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 import traceback
  12
  13 #old_handler = signal.getsignal(signal.SIGCHLD)
  14
  15 #def sig_handler(signum, stack):
  16 #       """ Handle SIGCHLD signal """
  17 #       global old_handler
  18 #       if signum == signal.SIGCHLD:
  19 #               try:
  20 #                       os.wait()
  21 #               except:
  22 #                       pass
  23 #       if old_handler != signal.SIG_DFL:
  24 #               old_handler(signum, stack)
  25 #
  26 #orig_sig_handler = signal.signal(signal.SIGCHLD, sig_handler)
  27
  28
  29 # QUERY all nodes.
  30 COMON_COTOPURL= "http://summer.cs.princeton.edu/status/tabulator.cgi?" + \
  31                                         "table=table_nodeview&" + \
  32                                     "dumpcols='name,resptime,sshstatus,uptime,lastcotop'&" + \
  33                                     "formatcsv"
  34                                     #"formatcsv&" + \
  35                                         #"select='lastcotop!=0'"
  36
  37 import threading
  38 plc_lock = threading.Lock()
  39 round = 1
  40 externalState = {'round': round, 'nodes': {'a': None}}
  41 errorState = {}
  42 count = 0
  43
  44 import reboot
  45 from reboot import pcu_name
  46
  47 import database
  48 import moncommands
  49 import plc
  50 import comon
  51 import threadpool
  52 import syncplcdb
  53
  54 def nmap_portstatus(status):
  55         ps = {}
  56         l_nmap = status.split()
  57         ports = l_nmap[4:]
  58
  59         continue_probe = False
  60         for port in ports:
  61                 results = port.split('/')
  62                 ps[results[0]] = results[1]
  63                 if results[1] == "open":
  64                         continue_probe = True
  65         return (ps, continue_probe)
  66
  67 def get_pcu(pcuname):
  68         plc_lock.acquire()
  69         try:
  70                 print "GetPCU from PLC %s" % pcuname
  71                 l_pcu  = plc.GetPCUs({'pcu_id' : pcuname})
  72                 print l_pcu
  73                 if len(l_pcu) > 0:
  74                         l_pcu = l_pcu[0]
  75         except:
  76                 try:
  77                         print "GetPCU from file %s" % pcuname
  78                         l_pcus = database.dbLoad("pculist")
  79                         for i in l_pcus:
  80                                 if i['pcu_id'] == pcuname:
  81                                         l_pcu = i
  82                 except:
  83                         traceback.print_exc()
  84                         l_pcu = None
  85
  86         plc_lock.release()
  87         return l_pcu
  88
  89 def get_nodes(node_ids):
  90         plc_lock.acquire()
  91         l_node = []
  92         try:
  93                 l_node = plc.getNodes(node_ids, ['hostname', 'last_contact', 'node_id', 'ports'])
  94         except:
  95                 try:
  96                         plc_nodes = database.dbLoad("l_plcnodes")
  97                         for n in plc_nodes:
  98                                 if n['node_id'] in node_ids:
  99                                         l_node.append(n)
 100                 except:
 101                         traceback.print_exc()
 102                         l_node = None
 103
 104         plc_lock.release()
 105         if l_node == []:
 106                 l_node = None
 107         return l_node
 108
 109
 110 def get_plc_pcu_values(pcuname):
 111         """
 112                 Try to contact PLC to get the PCU info.
 113                 If that fails, try a backup copy from the last run.
 114                 If that fails, return None
 115         """
 116         values = {}
 117
 118         l_pcu = get_pcu(pcuname)
 119
 120         if l_pcu is not None:
 121                 site_id = l_pcu['site_id']
 122                 node_ids = l_pcu['node_ids']
 123                 l_node = get_nodes(node_ids)
 124
 125                 if l_node is not None:
 126                         for node in l_node:
 127                                 values[node['hostname']] = node['ports'][0]
 128
 129                         values['nodenames'] = [node['hostname'] for node in l_node]
 130
 131                         # NOTE: this is for a dry run later. It doesn't matter which node.
