#!/usr/bin/python # # Copyright (C) 2006, 2007 Google Inc. # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation; either version 2 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, but # WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU # General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program; if not, write to the Free Software # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA # 02110-1301, USA. """Master daemon program. Some classes deviates from the standard style guide since the inheritance from parent classes requires it. """ # pylint: disable-msg=C0103 # C0103: Invalid name ganeti-masterd import sys import SocketServer import time import collections import signal import logging from optparse import OptionParser from ganeti import config from ganeti import constants from ganeti import daemon from ganeti import mcpu from ganeti import opcodes from ganeti import jqueue from ganeti import locking from ganeti import luxi from ganeti import utils from ganeti import errors from ganeti import ssconf from ganeti import workerpool from ganeti import rpc from ganeti import bootstrap CLIENT_REQUEST_WORKERS = 16 EXIT_NOTMASTER = constants.EXIT_NOTMASTER EXIT_NODESETUP_ERROR = constants.EXIT_NODESETUP_ERROR class ClientRequestWorker(workerpool.BaseWorker): # pylint: disable-msg=W0221 def RunTask(self, server, request, client_address): """Process the request. This is copied from the code in ThreadingMixIn. """ try: server.finish_request(request, client_address) server.close_request(request) except: # pylint: disable-msg=W0702 server.handle_error(request, client_address) server.close_request(request) class IOServer(SocketServer.UnixStreamServer): """IO thread class. This class takes care of initializing the other threads, setting signal handlers (which are processed only in this thread), and doing cleanup at shutdown. """ def __init__(self, address, rqhandler): """IOServer constructor @param address: the address to bind this IOServer to @param rqhandler: RequestHandler type object """ SocketServer.UnixStreamServer.__init__(self, address, rqhandler) # We'll only start threads once we've forked. self.context = None self.request_workers = None def setup_queue(self): self.context = GanetiContext() self.request_workers = workerpool.WorkerPool("ClientReq", CLIENT_REQUEST_WORKERS, ClientRequestWorker) def process_request(self, request, client_address): """Add task to workerpool to process request. """ (pid, uid, gid) = utils.GetSocketCredentials(request) logging.info("Accepted connection from pid=%s, uid=%s, gid=%s", pid, uid, gid) self.request_workers.AddTask(self, request, client_address) def handle_error(self, request, client_address): logging.exception("Error while handling request") @utils.SignalHandled([signal.SIGINT, signal.SIGTERM]) def serve_forever(self, signal_handlers=None): # pylint: disable-msg=W0221 """Handle one request at a time until told to quit.""" assert isinstance(signal_handlers, dict) and \ len(signal_handlers) > 0, \ "Broken SignalHandled decorator" # Since we use SignalHandled only once, the resulting dict will map all # signals to the same handler. We'll just use the first one. sighandler = signal_handlers.values()[0] while not sighandler.called: self.handle_request() def server_cleanup(self): """Cleanup the server. This involves shutting down the processor threads and the master socket. """ try: self.server_close() finally: if self.request_workers: self.request_workers.TerminateWorkers() if self.context: self.context.jobqueue.Shutdown() class ClientRqHandler(SocketServer.BaseRequestHandler): """Client handler""" EOM = '\3' READ_SIZE = 4096 def setup(self): # pylint: disable-msg=W0201 # setup() is the api for initialising for this class self._buffer = "" self._msgs = collections.deque() self._ops = ClientOps(self.server) def handle(self): while True: msg = self.read_message() if msg is None: logging.debug("client closed connection") break (method, args) = luxi.ParseRequest(msg) success = False try: result = self._ops.handle_request(method, args) success = True except errors.GenericError, err: result = errors.EncodeException(err) except: logging.error("Unexpected exception", exc_info=True) result = "Caught exception: %s" % str(sys.exc_info()[1]) self.send_message(luxi.FormatResponse(success, result)) def read_message(self): while not self._msgs: data = self.request.recv(self.READ_SIZE) if not data: return None new_msgs = (self._buffer + data).split(self.EOM) self._buffer = new_msgs.pop() self._msgs.extend(new_msgs) return self._msgs.popleft() def send_message(self, msg): # TODO: sendall is not guaranteed to send everything self.request.sendall(msg + self.EOM) class ClientOps: """Class holding high-level client operations.""" def __init__(self, server): self.server = server def handle_request(self, method, args): # pylint: disable-msg=R0911 queue = self.server.context.jobqueue # TODO: Parameter validation # TODO: Rewrite to not exit in each 'if/elif' branch if method == luxi.REQ_SUBMIT_JOB: logging.info("Received new job") ops = [opcodes.OpCode.LoadOpCode(state) for state in