monitor/scanapi.py

   1 #!/usr/bin/python
   2
   3 import os
   4 import sys
   5 import string
   6 import time
   7 from datetime import datetime,timedelta
   8 import threadpool
   9 import threading
  10
  11 import socket
  12 from pcucontrol import reboot
  13
  14 from pcucontrol.util import command
  15 from monitor import config
  16
  17 from monitor.database.info.model import *
  18
  19 from monitor.sources import comon
  20 from monitor.wrapper import plc, plccache
  21
  22 import traceback
  23 from monitor.common import nmap_port_status, email_exception
  24
  25 COMON_COTOPURL= "http://summer.cs.princeton.edu/status/tabulator.cgi?" + \
  26                         "table=table_nodeview&" + \
  27                         "dumpcols='name,resptime,sshstatus,uptime,lastcotop,cpuspeed,memsize,disksize'&" + \
  28                         "formatcsv"
  29
  30 api = plc.getAuthAPI()
  31 plc_lock = threading.Lock()
  32 round = 1
  33 global_round = round
  34 count = 0
  35
  36
  37 def get_pcu(pcuname):
  38         plc_lock.acquire()
  39         try:
  40                 #print "GetPCU from PLC %s" % pcuname
  41                 l_pcu  = plc.GetPCUs({'pcu_id' : pcuname})
  42                 #print l_pcu
  43                 if len(l_pcu) > 0:
  44                         l_pcu = l_pcu[0]
  45         except:
  46                 try:
  47                         #print "GetPCU from file %s" % pcuname
  48                         l_pcus = plccache.l_pcus
  49                         for i in l_pcus:
  50                                 if i['pcu_id'] == pcuname:
  51                                         l_pcu = i
  52                 except:
  53                         traceback.print_exc()
  54                         l_pcu = None
  55
  56         plc_lock.release()
  57         return l_pcu
  58
  59 def get_nodes(node_ids):
  60         plc_lock.acquire()
  61         l_node = []
  62         try:
  63                 l_node = plc.getNodes(node_ids, ['hostname', 'last_contact', 'node_id', 'ports'])
  64         except:
  65                 try:
  66                         plc_nodes = plccache.l_plcnodes
  67                         for n in plc_nodes:
  68                                 if n['node_id'] in node_ids:
  69                                         l_node.append(n)
  70                 except:
  71                         traceback.print_exc()
  72                         l_node = None
  73
  74         plc_lock.release()
  75         if l_node == []:
  76                 l_node = None
  77         return l_node
  78
  79
  80 def get_plc_pcu_values(pcuname):
  81         """
  82                 Try to contact PLC to get the PCU info.
  83                 If that fails, try a backup copy from the last run.
  84                 If that fails, return None
  85         """
  86         values = {}
  87
  88         l_pcu = get_pcu(pcuname)
  89
  90         if l_pcu is not None:
  91                 site_id = l_pcu['site_id']
  92                 node_ids = l_pcu['node_ids']
  93                 l_node = get_nodes(node_ids)
  94
  95                 if l_node is not None:
  96                         for node in l_node:
  97                                 values[node['hostname']] = node['ports'][0]
  98
  99                         values['nodenames'] = [node['hostname'] for node in l_node]
 100
 101                         # NOTE: this is for a dry run later. It doesn't matter which node.
 102                         values['node_id'] = l_node[0]['node_id']
 103
 104                 values.update(l_pcu)
 105         else:
 106                 values = None
 107
 108         return values
 109
 110 class ScanInterface(object):
 111         recordclass = None
 112         syncclass = None
 113         primarykey = 'hostname'
 114
 115         def __init__(self, round=1):
 116                 self.round = round
 117                 self.count = 1
 118
 119         def __getattr__(self, name):
 120                 if 'collect' in name or 'record' in name:
 121                         method = getattr(self, name, None)
 122                         if method is None:
 123                                 raise Exception("No such method %s" % name)
 124                         return method
 125                 else:
 126                         raise Exception("No such method %s" % name)
 127
 128         def collect(self, nodename, data):
 129                 pass
 130
 131         def record(self, request, (nodename, values) ):
 132
 133                 try:
 134                         if values is None:
 135                                 return
 136
 137                         if self.syncclass:
 138                                 fbnodesync = self.syncclass.findby_or_create(
 139                                                                                                 #if_new_set={'round' : self.round},
 140                                                                                                 **{ self.primarykey : nodename})
 141                         # NOTE: This code will either add a new record for the new self.round,
 142                         #       OR it will find the previous value, and update it with new information.
