }
        args[0] = &ztimer_id;

        if (sw_call_user_function_ex(EG(function_table), NULL, cb->callback, &retval, argc, args, 0, NULL TSRMLS_CC) == FAILURE)
        {
            swoole_php_fatal_error(E_WARNING, "swoole_timer: onTimerCallback handler error");
            return;
        }
    }

    if (tnode->remove)
    {
        php_swoole_del_timer(tnode TSRMLS_CC);
    }
}

php_swoole_onTimeout 函數

與上一個函數相似，只是此次直接從 timer 中刪除對應的元素。

void php_swoole_onTimeout(swTimer *timer, swTimer_node *tnode)
{
    {
        swTimer_callback *cb = tnode->data;
        zval *retval = NULL;

        {
            zval **args[2];
            int argc;

            if (NULL == cb->data)
            {
                argc = 0;
                args[0] = NULL;
            }
            else
            {
                argc = 1;
                args[0] = &cb->data;
            }

            if (sw_call_user_function_ex(EG(function_table), NULL, cb->callback, &retval, argc, args, 0, NULL TSRMLS_CC) == FAILURE)
            {
                swoole_php_fatal_error(E_WARNING, "swoole_timer: onTimeout handler error");
                return;
            }
        }

        php_swoole_del_timer(tnode TSRMLS_CC);
    }
}

Timer 模塊時間輪算法

時間輪算法是各大網絡模塊採用的剔除空閒鏈接的方法，原理是構建一個首尾相連的循環數組，每隔數組元素中有若干個鏈接。若是某個鏈接有數據發送過來，將鏈接從所在的數組元素中刪除，將鏈接放入最新的數組元素中，這樣有數據來往的鏈接會一直在新數組元素中，空閒的鏈接所在的數組元素漸漸的變成了舊數組元素。每隔一段時間就按順序清空舊數組元素的所有鏈接。

swTimeWheel_new 建立時間輪

時間輪的數據結構比較簡單，由哈希表、size（循環數組總數量），current （循環數組當前最舊的數組元素，current-1 是循環數組中最新的數組元素）。swTimeWheel_new 函數很簡單，就是建立這三個屬性。

typedef struct
{
    uint16_t current;
    uint16_t size;
    swHashMap **wheel;

} swTimeWheel;

swTimeWheel* swTimeWheel_new(uint16_t size)
{
    swTimeWheel *tw = sw_malloc(sizeof(swTimeWheel));
    if (!tw)
    {
        swWarn("malloc(%ld) failed.", sizeof(swTimeWheel));
        return NULL;
    }

    tw->size = size;
    tw->current = 0;
    tw->wheel = sw_calloc(size, sizeof(void*));
    if (tw->wheel == NULL)
    {
        swWarn("malloc(%ld) failed.", sizeof(void*) * size);
        sw_free(tw);
        return NULL;
    }

    int i;
    for (i = 0; i < size; i++)
    {
        tw->wheel[i] = swHashMap_new(16, NULL);
        if (tw->wheel[i] == NULL)
        {
            swTimeWheel_free(tw);
            return NULL;
        }
    }
    return tw;
}

swTimeWheel_add 添加鏈接

當 main_reactor 有新鏈接進入的時候，須要將新的鏈接添加到時間輪中，新的鏈接會被放到最新的數組元素中，也就是 current-1 的元素中，而後設置 swConnection 中的 timewheel_index。

void swTimeWheel_add(swTimeWheel *tw, swConnection *conn)
{
    uint16_t index = tw->current == 0 ? tw->size - 1 : tw->current - 1;
    swHashMap *new_set = tw->wheel[index];
    swHashMap_add_int(new_set, conn->fd, conn);

    conn->timewheel_index = index;

    swTraceLog(SW_TRACE_REACTOR, "current=%d, fd=%d, index=%d.", tw->current, conn->fd, index);
}

swTimeWheel_update 函數

當鏈接有數據傳輸的時候，須要更新該鏈接在時間輪中的位置，將該鏈接從原有的數組元素中刪除，而後添加到最新的數組元素中，也就是 current-1 中，而後更新 swConnection 中的 timewheel_index。

