Jacky Wu bio photo

Jacky Wu

Software Architect in China

Email Twitter Facebook LinkedIn Github Weibo

1. 前言

  1. 本代码分析基于 saltstack-2015.8 版本
  2. saltstack中大量使用了ZeroMQ,所以需要预先储备 ZeroMQ方面的知识
  3. 本文以运行 salt '*' test.ping 场景为例研究任务的处理机制。

2. 概念解释

Salt Client

运行在Salt Master上,给Master发送请求,并且得到结果返回的LocalClient类。

Salt Master有几个重要的类

  1. Publisher:给Minion发送指令
  2. EventPublisher:Master上的事件发布总线
  3. MWorkers:Salt Master的工作进程
  4. ReqServer:实现了一个MWorkers的多进程模型。接收Salt Client的请求或者Minion返回的结果,发送给MWorker去处理

3. 任务处理机制

  1. 在Master上运行salt '*' test.ping, 原理是通过SaltClient提交任务到ReqServer的TCP:4506端口,并且监听在EventPublisher上获取结果。
  2. ReqServer实现了一个MWorkers多进程模型。ReqServer收到请求后发给其中一个MWorker进行处理。
  3. MWorker做权限检查后将任务加密发送给Publisher,并且把这个事件发送给EventPublisher。
  4. 所有Minion都事先连接到了Publisher的TCP:4505端口上,获取到任务后解密处理。Minion处理完毕后将结果发送到Master的ReqServer的TCP:4506端口。
  5. ReqServer又把这个结果发给了其中一个MWorker进行处理。
  6. MWorker拿到结果后进行解密,然后发送到EventPublisher。
  7. 此时SaltClient监听在EventPublisher上等待获取结果,碰到结果来了或者超时,命令行返回。

3.1. 任务流程图

3.2. 提交任务

  • 入口 salt/scripts/salt.py
  • 调用 salt/salt/scripts.py: salt_main()
  • 调用 salt/salt/cli/salt.py: SaltCMD:run()
    1. 判断是否为异步任务,若是,jid = local.cmd_async(**kwargs)
    2. 若否,则进行同步执行
  • 同步执行
1
2
3
4
    #
    for full_ret in local.cmd_cli(**kwargs):
         ret_, out, retcode = self._format_ret(full_ret)
         ret.update(ret_)
  • local.cmd_cli调用 salt/salt/client/__init__.py: LocalClient:cmd_cli()
  • cmd_cli()调用pub_data = self.run_job() -> self.pub() 往ReqServer的ret_port:tcp4506端口提交任务请求。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
    #
    def pub():
        master_uri = 'tcp://' + salt.utils.ip_bracket(self.opts['interface']) + \
                     ':' + str(self.opts['ret_port'])
        channel = salt.transport.Channel.factory(self.opts,
                                                 crypt='clear',
                                                 master_uri=master_uri)

        try:
            payload = channel.send(payload_kwargs, timeout=timeout)
        except SaltReqTimeoutError:
            raise SaltReqTimeoutError(
                'Salt request timed out. The master is not responding. '
                'If this error persists after verifying the master is up, '
                'worker_threads may need to be increased.'
            )

这里的salt.transport.Channel.factory最终会被实例化为AsyncZeroMQReqChannel,而且实例化的参数中有crypt=’clear’, 这就意味着,LocalClient给Master发送的消息都是Clear的,也就是未AES加密的。

任务内容的组装在salt/client/__init__.pydef _prep_pub方法里。其中cmd: 'publish'会在接下来的流程里用到。

1
2
3
4
5
6
7
8
9
10
11
    #
    def _prep_pub():
    ...
        payload_kwargs = {'cmd': 'publish',
                          'tgt': tgt,
                          'fun': fun,
                          'arg': arg,
                          'key': self.key,
                          'tgt_type': expr_form,
                          'ret': ret,
                          'jid': jid}

最终在payload外包装上’enc’: ‘clear’是在AsyncZeroMQReqChannel 里的send方法实现的,然后调用_crypted_transfer函数里加上。

