findbadpcu.py

   1 #!/usr/bin/python
   2
   3 import os
   4 import sys
   5 import string
   6 import time
   7 import socket
   8
   9
  10 import signal
  11
  12 #old_handler = signal.getsignal(signal.SIGCHLD)
  13
  14 #def sig_handler(signum, stack):
  15 #       """ Handle SIGCHLD signal """
  16 #       global old_handler
  17 #       if signum == signal.SIGCHLD:
  18 #               try:
  19 #                       os.wait()
  20 #               except:
  21 #                       pass
  22 #       if old_handler != signal.SIG_DFL:
  23 #               old_handler(signum, stack)
  24 #
  25 #orig_sig_handler = signal.signal(signal.SIGCHLD, sig_handler)
  26
  27 from config import config
  28 from optparse import OptionParser
  29 parser = OptionParser()
  30 parser.set_defaults(filename="",
  31                                         increment=False,
  32                                         dbname="findbadpcus",
  33                                         cachenodes=False,
  34                                         refresh=False,
  35                                         )
  36 parser.add_option("-f", "--nodelist", dest="filename", metavar="FILE",
  37                                         help="Provide the input file for the node list")
  38 parser.add_option("", "--cachenodes", action="store_true",
  39                                         help="Cache node lookup from PLC")
  40 parser.add_option("", "--dbname", dest="dbname", metavar="FILE",
  41                                         help="Specify the name of the database to which the information is saved")
  42 parser.add_option("", "--refresh", action="store_true", dest="refresh",
  43                                         help="Refresh the cached values")
  44 parser.add_option("-i", "--increment", action="store_true", dest="increment",
  45                                         help="Increment round number to force refresh or retry")
  46 config = config(parser)
  47 config.parse_args()
  48
  49 # QUERY all nodes.
  50 COMON_COTOPURL= "http://summer.cs.princeton.edu/status/tabulator.cgi?" + \
  51                                         "table=table_nodeview&" + \
  52                                     "dumpcols='name,resptime,sshstatus,uptime,lastcotop'&" + \
  53                                     "formatcsv"
  54                                     #"formatcsv&" + \
  55                                         #"select='lastcotop!=0'"
  56
  57 import threading
  58 plc_lock = threading.Lock()
  59 round = 1
  60 externalState = {'round': round, 'nodes': {'a': None}}
  61 count = 0
  62
  63 import reboot
  64 from reboot import pcu_name
  65
  66 import soltesz
  67 import plc
  68 import comon
  69 import threadpool
  70 import syncplcdb
  71
  72 def nmap_portstatus(status):
  73         ps = {}
  74         l_nmap = status.split()
  75         ports = l_nmap[4:]
  76
  77         continue_probe = False
  78         for port in ports:
  79                 results = port.split('/')
  80                 ps[results[0]] = results[1]
  81                 if results[1] == "open":
  82                         continue_probe = True
  83         return (ps, continue_probe)
  84
  85 def collectPingAndSSH(pcuname, cohash):
  86
  87         continue_probe = True
  88         values = {}
  89         ### GET PCU ######################
  90         try:
  91                 b_except = False
  92                 plc_lock.acquire()
  93
  94                 try:
  95                         l_pcu  = plc.GetPCUs({'pcu_id' : pcuname})
  96
  97                         if len(l_pcu) > 0:
  98                                 node_ids = l_pcu[0]['node_ids']
  99                                 l_node = plc.getNodes(node_ids, ['hostname', 'last_contact', 'node_id'])
 100                                 site_id = l_pcu[0]['site_id']
 101
 102                                 values['pcu_id'] = l_pcu[0]['pcu_id']
 103
 104                         if len(l_node) > 0:
 105                                 values['nodenames'] = [node['hostname'] for node in l_node]
 106                                 # NOTE: this is for a dry run later. It doesn't matter which node.