#define swTimeWheel_new_index(tw)   (tw->current == 0 ? tw->size - 1 : tw->current - 1)

void swTimeWheel_update(swTimeWheel *tw, swConnection *conn)
{
    uint16_t new_index = swTimeWheel_new_index(tw);
    swHashMap *new_set = tw->wheel[new_index];
    swHashMap_add_int(new_set, conn->fd, conn);

    swHashMap *old_set = tw->wheel[conn->timewheel_index];
    swHashMap_del_int(old_set, conn->fd);

    swTraceLog(SW_TRACE_REACTOR, "current=%d, fd=%d, old_index=%d, new_index=%d.", tw->current, conn->fd, new_index, conn->timewheel_index);

    conn->timewheel_index = new_index;
}

swTimeWheel_remove 函數

在時間輪中刪除該鏈接，

void swTimeWheel_remove(swTimeWheel *tw, swConnection *conn)
{
    swHashMap *set = tw->wheel[conn->timewheel_index];
    swHashMap_del_int(set, conn->fd);
    swTraceLog(SW_TRACE_REACTOR, "current=%d, fd=%d.", tw->current, conn->fd);
}

swTimeWheel_forward 刪除空閒鏈接

swTimeWheel_forward 將最舊的數組元素 current 中全部鏈接都關閉掉，而後將 current 遞增。

void swTimeWheel_forward(swTimeWheel *tw, swReactor *reactor)
{
    swHashMap *set = tw->wheel[tw->current];
    tw->current = tw->current == tw->size - 1 ? 0 : tw->current + 1;

    swTraceLog(SW_TRACE_REACTOR, "current=%d.", tw->current);

    swConnection *conn;
    uint64_t fd;

    while (1)
    {
        conn = swHashMap_each_int(set, &fd);
        if (conn == NULL)
        {
            break;
        }

        conn->close_force = 1;
        conn->close_notify = 1;
        conn->close_wait = 1;
        conn->close_actively = 1;

        //notify to reactor thread
        if (conn->removed)
        {
            reactor->close(reactor, (int) fd);
        }
        else
        {
            reactor->set(reactor, fd, SW_FD_TCP | SW_EVENT_WRITE);
        }
    }
}

reactor 線程中時間輪的建立

• 時間輪的建立在 reactor 線程進行事件循環以前，按照用戶設置的鏈接最大空閒時間設置不一樣大小的時間輪，值得注意的是，時間輪最大是 SW_TIMEWHEEL_SIZE，也就是循環數組大小最大是 60。若是超過 60s 空閒時間，也僅僅創建 60 個元素的數組，可是這樣會形成每一個數組元素存放更多的鏈接。
• 值得注意的是，當容許空閒時間超過 60s 時，heartbeat_interval * 1000 是 reactor 的超時時間，例如空閒時間是 60s，那麼每隔 6s，reactor 都會超時來檢測空閒鏈接。當容許空閒時間小於 60s 時，reactor 統一每隔 1s 檢測空閒鏈接。
• 不一樣於 master 進程和 worker 線程，reactor 的 onFinish 和 onTimeout 再也不採用默認的 swReactor_onTimeout 與 swReactor_onFinish 函數，而是採用空閒鏈接檢測的 swReactorThread_onReactorCompleted 函數，該函數會調用 swTimeWheel_forward 來剔除空閒鏈接。

#define SW_TIMEWHEEL_SIZE          60

static int swReactorThread_loop(swThreadParam *param)
{
    ...
    