 143                         #       The data that is 'lost' is not that important, b/c older
 144                         #       history still exists.
 145                         fbrec = self.recordclass.findby_or_create(
 146                                                 **{ self.primarykey:nodename})
 147
 148                         fbrec.set( **values )
 149
 150                         fbrec.flush()
 151                         if self.syncclass:
 152                                 fbnodesync.round = self.round
 153                                 fbnodesync.flush()
 154
 155                         print "%d %s %s" % (self.count, nodename, values)
 156                         self.count += 1
 157
 158                 except:
 159                         print "ERROR:"
 160                         email_exception(nodename)
 161                         print traceback.print_exc()
 162                         pass
 163
 164 class ScanNodeInternal(ScanInterface):
 165         recordclass = FindbadNodeRecord
 166         #syncclass = FindbadNodeRecordSync
 167         syncclass = None
 168         primarykey = 'hostname'
 169
 170         def collectNMAP(self, nodename, cohash):
 171                 #### RUN NMAP ###############################
 172                 values = {}
 173                 nmap = command.CMD()
 174                 print "nmap -oG - -P0 -p22,80,806 %s | grep Host:" % nodename
 175                 (oval,eval) = nmap.run_noexcept("nmap -oG - -P0 -p22,80,806 %s | grep Host:" % nodename)
 176                 # NOTE: an empty / error value for oval, will still work.
 177                 (values['port_status'], continue_probe) = nmap_port_status(oval)
 178
 179                 values['date_checked'] = datetime.now()
 180
 181                 return (nodename, values)
 182
 183         def collectInternal(self, nodename, cohash):
 184                 ### RUN PING ######################
 185                 ping = command.CMD()
 186                 (oval,errval) = ping.run_noexcept("ping -c 1 -q %s | grep rtt" % nodename)
 187
 188                 try:
 189                         values = {}
 190
 191                         if oval == "":
 192                                 # An error occurred
 193                                 values['ping_status'] = False
 194                         else:
 195                                 values['ping_status'] = True
 196
 197                         try:
 198                                 for port in [22, 806]:
 199                                         ssh = command.SSH('root', nodename, port)
 200
 201                                         (oval, errval) = ssh.run_noexcept2(""" <<\EOF
 202                                                 echo "{"
 203                                                 echo '  "kernel_version":"'`uname -a`'",'
 204                                                 echo '  "bmlog":"'`ls /tmp/bm.log`'",'
 205                                                 echo '  "bootcd_version":"'`cat /mnt/cdrom/bootme/ID`'",'
 206                                                 echo '  "nm_status":"'`ps ax | grep nm.py | grep -v grep`'",'
 207                                                 echo '  "fs_status":"'`touch /var/log/monitor 2>&1`'",'
 208                                                 echo '  "dns_status":"'`host boot.planet-lab.org 2>&1`'",'
 209                                                 echo '  "princeton_comon_dir":"'`ls -d /vservers/princeton_comon`'",'
 210
 211                                                 ID=`grep princeton_comon /etc/passwd | awk -F : '{if ( $3 > 500 ) { print $3}}'`
 212                                                 echo '  "princeton_comon_running":"'`ls -d /proc/virtual/$ID`'",'
 213                                                 echo '  "princeton_comon_procs":"'`vps ax | grep $ID | grep -v grep | wc -l`'",'
 214                                                 echo "}"
 215 EOF                             """)
 216
 217                                         values['ssh_error'] = errval
 218                                         if len(oval) > 0:
 219                                                 #print "OVAL: %s" % oval
 220                                                 values.update(eval(oval))
 221                                                 values['ssh_portused'] = port
 222                                                 break
 223                                         else:
 224                                                 values.update({'kernel_version': "", 'bmlog' : "", 'bootcd_version' : '',
 225                                                                                 'nm_status' : '',
 226                                                                                 'fs_status' : '',
 227                                                                                 'dns_status' : '',
 228                                                                                 'princeton_comon_dir' : "",
 229                                                                                 'princeton_comon_running' : "",
 230                                                                                 'princeton_comon_procs' : "", 'ssh_portused' : None})
 231                         except:
 232                                 print traceback.print_exc()
 233                                 sys.exit(1)
 234
 235                         ### RUN SSH ######################
 236                         b_getbootcd_id = True
 237
 238                         oval = values['kernel_version']
 239                         if "2.6.17" in oval or "2.6.2" in oval:
 240                                 values['ssh_status'] = True
 241                                 values['observed_category'] = 'PROD'
 242                                 if "bm.log" in values['bmlog']:
 243                                         values['observed_status'] = 'DEBUG'
 244                                 else:
 245                                         values['observed_status'] = 'BOOT'
 246                         elif "2.6.12" in oval or "2.6.10" in oval:
 247                                 values['ssh_status'] = True
 248                                 values['observed_category'] = 'OLDPROD'
 249                                 if "bm.log" in values['bmlog']:
 250                                         values['observed_status'] = 'DEBUG'
 251                                 else:
 252                                         values['observed_status'] = 'BOOT'
 253
 254                         # NOTE: on 2.6.8 kernels, with 4.2 bootstrapfs, the chroot
 255                         #       command fails.  I have no idea why.