 132                         values['node_id'] = l_node[0]['node_id']
 133
 134                 values.update(l_pcu)
 135         else:
 136                 values = None
 137
 138         return values
 139
 140 def get_plc_site_values(site_id):
 141         ### GET PLC SITE ######################
 142         plc_lock.acquire()
 143         values = {}
 144         d_site = None
 145
 146         try:
 147                 d_site = plc.getSites({'site_id': site_id}, ['max_slices', 'slice_ids', 'node_ids', 'login_base'])
 148                 if len(d_site) > 0:
 149                         d_site = d_site[0]
 150         except:
 151                 try:
 152                         plc_sites = database.dbLoad("l_plcsites")
 153                         for site in plc_sites:
 154                                 if site['site_id'] == site_id:
 155                                         d_site = site
 156                                         break
 157                 except:
 158                         traceback.print_exc()
 159                         values = None
 160
 161         plc_lock.release()
 162
 163         if d_site is not None:
 164                 max_slices = d_site['max_slices']
 165                 num_slices = len(d_site['slice_ids'])
 166                 num_nodes = len(d_site['node_ids'])
 167                 loginbase = d_site['login_base']
 168                 values['plcsite'] = {'num_nodes' : num_nodes,
 169                                                         'max_slices' : max_slices,
 170                                                         'num_slices' : num_slices,
 171                                                         'login_base' : loginbase,
 172                                                         'status'     : 'SUCCESS'}
 173         else:
 174                 values = None
 175
 176
 177         return values
 178
 179
 180 def collectPingAndSSH(pcuname, cohash):
 181
 182         continue_probe = True
 183         errors = None
 184         values = {}
 185         ### GET PCU ######################
 186         try:
 187                 b_except = False
 188                 try:
 189                         v = get_plc_pcu_values(pcuname)
 190                         if v is not None:
 191                                 values.update(v)
 192                         else:
 193                                 continue_probe = False
 194                 except:
 195                         b_except = True
 196                         traceback.print_exc()
 197                         continue_probe = False
 198
 199                 if b_except or not continue_probe: return (None, None, None)
 200
 201                 if values['hostname'] is not None:
 202                         values['hostname'] = values['hostname'].strip()
 203
 204                 if values['ip'] is not None:
 205                         values['ip'] = values['ip'].strip()
 206
 207                 #### COMPLETE ENTRY   #######################
 208
 209                 values['complete_entry'] = []
 210                 #if values['protocol'] is None or values['protocol'] is "":
 211                 #       values['complete_entry'] += ["protocol"]
 212                 if values['model'] is None or values['model'] is "":
 213                         values['complete_entry'] += ["model"]
 214                         # Cannot continue due to this condition
 215                         continue_probe = False
 216
 217                 if values['password'] is None or values['password'] is "":
 218                         values['complete_entry'] += ["password"]
 219                         # Cannot continue due to this condition
 220                         continue_probe = False
 221
 222                 if len(values['complete_entry']) > 0:
 223                         continue_probe = False
 224
 225                 if values['hostname'] is None or values['hostname'] is "":
 226                         values['complete_entry'] += ["hostname"]
 227                 if values['ip'] is None or values['ip'] is "":
 228                         values['complete_entry'] += ["ip"]
 229
 230                 # If there are no nodes associated with this PCU, then we cannot continue.