    if (serv->heartbeat_idle_time > 0)
    {
        if (serv->heartbeat_idle_time < SW_TIMEWHEEL_SIZE)
        {
            reactor->timewheel = swTimeWheel_new(serv->heartbeat_idle_time);
            reactor->heartbeat_interval = 1;
        }
        else
        {
            reactor->timewheel = swTimeWheel_new(SW_TIMEWHEEL_SIZE);
            reactor->heartbeat_interval = serv->heartbeat_idle_time / SW_TIMEWHEEL_SIZE;
        }
        reactor->last_heartbeat_time = 0;
        if (reactor->timewheel == NULL)
        {
            swSysError("thread->timewheel create failed.");
            return SW_ERR;
        }
        reactor->timeout_msec = reactor->heartbeat_interval * 1000;
        reactor->onFinish = swReactorThread_onReactorCompleted;
        reactor->onTimeout = swReactorThread_onReactorCompleted;
    }
    
    reactor->wait(reactor, NULL);
}

reactor 線程中時間輪的添加

當有新鏈接的時候，conn->connect_notify 會被置爲 1，此時該鏈接文件描述符寫就緒，而後就會調用 swReactorThread_onWrite，此時 reactor 線程將該鏈接添加到時間輪中。

static int swReactorThread_onWrite(swReactor *reactor, swEvent *ev)
{
    ...
    
    if (conn->connect_notify)
    {
        conn->connect_notify = 0;
        if (reactor->timewheel)
        {
            swTimeWheel_add(reactor->timewheel, conn);
        }
        
        ...
    }
    ...
}

reactor 線程中時間輪的更新

static int swReactorThread_onRead(swReactor *reactor, swEvent *event)
{
    ...
    if (reactor->timewheel && swTimeWheel_new_index(reactor->timewheel) != event->socket->timewheel_index)
    {
        swTimeWheel_update(reactor->timewheel, event->socket);
    }
    ...
}

reactor 線程中時間輪的剔除

當鏈接在容許的空閒時間以內沒有任何數據發送，那麼時間輪算法就要關閉該鏈接。關閉鏈接並非直接 close 套接字，而是須要通知對應的 worker 進程調用 onClose 函數，而後才能關閉。具體的作法是設置 swConnection 的 close_force、close_notify 等成員變量爲 1，而且關閉該鏈接的讀就緒監聽事件。

static void swReactorThread_onReactorCompleted(swReactor *reactor)
{
    swServer *serv = reactor->ptr;
    if (reactor->heartbeat_interval > 0 && reactor->last_heartbeat_time < serv->gs->now - reactor->heartbeat_interval)
    {
        swTimeWheel_forward(reactor->timewheel, reactor);
        reactor->last_heartbeat_time = serv->gs->now;
    }
}

void swTimeWheel_forward(swTimeWheel *tw, swReactor *reactor)
{
    ...
    
    conn->close_force = 1;
    conn->close_notify = 1;
    conn->close_wait = 1;
    conn->close_actively = 1;
    
    if (conn->removed)
    {
        reactor->close(reactor, (int) fd);
    }
    else
    {
        reactor->set(reactor, fd, SW_FD_TCP | SW_EVENT_WRITE);
    }
    ...
}

當該鏈接寫就緒的時候，會調用 swReactorThread_onWrite 函數。這個時候就會調用 swServer_tcp_notify 函數，進而調用 swFactoryProcess_notify、swFactoryProcess_dispatch，最後調用 swReactorThread_send2worker 發送給了 worker 進程。

因爲 reactor 啓用的是水平觸發，因爲並未向該鏈接寫入數據，所以很快又會觸發寫就緒事件調用 swReactorThread_onWrite 函數，這時若是 disable_notify 爲 1（dispatch_mode 爲 1 或 3），會直接執行 swReactorThread_close 函數關閉鏈接，假如此時 conn->out_buffer 中還有數據未發送，也會被拋棄。若是 disable_notify 爲 0，則會繼續向將要關閉的鏈接發送數據，直到接收到 SW_CHUNK_CLOSE 類型的消息。

static int swReactorThread_onWrite(swReactor *reactor, swEvent *ev)
{
    ...
    else if (conn->close_notify)
    {
        swServer_tcp_notify(serv, conn, SW_EVENT_CLOSE);
        conn->close_notify = 0;
        return SW_OK;
    }
    else if (serv->disable_notify && conn->close_force)
    {
        return swReactorThread_close(reactor, fd);
    }
    ...
}

