Factory这个命名让我一度认为这是一个工厂模型……这个工厂实际上并不负责生产实例,而是根据类型的不同执行两项任务:Factory实现的功能是一个任务中心,一个task请求进入Factory,会进过dispatch分配、onTask处理、onFinish交付结果一系列流程;FactoryProcess用于管理manager和worker进程,也有对单独的writer线程的管理。 (PS:Swoole源码中有FactoryThread模块,该模块是一个多线程模型,根据开发者Rango韩少的解释,因为PHP不支持多线程,所以无法使用这个模块,因此该模块被废弃了。而实际上,FactoryThread比FactoryProcess要更简洁……)
Factory模块的相关声明在Server.h头文件中。首先是一个在Factory模块中被用到的结构体swDispatchData,该结构体声明在Server.h的145 – 149 行,声明如下:
typedef struct
{
long target_worker_id;
swEventData data;
} swDispatchData;
swDispatchData存放了一个目标worker进程的id和一条数据,该结构体用于传递数据给task进程进行处理。 Swoole中用swFactory结构体封装了Factory模块的相关操作,其声明在Server.h文件中的151 – 168 行,其声明如下:
struct _swFactory
{
void *object;
void *ptr; //server object
int last_from_id;
swReactor *reactor; //reserve for reactor
int (*start)(struct _swFactory *);
int (*shutdown)(struct _swFactory *);
int (*dispatch)(struct _swFactory *, swDispatchData *);
int (*finish)(struct _swFactory *, swSendData *);
int (*notify)(struct _swFactory *, swDataHead *); //send a event notify
int (*end)(struct _swFactory *, swDataHead *);
int (*onTask)(struct _swFactory *, swEventData *task); //worker function.get a task,goto to work
int (*onFinish)(struct _swFactory *, swSendData *result); //factory worker finish.callback
};
其中object用于存放一个具体的Factory类型(FactoryProcess or FactoryThread),last_from_id存放了最近一个通过该Factory发送消息的reactor的id。 swFactory的全部操作函数都声明在Server.h的176 – 183行,声明如下:
int swFactory_create(swFactory *factory);
int swFactory_start(swFactory *factory);
int swFactory_shutdown(swFactory *factory);
int swFactory_dispatch(swFactory *factory, swDispatchData *req);
int swFactory_finish(swFactory *factory, swSendData *_send);
int swFactory_notify(swFactory *factory, swDataHead *event);
int swFactory_end(swFactory *factory, swDataHead *cev);
int swFactory_check_callback(swFactory *factory);
这些函数在Factory.c中被定义。其中,start和shutdown函数简单返回SW_OK,create函数仅将传入的factory中的各个函数赋值(其实onTask和onFinish还是外部赋值……),callback函数仅仅检查onTask和onFinish两个函数指针是否为空,剩下的函数都是根据传入的参数调用对应的PHP回调函数(dispatch调用onTask,notify调用onClose和onConnect,end调用onClose并调用swServer_connection_close函数(详细分析见附录)关闭对应的connect连接) 这里需要分析一下swFactory_finish函数。swFactory_finish是一个通道,它的作用是将task运行结束后的数据发送给对应的Reactor。其核心源码如下:
//unix dgram
if (resp->info.type == SW_EVENT_UNIX_DGRAM)
{
socklen_t len;
struct sockaddr_un addr_un;
int from_sock = resp->info.from_fd;
addr_un.sun_family = AF_UNIX;
memcpy(addr_un.sun_path, resp->sun_path, resp->sun_path_len);
len = sizeof(addr_un);
ret = swSendto(from_sock, resp->data, resp->info.len, 0, (struct sockaddr *) &addr_un, len);
goto finish;
}
//UDP pacakge
else if (resp->info.type == SW_EVENT_UDP || resp->info.type == SW_EVENT_UDP6)
{
ret = swServer_udp_send(serv, resp);
goto finish;
}
else
{
resp->length = resp->info.len;
ret = swReactorThread_send(resp);
}
源码解释:这里首先判定swDataHead* resp应答中的type类型,如果type类型是Unix sock的数据包类型,则调用swSendto函数(详细分析见附录)发送数据到指定的Reactor;如果type类型是UDP,则调用swServer_udp_send(详细分析见附录)发送数据;否则,调用swReactorThread_send函数发送数据。
FactoryProcess模块 FactoryProcess模块是Swoole的进程管理模块,是Swoole另一个核心组件。通过该模块,Swoole能有效的调度和管理Master进程和多个Worker进程。 FactoryProcess的结构体swFactoryProcess声明在Server.h文件的170 – 174行,声明如下:
typedef struct _swFactoryProcess
{
swPipe *pipes;
int writer_pti; //current writer id
} swFactoryProcess;