args] return queue.SubmitJob(ops) if method == luxi.REQ_SUBMIT_MANY_JOBS: logging.info("Received multiple jobs") jobs = [] for ops in args: jobs.append([opcodes.OpCode.LoadOpCode(state) for state in ops]) return queue.SubmitManyJobs(jobs) elif method == luxi.REQ_CANCEL_JOB: job_id = args logging.info("Received job cancel request for %s", job_id) return queue.CancelJob(job_id) elif method == luxi.REQ_ARCHIVE_JOB: job_id = args logging.info("Received job archive request for %s", job_id) return queue.ArchiveJob(job_id) elif method == luxi.REQ_AUTOARCHIVE_JOBS: (age, timeout) = args logging.info("Received job autoarchive request for age %s, timeout %s", age, timeout) return queue.AutoArchiveJobs(age, timeout) elif method == luxi.REQ_WAIT_FOR_JOB_CHANGE: (job_id, fields, prev_job_info, prev_log_serial, timeout) = args logging.info("Received job poll request for %s", job_id) return queue.WaitForJobChanges(job_id, fields, prev_job_info, prev_log_serial, timeout) elif method == luxi.REQ_QUERY_JOBS: (job_ids, fields) = args if isinstance(job_ids, (tuple, list)) and job_ids: msg = utils.CommaJoin(job_ids) else: msg = str(job_ids) logging.info("Received job query request for %s", msg) return queue.QueryJobs(job_ids, fields) elif method == luxi.REQ_QUERY_INSTANCES: (names, fields, use_locking) = args logging.info("Received instance query request for %s", names) if use_locking: raise errors.OpPrereqError("Sync queries are not allowed", errors.ECODE_INVAL) op = opcodes.OpQueryInstances(names=names, output_fields=fields, use_locking=use_locking) return self._Query(op) elif method == luxi.REQ_QUERY_NODES: (names, fields, use_locking) = args logging.info("Received node query request for %s", names) if use_locking: raise errors.OpPrereqError("Sync queries are not allowed", errors.ECODE_INVAL) op = opcodes.OpQueryNodes(names=names, output_fields=fields, use_locking=use_locking) return self._Query(op) elif method == luxi.REQ_QUERY_EXPORTS: nodes, use_locking = args if use_locking: raise errors.OpPrereqError("Sync queries are not allowed", errors.ECODE_INVAL) logging.info("Received exports query request") op = opcodes.OpQueryExports(nodes=nodes, use_locking=use_locking) return self._Query(op) elif method == luxi.REQ_QUERY_CONFIG_VALUES: fields = args logging.info("Received config values query request for %s", fields) op = opcodes.OpQueryConfigValues(output_fields=fields) return self._Query(op) elif method == luxi.REQ_QUERY_CLUSTER_INFO: logging.info("Received cluster info query request") op = opcodes.OpQueryClusterInfo() return self._Query(op) elif method == luxi.REQ_QUERY_TAGS: kind, name = args logging.info("Received tags query request") op = opcodes.OpGetTags(kind=kind, name=name) return self._Query(op) elif method == luxi.REQ_QUEUE_SET_DRAIN_FLAG: drain_flag = args logging.info("Received queue drain flag change request to %s", drain_flag) return queue.SetDrainFlag(drain_flag) elif method == luxi.REQ_SET_WATCHER_PAUSE: (until, ) = args if until is None: logging.info("Received request to no longer pause the watcher") else: if not isinstance(until, (int, float)): raise TypeError("Duration must be an integer or float") if until < time.time(): raise errors.GenericError("Unable to set pause end time in the past") logging.info("Received request to pause the watcher until %s", until) return _SetWatcherPause(until) else: logging.info("Received invalid request '%s'", method) raise ValueError("Invalid operation '%s'" % method) def _Query(self, op): """Runs the specified opcode and returns the result. """ # Queries don't have a job id proc = mcpu.Processor(self.server.context, None) return proc.ExecOpCode(op, None) class GanetiContext(object): """Context common to all ganeti threads. This class creates and holds common objects shared by all threads. """ # pylint: disable-msg=W0212 # we do want to ensure a singleton here _instance = None def __init__(self): """Constructs a new GanetiContext object. There should be only a GanetiContext object at any time, so this function raises an error if this is not the case. """ assert self.__class__._instance is None, "double GanetiContext instance" # Create global configuration object self.cfg = config.ConfigWriter() # Locking manager self.glm = locking.GanetiLockManager( self.cfg.GetNodeList(), self.cfg.GetInstanceList()) # Job queue self.jobqueue = jqueue.JobQueue(self) # setting this also locks the class against attribute modifications self.__class__._instance = self def __setattr__(self, name, value): """Setting GanetiContext attributes is forbidden after initialization. """ assert self.__class__._instance is None, "Attempt to modify Ganeti Context" object.