int swServer_tcp_notify(swServer *serv, swConnection *conn, int event)
{
    swDataHead notify_event;
    notify_event.type = event;
    notify_event.from_id = conn->from_id;
    notify_event.fd = conn->fd;
    notify_event.from_fd = conn->from_fd;
    notify_event.len = 0;
    return serv->factory.notify(&serv->factory, &notify_event);
}

static int swFactoryProcess_notify(swFactory *factory, swDataHead *ev)
{
    memcpy(&sw_notify_data._send, ev, sizeof(swDataHead));
    sw_notify_data._send.len = 0;
    sw_notify_data.target_worker_id = -1;
    return factory->dispatch(factory, (swDispatchData *) &sw_notify_data);
}

static int swFactoryProcess_dispatch(swFactory *factory, swDispatchData *task)
{ 
   ...
   
   if (swEventData_is_stream(task->data.info.type))
   {
       swConnection *conn = swServer_connection_get(serv, fd);
       if (conn->closed)
        {
            //Connection has been clsoed by server
            if (!(task->data.info.type == SW_EVENT_CLOSE && conn->close_force))
            {
                return SW_OK;
            }
        }
        //converted fd to session_id
        task->data.info.fd = conn->session_id;
        task->data.info.from_fd = conn->from_fd;
   }
   
   return swReactorThread_send2worker((void *) &(task->data), send_len, target_worker_id);
}

worker 進程收到消息後會調用 swWorker_onTask 函數，進而調用 swFactoryProcess_end 函數，調用 serv->onClose 函數，並設置 swConnection 對象的 closed 爲 1，而後調用 swFactoryProcess_finish 函數將數據包發送給 reactor 線程。

int swWorker_onTask(swFactory *factory, swEventData *task)
{
    switch (task->info.type)
    {
        ... 
        factory->end(factory, task->info.fd);
        break;
        ...
    }
}

static int swFactoryProcess_end(swFactory *factory, int fd)
{
    bzero(&_send, sizeof(_send));
    _send.info.fd = fd;
    _send.info.len = 0;
    _send.info.type = SW_EVENT_CLOSE;
   
   swConnection *conn = swWorker_get_connection(serv, fd);
   
   if (conn->close_force)
   {
       goto do_close;
   }
   else if (conn->closing)
   {
       swoole_error_log(SW_LOG_NOTICE, SW_ERROR_SESSION_CLOSING, "The connection[%d] is closing.", fd);
       return SW_ERR;
   }
   else if (conn->closed)
   {
       return SW_ERR;
   }
   else
   {
        do_close:
        conn->closing = 1;
        if (serv->onClose != NULL)
        {
            info.fd = fd;
            if (conn->close_actively)
            {
                info.from_id = -1;
            }
            else
            {
                info.from_id = conn->from_id;
            }
            info.from_fd = conn->from_fd;
            serv->onClose(serv, &info);
        }
        conn->closing = 0;
        conn->closed = 1;
        conn->close_errno = 0;
        return factory->finish(factory, &_send);
   }

}

reactor 經過 swReactorThread_onPipeReceive 收到 worker 進程的鏈接關閉通知後，調用 swReactorThread_send 函數。若是鏈接已經被關閉，或者緩衝區中沒有任何數據的時候，直接調用 reactor->close 函數，也就是 swReactorThread_close 函數；若是緩衝區還有數據，那麼須要將消息放到 conn->out_buffer 中等待着該鏈接寫就緒回調 swReactorThread_close 函數（此時 close_notify 已經爲 0）。

int swReactorThread_send(swSendData *_send)
{
    ...
    if (_send->info.type == SW_EVENT_CLOSE && (conn->close_reset || conn->removed))
    {
        goto close_fd;
    }
    
