Synnefo » snf-ganeti

#

#

# Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 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.

"""Tool to restart erroneously downed virtual machines.

This program and set of classes implement a watchdog to restart

virtual machines in a Ganeti cluster that have crashed or been killed

by a node reboot.  Run from cron or similar.

"""

import os

import os.path

import sys

import time

import logging

import operator

import errno

from optparse import OptionParser

from ganeti import utils

from ganeti import constants

from ganeti import compat

from ganeti import errors

from ganeti import opcodes

from ganeti import cli

from ganeti import luxi

from ganeti import rapi

from ganeti import netutils

from ganeti import qlang

from ganeti import objects

from ganeti import ssconf

from ganeti import ht

import ganeti.rapi.client # pylint: disable-msg=W0611

from ganeti.watcher import nodemaint

from ganeti.watcher import state

MAXTRIES = 5

BAD_STATES = frozenset([

  constants.INSTST_ERRORDOWN,

  ])

HELPLESS_STATES = frozenset([

  constants.INSTST_NODEDOWN,

  constants.INSTST_NODEOFFLINE,

  ])

NOTICE = "NOTICE"

ERROR = "ERROR"

#: Number of seconds to wait between starting child processes for node groups

CHILD_PROCESS_DELAY = 1.0

#: How many seconds to wait for instance status file lock

INSTANCE_STATUS_LOCK_TIMEOUT = 10.0

class NotMasterError(errors.GenericError):

  """Exception raised when this host is not the master."""

def ShouldPause():

  """Check whether we should pause.

  """

  return bool(utils.ReadWatcherPauseFile(constants.WATCHER_PAUSEFILE))

def StartNodeDaemons():

  """Start all the daemons that should be running on all nodes.

  """

  # on master or not, try to start the node daemon

  utils.EnsureDaemon(constants.NODED)

  # start confd as well. On non candidates it will be in disabled mode.

  utils.EnsureDaemon(constants.CONFD)

def RunWatcherHooks():

  """Run the watcher hooks.

  """

  hooks_dir = utils.PathJoin(constants.HOOKS_BASE_DIR,

                             constants.HOOKS_NAME_WATCHER)

  if not os.path.isdir(hooks_dir):

    return

  try:

    results = utils.RunParts(hooks_dir)

  except Exception: # pylint: disable-msg=W0703

    logging.exception("RunParts %s failed: %s", hooks_dir)

    return

  for (relname, status, runresult) in results:

    if status == constants.RUNPARTS_SKIP:

      logging.debug("Watcher hook %s: skipped", relname)

    elif status == constants.RUNPARTS_ERR:

      logging.warning("Watcher hook %s: error (%s)", relname, runresult)

    elif status == constants.RUNPARTS_RUN:

      if runresult.failed:

        logging.warning("Watcher hook %s: failed (exit: %d) (output: %s)",

                        relname, runresult.exit_code, runresult.output)

      else:

        logging.debug("Watcher hook %s: success (output: %s)", relname,

                      runresult.output)

    else:

      raise errors.ProgrammerError("Unknown status %s returned by RunParts",

                                   status)

class Instance(object):

  """Abstraction for a Virtual Machine instance.

  """

  def __init__(self, name, status, autostart, snodes):

    self.name = name

    self.status = status

    self.autostart = autostart

    self.snodes = snodes

  def Restart(self, cl):

    """Encapsulates the start of an instance.

    """

    op = opcodes.OpInstanceStartup(instance_name=self.name, force=False)

    cli.SubmitOpCode(op, cl=cl)

  def ActivateDisks(self, cl):

    """Encapsulates the activation of all disks of an instance.

    """

    op = opcodes.OpInstanceActivateDisks(instance_name=self.name)

    cli.SubmitOpCode(op, cl=cl)

class Node:

  """Data container representing cluster node.

  """

  def __init__(self, name, bootid, offline, secondaries):

    """Initializes this class.

    """

    self.name = name

    self.bootid = bootid

    self.offline = offline

    self.secondaries = secondaries

def _CheckInstances(cl, notepad, instances):

  """Make a pass over the list of instances, restarting downed ones.