 107                                 values['node_id'] = l_node[0]['node_id']
 108
 109                         if len(l_pcu) > 0:
 110                                 values.update(l_pcu[0])
 111                         else:
 112                                 continue_probe = False
 113
 114                 except:
 115                         b_except = True
 116                         import traceback
 117                         traceback.print_exc()
 118
 119                         continue_probe = False
 120
 121                 plc_lock.release()
 122                 if b_except: return (None, None)
 123
 124                 if values['hostname'] is not None:
 125                         values['hostname'] = values['hostname'].strip()
 126
 127                 if values['ip'] is not None:
 128                         values['ip'] = values['ip'].strip()
 129
 130                 #### COMPLETE ENTRY   #######################
 131
 132                 values['complete_entry'] = []
 133                 if values['protocol'] is None or values['protocol'] is "":
 134                         values['complete_entry'] += ["protocol"]
 135                 if values['model'] is None or values['model'] is "":
 136                         values['complete_entry'] += ["model"]
 137                         # Cannot continue due to this condition
 138                         continue_probe = False
 139
 140                 if values['password'] is None or values['password'] is "":
 141                         values['complete_entry'] += ["password"]
 142                         # Cannot continue due to this condition
 143                         continue_probe = False
 144
 145                 if len(values['complete_entry']) > 0:
 146                         continue_probe = False
 147
 148                 if values['hostname'] is None or values['hostname'] is "":
 149                         values['complete_entry'] += ["hostname"]
 150                 if values['ip'] is None or values['ip'] is "":
 151                         values['complete_entry'] += ["ip"]
 152
 153                 # If there are no nodes associated with this PCU, then we cannot continue.
 154                 if len(values['node_ids']) == 0:
 155                         continue_probe = False
 156                         values['complete_entry'] += ['NoNodeIds']
 157
 158                 #### DNS and IP MATCH #######################
 159                 if values['hostname'] is not None and values['hostname'] is not "" and \
 160                    values['ip'] is not None and values['ip'] is not "":
 161                         #print "Calling socket.gethostbyname(%s)" % values['hostname']
 162                         try:
 163                                 ipaddr = socket.gethostbyname(values['hostname'])
 164                                 if ipaddr == values['ip']:
 165                                         values['dnsmatch'] = "DNS-OK"
 166                                 else:
 167                                         values['dnsmatch'] = "DNS-MISMATCH"
 168                                         continue_probe = False
 169
 170                         except Exception, err:
 171                                 values['dnsmatch'] = "DNS-NOENTRY"
 172                                 values['hostname'] = values['ip']
 173                                 #print err
 174                 else:
 175                         if values['ip'] is not None and values['ip'] is not "":
 176                                 values['dnsmatch'] = "NOHOSTNAME"
 177                                 values['hostname'] = values['ip']
 178                         else:
 179                                 values['dnsmatch'] = "NO-DNS-OR-IP"
 180                                 values['hostname'] = "No_entry_in_DB"
 181                                 continue_probe = False
 182
 183                 #### RUN NMAP ###############################
 184                 if continue_probe:
 185                         nmap = soltesz.CMD()
 186                         (oval,eval) = nmap.run_noexcept("nmap -oG - -p22,23,80,443,16992 %s | grep Host:" % pcu_name(values))
 187                         # NOTE: an empty / error value for oval, will still work.
 188                         (values['portstatus'], continue_probe) = nmap_portstatus(oval)
 189
 190                 ######  DRY RUN  ############################
 191                 node_ids = values['node_ids']
 192                 ports = values['ports']
 193                 nid2port = {}
 194                 i = 0
 195
 196                 for id in node_ids:
 197                         nid2port[id] = ports[i]
 198                         i += 1
 199
 200                 # ####
 201                 # TODO: check port status above for whether or not to try...