 231                 if len(values['node_ids']) == 0:
 232                         continue_probe = False
 233                         values['complete_entry'] += ['NoNodeIds']
 234
 235                 #### DNS and IP MATCH #######################
 236                 if values['hostname'] is not None and values['hostname'] is not "" and \
 237                    values['ip'] is not None and values['ip'] is not "":
 238                         #print "Calling socket.gethostbyname(%s)" % values['hostname']
 239                         try:
 240                                 ipaddr = socket.gethostbyname(values['hostname'])
 241                                 if ipaddr == values['ip']:
 242                                         values['dnsmatch'] = "DNS-OK"
 243                                 else:
 244                                         values['dnsmatch'] = "DNS-MISMATCH"
 245                                         continue_probe = False
 246
 247                         except Exception, err:
 248                                 values['dnsmatch'] = "DNS-NOENTRY"
 249                                 values['hostname'] = values['ip']
 250                                 #print err
 251                 else:
 252                         if values['ip'] is not None and values['ip'] is not "":
 253                                 values['dnsmatch'] = "NOHOSTNAME"
 254                                 values['hostname'] = values['ip']
 255                         else:
 256                                 values['dnsmatch'] = "NO-DNS-OR-IP"
 257                                 values['hostname'] = "No_entry_in_DB"
 258                                 continue_probe = False
 259
 260                 #### RUN NMAP ###############################
 261                 if continue_probe:
 262                         nmap = moncommands.CMD()
 263                         (oval,eval) = nmap.run_noexcept("nmap -oG - -P0 -p22,23,80,443,5869,9100,16992 %s | grep Host:" % pcu_name(values))
 264                         # NOTE: an empty / error value for oval, will still work.
 265                         (values['portstatus'], continue_probe) = nmap_portstatus(oval)
 266                 else:
 267                         values['portstatus'] = None
 268
 269
 270                 ######  DRY RUN  ############################
 271                 if 'node_ids' in values and len(values['node_ids']) > 0:
 272                         rb_ret = reboot.reboot_test(values['nodenames'][0], values, continue_probe, 1, True)
 273                 else:
 274                         rb_ret = "Not_Run" # No nodes to test"
 275
 276                 values['reboot'] = rb_ret
 277
 278                 ### GET PLC SITE ######################
 279                 v = get_plc_site_values(values['site_id'])
 280                 if v is not None:
 281                         values.update(v)
 282                 else:
 283                         values['plcsite'] = {'status' : "GS_FAILED"}
 284
 285         except:
 286                 print "____________________________________"
 287                 print values
 288                 errors = values
 289                 print "____________________________________"
 290                 errors['traceback'] = traceback.format_exc()
 291                 print errors['traceback']
 292
 293         values['checked'] = time.time()
 294         return (pcuname, values, errors)
 295
 296 def recordPingAndSSH(request, result):
 297         global errorState
 298         global externalState
 299         global count
 300         (nodename, values, errors) = result
 301
 302         if values is not None:
 303                 global_round = externalState['round']
 304                 pcu_id = "id_%s" % nodename
 305                 externalState['nodes'][pcu_id]['values'] = values
 306                 externalState['nodes'][pcu_id]['round'] = global_round
 307
 308                 count += 1
 309                 print "%d %s %s" % (count, nodename, externalState['nodes'][pcu_id]['values'])
 310                 database.dbDump(config.dbname, externalState)
 311
 312         if errors is not None:
 313                 pcu_id = "id_%s" % nodename
 314                 errorState[pcu_id] = errors
 315                 database.dbDump("findbadpcu_errors", errorState)
 316
 317 # this will be called when an exception occurs within a thread
 318 def handle_exception(request, result):
 319         print "Exception occured in request %s" % request.requestID
 320         for i in result:
 321                 print "Result: %s" % i
 322
 323
 324 def checkAndRecordState(l_pcus, cohash):
 325         global externalState
 326         global count
 327         global_round = externalState['round']
 328
 329         tp = threadpool.ThreadPool(20)
 330
 331         # CREATE all the work requests
 332         for pcuname in l_pcus:
 333                 pcu_id = "id_%s" % pcuname
 334                 if pcuname not in externalState['nodes']:
 335                         #print type(externalState['nodes'])
 336
 337                         externalState['nodes'][pcu_id] = {'round': 0, 'values': []}
 338
 339                 node_round   = externalState['nodes'][pcu_id]['round']
 340                 if node_round < global_round:
 341                         # recreate node stats when refreshed
 342                         #print "%s" % nodename
 343                         req = threadpool.WorkRequest(collectPingAndSSH, [pcuname, cohash], {},
 344                                                                                  None, recordPingAndSSH, handle_exception)
 345                         tp.putRequest(req)
 346                 else:
 347                         # We just skip it, since it's "up to date"
 348                         count += 1
 349                         print "%d %s %s" % (count, pcu_id, externalState['nodes'][pcu_id]['values'])
 350                         pass
 351
 352         # WAIT while all the work requests are processed.