  """

  notepad.MaintainInstanceList(instances.keys())

  started = set()

  for inst in instances.values():

    if inst.status in BAD_STATES:

      n = notepad.NumberOfRestartAttempts(inst.name)

      if n > MAXTRIES:

        logging.warning("Not restarting instance '%s', retries exhausted",

                        inst.name)

        continue

      if n == MAXTRIES:

        notepad.RecordRestartAttempt(inst.name)

        logging.error("Could not restart instance '%s' after %s attempts,"

                      " giving up", inst.name, MAXTRIES)

        continue

      try:

        logging.info("Restarting instance '%s' (attempt #%s)",

                     inst.name, n + 1)

        inst.Restart(cl)

      except Exception: # pylint: disable-msg=W0703

        logging.exception("Error while restarting instance '%s'", inst.name)

      else:

        started.add(inst.name)

      notepad.RecordRestartAttempt(inst.name)

    else:

      if notepad.NumberOfRestartAttempts(inst.name):

        notepad.RemoveInstance(inst.name)

        if inst.status not in HELPLESS_STATES:

          logging.info("Restart of instance '%s' succeeded", inst.name)

  return started

def _CheckDisks(cl, notepad, nodes, instances, started):

  """Check all nodes for restarted ones.

  """

  check_nodes = []

  for node in nodes.values():

    old = notepad.GetNodeBootID(node.name)

    if not node.bootid:

      # Bad node, not returning a boot id

      if not node.offline:

        logging.debug("Node '%s' missing boot ID, skipping secondary checks",

                      node.name)

      continue

    if old != node.bootid:

      # Node's boot ID has changed, probably through a reboot

      check_nodes.append(node)

  if check_nodes:

    # Activate disks for all instances with any of the checked nodes as a

    # secondary node.

    for node in check_nodes:

      for instance_name in node.secondaries:

        try:

          inst = instances[instance_name]

        except KeyError:

          logging.info("Can't find instance '%s', maybe it was ignored",

                       instance_name)

          continue

        if not inst.autostart:

          logging.info("Skipping disk activation for non-autostart"

                       " instance '%s'", inst.name)

          continue

        if inst.name in started:

          # we already tried to start the instance, which should have

          # activated its drives (if they can be at all)

          logging.debug("Skipping disk activation for instance '%s' as"

                        " it was already started", inst.name)

          continue

        try:

          logging.info("Activating disks for instance '%s'", inst.name)

          inst.ActivateDisks(cl)

        except Exception: # pylint: disable-msg=W0703

          logging.exception("Error while activating disks for instance '%s'",

                            inst.name)

    # Keep changed boot IDs

    for node in check_nodes:

      notepad.SetNodeBootID(node.name, node.bootid)

def _CheckForOfflineNodes(nodes, instance):

  """Checks if given instances has any secondary in offline status.

  @param instance: The instance object

  @return: True if any of the secondary is offline, False otherwise

  """

  return compat.any(nodes[node_name].offline for node_name in instance.snodes)

def _VerifyDisks(cl, uuid, nodes, instances):

  """Run a per-group "gnt-cluster verify-disks".

  """

  job_id = cl.SubmitJob([opcodes.OpGroupVerifyDisks(group_name=uuid)])

  ((_, offline_disk_instances, _), ) = \

    cli.PollJob(job_id, cl=cl, feedback_fn=logging.debug)

  cl.ArchiveJob(job_id)

  if not offline_disk_instances:

    # nothing to do

    logging.debug("Verify-disks reported no offline disks, nothing to do")

    return

  logging.debug("Will activate disks for instance(s) %s",

                utils.CommaJoin(offline_disk_instances))

  # We submit only one job, and wait for it. Not optimal, but this puts less

  # load on the job queue.

  job = []

  for name in offline_disk_instances:

    try:

      inst = instances[name]

    except KeyError:

      logging.info("Can't find instance '%s', maybe it was ignored", name)

      continue

    if inst.status in HELPLESS_STATES or _CheckForOfflineNodes(nodes, inst):

      logging.info("Skipping instance '%s' because it is in a helpless state or"

