#
# Copyright (c) 2011-2014 Red Hat, Inc. <http://www.redhat.com>
# This file is part of GlusterFS.
# This file is licensed to you under your choice of the GNU Lesser
# General Public License, version 3 or any later version (LGPLv3 or
# later), or the GNU General Public License, version 2 (GPLv2), in all
# cases as published by the Free Software Foundation.
#
import os
import sys
import time
import signal
import logging
import xml.etree.ElementTree as XET
from threading import Lock
from errno import ECHILD, ESRCH
import random
from resource import SSH
import gsyncdconfig as gconf
from rconf import rconf
from syncdutils import select, waitpid, errno_wrap, lf, grabpidfile
from syncdutils import set_term_handler, GsyncdError
from syncdutils import Thread, finalize, Volinfo, VolinfoFromGconf
from syncdutils import gf_event, EVENT_GEOREP_FAULTY, get_up_nodes
from gsyncdstatus import GeorepStatus, set_monitor_status
from syncdutils import unshare_propagation_supported
import py2py3
from py2py3 import pipe
ParseError = XET.ParseError if hasattr(XET, 'ParseError') else SyntaxError
def get_subvol_num(brick_idx, vol, hot):
tier = vol.is_tier()
disperse_count = vol.disperse_count(tier, hot)
replica_count = vol.replica_count(tier, hot)
distribute_count = vol.distribution_count(tier, hot)
gconf.setconfig("master-distribution-count", distribute_count)
if (tier and not hot):
brick_idx = brick_idx - vol.get_hot_bricks_count(tier)
subvol_size = disperse_count if disperse_count > 0 else replica_count
cnt = int((brick_idx + 1) / subvol_size)
rem = (brick_idx + 1) % subvol_size
if rem > 0:
cnt = cnt + 1
if (tier and hot):
return "hot_" + str(cnt)
elif (tier and not hot):
return "cold_" + str(cnt)
else:
return str(cnt)
class Monitor(object):
"""class which spawns and manages gsyncd workers"""
ST_INIT = 'Initializing...'
ST_STARTED = 'Started'
ST_STABLE = 'Active'
ST_FAULTY = 'Faulty'
ST_INCON = 'inconsistent'
_ST_ORD = [ST_STABLE, ST_INIT, ST_FAULTY, ST_INCON]
def __init__(self):
self.lock = Lock()
self.state = {}
self.status = {}
@staticmethod
def terminate():
# relax one SIGTERM by setting a handler that sets back
# standard handler
set_term_handler(lambda *a: set_term_handler())
# give a chance to graceful exit
errno_wrap(os.kill, [-os.getpid(), signal.SIGTERM], [ESRCH])
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
def multiplex(self, wspx, suuid, slave_vol, slave_host, master, slavenodes):
argv = [os.path.basename(sys.executable), sys.argv[0]]
cpids = set()
agents = set()
ta = []
for wx in wspx:
def wmon(w):
cpid, _ = self.monitor(w, argv, cpids, agents, slave_vol,
slave_host, master, suuid, slavenodes)
time.sleep(1)
self.lock.acquire()
for cpid in cpids:
errno_wrap(os.kill, [cpid, signal.SIGKILL], [ESRCH])
for apid in agents:
errno_wrap(os.kill, [apid, signal.SIGKILL], [ESRCH])
self.lock.release()
finalize(exval=1)
t = Thread(target=wmon, args=[wx])
t.start()
ta.append(t)
# monitor status was being updated in each monitor thread. It
# should not be done as it can cause deadlock for a worker start.
# set_monitor_status uses flock to synchronize multple instances
# updating the file. Since each monitor thread forks worker and
# agent, these processes can hold the reference to fd of status
# file causing deadlock to workers which starts later as flock
# will not be release until all references to same fd is closed.
# It will also cause fd leaks.
self.lock.acquire()
set_monitor_status(gconf.get("state-file"), self.ST_STARTED)
self.lock.release()
for t in ta:
t.join()
def distribute(master, slave):
if rconf.args.use_gconf_volinfo:
mvol = VolinfoFromGconf(master.volume, master=True)
else:
mvol = Volinfo(master.volume, master.host, master=True)
logging.debug('master bricks: ' + repr(mvol.bricks))
prelude = []
slave_host = None
slave_vol = None
prelude = [gconf.get("ssh-command")] + \
gconf.get("ssh-options").split() + \
["-p", str(gconf.get("ssh-port"))] + \
[slave.remote_addr]
logging.debug('slave SSH gateway: ' + slave.remote_addr)
if rconf.args.use_gconf_volinfo:
svol = VolinfoFromGconf(slave.volume, master=False)
else:
svol = Volinfo(slave.volume, "localhost", prelude, master=False)
sbricks = svol.bricks
suuid = svol.uuid
slave_host = slave.remote_addr.split('@')[-1]
slave_vol = slave.volume
# save this xattr for the session delete command
old_stime_xattr_prefix = gconf.get("stime-xattr-prefix", None)
new_stime_xattr_prefix = "trusted.glusterfs." + mvol.uuid + "." + \
svol.uuid
if not old_stime_xattr_prefix or \
old_stime_xattr_prefix != new_stime_xattr_prefix:
gconf.setconfig("stime-xattr-prefix", new_stime_xattr_prefix)
logging.debug('slave bricks: ' + repr(sbricks))
slavenodes = set((b['host'], b["uuid"]) for b in sbricks)
rap = SSH.parse_ssh_address(slave)
slaves = [(rap['user'] + '@' + h[0], h[1]) for h in slavenodes]
workerspex = []
for idx, brick in enumerate(mvol.bricks):
if rconf.args.local_node_id == brick['uuid']:
is_hot = mvol.is_hot(":".join([brick['host'], brick['dir']]))
workerspex.append((brick,
slaves[idx % len(slaves)],
get_subvol_num(idx, mvol, is_hot),
is_hot))
logging.debug('worker specs: ' + repr(workerspex))
return workerspex, suuid, slave_vol, slave_host, master, slavenodes
def monitor(local, remote):
# Check if gsyncd restarted in pause state. If
# yes, send SIGSTOP to negative of monitor pid
# to go back to pause state.
if rconf.args.pause_on_start:
errno_wrap(os.kill, [-os.getpid(), signal.SIGSTOP], [ESRCH])
"""oh yeah, actually Monitor is used as singleton, too"""
return Monitor().multiplex(*distribute(local, remote))
def startup(go_daemon=True):
"""set up logging, pidfile grabbing, daemonization"""
pid_file = gconf.get("pid-file")
if not grabpidfile():
sys.stderr.write("pidfile is taken, exiting.\n")
sys.exit(2)
rconf.pid_file_owned = True
if not go_daemon:
return
x, y = pipe()
cpid = os.fork()
if cpid:
os.close(x)
sys.exit()
os.close(y)
os.setsid()
dn = os.open(os.devnull, os.O_RDWR)
for f in (sys.stdin, sys.stdout, sys.stderr):
os.dup2(dn, f.fileno())
if not grabpidfile(pid_file + '.tmp'):
raise GsyncdError("cannot grab temporary pidfile")
os.rename(pid_file + '.tmp', pid_file)
# wait for parent to terminate
# so we can start up with
# no messing from the dirty
# ol' bustard
select((x,), (), ())
os.close(x)