コード例 #1
0
ファイル: gsyncdstatus.py プロジェクト: wthuan/glusterfs
 def set_passive(self):
     if self.set_field("worker_status", "Passive"):
         gf_event(EVENT_GEOREP_PASSIVE,
                  master_volume=self.master,
                  slave_host=self.slave_host,
                  slave_volume=self.slave_volume,
                  brick_path=self.brick)
コード例 #2
0
ファイル: gsyncdstatus.py プロジェクト: fmpnate/glusterfs
 def set_passive(self):
     if self.set_field("worker_status", "Passive"):
         logging.info("Worker Status: Passive")
         gf_event(EVENT_GEOREP_PASSIVE,
                  master_volume=self.master,
                  slave_host=self.slave_host,
                  slave_volume=self.slave_volume,
                  brick_path=self.brick)
コード例 #3
0
 def set_active(self):
     if self.set_field("worker_status", "Active"):
         logging.info("Worker Status: Active")
         gf_event(EVENT_GEOREP_ACTIVE,
                  master_volume=self.master,
                  slave_host=self.slave_host,
                  slave_volume=self.slave_volume,
                  brick_path=self.brick)
コード例 #4
0
 def send_event(self, event_type, **kwargs):
     gf_event(event_type,
              master_volume=self.master,
              master_node=self.master_node,
              master_node_id=self.master_node_id,
              slave_host=self.slave_host,
              slave_volume=self.slave_volume,
              brick_path=self.brick,
              **kwargs)
コード例 #5
0
ファイル: gsyncdstatus.py プロジェクト: wthuan/glusterfs
 def trigger_gf_event_checkpoint_completion(self, checkpoint_time,
                                            checkpoint_completion_time):
     gf_event(EVENT_GEOREP_CHECKPOINT_COMPLETED,
              master_volume=self.master,
              slave_host=self.slave_host,
              slave_volume=self.slave_volume,
              brick_path=self.brick,
              checkpoint_time=checkpoint_time,
              checkpoint_completion_time=checkpoint_completion_time)
コード例 #6
0
ファイル: gsyncdstatus.py プロジェクト: fmpnate/glusterfs
 def trigger_gf_event_checkpoint_completion(self, checkpoint_time,
                                            checkpoint_completion_time):
     gf_event(EVENT_GEOREP_CHECKPOINT_COMPLETED,
              master_volume=self.master,
              slave_host=self.slave_host,
              slave_volume=self.slave_volume,
              brick_path=self.brick,
              checkpoint_time=checkpoint_time,
              checkpoint_completion_time=checkpoint_completion_time)
コード例 #7
0
    def monitor(self, w, argv, cpids, agents, slave_vol, slave_host, master,
                suuid, slavenodes):
        """the monitor loop

        Basic logic is a blantantly simple blunt heuristics:
        if spawned client survives 60 secs, it's considered OK.
        This servers us pretty well as it's not vulneralbe to
        any kind of irregular behavior of the child...

        ... well, except for one: if children is hung up on
        waiting for some event, it can survive aeons, still
        will be defunct. So we tweak the above logic to
        expect the worker to send us a signal within 60 secs
        (in the form of closing its end of a pipe). The worker
        does this when it's done with the setup stage
        ready to enter the service loop (note it's the setup
        stage which is vulnerable to hangs -- the full
        blown worker blows up on EPIPE if the net goes down,
        due to the keep-alive thread)
        """
        if not self.status.get(w[0]['dir'], None):
            self.status[w[0]['dir']] = GeorepStatus(
                gconf.get("state-file"), w[0]['host'], w[0]['dir'],
                w[0]['uuid'], master, "%s::%s" % (slave_host, slave_vol))

        set_monitor_status(gconf.get("state-file"), self.ST_STARTED)
        self.status[w[0]['dir']].set_worker_status(self.ST_INIT)

        ret = 0

        def nwait(p, o=0):
            try:
                p2, r = waitpid(p, o)
                if not p2:
                    return
                return r
            except OSError as e:
                # no child process, this happens if the child process
                # already died and has been cleaned up
                if e.errno == ECHILD:
                    return -1
                else:
                    raise