                   " has offline secondaries", name)

      continue

    job.append(opcodes.OpInstanceActivateDisks(instance_name=name))

  if job:

    job_id = cli.SendJob(job, cl=cl)

    try:

      cli.PollJob(job_id, cl=cl, feedback_fn=logging.debug)

    except Exception: # pylint: disable-msg=W0703

      logging.exception("Error while activating disks")

def IsRapiResponding(hostname):

  """Connects to RAPI port and does a simple test.

  Connects to RAPI port of hostname and does a simple test. At this time, the

  test is GetVersion.

  @type hostname: string

  @param hostname: hostname of the node to connect to.

  @rtype: bool

  @return: Whether RAPI is working properly

  """

  curl_config = rapi.client.GenericCurlConfig()

  rapi_client = rapi.client.GanetiRapiClient(hostname,

                                             curl_config_fn=curl_config)

  try:

    master_version = rapi_client.GetVersion()

  except rapi.client.CertificateError, err:

    logging.warning("RAPI certificate error: %s", err)

    return False

  except rapi.client.GanetiApiError, err:

    logging.warning("RAPI error: %s", err)

    return False

  else:

    logging.debug("Reported RAPI version %s", master_version)

    return master_version == constants.RAPI_VERSION

def ParseOptions():

  """Parse the command line options.

  @return: (options, args) as from OptionParser.parse_args()

  """

  parser = OptionParser(description="Ganeti cluster watcher",

                        usage="%prog [-d]",

                        version="%%prog (ganeti) %s" %

                        constants.RELEASE_VERSION)

  parser.add_option(cli.DEBUG_OPT)

  parser.add_option(cli.NODEGROUP_OPT)

  parser.add_option("-A", "--job-age", dest="job_age", default=6 * 3600,

                    help="Autoarchive jobs older than this age (default"

                          " 6 hours)")

  parser.add_option("--ignore-pause", dest="ignore_pause", default=False,

                    action="store_true", help="Ignore cluster pause setting")

  parser.add_option("--wait-children", dest="wait_children", default=False,

                    action="store_true", help="Wait for child processes")

  options, args = parser.parse_args()

  options.job_age = cli.ParseTimespec(options.job_age)

  if args:

    parser.error("No arguments expected")

  return (options, args)

def _WriteInstanceStatus(filename, data):

  """Writes the per-group instance status file.

  The entries are sorted.

  @type filename: string

  @param filename: Path to instance status file

  @type data: list of tuple; (instance name as string, status as string)

  @param data: Instance name and status

  """

  logging.debug("Updating instance status file '%s' with %s instances",

                filename, len(data))

  utils.WriteFile(filename,

                  data="".join(map(compat.partial(operator.mod, "%s %s\n"),

                                   sorted(data))))

def _UpdateInstanceStatus(filename, instances):

  """Writes an instance status file from L{Instance} objects.

  @type filename: string

  @param filename: Path to status file

  @type instances: list of L{Instance}

  """

  _WriteInstanceStatus(filename, [(inst.name, inst.status)

                                  for inst in instances])

class _StatCb:

  """Helper to store file handle's C{fstat}.

  """

  def __init__(self):

    """Initializes this class.

    """

    self.st = None

  def __call__(self, fh):

    """Calls C{fstat} on file handle.

    """

    self.st = os.fstat(fh.fileno())

def _ReadInstanceStatus(filename):

  """Reads an instance status file.

  @type filename: string

  @param filename: Path to status file

  @rtype: tuple; (None or number, list of lists containing instance name and

    status)

  @return: File's mtime and instance status contained in the file; mtime is

    C{None} if file can't be read

  """

  logging.debug("Reading per-group instance status from '%s'", filename)

  statcb = _StatCb()

  try:

    content = utils.ReadFile(filename, preread=statcb)

  except EnvironmentError, err:

    if err.errno == errno.ENOENT:

      logging.error("Can't read '%s', does not exist (yet)", filename)

    else:

      logging.exception("Unable to read '%s', ignoring", filename)

    return (None, None)

  else:

    return (statcb.st.st_mtime, [line.split(1)

                                 for line in content.splitlines()])

def _MergeInstanceStatus(filename, pergroup_filename, groups):

  """Merges all per-group instance status files into a global one.