 202                 # ####
 203                 # DataProbe iPal (many sites)
 204                 if  continue_probe and values['model'].find("Dataprobe IP-41x/IP-81x") >= 0:
 205                         if values['portstatus']['23'] == "open":
 206                                 rb_ret = reboot.ipal_reboot(pcu_name(values),
 207                                                                                 values['password'],
 208                                                                                 nid2port[values['node_id']],
 209                                                                                 True)
 210                         else:
 211                                 rb_ret = "Unsupported_Port"
 212
 213
 214                 # APC Masterswitch (Berkeley)
 215                 elif continue_probe and values['model'].find("APC AP79xx/Masterswitch") >= 0:
 216                         if  values['portstatus']['22'] == "open" or \
 217                                 values['portstatus']['23'] == "open":
 218                                 rb_ret = reboot.apc_reboot(pcu_name(values),
 219                                                                                 values['username'],
 220                                                                                 values['password'],
 221                                                                                 nid2port[values['node_id']],
 222                                                                                 values['portstatus'],
 223                                                                                 True)
 224                         else:
 225                                 rb_ret = "Unsupported_Port"
 226                 # BayTech DS4-RPC
 227                 elif continue_probe and values['model'].find("Baytech DS4-RPC") >= 0:
 228                         if values['portstatus']['22'] == "open":
 229                                 rb_ret = reboot.baytech_reboot(pcu_name(values),
 230                                                                                    values['username'],
 231                                                                                    values['password'],
 232                                                                                    nid2port[values['node_id']],
 233                                                                                    True)
 234                         else:
 235                                 rb_ret = "Unsupported_Port"
 236
 237
 238                 # iLO
 239                 elif continue_probe and values['model'].find("HP iLO") >= 0:
 240                         if values['portstatus']['22'] == "open":
 241                                 rb_ret = reboot.ilo_reboot(pcu_name(values),
 242                                                                                    values['username'],
 243                                                                                    values['password'],
 244                                                                                    True)
 245                         else:
 246                                 rb_ret = "Unsupported_Port"
 247
 248                 # DRAC ssh
 249                 elif continue_probe and values['model'].find("Dell RAC") >= 0:
 250                         if values['portstatus']['22'] == "open":
 251                                 rb_ret = reboot.drac_reboot(pcu_name(values),
 252                                                                                    values['username'],
 253                                                                                    values['password'],
 254                                                                                    True)
 255                         else:
 256                                 rb_ret = "Unsupported_Port"
 257
 258
 259                 # BlackBox PSExxx-xx (e.g. PSE505-FR)
 260                 elif continue_probe and \
 261                         (values['model'].find("BlackBox PS5xx") >= 0 or
 262                          values['model'].find("ePowerSwitch 1/4/8x") >=0 ):
 263                         if values['portstatus']['80'] == "open":
 264                                 rb_ret = reboot.bbpse_reboot(pcu_name(values),
 265                                                                 values['username'],
 266                                                                 values['password'],
 267                                                                 nid2port[values['node_id']],
 268                                                                 80,
 269                                                                 True)
 270                         else:
 271                                 rb_ret = "Unsupported_PCU"
 272
 273                 # x10toggle
 274                 elif    continue_probe and values['protocol'] == "ssh" and \
 275                                 values['model'] == "x10toggle":
 276                         rb_ret = reboot.x10toggle_reboot(pcu_name(values),
 277                                                                                         values['username'],
 278                                                                                         values['password'],
 279                                                                                         nid2port[values['node_id']],
 280                                                                                         True)
 281                 # ????
 282                 elif continue_probe and values['protocol'] == "racadm" and \
 283                                 values['model'] == "RAC":
 284                         rb_ret = reboot.racadm_reboot(pcu_name(values),
 285                                                                                   values['username'],
 286                                                                                   values['password'],
 287                                                                                   nid2port[values['node_id']],
 288                                                                                   True)
 289                 elif continue_probe:
 290                         rb_ret = "Unsupported_PCU"
 291
 292                 elif continue_probe == False:
 293                         if 'portstatus' in values:
 294                                 rb_ret = "NetDown"
 295                         else:
 296                                 rb_ret = "Not_Run"