1
2
3
4
5
6
7
8
9
10
11
    #
     @tornado.gen.coroutine
        def send(self, load, tries=3, timeout=60):
            '''
            Send a request, return a future which will complete when we send the message
            '''
            if self.crypt == 'clear':
                ret = yield self._uncrypted_transfer(load, tries=tries, timeout=timeout)
            else:
                ret = yield self._crypted_transfer(load, tries=tries, timeout=timeout)
            raise tornado.gen.Return(ret)

最终提交到4506端口的数据变成了类似这样的:

    #
    {'enc': 'clear',
     'load': {'arg': [],
              'cmd': 'publish',
              'fun': 'test.ping',
              'jid': '',
              'key': 'alsdkjfa.,maljf-==adflkjadflkjalkjadfadflkajdflkj',
              'kwargs': {'show_jid': False, 'show_timeout': False},
              'ret': '',
              'tgt': 'client.jacky.com',
              'tgt_type': 'glob',
              'user': 'sudo_vagrant'}}
  • 提交完任务后,cmd_cli()就调用 self.get_cli_event_returns()
  • 调用 self.get_iter_returns()
    1. 判断是否是MoM结构,若是ret_iter = self.get_returns_no_block('(salt/job|syndic/.*)/{0}'.format(jid), 'regex')
    2. 若否,ret_iter = self.get_returns_no_block('salt/job/{0}'.format(jid))
  • self.get_returns_no_block
1
2
3
4
    #
    while True:
        raw = self.event.get_event(wait=0.01, tag=tag, match_type=match_type, full=True, no_block=True)
        yield raw
  • 调用了salt/utils/event.py:SaltEvent:get_event(), get_event又调用了_get_event(), 在里面其以SUB的角色连接到了EventPublish监听任务结果的消息。

提交的任务的event内容是

1
2
3
    #
    [DEBUG   ] Sending event - data = {'tgt_type': 'glob', 'jid': '20151123220104386580', 'tgt': 'client.jacky.com', '_stamp'
    : '2015-11-23T14:01:04.386948', 'user': 'sudo_vagrant', 'arg': [], 'fun': 'test.ping', 'minions': ['client.jacky.com']}

这是salt命令后提交任务的debug输出

    #
     [root@master base]# salt 'client.jacky.com' test.ping -l debug
    [DEBUG   ] Reading configuration from /etc/salt/master
    [DEBUG   ] Missing configuration file: /root/.saltrc
    [DEBUG   ] Configuration file path: /etc/salt/master
    [WARNING ] Insecure logging configuration detected! Sensitive data may be logged.
    [DEBUG   ] Reading configuration from /etc/salt/master
    [DEBUG   ] Missing configuration file: /root/.saltrc
    [DEBUG   ] MasterEvent PUB socket URI: ipc:///var/run/salt/master/master_event_pub.ipc
    [DEBUG   ] MasterEvent PULL socket URI: ipc:///var/run/salt/master/master_event_pull.ipc
    [DEBUG   ] Initializing new AsyncZeroMQReqChannel for ('/etc/salt/pki/master', 'master.jacky.com_master', 'tcp://127.0.0.1:4506', 'clear')
    [DEBUG   ] LazyLoaded config.option
    [DEBUG   ] get_iter_returns for jid 20151124092621201574 sent to set(['client.jacky.com']) will timeout at 09:26:26.241828
    [DEBUG   ] jid 20151124092621201574 return from client.jacky.com

    client.jacky.com:
    	True
    [DEBUG   ] jid 20151124092621201574 found all minions set(['client.jacky.com'])

3.3. ReqServer实现的多进程模型

ReqServer的启动入口在salt/master.py里。

1
2
3
4
5
6
    #
    req_channels = []
    for transport, opts in iter_transport_opts(self.opts):
        chan = salt.transport.server.ReqServerChannel.factory(opts)
        chan.pre_fork(self.process_manager)
        req_channels.append(chan)