第一个变量pipes存放用于进程间通信的管道,从后续实现分析这个Pipe仅用于接收其他进程发来的消息;第二个参数定义了当前writer线程的id,这个writer线程用于进程发送数据。 swFactoryProcess的四个公共操作函数声明在Server.h文件中的185 – 188行,其声明如下:
int swFactoryProcess_create(swFactory *factory, int writer_num, int worker_num);
int swFactoryProcess_start(swFactory *factory);
int swFactoryProcess_shutdown(swFactory *factory);
int swFactoryProcess_end(swFactory *factory, swDataHead *event);
这四个函数都在FactoryProcess.c文件中具体定义。其中,swFactoryProcess_create函数用于创建一个swFactoryProcess实例,其核心代码如下:
swFactoryProcess *object;
swServer *serv = SwooleG.serv;
object = SwooleG.memory_pool->alloc(SwooleG.memory_pool, sizeof(swFactoryProcess));
if (object == NULL)
{
swWarn("[Master] malloc[object] failed");
return SW_ERR;
}
serv->writer_threads = SwooleG.memory_pool->alloc(SwooleG.memory_pool, serv->reactor_num * sizeof(swWorkerThread));
if (serv->writer_threads == NULL)
{
swWarn("[Master] malloc[object->writers] failed");
return SW_ERR;
}
object->writer_pti = 0;
源码解释:从全局变量SwooleG的内存池中申请一个swFactoryProcess的内存,并同样从这个内存池中申请和reactor_num等量的writer线程,并初始化swFactoryProcess的writer_pti为0. swFactoryProcess_shutdown函数不是关闭一个swFactoryProcess,而是关闭SwooleG全局变量中的swServer实例。其核心源码如下:
swServer *serv = SwooleG.serv;
int i;
//kill manager process
kill(SwooleGS->manager_pid, SIGTERM);
//kill all child process
for (i = 0; i < serv->worker_num; i++)
{
swTrace("[Main]kill worker processor");
kill(serv->workers[i].pid, SIGTERM);
}
if (serv->ipc_mode == SW_IPC_MSGQUEUE)
{
serv->read_queue.free(&serv->read_queue);
serv->write_queue.free(&serv->write_queue);
}
//close pipes
return SW_OK;
源码解释:杀死manager进程,遍历swServer的worker列表并分别杀死每一个worker进程。如果使用了消息队列模式,则还需将read队列和write队列释放。 swFactoryProcess_start函数用于启动一个swFactoryProcess实例。这里需要注意,在该函数中涉及到一些没有分析过的函数和结构体,这些对象的实际分析将于后面的章节进行,在此我只简单说明这些函数的作用。swFactoryProcess_start函数核心源码如下:
if (swFactory_check_callback(factory) < 0)
{
swWarn("swFactory_check_callback failed");
return SW_ERR;
}
int i;
swServer *serv = factory->ptr;
swWorker *worker;
for (i = 0; i < serv->worker_num; i++)
{
worker = swServer_get_worker(serv, i);
if (swWorker_create(worker) < 0)
{
return SW_ERR;
}
}
//必须先启动manager进程组,否则会带线程fork
if (swFactoryProcess_manager_start(factory) < 0)
{
swWarn("swFactoryProcess_manager_start fail");
return SW_ERR;
}
if (serv->ipc_mode == SW_IPC_MSGQUEUE)
{
swQueueMsg_set_blocking(&serv->read_queue, 0);
//tcp & message queue require writer pthread
if (serv->have_tcp_sock == 1)
{
int ret = swFactoryProcess_writer_start(factory);
if (ret < 0)
{
return SW_ERR;
}
}
}
//主进程需要设置为直写模式
factory->finish = swFactory_finish;
源码解释:首先判断swFactory是否设置了相应的PHP回调函数(onTask,onFinish)。然后创建worker_num个worker对象,随后调用swFactoryProcess_manager_start函数启动manager进程和对应的worker子进程。如果swServer设置了消息队列属性,则设置swServer的read队列为非阻塞,并启动writer线程。最后设置manager主进程的finish函数。
swFactoryProcess_end函数用于关闭一个swFactoryProcess实例。其核心代码入下:
swServer *serv = factory->ptr;
swSendData _send;
bzero(&_send, sizeof(_send));
_send.info.fd = event->fd;
/**
* length == 0, close the connection
*/
_send.info.len = 0;
/**
* passive or initiative
*/
_send.info.type = event->type;
swConnection *conn = swServer_connection_get(serv, _send.info.fd);
if (conn == NULL || conn->active == 0)
{
swWarn("can not close. Connection[%d] not found.", _send.info.fd);
return SW_ERR;
}
if (serv->onClose != NULL)
{
serv->onClose(serv, event->fd, conn->from_id);