  @type filename: string

  @param filename: Path to global instance status file

  @type pergroup_filename: string

  @param pergroup_filename: Path to per-group status files, must contain "%s"

    to be replaced with group UUID

  @type groups: sequence

  @param groups: UUIDs of known groups

  """

  # Lock global status file in exclusive mode

  lock = utils.FileLock.Open(filename)

  try:

    lock.Exclusive(blocking=True, timeout=INSTANCE_STATUS_LOCK_TIMEOUT)

  except errors.LockError, err:

    # All per-group processes will lock and update the file. None of them

    # should take longer than 10 seconds (the value of

    # INSTANCE_STATUS_LOCK_TIMEOUT).

    logging.error("Can't acquire lock on instance status file '%s', not"

                  " updating: %s", filename, err)

    return

  logging.debug("Acquired exclusive lock on '%s'", filename)

  data = {}

  # Load instance status from all groups

  for group_uuid in groups:

    (mtime, instdata) = _ReadInstanceStatus(pergroup_filename % group_uuid)

    if mtime is not None:

      for (instance_name, status) in instdata:

        data.setdefault(instance_name, []).append((mtime, status))

  # Select last update based on file mtime

  inststatus = [(instance_name, sorted(status, reverse=True)[0][1])

                for (instance_name, status) in data.items()]

  # Write the global status file. Don't touch file after it's been

  # updated--there is no lock anymore.

  _WriteInstanceStatus(filename, inststatus)

def GetLuxiClient(try_restart):

  """Tries to connect to the master daemon.

  @type try_restart: bool

  @param try_restart: Whether to attempt to restart the master daemon

  """

  try:

    return cli.GetClient()

  except errors.OpPrereqError, err:

    # this is, from cli.GetClient, a not-master case

    raise NotMasterError("Not on master node (%s)" % err)

  except luxi.NoMasterError, err:

    if not try_restart:

      raise

    logging.warning("Master daemon seems to be down (%s), trying to restart",

                    err)

    if not utils.EnsureDaemon(constants.MASTERD):

      raise errors.GenericError("Can't start the master daemon")

    # Retry the connection

    return cli.GetClient()

def _StartGroupChildren(cl, wait):

  """Starts a new instance of the watcher for every node group.

  """

  assert not compat.any(arg.startswith(cli.NODEGROUP_OPT_NAME)

                        for arg in sys.argv)

  result = cl.QueryGroups([], ["name", "uuid"], False)

  children = []

  for (idx, (name, uuid)) in enumerate(result):

    args = sys.argv + [cli.NODEGROUP_OPT_NAME, uuid]

    if idx > 0:

      # Let's not kill the system

      time.sleep(CHILD_PROCESS_DELAY)

    logging.debug("Spawning child for group '%s' (%s), arguments %s",

                  name, uuid, args)

    try:

      # TODO: Should utils.StartDaemon be used instead?

      pid = os.spawnv(os.P_NOWAIT, args[0], args)

    except Exception: # pylint: disable-msg=W0703

      logging.exception("Failed to start child for group '%s' (%s)",

                        name, uuid)

    else:

      logging.debug("Started with PID %s", pid)

      children.append(pid)

  if wait:

    for pid in children:

      logging.debug("Waiting for child PID %s", pid)

      try:

        result = utils.RetryOnSignal(os.waitpid, pid, 0)

      except EnvironmentError, err:

        result = str(err)

      logging.debug("Child PID %s exited with status %s", pid, result)

def _ArchiveJobs(cl, age):

  """Archives old jobs.

