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