利用ReqServerChannel的工厂函数实例化ZeroMQ或Raet等其他协议的通信通道,具体的实现在salt/transport/zeromq.py里。 chan.pre_fork(self.process_manager) -> process_manager.add_process(self.zmq_device), 在zmq_device实现了Router+Dealer的模型,并且Client(Router)绑定了tcp://ip:4506端口,Worker(Dealer)根据ipc_mode的配置绑定了tcp://127.0.0.1:4515或者ipc://workers.ipc

1
2
3
4
5
6
7
8
    #
    while True:
        try:
            zmq.device(zmq.QUEUE, self.clients, self.workers)
        except zmq.ZMQError as exc:
            if exc.errno == errno.EINTR:
                continue
            raise exc

zmq.device对clients和worker之间的消息传递进行了代理。

1
2
3
4
5
6
7
8
9
10
    #
    for ind in range(int(self.opts['worker_threads'])):
        self.process_manager.add_process(MWorker,
                                         args=(self.opts,
                                               self.master_key,
                                               self.key,
                                               req_channels,
                                               ),
                                         )
    self.process_manager.run()

根据配置文件里worker_threads 的配置启动相应个数的MWorker。

关于ZeroMQReqServerChannel的实现在salt/transport/zeromq.py中。客户端会通过AsyncReqMessageClient往Router:4506端口提交请求。MWorker继承自multiprocessing.Process,每个MWorker被创建的时候都传入了req_channels ,在run()方法里调用了self.__bind(), 通过post_fork()创建了REP类型的连接到Dealer上,等待处理任务。

1
2
3
4
5
6
7
8
9
10
11
    #
    def __bind(self):
        '''
        Bind to the local port
        '''
        # using ZMQIOLoop since we *might* need zmq in there
        zmq.eventloop.ioloop.install()
        self.io_loop = zmq.eventloop.ioloop.ZMQIOLoop()
        for req_channel in self.req_channels:
            req_channel.post_fork(self._handle_payload, io_loop=self.io_loop)  # TODO: cleaner? Maybe lazily?
        self.io_loop.start()

至此,这样一个多进程模型就创建完毕。(该图来自ZeroMQ官网)

3.4. MWorker的处理逻辑

MWorker的启动入口在salt/master.py里。

程序入口是run()方法。设置了self.clear_funcsself.aes_funcs方法,并且调用__bind方法,通过post_fork()传入了self._handle_payload任务处理函数,并且创建了REP类型的连接到Dealer上,等待处理任务。

1
2
3
4
5
6
7
8
9
10
11
    
    def __bind(self):
        '''
        Bind to the local port
        '''
        # using ZMQIOLoop since we *might* need zmq in there
        zmq.eventloop.ioloop.install()
        self.io_loop = zmq.eventloop.ioloop.ZMQIOLoop()
        for req_channel in self.req_channels:
            req_channel.post_fork(self._handle_payload, io_loop=self.io_loop)  # TODO: cleaner? Maybe lazily?
        self.io_loop.start()

post_fork()里调用ZeroMQReqServerChannel或者其他ReqServerChannel的实现,以ZeroMQReqServerChannel的post_fork为例,处理任务的入口在这几行代码, 设置了任务处理函数 payload_handler。self.handle_message是MWorker在接收到消息时的一个回调函数,在里面调用了payload_handler去处理任务。

1
2
3
4
5
    
    salt.transport.mixins.auth.AESReqServerMixin.post_fork(self, payload_handler, io_loop)

    self.stream = zmq.eventloop.zmqstream.ZMQStream(self._socket, io_loop=self.io_loop)
    self.stream.on_recv_stream(self.handle_message)

通过_handle_payload方法,我们可以看到cmd: 'publish', 那么MWorker就会调用self._handle_clear.publish方法,将任务load发送到PubServerChannel的PULL接口。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
    
    @tornado.gen.coroutine
    def _handle_payload(self, payload):
        '''
        The _handle_payload method is the key method used to figure out what
        needs to be done with communication to the server

        Example cleartext payload generated for 'salt myminion test.ping':

