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