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