findbadpcu.py

   1 #!/usr/bin/python
   2
   3 import os
   4 import sys
   5 import string
   6 import time
   7 import socket
   8 import util.file
   9 import plc
  10 import sets
  11
  12
  13 import signal
  14 import traceback
  15
  16 #old_handler = signal.getsignal(signal.SIGCHLD)
  17
  18 #def sig_handler(signum, stack):
  19 #       """ Handle SIGCHLD signal """
  20 #       global old_handler
  21 #       if signum == signal.SIGCHLD:
  22 #               try:
  23 #                       os.wait()
  24 #               except:
  25 #                       pass
  26 #       if old_handler != signal.SIG_DFL:
  27 #               old_handler(signum, stack)
  28 #
  29 #orig_sig_handler = signal.signal(signal.SIGCHLD, sig_handler)
  30
  31
  32 # QUERY all nodes.
  33 COMON_COTOPURL= "http://summer.cs.princeton.edu/status/tabulator.cgi?" + \
  34                                         "table=table_nodeview&" + \
  35                                     "dumpcols='name,resptime,sshstatus,uptime,lastcotop'&" + \
  36                                     "formatcsv"
  37                                     #"formatcsv&" + \
  38                                         #"select='lastcotop!=0'"
  39
  40 import threading
  41 plc_lock = threading.Lock()
  42 round = 1
  43 externalState = {'round': round, 'nodes': {'a': None}}
  44 errorState = {}
  45 count = 0
  46
  47 import reboot
  48 from reboot import pcu_name
  49
  50 import database
  51 import moncommands
  52 import plc
  53 import comon
  54 import threadpool
  55 import syncplcdb
  56
  57 def nmap_portstatus(status):
  58         ps = {}
  59         l_nmap = status.split()
  60         ports = l_nmap[4:]
  61
  62         continue_probe = False
  63         for port in ports:
  64                 results = port.split('/')
  65                 ps[results[0]] = results[1]
  66                 if results[1] == "open":
  67                         continue_probe = True
  68         return (ps, continue_probe)
  69
  70 def get_pcu(pcuname):
  71         plc_lock.acquire()
  72         try:
  73                 print "GetPCU from PLC %s" % pcuname
  74                 l_pcu  = plc.GetPCUs({'pcu_id' : pcuname})
  75                 print l_pcu
  76                 if len(l_pcu) > 0:
  77                         l_pcu = l_pcu[0]
  78         except:
  79                 try:
  80                         print "GetPCU from file %s" % pcuname
  81                         l_pcus = database.dbLoad("pculist")
  82                         for i in l_pcus:
  83                                 if i['pcu_id'] == pcuname:
  84                                         l_pcu = i
  85                 except:
  86                         traceback.print_exc()
  87                         l_pcu = None
  88
  89         plc_lock.release()
  90         return l_pcu
  91
  92 def get_nodes(node_ids):
  93         plc_lock.acquire()
  94         l_node = []
  95         try:
  96                 l_node = plc.getNodes(node_ids, ['hostname', 'last_contact', 'node_id', 'ports'])
  97         except:
  98                 try:
  99                         plc_nodes = database.dbLoad("l_plcnodes")
 100                         for n in plc_nodes:
 101                                 if n['node_id'] in node_ids:
 102                                         l_node.append(n)
 103                 except:
 104                         traceback.print_exc()
 105                         l_node = None
 106
 107         plc_lock.release()
 108         if l_node == []:
 109                 l_node = None
 110         return l_node
 111
 112
 113 def get_plc_pcu_values(pcuname):
 114         """
 115                 Try to contact PLC to get the PCU info.
 116                 If that fails, try a backup copy from the last run.
 117                 If that fails, return None
 118         """
 119         values = {}
 120
 121         l_pcu = get_pcu(pcuname)
 122
 123         if l_pcu is not None:
 124                 site_id = l_pcu['site_id']
 125                 node_ids = l_pcu['node_ids']
 126                 l_node = get_nodes(node_ids)
 127
 128                 if l_node is not None:
 129                         for node in l_node:
 130                                 values[node['hostname']] = node['ports'][0]
 131
 132                         values['nodenames'] = [node['hostname'] for node in l_node]
 133
 134                         # NOTE: this is for a dry run later. It doesn't matter which node.