    ...
    if (swBuffer_empty(conn->out_buffer))
    {
        if (_send->info.type == SW_EVENT_CLOSE)
        {
            close_fd:
            reactor->close(reactor, fd);
            return SW_OK;
        }
    }
    
    swBuffer_chunk *chunk;
    //close connection
    if (_send->info.type == SW_EVENT_CLOSE)
    {
        chunk = swBuffer_new_chunk(conn->out_buffer, SW_CHUNK_CLOSE, 0);
        chunk->store.data.val1 = _send->info.type;
    }
    
    if (reactor->set(reactor, fd, SW_EVENT_TCP | SW_EVENT_WRITE | SW_EVENT_READ) < 0
            && (errno == EBADF || errno == ENOENT))
    {
        goto close_fd;
    }

    ...
    close_fd:
        reactor->close(reactor, fd);
        return SW_OK;
}

static int swReactorThread_onWrite(swReactor *reactor, swEvent *ev)
{
    ...
    else if (conn->close_notify)
    {
        swServer_tcp_notify(serv, conn, SW_EVENT_CLOSE);
        conn->close_notify = 0;
        return SW_OK;
    }
    else if (serv->disable_notify && conn->close_force)
    {
        return swReactorThread_close(reactor, fd);
    }

    _pop_chunk: while (!swBuffer_empty(conn->out_buffer))
    {
        chunk = swBuffer_get_chunk(conn->out_buffer);
        if (chunk->type == SW_CHUNK_CLOSE)
        {
            close_fd: reactor->close(reactor, fd);
            return SW_OK;
        }
        ...
    }
    ...
}

swReactorThread_close 函數會刪除 swConnection 在 server 中的全部痕跡，包括 reactor 中的監控，serv->stats 的成員變量，port->connection_num 遞減，從時間輪中刪除、session 中 fd 置空等等工做。並且，還要清空套接字緩存中的全部數據，直接向客戶端發送關閉請求。swReactor_close 函數釋放內存，關閉套接字文件描述符。

int swReactorThread_close(swReactor *reactor, int fd)
{
    swServer *serv = SwooleG.serv;

    if (conn->removed == 0 && reactor->del(reactor, fd) < 0)
    {
        return SW_ERR;
    }

    sw_atomic_fetch_add(&serv->stats->close_count, 1);
    sw_atomic_fetch_sub(&serv->stats->connection_num, 1);

    swTrace("Close Event.fd=%d|from=%d", fd, reactor->id);

    //free the receive memory buffer
    swServer_free_buffer(serv, fd);

    swListenPort *port = swServer_get_port(serv, fd);
    sw_atomic_fetch_sub(&port->connection_num, 1);

#ifdef SW_USE_SOCKET_LINGER
    if (conn->close_force)
    {
        struct linger linger;
        linger.l_onoff = 1;
        linger.l_linger = 0;
        if (setsockopt(fd, SOL_SOCKET, SO_LINGER, &linger, sizeof(struct linger)) == -1)
        {
            swWarn("setsockopt(SO_LINGER) failed. Error: %s[%d]", strerror(errno), errno);
        }
    }
#endif

#ifdef SW_REACTOR_USE_SESSION
    swSession *session = swServer_get_session(serv, conn->session_id);
    session->fd = 0;
#endif

#ifdef SW_USE_TIMEWHEEL
    if (reactor->timewheel)
    {
        swTimeWheel_remove(reactor->timewheel, conn);
    }
#endif

    return swReactor_close(reactor, fd);
}

int swReactor_close(swReactor *reactor, int fd)
{
    swConnection *socket = swReactor_get(reactor, fd);
    if (socket->out_buffer)
    {
        swBuffer_free(socket->out_buffer);
    }
    if (socket->in_buffer)
    {
        swBuffer_free(socket->in_buffer);
    }
    if (socket->websocket_buffer)
    {
        swString_free(socket->websocket_buffer);
    }
    bzero(socket, sizeof(swConnection));
    socket->removed = 1;
    swTraceLog(SW_TRACE_CLOSE, "fd=%d.", fd);
    return close(fd);
}