 256                         elif "2.4" in oval or "2.6.8" in oval:
 257                                 b_getbootcd_id = False
 258                                 values['ssh_status'] = True
 259                                 values['observed_category'] = 'OLDBOOTCD'
 260                                 values['observed_status'] = 'DEBUG'
 261                         elif oval != "":
 262                                 values['ssh_status'] = True
 263                                 values['observed_category'] = 'UNKNOWN'
 264                                 if "bm.log" in values['bmlog']:
 265                                         values['observed_status'] = 'DEBUG'
 266                                 else:
 267                                         values['observed_status'] = 'BOOT'
 268                         else:
 269                                 # An error occurred.
 270                                 b_getbootcd_id = False
 271                                 values['ssh_status'] = False
 272                                 values['observed_category'] = 'ERROR'
 273                                 values['observed_status'] = 'DOWN'
 274                                 val = errval.strip()
 275                                 values['ssh_error'] = val
 276                                 values['kernel_version'] = ""
 277
 278                         if b_getbootcd_id:
 279                                 # try to get BootCD for all nodes that are not 2.4 nor inaccessible
 280                                 oval = values['bootcd_version']
 281                                 if "BootCD" in oval:
 282                                         values['bootcd_version'] = oval
 283                                         if "v2" in oval and \
 284                                                 ( nodename is not "planetlab1.cs.unc.edu" and \
 285                                                   nodename is not "planetlab2.cs.unc.edu" ):
 286                                                 values['observed_category'] = 'OLDBOOTCD'
 287                                 else:
 288                                         values['bootcd_version'] = ""
 289                         else:
 290                                 values['bootcd_version'] = ""
 291
 292                         oval = values['nm_status']
 293                         if "nm.py" in oval:
 294                                 values['nm_status'] = "Y"
 295                         else:
 296                                 values['nm_status'] = "N"
 297
 298                         continue_slice_check = True
 299                         oval = values['princeton_comon_dir']
 300                         if "princeton_comon_dir" in oval:
 301                                 values['princeton_comon_dir'] = True
 302                         else:
 303                                 values['princeton_comon_dir'] = False
 304                                 continue_slice_check = False
 305
 306                         if continue_slice_check:
 307                                 oval = values['princeton_comon_running']
 308                                 if len(oval) > len('/proc/virtual/'):
 309                                         values['princeton_comon_running'] = True
 310                                 else:
 311                                         values['princeton_comon_running'] = False
 312                                         continue_slice_check = False
 313                         else:
 314                                 values['princeton_comon_running'] = False
 315
 316                         if continue_slice_check:
 317                                 oval = values['princeton_comon_procs']
 318                                 values['princeton_comon_procs'] = int(oval)
 319                         else:
 320                                 values['princeton_comon_procs'] = None
 321
 322
 323                         if nodename in cohash:
 324                                 values['comon_stats'] = cohash[nodename]
 325                         else:
 326                                 values['comon_stats'] = {'resptime':  '-1',
 327                                                                                 'uptime':    '-1',
 328                                                                                 'sshstatus': '-1',
 329                                                                                 'lastcotop': '-1',
 330                                                                                 'cpuspeed' : "null",
 331                                                                                 'disksize' : 'null',
 332                                                                                 'memsize'  : 'null'}
 333                         # include output value
 334                         ### GET PLC NODE ######################
 335                         plc_lock.acquire()
 336                         d_node = None
 337                         try:
 338                                 d_node = plccache.GetNodeByName(nodename)
 339                                 #d_node = plc.getNodes({'hostname': nodename}, ['pcu_ids', 'site_id',
 340                                 #                                               'date_created', 'last_updated',