        def exit_signalled(s):
            """ child teminated due to receipt of SIGUSR1 """
            return (os.WIFSIGNALED(s) and (os.WTERMSIG(s) == signal.SIGUSR1))

        def exit_status(s):
            if os.WIFEXITED(s):
                return os.WEXITSTATUS(s)
            return 1

        conn_timeout = gconf.get("connection-timeout")
        while ret in (0, 1):
            remote_user, remote_host = w[1][0].split("@")
            remote_id = w[1][1]
            # Check the status of the connected slave node
            # If the connected slave node is down then try to connect to
            # different up node.
            current_slave_host = remote_host
            slave_up_hosts = get_up_nodes(slavenodes, gconf.get("ssh-port"))

            if (current_slave_host, remote_id) not in slave_up_hosts:
                if len(slave_up_hosts) > 0:
                    remote_new = random.choice(slave_up_hosts)
                    remote_host = "%s@%s" % (remote_user, remote_new[0])
                    remote_id = remote_new[1]

            # Spawn the worker and agent in lock to avoid fd leak
            self.lock.acquire()

            logging.info(
                lf('starting gsyncd worker',
                   brick=w[0]['dir'],
                   slave_node=remote_host))

            # Couple of pipe pairs for RPC communication b/w
            # worker and changelog agent.

            # read/write end for agent
            (ra, ww) = os.pipe()
            # read/write end for worker
            (rw, wa) = os.pipe()

            # spawn the agent process
            apid = os.fork()
            if apid == 0:
                os.close(rw)
                os.close(ww)
                args_to_agent = argv + [
                    'agent', rconf.args.master, rconf.args.slave,
                    '--local-path', w[0]['dir'], '--local-node', w[0]['host'],
                    '--local-node-id', w[0]['uuid'], '--slave-id', suuid,
                    '--rpc-fd', ','.join(
                        [str(ra), str(wa), str(rw),
                         str(ww)])
                ]

                if rconf.args.config_file is not None:
                    args_to_agent += ['-c', rconf.args.config_file]

                if rconf.args.debug:
                    args_to_agent.append("--debug")

                os.execv(sys.executable, args_to_agent)

            pr, pw = os.pipe()
            cpid = os.fork()
            if cpid == 0:
                os.close(pr)
                os.close(ra)
                os.close(wa)

                args_to_worker = argv + [
                    'worker', rconf.args.master, rconf.args.slave,
                    '--feedback-fd',
                    str(pw), '--local-path', w[0]['dir'], '--local-node',
                    w[0]['host'], '--local-node-id', w[0]['uuid'],
                    '--slave-id', suuid, '--rpc-fd', ','.join(
                        [str(rw), str(ww), str(ra),
                         str(wa)]), '--subvol-num',
                    str(w[2]), '--resource-remote', remote_host,
                    '--resource-remote-id', remote_id
                ]

                if rconf.args.config_file is not None:
                    args_to_worker += ['-c', rconf.args.config_file]

                if w[3]:
                    args_to_worker.append("--is-hottier")

                if rconf.args.debug:
                    args_to_worker.append("--debug")

                os.execv(sys.executable, args_to_worker)

            cpids.add(cpid)
            agents.add(apid)
            os.close(pw)

            # close all RPC pipes in monitor
            os.close(ra)
            os.close(wa)
            os.close(rw)
            os.close(ww)
            self.lock.release()

            t0 = time.time()
            so = select((pr, ), (), (), conn_timeout)[0]
            os.close(pr)

            if so:
                ret = nwait(cpid, os.WNOHANG)
                ret_agent = nwait(apid, os.WNOHANG)

                if ret_agent is not None:
                    # Agent is died Kill Worker
                    logging.info(
                        lf("Changelog Agent died, Aborting Worker",
                           brick=w[0]['dir']))
                    errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                    nwait(cpid)
                    nwait(apid)

                if ret is not None:
                    logging.info(
                        lf("worker died before establishing "
                           "connection",
                           brick=w[0]['dir']))
                    nwait(apid)  # wait for agent
                else:
                    logging.debug("worker(%s) connected" % w[0]['dir'])
                    while time.time() < t0 + conn_timeout:
                        ret = nwait(cpid, os.WNOHANG)
                        ret_agent = nwait(apid, os.WNOHANG)