 297                 else:
 298                         rb_ret = -1
 299
 300                 values['reboot'] = rb_ret
 301
 302                 ### GET PLC SITE ######################
 303                 b_except = False
 304                 plc_lock.acquire()
 305
 306                 try:
 307                         d_site = plc.getSites({'site_id': site_id},
 308                                                                 ['max_slices', 'slice_ids', 'node_ids', 'login_base'])
 309                 except:
 310                         b_except = True
 311                         import traceback
 312                         traceback.print_exc()
 313
 314                 plc_lock.release()
 315                 if b_except: return (None, None)
 316
 317                 if d_site and len(d_site) > 0:
 318                         max_slices = d_site[0]['max_slices']
 319                         num_slices = len(d_site[0]['slice_ids'])
 320                         num_nodes = len(d_site[0]['node_ids'])
 321                         loginbase = d_site[0]['login_base']
 322                         values['plcsite'] = {'num_nodes' : num_nodes,
 323                                                                 'max_slices' : max_slices,
 324                                                                 'num_slices' : num_slices,
 325                                                                 'login_base' : loginbase,
 326                                                                 'status'     : 'SUCCESS'}
 327                 else:
 328                         values['plcsite'] = {'status' : "GS_FAILED"}
 329         except:
 330                 print "____________________________________"
 331                 print values
 332                 print "____________________________________"
 333                 import traceback
 334                 traceback.print_exc()
 335
 336         return (pcuname, values)
 337
 338 def recordPingAndSSH(request, result):
 339         global externalState
 340         global count
 341         (nodename, values) = result
 342
 343         if values is not None:
 344                 global_round = externalState['round']
 345                 pcu_id = "id_%s" % nodename
 346                 externalState['nodes'][pcu_id]['values'] = values
 347                 externalState['nodes'][pcu_id]['round'] = global_round
 348
 349                 count += 1
 350                 print "%d %s %s" % (count, nodename, externalState['nodes'][pcu_id]['values'])
 351                 soltesz.dbDump(config.dbname, externalState)
 352
 353 # this will be called when an exception occurs within a thread
 354 def handle_exception(request, result):
 355         print "Exception occured in request %s" % request.requestID
 356         for i in result:
 357                 print "Result: %s" % i
 358
 359
 360 def checkAndRecordState(l_pcus, cohash):
 361         global externalState
 362         global count
 363         global_round = externalState['round']
 364
 365         tp = threadpool.ThreadPool(20)
 366
 367         # CREATE all the work requests
 368         for pcuname in l_pcus:
 369                 pcu_id = "id_%s" % pcuname
 370                 if pcuname not in externalState['nodes']:
 371                         #print type(externalState['nodes'])
 372
 373                         externalState['nodes'][pcu_id] = {'round': 0, 'values': []}
 374
 375                 node_round   = externalState['nodes'][pcu_id]['round']
 376                 if node_round < global_round:
 377                         # recreate node stats when refreshed
 378                         #print "%s" % nodename
 379                         req = threadpool.WorkRequest(collectPingAndSSH, [pcuname, cohash], {},
 380                                                                                  None, recordPingAndSSH, handle_exception)
 381                         tp.putRequest(req)
 382                 else:
 383                         # We just skip it, since it's "up to date"
 384                         count += 1
 385                         print "%d %s %s" % (count, pcu_id, externalState['nodes'][pcu_id]['values'])
 386                         pass
 387
 388         # WAIT while all the work requests are processed.
 389         while 1:
 390                 try:
 391                         time.sleep(1)
 392                         tp.poll()
 393                 except KeyboardInterrupt:
 394                         print "Interrupted!"
 395                         break
 396                 except threadpool.NoResultsPending:
 397                         print "All results collected."
 398                         break
 399
 400
 401
 402 def main():
 403         global externalState
 404
 405         externalState = soltesz.if_cached_else(1, config.dbname, lambda : externalState)
 406         cohash = {}
 407
 408         if config.increment:
 409                 # update global round number to force refreshes across all nodes
 410                 externalState['round'] += 1
 411
 412         if config.filename == "":
 413                 print "Calling API GetPCUs() : refresh(%s)" % config.refresh
 414                 l_pcus = soltesz.if_cached_else_refresh(1,
 415                                                                 config.refresh, "pculist", lambda : plc.GetPCUs())
 416                 l_pcus  = [pcu['pcu_id'] for pcu in l_pcus]
 417         else:
 418                 l_pcus = config.getListFromFile(config.filename)
 419                 l_pcus = [int(pcu) for pcu in l_pcus]
 420
 421         checkAndRecordState(l_pcus, cohash)
 422
 423         return 0
 424
 425 import logging
 426 logger = logging.getLogger("monitor")
 427 logger.setLevel(logging.DEBUG)
 428 fh = logging.FileHandler("monitor.log", mode = 'a')
 429 fh.setLevel(logging.DEBUG)
 430 formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
 431 fh.setFormatter(formatter)
 432 logger.addHandler(fh)
 433
 434
 435 if __name__ == '__main__':
 436         try:
 437                 # NOTE: evidently, there is a bizarre interaction between iLO and ssh
 438                 # when LANG is set... Do not know why.  Unsetting LANG, fixes the problem.
 439                 if 'LANG' in os.environ:
 440                         del os.environ['LANG']
 441                 main()
 442                 time.sleep(1)
 443         except Exception, err:
 444                 print "Exception: %s" % err
 445                 print "Saving data... exitting."
 446                 soltesz.dbDump(config.dbname, externalState)
 447                 sys.exit(0)