        {'enc': 'clear',
         'load': {'arg': [],
                  'cmd': 'publish',
                  'fun': 'test.ping',
                  'jid': '',
                  'key': 'alsdkjfa.,maljf-==adflkjadflkjalkjadfadflkajdflkj',
                  'kwargs': {'show_jid': False, 'show_timeout': False},
                  'ret': '',
                  'tgt': 'myminion',
                  'tgt_type': 'glob',
                  'user': 'root'}}

        :param dict payload: The payload route to the appropriate handler
        '''
        key = payload['enc']
        load = payload['load']
        ret = {'aes': self._handle_aes,
               'clear': self._handle_clear}[key](load)
        raise tornado.gen.Return(ret)

    def _handle_clear(self, load):
        '''
        Process a cleartext command

        :param dict load: Cleartext payload
        :return: The result of passing the load to a function in ClearFuncs corresponding to
                 the command specified in the load's 'cmd' key.
        '''
        log.trace('Clear payload received with command {cmd}'.format(**load))
        if load['cmd'].startswith('__'):
            return False
        return getattr(self.clear_funcs, load['cmd'])(load), {'fun': 'send_clear'}

3.5. Publisher实现的Pub-Sub任务分发机制

Publisher的入口在salt/master.py里。

在Master的启动过程中start()函数里,创建了PubServerChannel,根据传入参数不同创建ZeroMQPubServerChannel或者其他协议的实现,具体的实现在salt/transport/zeromq.py里。这里以PubServerChannel为例。

1
2
3
4
5
    
    for transport, opts in iter_transport_opts(self.opts):
        chan = salt.transport.server.PubServerChannel.factory(opts)
        chan.pre_fork(process_manager)
        pub_channels.append(chan)
1
2
3
4
5
6
7
8
9
10
    
    def pre_fork(self, process_manager):
        '''
        Do anything necessary pre-fork. Since this is on the master side this will
        primarily be used to create IPC channels and create our daemon process to
        do the actual publishing

        :param func process_manager: A ProcessManager, from salt.utils.process.ProcessManager
        '''
        process_manager.add_process(self._publish_daemon)

实例化chan后,调用pre_fork方法,利用_publish_daemon方法启动新Daemon进程,监听了pull_uri, pub_uri,并且在while循环里将pull_uri接收到的任务转发到pub_uri里。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
    #
    def _publish_daemon(self):
    ...
       while True:
            # Catch and handle EINTR from when this process is sent
            # SIGUSR1 gracefully so we don't choke and die horribly
            try:
                package = pull_sock.recv()
                unpacked_package = salt.payload.unpackage(package)
                payload = unpacked_package['payload']
                if self.opts['zmq_filtering']:
                    # if you have a specific topic list, use that
                    if 'topic_lst' in unpacked_package:
                        for topic in unpacked_package['topic_lst']:
                            # zmq filters are substring match, hash the topic
                            # to avoid collisions
                            htopic = hashlib.sha1(topic).hexdigest()
                            pub_sock.send(htopic, flags=zmq.SNDMORE)
                            pub_sock.send(payload)
                            # otherwise its a broadcast
                    else:
                        # TODO: constants file for "broadcast"
                        pub_sock.send('broadcast', flags=zmq.SNDMORE)
                        pub_sock.send(payload)
                else:
                    pub_sock.send(payload)
            except zmq.ZMQError as exc:
                if exc.errno == errno.EINTR:
                    continue
                raise exc

MWorker将任务提交到Publisher是在self._handle_clear.publish方法里实现的。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
    #
    def publish():
    ...
         # Send it!
        self._send_pub(payload)
    ...

    def _send_pub(self, load):
        '''
        Take a load and send it across the network to connected minions
        '''
        for transport, opts in iter_transport_opts(self.opts):
            chan = salt.transport.server.PubServerChannel.factory(opts)
            chan.publish(load)

这里的chan就是ZeroMQPubServerChannel的实例,调用该实例的publish方法,将payload加密,然后发送给Publisher的pull_uri,即publish_pull.ipc端口。

1
2
3
4
5
6
7
8
9
10
11
12
13
    #
    payload = {'enc': 'aes'}

    crypticle = salt.crypt.Crypticle(self.opts, salt.master.SMaster.secrets['aes']['secret'].value)
    payload['load'] = crypticle.dumps(load)

    ...
    pub_sock.connect(pull_uri)

    ...

    pub_sock.send(self.serial.dumps(int_payload))
    ...