                        if ret is not None:
                            logging.info(
                                lf("worker died in startup phase",
                                   brick=w[0]['dir']))
                            nwait(apid)  # wait for agent
                            break

                        if ret_agent is not None:
                            # Agent is died Kill Worker
                            logging.info(
                                lf("Changelog Agent died, Aborting "
                                   "Worker",
                                   brick=w[0]['dir']))
                            errno_wrap(os.kill, [cpid, signal.SIGKILL],
                                       [ESRCH])
                            nwait(cpid)
                            nwait(apid)
                            break

                        time.sleep(1)
            else:
                logging.info(
                    lf(
                        "Worker not confirmed after wait, aborting it. "
                        "Gsyncd invocation on remote slave via SSH or "
                        "gluster master mount might have hung. Please "
                        "check the above logs for exact issue and check "
                        "master or slave volume for errors. Restarting "
                        "master/slave volume accordingly might help.",
                        brick=w[0]['dir'],
                        timeout=conn_timeout))
                errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                nwait(apid)  # wait for agent
                ret = nwait(cpid)
            if ret is None:
                # If worker dies, agent terminates on EOF.
                # So lets wait for agent first.
                nwait(apid)
                ret = nwait(cpid)
            if exit_signalled(ret):
                ret = 0
            else:
                ret = exit_status(ret)
                if ret in (0, 1):
                    self.status[w[0]['dir']].set_worker_status(self.ST_FAULTY)
                    gf_event(EVENT_GEOREP_FAULTY,
                             master_volume=master.volume,
                             master_node=w[0]['host'],
                             master_node_id=w[0]['uuid'],
                             slave_host=slave_host,
                             slave_volume=slave_vol,
                             current_slave_host=current_slave_host,
                             brick_path=w[0]['dir'])
            time.sleep(10)
        self.status[w[0]['dir']].set_worker_status(self.ST_INCON)
        return ret
コード例 #8
0
ファイル: monitor.py プロジェクト: zagrider/glusterfs
    def monitor(self, w, argv, cpids, agents, slave_vol, slave_host, master):
        """the monitor loop

        Basic logic is a blantantly simple blunt heuristics:
        if spawned client survives 60 secs, it's considered OK.
        This servers us pretty well as it's not vulneralbe to
        any kind of irregular behavior of the child...

        ... well, except for one: if children is hung up on
        waiting for some event, it can survive aeons, still
        will be defunct. So we tweak the above logic to
        expect the worker to send us a signal within 60 secs
        (in the form of closing its end of a pipe). The worker
        does this when it's done with the setup stage
        ready to enter the service loop (note it's the setup
        stage which is vulnerable to hangs -- the full
        blown worker blows up on EPIPE if the net goes down,
        due to the keep-alive thread)
        """
        if not self.status.get(w[0]['dir'], None):
            self.status[w[0]['dir']] = GeorepStatus(
                gconf.state_file, w[0]['dir'], master,
                "%s::%s" % (slave_host, slave_vol))

        set_monitor_status(gconf.state_file, self.ST_STARTED)
        self.status[w[0]['dir']].set_worker_status(self.ST_INIT)

        ret = 0

        def nwait(p, o=0):
            try:
                p2, r = waitpid(p, o)
                if not p2:
                    return
                return r
            except OSError as e:
                # no child process, this happens if the child process
                # already died and has been cleaned up
                if e.errno == ECHILD:
                    return -1
                else:
                    raise

        def exit_signalled(s):
            """ child teminated due to receipt of SIGUSR1 """
            return (os.WIFSIGNALED(s) and (os.WTERMSIG(s) == signal.SIGUSR1))

        def exit_status(s):
            if os.WIFEXITED(s):
                return os.WEXITSTATUS(s)
            return 1