}
return swFactoryProcess_finish(factory, &_send);
源码解释:获取swServer对象,创建需要发送的swSendData并设置fd、长度(长度为0代表关闭连接)和类型。并获取swServer中fd对应的connection连接。随后调用onClose回调函数(如果有),并调用swFactoryProcess_finish函数发送数据。
除了以上四个共有的操作函数外,swFactoryProcess还拥有三个不同的操作函数,声明如下:
static int swFactoryProcess_notify(swFactory *factory, swDataHead *event);
static int swFactoryProcess_dispatch(swFactory *factory, swDispatchData *buf);
static int swFactoryProcess_finish(swFactory *factory, swSendData *data);
这三个函数是核心的进程通信函数,notify函数用于主进程通知worker进程,dispatch用于reactor向worker进程投递请求,finish用于worker将执行后的数据发送给client。 一个个来看,首先是swFactoryProcess_notify函数,其核心代码如下:
int swFactoryProcess_notify(swFactory *factory, swDataHead *ev)
{
memcpy(&sw_notify_data._send, ev, sizeof(swDataHead));
sw_notify_data._send.len = 0;
return factory->dispatch(factory, (swDispatchData *) &sw_notify_data);
}
源码解释:通过factory的dispatch函数将notify_data发送给worker进程。 其中sw_notify_data的声明如下:
static __thread struct
{
long target_worker_id;
swDataHead _send;
} sw_notify_data;
这个结构体是用来匹配swDispatchData结构体的,并没有特殊的意义。 swFactoryDespatch函数的调用来自于ReactorThread或者ReactorProcess(根据Rango的说法,所有线程相关的模块都已经停用了……),其核心源码如下:
uint32_t schedule_key;
uint32_t send_len = sizeof(task->data.info) + task->data.info.len;
uint16_t target_worker_id;
swServer *serv = SwooleG.serv;
if (task->target_worker_id < 0)
{
//udp use remote port
if (task->data.info.type == SW_EVENT_UDP || task->data.info.type == SW_EVENT_UDP6
|| task->data.info.type == SW_EVENT_UNIX_DGRAM)
{
schedule_key = task->data.info.from_id;
}
else
{
schedule_key = task->data.info.fd;
}
#ifndef SW_USE_RINGBUFFER
if (SwooleTG.factory_lock_target)
{
if (SwooleTG.factory_target_worker < 0)
{
target_worker_id = swServer_worker_schedule(serv, schedule_key);
SwooleTG.factory_target_worker = target_worker_id;
}
else
{
target_worker_id = SwooleTG.factory_target_worker;
}
}
else
#endif
{
target_worker_id = swServer_worker_schedule(serv, schedule_key);
}
}
else
{
target_worker_id = task->target_worker_id;
}
if (SwooleTG.type == SW_THREAD_REACTOR)
{
return swReactorThread_send2worker((void *) &(task->data), send_len, target_worker_id);
}
else
{
swTrace("dispatch to worker#%d", target_worker_id);
return swServer_send2worker_blocking(serv, (void *) &(task->data), send_len, target_worker_id);
}
源码解释:首先判定task是否已经指定了worker,如果task的target_worker_id小于0(代表没有指定),则需要为其分配worker进程。首先根据task的data域中info的type字段判断task是属于哪个fd,这个fd用于在随后的调用中判断如何分配worker进程。之后,判断是否指定使用RingBuffer,判断SwooleTG(Swoole Thread Global,线程相关全局变量,以__thread关键字修饰,每个线程拥有自己独立的一份,线程之间互不影响)是否指定了当前线程所有的worker_id,如果没有指定,调用swServer_worker_schedule函数分配worker;否则直接使用指定的worker。如果以上条件都未能满足,则直接使用swServer_worker_schedule函数分配worker。如果task已经指定了worker,则默认使用该worker。随后判断SwooleTG的type字段,如果为SW_THREAD_REACTOR(线程Reacotr模式),则调用swReactorThread_send2worker函数(非阻塞)发送task请求;否则(type为SW_PROCESS_REACTOR模式),调用swServer_send2worker_blocking函数(阻塞)发送请求。
这里解释一下这两个函数的阻塞和非阻塞的原因。swReactorThread_send2worker调用write函数如果失败,会将数据存放进pipe的缓存中,等待下一次发送;而swServer_send2worker_blocking调用write失败后,会重新发送或调用swSocket_wait等待一段时间后再次发送。
接下来是swFactoryProcess_finish函数,该函数用于将一个resp应答发送给task的发起者。由于该函数较长,在此分段进行分析:其源码如下:
//unix dgram
if (resp->info.type == SW_EVENT_UNIX_DGRAM)
{
socklen_t len;
struct sockaddr_un addr_un;
int from_sock = resp->info.from_fd;