 353         begin = time.time()
 354         while 1:
 355                 try:
 356                         time.sleep(1)
 357                         tp.poll()
 358                         # if more than two hours
 359                         if time.time() - begin > (60*60*1):
 360                                 print "findbadpcus.py has run out of time!!!!!!"
 361                                 database.dbDump(config.dbname, externalState)
 362                                 os._exit(1)
 363                 except KeyboardInterrupt:
 364                         print "Interrupted!"
 365                         break
 366                 except threadpool.NoResultsPending:
 367                         print "All results collected."
 368                         break
 369
 370
 371
 372 def main():
 373         global externalState
 374
 375         l_pcus = database.if_cached_else_refresh(1, config.refresh, "pculist", lambda : plc.GetPCUs())
 376         externalState = database.if_cached_else(1, config.dbname, lambda : externalState)
 377         cohash = {}
 378
 379         if config.increment:
 380                 # update global round number to force refreshes across all nodes
 381                 externalState['round'] += 1
 382
 383         if config.nodelist == None and config.pcuid == None:
 384                 print "Calling API GetPCUs() : refresh(%s)" % config.refresh
 385                 l_pcus  = [pcu['pcu_id'] for pcu in l_pcus]
 386         elif config.nodelist is not None:
 387                 l_pcus = config.getListFromFile(config.nodelist)
 388                 l_pcus = [int(pcu) for pcu in l_pcus]
 389         elif config.pcuid is not None:
 390                 l_pcus = [ config.pcuid ]
 391                 l_pcus = [int(pcu) for pcu in l_pcus]
 392
 393         checkAndRecordState(l_pcus, cohash)
 394
 395         return 0
 396
 397
 398 if __name__ == '__main__':
 399         import logging
 400         logger = logging.getLogger("monitor")
 401         logger.setLevel(logging.DEBUG)
 402         fh = logging.FileHandler("monitor.log", mode = 'a')
 403         fh.setLevel(logging.DEBUG)
 404         formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
 405         fh.setFormatter(formatter)
 406         logger.addHandler(fh)
 407         import parser as parsermodule
 408         parser = parsermodule.getParser()
 409         parser.set_defaults(nodelist=None,
 410                                                 increment=False,
 411                                                 pcuid=None,
 412                                                 dbname="findbadpcus",
 413                                                 cachenodes=False,
 414                                                 refresh=False,
 415                                                 )
 416         parser.add_option("-f", "--nodelist", dest="nodelist", metavar="FILE",
 417                                                 help="Provide the input file for the node list")
 418         parser.add_option("", "--pcuid", dest="pcuid", metavar="id",
 419                                                 help="Provide the id for a single pcu")
 420         parser.add_option("", "--cachenodes", action="store_true",
 421                                                 help="Cache node lookup from PLC")
 422         parser.add_option("", "--dbname", dest="dbname", metavar="FILE",
 423                                                 help="Specify the name of the database to which the information is saved")
 424         parser.add_option("", "--refresh", action="store_true", dest="refresh",
 425                                                 help="Refresh the cached values")
 426         parser.add_option("-i", "--increment", action="store_true", dest="increment",
 427                                                 help="Increment round number to force refresh or retry")
 428         parser = parsermodule.getParser(['defaults'], parser)
 429         config = parsermodule.parse_args(parser)
 430         try:
 431                 # NOTE: evidently, there is a bizarre interaction between iLO and ssh
 432                 # when LANG is set... Do not know why.  Unsetting LANG, fixes the problem.
 433                 if 'LANG' in os.environ:
 434                         del os.environ['LANG']
 435                 main()
 436                 time.sleep(1)
 437         except Exception, err:
 438                 traceback.print_exc()
 439                 print "Exception: %s" % err
 440                 print "Saving data... exitting."
 441                 database.dbDump(config.dbname, externalState)
 442                 sys.exit(0)