        conn_timeout = int(gconf.connection_timeout)
        while ret in (0, 1):
            remote_host = w[1]
            # Check the status of the connected slave node
            # If the connected slave node is down then try to connect to
            # different up node.
            m = re.match("(ssh|gluster|file):\/\/(.+)@([^:]+):(.+)",
                         remote_host)
            if m:
                current_slave_host = m.group(3)
                slave_up_hosts = get_slave_bricks_status(slave_host, slave_vol)

                if current_slave_host not in slave_up_hosts:
                    if len(slave_up_hosts) > 0:
                        remote_host = "%s://%s@%s:%s" % (m.group(1), m.group(
                            2), random.choice(slave_up_hosts), m.group(4))

            # Spawn the worker and agent in lock to avoid fd leak
            self.lock.acquire()

            logging.info('starting gsyncd worker(%s). Slave node: %s' %
                         (w[0]['dir'], remote_host))

            # Couple of pipe pairs for RPC communication b/w
            # worker and changelog agent.

            # read/write end for agent
            (ra, ww) = os.pipe()
            # read/write end for worker
            (rw, wa) = os.pipe()

            # spawn the agent process
            apid = os.fork()
            if apid == 0:
                os.close(rw)
                os.close(ww)
                os.execv(
                    sys.executable, argv + [
                        '--local-path', w[0]['dir'], '--agent', '--rpc-fd',
                        ','.join([str(ra), str(wa),
                                  str(rw), str(ww)])
                    ])
            pr, pw = os.pipe()
            cpid = os.fork()
            if cpid == 0:
                os.close(pr)
                os.close(ra)
                os.close(wa)
                os.execv(
                    sys.executable, argv + [
                        '--feedback-fd',
                        str(pw), '--local-path', w[0]['dir'], '--local-id',
                        '.' + escape(w[0]['dir']), '--rpc-fd', ','.join(
                            [str(rw), str(ww),
                             str(ra), str(wa)]), '--subvol-num',
                        str(w[2])
                    ] + (['--is-hottier'] if w[3] else []) +
                    ['--resource-remote', remote_host])

            cpids.add(cpid)
            agents.add(apid)
            os.close(pw)

            # close all RPC pipes in monitor
            os.close(ra)
            os.close(wa)
            os.close(rw)
            os.close(ww)
            self.lock.release()

            t0 = time.time()
            so = select((pr, ), (), (), conn_timeout)[0]
            os.close(pr)

            if so:
                ret = nwait(cpid, os.WNOHANG)
                ret_agent = nwait(apid, os.WNOHANG)

                if ret_agent is not None:
                    # Agent is died Kill Worker
                    logging.info("Changelog Agent died, "
                                 "Aborting Worker(%s)" % w[0]['dir'])
                    errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                    nwait(cpid)
                    nwait(apid)

                if ret is not None:
                    logging.info("worker(%s) died before establishing "
                                 "connection" % w[0]['dir'])
                    nwait(apid)  # wait for agent
                else:
                    logging.debug("worker(%s) connected" % w[0]['dir'])
                    while time.time() < t0 + conn_timeout:
                        ret = nwait(cpid, os.WNOHANG)
                        ret_agent = nwait(apid, os.WNOHANG)

                        if ret is not None:
                            logging.info("worker(%s) died in startup "
                                         "phase" % w[0]['dir'])
                            nwait(apid)  # wait for agent
                            break

                        if ret_agent is not None:
                            # Agent is died Kill Worker
                            logging.info("Changelog Agent died, Aborting "
                                         "Worker(%s)" % w[0]['dir'])
                            errno_wrap(os.kill, [cpid, signal.SIGKILL],
                                       [ESRCH])
                            nwait(cpid)
                            nwait(apid)
                            break