这是Master上的debug输出

    #
    [DEBUG   ] Sending event - data = {'_stamp': '2015-11-24T00:39:00.185467', 'minions': ['client.jacky.com']}
    [DEBUG   ] Sending event - data = {'tgt_type': 'glob', 'jid': '20151124083900185051', 'tgt': 'client.jacky.com', '_stamp': '2015-11-24T00:39:00.186134', 'user': 'sudo_vagrant', 'arg': [], 'fun': 'test.ping', 'minions': ['client.jacky.com']}
    [INFO    ] User sudo_vagrant Published command test.ping with jid 20151124083900185051
    [DEBUG   ] Published command details {'tgt_type': 'glob', 'jid': '20151124083900185051', 'tgt': 'client.jacky.com', 'ret': 'mysql', 'user': 'sudo_vagrant', 'arg': [], 'fun': 'test.ping'}
    [INFO    ] Got return from client.jacky.com for job 20151124083900185051
    [DEBUG   ] Sending event - data = {'fun_args': [], 'jid': '20151124083900185051', 'return': True, 'retcode': 0, 'success': True, 'cmd': '_return', '_stamp': '2015-11-24T00:39:00.303857', 'fun': 'test.ping', 'id': 'client.jacky.com'}

3.6. Minion连接到Publisher获取和处理任务

Minion在启动过程中最终会进入tune_in方法,然后陷入无限循坏,不断处理任务和返回结果。实现在salt/minion.py里。

  • tune_in调用self.connect_master
  • connect_master调用self.eval_master(self.opts, self.timeout, self.safe)
  • eval_master中利用AsyncPubChannel的factory方法,根据不同参数实例化相应的pub_channel,这里以AsyncZeroMQPubChannel为例,具体的实现在salt/transport/zeromq.py里。
1
2
3
4
5
6
7
8
9
10
11
    #
    try:
        pub_channel = salt.transport.client.AsyncPubChannel.factory(opts, **factory_kwargs)
        yield pub_channel.connect()
        conn = True
        break
    except SaltClientError:
        msg = ('Master {0} could not be reached, trying '
               'next master (if any)'.format(opts['master']))
        log.info(msg)
        continue
  • 调用AsyncZeroMQPubChannel的connect方法,以SUB的角色连接到Master的PUB端口4505。
1
2
3
4
5
6
7
8
        #
        def connect(self):
            if not self.auth.authenticated:
                yield self.auth.authenticate()
            self.publish_port = self.auth.creds['publish_port']
            self._socket.connect(self.master_pub)

    
  • tune_in方法里,连接到Master之后,设置任务处理的回调函数
1
2
3
    #
    # add handler to subscriber
    self.pub_channel.on_recv(self._handle_payload)
  • pub_channel.on_recv其实就是AsyncZeroMQPubChannel.on_recv()方法,封装了对收到的message进行解密。
1
2
3
4
5
6
7
     #
     def on_recv():
             @tornado.gen.coroutine
            def wrap_callback(messages):
                payload = yield self._decode_messages(messages)
                if payload is not None:
                    callback(payload)

self._decode_messages()调用了self._decode_payload()进行解密。见salt/transport/mixins/auth.pyclass AESPubClientMixin(object)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
    #
    @tornado.gen.coroutine
    def _decode_payload(self, payload):
        # we need to decrypt it
        log.trace('Decoding payload: {0}'.format(payload))
        if payload['enc'] == 'aes':
            self._verify_master_signature(payload)
            try:
                payload['load'] = self.auth.crypticle.loads(payload['load'])
            except salt.crypt.AuthenticationError:
                yield self.auth.authenticate()
                payload['load'] = self.auth.crypticle.loads(payload['load'])