 135                         values['node_id'] = l_node[0]['node_id']
 136
 137                 values.update(l_pcu)
 138         else:
 139                 values = None
 140
 141         return values
 142
 143 def get_plc_site_values(site_id):
 144         ### GET PLC SITE ######################
 145         plc_lock.acquire()
 146         values = {}
 147         d_site = None
 148
 149         try:
 150                 d_site = plc.getSites({'site_id': site_id}, ['max_slices', 'slice_ids', 'node_ids', 'login_base'])
 151                 if len(d_site) > 0:
 152                         d_site = d_site[0]
 153         except:
 154                 try:
 155                         plc_sites = database.dbLoad("l_plcsites")
 156                         for site in plc_sites:
 157                                 if site['site_id'] == site_id:
 158                                         d_site = site
 159                                         break
 160                 except:
 161                         traceback.print_exc()
 162                         values = None
 163
 164         plc_lock.release()
 165
 166         if d_site is not None:
 167                 max_slices = d_site['max_slices']
 168                 num_slices = len(d_site['slice_ids'])
 169                 num_nodes = len(d_site['node_ids'])
 170                 loginbase = d_site['login_base']
 171                 values['plcsite'] = {'num_nodes' : num_nodes,
 172                                                         'max_slices' : max_slices,
 173                                                         'num_slices' : num_slices,
 174                                                         'login_base' : loginbase,
 175                                                         'status'     : 'SUCCESS'}
 176         else:
 177                 values = None
 178
 179
 180         return values
 181
 182
 183 def collectPingAndSSH(pcuname, cohash):
 184
 185         continue_probe = True
 186         errors = None
 187         values = {}
 188         ### GET PCU ######################
 189         try:
 190                 b_except = False
 191                 try:
 192                         v = get_plc_pcu_values(pcuname)
 193                         if v is not None:
 194                                 values.update(v)
 195                         else:
 196                                 continue_probe = False
 197                 except:
 198                         b_except = True
 199                         traceback.print_exc()
 200                         continue_probe = False
 201
 202                 if b_except or not continue_probe: return (None, None, None)
 203
 204                 if values['hostname'] is not None:
 205                         values['hostname'] = values['hostname'].strip()
 206
 207                 if values['ip'] is not None:
 208                         values['ip'] = values['ip'].strip()
 209
 210                 #### COMPLETE ENTRY   #######################
 211
 212                 values['complete_entry'] = []
 213                 #if values['protocol'] is None or values['protocol'] is "":
 214                 #       values['complete_entry'] += ["protocol"]
 215                 if values['model'] is None or values['model'] is "":
 216                         values['complete_entry'] += ["model"]
 217                         # Cannot continue due to this condition
 218                         continue_probe = False
 219
 220                 if values['password'] is None or values['password'] is "":
 221                         values['complete_entry'] += ["password"]
 222                         # Cannot continue due to this condition
 223                         continue_probe = False
 224
 225                 if len(values['complete_entry']) > 0:
 226                         continue_probe = False
 227
 228                 if values['hostname'] is None or values['hostname'] is "":
 229                         values['complete_entry'] += ["hostname"]
 230                 if values['ip'] is None or values['ip'] is "":
 231                         values['complete_entry'] += ["ip"]
 232
 233                 # If there are no nodes associated with this PCU, then we cannot continue.
 234                 if len(values['node_ids']) == 0:
 235                         continue_probe = False
 236                         values['complete_entry'] += ['NoNodeIds']
 237
 238                 #### DNS and IP MATCH #######################
 239                 if values['hostname'] is not None and values['hostname'] is not "" and \
 240                    values['ip'] is not None and values['ip'] is not "":
 241                         #print "Calling socket.gethostbyname(%s)" % values['hostname']
 242                         try:
 243                                 ipaddr = socket.gethostbyname(values['hostname'])
 244                                 if ipaddr == values['ip']:
 245                                         values['dnsmatch'] = "DNS-OK"
 246                                 else:
 247                                         values['dnsmatch'] = "DNS-MISMATCH"
 248                                         continue_probe = False
 249
 250                         except Exception, err:
 251                                 values['dnsmatch'] = "DNS-NOENTRY"
 252                                 values['hostname'] = values['ip']
 253                                 #print err
 254                 else:
 255                         if values['ip'] is not None and values['ip'] is not "":
 256                                 values['dnsmatch'] = "NOHOSTNAME"
 257                                 values['hostname'] = values['ip']
 258                         else:
 259                                 values['dnsmatch'] = "NO-DNS-OR-IP"
 260                                 values['hostname'] = "No_entry_in_DB"
 261                                 continue_probe = False
 262
 263                 #### RUN NMAP ###############################
 264                 if continue_probe:
 265                         nmap = moncommands.CMD()
 266                         (oval,eval) = nmap.run_noexcept("nmap -oG - -P0 -p22,23,80,443,5869,9100,16992 %s | grep Host:" % pcu_name(values))
 267                         # NOTE: an empty / error value for oval, will still work.