  """

  (arch_count, left_count) = cl.AutoArchiveJobs(age)

  logging.debug("Archived %s jobs, left %s", arch_count, left_count)

def _CheckMaster(cl):

  """Ensures current host is master node.

  """

  (master, ) = cl.QueryConfigValues(["master_node"])

  if master != netutils.Hostname.GetSysName():

    raise NotMasterError("This is not the master node")

@rapi.client.UsesRapiClient

def _GlobalWatcher(opts):

  """Main function for global watcher.

  At the end child processes are spawned for every node group.

  """

  StartNodeDaemons()

  RunWatcherHooks()

  # Run node maintenance in all cases, even if master, so that old masters can

  # be properly cleaned up

  if nodemaint.NodeMaintenance.ShouldRun(): # pylint: disable-msg=E0602

    nodemaint.NodeMaintenance().Exec() # pylint: disable-msg=E0602

  try:

    client = GetLuxiClient(True)

  except NotMasterError:

    # Don't proceed on non-master nodes

    return constants.EXIT_SUCCESS

  # we are on master now

  utils.EnsureDaemon(constants.RAPI)

  # If RAPI isn't responding to queries, try one restart

  logging.debug("Attempting to talk to remote API on %s",

                constants.IP4_ADDRESS_LOCALHOST)

  if not IsRapiResponding(constants.IP4_ADDRESS_LOCALHOST):

    logging.warning("Couldn't get answer from remote API, restaring daemon")

    utils.StopDaemon(constants.RAPI)

    utils.EnsureDaemon(constants.RAPI)

    logging.debug("Second attempt to talk to remote API")

    if not IsRapiResponding(constants.IP4_ADDRESS_LOCALHOST):

      logging.fatal("RAPI is not responding")

  logging.debug("Successfully talked to remote API")

  _CheckMaster(client)

  _ArchiveJobs(client, opts.job_age)

  # Spawn child processes for all node groups

  _StartGroupChildren(client, opts.wait_children)

  return constants.EXIT_SUCCESS

def _GetGroupData(cl, uuid):

  """Retrieves instances and nodes per node group.

  """

  # TODO: Implement locking

  job = [

    # Get all primary instances in group

    opcodes.OpQuery(what=constants.QR_INSTANCE,

                    fields=["name", "status", "admin_state", "snodes",

                            "pnode.group.uuid", "snodes.group.uuid"],

                    filter=[qlang.OP_EQUAL, "pnode.group.uuid", uuid]),

    # Get all nodes in group

    opcodes.OpQuery(what=constants.QR_NODE,

                    fields=["name", "bootid", "offline"],

                    filter=[qlang.OP_EQUAL, "group.uuid", uuid]),

    ]

  job_id = cl.SubmitJob(job)

  results = map(objects.QueryResponse.FromDict,

                cli.PollJob(job_id, cl=cl, feedback_fn=logging.debug))

  cl.ArchiveJob(job_id)

  results_data = map(operator.attrgetter("data"), results)

  # Ensure results are tuples with two values

  assert compat.all(map(ht.TListOf(ht.TListOf(ht.TIsLength(2))), results_data))

  # Extract values ignoring result status

  (raw_instances, raw_nodes) = [[map(compat.snd, values)

                                 for values in res]

                                for res in results_data]

  secondaries = {}

  instances = []

  # Load all instances

  for (name, status, autostart, snodes, pnode_group_uuid,

       snodes_group_uuid) in raw_instances:

    if snodes and set([pnode_group_uuid]) != set(snodes_group_uuid):

      logging.error("Ignoring split instance '%s', primary group %s, secondary"

                    " groups %s", name, pnode_group_uuid,

                    utils.CommaJoin(snodes_group_uuid))

    else:

      instances.append(Instance(name, status, autostart, snodes))

      for node in snodes:

        secondaries.setdefault(node, set()).add(name)

  # Load all nodes

  nodes = [Node(name, bootid, offline, secondaries.get(name, set()))

           for (name, bootid, offline) in raw_nodes]

  return (dict((node.name, node) for node in nodes),

          dict((inst.name, inst) for inst in instances))

def _LoadKnownGroups():

  """Returns a list of all node groups known by L{ssconf}.