addr_un.sun_family = AF_UNIX;
memcpy(addr_un.sun_path, resp->sun_path, resp->sun_path_len);
len = sizeof(addr_un);
ret = swSendto(from_sock, resp->data, resp->info.len, 0, (struct sockaddr *) &addr_un, len);
goto finish;
}
源码解释:如果判定resp的类型为unixsock的报文模式,则构建相应sockaddr_un结构体并调用swSendto函数发送resp(swSendto函数分析见附录)。发送完成后,goto到finish标签继续运行。
//UDP pacakge
else if (resp->info.type == SW_EVENT_UDP || resp->info.type == SW_EVENT_UDP6)
{
ret = swServer_udp_send(serv, resp);
goto finish;
}
源码解释:如果是UDP模式,调用swServer_udp_send函数发送数据并goto到finish标签。
struct
{
long pti;
swEventData _send;
} sdata;
//for message queue
sdata.pti = (SwooleWG.id % serv->writer_num) + 1;
swConnection *conn = swServer_connection_get(serv, fd);
if (conn == NULL || conn->active == 0)
{
swWarn("connection[%d] not found.", fd);
return SW_ERR;
}
sdata._send.info.fd = fd;
sdata._send.info.type = resp->info.type;
swWorker *worker = swServer_get_worker(serv, SwooleWG.id);
源码解释:构建了一个临时结构体用于匹配swQueue_data结构体,并指定对应的writer线程id。随后获取fd对应的connection连接和SwooleWG(Worker进程的全局变量)中存放的当前进程对应的worker_id。
if (resp->length > 0)
{
int64_t wait_reactor;
/**
* Storage is in use right now, wait notify.
*/
if (worker->store.lock == 1)
{
worker->notify->read(worker->notify, &wait_reactor, sizeof(wait_reactor));
}
swPackage_response response;
response.length = resp->length;
response.worker_id = SwooleWG.id;
//swWarn("BigPackage, length=%d|worker_id=%d", response.length, response.worker_id);
sdata._send.info.from_fd = SW_RESPONSE_BIG;
sdata._send.info.len = sizeof(response);
memcpy(sdata._send.data, &response, sizeof(response));
/**
* Lock the worker storage
*/
worker->store.lock = 1;
memcpy(worker->store.ptr, resp->data, resp->length);
}
源码解释:如果resp的长度大于0,表示这是个较大的应答包(完全不理解为什么大于0就代表big response了……),需要使用共享内存。首先判定worker的store是否已经被占用,如果被占用,则通过notify管道等待唤醒。当store可用后,用swPackage_response结构体封装数据信息并代替原有的data域,设置from_id为SW_RESPONSE_BIG,并将data复制到worker的store共享内存中并上锁。
#if SW_REACTOR_SCHEDULE == 2
sdata._send.info.from_id = fd % serv->reactor_num;
#else
sdata._send.info.from_id = conn->from_id;
#endif
源码解释:根据Reactor的Schedule模型确定from_id, 2代表的是取模。
for (count = 0; count < SW_WORKER_SENDTO_COUNT; count++)
{
if (serv->ipc_mode == SW_IPC_MSGQUEUE)
{
ret = serv->write_queue.in(&serv->write_queue, (swQueue_data *) &sdata, sendn);
}
else
{
int master_pipe = swWorker_get_write_pipe(serv, fd);
//swWarn("send to reactor. fd=%d|pipe=%d|reactor_id=%d|reactor_pipe_num=%d", fd, master_pipe, conn->from_id, serv->reactor_pipe_num);
ret = write(master_pipe, &sdata._send, sendn);
#ifdef SW_WORKER_WAIT_PIPE
if (ret < 0 && errno == EAGAIN)
{
/**
* Wait pipe can be written.
*/
if (swSocket_wait(master_pipe, SW_WORKER_WAIT_TIMEOUT, SW_EVENT_WRITE) == SW_OK)
{
continue;
}
else
{
goto finish;
}
}
#endif
}
//swTraceLog("wt_queue->in: fd=%d|from_id=%d|data=%s|ret=%d|errno=%d", sdata._send.info.fd, sdata._send.info.from_id, sdata._send.data, ret, errno);
if (ret >= 0)
{
break;
}
else if (errno == EINTR)
{
continue;
}
else if (errno == EAGAIN)
{
swYield();
}
else
{
break;
}
}
源码解释:循环写出数据直到写出成功,最大重写次数为SW_WORKER_SENDTO_COUNT(32)次。每次写出时,先判断是否为消息队列模式,如果是,将数据写入消息队列;如果不是,获取worker的write_pipe,并通过该管道将数据发送给master进程。
对于FactoryProcess的基本分析到这里就结束了,下一章将分析FactoryProcess.c文件中剩下的操作函数。
- 目录
- 上一节: 8.Reactor模块-epoll
- 下一节: 10.Factory模块(下)