                        time.sleep(1)
            else:
                logging.info("worker(%s) not confirmed in %d sec, "
                             "aborting it" % (w[0]['dir'], conn_timeout))
                errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                nwait(apid)  # wait for agent
                ret = nwait(cpid)
            if ret is None:
                self.status[w[0]['dir']].set_worker_status(self.ST_STABLE)
                # If worker dies, agent terminates on EOF.
                # So lets wait for agent first.
                nwait(apid)
                ret = nwait(cpid)
            if exit_signalled(ret):
                ret = 0
            else:
                ret = exit_status(ret)
                if ret in (0, 1):
                    self.status[w[0]['dir']].set_worker_status(self.ST_FAULTY)
                    gf_event(EVENT_GEOREP_FAULTY,
                             master_volume=master.volume,
                             master_node=w[0]['host'],
                             slave_host=slave_host,
                             slave_volume=slave_vol,
                             current_slave_host=current_slave_host,
                             brick_path=w[0]['dir'])
            time.sleep(10)
        self.status[w[0]['dir']].set_worker_status(self.ST_INCON)
        return ret
コード例 #9
0
ファイル: monitor.py プロジェクト: rlugojr/glusterfs
    def monitor(self, w, argv, cpids, agents, slave_vol, slave_host, master):
        """the monitor loop

        Basic logic is a blantantly simple blunt heuristics:
        if spawned client survives 60 secs, it's considered OK.
        This servers us pretty well as it's not vulneralbe to
        any kind of irregular behavior of the child...

        ... well, except for one: if children is hung up on
        waiting for some event, it can survive aeons, still
        will be defunct. So we tweak the above logic to
        expect the worker to send us a signal within 60 secs
        (in the form of closing its end of a pipe). The worker
        does this when it's done with the setup stage
        ready to enter the service loop (note it's the setup
        stage which is vulnerable to hangs -- the full
        blown worker blows up on EPIPE if the net goes down,
        due to the keep-alive thread)
        """
        if not self.status.get(w[0]['dir'], None):
            self.status[w[0]['dir']] = GeorepStatus(gconf.state_file,
                                                    w[0]['host'],
                                                    w[0]['dir'],
                                                    w[0]['uuid'],
                                                    master,
                                                    "%s::%s" % (slave_host,
                                                                slave_vol))

        set_monitor_status(gconf.state_file, self.ST_STARTED)
        self.status[w[0]['dir']].set_worker_status(self.ST_INIT)

        ret = 0

        def nwait(p, o=0):
            try:
                p2, r = waitpid(p, o)
                if not p2:
                    return
                return r
            except OSError as e:
                # no child process, this happens if the child process
                # already died and has been cleaned up
                if e.errno == ECHILD:
                    return -1
                else:
                    raise

        def exit_signalled(s):
            """ child teminated due to receipt of SIGUSR1 """
            return (os.WIFSIGNALED(s) and (os.WTERMSIG(s) == signal.SIGUSR1))

        def exit_status(s):
            if os.WIFEXITED(s):
                return os.WEXITSTATUS(s)
            return 1

        conn_timeout = int(gconf.connection_timeout)
        while ret in (0, 1):
            remote_host = w[1]
            # Check the status of the connected slave node
            # If the connected slave node is down then try to connect to
            # different up node.
            m = re.match("(ssh|gluster|file):\/\/(.+)@([^:]+):(.+)",
                         remote_host)
            if m:
                current_slave_host = m.group(3)
                slave_up_hosts = get_slave_bricks_status(
                    slave_host, slave_vol)

                if current_slave_host not in slave_up_hosts:
                    if len(slave_up_hosts) > 0:
                        remote_host = "%s://%s@%s:%s" % (m.group(1),
                                                         m.group(2),
                                                         random.choice(
                                                             slave_up_hosts),
                                                         m.group(4))

            # Spawn the worker and agent in lock to avoid fd leak
            self.lock.acquire()

            logging.info('starting gsyncd worker(%s). Slave node: %s' %
                         (w[0]['dir'], remote_host))

            # Couple of pipe pairs for RPC communication b/w
            # worker and changelog agent.

            # read/write end for agent
            (ra, ww) = os.pipe()
            # read/write end for worker
            (rw, wa) = os.pipe()