  """

  groups = ssconf.SimpleStore().GetNodegroupList()

  result = list(line.split(None, 1)[0] for line in groups

                if line.strip())

  if not compat.all(map(utils.UUID_RE.match, result)):

    raise errors.GenericError("Ssconf contains invalid group UUID")

  return result

def _GroupWatcher(opts):

  """Main function for per-group watcher process.

  """

  group_uuid = opts.nodegroup.lower()

  if not utils.UUID_RE.match(group_uuid):

    raise errors.GenericError("Node group parameter (%s) must be given a UUID,"

                              " got '%s'" %

                              (cli.NODEGROUP_OPT_NAME, group_uuid))

  logging.info("Watcher for node group '%s'", group_uuid)

  known_groups = _LoadKnownGroups()

  # Check if node group is known

  if group_uuid not in known_groups:

    raise errors.GenericError("Node group '%s' is not known by ssconf" %

                              group_uuid)

  # Group UUID has been verified and should not contain any dangerous characters

  state_path = constants.WATCHER_GROUP_STATE_FILE % group_uuid

  inst_status_path = constants.WATCHER_GROUP_INSTANCE_STATUS_FILE % group_uuid

  logging.debug("Using state file %s", state_path)

  # Global watcher

  statefile = state.OpenStateFile(state_path) # pylint: disable-msg=E0602

  if not statefile:

    return constants.EXIT_FAILURE

  notepad = state.WatcherState(statefile) # pylint: disable-msg=E0602

  try:

    # Connect to master daemon

    client = GetLuxiClient(False)

    _CheckMaster(client)

    (nodes, instances) = _GetGroupData(client, group_uuid)

    # Update per-group instance status file

    _UpdateInstanceStatus(inst_status_path, instances.values())

    _MergeInstanceStatus(constants.INSTANCE_STATUS_FILE,

                         constants.WATCHER_GROUP_INSTANCE_STATUS_FILE,

                         known_groups)

    started = _CheckInstances(client, notepad, instances)

    _CheckDisks(client, notepad, nodes, instances, started)

    _VerifyDisks(client, group_uuid, nodes, instances)

  except Exception, err:

    logging.info("Not updating status file due to failure: %s", err)

    raise

  else:

    # Save changes for next run

    notepad.Save(state_path)

  return constants.EXIT_SUCCESS

def Main():

  """Main function.

  """

  (options, _) = ParseOptions()

  utils.SetupLogging(constants.LOG_WATCHER, sys.argv[0],

                     debug=options.debug, stderr_logging=options.debug)

  if ShouldPause() and not options.ignore_pause:

    logging.debug("Pause has been set, exiting")

    return constants.EXIT_SUCCESS

  # Try to acquire global watcher lock in shared mode

  lock = utils.FileLock.Open(constants.WATCHER_LOCK_FILE)

  try:

    lock.Shared(blocking=False)

  except (EnvironmentError, errors.LockError), err:

    logging.error("Can't acquire lock on %s: %s",

                  constants.WATCHER_LOCK_FILE, err)

    return constants.EXIT_SUCCESS

  if options.nodegroup is None:

    fn = _GlobalWatcher

  else:

    # Per-nodegroup watcher

    fn = _GroupWatcher

  try:

    return fn(options)

  except (SystemExit, KeyboardInterrupt):

    raise

  except NotMasterError:

    logging.debug("Not master, exiting")

    return constants.EXIT_NOTMASTER

  except errors.ResolverError, err:

    logging.error("Cannot resolve hostname '%s', exiting", err.args[0])

    return constants.EXIT_NODESETUP_ERROR

  except errors.JobQueueFull:

    logging.error("Job queue is full, can't query cluster state")

  except errors.JobQueueDrainError:

    logging.error("Job queue is drained, can't maintain cluster state")

  except Exception, err:

    logging.exception(str(err))

    return constants.EXIT_FAILURE

  return constants.EXIT_SUCCESS