        raise tornado.gen.Return(payload)
  • self._handle_payload调用了_handle_decoded_payload,根据data['fun']的类型初始化target,接下来以target创建新的进程或者线程去处理这个任务(根据master里multiprocessing参数来判断是创建新进程还是线程)。具体处理逻辑在_thread_return或者_thread_multi_return里,这里以_thread_return为例。
1
2
3
4
5
6
7
        #
        if isinstance(data['fun'], tuple) or isinstance(data['fun'], list):
            target = Minion._thread_multi_return
        else:
            target = Minion._thread_return

    
  • _thread_return中对任务进行处理,然后用minion_instance._return_pub返回结果给Master的4506端口。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
    #
    try:
        func = minion_instance.functions[data['fun']]
        args, kwargs = load_args_and_kwargs(
            func,
            data['arg'],
            data)
        minion_instance.functions.pack['__context__']['retcode'] = 0
        if opts.get('sudo_user', ''):
            sudo_runas = opts.get('sudo_user')
            if 'sudo.salt_call' in minion_instance.functions:
                return_data = minion_instance.functions['sudo.salt_call'](
                        sudo_runas,
                        data['fun'],
                        *args,
                        **kwargs)
        else:
            return_data = func(*args, **kwargs)
  • _return_pub方法中调用salt.transport.Channel的factory工厂方法实例化到master到连接,这里的channel就是AsyncZeroMQReqChannel,使用channel.send方法将结果发送。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
    #
    def _return_pub():
    ...
        channel = salt.transport.Channel.factory(self.opts)
    ...
            try:
            ret_val = channel.send(load, timeout=timeout)
        except SaltReqTimeoutError:
            msg = ('The minion failed to return the job information for job '
                   '{0}. This is often due to the master being shut down or '
                   'overloaded. If the master is running consider increasing '
                   'the worker_threads value.').format(jid)
            log.warn(msg)
            return ''

这里的AsyncZeroMQReqChannel.send默认会使用aes加密的方式对消息进行发送。

1
2
3
4
5
6
7
8
9
10
11
    #
     @tornado.gen.coroutine
        def send(self, load, tries=3, timeout=60):
            '''
            Send a request, return a future which will complete when we send the message
            '''
            if self.crypt == 'clear':
                ret = yield self._uncrypted_transfer(load, tries=tries, timeout=timeout)
            else:
                ret = yield self._crypted_transfer(load, tries=tries, timeout=timeout)
            raise tornado.gen.Return(ret)

这是Minion上的debug输出

1
2
3
4
5
6
7
8
9
    #
     [INFO    ] User sudo_vagrant Executing command test.ping with jid 20151124083900185051
    [DEBUG   ] Command details {'tgt_type': 'glob', 'jid': '20151124083900185051', 'tgt': 'client.jacky.com', 'ret': 'mysql', 'user': 'sudo_vagrant', 'arg': [], 'fun': 'test.ping'}
    [INFO    ] Starting a new job with PID 27556
    [DEBUG   ] LazyLoaded test.ping
    [DEBUG   ] Minion return retry timer set to 2 seconds (randomized)
    [INFO    ] Returning information for job: 20151124083900185051
    [DEBUG   ] Initializing new AsyncZeroMQReqChannel for ('/etc/salt/pki/minion', 'client.jacky.com', 'tcp://192.168.33.20:4506', 'aes')
    [DEBUG   ] Initializing new SAuth for ('/etc/salt/pki/minion', 'client.jacky.com', 'tcp://192.168.33.20:4506')

3.7. SaltClient将结果返回

在3.1里讲到了Client以SUB的角色连接到了Master的EventPublisher上监听结果。自此,执行一个命令的闭环就完成了。

4. 小结

分析完这个流程,几点值得学习的地方是

  1. 如何利用ZeroMQ实现各种进程通信模型
  2. 如何用Python实现大型基础架构软件

5. 参考资料

  1. Salt Development Architecture Overview
  2. Saltstack github