 268                         (values['portstatus'], continue_probe) = nmap_portstatus(oval)
 269                 else:
 270                         values['portstatus'] = None
 271
 272
 273                 ######  DRY RUN  ############################
 274                 if 'node_ids' in values and len(values['node_ids']) > 0:
 275                         rb_ret = reboot.reboot_test(values['nodenames'][0], values, continue_probe, 1, True)
 276                 else:
 277                         rb_ret = "Not_Run" # No nodes to test"
 278
 279                 values['reboot'] = rb_ret
 280
 281                 ### GET PLC SITE ######################
 282                 v = get_plc_site_values(values['site_id'])
 283                 if v is not None:
 284                         values.update(v)
 285                 else:
 286                         values['plcsite'] = {'status' : "GS_FAILED"}
 287
 288         except:
 289                 print "____________________________________"
 290                 print values
 291                 errors = values
 292                 print "____________________________________"
 293                 errors['traceback'] = traceback.format_exc()
 294                 print errors['traceback']
 295
 296         values['checked'] = time.time()
 297         return (pcuname, values, errors)
 298
 299 def recordPingAndSSH(request, result):
 300         global errorState
 301         global externalState
 302         global count
 303         (nodename, values, errors) = result
 304
 305         if values is not None:
 306                 global_round = externalState['round']
 307                 pcu_id = "id_%s" % nodename
 308                 externalState['nodes'][pcu_id]['values'] = values
 309                 externalState['nodes'][pcu_id]['round'] = global_round
 310
 311                 count += 1
 312                 print "%d %s %s" % (count, nodename, externalState['nodes'][pcu_id]['values'])
 313                 database.dbDump(config.dbname, externalState)
 314
 315         if errors is not None:
 316                 pcu_id = "id_%s" % nodename
 317                 errorState[pcu_id] = errors
 318                 database.dbDump("findbadpcu_errors", errorState)
 319
 320 # this will be called when an exception occurs within a thread
 321 def handle_exception(request, result):
 322         print "Exception occured in request %s" % request.requestID
 323         for i in result:
 324                 print "Result: %s" % i
 325
 326
 327 def checkAndRecordState(l_pcus, cohash):
 328         global externalState
 329         global count
 330         global_round = externalState['round']
 331
 332         tp = threadpool.ThreadPool(20)
 333
 334         # CREATE all the work requests
 335         for pcuname in l_pcus:
 336                 pcu_id = "id_%s" % pcuname
 337                 if pcuname not in externalState['nodes']:
 338                         #print type(externalState['nodes'])
 339
 340                         externalState['nodes'][pcu_id] = {'round': 0, 'values': []}
 341
 342                 node_round   = externalState['nodes'][pcu_id]['round']
 343                 if node_round < global_round:
 344                         # recreate node stats when refreshed
 345                         #print "%s" % nodename
 346                         req = threadpool.WorkRequest(collectPingAndSSH, [pcuname, cohash], {},
 347                                                                                  None, recordPingAndSSH, handle_exception)
 348                         tp.putRequest(req)
 349                 else:
 350                         # We just skip it, since it's "up to date"
 351                         count += 1
 352                         print "%d %s %s" % (count, pcu_id, externalState['nodes'][pcu_id]['values'])
 353                         pass
 354
 355         # WAIT while all the work requests are processed.
 356         begin = time.time()
 357         while 1:
 358                 try:
 359                         time.sleep(1)
 360                         tp.poll()
 361                         # if more than two hours
 362                         if time.time() - begin > (60*60*1):
 363                                 print "findbadpcus.py has run out of time!!!!!!"