            # spawn the agent process
            apid = os.fork()
            if apid == 0:
                os.close(rw)
                os.close(ww)
                os.execv(sys.executable, argv + ['--local-path', w[0]['dir'],
                                                 '--local-node', w[0]['host'],
                                                 '--local-node-id',
                                                 w[0]['uuid'],
                                                 '--agent',
                                                 '--rpc-fd',
                                                 ','.join([str(ra), str(wa),
                                                           str(rw), str(ww)])])
            pr, pw = os.pipe()
            cpid = os.fork()
            if cpid == 0:
                os.close(pr)
                os.close(ra)
                os.close(wa)
                os.execv(sys.executable, argv + ['--feedback-fd', str(pw),
                                                 '--local-path', w[0]['dir'],
                                                 '--local-node', w[0]['host'],
                                                 '--local-node-id',
                                                 w[0]['uuid'],
                                                 '--local-id',
                                                 '.' + escape(w[0]['dir']),
                                                 '--rpc-fd',
                                                 ','.join([str(rw), str(ww),
                                                           str(ra), str(wa)]),
                                                 '--subvol-num', str(w[2])] +
                         (['--is-hottier'] if w[3] else []) +
                         ['--resource-remote', remote_host])

            cpids.add(cpid)
            agents.add(apid)
            os.close(pw)

            # close all RPC pipes in monitor
            os.close(ra)
            os.close(wa)
            os.close(rw)
            os.close(ww)
            self.lock.release()

            t0 = time.time()
            so = select((pr,), (), (), conn_timeout)[0]
            os.close(pr)

            if so:
                ret = nwait(cpid, os.WNOHANG)
                ret_agent = nwait(apid, os.WNOHANG)

                if ret_agent is not None:
                    # Agent is died Kill Worker
                    logging.info("Changelog Agent died, "
                                 "Aborting Worker(%s)" % w[0]['dir'])
                    errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                    nwait(cpid)
                    nwait(apid)

                if ret is not None:
                    logging.info("worker(%s) died before establishing "
                                 "connection" % w[0]['dir'])
                    nwait(apid)  # wait for agent
                else:
                    logging.debug("worker(%s) connected" % w[0]['dir'])
                    while time.time() < t0 + conn_timeout:
                        ret = nwait(cpid, os.WNOHANG)
                        ret_agent = nwait(apid, os.WNOHANG)

                        if ret is not None:
                            logging.info("worker(%s) died in startup "
                                         "phase" % w[0]['dir'])
                            nwait(apid)  # wait for agent
                            break

                        if ret_agent is not None:
                            # Agent is died Kill Worker
                            logging.info("Changelog Agent died, Aborting "
                                         "Worker(%s)" % w[0]['dir'])
                            errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                            nwait(cpid)
                            nwait(apid)
                            break

                        time.sleep(1)
            else:
                logging.info("worker(%s) not confirmed in %d sec, "
                             "aborting it" % (w[0]['dir'], conn_timeout))
                errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                nwait(apid)  # wait for agent
                ret = nwait(cpid)
            if ret is None:
                self.status[w[0]['dir']].set_worker_status(self.ST_STABLE)
                # If worker dies, agent terminates on EOF.
                # So lets wait for agent first.
                nwait(apid)
                ret = nwait(cpid)
            if exit_signalled(ret):
                ret = 0
            else:
                ret = exit_status(ret)
                if ret in (0, 1):
                    self.status[w[0]['dir']].set_worker_status(self.ST_FAULTY)
                    gf_event(EVENT_GEOREP_FAULTY,
                             master_volume=master.volume,
                             master_node=w[0]['host'],
                             master_node_id=w[0]['uuid'],
                             slave_host=slave_host,
                             slave_volume=slave_vol,
                             current_slave_host=current_slave_host,
                             brick_path=w[0]['dir'])
            time.sleep(10)
        self.status[w[0]['dir']].set_worker_status(self.ST_INCON)
        return ret
コード例 #10
0
ファイル: monitor.py プロジェクト: amarts/glusterfs
    def monitor(self, w, argv, cpids, agents, slave_vol, slave_host, master,
                suuid, slavenodes):
        """the monitor loop

        Basic logic is a blantantly simple blunt heuristics:
        if spawned client survives 60 secs, it's considered OK.
        This servers us pretty well as it's not vulneralbe to
        any kind of irregular behavior of the child...