__setattr__(self, name, value) def AddNode(self, node, ec_id): """Adds a node to the configuration and lock manager. """ # Add it to the configuration self.cfg.AddNode(node, ec_id) # If preseeding fails it'll not be added self.jobqueue.AddNode(node) # Add the new node to the Ganeti Lock Manager self.glm.add(locking.LEVEL_NODE, node.name) def ReaddNode(self, node): """Updates a node that's already in the configuration """ # Synchronize the queue again self.jobqueue.AddNode(node) def RemoveNode(self, name): """Removes a node from the configuration and lock manager. """ # Remove node from configuration self.cfg.RemoveNode(name) # Notify job queue self.jobqueue.RemoveNode(name) # Remove the node from the Ganeti Lock Manager self.glm.remove(locking.LEVEL_NODE, name) def _SetWatcherPause(until): """Creates or removes the watcher pause file. @type until: None or int @param until: Unix timestamp saying until when the watcher shouldn't run """ if until is None: utils.RemoveFile(constants.WATCHER_PAUSEFILE) else: utils.WriteFile(constants.WATCHER_PAUSEFILE, data="%d\n" % (until, )) return until def CheckAgreement(): """Check the agreement on who is the master. The function uses a very simple algorithm: we must get more positive than negative answers. Since in most of the cases we are the master, we'll use our own config file for getting the node list. In the future we could collect the current node list from our (possibly obsolete) known nodes. In order to account for cold-start of all nodes, we retry for up to a minute until we get a real answer as the top-voted one. If the nodes are more out-of-sync, for now manual startup of the master should be attempted. Note that for a even number of nodes cluster, we need at least half of the nodes (beside ourselves) to vote for us. This creates a problem on two-node clusters, since in this case we require the other node to be up too to confirm our status. """ myself = utils.HostInfo().name #temp instantiation of a config writer, used only to get the node list cfg = config.ConfigWriter() node_list = cfg.GetNodeList() del cfg retries = 6 while retries > 0: votes = bootstrap.GatherMasterVotes(node_list) if not votes: # empty node list, this is a one node cluster return True if votes[0][0] is None: retries -= 1 time.sleep(10) continue break if retries == 0: logging.critical("Cluster inconsistent, most of the nodes didn't answer" " after multiple retries. Aborting startup") logging.critical("Use the --no-voting option if you understand what" " effects it has on the cluster state") return False # here a real node is at the top of the list all_votes = sum(item[1] for item in votes) top_node, top_votes = votes[0] result = False if top_node != myself: logging.critical("It seems we are not the master (top-voted node" " is %s with %d out of %d votes)", top_node, top_votes, all_votes) elif top_votes < all_votes - top_votes: logging.critical("It seems we are not the master (%d votes for," " %d votes against)", top_votes, all_votes - top_votes) else: result = True return result def CheckAgreementWithRpc(): rpc.Init() try: return CheckAgreement() finally: rpc.Shutdown() def CheckMasterd(options, args): """Initial checks whether to run or exit with a failure. """ if args: # masterd doesn't take any arguments print >> sys.stderr, ("Usage: %s [-f] [-d]" % sys.argv[0]) sys.exit(constants.EXIT_FAILURE) ssconf.CheckMaster(options.debug) # If CheckMaster didn't fail we believe we are the master, but we have to # confirm with the other nodes. if options.no_voting: if options.yes_do_it: return sys.stdout.write("The 'no voting' option has been selected.\n") sys.stdout.write("This is dangerous, please confirm by" " typing uppercase 'yes': ") sys.stdout.flush() confirmation = sys.stdin.readline().strip() if confirmation != "YES": print >> sys.stderr, "Aborting." sys.exit(constants.EXIT_FAILURE) return # CheckAgreement uses RPC and threads, hence it needs to be run in a separate # process before we call utils.Daemonize in the current process. if not utils.RunInSeparateProcess(CheckAgreementWithRpc): sys.exit(constants.EXIT_FAILURE) def ExecMasterd (options, args): # pylint: disable-msg=W0613 """Main master daemon function, executed with the PID file held. """ # This is safe to do as the pid file guarantees against # concurrent execution. utils.RemoveFile(constants.MASTER_SOCKET) master = IOServer(constants.MASTER_SOCKET, ClientRqHandler) try: rpc.Init() try: # activate ip master_node = ssconf.SimpleStore().GetMasterNode() result = rpc.RpcRunner.call_node_start_master(master_node, False, False) msg = result.fail_msg if msg: logging.error("Can't activate master IP address: %s", msg) master.setup_queue() try: master.serve_forever() finally: master.server_cleanup() finally: rpc.Shutdown() finally: utils.RemoveFile(constants.MASTER_SOCKET) def main(): """Main function""" parser = OptionParser(description="Ganeti master daemon", usage="%prog [-f] [-d]", version="%%prog (ganeti) %s" % constants.RELEASE_VERSION) parser.add_option("--no-voting", dest="no_voting", help="Do not check that the nodes agree on this node" " being the master and start the daemon unconditionally", default=False, action="store_true") parser.add_option("--yes-do-it", dest="yes_do_it", help="Override interactive check for --no-voting", default=False, action="store_true") dirs = [(constants.RUN_GANETI_DIR, constants.RUN_DIRS_MODE), (constants.SOCKET_DIR, constants.SOCKET_DIR_MODE), ] daemon.GenericMain(constants.MASTERD, parser, dirs, CheckMasterd, ExecMasterd, multithreaded=True) if __name__ == "__main__": main()