 364                                 database.dbDump(config.dbname, externalState)
 365                                 os._exit(1)
 366                 except KeyboardInterrupt:
 367                         print "Interrupted!"
 368                         break
 369                 except threadpool.NoResultsPending:
 370                         print "All results collected."
 371                         break
 372
 373
 374
 375 def main():
 376         global externalState
 377
 378         l_pcus = database.if_cached_else_refresh(1, config.refresh, "pculist", lambda : plc.GetPCUs())
 379         externalState = database.if_cached_else(1, config.dbname, lambda : externalState)
 380         cohash = {}
 381
 382         if config.increment:
 383                 # update global round number to force refreshes across all nodes
 384                 externalState['round'] += 1
 385
 386         if config.site is not None:
 387                 api = plc.getAuthAPI()
 388                 site = api.GetSites(config.site)
 389                 l_nodes = api.GetNodes(site[0]['node_ids'], ['pcu_ids'])
 390                 pcus = []
 391                 for node in l_nodes:
 392                         pcus += node['pcu_ids']
 393                 l_pcus = [pcu for pcu in sets.Set(pcus)]
 394
 395         elif config.nodelist == None and config.pcuid == None:
 396                 print "Calling API GetPCUs() : refresh(%s)" % config.refresh
 397                 l_pcus  = [pcu['pcu_id'] for pcu in l_pcus]
 398         elif config.nodelist is not None:
 399                 l_pcus = util.file.getListFromFile(config.nodelist)
 400                 l_pcus = [int(pcu) for pcu in l_pcus]
 401         elif config.pcuid is not None:
 402                 l_pcus = [ config.pcuid ]
 403                 l_pcus = [int(pcu) for pcu in l_pcus]
 404
 405         checkAndRecordState(l_pcus, cohash)
 406
 407         return 0
 408
 409
 410 if __name__ == '__main__':
 411         import logging
 412         logger = logging.getLogger("monitor")
 413         logger.setLevel(logging.DEBUG)
 414         fh = logging.FileHandler("monitor.log", mode = 'a')
 415         fh.setLevel(logging.DEBUG)
 416         formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
 417         fh.setFormatter(formatter)
 418         logger.addHandler(fh)
 419         import parser as parsermodule
 420         parser = parsermodule.getParser()
 421         parser.set_defaults(nodelist=None,
 422                                                 increment=False,
 423                                                 pcuid=None,
 424                                                 site=None,
 425                                                 dbname="findbadpcus",
 426                                                 cachenodes=False,
 427                                                 refresh=False,
 428                                                 )
 429         parser.add_option("-f", "--nodelist", dest="nodelist", metavar="FILE",
 430                                                 help="Provide the input file for the node list")
 431         parser.add_option("", "--site", dest="site", metavar="FILE",
 432                                                 help="Get all pcus associated with the given site's nodes")
 433         parser.add_option("", "--pcuid", dest="pcuid", metavar="id",
 434                                                 help="Provide the id for a single pcu")
 435
 436         parser.add_option("", "--cachenodes", action="store_true",
 437                                                 help="Cache node lookup from PLC")
 438         parser.add_option("", "--dbname", dest="dbname", metavar="FILE",
 439                                                 help="Specify the name of the database to which the information is saved")
 440         parser.add_option("", "--refresh", action="store_true", dest="refresh",
 441                                                 help="Refresh the cached values")
 442         parser.add_option("-i", "--increment", action="store_true", dest="increment",
 443                                                 help="Increment round number to force refresh or retry")
 444         parser = parsermodule.getParser(['defaults'], parser)
 445         config = parsermodule.parse_args(parser)
 446         try:
 447                 # NOTE: evidently, there is a bizarre interaction between iLO and ssh
 448                 # when LANG is set... Do not know why.  Unsetting LANG, fixes the problem.
 449                 if 'LANG' in os.environ:
 450                         del os.environ['LANG']
 451                 main()
 452                 time.sleep(1)
 453         except Exception, err:
 454                 traceback.print_exc()
 455                 print "Exception: %s" % err
 456                 print "Saving data... exitting."
 457                 database.dbDump(config.dbname, externalState)
 458                 sys.exit(0)