        ... well, except for one: if children is hung up on
        waiting for some event, it can survive aeons, still
        will be defunct. So we tweak the above logic to
        expect the worker to send us a signal within 60 secs
        (in the form of closing its end of a pipe). The worker
        does this when it's done with the setup stage
        ready to enter the service loop (note it's the setup
        stage which is vulnerable to hangs -- the full
        blown worker blows up on EPIPE if the net goes down,
        due to the keep-alive thread)
        """
        if not self.status.get(w[0]['dir'], None):
            self.status[w[0]['dir']] = GeorepStatus(gconf.get("state-file"),
                                                    w[0]['host'],
                                                    w[0]['dir'],
                                                    w[0]['uuid'],
                                                    master,
                                                    "%s::%s" % (slave_host,
                                                                slave_vol))
        ret = 0

        def nwait(p, o=0):
            try:
                p2, r = waitpid(p, o)
                if not p2:
                    return
                return r
            except OSError as e:
                # no child process, this happens if the child process
                # already died and has been cleaned up
                if e.errno == ECHILD:
                    return -1
                else:
                    raise

        def exit_signalled(s):
            """ child terminated due to receipt of SIGUSR1 """
            return (os.WIFSIGNALED(s) and (os.WTERMSIG(s) == signal.SIGUSR1))

        def exit_status(s):
            if os.WIFEXITED(s):
                return os.WEXITSTATUS(s)
            return 1

        conn_timeout = gconf.get("connection-timeout")
        while ret in (0, 1):
            remote_user, remote_host = w[1][0].split("@")
            remote_id = w[1][1]
            # Check the status of the connected slave node
            # If the connected slave node is down then try to connect to
            # different up node.
            current_slave_host = remote_host
            slave_up_hosts = get_up_nodes(slavenodes, gconf.get("ssh-port"))

            if (current_slave_host, remote_id) not in slave_up_hosts:
                if len(slave_up_hosts) > 0:
                    remote_new = random.choice(slave_up_hosts)
                    remote_host = "%s@%s" % (remote_user, remote_new[0])
                    remote_id = remote_new[1]

            # Spawn the worker and agent in lock to avoid fd leak
            self.lock.acquire()

            self.status[w[0]['dir']].set_worker_status(self.ST_INIT)
            logging.info(lf('starting gsyncd worker',
                            brick=w[0]['dir'],
                            slave_node=remote_host))

            # Couple of pipe pairs for RPC communication b/w
            # worker and changelog agent.

            # read/write end for agent
            (ra, ww) = pipe()
            # read/write end for worker
            (rw, wa) = pipe()

            # spawn the agent process
            apid = os.fork()
            if apid == 0:
                os.close(rw)
                os.close(ww)
                args_to_agent = argv + [
                    'agent',
                    rconf.args.master,
                    rconf.args.slave,
                    '--local-path', w[0]['dir'],
                    '--local-node', w[0]['host'],
                    '--local-node-id', w[0]['uuid'],
                    '--slave-id', suuid,
                    '--rpc-fd', ','.join([str(ra), str(wa), str(rw), str(ww)])
                ]

                if rconf.args.config_file is not None:
                    args_to_agent += ['-c', rconf.args.config_file]

                if rconf.args.debug:
                    args_to_agent.append("--debug")

                os.execv(sys.executable, args_to_agent)

            pr, pw = pipe()
            cpid = os.fork()
            if cpid == 0:
                os.close(pr)
                os.close(ra)
                os.close(wa)

                args_to_worker = argv + [
                    'worker',
                    rconf.args.master,
                    rconf.args.slave,
                    '--feedback-fd', str(pw),
                    '--local-path', w[0]['dir'],
                    '--local-node', w[0]['host'],
                    '--local-node-id', w[0]['uuid'],
                    '--slave-id', suuid,
                    '--rpc-fd',
                    ','.join([str(rw), str(ww), str(ra), str(wa)]),
                    '--subvol-num', str(w[2]),
                    '--resource-remote', remote_host,
                    '--resource-remote-id', remote_id
                ]