 341                                 #                                               'last_contact', 'boot_state', 'nodegroup_ids'])[0]
 342                         except:
 343                                 traceback.print_exc()
 344                         plc_lock.release()
 345                         values['plc_node_stats'] = d_node
 346
 347                         ##### NMAP  ###################
 348                         (n, v) = self.collectNMAP(nodename, None)
 349                         values.update(v)
 350
 351                         ### GET PLC PCU ######################
 352                         site_id = -1
 353                         d_pcu = None
 354                         if d_node:
 355                                 pcu = d_node['pcu_ids']
 356                                 if len(pcu) > 0:
 357                                         d_pcu = pcu[0]
 358
 359                                 site_id = d_node['site_id']
 360
 361                         values['plc_pcuid'] = d_pcu
 362
 363                         ### GET PLC SITE ######################
 364                         plc_lock.acquire()
 365                         d_site = None
 366                         values['loginbase'] = ""
 367                         try:
 368                                 d_site = plccache.GetSitesById([ site_id ])[0]
 369                                 #d_site = plc.getSites({'site_id': site_id},
 370                                 #                                       ['max_slices', 'slice_ids', 'node_ids', 'login_base'])[0]
 371                                 values['loginbase'] = d_site['login_base']
 372                         except:
 373                                 traceback.print_exc()
 374                         plc_lock.release()
 375
 376                         values['plc_site_stats'] = d_site
 377                         values['date_checked'] = datetime.now()
 378                 except:
 379                         print traceback.print_exc()
 380
 381                 return (nodename, values)
 382
 383 def internalprobe(hostname):
 384         #fbsync = FindbadNodeRecordSync.findby_or_create(hostname="global",
 385         #                                                                                               if_new_set={'round' : 1})
 386         scannode = ScanNodeInternal() # fbsync.round)
 387         try:
 388                 (nodename, values) = scannode.collectInternal(hostname, {})
 389                 scannode.record(None, (nodename, values))
 390                 session.flush()
 391                 return True
 392         except:
 393                 print traceback.print_exc()
 394                 return False
 395
 396 def externalprobe(hostname):
 397         #fbsync = FindbadNodeRecordSync.findby_or_create(hostname="global",
 398         #                                                                                               if_new_set={'round' : 1})
 399         scannode = ScanNodeInternal() # fbsync.round)
 400         try:
 401                 (nodename, values) = scannode.collectNMAP(hostname, {})
 402                 scannode.record(None, (nodename, values))
 403                 session.flush()
 404                 return True
 405         except:
 406                 print traceback.print_exc()
 407                 return False
 408
 409 class ScanPCU(ScanInterface):
 410         recordclass = FindbadPCURecord
 411         syncclass = None
 412         primarykey = 'plc_pcuid'
 413
 414         def collectInternal(self, pcuname, cohash):
 415
 416                 continue_probe = True
 417                 errors = None
 418                 values = {'reboot_trial_status' : 'novalue'}
 419                 ### GET PCU ######################
 420                 try:
 421                         b_except = False
 422                         try:
 423                                 v = get_plc_pcu_values(pcuname)
 424                                 if v['hostname'] is not None: v['hostname'] = v['hostname'].strip()
 425                                 if v['ip'] is not None: v['ip'] = v['ip'].strip()
 426
 427                                 if v is not None:
 428                                         values['plc_pcu_stats'] = v
 429                                 else:
 430                                         continue_probe = False
 431                         except:
 432                                 b_except = True
 433                                 traceback.print_exc()
 434                                 continue_probe = False
 435
 436                         if b_except or not continue_probe: return (None, None, None)
 437
 438                         #### RUN NMAP ###############################
 439                         if continue_probe:
 440                                 nmap = command.CMD()
 441                                 print "nmap -oG - -P0 -p22,23,80,443,5869,9100,16992 %s | grep Host:" % reboot.pcu_name(values['plc_pcu_stats'])
 442                                 (oval,eval) = nmap.run_noexcept("nmap -oG - -P0 -p22,23,80,443,5869,9100,16992 %s | grep Host:" % reboot.pcu_name(values['plc_pcu_stats']))
 443                                 # NOTE: an empty / error value for oval, will still work.