                if rconf.args.config_file is not None:
                    args_to_worker += ['-c', rconf.args.config_file]

                if w[3]:
                    args_to_worker.append("--is-hottier")

                if rconf.args.debug:
                    args_to_worker.append("--debug")

                access_mount = gconf.get("access-mount")
                if access_mount:
                    os.execv(sys.executable, args_to_worker)
                else:
                    if unshare_propagation_supported():
                        logging.debug("Worker would mount volume privately")
                        unshare_cmd = ['unshare', '-m', '--propagation',
                                       'private']
                        cmd = unshare_cmd + args_to_worker
                        os.execvp("unshare", cmd)
                    else:
                        logging.debug("Mount is not private. It would be lazy"
                                      " umounted")
                        os.execv(sys.executable, args_to_worker)

            cpids.add(cpid)
            agents.add(apid)
            os.close(pw)

            # close all RPC pipes in monitor
            os.close(ra)
            os.close(wa)
            os.close(rw)
            os.close(ww)
            self.lock.release()

            t0 = time.time()
            so = select((pr,), (), (), conn_timeout)[0]
            os.close(pr)

            if so:
                ret = nwait(cpid, os.WNOHANG)
                ret_agent = nwait(apid, os.WNOHANG)

                if ret_agent is not None:
                    # Agent is died Kill Worker
                    logging.info(lf("Changelog Agent died, Aborting Worker",
                                    brick=w[0]['dir']))
                    errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                    nwait(cpid)
                    nwait(apid)

                if ret is not None:
                    logging.info(lf("worker died before establishing "
                                    "connection",
                                    brick=w[0]['dir']))
                    nwait(apid)  # wait for agent
                else:
                    logging.debug("worker(%s) connected" % w[0]['dir'])
                    while time.time() < t0 + conn_timeout:
                        ret = nwait(cpid, os.WNOHANG)
                        ret_agent = nwait(apid, os.WNOHANG)

                        if ret is not None:
                            logging.info(lf("worker died in startup phase",
                                            brick=w[0]['dir']))
                            nwait(apid)  # wait for agent
                            break

                        if ret_agent is not None:
                            # Agent is died Kill Worker
                            logging.info(lf("Changelog Agent died, Aborting "
                                            "Worker",
                                            brick=w[0]['dir']))
                            errno_wrap(os.kill, [cpid, signal.SIGKILL],
                                       [ESRCH])
                            nwait(cpid)
                            nwait(apid)
                            break

                        time.sleep(1)
            else:
                logging.info(
                    lf("Worker not confirmed after wait, aborting it. "
                       "Gsyncd invocation on remote slave via SSH or "
                       "gluster master mount might have hung. Please "
                       "check the above logs for exact issue and check "
                       "master or slave volume for errors. Restarting "
                       "master/slave volume accordingly might help.",
                       brick=w[0]['dir'],
                       timeout=conn_timeout))
                errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
                nwait(apid)  # wait for agent
                ret = nwait(cpid)
            if ret is None:
                # If worker dies, agent terminates on EOF.
                # So lets wait for agent first.
                nwait(apid)
                ret = nwait(cpid)
            if exit_signalled(ret):
                ret = 0
            else:
                ret = exit_status(ret)
                if ret in (0, 1):
                    self.status[w[0]['dir']].set_worker_status(self.ST_FAULTY)
                    gf_event(EVENT_GEOREP_FAULTY,
                             master_volume=master.volume,
                             master_node=w[0]['host'],
                             master_node_id=w[0]['uuid'],
                             slave_host=slave_host,
                             slave_volume=slave_vol,
                             current_slave_host=current_slave_host,
                             brick_path=w[0]['dir'])
            time.sleep(10)
        self.status[w[0]['dir']].set_worker_status(self.ST_INCON)
        return ret