 444                                 (values['port_status'], continue_probe) = nmap_port_status(oval)
 445                         else:
 446                                 values['port_status'] = None
 447
 448                         #### COMPLETE ENTRY   #######################
 449
 450                         values['entry_complete'] = []
 451                         #if values['protocol'] is None or values['protocol'] is "":
 452                         #       values['entry_complete'] += ["protocol"]
 453                         if values['plc_pcu_stats']['model'] is None or values['plc_pcu_stats']['model'] is "":
 454                                 values['entry_complete'] += ["model"]
 455                                 # Cannot continue due to this condition
 456                                 continue_probe = False
 457
 458                         if values['plc_pcu_stats']['password'] is None or values['plc_pcu_stats']['password'] is "":
 459                                 values['entry_complete'] += ["password"]
 460                                 # Cannot continue due to this condition
 461                                 continue_probe = False
 462
 463                         if len(values['entry_complete']) > 0:
 464                                 continue_probe = False
 465
 466                         if values['plc_pcu_stats']['hostname'] is None or values['plc_pcu_stats']['hostname'] is "":
 467                                 values['entry_complete'] += ["hostname"]
 468                         if values['plc_pcu_stats']['ip'] is None or values['plc_pcu_stats']['ip'] is "":
 469                                 values['entry_complete'] += ["ip"]
 470
 471                         # If there are no nodes associated with this PCU, then we cannot continue.
 472                         if len(values['plc_pcu_stats']['node_ids']) == 0:
 473                                 continue_probe = False
 474                                 values['entry_complete'] += ['nodeids']
 475
 476
 477                         #### DNS and IP MATCH #######################
 478                         if values['plc_pcu_stats']['hostname'] is not None and values['plc_pcu_stats']['hostname'] is not "" and \
 479                            values['plc_pcu_stats']['ip'] is not None and values['plc_pcu_stats']['ip'] is not "":
 480                                 try:
 481                                         ipaddr = socket.gethostbyname(values['plc_pcu_stats']['hostname'])
 482                                         if ipaddr == values['plc_pcu_stats']['ip']:
 483                                                 values['dns_status'] = "DNS-OK"
 484                                         else:
 485                                                 values['dns_status'] = "DNS-MISMATCH"
 486                                                 continue_probe = False
 487
 488                                 except Exception, err:
 489                                         values['dns_status'] = "DNS-NOENTRY"
 490                                         values['plc_pcu_stats']['hostname'] = values['plc_pcu_stats']['ip']
 491                         else:
 492                                 if values['plc_pcu_stats']['ip'] is not None and values['plc_pcu_stats']['ip'] is not "":
 493                                         values['dns_status'] = "NOHOSTNAME"
 494                                         values['plc_pcu_stats']['hostname'] = values['plc_pcu_stats']['ip']
 495                                 else:
 496                                         values['dns_status'] = "NO-DNS-OR-IP"
 497                                         values['plc_pcu_stats']['hostname'] = "No_entry_in_DB"
 498                                         continue_probe = False
 499
 500
 501                         ######  DRY RUN  ############################
 502                         if continue_probe and 'node_ids' in values['plc_pcu_stats'] and \
 503                                 len(values['plc_pcu_stats']['node_ids']) > 0:
 504                                 rb_ret = reboot.reboot_test_new(values['plc_pcu_stats']['nodenames'][0],
 505                                                                                                 values, 1, True)
 506                         else:
 507                                 rb_ret = "Not_Run" # No nodes to test"
 508
 509                         values['reboot_trial_status'] = rb_ret
 510
 511                 except:
 512                         print "____________________________________"
 513                         print values
 514                         errors = values
 515                         print "____________________________________"
 516                         errors['traceback'] = traceback.format_exc()
 517                         print errors['traceback']
 518                         values['reboot_trial_status'] = str(errors['traceback'])
 519                         print values
 520
 521                 values['entry_complete']=" ".join(values['entry_complete'])
 522
 523                 values['date_checked'] = datetime.now()
 524                 return (pcuname, values)
 525